sfc: Move and rename Falcon/Siena common NIC operations
authorBen Hutchings <bhutchings@solarflare.com>
Wed, 21 Aug 2013 18:51:04 +0000 (19:51 +0100)
committerBen Hutchings <bhutchings@solarflare.com>
Wed, 21 Aug 2013 19:19:05 +0000 (20:19 +0100)
Add efx_nic_type operations for the many efx_nic functions that need
to be implemented different on EF10.  For now, change most of the
existing efx_nic_*() functions into inline wrappers.  As a later step,
we may be able to improve branch prediction for operations used on the
fast path by copying the pointers into each queue/channel structure.

Move the Falcon/Siena implementations to new file farch.c and rename
the functions and static data to use a prefix of 'efx_farch_'.

Move efx_may_push_tx_desc() to nic.h, as the EF10 TX code will also
use it.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
drivers/net/ethernet/sfc/Makefile
drivers/net/ethernet/sfc/efx.c
drivers/net/ethernet/sfc/falcon.c
drivers/net/ethernet/sfc/farch.c [new file with mode: 0644]
drivers/net/ethernet/sfc/net_driver.h
drivers/net/ethernet/sfc/nic.c
drivers/net/ethernet/sfc/nic.h
drivers/net/ethernet/sfc/siena.c
drivers/net/ethernet/sfc/siena_sriov.c

index 5b31d8a4ae5e886d1aae3e56402c5a63461e1277..ef7410f014d64220735020f01a75873c05d388cd 100644 (file)
@@ -1,4 +1,5 @@
-sfc-y                  += efx.o nic.o falcon.o siena.o tx.o rx.o filter.o \
+sfc-y                  += efx.o nic.o farch.o falcon.o siena.o tx.o rx.o \
+                          filter.o \
                           selftest.o ethtool.o qt202x_phy.o mdio_10g.o \
                           tenxpress.o txc43128_phy.o falcon_boards.o \
                           mcdi.o mcdi_port.o mcdi_mon.o ptp.o
index 9c6555c12acf40cf51de820c8b0bc81b16c32d95..872b9f5b38a3118c374b529b5bce48d0d1bac900 100644 (file)
@@ -1386,7 +1386,7 @@ static void efx_enable_interrupts(struct efx_nic *efx)
                efx->eeh_disabled_legacy_irq = false;
        }
 
-       efx_nic_enable_interrupts(efx);
+       efx->type->irq_enable_master(efx);
 
        efx_for_each_channel(channel, efx) {
                if (channel->type->keep_eventq)
@@ -1407,7 +1407,7 @@ static void efx_disable_interrupts(struct efx_nic *efx)
                        efx_fini_eventq(channel);
        }
 
-       efx_nic_disable_interrupts(efx);
+       efx->type->irq_disable_non_ev(efx);
 }
 
 static void efx_remove_interrupts(struct efx_nic *efx)
index c8efcb0efdedfe6bb01aabab7ac4eb4fbf067b9c..fe83c26c4b8a90201f319dc99495e6ce02a9ca43 100644 (file)
@@ -346,7 +346,7 @@ static inline void falcon_irq_ack_a1(struct efx_nic *efx)
 }
 
 
-irqreturn_t falcon_legacy_interrupt_a1(int irq, void *dev_id)
+static irqreturn_t falcon_legacy_interrupt_a1(int irq, void *dev_id)
 {
        struct efx_nic *efx = dev_id;
        efx_oword_t *int_ker = efx->irq_status.addr;
@@ -373,7 +373,7 @@ irqreturn_t falcon_legacy_interrupt_a1(int irq, void *dev_id)
        /* Check to see if we have a serious error condition */
        syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
        if (unlikely(syserr))
-               return efx_nic_fatal_interrupt(efx);
+               return efx_farch_fatal_interrupt(efx);
 
        /* Determine interrupting queues, clear interrupt status
         * register and acknowledge the device interrupt.
@@ -1558,7 +1558,7 @@ static int falcon_test_nvram(struct efx_nic *efx)
        return falcon_read_nvram(efx, NULL);
 }
 
-static const struct efx_nic_register_test falcon_b0_register_tests[] = {
+static const struct efx_farch_register_test falcon_b0_register_tests[] = {
        { FR_AZ_ADR_REGION,
          EFX_OWORD32(0x0003FFFF, 0x0003FFFF, 0x0003FFFF, 0x0003FFFF) },
        { FR_AZ_RX_CFG,
@@ -1618,8 +1618,8 @@ falcon_b0_test_chip(struct efx_nic *efx, struct efx_self_tests *tests)
        efx_reset_down(efx, reset_method);
 
        tests->registers =
-               efx_nic_test_registers(efx, falcon_b0_register_tests,
-                                      ARRAY_SIZE(falcon_b0_register_tests))
+               efx_farch_test_registers(efx, falcon_b0_register_tests,
+                                        ARRAY_SIZE(falcon_b0_register_tests))
                ? -1 : 1;
 
        rc = falcon_reset_hw(efx, reset_method);
@@ -1984,7 +1984,7 @@ static int falcon_probe_nic(struct efx_nic *efx)
 
        rc = -ENODEV;
 
-       if (efx_nic_fpga_ver(efx) != 0) {
+       if (efx_farch_fpga_ver(efx) != 0) {
                netif_err(efx, probe, efx->net_dev,
                          "Falcon FPGA not supported\n");
                goto fail1;
@@ -2218,7 +2218,7 @@ static int falcon_init_nic(struct efx_nic *efx)
                efx_writeo(efx, &temp, FR_BZ_DP_CTRL);
        }
 
-       efx_nic_init_common(efx);
+       efx_farch_init_common(efx);
 
        return 0;
 }
@@ -2367,6 +2367,28 @@ const struct efx_nic_type falcon_a1_nic_type = {
        .set_wol = falcon_set_wol,
        .resume_wol = efx_port_dummy_op_void,
        .test_nvram = falcon_test_nvram,
+       .irq_enable_master = efx_farch_irq_enable_master,
+       .irq_test_generate = efx_farch_irq_test_generate,
+       .irq_disable_non_ev = efx_farch_irq_disable_master,
+       .irq_handle_msi = efx_farch_msi_interrupt,
+       .irq_handle_legacy = falcon_legacy_interrupt_a1,
+       .tx_probe = efx_farch_tx_probe,
+       .tx_init = efx_farch_tx_init,
+       .tx_remove = efx_farch_tx_remove,
+       .tx_write = efx_farch_tx_write,
+       .rx_push_indir_table = efx_farch_rx_push_indir_table,
+       .rx_probe = efx_farch_rx_probe,
+       .rx_init = efx_farch_rx_init,
+       .rx_remove = efx_farch_rx_remove,
+       .rx_write = efx_farch_rx_write,
+       .rx_defer_refill = efx_farch_rx_defer_refill,
+       .ev_probe = efx_farch_ev_probe,
+       .ev_init = efx_farch_ev_init,
+       .ev_fini = efx_farch_ev_fini,
+       .ev_remove = efx_farch_ev_remove,
+       .ev_process = efx_farch_ev_process,
+       .ev_read_ack = efx_farch_ev_read_ack,
+       .ev_test_generate = efx_farch_ev_test_generate,
 
        .revision = EFX_REV_FALCON_A1,
        .mem_map_size = 0x20000,
@@ -2414,6 +2436,28 @@ const struct efx_nic_type falcon_b0_nic_type = {
        .resume_wol = efx_port_dummy_op_void,
        .test_chip = falcon_b0_test_chip,
        .test_nvram = falcon_test_nvram,
+       .irq_enable_master = efx_farch_irq_enable_master,
+       .irq_test_generate = efx_farch_irq_test_generate,
+       .irq_disable_non_ev = efx_farch_irq_disable_master,
+       .irq_handle_msi = efx_farch_msi_interrupt,
+       .irq_handle_legacy = efx_farch_legacy_interrupt,
+       .tx_probe = efx_farch_tx_probe,
+       .tx_init = efx_farch_tx_init,
+       .tx_remove = efx_farch_tx_remove,
+       .tx_write = efx_farch_tx_write,
+       .rx_push_indir_table = efx_farch_rx_push_indir_table,
+       .rx_probe = efx_farch_rx_probe,
+       .rx_init = efx_farch_rx_init,
+       .rx_remove = efx_farch_rx_remove,
+       .rx_write = efx_farch_rx_write,
+       .rx_defer_refill = efx_farch_rx_defer_refill,
+       .ev_probe = efx_farch_ev_probe,
+       .ev_init = efx_farch_ev_init,
+       .ev_fini = efx_farch_ev_fini,
+       .ev_remove = efx_farch_ev_remove,
+       .ev_process = efx_farch_ev_process,
+       .ev_read_ack = efx_farch_ev_read_ack,
+       .ev_test_generate = efx_farch_ev_test_generate,
 
