dmaengine: add an shdma-base library
authorGuennadi Liakhovetski <g.liakhovetski@gmx.de>
Wed, 9 May 2012 15:09:13 +0000 (17:09 +0200)
committerVinod Koul <vinod.koul@linux.intel.com>
Fri, 13 Jul 2012 03:43:07 +0000 (09:13 +0530)
This patch extracts code from shdma.c, that does not directly deal with
hardware implementation details and can be re-used with diverse DMA
controller variants, found on SH-based SoCs.

Signed-off-by: Guennadi Liakhovetski <g.liakhovetski@gmx.de>
Cc: Sascha Hauer <s.hauer@pengutronix.de>
Signed-off-by: Vinod Koul <vinod.koul@linux.intel.com>
drivers/dma/sh/Makefile
drivers/dma/sh/shdma-base.c [new file with mode: 0644]
include/linux/shdma-base.h [new file with mode: 0644]

index ad4981af38b24da62f2b441d644ea0a223c5fa68..54ae9572b0ac1f2735d7a8fe3e47adb4c87a6373 100644 (file)
@@ -1 +1,2 @@
+obj-$(CONFIG_SH_DMAE) += shdma-base.o
 obj-$(CONFIG_SH_DMAE) += shdma.o
diff --git a/drivers/dma/sh/shdma-base.c b/drivers/dma/sh/shdma-base.c
new file mode 100644 (file)
index 0000000..ff060d0
--- /dev/null
@@ -0,0 +1,868 @@
+/*
+ * Dmaengine driver base library for DMA controllers, found on SH-based SoCs
+ *
+ * extracted from shdma.c
+ *
+ * Copyright (C) 2011-2012 Guennadi Liakhovetski <g.liakhovetski@gmx.de>
+ * Copyright (C) 2009 Nobuhiro Iwamatsu <iwamatsu.nobuhiro@renesas.com>
+ * Copyright (C) 2009 Renesas Solutions, Inc. All rights reserved.
+ * Copyright (C) 2007 Freescale Semiconductor, Inc. All rights reserved.
+ *
+ * This is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ */
+
+#include <linux/delay.h>
+#include <linux/shdma-base.h>
+#include <linux/dmaengine.h>
+#include <linux/init.h>
+#include <linux/interrupt.h>
+#include <linux/module.h>
+#include <linux/pm_runtime.h>
+#include <linux/slab.h>
+#include <linux/spinlock.h>
+
+#include "../dmaengine.h"
+
+/* DMA descriptor control */
+enum shdma_desc_status {
+       DESC_IDLE,
+       DESC_PREPARED,
+       DESC_SUBMITTED,
+       DESC_COMPLETED, /* completed, have to call callback */
+       DESC_WAITING,   /* callback called, waiting for ack / re-submit */
+};
+
+#define NR_DESCS_PER_CHANNEL 32
+
+#define to_shdma_chan(c) container_of(c, struct shdma_chan, dma_chan)
+#define to_shdma_dev(d) container_of(d, struct shdma_dev, dma_dev)
+
+/*
+ * For slave DMA we assume, that there is a finite number of DMA slaves in the
+ * system, and that each such slave can only use a finite number of channels.
+ * We use slave channel IDs to make sure, that no such slave channel ID is
+ * allocated more than once.