        .revision = EFX_REV_FALCON_B0,
        /* Map everything up to and including the RSS indirection
diff --git a/drivers/net/ethernet/sfc/farch.c b/drivers/net/ethernet/sfc/farch.c
new file mode 100644 (file)
index 0000000..c3d07c5
--- /dev/null
@@ -0,0 +1,1781 @@
+/****************************************************************************
+ * Driver for Solarflare Solarstorm network controllers and boards
+ * Copyright 2005-2006 Fen Systems Ltd.
+ * Copyright 2006-2011 Solarflare Communications Inc.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published
+ * by the Free Software Foundation, incorporated herein by reference.
+ */
+
+#include <linux/bitops.h>
+#include <linux/delay.h>
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/module.h>
+#include <linux/seq_file.h>
+#include "net_driver.h"
+#include "bitfield.h"
+#include "efx.h"
+#include "nic.h"
+#include "farch_regs.h"
+#include "io.h"
+#include "workarounds.h"
+
+/* Falcon-architecture (SFC4000 and SFC9000-family) support */
+
+/**************************************************************************
+ *
+ * Configurable values
+ *
+ **************************************************************************
+ */
+
+/* This is set to 16 for a good reason.  In summary, if larger than
+ * 16, the descriptor cache holds more than a default socket
+ * buffer's worth of packets (for UDP we can only have at most one
+ * socket buffer's worth outstanding).  This combined with the fact
+ * that we only get 1 TX event per descriptor cache means the NIC
+ * goes idle.
+ */
+#define TX_DC_ENTRIES 16
+#define TX_DC_ENTRIES_ORDER 1
+
+#define RX_DC_ENTRIES 64
+#define RX_DC_ENTRIES_ORDER 3
+
+/* If EFX_MAX_INT_ERRORS internal errors occur within
+ * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
+ * disable it.
+ */
+#define EFX_INT_ERROR_EXPIRE 3600
+#define EFX_MAX_INT_ERRORS 5
+
+/* Depth of RX flush request fifo */
+#define EFX_RX_FLUSH_COUNT 4
+
+/* Driver generated events */
+#define _EFX_CHANNEL_MAGIC_TEST                0x000101
+#define _EFX_CHANNEL_MAGIC_FILL                0x000102
+#define _EFX_CHANNEL_MAGIC_RX_DRAIN    0x000103
+#define _EFX_CHANNEL_MAGIC_TX_DRAIN    0x000104
+
+#define _EFX_CHANNEL_MAGIC(_code, _data)       ((_code) << 8 | (_data))
+#define _EFX_CHANNEL_MAGIC_CODE(_magic)                ((_magic) >> 8)
+
+#define EFX_CHANNEL_MAGIC_TEST(_channel)                               \
+       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
+#define EFX_CHANNEL_MAGIC_FILL(_rx_queue)                              \
+       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL,                     \
+                          efx_rx_queue_index(_rx_queue))
+#define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue)                          \
+       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN,                 \
+                          efx_rx_queue_index(_rx_queue))
+#define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue)                          \
+       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN,                 \
+                          (_tx_queue)->queue)
+
+static void efx_farch_magic_event(struct efx_channel *channel, u32 magic);
+
+/**************************************************************************
+ *
+ * Hardware access
+ *
+ **************************************************************************/
+
+static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
+                                    unsigned int index)
+{
+       efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
+                       value, index);
+}
+
+static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
+                                    const efx_oword_t *mask)
+{
+       return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
+               ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
+}
+
+int efx_farch_test_registers(struct efx_nic *efx,
+                            const struct efx_farch_register_test *regs,
+                            size_t n_regs)
+{
+       unsigned address = 0, i, j;
+       efx_oword_t mask, imask, original, reg, buf;
+
+       for (i = 0; i < n_regs; ++i) {
+               address = regs[i].address;
+               mask = imask = regs[i].mask;
+               EFX_INVERT_OWORD(imask);
+
+               efx_reado(efx, &original, address);
+
+               /* bit sweep on and off */
+               for (j = 0; j < 128; j++) {
+                       if (!EFX_EXTRACT_OWORD32(mask, j, j))
+                               continue;
+
+                       /* Test this testable bit can be set in isolation */
+                       EFX_AND_OWORD(reg, original, mask);
+                       EFX_SET_OWORD32(reg, j, j, 1);
+
+                       efx_writeo(efx, &reg, address);
+                       efx_reado(efx, &buf, address);
+
+                       if (efx_masked_compare_oword(&reg, &buf, &mask))
+                               goto fail;
+
+                       /* Test this testable bit can be cleared in isolation */
+                       EFX_OR_OWORD(reg, original, mask);
+                       EFX_SET_OWORD32(reg, j, j, 0);
+
+                       efx_writeo(efx, &reg, address);
+                       efx_reado(efx, &buf, address);
+
+                       if (efx_masked_compare_oword(&reg, &buf, &mask))
+                               goto fail;
+               }
+
+               efx_writeo(efx, &original, address);
+       }
+
+       return 0;
+
+fail:
+       netif_err(efx, hw, efx->net_dev,
+                 "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
+                 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
+                 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
+       return -EIO;
+}
+
+/**************************************************************************
+ *
+ * Special buffer handling
+ * Special buffers are used for event queues and the TX and RX
+ * descriptor rings.
+ *
+ *************************************************************************/
+
+/*
+ * Initialise a special buffer
+ *
+ * This will define a buffer (previously allocated via
+ * efx_alloc_special_buffer()) in the buffer table, allowing
+ * it to be used for event queues, descriptor rings etc.
+ */
+static void
+efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
+{
+       efx_qword_t buf_desc;
+       unsigned int index;
+       dma_addr_t dma_addr;
+       int i;
+
+       EFX_BUG_ON_PARANOID(!buffer->buf.addr);
+
+       /* Write buffer descriptors to NIC */
+       for (i = 0; i < buffer->entries; i++) {
+               index = buffer->index + i;
+               dma_addr = buffer->buf.dma_addr + (i * EFX_BUF_SIZE);
+               netif_dbg(efx, probe, efx->net_dev,
+                         "mapping special buffer %d at %llx\n",
+                         index, (unsigned long long)dma_addr);
+               EFX_POPULATE_QWORD_3(buf_desc,
+                                    FRF_AZ_BUF_ADR_REGION, 0,
+                                    FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
+                                    FRF_AZ_BUF_OWNER_ID_FBUF, 0);
+               efx_write_buf_tbl(efx, &buf_desc, index);
+       }
+}
+
+/* Unmaps a buffer and clears the buffer table entries */
+static void
+efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
+{
+       efx_oword_t buf_tbl_upd;
+       unsigned int start = buffer->index;
+       unsigned int end = (buffer->index + buffer->entries - 1);
+
+       if (!buffer->entries)
+               return;
+
+       netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
+                 buffer->index, buffer->index + buffer->entries - 1);
+
+       EFX_POPULATE_OWORD_4(buf_tbl_upd,
+                            FRF_AZ_BUF_UPD_CMD, 0,
+                            FRF_AZ_BUF_CLR_CMD, 1,
+                            FRF_AZ_BUF_CLR_END_ID, end,
+                            FRF_AZ_BUF_CLR_START_ID, start);
+       efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
+}
+
+/*
+ * Allocate a new special buffer
+ *
+ * This allocates memory for a new buffer, clears it and allocates a
+ * new buffer ID range.  It does not write into the buffer table.
+ *
+ * This call will allocate 4KB buffers, since 8KB buffers can't be
+ * used for event queues and descriptor rings.
+ */
+static int efx_alloc_special_buffer(struct efx_nic *efx,
+                                   struct efx_special_buffer *buffer,
+                                   unsigned int len)
+{
+       len = ALIGN(len, EFX_BUF_SIZE);
+
+       if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
+               return -ENOMEM;
+       buffer->entries = len / EFX_BUF_SIZE;
+       BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
+
+       /* Select new buffer ID */
+       buffer->index = efx->next_buffer_table;
+       efx->next_buffer_table += buffer->entries;
+#ifdef CONFIG_SFC_SRIOV
+       BUG_ON(efx_sriov_enabled(efx) &&
+              efx->vf_buftbl_base < efx->next_buffer_table);
+#endif
+
+       netif_dbg(efx, probe, efx->net_dev,
+                 "allocating special buffers %d-%d at %llx+%x "
+                 "(virt %p phys %llx)\n", buffer->index,
+                 buffer->index + buffer->entries - 1,
+                 (u64)buffer->buf.dma_addr, len,
+                 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
+
+       return 0;
+}
+
+static void
+efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
+{
+       if (!buffer->buf.addr)
+               return;
+
+       netif_dbg(efx, hw, efx->net_dev,
+                 "deallocating special buffers %d-%d at %llx+%x "
+                 "(virt %p phys %llx)\n", buffer->index,
+                 buffer->index + buffer->entries - 1,
+                 (u64)buffer->buf.dma_addr, buffer->buf.len,
+                 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
+
+       efx_nic_free_buffer(efx, &buffer->buf);
+       buffer->entries = 0;
+}
+
+/**************************************************************************
+ *
+ * TX path
+ *
+ **************************************************************************/
+
+/* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
+static inline void efx_farch_notify_tx_desc(struct efx_tx_queue *tx_queue)
+{
+       unsigned write_ptr;
+       efx_dword_t reg;
+
+       write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
+       EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
+       efx_writed_page(tx_queue->efx, &reg,
+                       FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
+}
+
+/* Write pointer and first descriptor for TX descriptor ring */
+static inline void efx_farch_push_tx_desc(struct efx_tx_queue *tx_queue,
+                                         const efx_qword_t *txd)
+{
+       unsigned write_ptr;
+       efx_oword_t reg;
+
+       BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
+       BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
+
+       write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
+       EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
+                            FRF_AZ_TX_DESC_WPTR, write_ptr);
+       reg.qword[0] = *txd;
+       efx_writeo_page(tx_queue->efx, &reg,
+                       FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
+}
+
+
+/* For each entry inserted into the software descriptor ring, create a
+ * descriptor in the hardware TX descriptor ring (in host memory), and
+ * write a doorbell.
+ */
+void efx_farch_tx_write(struct efx_tx_queue *tx_queue)
+{
+
+       struct efx_tx_buffer *buffer;
+       efx_qword_t *txd;
+       unsigned write_ptr;
+       unsigned old_write_count = tx_queue->write_count;
+
+       BUG_ON(tx_queue->write_count == tx_queue->insert_count);
+
+       do {
+               write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
+               buffer = &tx_queue->buffer[write_ptr];
+               txd = efx_tx_desc(tx_queue, write_ptr);
+               ++tx_queue->write_count;
+
+               /* Create TX descriptor ring entry */
+               BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
+               EFX_POPULATE_QWORD_4(*txd,
+                                    FSF_AZ_TX_KER_CONT,
+                                    buffer->flags & EFX_TX_BUF_CONT,
+                                    FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
+                                    FSF_AZ_TX_KER_BUF_REGION, 0,
+                                    FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
+       } while (tx_queue->write_count != tx_queue->insert_count);
+
+       wmb(); /* Ensure descriptors are written before they are fetched */
+
+       if (efx_nic_may_push_tx_desc(tx_queue, old_write_count)) {
+               txd = efx_tx_desc(tx_queue,
+                                 old_write_count & tx_queue->ptr_mask);
+               efx_farch_push_tx_desc(tx_queue, txd);
+               ++tx_queue->pushes;
+       } else {
+               efx_farch_notify_tx_desc(tx_queue);
+       }
+}
+
+/* Allocate hardware resources for a TX queue */
+int efx_farch_tx_probe(struct efx_tx_queue *tx_queue)
+{
+       struct efx_nic *efx = tx_queue->efx;
+       unsigned entries;
+
+       entries = tx_queue->ptr_mask + 1;
+       return efx_alloc_special_buffer(efx, &tx_queue->txd,
+                                       entries * sizeof(efx_qword_t));
+}
+
+void efx_farch_tx_init(struct efx_tx_queue *tx_queue)
+{
+       struct efx_nic *efx = tx_queue->efx;
+       efx_oword_t reg;
+
+       /* Pin TX descriptor ring */
+       efx_init_special_buffer(efx, &tx_queue->txd);
+
+       /* Push TX descriptor ring to card */
+       EFX_POPULATE_OWORD_10(reg,
+                             FRF_AZ_TX_DESCQ_EN, 1,
+                             FRF_AZ_TX_ISCSI_DDIG_EN, 0,
+                             FRF_AZ_TX_ISCSI_HDIG_EN, 0,
+                             FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
+                             FRF_AZ_TX_DESCQ_EVQ_ID,
+                             tx_queue->channel->channel,
+                             FRF_AZ_TX_DESCQ_OWNER_ID, 0,
+                             FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
+                             FRF_AZ_TX_DESCQ_SIZE,
+                             __ffs(tx_queue->txd.entries),
+                             FRF_AZ_TX_DESCQ_TYPE, 0,
+                             FRF_BZ_TX_NON_IP_DROP_DIS, 1);
+
+       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
+               int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
+               EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
+               EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
+                                   !csum);
+       }
+
+       efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
+                        tx_queue->queue);
+
+       if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
+               /* Only 128 bits in this register */
+               BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
+
+               efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
+               if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
+                       __clear_bit_le(tx_queue->queue, &reg);
+               else
+                       __set_bit_le(tx_queue->queue, &reg);
+               efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
+       }
+
+       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
+               EFX_POPULATE_OWORD_1(reg,
+                                    FRF_BZ_TX_PACE,
+                                    (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
+                                    FFE_BZ_TX_PACE_OFF :
+                                    FFE_BZ_TX_PACE_RESERVED);
+               efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
+                                tx_queue->queue);
+       }
+}
+
+static void efx_farch_flush_tx_queue(struct efx_tx_queue *tx_queue)
+{
+       struct efx_nic *efx = tx_queue->efx;
+       efx_oword_t tx_flush_descq;
+
+       WARN_ON(atomic_read(&tx_queue->flush_outstanding));
+       atomic_set(&tx_queue->flush_outstanding, 1);
+
+       EFX_POPULATE_OWORD_2(tx_flush_descq,
+                            FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
+                            FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
+       efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
+}
+
+void efx_farch_tx_fini(struct efx_tx_queue *tx_queue)
+{
+       struct efx_nic *efx = tx_queue->efx;
+       efx_oword_t tx_desc_ptr;
+
+       /* Remove TX descriptor ring from card */
+       EFX_ZERO_OWORD(tx_desc_ptr);
+       efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
+                        tx_queue->queue);
+
+       /* Unpin TX descriptor ring */
+       efx_fini_special_buffer(efx, &tx_queue->txd);
+}
+
+/* Free buffers backing TX queue */
+void efx_farch_tx_remove(struct efx_tx_queue *tx_queue)
+{
+       efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
+}
+
+/**************************************************************************
+ *
+ * RX path
+ *
+ **************************************************************************/
+
+/* This creates an entry in the RX descriptor queue */
+static inline void
+efx_farch_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
+{
+       struct efx_rx_buffer *rx_buf;
+       efx_qword_t *rxd;
+
+       rxd = efx_rx_desc(rx_queue, index);
+       rx_buf = efx_rx_buffer(rx_queue, index);
+       EFX_POPULATE_QWORD_3(*rxd,
+                            FSF_AZ_RX_KER_BUF_SIZE,
+                            rx_buf->len -
+                            rx_queue->efx->type->rx_buffer_padding,
+                            FSF_AZ_RX_KER_BUF_REGION, 0,
+                            FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
+}
+
+/* This writes to the RX_DESC_WPTR register for the specified receive
+ * descriptor ring.
+ */
+void efx_farch_rx_write(struct efx_rx_queue *rx_queue)
+{
+       struct efx_nic *efx = rx_queue->efx;
+       efx_dword_t reg;
+       unsigned write_ptr;
+
+       while (rx_queue->notified_count != rx_queue->added_count) {
+               efx_farch_build_rx_desc(
+                       rx_queue,
+                       rx_queue->notified_count & rx_queue->ptr_mask);
+               ++rx_queue->notified_count;
+       }
+
+       wmb();
+       write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
+       EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
+       efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
+                       efx_rx_queue_index(rx_queue));
+}
+
+int efx_farch_rx_probe(struct efx_rx_queue *rx_queue)
+{
+       struct efx_nic *efx = rx_queue->efx;
+       unsigned entries;
+
+       entries = rx_queue->ptr_mask + 1;
+       return efx_alloc_special_buffer(efx, &rx_queue->rxd,
+                                       entries * sizeof(efx_qword_t));
+}
+
+void efx_farch_rx_init(struct efx_rx_queue *rx_queue)
+{
+       efx_oword_t rx_desc_ptr;
+       struct efx_nic *efx = rx_queue->efx;
+       bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
+       bool iscsi_digest_en = is_b0;
+       bool jumbo_en;
+
+       /* For kernel-mode queues in Falcon A1, the JUMBO flag enables
+        * DMA to continue after a PCIe page boundary (and scattering
+        * is not possible).  In Falcon B0 and Siena, it enables
+        * scatter.
+        */
+       jumbo_en = !is_b0 || efx->rx_scatter;
+
+       netif_dbg(efx, hw, efx->net_dev,
+                 "RX queue %d ring in special buffers %d-%d\n",
+                 efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
+                 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
+
+       rx_queue->scatter_n = 0;
+
+       /* Pin RX descriptor ring */
+       efx_init_special_buffer(efx, &rx_queue->rxd);
+
+       /* Push RX descriptor ring to card */
+       EFX_POPULATE_OWORD_10(rx_desc_ptr,
+                             FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
+                             FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
+                             FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
+                             FRF_AZ_RX_DESCQ_EVQ_ID,
+                             efx_rx_queue_channel(rx_queue)->channel,
+                             FRF_AZ_RX_DESCQ_OWNER_ID, 0,
+                             FRF_AZ_RX_DESCQ_LABEL,
+                             efx_rx_queue_index(rx_queue),
+                             FRF_AZ_RX_DESCQ_SIZE,
+                             __ffs(rx_queue->rxd.entries),
+                             FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
+                             FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
+                             FRF_AZ_RX_DESCQ_EN, 1);
+       efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
+                        efx_rx_queue_index(rx_queue));
+}
+
+static void efx_farch_flush_rx_queue(struct efx_rx_queue *rx_queue)
+{
+       struct efx_nic *efx = rx_queue->efx;
+       efx_oword_t rx_flush_descq;
+
+       EFX_POPULATE_OWORD_2(rx_flush_descq,
+                            FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
+                            FRF_AZ_RX_FLUSH_DESCQ,
+                            efx_rx_queue_index(rx_queue));
+       efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
+}
+
+void efx_farch_rx_fini(struct efx_rx_queue *rx_queue)
+{
+       efx_oword_t rx_desc_ptr;
+       struct efx_nic *efx = rx_queue->efx;
+
+       /* Remove RX descriptor ring from card */
+       EFX_ZERO_OWORD(rx_desc_ptr);
+       efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
+                        efx_rx_queue_index(rx_queue));
+
+       /* Unpin RX descriptor ring */
+       efx_fini_special_buffer(efx, &rx_queue->rxd);
+}
+
+/* Free buffers backing RX queue */
+void efx_farch_rx_remove(struct efx_rx_queue *rx_queue)
+{
+       efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
+}
+
+/**************************************************************************
+ *
+ * Flush handling
+ *
+ **************************************************************************/
+
+/* efx_farch_flush_queues() must be woken up when all flushes are completed,
+ * or more RX flushes can be kicked off.
+ */
+static bool efx_farch_flush_wake(struct efx_nic *efx)
+{
+       /* Ensure that all updates are visible to efx_farch_flush_queues() */
+       smp_mb();
+
+       return (atomic_read(&efx->drain_pending) == 0 ||
+               (atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
+                && atomic_read(&efx->rxq_flush_pending) > 0));
+}
+
+static bool efx_check_tx_flush_complete(struct efx_nic *efx)
+{
+       bool i = true;
+       efx_oword_t txd_ptr_tbl;
+       struct efx_channel *channel;
+       struct efx_tx_queue *tx_queue;
+
+       efx_for_each_channel(channel, efx) {
+               efx_for_each_channel_tx_queue(tx_queue, channel) {
+                       efx_reado_table(efx, &txd_ptr_tbl,
+                                       FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
+                       if (EFX_OWORD_FIELD(txd_ptr_tbl,
+                                           FRF_AZ_TX_DESCQ_FLUSH) ||
+                           EFX_OWORD_FIELD(txd_ptr_tbl,
+                                           FRF_AZ_TX_DESCQ_EN)) {
+                               netif_dbg(efx, hw, efx->net_dev,
+                                         "flush did not complete on TXQ %d\n",
+                                         tx_queue->queue);
+                               i = false;
+                       } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
+                                                 1, 0)) {
+                               /* The flush is complete, but we didn't
+                                * receive a flush completion event
+                                */
+                               netif_dbg(efx, hw, efx->net_dev,
+                                         "flush complete on TXQ %d, so drain "
+                                         "the queue\n", tx_queue->queue);
+                               /* Don't need to increment drain_pending as it
+                                * has already been incremented for the queues
+                                * which did not drain
+                                */
+                               efx_farch_magic_event(channel,
+                                                     EFX_CHANNEL_MAGIC_TX_DRAIN(
+                                                             tx_queue));
+                       }
+               }
+       }
+
+       return i;
+}
+
+/* Flush all the transmit queues, and continue flushing receive queues until
+ * they're all flushed. Wait for the DRAIN events to be recieved so that there
+ * are no more RX and TX events left on any channel. */
+static int efx_farch_do_flush(struct efx_nic *efx)
+{
+       unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
+       struct efx_channel *channel;
+       struct efx_rx_queue *rx_queue;
+       struct efx_tx_queue *tx_queue;
+       int rc = 0;
+
+       efx_for_each_channel(channel, efx) {
+               efx_for_each_channel_tx_queue(tx_queue, channel) {
+                       atomic_inc(&efx->drain_pending);
+                       efx_farch_flush_tx_queue(tx_queue);
+               }
+               efx_for_each_channel_rx_queue(rx_queue, channel) {
+                       atomic_inc(&efx->drain_pending);
+                       rx_queue->flush_pending = true;
+                       atomic_inc(&efx->rxq_flush_pending);
+               }
+       }
+
+       while (timeout && atomic_read(&efx->drain_pending) > 0) {
+               /* If SRIOV is enabled, then offload receive queue flushing to
+                * the firmware (though we will still have to poll for
+                * completion). If that fails, fall back to the old scheme.
+                */
+               if (efx_sriov_enabled(efx)) {
+                       rc = efx_mcdi_flush_rxqs(efx);
+                       if (!rc)
+                               goto wait;
+               }
+
+               /* The hardware supports four concurrent rx flushes, each of
+                * which may need to be retried if there is an outstanding
+                * descriptor fetch
+                */
+               efx_for_each_channel(channel, efx) {
+                       efx_for_each_channel_rx_queue(rx_queue, channel) {
+                               if (atomic_read(&efx->rxq_flush_outstanding) >=
+                                   EFX_RX_FLUSH_COUNT)
+                                       break;
+
+                               if (rx_queue->flush_pending) {
+                                       rx_queue->flush_pending = false;
+                                       atomic_dec(&efx->rxq_flush_pending);
+                                       atomic_inc(&efx->rxq_flush_outstanding);
+                                       efx_farch_flush_rx_queue(rx_queue);
+                               }
+                       }
+               }
+
+       wait:
+               timeout = wait_event_timeout(efx->flush_wq,
+                                            efx_farch_flush_wake(efx),
+                                            timeout);
+       }
+
+       if (atomic_read(&efx->drain_pending) &&
+           !efx_check_tx_flush_complete(efx)) {
+               netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
+                         "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
+                         atomic_read(&efx->rxq_flush_outstanding),
+                         atomic_read(&efx->rxq_flush_pending));
+               rc = -ETIMEDOUT;
+
+               atomic_set(&efx->drain_pending, 0);
+               atomic_set(&efx->rxq_flush_pending, 0);
+               atomic_set(&efx->rxq_flush_outstanding, 0);
+       }
+
+       return rc;
+}
+
+int efx_farch_fini_dmaq(struct efx_nic *efx)
+{
+       struct efx_channel *channel;
+       struct efx_tx_queue *tx_queue;
+       struct efx_rx_queue *rx_queue;
+       int rc = 0;
+
+       /* Do not attempt to write to the NIC during EEH recovery */
+       if (efx->state != STATE_RECOVERY) {
+               /* Only perform flush if DMA is enabled */
+               if (efx->pci_dev->is_busmaster) {
+                       efx->type->prepare_flush(efx);
+                       rc = efx_farch_do_flush(efx);
+                       efx->type->finish_flush(efx);
+               }
+
+               efx_for_each_channel(channel, efx) {
+                       efx_for_each_channel_rx_queue(rx_queue, channel)
+                               efx_farch_rx_fini(rx_queue);
+                       efx_for_each_channel_tx_queue(tx_queue, channel)
+                               efx_farch_tx_fini(tx_queue);
+               }
+       }
+
+       return rc;
+}
+
+/**************************************************************************
+ *
+ * Event queue processing
+ * Event queues are processed by per-channel tasklets.
+ *
+ **************************************************************************/
+
+/* Update a channel's event queue's read pointer (RPTR) register
+ *
+ * This writes the EVQ_RPTR_REG register for the specified channel's
+ * event queue.
+ */
+void efx_farch_ev_read_ack(struct efx_channel *channel)
+{
+       efx_dword_t reg;
+       struct efx_nic *efx = channel->efx;
+
+       EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
+                            channel->eventq_read_ptr & channel->eventq_mask);
+
+       /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
+        * of 4 bytes, but it is really 16 bytes just like later revisions.
+        */
+       efx_writed(efx, &reg,
+                  efx->type->evq_rptr_tbl_base +
+                  FR_BZ_EVQ_RPTR_STEP * channel->channel);
+}
+
+/* Use HW to insert a SW defined event */
+void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
+                             efx_qword_t *event)
+{
+       efx_oword_t drv_ev_reg;
+
+       BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
+                    FRF_AZ_DRV_EV_DATA_WIDTH != 64);
+       drv_ev_reg.u32[0] = event->u32[0];
+       drv_ev_reg.u32[1] = event->u32[1];
+       drv_ev_reg.u32[2] = 0;
+       drv_ev_reg.u32[3] = 0;
+       EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
+       efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
+}
+
+static void efx_farch_magic_event(struct efx_channel *channel, u32 magic)
+{
+       efx_qword_t event;
+
+       EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
+                            FSE_AZ_EV_CODE_DRV_GEN_EV,
+                            FSF_AZ_DRV_GEN_EV_MAGIC, magic);
+       efx_farch_generate_event(channel->efx, channel->channel, &event);
+}
+
+/* Handle a transmit completion event
+ *
+ * The NIC batches TX completion events; the message we receive is of
+ * the form "complete all TX events up to this index".
+ */
+static int
+efx_farch_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
+{
+       unsigned int tx_ev_desc_ptr;
+       unsigned int tx_ev_q_label;
+       struct efx_tx_queue *tx_queue;
+       struct efx_nic *efx = channel->efx;
+       int tx_packets = 0;
+
+       if (unlikely(ACCESS_ONCE(efx->reset_pending)))
+               return 0;
+
+       if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
+               /* Transmit completion */
+               tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
+               tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
+               tx_queue = efx_channel_get_tx_queue(
+                       channel, tx_ev_q_label % EFX_TXQ_TYPES);
+               tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
+                             tx_queue->ptr_mask);
+               efx_xmit_done(tx_queue, tx_ev_desc_ptr);
+       } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
+               /* Rewrite the FIFO write pointer */
+               tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
+               tx_queue = efx_channel_get_tx_queue(
+                       channel, tx_ev_q_label % EFX_TXQ_TYPES);
+
+               netif_tx_lock(efx->net_dev);
+               efx_farch_notify_tx_desc(tx_queue);
+               netif_tx_unlock(efx->net_dev);
+       } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
+                  EFX_WORKAROUND_10727(efx)) {
+               efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
+       } else {
+               netif_err(efx, tx_err, efx->net_dev,
+                         "channel %d unexpected TX event "
+                         EFX_QWORD_FMT"\n", channel->channel,
+                         EFX_QWORD_VAL(*event));
+       }
+
+       return tx_packets;
+}
+
+/* Detect errors included in the rx_evt_pkt_ok bit. */
+static u16 efx_farch_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
+                                     const efx_qword_t *event)
+{
+       struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
+       struct efx_nic *efx = rx_queue->efx;
+       bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
+       bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
+       bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
+       bool rx_ev_other_err, rx_ev_pause_frm;
+       bool rx_ev_hdr_type, rx_ev_mcast_pkt;
+       unsigned rx_ev_pkt_type;
+
+       rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
+       rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
+       rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
+       rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
+       rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
+                                                FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
+       rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
+                                                 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
+       rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
+                                                  FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
+       rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
+       rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
+       rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
+                         0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
+       rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
+
+       /* Every error apart from tobe_disc and pause_frm */
+       rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
+                          rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
+                          rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
+
+       /* Count errors that are not in MAC stats.  Ignore expected
+        * checksum errors during self-test. */
+       if (rx_ev_frm_trunc)
+               ++channel->n_rx_frm_trunc;
+       else if (rx_ev_tobe_disc)
+               ++channel->n_rx_tobe_disc;
+       else if (!efx->loopback_selftest) {
+               if (rx_ev_ip_hdr_chksum_err)
+                       ++channel->n_rx_ip_hdr_chksum_err;
+               else if (rx_ev_tcp_udp_chksum_err)
+                       ++channel->n_rx_tcp_udp_chksum_err;
+       }
+
+       /* TOBE_DISC is expected on unicast mismatches; don't print out an
+        * error message.  FRM_TRUNC indicates RXDP dropped the packet due
+        * to a FIFO overflow.
+        */
+#ifdef DEBUG
+       if (rx_ev_other_err && net_ratelimit()) {
+               netif_dbg(efx, rx_err, efx->net_dev,
+                         " RX queue %d unexpected RX event "
+                         EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
+                         efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
+                         rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
+                         rx_ev_ip_hdr_chksum_err ?
+                         " [IP_HDR_CHKSUM_ERR]" : "",
+                         rx_ev_tcp_udp_chksum_err ?
+                         " [TCP_UDP_CHKSUM_ERR]" : "",
+                         rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
+                         rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
+                         rx_ev_drib_nib ? " [DRIB_NIB]" : "",
+                         rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
+                         rx_ev_pause_frm ? " [PAUSE]" : "");
+       }
+#endif
+
+       /* The frame must be discarded if any of these are true. */
+       return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
+               rx_ev_tobe_disc | rx_ev_pause_frm) ?
+               EFX_RX_PKT_DISCARD : 0;
+}
+
+/* Handle receive events that are not in-order. Return true if this
+ * can be handled as a partial packet discard, false if it's more
+ * serious.
+ */
+static bool
+efx_farch_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
+{
+       struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
+       struct efx_nic *efx = rx_queue->efx;
+       unsigned expected, dropped;
+
+       if (rx_queue->scatter_n &&
+           index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
+                     rx_queue->ptr_mask)) {
+               ++channel->n_rx_nodesc_trunc;
+               return true;
+       }
+
+       expected = rx_queue->removed_count & rx_queue->ptr_mask;
+       dropped = (index - expected) & rx_queue->ptr_mask;
+       netif_info(efx, rx_err, efx->net_dev,
+                  "dropped %d events (index=%d expected=%d)\n",
+                  dropped, index, expected);
+
+       efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
+                          RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
+       return false;
+}
+
+/* Handle a packet received event
+ *
+ * The NIC gives a "discard" flag if it's a unicast packet with the
+ * wrong destination address
+ * Also "is multicast" and "matches multicast filter" flags can be used to
+ * discard non-matching multicast packets.
+ */
+static void
+efx_farch_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
+{
+       unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
+       unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
+       unsigned expected_ptr;
+       bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
+       u16 flags;
+       struct efx_rx_queue *rx_queue;
+       struct efx_nic *efx = channel->efx;
+
+       if (unlikely(ACCESS_ONCE(efx->reset_pending)))
+               return;
+
+       rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
+       rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
+       WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
+               channel->channel);
+
+       rx_queue = efx_channel_get_rx_queue(channel);
+
+       rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
+       expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
+                       rx_queue->ptr_mask);
+
+       /* Check for partial drops and other errors */
+       if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
+           unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
+               if (rx_ev_desc_ptr != expected_ptr &&
+                   !efx_farch_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
+                       return;
+
+               /* Discard all pending fragments */
+               if (rx_queue->scatter_n) {
+                       efx_rx_packet(
+                               rx_queue,
+                               rx_queue->removed_count & rx_queue->ptr_mask,
+                               rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
+                       rx_queue->removed_count += rx_queue->scatter_n;
+                       rx_queue->scatter_n = 0;
+               }
+
+               /* Return if there is no new fragment */
+               if (rx_ev_desc_ptr != expected_ptr)
+                       return;
+
+               /* Discard new fragment if not SOP */
+               if (!rx_ev_sop) {
+                       efx_rx_packet(
+                               rx_queue,
+                               rx_queue->removed_count & rx_queue->ptr_mask,
+                               1, 0, EFX_RX_PKT_DISCARD);
+                       ++rx_queue->removed_count;
+                       return;
+               }
+       }
+
+       ++rx_queue->scatter_n;
+       if (rx_ev_cont)
+               return;
+
+       rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
+       rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
+       rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
+
+       if (likely(rx_ev_pkt_ok)) {
+               /* If packet is marked as OK then we can rely on the
+                * hardware checksum and classification.
+                */
+               flags = 0;
+               switch (rx_ev_hdr_type) {
+               case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
+                       flags |= EFX_RX_PKT_TCP;
+                       /* fall through */
+               case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
+                       flags |= EFX_RX_PKT_CSUMMED;
+                       /* fall through */
+               case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
+               case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
+                       break;
+               }
+       } else {
+               flags = efx_farch_handle_rx_not_ok(rx_queue, event);
+       }
+
+       /* Detect multicast packets that didn't match the filter */
+       rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
+       if (rx_ev_mcast_pkt) {
+               unsigned int rx_ev_mcast_hash_match =
+                       EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
+
+               if (unlikely(!rx_ev_mcast_hash_match)) {
+                       ++channel->n_rx_mcast_mismatch;
+                       flags |= EFX_RX_PKT_DISCARD;
+               }
+       }
+
+       channel->irq_mod_score += 2;
+
+       /* Handle received packet */
+       efx_rx_packet(rx_queue,
+                     rx_queue->removed_count & rx_queue->ptr_mask,
+                     rx_queue->scatter_n, rx_ev_byte_cnt, flags);
+       rx_queue->removed_count += rx_queue->scatter_n;
+       rx_queue->scatter_n = 0;
+}
+
+/* If this flush done event corresponds to a &struct efx_tx_queue, then
+ * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
+ * of all transmit completions.
+ */
+static void
+efx_farch_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
+{
+       struct efx_tx_queue *tx_queue;
+       int qid;
+
+       qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
+       if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
+               tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
+                                           qid % EFX_TXQ_TYPES);
+               if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
+                       efx_farch_magic_event(tx_queue->channel,
+                                             EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
+               }
+       }
+}
+
+/* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
+ * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
+ * the RX queue back to the mask of RX queues in need of flushing.
+ */
+static void
+efx_farch_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
+{
+       struct efx_channel *channel;
+       struct efx_rx_queue *rx_queue;
+       int qid;
+       bool failed;
+
+       qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
+       failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
+       if (qid >= efx->n_channels)
+               return;
+       channel = efx_get_channel(efx, qid);
+       if (!efx_channel_has_rx_queue(channel))
+               return;
+       rx_queue = efx_channel_get_rx_queue(channel);
+
+       if (failed) {
+               netif_info(efx, hw, efx->net_dev,
+                          "RXQ %d flush retry\n", qid);
+               rx_queue->flush_pending = true;
+               atomic_inc(&efx->rxq_flush_pending);
+       } else {
+               efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
+                                     EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
+       }
+       atomic_dec(&efx->rxq_flush_outstanding);
+       if (efx_farch_flush_wake(efx))
+               wake_up(&efx->flush_wq);
+}
+
+static void
+efx_farch_handle_drain_event(struct efx_channel *channel)
+{
+       struct efx_nic *efx = channel->efx;
+
+       WARN_ON(atomic_read(&efx->drain_pending) == 0);
+       atomic_dec(&efx->drain_pending);
+       if (efx_farch_flush_wake(efx))
+               wake_up(&efx->flush_wq);
+}
+
+static void efx_farch_handle_generated_event(struct efx_channel *channel,
+                                            efx_qword_t *event)
+{
+       struct efx_nic *efx = channel->efx;
+       struct efx_rx_queue *rx_queue =
+               efx_channel_has_rx_queue(channel) ?
+               efx_channel_get_rx_queue(channel) : NULL;
+       unsigned magic, code;
+
+       magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
+       code = _EFX_CHANNEL_MAGIC_CODE(magic);
+
+       if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
+               channel->event_test_cpu = raw_smp_processor_id();
+       } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
+               /* The queue must be empty, so we won't receive any rx
+                * events, so efx_process_channel() won't refill the
+                * queue. Refill it here */
+               efx_fast_push_rx_descriptors(rx_queue);
+       } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
+               efx_farch_handle_drain_event(channel);
+       } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
+               efx_farch_handle_drain_event(channel);
+       } else {
+               netif_dbg(efx, hw, efx->net_dev, "channel %d received "
+                         "generated event "EFX_QWORD_FMT"\n",
+                         channel->channel, EFX_QWORD_VAL(*event));
+       }
+}
+
+static void
+efx_farch_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
+{
+       struct efx_nic *efx = channel->efx;
+       unsigned int ev_sub_code;
+       unsigned int ev_sub_data;
+
+       ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
+       ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
+
+       switch (ev_sub_code) {
+       case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
+               netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
+                          channel->channel, ev_sub_data);
+               efx_farch_handle_tx_flush_done(efx, event);
+               efx_sriov_tx_flush_done(efx, event);
+               break;
+       case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
+               netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
+                          channel->channel, ev_sub_data);
+               efx_farch_handle_rx_flush_done(efx, event);
+               efx_sriov_rx_flush_done(efx, event);
+               break;
+       case FSE_AZ_EVQ_INIT_DONE_EV:
+               netif_dbg(efx, hw, efx->net_dev,
+                         "channel %d EVQ %d initialised\n",
+                         channel->channel, ev_sub_data);
+               break;
+       case FSE_AZ_SRM_UPD_DONE_EV:
+               netif_vdbg(efx, hw, efx->net_dev,
+                          "channel %d SRAM update done\n", channel->channel);
+               break;
+       case FSE_AZ_WAKE_UP_EV:
+               netif_vdbg(efx, hw, efx->net_dev,
+                          "channel %d RXQ %d wakeup event\n",
+                          channel->channel, ev_sub_data);
+               break;
+       case FSE_AZ_TIMER_EV:
+               netif_vdbg(efx, hw, efx->net_dev,
+                          "channel %d RX queue %d timer expired\n",
+                          channel->channel, ev_sub_data);
+               break;
+       case FSE_AA_RX_RECOVER_EV:
+               netif_err(efx, rx_err, efx->net_dev,
+                         "channel %d seen DRIVER RX_RESET event. "
+                       "Resetting.\n", channel->channel);
+               atomic_inc(&efx->rx_reset);
+               efx_schedule_reset(efx,
+                                  EFX_WORKAROUND_6555(efx) ?
+                                  RESET_TYPE_RX_RECOVERY :
+                                  RESET_TYPE_DISABLE);
+               break;
+       case FSE_BZ_RX_DSC_ERROR_EV:
+               if (ev_sub_data < EFX_VI_BASE) {
+                       netif_err(efx, rx_err, efx->net_dev,
+                                 "RX DMA Q %d reports descriptor fetch error."
+                                 " RX Q %d is disabled.\n", ev_sub_data,
+                                 ev_sub_data);
+                       efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
+               } else
+                       efx_sriov_desc_fetch_err(efx, ev_sub_data);
+               break;
+       case FSE_BZ_TX_DSC_ERROR_EV:
+               if (ev_sub_data < EFX_VI_BASE) {
+                       netif_err(efx, tx_err, efx->net_dev,
+                                 "TX DMA Q %d reports descriptor fetch error."
+                                 " TX Q %d is disabled.\n", ev_sub_data,
+                                 ev_sub_data);
+                       efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
+               } else
+                       efx_sriov_desc_fetch_err(efx, ev_sub_data);
+               break;
+       default:
+               netif_vdbg(efx, hw, efx->net_dev,
+                          "channel %d unknown driver event code %d "
+                          "data %04x\n", channel->channel, ev_sub_code,
+                          ev_sub_data);
+               break;
+       }
+}
+
+int efx_farch_ev_process(struct efx_channel *channel, int budget)
+{
+       struct efx_nic *efx = channel->efx;
+       unsigned int read_ptr;
+       efx_qword_t event, *p_event;
+       int ev_code;
+       int tx_packets = 0;
+       int spent = 0;
+
+       read_ptr = channel->eventq_read_ptr;
+
+       for (;;) {
+               p_event = efx_event(channel, read_ptr);
+               event = *p_event;
+
+               if (!efx_event_present(&event))
+                       /* End of events */
+                       break;
+
+               netif_vdbg(channel->efx, intr, channel->efx->net_dev,
+                          "channel %d event is "EFX_QWORD_FMT"\n",
+                          channel->channel, EFX_QWORD_VAL(event));
+
+               /* Clear this event by marking it all ones */
+               EFX_SET_QWORD(*p_event);
+
+               ++read_ptr;
+
+               ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
+
+               switch (ev_code) {
+               case FSE_AZ_EV_CODE_RX_EV:
+                       efx_farch_handle_rx_event(channel, &event);
+                       if (++spent == budget)
+                               goto out;
+                       break;
+               case FSE_AZ_EV_CODE_TX_EV:
+                       tx_packets += efx_farch_handle_tx_event(channel,
+                                                               &event);
+                       if (tx_packets > efx->txq_entries) {
+                               spent = budget;
+                               goto out;
+                       }
+                       break;
+               case FSE_AZ_EV_CODE_DRV_GEN_EV:
+                       efx_farch_handle_generated_event(channel, &event);
+                       break;
+               case FSE_AZ_EV_CODE_DRIVER_EV:
+                       efx_farch_handle_driver_event(channel, &event);
+                       break;
+               case FSE_CZ_EV_CODE_USER_EV:
+                       efx_sriov_event(channel, &event);
+                       break;
+               case FSE_CZ_EV_CODE_MCDI_EV:
+                       efx_mcdi_process_event(channel, &event);
+                       break;
+               case FSE_AZ_EV_CODE_GLOBAL_EV:
+                       if (efx->type->handle_global_event &&
+                           efx->type->handle_global_event(channel, &event))
+                               break;
+                       /* else fall through */
+               default:
+                       netif_err(channel->efx, hw, channel->efx->net_dev,
+                                 "channel %d unknown event type %d (data "
+                                 EFX_QWORD_FMT ")\n", channel->channel,
+                                 ev_code, EFX_QWORD_VAL(event));
+               }
+       }
+
+out:
+       channel->eventq_read_ptr = read_ptr;
+       return spent;
+}
+
+/* Allocate buffer table entries for event queue */
+int efx_farch_ev_probe(struct efx_channel *channel)
+{
+       struct efx_nic *efx = channel->efx;
+       unsigned entries;
+
+       entries = channel->eventq_mask + 1;
+       return efx_alloc_special_buffer(efx, &channel->eventq,
+                                       entries * sizeof(efx_qword_t));
+}
+
+void efx_farch_ev_init(struct efx_channel *channel)
+{
+       efx_oword_t reg;
+       struct efx_nic *efx = channel->efx;
+
+       netif_dbg(efx, hw, efx->net_dev,
+                 "channel %d event queue in special buffers %d-%d\n",
+                 channel->channel, channel->eventq.index,
+                 channel->eventq.index + channel->eventq.entries - 1);
+
+       if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
+               EFX_POPULATE_OWORD_3(reg,
+                                    FRF_CZ_TIMER_Q_EN, 1,
+                                    FRF_CZ_HOST_NOTIFY_MODE, 0,
+                                    FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
+               efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
+       }
+
+       /* Pin event queue buffer */
+       efx_init_special_buffer(efx, &channel->eventq);
+
+       /* Fill event queue with all ones (i.e. empty events) */
+       memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
+
+       /* Push event queue to card */
+       EFX_POPULATE_OWORD_3(reg,
+                            FRF_AZ_EVQ_EN, 1,
+                            FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
+                            FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
+       efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
+                        channel->channel);
+
+       efx->type->push_irq_moderation(channel);
+}
+
+void efx_farch_ev_fini(struct efx_channel *channel)
+{
+       efx_oword_t reg;
+       struct efx_nic *efx = channel->efx;
+
+       /* Remove event queue from card */
+       EFX_ZERO_OWORD(reg);
+       efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
+                        channel->channel);
+       if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
+               efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
+
+       /* Unpin event queue */
+       efx_fini_special_buffer(efx, &channel->eventq);
+}
+
+/* Free buffers backing event queue */
+void efx_farch_ev_remove(struct efx_channel *channel)
+{
+       efx_free_special_buffer(channel->efx, &channel->eventq);
+}
+
+
+void efx_farch_ev_test_generate(struct efx_channel *channel)
+{
+       efx_farch_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
+}
+
+void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue)
+{
+       efx_farch_magic_event(efx_rx_queue_channel(rx_queue),
+                             EFX_CHANNEL_MAGIC_FILL(rx_queue));
+}
+
+/**************************************************************************
+ *
+ * Hardware interrupts
+ * The hardware interrupt handler does very little work; all the event
+ * queue processing is carried out by per-channel tasklets.
+ *
+ **************************************************************************/
+
+/* Enable/disable/generate interrupts */
+static inline void efx_farch_interrupts(struct efx_nic *efx,
+                                     bool enabled, bool force)
+{
+       efx_oword_t int_en_reg_ker;
+
+       EFX_POPULATE_OWORD_3(int_en_reg_ker,
+                            FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
+                            FRF_AZ_KER_INT_KER, force,
+                            FRF_AZ_DRV_INT_EN_KER, enabled);
+       efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
+}
+
+void efx_farch_irq_enable_master(struct efx_nic *efx)
+{
+       EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
+       wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
+
+       efx_farch_interrupts(efx, true, false);
+}
+
+void efx_farch_irq_disable_master(struct efx_nic *efx)
+{
+       /* Disable interrupts */
+       efx_farch_interrupts(efx, false, false);
+}
+
+/* Generate a test interrupt
+ * Interrupt must already have been enabled, otherwise nasty things
+ * may happen.
+ */
+void efx_farch_irq_test_generate(struct efx_nic *efx)
+{
+       efx_farch_interrupts(efx, true, true);
+}
+
+/* Process a fatal interrupt
+ * Disable bus mastering ASAP and schedule a reset
+ */
+irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx)
+{
+       struct falcon_nic_data *nic_data = efx->nic_data;
+       efx_oword_t *int_ker = efx->irq_status.addr;
+       efx_oword_t fatal_intr;
+       int error, mem_perr;
+
+       efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
+       error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
+
+       netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
+                 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
+                 EFX_OWORD_VAL(fatal_intr),
+                 error ? "disabling bus mastering" : "no recognised error");
+
+       /* If this is a memory parity error dump which blocks are offending */
+       mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
+                   EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
+       if (mem_perr) {
+               efx_oword_t reg;
+               efx_reado(efx, &reg, FR_AZ_MEM_STAT);
+               netif_err(efx, hw, efx->net_dev,
+                         "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
+                         EFX_OWORD_VAL(reg));
+       }
+
+       /* Disable both devices */
+       pci_clear_master(efx->pci_dev);
+       if (efx_nic_is_dual_func(efx))
+               pci_clear_master(nic_data->pci_dev2);
+       efx_farch_irq_disable_master(efx);
+
+       /* Count errors and reset or disable the NIC accordingly */
+       if (efx->int_error_count == 0 ||
+           time_after(jiffies, efx->int_error_expire)) {
+               efx->int_error_count = 0;
+               efx->int_error_expire =
+                       jiffies + EFX_INT_ERROR_EXPIRE * HZ;
+       }
+       if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
+               netif_err(efx, hw, efx->net_dev,
+                         "SYSTEM ERROR - reset scheduled\n");
+               efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
+       } else {
+               netif_err(efx, hw, efx->net_dev,
+                         "SYSTEM ERROR - max number of errors seen."
+                         "NIC will be disabled\n");
+               efx_schedule_reset(efx, RESET_TYPE_DISABLE);
+       }
+
+       return IRQ_HANDLED;
+}
+
+/* Handle a legacy interrupt
+ * Acknowledges the interrupt and schedule event queue processing.
+ */
+irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id)
+{
+       struct efx_nic *efx = dev_id;
+       bool soft_enabled = ACCESS_ONCE(efx->irq_soft_enabled);
+       efx_oword_t *int_ker = efx->irq_status.addr;
+       irqreturn_t result = IRQ_NONE;
+       struct efx_channel *channel;
+       efx_dword_t reg;
+       u32 queues;
+       int syserr;
+
+       /* Read the ISR which also ACKs the interrupts */
+       efx_readd(efx, &reg, FR_BZ_INT_ISR0);
+       queues = EFX_EXTRACT_DWORD(reg, 0, 31);
+
+       /* Legacy interrupts are disabled too late by the EEH kernel
+        * code. Disable them earlier.
+        * If an EEH error occurred, the read will have returned all ones.
+        */
+       if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
+           !efx->eeh_disabled_legacy_irq) {
+               disable_irq_nosync(efx->legacy_irq);
+               efx->eeh_disabled_legacy_irq = true;
+       }
+
+       /* Handle non-event-queue sources */
+       if (queues & (1U << efx->irq_level) && soft_enabled) {
+               syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
+               if (unlikely(syserr))
+                       return efx_farch_fatal_interrupt(efx);
+               efx->last_irq_cpu = raw_smp_processor_id();
+       }
+
+       if (queues != 0) {
+               if (EFX_WORKAROUND_15783(efx))
+                       efx->irq_zero_count = 0;
+
+               /* Schedule processing of any interrupting queues */
+               if (likely(soft_enabled)) {
+                       efx_for_each_channel(channel, efx) {
+                               if (queues & 1)
+                                       efx_schedule_channel_irq(channel);
+                               queues >>= 1;
+                       }
+               }
+               result = IRQ_HANDLED;
+
+       } else if (EFX_WORKAROUND_15783(efx)) {
+               efx_qword_t *event;
+
+               /* We can't return IRQ_HANDLED more than once on seeing ISR=0
+                * because this might be a shared interrupt. */
+               if (efx->irq_zero_count++ == 0)
+                       result = IRQ_HANDLED;
+
+               /* Ensure we schedule or rearm all event queues */
+               if (likely(soft_enabled)) {
+                       efx_for_each_channel(channel, efx) {
+                               event = efx_event(channel,
+                                                 channel->eventq_read_ptr);
+                               if (efx_event_present(event))
+                                       efx_schedule_channel_irq(channel);
+                               else
+                                       efx_farch_ev_read_ack(channel);
+                       }
+               }
+       }
+
+       if (result == IRQ_HANDLED)
+               netif_vdbg(efx, intr, efx->net_dev,
+                          "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
+                          irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
+
+       return result;
+}
+
+/* Handle an MSI interrupt
+ *
+ * Handle an MSI hardware interrupt.  This routine schedules event
+ * queue processing.  No interrupt acknowledgement cycle is necessary.
+ * Also, we never need to check that the interrupt is for us, since
+ * MSI interrupts cannot be shared.
+ */
+irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id)
+{
+       struct efx_msi_context *context = dev_id;
+       struct efx_nic *efx = context->efx;
+       efx_oword_t *int_ker = efx->irq_status.addr;
+       int syserr;
+
+       netif_vdbg(efx, intr, efx->net_dev,
+                  "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
+                  irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
+
+       if (!likely(ACCESS_ONCE(efx->irq_soft_enabled)))
+               return IRQ_HANDLED;
+
+       /* Handle non-event-queue sources */
+       if (context->index == efx->irq_level) {
+               syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
+               if (unlikely(syserr))
+                       return efx_farch_fatal_interrupt(efx);
+               efx->last_irq_cpu = raw_smp_processor_id();
+       }
+
+       /* Schedule processing of the channel */
+       efx_schedule_channel_irq(efx->channel[context->index]);
+
+       return IRQ_HANDLED;
+}
+
+
+/* Setup RSS indirection table.
+ * This maps from the hash value of the packet to RXQ
+ */
+void efx_farch_rx_push_indir_table(struct efx_nic *efx)
+{
+       size_t i = 0;
+       efx_dword_t dword;
+
+       if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
+               return;
+
+       BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
+                    FR_BZ_RX_INDIRECTION_TBL_ROWS);
+
+       for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
+               EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
+                                    efx->rx_indir_table[i]);
+               efx_writed(efx, &dword,
+                          FR_BZ_RX_INDIRECTION_TBL +
+                          FR_BZ_RX_INDIRECTION_TBL_STEP * i);
+       }
+}
+
+/* Looks at available SRAM resources and works out how many queues we
+ * can support, and where things like descriptor caches should live.
+ *
+ * SRAM is split up as follows:
+ * 0                          buftbl entries for channels
+ * efx->vf_buftbl_base        buftbl entries for SR-IOV
+ * efx->rx_dc_base            RX descriptor caches
+ * efx->tx_dc_base            TX descriptor caches
+ */
+void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
+{
+       unsigned vi_count, buftbl_min;
+
+       /* Account for the buffer table entries backing the datapath channels
+        * and the descriptor caches for those channels.
+        */
+       buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
+                      efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
+                      efx->n_channels * EFX_MAX_EVQ_SIZE)
+                     * sizeof(efx_qword_t) / EFX_BUF_SIZE);
+       vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
+
+#ifdef CONFIG_SFC_SRIOV
+       if (efx_sriov_wanted(efx)) {
+               unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
+
+               efx->vf_buftbl_base = buftbl_min;
+
+               vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
+               vi_count = max(vi_count, EFX_VI_BASE);
+               buftbl_free = (sram_lim_qw - buftbl_min -
+                              vi_count * vi_dc_entries);
+
+               entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
+                                 efx_vf_size(efx));
+               vf_limit = min(buftbl_free / entries_per_vf,
+                              (1024U - EFX_VI_BASE) >> efx->vi_scale);
+
+               if (efx->vf_count > vf_limit) {
+                       netif_err(efx, probe, efx->net_dev,
+                                 "Reducing VF count from from %d to %d\n",
+                                 efx->vf_count, vf_limit);
+                       efx->vf_count = vf_limit;
+               }
+               vi_count += efx->vf_count * efx_vf_size(efx);
+       }
+#endif
+
+       efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
+       efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
+}
+
+u32 efx_farch_fpga_ver(struct efx_nic *efx)
+{
+       efx_oword_t altera_build;
+       efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
+       return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
+}
+
+void efx_farch_init_common(struct efx_nic *efx)
+{
+       efx_oword_t temp;
+
+       /* Set positions of descriptor caches in SRAM. */
+       EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
+       efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
+       EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
+       efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
+
+       /* Set TX descriptor cache size. */
+       BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
+       EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
+       efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
+
+       /* Set RX descriptor cache size.  Set low watermark to size-8, as
+        * this allows most efficient prefetching.
+        */
+       BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
+       EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
+       efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
+       EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
+       efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
+
+       /* Program INT_KER address */
+       EFX_POPULATE_OWORD_2(temp,
+                            FRF_AZ_NORM_INT_VEC_DIS_KER,
+                            EFX_INT_MODE_USE_MSI(efx),
+                            FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
+       efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
+
+       if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
+               /* Use an interrupt level unused by event queues */
+               efx->irq_level = 0x1f;
+       else
+               /* Use a valid MSI-X vector */
+               efx->irq_level = 0;
+
+       /* Enable all the genuinely fatal interrupts.  (They are still
+        * masked by the overall interrupt mask, controlled by
+        * falcon_interrupts()).
+        *
+        * Note: All other fatal interrupts are enabled
+        */
+       EFX_POPULATE_OWORD_3(temp,
+                            FRF_AZ_ILL_ADR_INT_KER_EN, 1,
+                            FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
+                            FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
+       if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
+               EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
+       EFX_INVERT_OWORD(temp);
+       efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
+
+       efx_farch_rx_push_indir_table(efx);
+
+       /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
+        * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
+        */
+       efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
+       /* Enable SW_EV to inherit in char driver - assume harmless here */
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
+       /* Prefetch threshold 2 => fetch when descriptor cache half empty */
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
+       /* Disable hardware watchdog which can misfire */
+       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
+       /* Squash TX of packets of 16 bytes or less */
+       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
+               EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
+       efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
+
+       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
+               EFX_POPULATE_OWORD_4(temp,
+                                    /* Default values */
+                                    FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
+                                    FRF_BZ_TX_PACE_SB_AF, 0xb,
+                                    FRF_BZ_TX_PACE_FB_BASE, 0,
+                                    /* Allow large pace values in the
+                                     * fast bin. */
+                                    FRF_BZ_TX_PACE_BIN_TH,
+                                    FFE_BZ_TX_PACE_RESERVED);
+               efx_writeo(efx, &temp, FR_BZ_TX_PACE);
+       }
+}
index b382895901b1545976ab199d4a38e926fcbd2509..7283cc1a90fec1a298dcf0bd8d43ebc5b2b8e6be 100644 (file)
@@ -971,7 +971,7 @@ static inline unsigned int efx_port_num(struct efx_nic *efx)
  * @get_wol: Get WoL configuration from driver state
  * @set_wol: Push WoL configuration to the NIC
  * @resume_wol: Synchronise WoL state between driver and MC (e.g. after resume)
- * @test_chip: Test registers.  Should use efx_nic_test_registers(), and is
+ * @test_chip: Test registers.  May use efx_farch_test_registers(), and is
  *     expected to reset the NIC.
  * @test_nvram: Test validity of NVRAM contents
  * @mcdi_request: Send an MCDI request with the given header and SDU.
@@ -985,6 +985,32 @@ static inline unsigned int efx_port_num(struct efx_nic *efx)
  * @mcdi_poll_reboot: Test whether the MCDI has rebooted.  If so,
  *     return an appropriate error code for aborting any current
  *     request; otherwise return 0.
+ * @irq_enable_master: Enable IRQs on the NIC.  Each event queue must
+ *     be separately enabled after this.
+ * @irq_test_generate: Generate a test IRQ
+ * @irq_disable_non_ev: Disable non-event IRQs on the NIC.  Each event
+ *     queue must be separately disabled before this.
+ * @irq_handle_msi: Handle MSI for a channel.  The @dev_id argument is
+ *     a pointer to the &struct efx_msi_context for the channel.
+ * @irq_handle_legacy: Handle legacy interrupt.  The @dev_id argument
+ *     is a pointer to the &struct efx_nic.
+ * @tx_probe: Allocate resources for TX queue
+ * @tx_init: Initialise TX queue on the NIC
+ * @tx_remove: Free resources for TX queue
+ * @tx_write: Write TX descriptors and doorbell
+ * @rx_push_indir_table: Write RSS indirection table to the NIC
+ * @rx_probe: Allocate resources for RX queue
+ * @rx_init: Initialise RX queue on the NIC
+ * @rx_remove: Free resources for RX queue
+ * @rx_write: Write RX descriptors and doorbell
+ * @rx_defer_refill: Generate a refill reminder event
+ * @ev_probe: Allocate resources for event queue
+ * @ev_init: Initialise event queue on the NIC
+ * @ev_fini: Deinitialise event queue on the NIC
+ * @ev_remove: Free resources for event queue
+ * @ev_process: Process events for a queue, up to the given NAPI quota
+ * @ev_read_ack: Acknowledge read events on a queue, rearming its IRQ
+ * @ev_test_generate: Generate a test event
  * @revision: Hardware architecture revision
  * @mem_map_size: Memory BAR mapped size
  * @txd_ptr_tbl_base: TX descriptor ring base address
@@ -1041,6 +1067,28 @@ struct efx_nic_type {
        void (*mcdi_read_response)(struct efx_nic *efx, efx_dword_t *pdu,
                                   size_t pdu_offset, size_t pdu_len);
        int (*mcdi_poll_reboot)(struct efx_nic *efx);
+       void (*irq_enable_master)(struct efx_nic *efx);
+       void (*irq_test_generate)(struct efx_nic *efx);
+       void (*irq_disable_non_ev)(struct efx_nic *efx);
+       irqreturn_t (*irq_handle_msi)(int irq, void *dev_id);
+       irqreturn_t (*irq_handle_legacy)(int irq, void *dev_id);
+       int (*tx_probe)(struct efx_tx_queue *tx_queue);
+       void (*tx_init)(struct efx_tx_queue *tx_queue);
+       void (*tx_remove)(struct efx_tx_queue *tx_queue);
+       void (*tx_write)(struct efx_tx_queue *tx_queue);
+       void (*rx_push_indir_table)(struct efx_nic *efx);
+       int (*rx_probe)(struct efx_rx_queue *rx_queue);
+       void (*rx_init)(struct efx_rx_queue *rx_queue);
+       void (*rx_remove)(struct efx_rx_queue *rx_queue);
+       void (*rx_write)(struct efx_rx_queue *rx_queue);
+       void (*rx_defer_refill)(struct efx_rx_queue *rx_queue);
+       int (*ev_probe)(struct efx_channel *channel);
+       void (*ev_init)(struct efx_channel *channel);
+       void (*ev_fini)(struct efx_channel *channel);
+       void (*ev_remove)(struct efx_channel *channel);
+       int (*ev_process)(struct efx_channel *channel, int quota);
+       void (*ev_read_ack)(struct efx_channel *channel);
+       void (*ev_test_generate)(struct efx_channel *channel);
 