+ */
+static unsigned int slave_num = 256;
+module_param(slave_num, uint, 0444);
+
+/* A bitmask with slave_num bits */
+static unsigned long *shdma_slave_used;
+
+/* Called under spin_lock_irq(&schan->chan_lock") */
+static void shdma_chan_xfer_ld_queue(struct shdma_chan *schan)
+{
+       struct shdma_dev *sdev = to_shdma_dev(schan->dma_chan.device);
+       const struct shdma_ops *ops = sdev->ops;
+       struct shdma_desc *sdesc;
+
+       /* DMA work check */
+       if (ops->channel_busy(schan))
+               return;
+
+       /* Find the first not transferred descriptor */
+       list_for_each_entry(sdesc, &schan->ld_queue, node)
+               if (sdesc->mark == DESC_SUBMITTED) {
+                       ops->start_xfer(schan, sdesc);
+                       break;
+               }
+}
+
+static dma_cookie_t shdma_tx_submit(struct dma_async_tx_descriptor *tx)
+{
+       struct shdma_desc *chunk, *c, *desc =
+               container_of(tx, struct shdma_desc, async_tx),
+               *last = desc;
+       struct shdma_chan *schan = to_shdma_chan(tx->chan);
+       struct shdma_slave *slave = tx->chan->private;
+       dma_async_tx_callback callback = tx->callback;
+       dma_cookie_t cookie;
+       bool power_up;
+
+       spin_lock_irq(&schan->chan_lock);
+
+       power_up = list_empty(&schan->ld_queue);
+
+       cookie = dma_cookie_assign(tx);
+
+       /* Mark all chunks of this descriptor as submitted, move to the queue */
+       list_for_each_entry_safe(chunk, c, desc->node.prev, node) {
+               /*
+                * All chunks are on the global ld_free, so, we have to find
+                * the end of the chain ourselves
+                */
+               if (chunk != desc && (chunk->mark == DESC_IDLE ||
+                                     chunk->async_tx.cookie > 0 ||
+                                     chunk->async_tx.cookie == -EBUSY ||
+                                     &chunk->node == &schan->ld_free))
+                       break;
+               chunk->mark = DESC_SUBMITTED;
+               /* Callback goes to the last chunk */
+               chunk->async_tx.callback = NULL;
+               chunk->cookie = cookie;
+               list_move_tail(&chunk->node, &schan->ld_queue);
+               last = chunk;
+
+               dev_dbg(schan->dev, "submit #%d@%p on %d\n",
+                       tx->cookie, &last->async_tx, schan->id);
+       }
+
+       last->async_tx.callback = callback;
+       last->async_tx.callback_param = tx->callback_param;
+
+       if (power_up) {
+               int ret;
+               schan->pm_state = SHDMA_PM_BUSY;
+
+               ret = pm_runtime_get(schan->dev);
+
+               spin_unlock_irq(&schan->chan_lock);
+               if (ret < 0)
+                       dev_err(schan->dev, "%s(): GET = %d\n", __func__, ret);
+
+               pm_runtime_barrier(schan->dev);
+
+               spin_lock_irq(&schan->chan_lock);
+
+               /* Have we been reset, while waiting? */
+               if (schan->pm_state != SHDMA_PM_ESTABLISHED) {
+                       struct shdma_dev *sdev =
+                               to_shdma_dev(schan->dma_chan.device);
+                       const struct shdma_ops *ops = sdev->ops;
+                       dev_dbg(schan->dev, "Bring up channel %d\n",
+                               schan->id);
+                       /*
+                        * TODO: .xfer_setup() might fail on some platforms.
+                        * Make it int then, on error remove chunks from the
+                        * queue again
+                        */
+                       ops->setup_xfer(schan, slave);
+
+                       if (schan->pm_state == SHDMA_PM_PENDING)
+                               shdma_chan_xfer_ld_queue(schan);
+                       schan->pm_state = SHDMA_PM_ESTABLISHED;
+               }
+       } else {
+               /*
+                * Tell .device_issue_pending() not to run the queue, interrupts
+                * will do it anyway
+                */
+               schan->pm_state = SHDMA_PM_PENDING;
+       }
+
+       spin_unlock_irq(&schan->chan_lock);
+
+       return cookie;
+}
+
+/* Called with desc_lock held */
+static struct shdma_desc *shdma_get_desc(struct shdma_chan *schan)
+{
+       struct shdma_desc *sdesc;
+
+       list_for_each_entry(sdesc, &schan->ld_free, node)
+               if (sdesc->mark != DESC_PREPARED) {
+                       BUG_ON(sdesc->mark != DESC_IDLE);
+                       list_del(&sdesc->node);
+                       return sdesc;
+               }
+
+       return NULL;
+}
+
+static int shdma_alloc_chan_resources(struct dma_chan *chan)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+       struct shdma_dev *sdev = to_shdma_dev(schan->dma_chan.device);
+       const struct shdma_ops *ops = sdev->ops;
+       struct shdma_desc *desc;
+       struct shdma_slave *slave = chan->private;
+       int ret, i;
+
+       /*
+        * This relies on the guarantee from dmaengine that alloc_chan_resources
+        * never runs concurrently with itself or free_chan_resources.