        int revision;
        unsigned int mem_map_size;
index 7c52691e9d26803639bdf34640507d8d56787b00..66c71ede3a98bf9826b6f1e7af576b2aad8a73e5 100644 (file)
 #include "io.h"
 #include "workarounds.h"
 
-/**************************************************************************
- *
- * Configurable values
- *
- **************************************************************************
- */
-
-/* This is set to 16 for a good reason.  In summary, if larger than
- * 16, the descriptor cache holds more than a default socket
- * buffer's worth of packets (for UDP we can only have at most one
- * socket buffer's worth outstanding).  This combined with the fact
- * that we only get 1 TX event per descriptor cache means the NIC
- * goes idle.
- */
-#define TX_DC_ENTRIES 16
-#define TX_DC_ENTRIES_ORDER 1
-
-#define RX_DC_ENTRIES 64
-#define RX_DC_ENTRIES_ORDER 3
-
-/* If EFX_MAX_INT_ERRORS internal errors occur within
- * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
- * disable it.
- */
-#define EFX_INT_ERROR_EXPIRE 3600
-#define EFX_MAX_INT_ERRORS 5
-
-/* Depth of RX flush request fifo */
-#define EFX_RX_FLUSH_COUNT 4
-
-/* Driver generated events */
-#define _EFX_CHANNEL_MAGIC_TEST                0x000101
-#define _EFX_CHANNEL_MAGIC_FILL                0x000102
-#define _EFX_CHANNEL_MAGIC_RX_DRAIN    0x000103
-#define _EFX_CHANNEL_MAGIC_TX_DRAIN    0x000104
-
-#define _EFX_CHANNEL_MAGIC(_code, _data)       ((_code) << 8 | (_data))
-#define _EFX_CHANNEL_MAGIC_CODE(_magic)                ((_magic) >> 8)
-
-#define EFX_CHANNEL_MAGIC_TEST(_channel)                               \
-       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TEST, (_channel)->channel)
-#define EFX_CHANNEL_MAGIC_FILL(_rx_queue)                              \
-       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_FILL,                     \
-                          efx_rx_queue_index(_rx_queue))
-#define EFX_CHANNEL_MAGIC_RX_DRAIN(_rx_queue)                          \
-       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_RX_DRAIN,                 \
-                          efx_rx_queue_index(_rx_queue))
-#define EFX_CHANNEL_MAGIC_TX_DRAIN(_tx_queue)                          \
-       _EFX_CHANNEL_MAGIC(_EFX_CHANNEL_MAGIC_TX_DRAIN,                 \
-                          (_tx_queue)->queue)
-
-static void efx_magic_event(struct efx_channel *channel, u32 magic);
-
-/**************************************************************************
- *
- * Solarstorm hardware access
- *
- **************************************************************************/
-
-static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
-                                    unsigned int index)
-{
-       efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
-                       value, index);
-}
-
-/* Read the current event from the event queue */
-static inline efx_qword_t *efx_event(struct efx_channel *channel,
-                                    unsigned int index)
-{
-       return ((efx_qword_t *) (channel->eventq.buf.addr)) +
-               (index & channel->eventq_mask);
-}
-
-/* See if an event is present
- *
- * We check both the high and low dword of the event for all ones.  We
- * wrote all ones when we cleared the event, and no valid event can
- * have all ones in either its high or low dwords.  This approach is
- * robust against reordering.
- *
- * Note that using a single 64-bit comparison is incorrect; even
- * though the CPU read will be atomic, the DMA write may not be.
- */
-static inline int efx_event_present(efx_qword_t *event)
-{
-       return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
-                 EFX_DWORD_IS_ALL_ONES(event->dword[1]));
-}
-
-static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
-                                    const efx_oword_t *mask)
-{
-       return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
-               ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
-}
-
-int efx_nic_test_registers(struct efx_nic *efx,
-                          const struct efx_nic_register_test *regs,
-                          size_t n_regs)
-{
-       unsigned address = 0, i, j;
-       efx_oword_t mask, imask, original, reg, buf;
-
-       for (i = 0; i < n_regs; ++i) {
-               address = regs[i].address;
-               mask = imask = regs[i].mask;
-               EFX_INVERT_OWORD(imask);
-
-               efx_reado(efx, &original, address);
-
-               /* bit sweep on and off */
-               for (j = 0; j < 128; j++) {
-                       if (!EFX_EXTRACT_OWORD32(mask, j, j))
-                               continue;
-
-                       /* Test this testable bit can be set in isolation */
-                       EFX_AND_OWORD(reg, original, mask);
-                       EFX_SET_OWORD32(reg, j, j, 1);
-
-                       efx_writeo(efx, &reg, address);
-                       efx_reado(efx, &buf, address);
-
-                       if (efx_masked_compare_oword(&reg, &buf, &mask))
-                               goto fail;
-
-                       /* Test this testable bit can be cleared in isolation */
-                       EFX_OR_OWORD(reg, original, mask);
-                       EFX_SET_OWORD32(reg, j, j, 0);
-
-                       efx_writeo(efx, &reg, address);
-                       efx_reado(efx, &buf, address);
-
-                       if (efx_masked_compare_oword(&reg, &buf, &mask))
-                               goto fail;
-               }
-
-               efx_writeo(efx, &original, address);
-       }
-
-       return 0;
-
-fail:
-       netif_err(efx, hw, efx->net_dev,
-                 "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
-                 " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
-                 EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
-       return -EIO;
-}
-
-/**************************************************************************
- *
- * Special buffer handling
- * Special buffers are used for event queues and the TX and RX
- * descriptor rings.
- *
- *************************************************************************/
-
-/*
- * Initialise a special buffer
- *
- * This will define a buffer (previously allocated via
- * efx_alloc_special_buffer()) in the buffer table, allowing
- * it to be used for event queues, descriptor rings etc.
- */
-static void
-efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
-{
-       efx_qword_t buf_desc;
-       unsigned int index;
-       dma_addr_t dma_addr;
-       int i;
-
-       EFX_BUG_ON_PARANOID(!buffer->buf.addr);
-
-       /* Write buffer descriptors to NIC */
-       for (i = 0; i < buffer->entries; i++) {
-               index = buffer->index + i;
-               dma_addr = buffer->buf.dma_addr + (i * EFX_BUF_SIZE);
-               netif_dbg(efx, probe, efx->net_dev,
-                         "mapping special buffer %d at %llx\n",
-                         index, (unsigned long long)dma_addr);
-               EFX_POPULATE_QWORD_3(buf_desc,
-                                    FRF_AZ_BUF_ADR_REGION, 0,
-                                    FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
-                                    FRF_AZ_BUF_OWNER_ID_FBUF, 0);
-               efx_write_buf_tbl(efx, &buf_desc, index);
-       }
-}
-
-/* Unmaps a buffer and clears the buffer table entries */
-static void
-efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
-{
-       efx_oword_t buf_tbl_upd;
-       unsigned int start = buffer->index;
-       unsigned int end = (buffer->index + buffer->entries - 1);
-
-       if (!buffer->entries)
-               return;
-
-       netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
-                 buffer->index, buffer->index + buffer->entries - 1);
-
-       EFX_POPULATE_OWORD_4(buf_tbl_upd,
-                            FRF_AZ_BUF_UPD_CMD, 0,
-                            FRF_AZ_BUF_CLR_CMD, 1,
-                            FRF_AZ_BUF_CLR_END_ID, end,
-                            FRF_AZ_BUF_CLR_START_ID, start);
-       efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
-}
-
-/*
- * Allocate a new special buffer
- *
- * This allocates memory for a new buffer, clears it and allocates a
- * new buffer ID range.  It does not write into the buffer table.
- *
- * This call will allocate 4KB buffers, since 8KB buffers can't be
- * used for event queues and descriptor rings.
- */
-static int efx_alloc_special_buffer(struct efx_nic *efx,
-                                   struct efx_special_buffer *buffer,
-                                   unsigned int len)
-{
-       len = ALIGN(len, EFX_BUF_SIZE);
-
-       if (efx_nic_alloc_buffer(efx, &buffer->buf, len, GFP_KERNEL))
-               return -ENOMEM;
-       buffer->entries = len / EFX_BUF_SIZE;
-       BUG_ON(buffer->buf.dma_addr & (EFX_BUF_SIZE - 1));
-
-       /* Select new buffer ID */
-       buffer->index = efx->next_buffer_table;
-       efx->next_buffer_table += buffer->entries;
-#ifdef CONFIG_SFC_SRIOV
-       BUG_ON(efx_sriov_enabled(efx) &&
-              efx->vf_buftbl_base < efx->next_buffer_table);
-#endif
-
-       netif_dbg(efx, probe, efx->net_dev,
-                 "allocating special buffers %d-%d at %llx+%x "
-                 "(virt %p phys %llx)\n", buffer->index,
-                 buffer->index + buffer->entries - 1,
-                 (u64)buffer->buf.dma_addr, len,
-                 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
-
-       return 0;
-}
-
-static void
-efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
-{
-       if (!buffer->buf.addr)
-               return;
-
-       netif_dbg(efx, hw, efx->net_dev,
-                 "deallocating special buffers %d-%d at %llx+%x "
-                 "(virt %p phys %llx)\n", buffer->index,
-                 buffer->index + buffer->entries - 1,
-                 (u64)buffer->buf.dma_addr, buffer->buf.len,
-                 buffer->buf.addr, (u64)virt_to_phys(buffer->buf.addr));
-
-       efx_nic_free_buffer(efx, &buffer->buf);
-       buffer->entries = 0;
-}
-
 /**************************************************************************
  *
  * Generic buffer handling
@@ -318,1079 +51,6 @@ void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
        }
 }
 
-/**************************************************************************
- *
- * TX path
- *
- **************************************************************************/
-
-/* Returns a pointer to the specified transmit descriptor in the TX
- * descriptor queue belonging to the specified channel.
- */
-static inline efx_qword_t *
-efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
-{
-       return ((efx_qword_t *) (tx_queue->txd.buf.addr)) + index;
-}
-
-/* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
-static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
-{
-       unsigned write_ptr;
-       efx_dword_t reg;
-
-       write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
-       EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
-       efx_writed_page(tx_queue->efx, &reg,
-                       FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
-}
-
-/* Write pointer and first descriptor for TX descriptor ring */
-static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
-                                   const efx_qword_t *txd)
-{
-       unsigned write_ptr;
-       efx_oword_t reg;
-
-       BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
-       BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
-
-       write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
-       EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
-                            FRF_AZ_TX_DESC_WPTR, write_ptr);
-       reg.qword[0] = *txd;
-       efx_writeo_page(tx_queue->efx, &reg,
-                       FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
-}
-
-static inline bool
-efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
-{
-       unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
-
-       if (empty_read_count == 0)
-               return false;
-
-       tx_queue->empty_read_count = 0;
-       return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0
-               && tx_queue->write_count - write_count == 1;
-}
-
-/* For each entry inserted into the software descriptor ring, create a
- * descriptor in the hardware TX descriptor ring (in host memory), and
- * write a doorbell.
- */
-void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
-{
-
-       struct efx_tx_buffer *buffer;
-       efx_qword_t *txd;
-       unsigned write_ptr;
-       unsigned old_write_count = tx_queue->write_count;
-
-       BUG_ON(tx_queue->write_count == tx_queue->insert_count);
-
-       do {
-               write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
-               buffer = &tx_queue->buffer[write_ptr];
-               txd = efx_tx_desc(tx_queue, write_ptr);
-               ++tx_queue->write_count;
-
-               /* Create TX descriptor ring entry */
-               BUILD_BUG_ON(EFX_TX_BUF_CONT != 1);
-               EFX_POPULATE_QWORD_4(*txd,
-                                    FSF_AZ_TX_KER_CONT,
-                                    buffer->flags & EFX_TX_BUF_CONT,
-                                    FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
-                                    FSF_AZ_TX_KER_BUF_REGION, 0,
-                                    FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
-       } while (tx_queue->write_count != tx_queue->insert_count);
-
-       wmb(); /* Ensure descriptors are written before they are fetched */
-
-       if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
-               txd = efx_tx_desc(tx_queue,
-                                 old_write_count & tx_queue->ptr_mask);
-               efx_push_tx_desc(tx_queue, txd);
-               ++tx_queue->pushes;
-       } else {
-               efx_notify_tx_desc(tx_queue);
-       }
-}
-
-/* Allocate hardware resources for a TX queue */
-int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
-{
-       struct efx_nic *efx = tx_queue->efx;
-       unsigned entries;
-
-       entries = tx_queue->ptr_mask + 1;
-       return efx_alloc_special_buffer(efx, &tx_queue->txd,
-                                       entries * sizeof(efx_qword_t));
-}
-
-void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
-{
-       struct efx_nic *efx = tx_queue->efx;
-       efx_oword_t reg;
-
-       /* Pin TX descriptor ring */
-       efx_init_special_buffer(efx, &tx_queue->txd);
-
-       /* Push TX descriptor ring to card */
-       EFX_POPULATE_OWORD_10(reg,
-                             FRF_AZ_TX_DESCQ_EN, 1,
-                             FRF_AZ_TX_ISCSI_DDIG_EN, 0,
-                             FRF_AZ_TX_ISCSI_HDIG_EN, 0,
-                             FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
-                             FRF_AZ_TX_DESCQ_EVQ_ID,
-                             tx_queue->channel->channel,
-                             FRF_AZ_TX_DESCQ_OWNER_ID, 0,
-                             FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
-                             FRF_AZ_TX_DESCQ_SIZE,
-                             __ffs(tx_queue->txd.entries),
-                             FRF_AZ_TX_DESCQ_TYPE, 0,
-                             FRF_BZ_TX_NON_IP_DROP_DIS, 1);
-
-       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
-               int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
-               EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
-               EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
-                                   !csum);
-       }
-
-       efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
-                        tx_queue->queue);
-
-       if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
-               /* Only 128 bits in this register */
-               BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
-
-               efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
-               if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
-                       __clear_bit_le(tx_queue->queue, &reg);
-               else
-                       __set_bit_le(tx_queue->queue, &reg);
-               efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
-       }
-
-       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
-               EFX_POPULATE_OWORD_1(reg,
-                                    FRF_BZ_TX_PACE,
-                                    (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
-                                    FFE_BZ_TX_PACE_OFF :
-                                    FFE_BZ_TX_PACE_RESERVED);
-               efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
-                                tx_queue->queue);
-       }
-}
-
-static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
-{
-       struct efx_nic *efx = tx_queue->efx;
-       efx_oword_t tx_flush_descq;
-
-       WARN_ON(atomic_read(&tx_queue->flush_outstanding));
-       atomic_set(&tx_queue->flush_outstanding, 1);
-
-       EFX_POPULATE_OWORD_2(tx_flush_descq,
-                            FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
-                            FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
-       efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
-}
-
-void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
-{
-       struct efx_nic *efx = tx_queue->efx;
-       efx_oword_t tx_desc_ptr;
-
-       /* Remove TX descriptor ring from card */
-       EFX_ZERO_OWORD(tx_desc_ptr);
-       efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
-                        tx_queue->queue);
-
-       /* Unpin TX descriptor ring */
-       efx_fini_special_buffer(efx, &tx_queue->txd);
-}
-
-/* Free buffers backing TX queue */
-void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
-{
-       efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
-}
-
-/**************************************************************************
- *
- * RX path
- *
- **************************************************************************/
-
-/* Returns a pointer to the specified descriptor in the RX descriptor queue */
-static inline efx_qword_t *
-efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
-{
-       return ((efx_qword_t *) (rx_queue->rxd.buf.addr)) + index;
-}
-
-/* This creates an entry in the RX descriptor queue */
-static inline void
-efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
-{
-       struct efx_rx_buffer *rx_buf;
-       efx_qword_t *rxd;
-
-       rxd = efx_rx_desc(rx_queue, index);
-       rx_buf = efx_rx_buffer(rx_queue, index);
-       EFX_POPULATE_QWORD_3(*rxd,
-                            FSF_AZ_RX_KER_BUF_SIZE,
-                            rx_buf->len -
-                            rx_queue->efx->type->rx_buffer_padding,
-                            FSF_AZ_RX_KER_BUF_REGION, 0,
-                            FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
-}
-
-/* This writes to the RX_DESC_WPTR register for the specified receive
- * descriptor ring.
- */
-void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
-{
-       struct efx_nic *efx = rx_queue->efx;
-       efx_dword_t reg;
-       unsigned write_ptr;
-
-       while (rx_queue->notified_count != rx_queue->added_count) {
-               efx_build_rx_desc(
-                       rx_queue,
-                       rx_queue->notified_count & rx_queue->ptr_mask);
-               ++rx_queue->notified_count;
-       }
-
-       wmb();
-       write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
-       EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
-       efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
-                       efx_rx_queue_index(rx_queue));
-}
-
-int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
-{
-       struct efx_nic *efx = rx_queue->efx;
-       unsigned entries;
-
-       entries = rx_queue->ptr_mask + 1;
-       return efx_alloc_special_buffer(efx, &rx_queue->rxd,
-                                       entries * sizeof(efx_qword_t));
-}
-
-void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
-{
-       efx_oword_t rx_desc_ptr;
-       struct efx_nic *efx = rx_queue->efx;
-       bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
-       bool iscsi_digest_en = is_b0;
-       bool jumbo_en;
-
-       /* For kernel-mode queues in Falcon A1, the JUMBO flag enables
-        * DMA to continue after a PCIe page boundary (and scattering
-        * is not possible).  In Falcon B0 and Siena, it enables
-        * scatter.
-        */
-       jumbo_en = !is_b0 || efx->rx_scatter;
-
-       netif_dbg(efx, hw, efx->net_dev,
-                 "RX queue %d ring in special buffers %d-%d\n",
-                 efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
-                 rx_queue->rxd.index + rx_queue->rxd.entries - 1);
-
-       rx_queue->scatter_n = 0;
-
-       /* Pin RX descriptor ring */
-       efx_init_special_buffer(efx, &rx_queue->rxd);
-
-       /* Push RX descriptor ring to card */
-       EFX_POPULATE_OWORD_10(rx_desc_ptr,
-                             FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
-                             FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
-                             FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
-                             FRF_AZ_RX_DESCQ_EVQ_ID,
-                             efx_rx_queue_channel(rx_queue)->channel,
-                             FRF_AZ_RX_DESCQ_OWNER_ID, 0,
-                             FRF_AZ_RX_DESCQ_LABEL,
-                             efx_rx_queue_index(rx_queue),
-                             FRF_AZ_RX_DESCQ_SIZE,
-                             __ffs(rx_queue->rxd.entries),
-                             FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
-                             FRF_AZ_RX_DESCQ_JUMBO, jumbo_en,
-                             FRF_AZ_RX_DESCQ_EN, 1);
-       efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
-                        efx_rx_queue_index(rx_queue));
-}
-
-static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
-{
-       struct efx_nic *efx = rx_queue->efx;
-       efx_oword_t rx_flush_descq;
-
-       EFX_POPULATE_OWORD_2(rx_flush_descq,
-                            FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
-                            FRF_AZ_RX_FLUSH_DESCQ,
-                            efx_rx_queue_index(rx_queue));
-       efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
-}
-
-void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
-{
-       efx_oword_t rx_desc_ptr;
-       struct efx_nic *efx = rx_queue->efx;
-
-       /* Remove RX descriptor ring from card */
-       EFX_ZERO_OWORD(rx_desc_ptr);
-       efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
-                        efx_rx_queue_index(rx_queue));
-
-       /* Unpin RX descriptor ring */
-       efx_fini_special_buffer(efx, &rx_queue->rxd);
-}
-
-/* Free buffers backing RX queue */
-void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
-{
-       efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
-}
-
-/**************************************************************************
- *
- * Flush handling
- *
- **************************************************************************/
-
-/* efx_nic_flush_queues() must be woken up when all flushes are completed,
- * or more RX flushes can be kicked off.
- */
-static bool efx_flush_wake(struct efx_nic *efx)
-{
-       /* Ensure that all updates are visible to efx_nic_flush_queues() */
-       smp_mb();
-
-       return (atomic_read(&efx->drain_pending) == 0 ||
-               (atomic_read(&efx->rxq_flush_outstanding) < EFX_RX_FLUSH_COUNT
-                && atomic_read(&efx->rxq_flush_pending) > 0));
-}
-
-static bool efx_check_tx_flush_complete(struct efx_nic *efx)
-{
-       bool i = true;
-       efx_oword_t txd_ptr_tbl;
-       struct efx_channel *channel;
-       struct efx_tx_queue *tx_queue;
-
-       efx_for_each_channel(channel, efx) {
-               efx_for_each_channel_tx_queue(tx_queue, channel) {
-                       efx_reado_table(efx, &txd_ptr_tbl,
-                                       FR_BZ_TX_DESC_PTR_TBL, tx_queue->queue);
-                       if (EFX_OWORD_FIELD(txd_ptr_tbl,
-                                           FRF_AZ_TX_DESCQ_FLUSH) ||
-                           EFX_OWORD_FIELD(txd_ptr_tbl,
-                                           FRF_AZ_TX_DESCQ_EN)) {
-                               netif_dbg(efx, hw, efx->net_dev,
-                                         "flush did not complete on TXQ %d\n",
-                                         tx_queue->queue);
-                               i = false;
-                       } else if (atomic_cmpxchg(&tx_queue->flush_outstanding,
-                                                 1, 0)) {
-                               /* The flush is complete, but we didn't
-                                * receive a flush completion event
-                                */
-                               netif_dbg(efx, hw, efx->net_dev,
-                                         "flush complete on TXQ %d, so drain "
-                                         "the queue\n", tx_queue->queue);
-                               /* Don't need to increment drain_pending as it
-                                * has already been incremented for the queues
-                                * which did not drain
-                                */
-                               efx_magic_event(channel,
-                                               EFX_CHANNEL_MAGIC_TX_DRAIN(
-                                                       tx_queue));
-                       }
-               }
-       }
-
-       return i;
-}
-
-/* Flush all the transmit queues, and continue flushing receive queues until
- * they're all flushed. Wait for the DRAIN events to be recieved so that there
- * are no more RX and TX events left on any channel. */
-static int efx_farch_do_flush(struct efx_nic *efx)
-{
-       unsigned timeout = msecs_to_jiffies(5000); /* 5s for all flushes and drains */
-       struct efx_channel *channel;
-       struct efx_rx_queue *rx_queue;
-       struct efx_tx_queue *tx_queue;
-       int rc = 0;
-
-       efx_for_each_channel(channel, efx) {
-               efx_for_each_channel_tx_queue(tx_queue, channel) {
-                       atomic_inc(&efx->drain_pending);
-                       efx_flush_tx_queue(tx_queue);
-               }
-               efx_for_each_channel_rx_queue(rx_queue, channel) {
-                       atomic_inc(&efx->drain_pending);
-                       rx_queue->flush_pending = true;
-                       atomic_inc(&efx->rxq_flush_pending);
-               }
-       }
-
-       while (timeout && atomic_read(&efx->drain_pending) > 0) {
-               /* If SRIOV is enabled, then offload receive queue flushing to
-                * the firmware (though we will still have to poll for
-                * completion). If that fails, fall back to the old scheme.
-                */
-               if (efx_sriov_enabled(efx)) {
-                       rc = efx_mcdi_flush_rxqs(efx);
-                       if (!rc)
-                               goto wait;
-               }
-
-               /* The hardware supports four concurrent rx flushes, each of
-                * which may need to be retried if there is an outstanding
-                * descriptor fetch
-                */
-               efx_for_each_channel(channel, efx) {
-                       efx_for_each_channel_rx_queue(rx_queue, channel) {
-                               if (atomic_read(&efx->rxq_flush_outstanding) >=
-                                   EFX_RX_FLUSH_COUNT)
-                                       break;
-
-                               if (rx_queue->flush_pending) {
-                                       rx_queue->flush_pending = false;
-                                       atomic_dec(&efx->rxq_flush_pending);
-                                       atomic_inc(&efx->rxq_flush_outstanding);
-                                       efx_flush_rx_queue(rx_queue);
-                               }
-                       }
-               }
-
-       wait:
-               timeout = wait_event_timeout(efx->flush_wq, efx_flush_wake(efx),
-                                            timeout);
-       }
-
-       if (atomic_read(&efx->drain_pending) &&
-           !efx_check_tx_flush_complete(efx)) {
-               netif_err(efx, hw, efx->net_dev, "failed to flush %d queues "
-                         "(rx %d+%d)\n", atomic_read(&efx->drain_pending),
-                         atomic_read(&efx->rxq_flush_outstanding),
-                         atomic_read(&efx->rxq_flush_pending));
-               rc = -ETIMEDOUT;
-
-               atomic_set(&efx->drain_pending, 0);
-               atomic_set(&efx->rxq_flush_pending, 0);
-               atomic_set(&efx->rxq_flush_outstanding, 0);
-       }
-
-       return rc;
-}
-
-int efx_farch_fini_dmaq(struct efx_nic *efx)
-{
-       struct efx_channel *channel;
-       struct efx_tx_queue *tx_queue;
-       struct efx_rx_queue *rx_queue;
-       int rc = 0;
-
-       /* Do not attempt to write to the NIC during EEH recovery */
-       if (efx->state != STATE_RECOVERY) {
-               /* Only perform flush if DMA is enabled */
-               if (efx->pci_dev->is_busmaster) {
-                       efx->type->prepare_flush(efx);
-                       rc = efx_farch_do_flush(efx);
-                       efx->type->finish_flush(efx);
-               }
-
-               efx_for_each_channel(channel, efx) {
-                       efx_for_each_channel_rx_queue(rx_queue, channel)
-                               efx_nic_fini_rx(rx_queue);
-                       efx_for_each_channel_tx_queue(tx_queue, channel)
-                               efx_nic_fini_tx(tx_queue);
-               }
-       }
-
-       return rc;
-}
-
-/**************************************************************************
- *
- * Event queue processing
- * Event queues are processed by per-channel tasklets.
- *
- **************************************************************************/
-
-/* Update a channel's event queue's read pointer (RPTR) register
- *
- * This writes the EVQ_RPTR_REG register for the specified channel's
- * event queue.
- */
-void efx_nic_eventq_read_ack(struct efx_channel *channel)
-{
-       efx_dword_t reg;
-       struct efx_nic *efx = channel->efx;
-
-       EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
-                            channel->eventq_read_ptr & channel->eventq_mask);
-
-       /* For Falcon A1, EVQ_RPTR_KER is documented as having a step size
-        * of 4 bytes, but it is really 16 bytes just like later revisions.
-        */
-       efx_writed(efx, &reg,
-                  efx->type->evq_rptr_tbl_base +
-                  FR_BZ_EVQ_RPTR_STEP * channel->channel);
-}
-
-/* Use HW to insert a SW defined event */
-void efx_generate_event(struct efx_nic *efx, unsigned int evq,
-                       efx_qword_t *event)
-{
-       efx_oword_t drv_ev_reg;
-
-       BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
-                    FRF_AZ_DRV_EV_DATA_WIDTH != 64);
-       drv_ev_reg.u32[0] = event->u32[0];
-       drv_ev_reg.u32[1] = event->u32[1];
-       drv_ev_reg.u32[2] = 0;
-       drv_ev_reg.u32[3] = 0;
-       EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, evq);
-       efx_writeo(efx, &drv_ev_reg, FR_AZ_DRV_EV);
-}
-
-static void efx_magic_event(struct efx_channel *channel, u32 magic)
-{
-       efx_qword_t event;
-
-       EFX_POPULATE_QWORD_2(event, FSF_AZ_EV_CODE,
-                            FSE_AZ_EV_CODE_DRV_GEN_EV,
-                            FSF_AZ_DRV_GEN_EV_MAGIC, magic);
-       efx_generate_event(channel->efx, channel->channel, &event);
-}
-
-/* Handle a transmit completion event
- *
- * The NIC batches TX completion events; the message we receive is of
- * the form "complete all TX events up to this index".
- */
-static int
-efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
-{
-       unsigned int tx_ev_desc_ptr;
-       unsigned int tx_ev_q_label;
-       struct efx_tx_queue *tx_queue;
-       struct efx_nic *efx = channel->efx;
-       int tx_packets = 0;
-
-       if (unlikely(ACCESS_ONCE(efx->reset_pending)))
-               return 0;
-
-       if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
-               /* Transmit completion */
-               tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
-               tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
-               tx_queue = efx_channel_get_tx_queue(
-                       channel, tx_ev_q_label % EFX_TXQ_TYPES);
-               tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
-                             tx_queue->ptr_mask);
-               efx_xmit_done(tx_queue, tx_ev_desc_ptr);
-       } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
-               /* Rewrite the FIFO write pointer */
-               tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
-               tx_queue = efx_channel_get_tx_queue(
-                       channel, tx_ev_q_label % EFX_TXQ_TYPES);
-
-               netif_tx_lock(efx->net_dev);
-               efx_notify_tx_desc(tx_queue);
-               netif_tx_unlock(efx->net_dev);
-       } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
-                  EFX_WORKAROUND_10727(efx)) {
-               efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
-       } else {
-               netif_err(efx, tx_err, efx->net_dev,
-                         "channel %d unexpected TX event "
-                         EFX_QWORD_FMT"\n", channel->channel,
-                         EFX_QWORD_VAL(*event));
-       }
-
-       return tx_packets;
-}
-
-/* Detect errors included in the rx_evt_pkt_ok bit. */
-static u16 efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
-                               const efx_qword_t *event)
-{
-       struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
-       struct efx_nic *efx = rx_queue->efx;
-       bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
-       bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
-       bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
-       bool rx_ev_other_err, rx_ev_pause_frm;
-       bool rx_ev_hdr_type, rx_ev_mcast_pkt;
-       unsigned rx_ev_pkt_type;
-
-       rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
-       rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
-       rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
-       rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
-       rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
-                                                FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
-       rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
-                                                 FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
-       rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
-                                                  FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
-       rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
-       rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
-       rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
-                         0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
-       rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
-
-       /* Every error apart from tobe_disc and pause_frm */
-       rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
-                          rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
-                          rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
-
-       /* Count errors that are not in MAC stats.  Ignore expected
-        * checksum errors during self-test. */
-       if (rx_ev_frm_trunc)
-               ++channel->n_rx_frm_trunc;
-       else if (rx_ev_tobe_disc)
-               ++channel->n_rx_tobe_disc;
-       else if (!efx->loopback_selftest) {
-               if (rx_ev_ip_hdr_chksum_err)
-                       ++channel->n_rx_ip_hdr_chksum_err;
-               else if (rx_ev_tcp_udp_chksum_err)
-                       ++channel->n_rx_tcp_udp_chksum_err;
-       }
-
-       /* TOBE_DISC is expected on unicast mismatches; don't print out an
-        * error message.  FRM_TRUNC indicates RXDP dropped the packet due
-        * to a FIFO overflow.
-        */
-#ifdef DEBUG
-       if (rx_ev_other_err && net_ratelimit()) {
-               netif_dbg(efx, rx_err, efx->net_dev,
-                         " RX queue %d unexpected RX event "
-                         EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
-                         efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
-                         rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
-                         rx_ev_ip_hdr_chksum_err ?
-                         " [IP_HDR_CHKSUM_ERR]" : "",
-                         rx_ev_tcp_udp_chksum_err ?
-                         " [TCP_UDP_CHKSUM_ERR]" : "",
-                         rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
-                         rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
-                         rx_ev_drib_nib ? " [DRIB_NIB]" : "",
-                         rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
-                         rx_ev_pause_frm ? " [PAUSE]" : "");
-       }
-#endif
-
-       /* The frame must be discarded if any of these are true. */
-       return (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
-               rx_ev_tobe_disc | rx_ev_pause_frm) ?
-               EFX_RX_PKT_DISCARD : 0;
-}
-
-/* Handle receive events that are not in-order. Return true if this
- * can be handled as a partial packet discard, false if it's more
- * serious.
- */
-static bool
-efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
-{
-       struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
-       struct efx_nic *efx = rx_queue->efx;
-       unsigned expected, dropped;
-
-       if (rx_queue->scatter_n &&
-           index == ((rx_queue->removed_count + rx_queue->scatter_n - 1) &
-                     rx_queue->ptr_mask)) {
-               ++channel->n_rx_nodesc_trunc;
-               return true;
-       }
-
-       expected = rx_queue->removed_count & rx_queue->ptr_mask;
-       dropped = (index - expected) & rx_queue->ptr_mask;
-       netif_info(efx, rx_err, efx->net_dev,
-                  "dropped %d events (index=%d expected=%d)\n",
-                  dropped, index, expected);
-
-       efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
-                          RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
-       return false;
-}
-
-/* Handle a packet received event
- *
- * The NIC gives a "discard" flag if it's a unicast packet with the
- * wrong destination address
- * Also "is multicast" and "matches multicast filter" flags can be used to
- * discard non-matching multicast packets.
- */
-static void
-efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
-{
-       unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
-       unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
-       unsigned expected_ptr;
-       bool rx_ev_pkt_ok, rx_ev_sop, rx_ev_cont;
-       u16 flags;
-       struct efx_rx_queue *rx_queue;
-       struct efx_nic *efx = channel->efx;
-
-       if (unlikely(ACCESS_ONCE(efx->reset_pending)))
-               return;
-
-       rx_ev_cont = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT);
-       rx_ev_sop = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP);
-       WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
-               channel->channel);
-
-       rx_queue = efx_channel_get_rx_queue(channel);
-
-       rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
-       expected_ptr = ((rx_queue->removed_count + rx_queue->scatter_n) &
-                       rx_queue->ptr_mask);
-
-       /* Check for partial drops and other errors */
-       if (unlikely(rx_ev_desc_ptr != expected_ptr) ||
-           unlikely(rx_ev_sop != (rx_queue->scatter_n == 0))) {
-               if (rx_ev_desc_ptr != expected_ptr &&
-                   !efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr))
-                       return;
-
-               /* Discard all pending fragments */
-               if (rx_queue->scatter_n) {
-                       efx_rx_packet(
-                               rx_queue,
-                               rx_queue->removed_count & rx_queue->ptr_mask,
-                               rx_queue->scatter_n, 0, EFX_RX_PKT_DISCARD);
-                       rx_queue->removed_count += rx_queue->scatter_n;
-                       rx_queue->scatter_n = 0;
-               }
-
-               /* Return if there is no new fragment */
-               if (rx_ev_desc_ptr != expected_ptr)
-                       return;
-
-               /* Discard new fragment if not SOP */
-               if (!rx_ev_sop) {
-                       efx_rx_packet(
-                               rx_queue,
-                               rx_queue->removed_count & rx_queue->ptr_mask,
-                               1, 0, EFX_RX_PKT_DISCARD);
-                       ++rx_queue->removed_count;
-                       return;
-               }
-       }
-
-       ++rx_queue->scatter_n;
-       if (rx_ev_cont)
-               return;
-
-       rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
-       rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
-       rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
-
-       if (likely(rx_ev_pkt_ok)) {
-               /* If packet is marked as OK then we can rely on the
-                * hardware checksum and classification.
-                */
-               flags = 0;
-               switch (rx_ev_hdr_type) {
-               case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP:
-                       flags |= EFX_RX_PKT_TCP;
-                       /* fall through */
-               case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP:
-                       flags |= EFX_RX_PKT_CSUMMED;
-                       /* fall through */
-               case FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_OTHER:
-               case FSE_AZ_RX_EV_HDR_TYPE_OTHER:
-                       break;
-               }
-       } else {
-               flags = efx_handle_rx_not_ok(rx_queue, event);
-       }
-
-       /* Detect multicast packets that didn't match the filter */
-       rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
-       if (rx_ev_mcast_pkt) {
-               unsigned int rx_ev_mcast_hash_match =
-                       EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
-
-               if (unlikely(!rx_ev_mcast_hash_match)) {
-                       ++channel->n_rx_mcast_mismatch;
-                       flags |= EFX_RX_PKT_DISCARD;
-               }
-       }
-
-       channel->irq_mod_score += 2;
-
-       /* Handle received packet */
-       efx_rx_packet(rx_queue,
-                     rx_queue->removed_count & rx_queue->ptr_mask,
-                     rx_queue->scatter_n, rx_ev_byte_cnt, flags);
-       rx_queue->removed_count += rx_queue->scatter_n;
-       rx_queue->scatter_n = 0;
-}
-
-/* If this flush done event corresponds to a &struct efx_tx_queue, then
- * send an %EFX_CHANNEL_MAGIC_TX_DRAIN event to drain the event queue
- * of all transmit completions.
- */
-static void
-efx_handle_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
-{
-       struct efx_tx_queue *tx_queue;
-       int qid;
-
-       qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
-       if (qid < EFX_TXQ_TYPES * efx->n_tx_channels) {
-               tx_queue = efx_get_tx_queue(efx, qid / EFX_TXQ_TYPES,
-                                           qid % EFX_TXQ_TYPES);
-               if (atomic_cmpxchg(&tx_queue->flush_outstanding, 1, 0)) {
-                       efx_magic_event(tx_queue->channel,
-                                       EFX_CHANNEL_MAGIC_TX_DRAIN(tx_queue));
-               }
-       }
-}
-
-/* If this flush done event corresponds to a &struct efx_rx_queue: If the flush
- * was succesful then send an %EFX_CHANNEL_MAGIC_RX_DRAIN, otherwise add
- * the RX queue back to the mask of RX queues in need of flushing.
- */
-static void
-efx_handle_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
-{
-       struct efx_channel *channel;
-       struct efx_rx_queue *rx_queue;
-       int qid;
-       bool failed;
-
-       qid = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
-       failed = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
-       if (qid >= efx->n_channels)
-               return;
-       channel = efx_get_channel(efx, qid);
-       if (!efx_channel_has_rx_queue(channel))
-               return;
-       rx_queue = efx_channel_get_rx_queue(channel);
-
-       if (failed) {
-               netif_info(efx, hw, efx->net_dev,
-                          "RXQ %d flush retry\n", qid);
-               rx_queue->flush_pending = true;
-               atomic_inc(&efx->rxq_flush_pending);
-       } else {
-               efx_magic_event(efx_rx_queue_channel(rx_queue),
-                               EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue));
-       }
-       atomic_dec(&efx->rxq_flush_outstanding);
-       if (efx_flush_wake(efx))
-               wake_up(&efx->flush_wq);
-}
-
-static void
-efx_handle_drain_event(struct efx_channel *channel)
-{
-       struct efx_nic *efx = channel->efx;
-
-       WARN_ON(atomic_read(&efx->drain_pending) == 0);
-       atomic_dec(&efx->drain_pending);
-       if (efx_flush_wake(efx))
-               wake_up(&efx->flush_wq);
-}
-
-static void
-efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
-{
-       struct efx_nic *efx = channel->efx;
-       struct efx_rx_queue *rx_queue =
-               efx_channel_has_rx_queue(channel) ?
-               efx_channel_get_rx_queue(channel) : NULL;
-       unsigned magic, code;
-
-       magic = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
-       code = _EFX_CHANNEL_MAGIC_CODE(magic);
-
-       if (magic == EFX_CHANNEL_MAGIC_TEST(channel)) {
-               channel->event_test_cpu = raw_smp_processor_id();
-       } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_FILL(rx_queue)) {
-               /* The queue must be empty, so we won't receive any rx
-                * events, so efx_process_channel() won't refill the
-                * queue. Refill it here */
-               efx_fast_push_rx_descriptors(rx_queue);
-       } else if (rx_queue && magic == EFX_CHANNEL_MAGIC_RX_DRAIN(rx_queue)) {
-               efx_handle_drain_event(channel);
-       } else if (code == _EFX_CHANNEL_MAGIC_TX_DRAIN) {
-               efx_handle_drain_event(channel);
-       } else {
-               netif_dbg(efx, hw, efx->net_dev, "channel %d received "
-                         "generated event "EFX_QWORD_FMT"\n",
-                         channel->channel, EFX_QWORD_VAL(*event));
-       }
-}
-
-static void
-efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
-{
-       struct efx_nic *efx = channel->efx;
-       unsigned int ev_sub_code;
-       unsigned int ev_sub_data;
-
-       ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
-       ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
-
-       switch (ev_sub_code) {
-       case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
-               netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
-                          channel->channel, ev_sub_data);
-               efx_handle_tx_flush_done(efx, event);
-               efx_sriov_tx_flush_done(efx, event);
-               break;
-       case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
-               netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
-                          channel->channel, ev_sub_data);
-               efx_handle_rx_flush_done(efx, event);
-               efx_sriov_rx_flush_done(efx, event);
-               break;
-       case FSE_AZ_EVQ_INIT_DONE_EV:
-               netif_dbg(efx, hw, efx->net_dev,
-                         "channel %d EVQ %d initialised\n",
-                         channel->channel, ev_sub_data);
-               break;
-       case FSE_AZ_SRM_UPD_DONE_EV:
-               netif_vdbg(efx, hw, efx->net_dev,
-                          "channel %d SRAM update done\n", channel->channel);
-               break;
-       case FSE_AZ_WAKE_UP_EV:
-               netif_vdbg(efx, hw, efx->net_dev,
-                          "channel %d RXQ %d wakeup event\n",
-                          channel->channel, ev_sub_data);
-               break;
-       case FSE_AZ_TIMER_EV:
-               netif_vdbg(efx, hw, efx->net_dev,
-                          "channel %d RX queue %d timer expired\n",
-                          channel->channel, ev_sub_data);
-               break;
-       case FSE_AA_RX_RECOVER_EV:
-               netif_err(efx, rx_err, efx->net_dev,
-                         "channel %d seen DRIVER RX_RESET event. "
-                       "Resetting.\n", channel->channel);
-               atomic_inc(&efx->rx_reset);
-               efx_schedule_reset(efx,
-                                  EFX_WORKAROUND_6555(efx) ?
-                                  RESET_TYPE_RX_RECOVERY :
-                                  RESET_TYPE_DISABLE);
-               break;
-       case FSE_BZ_RX_DSC_ERROR_EV:
-               if (ev_sub_data < EFX_VI_BASE) {
-                       netif_err(efx, rx_err, efx->net_dev,
-                                 "RX DMA Q %d reports descriptor fetch error."
-                                 " RX Q %d is disabled.\n", ev_sub_data,
-                                 ev_sub_data);
-                       efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
-               } else
-                       efx_sriov_desc_fetch_err(efx, ev_sub_data);
-               break;
-       case FSE_BZ_TX_DSC_ERROR_EV:
-               if (ev_sub_data < EFX_VI_BASE) {
-                       netif_err(efx, tx_err, efx->net_dev,
-                                 "TX DMA Q %d reports descriptor fetch error."
-                                 " TX Q %d is disabled.\n", ev_sub_data,
-                                 ev_sub_data);
-                       efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
-               } else
-                       efx_sriov_desc_fetch_err(efx, ev_sub_data);
-               break;
-       default:
-               netif_vdbg(efx, hw, efx->net_dev,
-                          "channel %d unknown driver event code %d "
-                          "data %04x\n", channel->channel, ev_sub_code,
-                          ev_sub_data);
-               break;
-       }
-}
-
-int efx_nic_process_eventq(struct efx_channel *channel, int budget)
-{
-       struct efx_nic *efx = channel->efx;
-       unsigned int read_ptr;
-       efx_qword_t event, *p_event;
-       int ev_code;
-       int tx_packets = 0;
-       int spent = 0;
-
-       read_ptr = channel->eventq_read_ptr;
-
-       for (;;) {
-               p_event = efx_event(channel, read_ptr);
-               event = *p_event;
-
-               if (!efx_event_present(&event))
-                       /* End of events */
-                       break;
-
-               netif_vdbg(channel->efx, intr, channel->efx->net_dev,
-                          "channel %d event is "EFX_QWORD_FMT"\n",
-                          channel->channel, EFX_QWORD_VAL(event));
-
-               /* Clear this event by marking it all ones */
-               EFX_SET_QWORD(*p_event);
-
-               ++read_ptr;
-
-               ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
-
-               switch (ev_code) {
-               case FSE_AZ_EV_CODE_RX_EV:
-                       efx_handle_rx_event(channel, &event);
-                       if (++spent == budget)
-                               goto out;
-                       break;
-               case FSE_AZ_EV_CODE_TX_EV:
-                       tx_packets += efx_handle_tx_event(channel, &event);
-                       if (tx_packets > efx->txq_entries) {
-                               spent = budget;
-                               goto out;
-                       }
-                       break;
-               case FSE_AZ_EV_CODE_DRV_GEN_EV:
-                       efx_handle_generated_event(channel, &event);
-                       break;
-               case FSE_AZ_EV_CODE_DRIVER_EV:
-                       efx_handle_driver_event(channel, &event);
-                       break;
-               case FSE_CZ_EV_CODE_USER_EV:
-                       efx_sriov_event(channel, &event);
-                       break;
-               case FSE_CZ_EV_CODE_MCDI_EV:
-                       efx_mcdi_process_event(channel, &event);
-                       break;
-               case FSE_AZ_EV_CODE_GLOBAL_EV:
-                       if (efx->type->handle_global_event &&
-                           efx->type->handle_global_event(channel, &event))
-                               break;
-                       /* else fall through */
-               default:
-                       netif_err(channel->efx, hw, channel->efx->net_dev,
-                                 "channel %d unknown event type %d (data "
-                                 EFX_QWORD_FMT ")\n", channel->channel,
-                                 ev_code, EFX_QWORD_VAL(event));
-               }
-       }
-
-out:
-       channel->eventq_read_ptr = read_ptr;
-       return spent;
-}
-
 /* Check whether an event is present in the eventq at the current
  * read pointer.  Only useful for self-test.
  */
@@ -1399,326 +59,18 @@ bool efx_nic_event_present(struct efx_channel *channel)
        return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
 }
 