+        */
+       if (slave) {
+               if (slave->slave_id >= slave_num) {
+                       ret = -EINVAL;
+                       goto evalid;
+               }
+
+               if (test_and_set_bit(slave->slave_id, shdma_slave_used)) {
+                       ret = -EBUSY;
+                       goto etestused;
+               }
+
+               ret = ops->set_slave(schan, slave);
+               if (ret < 0)
+                       goto esetslave;
+       }
+
+       schan->desc = kcalloc(NR_DESCS_PER_CHANNEL,
+                             sdev->desc_size, GFP_KERNEL);
+       if (!schan->desc) {
+               ret = -ENOMEM;
+               goto edescalloc;
+       }
+       schan->desc_num = NR_DESCS_PER_CHANNEL;
+
+       for (i = 0; i < NR_DESCS_PER_CHANNEL; i++) {
+               desc = ops->embedded_desc(schan->desc, i);
+               dma_async_tx_descriptor_init(&desc->async_tx,
+                                            &schan->dma_chan);
+               desc->async_tx.tx_submit = shdma_tx_submit;
+               desc->mark = DESC_IDLE;
+
+               list_add(&desc->node, &schan->ld_free);
+       }
+
+       return NR_DESCS_PER_CHANNEL;
+
+edescalloc:
+       if (slave)
+esetslave:
+               clear_bit(slave->slave_id, shdma_slave_used);
+etestused:
+evalid:
+       chan->private = NULL;
+       return ret;
+}
+
+static dma_async_tx_callback __ld_cleanup(struct shdma_chan *schan, bool all)
+{
+       struct shdma_desc *desc, *_desc;
+       /* Is the "exposed" head of a chain acked? */
+       bool head_acked = false;
+       dma_cookie_t cookie = 0;
+       dma_async_tx_callback callback = NULL;
+       void *param = NULL;
+       unsigned long flags;
+
+       spin_lock_irqsave(&schan->chan_lock, flags);
+       list_for_each_entry_safe(desc, _desc, &schan->ld_queue, node) {
+               struct dma_async_tx_descriptor *tx = &desc->async_tx;
+
+               BUG_ON(tx->cookie > 0 && tx->cookie != desc->cookie);
+               BUG_ON(desc->mark != DESC_SUBMITTED &&
+                      desc->mark != DESC_COMPLETED &&
+                      desc->mark != DESC_WAITING);
+
+               /*
+                * queue is ordered, and we use this loop to (1) clean up all
+                * completed descriptors, and to (2) update descriptor flags of
+                * any chunks in a (partially) completed chain
+                */
+               if (!all && desc->mark == DESC_SUBMITTED &&
+                   desc->cookie != cookie)
+                       break;
+
+               if (tx->cookie > 0)
+                       cookie = tx->cookie;
+
+               if (desc->mark == DESC_COMPLETED && desc->chunks == 1) {
+                       if (schan->dma_chan.completed_cookie != desc->cookie - 1)
+                               dev_dbg(schan->dev,
+                                       "Completing cookie %d, expected %d\n",
+                                       desc->cookie,
+                                       schan->dma_chan.completed_cookie + 1);
+                       schan->dma_chan.completed_cookie = desc->cookie;
+               }
+
+               /* Call callback on the last chunk */
+               if (desc->mark == DESC_COMPLETED && tx->callback) {
+                       desc->mark = DESC_WAITING;
+                       callback = tx->callback;
+                       param = tx->callback_param;
+                       dev_dbg(schan->dev, "descriptor #%d@%p on %d callback\n",
+                               tx->cookie, tx, schan->id);
+                       BUG_ON(desc->chunks != 1);
+                       break;
+               }
+
+               if (tx->cookie > 0 || tx->cookie == -EBUSY) {
+                       if (desc->mark == DESC_COMPLETED) {
+                               BUG_ON(tx->cookie < 0);
+                               desc->mark = DESC_WAITING;
+                       }
+                       head_acked = async_tx_test_ack(tx);
+               } else {
+                       switch (desc->mark) {
+                       case DESC_COMPLETED:
+                               desc->mark = DESC_WAITING;