-/* Allocate buffer table entries for event queue */
-int efx_nic_probe_eventq(struct efx_channel *channel)
-{
-       struct efx_nic *efx = channel->efx;
-       unsigned entries;
-
-       entries = channel->eventq_mask + 1;
-       return efx_alloc_special_buffer(efx, &channel->eventq,
-                                       entries * sizeof(efx_qword_t));
-}
-
-void efx_nic_init_eventq(struct efx_channel *channel)
-{
-       efx_oword_t reg;
-       struct efx_nic *efx = channel->efx;
-
-       netif_dbg(efx, hw, efx->net_dev,
-                 "channel %d event queue in special buffers %d-%d\n",
-                 channel->channel, channel->eventq.index,
-                 channel->eventq.index + channel->eventq.entries - 1);
-
-       if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
-               EFX_POPULATE_OWORD_3(reg,
-                                    FRF_CZ_TIMER_Q_EN, 1,
-                                    FRF_CZ_HOST_NOTIFY_MODE, 0,
-                                    FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
-               efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
-       }
-
-       /* Pin event queue buffer */
-       efx_init_special_buffer(efx, &channel->eventq);
-
-       /* Fill event queue with all ones (i.e. empty events) */
-       memset(channel->eventq.buf.addr, 0xff, channel->eventq.buf.len);
-
-       /* Push event queue to card */
-       EFX_POPULATE_OWORD_3(reg,
-                            FRF_AZ_EVQ_EN, 1,
-                            FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
-                            FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
-       efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
-                        channel->channel);
-
-       efx->type->push_irq_moderation(channel);
-}
-
-void efx_nic_fini_eventq(struct efx_channel *channel)
-{
-       efx_oword_t reg;
-       struct efx_nic *efx = channel->efx;
-
-       /* Remove event queue from card */
-       EFX_ZERO_OWORD(reg);
-       efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
-                        channel->channel);
-       if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
-               efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
-
-       /* Unpin event queue */
-       efx_fini_special_buffer(efx, &channel->eventq);
-}
-
-/* Free buffers backing event queue */
-void efx_nic_remove_eventq(struct efx_channel *channel)
-{
-       efx_free_special_buffer(channel->efx, &channel->eventq);
-}
-
-
 void efx_nic_event_test_start(struct efx_channel *channel)
 {
        channel->event_test_cpu = -1;
        smp_wmb();
-       efx_magic_event(channel, EFX_CHANNEL_MAGIC_TEST(channel));
-}
-
-void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
-{
-       efx_magic_event(efx_rx_queue_channel(rx_queue),
-                       EFX_CHANNEL_MAGIC_FILL(rx_queue));
-}
-
-/**************************************************************************
- *
- * Hardware interrupts
- * The hardware interrupt handler does very little work; all the event
- * queue processing is carried out by per-channel tasklets.
- *
- **************************************************************************/
-
-/* Enable/disable/generate interrupts */
-static inline void efx_nic_interrupts(struct efx_nic *efx,
-                                     bool enabled, bool force)
-{
-       efx_oword_t int_en_reg_ker;
-
-       EFX_POPULATE_OWORD_3(int_en_reg_ker,
-                            FRF_AZ_KER_INT_LEVE_SEL, efx->irq_level,
-                            FRF_AZ_KER_INT_KER, force,
-                            FRF_AZ_DRV_INT_EN_KER, enabled);
-       efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
+       channel->efx->type->ev_test_generate(channel);
 }
 
-void efx_nic_enable_interrupts(struct efx_nic *efx)
-{
-       EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
-       wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
-
-       efx_nic_interrupts(efx, true, false);
-}
-
-void efx_nic_disable_interrupts(struct efx_nic *efx)
-{
-       /* Disable interrupts */
-       efx_nic_interrupts(efx, false, false);
-}
-
-/* Generate a test interrupt
- * Interrupt must already have been enabled, otherwise nasty things
- * may happen.
- */
 void efx_nic_irq_test_start(struct efx_nic *efx)
 {
        efx->last_irq_cpu = -1;
        smp_wmb();
-       efx_nic_interrupts(efx, true, true);
-}
-
-/* Process a fatal interrupt
- * Disable bus mastering ASAP and schedule a reset
- */
-irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
-{
-       struct falcon_nic_data *nic_data = efx->nic_data;
-       efx_oword_t *int_ker = efx->irq_status.addr;
-       efx_oword_t fatal_intr;
-       int error, mem_perr;
-
-       efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
-       error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
-
-       netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
-                 EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
-                 EFX_OWORD_VAL(fatal_intr),
-                 error ? "disabling bus mastering" : "no recognised error");
-
-       /* If this is a memory parity error dump which blocks are offending */
-       mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
-                   EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
-       if (mem_perr) {
-               efx_oword_t reg;
-               efx_reado(efx, &reg, FR_AZ_MEM_STAT);
-               netif_err(efx, hw, efx->net_dev,
-                         "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
-                         EFX_OWORD_VAL(reg));
-       }
-
-       /* Disable both devices */
-       pci_clear_master(efx->pci_dev);
-       if (efx_nic_is_dual_func(efx))
-               pci_clear_master(nic_data->pci_dev2);
-       efx_nic_disable_interrupts(efx);
-
-       /* Count errors and reset or disable the NIC accordingly */
-       if (efx->int_error_count == 0 ||
-           time_after(jiffies, efx->int_error_expire)) {
-               efx->int_error_count = 0;
-               efx->int_error_expire =
-                       jiffies + EFX_INT_ERROR_EXPIRE * HZ;
-       }
-       if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
-               netif_err(efx, hw, efx->net_dev,
-                         "SYSTEM ERROR - reset scheduled\n");
-               efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
-       } else {
-               netif_err(efx, hw, efx->net_dev,
-                         "SYSTEM ERROR - max number of errors seen."
-                         "NIC will be disabled\n");
-               efx_schedule_reset(efx, RESET_TYPE_DISABLE);
-       }
-
-       return IRQ_HANDLED;
-}
-
-/* Handle a legacy interrupt
- * Acknowledges the interrupt and schedule event queue processing.
- */
-static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
-{
-       struct efx_nic *efx = dev_id;
-       bool soft_enabled = ACCESS_ONCE(efx->irq_soft_enabled);
-       efx_oword_t *int_ker = efx->irq_status.addr;
-       irqreturn_t result = IRQ_NONE;
-       struct efx_channel *channel;
-       efx_dword_t reg;
-       u32 queues;
-       int syserr;
-
-       /* Read the ISR which also ACKs the interrupts */
-       efx_readd(efx, &reg, FR_BZ_INT_ISR0);
-       queues = EFX_EXTRACT_DWORD(reg, 0, 31);
-
-       /* Legacy interrupts are disabled too late by the EEH kernel
-        * code. Disable them earlier.
-        * If an EEH error occurred, the read will have returned all ones.
-        */
-       if (EFX_DWORD_IS_ALL_ONES(reg) && efx_try_recovery(efx) &&
-           !efx->eeh_disabled_legacy_irq) {
-               disable_irq_nosync(efx->legacy_irq);
-               efx->eeh_disabled_legacy_irq = true;
-       }
-
-       /* Handle non-event-queue sources */
-       if (queues & (1U << efx->irq_level) && soft_enabled) {
-               syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
-               if (unlikely(syserr))
-                       return efx_nic_fatal_interrupt(efx);
-               efx->last_irq_cpu = raw_smp_processor_id();
-       }
-
-       if (queues != 0) {
-               if (EFX_WORKAROUND_15783(efx))
-                       efx->irq_zero_count = 0;
-
-               /* Schedule processing of any interrupting queues */
-               if (likely(soft_enabled)) {
-                       efx_for_each_channel(channel, efx) {
-                               if (queues & 1)
-                                       efx_schedule_channel_irq(channel);
-                               queues >>= 1;
-                       }
-               }
-               result = IRQ_HANDLED;
-
-       } else if (EFX_WORKAROUND_15783(efx)) {
-               efx_qword_t *event;
-
-               /* We can't return IRQ_HANDLED more than once on seeing ISR=0
-                * because this might be a shared interrupt. */
-               if (efx->irq_zero_count++ == 0)
-                       result = IRQ_HANDLED;
-
-               /* Ensure we schedule or rearm all event queues */
-               if (likely(soft_enabled)) {
-                       efx_for_each_channel(channel, efx) {
-                               event = efx_event(channel,
-                                                 channel->eventq_read_ptr);
-                               if (efx_event_present(event))
-                                       efx_schedule_channel_irq(channel);
-                               else
-                                       efx_nic_eventq_read_ack(channel);
-                       }
-               }
-       }
-
-       if (result == IRQ_HANDLED)
-               netif_vdbg(efx, intr, efx->net_dev,
-                          "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
-                          irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
-
-       return result;
-}
-
-/* Handle an MSI interrupt
- *
- * Handle an MSI hardware interrupt.  This routine schedules event
- * queue processing.  No interrupt acknowledgement cycle is necessary.
- * Also, we never need to check that the interrupt is for us, since
- * MSI interrupts cannot be shared.
- */
-static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
-{
-       struct efx_msi_context *context = dev_id;
-       struct efx_nic *efx = context->efx;
-       efx_oword_t *int_ker = efx->irq_status.addr;
-       int syserr;
-
-       netif_vdbg(efx, intr, efx->net_dev,
-                  "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
-                  irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
-
-       if (!likely(ACCESS_ONCE(efx->irq_soft_enabled)))
-               return IRQ_HANDLED;
-
-       /* Handle non-event-queue sources */
-       if (context->index == efx->irq_level) {
-               syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
-               if (unlikely(syserr))
-                       return efx_nic_fatal_interrupt(efx);
-               efx->last_irq_cpu = raw_smp_processor_id();
-       }
-
-       /* Schedule processing of the channel */
-       efx_schedule_channel_irq(efx->channel[context->index]);
-
-       return IRQ_HANDLED;
-}
-
-
-/* Setup RSS indirection table.
- * This maps from the hash value of the packet to RXQ
- */
-void efx_nic_push_rx_indir_table(struct efx_nic *efx)
-{
-       size_t i = 0;
-       efx_dword_t dword;
-
-       if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
-               return;
-
-       BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
-                    FR_BZ_RX_INDIRECTION_TBL_ROWS);
-
-       for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
-               EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
-                                    efx->rx_indir_table[i]);
-               efx_writed(efx, &dword,
-                          FR_BZ_RX_INDIRECTION_TBL +
-                          FR_BZ_RX_INDIRECTION_TBL_STEP * i);
-       }
+       efx->type->irq_test_generate(efx);
 }
 
 /* Hook interrupt handler(s)
@@ -1731,13 +83,8 @@ int efx_nic_init_interrupt(struct efx_nic *efx)
        int rc;
 
        if (!EFX_INT_MODE_USE_MSI(efx)) {
-               irq_handler_t handler;
-               if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
-                       handler = efx_legacy_interrupt;
-               else
-                       handler = falcon_legacy_interrupt_a1;
-
-               rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
+               rc = request_irq(efx->legacy_irq,
+                                efx->type->irq_handle_legacy, IRQF_SHARED,
                                 efx->name, efx);
                if (rc) {
                        netif_err(efx, drv, efx->net_dev,
@@ -1762,7 +109,7 @@ int efx_nic_init_interrupt(struct efx_nic *efx)
        /* Hook MSI or MSI-X interrupt */
        n_irqs = 0;
        efx_for_each_channel(channel, efx) {
-               rc = request_irq(channel->irq, efx_msi_interrupt,
+               rc = request_irq(channel->irq, efx->type->irq_handle_msi,
                                 IRQF_PROBE_SHARED, /* Not shared */
                                 efx->msi_context[channel->channel].name,
                                 &efx->msi_context[channel->channel]);
@@ -1818,154 +165,6 @@ void efx_nic_fini_interrupt(struct efx_nic *efx)
                free_irq(efx->legacy_irq, efx);
 }
 
-/* Looks at available SRAM resources and works out how many queues we
- * can support, and where things like descriptor caches should live.
- *
- * SRAM is split up as follows:
- * 0                          buftbl entries for channels
- * efx->vf_buftbl_base        buftbl entries for SR-IOV
- * efx->rx_dc_base            RX descriptor caches
- * efx->tx_dc_base            TX descriptor caches
- */
-void efx_nic_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw)
-{
-       unsigned vi_count, buftbl_min;
-
-       /* Account for the buffer table entries backing the datapath channels
-        * and the descriptor caches for those channels.
-        */
-       buftbl_min = ((efx->n_rx_channels * EFX_MAX_DMAQ_SIZE +
-                      efx->n_tx_channels * EFX_TXQ_TYPES * EFX_MAX_DMAQ_SIZE +
-                      efx->n_channels * EFX_MAX_EVQ_SIZE)
-                     * sizeof(efx_qword_t) / EFX_BUF_SIZE);
-       vi_count = max(efx->n_channels, efx->n_tx_channels * EFX_TXQ_TYPES);
-
-#ifdef CONFIG_SFC_SRIOV
-       if (efx_sriov_wanted(efx)) {
-               unsigned vi_dc_entries, buftbl_free, entries_per_vf, vf_limit;
-
-               efx->vf_buftbl_base = buftbl_min;
-
-               vi_dc_entries = RX_DC_ENTRIES + TX_DC_ENTRIES;
-               vi_count = max(vi_count, EFX_VI_BASE);
-               buftbl_free = (sram_lim_qw - buftbl_min -
-                              vi_count * vi_dc_entries);
-
-               entries_per_vf = ((vi_dc_entries + EFX_VF_BUFTBL_PER_VI) *
-                                 efx_vf_size(efx));
-               vf_limit = min(buftbl_free / entries_per_vf,
-                              (1024U - EFX_VI_BASE) >> efx->vi_scale);
-
-               if (efx->vf_count > vf_limit) {
-                       netif_err(efx, probe, efx->net_dev,
-                                 "Reducing VF count from from %d to %d\n",
-                                 efx->vf_count, vf_limit);
-                       efx->vf_count = vf_limit;
-               }
-               vi_count += efx->vf_count * efx_vf_size(efx);
-       }
-#endif
-
-       efx->tx_dc_base = sram_lim_qw - vi_count * TX_DC_ENTRIES;
-       efx->rx_dc_base = efx->tx_dc_base - vi_count * RX_DC_ENTRIES;
-}
-
-u32 efx_nic_fpga_ver(struct efx_nic *efx)
-{
-       efx_oword_t altera_build;
-       efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
-       return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
-}
-
-void efx_nic_init_common(struct efx_nic *efx)
-{
-       efx_oword_t temp;
-
-       /* Set positions of descriptor caches in SRAM. */
-       EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR, efx->tx_dc_base);
-       efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
-       EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR, efx->rx_dc_base);
-       efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
-
-       /* Set TX descriptor cache size. */
-       BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
-       EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
-       efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
-
-       /* Set RX descriptor cache size.  Set low watermark to size-8, as
-        * this allows most efficient prefetching.
-        */
-       BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
-       EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
-       efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
-       EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
-       efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
-
-       /* Program INT_KER address */
-       EFX_POPULATE_OWORD_2(temp,
-                            FRF_AZ_NORM_INT_VEC_DIS_KER,
-                            EFX_INT_MODE_USE_MSI(efx),
-                            FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
-       efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
-
-       if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
-               /* Use an interrupt level unused by event queues */
-               efx->irq_level = 0x1f;
-       else
-               /* Use a valid MSI-X vector */
-               efx->irq_level = 0;
-
-       /* Enable all the genuinely fatal interrupts.  (They are still
-        * masked by the overall interrupt mask, controlled by
-        * falcon_interrupts()).
-        *
-        * Note: All other fatal interrupts are enabled
-        */
-       EFX_POPULATE_OWORD_3(temp,
-                            FRF_AZ_ILL_ADR_INT_KER_EN, 1,
-                            FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
-                            FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
-       if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
-               EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
-       EFX_INVERT_OWORD(temp);
-       efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
-
-       efx_nic_push_rx_indir_table(efx);
-
-       /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
-        * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
-        */
-       efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
-       /* Enable SW_EV to inherit in char driver - assume harmless here */
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
-       /* Prefetch threshold 2 => fetch when descriptor cache half empty */
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
-       /* Disable hardware watchdog which can misfire */
-       EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
-       /* Squash TX of packets of 16 bytes or less */
-       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
-               EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
-       efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
-
-       if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
-               EFX_POPULATE_OWORD_4(temp,
-                                    /* Default values */
-                                    FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
-                                    FRF_BZ_TX_PACE_SB_AF, 0xb,
-                                    FRF_BZ_TX_PACE_FB_BASE, 0,
-                                    /* Allow large pace values in the
-                                     * fast bin. */
-                                    FRF_BZ_TX_PACE_BIN_TH,
-                                    FFE_BZ_TX_PACE_RESERVED);
-               efx_writeo(efx, &temp, FR_BZ_TX_PACE);
-       }
-}
-
 /* Register dump */
 