+                               /* Fall through */
+                       case DESC_WAITING:
+                               if (head_acked)
+                                       async_tx_ack(&desc->async_tx);
+                       }
+               }
+
+               dev_dbg(schan->dev, "descriptor %p #%d completed.\n",
+                       tx, tx->cookie);
+
+               if (((desc->mark == DESC_COMPLETED ||
+                     desc->mark == DESC_WAITING) &&
+                    async_tx_test_ack(&desc->async_tx)) || all) {
+                       /* Remove from ld_queue list */
+                       desc->mark = DESC_IDLE;
+
+                       list_move(&desc->node, &schan->ld_free);
+
+                       if (list_empty(&schan->ld_queue)) {
+                               dev_dbg(schan->dev, "Bring down channel %d\n", schan->id);
+                               pm_runtime_put(schan->dev);
+                               schan->pm_state = SHDMA_PM_ESTABLISHED;
+                       }
+               }
+       }
+
+       if (all && !callback)
+               /*
+                * Terminating and the loop completed normally: forgive
+                * uncompleted cookies
+                */
+               schan->dma_chan.completed_cookie = schan->dma_chan.cookie;
+
+       spin_unlock_irqrestore(&schan->chan_lock, flags);
+
+       if (callback)
+               callback(param);
+
+       return callback;
+}
+
+/*
+ * shdma_chan_ld_cleanup - Clean up link descriptors
+ *
+ * Clean up the ld_queue of DMA channel.
+ */
+static void shdma_chan_ld_cleanup(struct shdma_chan *schan, bool all)
+{
+       while (__ld_cleanup(schan, all))
+               ;
+}
+
+/*
+ * shdma_free_chan_resources - Free all resources of the channel.
+ */
+static void shdma_free_chan_resources(struct dma_chan *chan)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+       struct shdma_dev *sdev = to_shdma_dev(chan->device);
+       const struct shdma_ops *ops = sdev->ops;
+       LIST_HEAD(list);
+
+       /* Protect against ISR */
+       spin_lock_irq(&schan->chan_lock);
+       ops->halt_channel(schan);
+       spin_unlock_irq(&schan->chan_lock);
+
+       /* Now no new interrupts will occur */
+
+       /* Prepared and not submitted descriptors can still be on the queue */
+       if (!list_empty(&schan->ld_queue))
+               shdma_chan_ld_cleanup(schan, true);
+
+       if (chan->private) {
+               /* The caller is holding dma_list_mutex */
+               struct shdma_slave *slave = chan->private;
+               clear_bit(slave->slave_id, shdma_slave_used);
+               chan->private = NULL;
+       }
+
+       spin_lock_irq(&schan->chan_lock);
+
+       list_splice_init(&schan->ld_free, &list);
+       schan->desc_num = 0;
+
+       spin_unlock_irq(&schan->chan_lock);
+
+       kfree(schan->desc);
+}
+
+/**
+ * shdma_add_desc - get, set up and return one transfer descriptor
+ * @schan:     DMA channel
+ * @flags:     DMA transfer flags
+ * @dst:       destination DMA address, incremented when direction equals
+ *             DMA_DEV_TO_MEM or DMA_MEM_TO_MEM
+ * @src:       source DMA address, incremented when direction equals
+ *             DMA_MEM_TO_DEV or DMA_MEM_TO_MEM
+ * @len:       DMA transfer length
+ * @first:     if NULL, set to the current descriptor and cookie set to -EBUSY
+ * @direction: needed for slave DMA to decide which address to keep constant,
+ *             equals DMA_MEM_TO_MEM for MEMCPY
+ * Returns 0 or an error
+ * Locks: called with desc_lock held
+ */
+static struct shdma_desc *shdma_add_desc(struct shdma_chan *schan,
+       unsigned long flags, dma_addr_t *dst, dma_addr_t *src, size_t *len,
+       struct shdma_desc **first, enum dma_transfer_direction direction)
+{
+       struct shdma_dev *sdev = to_shdma_dev(schan->dma_chan.device);
+       const struct shdma_ops *ops = sdev->ops;
+       struct shdma_desc *new;
+       size_t copy_size = *len;
+
+       if (!copy_size)
+               return NULL;
+
+       /* Allocate the link descriptor from the free list */
+       new = shdma_get_desc(schan);
+       if (!new) {