 #define REGISTER_REVISION_A    1
index 21f662cc39a465cf1005f71bf6deb03a54d2f773..25e25b6357982c7089d4ec901601e11154e0dd4d 100644 (file)
@@ -34,7 +34,7 @@ static inline int efx_nic_rev(struct efx_nic *efx)
        return efx->type->revision;
 }
 
-extern u32 efx_nic_fpga_ver(struct efx_nic *efx);
+extern u32 efx_farch_fpga_ver(struct efx_nic *efx);
 
 /* NIC has two interlinked PCI functions for the same port. */
 static inline bool efx_nic_is_dual_func(struct efx_nic *efx)
@@ -42,6 +42,65 @@ static inline bool efx_nic_is_dual_func(struct efx_nic *efx)
        return efx_nic_rev(efx) < EFX_REV_FALCON_B0;
 }
 
+/* Read the current event from the event queue */
+static inline efx_qword_t *efx_event(struct efx_channel *channel,
+                                    unsigned int index)
+{
+       return ((efx_qword_t *) (channel->eventq.buf.addr)) +
+               (index & channel->eventq_mask);
+}
+
+/* See if an event is present
+ *
+ * We check both the high and low dword of the event for all ones.  We
+ * wrote all ones when we cleared the event, and no valid event can
+ * have all ones in either its high or low dwords.  This approach is
+ * robust against reordering.
+ *
+ * Note that using a single 64-bit comparison is incorrect; even
+ * though the CPU read will be atomic, the DMA write may not be.
+ */
+static inline int efx_event_present(efx_qword_t *event)
+{
+       return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
+                 EFX_DWORD_IS_ALL_ONES(event->dword[1]));
+}
+
+/* Returns a pointer to the specified transmit descriptor in the TX
+ * descriptor queue belonging to the specified channel.
+ */
+static inline efx_qword_t *
+efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
+{
+       return ((efx_qword_t *) (tx_queue->txd.buf.addr)) + index;
+}
+
+/* Decide whether to push a TX descriptor to the NIC vs merely writing
+ * the doorbell.  This can reduce latency when we are adding a single
+ * descriptor to an empty queue, but is otherwise pointless.  Further,
+ * Falcon and Siena have hardware bugs (SF bug 33851) that may be
+ * triggered if we don't check this.
+ */
+static inline bool efx_nic_may_push_tx_desc(struct efx_tx_queue *tx_queue,
+                                           unsigned int write_count)
+{
+       unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
+
+       if (empty_read_count == 0)
+               return false;
+
+       tx_queue->empty_read_count = 0;
+       return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0
+               && tx_queue->write_count - write_count == 1;
+}
+
+/* Returns a pointer to the specified descriptor in the RX descriptor queue */
+static inline efx_qword_t *
+efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
+{
+       return ((efx_qword_t *) (rx_queue->rxd.buf.addr)) + index;
+}
+
 enum {
        PHY_TYPE_NONE = 0,
        PHY_TYPE_TXC43128 = 1,
@@ -258,25 +317,93 @@ extern const struct efx_nic_type siena_a0_nic_type;
 extern int falcon_probe_board(struct efx_nic *efx, u16 revision_info);
 
 /* TX data path */
-extern int efx_nic_probe_tx(struct efx_tx_queue *tx_queue);
-extern void efx_nic_init_tx(struct efx_tx_queue *tx_queue);
-extern void efx_nic_remove_tx(struct efx_tx_queue *tx_queue);
-extern void efx_nic_push_buffers(struct efx_tx_queue *tx_queue);
+static inline int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
+{
+       return tx_queue->efx->type->tx_probe(tx_queue);
+}
+static inline void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
+{
+       tx_queue->efx->type->tx_init(tx_queue);
+}
+static inline void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
+{
+       tx_queue->efx->type->tx_remove(tx_queue);
+}
+static inline void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
+{
+       tx_queue->efx->type->tx_write(tx_queue);
+}
 
 /* RX data path */
-extern int efx_nic_probe_rx(struct efx_rx_queue *rx_queue);
-extern void efx_nic_init_rx(struct efx_rx_queue *rx_queue);
-extern void efx_nic_remove_rx(struct efx_rx_queue *rx_queue);
-extern void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue);
-extern void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue);
+static inline int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
+{
+       return rx_queue->efx->type->rx_probe(rx_queue);
+}
+static inline void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
+{
+       rx_queue->efx->type->rx_init(rx_queue);
+}
+static inline void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
+{
+       rx_queue->efx->type->rx_remove(rx_queue);
+}
+static inline void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
+{
+       rx_queue->efx->type->rx_write(rx_queue);
+}
+static inline void efx_nic_generate_fill_event(struct efx_rx_queue *rx_queue)
+{
+       rx_queue->efx->type->rx_defer_refill(rx_queue);
+}
 
 /* Event data path */
-extern int efx_nic_probe_eventq(struct efx_channel *channel);
-extern void efx_nic_init_eventq(struct efx_channel *channel);
-extern void efx_nic_fini_eventq(struct efx_channel *channel);
-extern void efx_nic_remove_eventq(struct efx_channel *channel);
-extern int efx_nic_process_eventq(struct efx_channel *channel, int rx_quota);
-extern void efx_nic_eventq_read_ack(struct efx_channel *channel);
+static inline int efx_nic_probe_eventq(struct efx_channel *channel)
+{
+       return channel->efx->type->ev_probe(channel);
+}
+static inline void efx_nic_init_eventq(struct efx_channel *channel)
+{
+       channel->efx->type->ev_init(channel);
+}
+static inline void efx_nic_fini_eventq(struct efx_channel *channel)
+{
+       channel->efx->type->ev_fini(channel);
+}
+static inline void efx_nic_remove_eventq(struct efx_channel *channel)
+{
+       channel->efx->type->ev_remove(channel);
+}
+static inline int
+efx_nic_process_eventq(struct efx_channel *channel, int quota)
+{
+       return channel->efx->type->ev_process(channel, quota);
+}
+static inline void efx_nic_eventq_read_ack(struct efx_channel *channel)
+{
+       channel->efx->type->ev_read_ack(channel);
+}
+extern void efx_nic_event_test_start(struct efx_channel *channel);
+
+/* Falcon/Siena queue operations */
+extern int efx_farch_tx_probe(struct efx_tx_queue *tx_queue);
+extern void efx_farch_tx_init(struct efx_tx_queue *tx_queue);
+extern void efx_farch_tx_fini(struct efx_tx_queue *tx_queue);
+extern void efx_farch_tx_remove(struct efx_tx_queue *tx_queue);
+extern void efx_farch_tx_write(struct efx_tx_queue *tx_queue);
+extern int efx_farch_rx_probe(struct efx_rx_queue *rx_queue);
+extern void efx_farch_rx_init(struct efx_rx_queue *rx_queue);
+extern void efx_farch_rx_fini(struct efx_rx_queue *rx_queue);
+extern void efx_farch_rx_remove(struct efx_rx_queue *rx_queue);
+extern void efx_farch_rx_write(struct efx_rx_queue *rx_queue);
+extern void efx_farch_rx_defer_refill(struct efx_rx_queue *rx_queue);
+extern int efx_farch_ev_probe(struct efx_channel *channel);
+extern void efx_farch_ev_init(struct efx_channel *channel);
+extern void efx_farch_ev_fini(struct efx_channel *channel);
+extern void efx_farch_ev_remove(struct efx_channel *channel);
+extern int efx_farch_ev_process(struct efx_channel *channel, int quota);
+extern void efx_farch_ev_read_ack(struct efx_channel *channel);
+extern void efx_farch_ev_test_generate(struct efx_channel *channel);
+
 extern bool efx_nic_event_present(struct efx_channel *channel);
 
 /* Some statistics are computed as A - B where A and B each increase
@@ -297,15 +424,18 @@ static inline void efx_update_diff_stat(u64 *stat, u64 diff)
                *stat = diff;
 }
 
-/* Interrupts and test events */
+/* Interrupts */
 extern int efx_nic_init_interrupt(struct efx_nic *efx);
-extern void efx_nic_enable_interrupts(struct efx_nic *efx);
-extern void efx_nic_event_test_start(struct efx_channel *channel);
 extern void efx_nic_irq_test_start(struct efx_nic *efx);
-extern void efx_nic_disable_interrupts(struct efx_nic *efx);
 extern void efx_nic_fini_interrupt(struct efx_nic *efx);
-extern irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx);
-extern irqreturn_t falcon_legacy_interrupt_a1(int irq, void *dev_id);
+
+/* Falcon/Siena interrupts */
+extern void efx_farch_irq_enable_master(struct efx_nic *efx);
+extern void efx_farch_irq_test_generate(struct efx_nic *efx);
+extern void efx_farch_irq_disable_master(struct efx_nic *efx);
+extern irqreturn_t efx_farch_msi_interrupt(int irq, void *dev_id);
+extern irqreturn_t efx_farch_legacy_interrupt(int irq, void *dev_id);
+extern irqreturn_t efx_farch_fatal_interrupt(struct efx_nic *efx);
 
 static inline int efx_nic_event_test_irq_cpu(struct efx_channel *channel)
 {
@@ -317,36 +447,40 @@ static inline int efx_nic_irq_test_irq_cpu(struct efx_nic *efx)
 }
 
 /* Global Resources */
-extern int efx_farch_fini_dmaq(struct efx_nic *efx);
+extern int efx_nic_flush_queues(struct efx_nic *efx);
 extern void siena_prepare_flush(struct efx_nic *efx);
+extern int efx_farch_fini_dmaq(struct efx_nic *efx);
 extern void siena_finish_flush(struct efx_nic *efx);
 extern void falcon_start_nic_stats(struct efx_nic *efx);
 extern void falcon_stop_nic_stats(struct efx_nic *efx);
 extern int falcon_reset_xaui(struct efx_nic *efx);
-extern void
-efx_nic_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw);
-extern void efx_nic_init_common(struct efx_nic *efx);
-extern void efx_nic_push_rx_indir_table(struct efx_nic *efx);
+extern void efx_farch_dimension_resources(struct efx_nic *efx, unsigned sram_lim_qw);
+extern void efx_farch_init_common(struct efx_nic *efx);
+static inline void efx_nic_push_rx_indir_table(struct efx_nic *efx)
+{
+       efx->type->rx_push_indir_table(efx);
+}
+extern void efx_farch_rx_push_indir_table(struct efx_nic *efx);
 
 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
                         unsigned int len, gfp_t gfp_flags);
 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer);
 
 /* Tests */
-struct efx_nic_register_test {
+struct efx_farch_register_test {
        unsigned address;
        efx_oword_t mask;
 };
-extern int efx_nic_test_registers(struct efx_nic *efx,
-                                 const struct efx_nic_register_test *regs,
-                                 size_t n_regs);
+extern int efx_farch_test_registers(struct efx_nic *efx,
+                                   const struct efx_farch_register_test *regs,
+                                   size_t n_regs);
 
 extern size_t efx_nic_get_regs_len(struct efx_nic *efx);
 extern void efx_nic_get_regs(struct efx_nic *efx, void *buf);
 
 #define EFX_MAX_FLUSH_TIME 5000
 
-extern void efx_generate_event(struct efx_nic *efx, unsigned int evq,
-                              efx_qword_t *event);
+extern void efx_farch_generate_event(struct efx_nic *efx, unsigned int evq,
+                                    efx_qword_t *event);
 
 #endif /* EFX_NIC_H */
index b4c1d4310afe39d3a13179beada3a7dac35fdcc4..6a833d531b86645778d2c1255bc5e5d240d892d4 100644 (file)
@@ -63,7 +63,7 @@ void siena_finish_flush(struct efx_nic *efx)
                efx_mcdi_set_mac(efx);
 }
 
-static const struct efx_nic_register_test siena_register_tests[] = {
+static const struct efx_farch_register_test siena_register_tests[] = {
        { FR_AZ_ADR_REGION,
          EFX_OWORD32(0x0003FFFF, 0x0003FFFF, 0x0003FFFF, 0x0003FFFF) },
        { FR_CZ_USR_EV_CFG,
@@ -107,8 +107,8 @@ static int siena_test_chip(struct efx_nic *efx, struct efx_self_tests *tests)
                goto out;
 
        tests->registers =
-               efx_nic_test_registers(efx, siena_register_tests,
-                                      ARRAY_SIZE(siena_register_tests))
+               efx_farch_test_registers(efx, siena_register_tests,
+                                        ARRAY_SIZE(siena_register_tests))
                ? -1 : 1;
 
        rc = efx_mcdi_reset(efx, reset_method);
@@ -184,7 +184,7 @@ static void siena_dimension_resources(struct efx_nic *efx)
         * the buffer table and descriptor caches.  In theory we can
         * map both blocks to one port, but we don't.
         */
-       efx_nic_dimension_resources(efx, FR_CZ_BUF_FULL_TBL_ROWS / 2);
+       efx_farch_dimension_resources(efx, FR_CZ_BUF_FULL_TBL_ROWS / 2);
 }
 
 static int siena_probe_nic(struct efx_nic *efx)
@@ -200,7 +200,7 @@ static int siena_probe_nic(struct efx_nic *efx)
                return -ENOMEM;
        efx->nic_data = nic_data;
 
-       if (efx_nic_fpga_ver(efx) != 0) {
+       if (efx_farch_fpga_ver(efx) != 0) {
                netif_err(efx, probe, efx->net_dev,
                          "Siena FPGA not supported\n");
                rc = -ENODEV;
@@ -351,7 +351,7 @@ static int siena_init_nic(struct efx_nic *efx)
        EFX_POPULATE_OWORD_1(temp, FRF_CZ_USREV_DIS, 1);
        efx_writeo(efx, &temp, FR_CZ_USR_EV_CFG);
 
-       efx_nic_init_common(efx);
+       efx_farch_init_common(efx);
        return 0;
 }
 
@@ -705,6 +705,28 @@ const struct efx_nic_type siena_a0_nic_type = {
        .mcdi_poll_response = siena_mcdi_poll_response,
        .mcdi_read_response = siena_mcdi_read_response,
        .mcdi_poll_reboot = siena_mcdi_poll_reboot,
+       .irq_enable_master = efx_farch_irq_enable_master,
+       .irq_test_generate = efx_farch_irq_test_generate,
+       .irq_disable_non_ev = efx_farch_irq_disable_master,
+       .irq_handle_msi = efx_farch_msi_interrupt,
+       .irq_handle_legacy = efx_farch_legacy_interrupt,
+       .tx_probe = efx_farch_tx_probe,
+       .tx_init = efx_farch_tx_init,
+       .tx_remove = efx_farch_tx_remove,
+       .tx_write = efx_farch_tx_write,
+       .rx_push_indir_table = efx_farch_rx_push_indir_table,
+       .rx_probe = efx_farch_rx_probe,
+       .rx_init = efx_farch_rx_init,
+       .rx_remove = efx_farch_rx_remove,
+       .rx_write = efx_farch_rx_write,
+       .rx_defer_refill = efx_farch_rx_defer_refill,
+       .ev_probe = efx_farch_ev_probe,
+       .ev_init = efx_farch_ev_init,
+       .ev_fini = efx_farch_ev_fini,
+       .ev_remove = efx_farch_ev_remove,
+       .ev_process = efx_farch_ev_process,
+       .ev_read_ack = efx_farch_ev_read_ack,
+       .ev_test_generate = efx_farch_ev_test_generate,
 
        .revision = EFX_REV_SIENA_A0,
        .mem_map_size = (FR_CZ_MC_TREG_SMEM +
index 4d214bce896946ef66ac87cbf5f85e36858b6d51..4b8eef962faa03ca65420fc2d7cf7f17dd5234f1 100644 (file)
@@ -464,8 +464,9 @@ static void __efx_sriov_push_vf_status(struct efx_vf *vf)
                             VFDI_EV_SEQ, (vf->msg_seqno & 0xff),
                             VFDI_EV_TYPE, VFDI_EV_TYPE_STATUS);
        ++vf->msg_seqno;
-       efx_generate_event(efx, EFX_VI_BASE + vf->index * efx_vf_size(efx),
-                             &event);
+       efx_farch_generate_event(efx,
+                                EFX_VI_BASE + vf->index * efx_vf_size(efx),
+                                &event);
 }
 
 static void efx_sriov_bufs(struct efx_nic *efx, unsigned offset,