+               dev_err(schan->dev, "No free link descriptor available\n");
+               return NULL;
+       }
+
+       ops->desc_setup(schan, new, *src, *dst, &copy_size);
+
+       if (!*first) {
+               /* First desc */
+               new->async_tx.cookie = -EBUSY;
+               *first = new;
+       } else {
+               /* Other desc - invisible to the user */
+               new->async_tx.cookie = -EINVAL;
+       }
+
+       dev_dbg(schan->dev,
+               "chaining (%u/%u)@%x -> %x with %p, cookie %d\n",
+               copy_size, *len, *src, *dst, &new->async_tx,
+               new->async_tx.cookie);
+
+       new->mark = DESC_PREPARED;
+       new->async_tx.flags = flags;
+       new->direction = direction;
+
+       *len -= copy_size;
+       if (direction == DMA_MEM_TO_MEM || direction == DMA_MEM_TO_DEV)
+               *src += copy_size;
+       if (direction == DMA_MEM_TO_MEM || direction == DMA_DEV_TO_MEM)
+               *dst += copy_size;
+
+       return new;
+}
+
+/*
+ * shdma_prep_sg - prepare transfer descriptors from an SG list
+ *
+ * Common routine for public (MEMCPY) and slave DMA. The MEMCPY case is also
+ * converted to scatter-gather to guarantee consistent locking and a correct
+ * list manipulation. For slave DMA direction carries the usual meaning, and,
+ * logically, the SG list is RAM and the addr variable contains slave address,
+ * e.g., the FIFO I/O register. For MEMCPY direction equals DMA_MEM_TO_MEM
+ * and the SG list contains only one element and points at the source buffer.
+ */
+static struct dma_async_tx_descriptor *shdma_prep_sg(struct shdma_chan *schan,
+       struct scatterlist *sgl, unsigned int sg_len, dma_addr_t *addr,
+       enum dma_transfer_direction direction, unsigned long flags)
+{
+       struct scatterlist *sg;
+       struct shdma_desc *first = NULL, *new = NULL /* compiler... */;
+       LIST_HEAD(tx_list);
+       int chunks = 0;
+       unsigned long irq_flags;
+       int i;
+
+       for_each_sg(sgl, sg, sg_len, i)
+               chunks += DIV_ROUND_UP(sg_dma_len(sg), schan->max_xfer_len);
+
+       /* Have to lock the whole loop to protect against concurrent release */
+       spin_lock_irqsave(&schan->chan_lock, irq_flags);
+
+       /*
+        * Chaining:
+        * first descriptor is what user is dealing with in all API calls, its
+        *      cookie is at first set to -EBUSY, at tx-submit to a positive
+        *      number
+        * if more than one chunk is needed further chunks have cookie = -EINVAL
+        * the last chunk, if not equal to the first, has cookie = -ENOSPC
+        * all chunks are linked onto the tx_list head with their .node heads
+        *      only during this function, then they are immediately spliced
+        *      back onto the free list in form of a chain
+        */
+       for_each_sg(sgl, sg, sg_len, i) {
+               dma_addr_t sg_addr = sg_dma_address(sg);
+               size_t len = sg_dma_len(sg);
+
+               if (!len)
+                       goto err_get_desc;
+
+               do {
+                       dev_dbg(schan->dev, "Add SG #%d@%p[%d], dma %llx\n",
+                               i, sg, len, (unsigned long long)sg_addr);
+
+                       if (direction == DMA_DEV_TO_MEM)
+                               new = shdma_add_desc(schan, flags,
+                                               &sg_addr, addr, &len, &first,
+                                               direction);
+                       else
+                               new = shdma_add_desc(schan, flags,
+                                               addr, &sg_addr, &len, &first,
+                                               direction);
+                       if (!new)
+                               goto err_get_desc;
+
+                       new->chunks = chunks--;
+                       list_add_tail(&new->node, &tx_list);
+               } while (len);
+       }
+
+       if (new != first)
+               new->async_tx.cookie = -ENOSPC;
+
+       /* Put them back on the free list, so, they don't get lost */
+       list_splice_tail(&tx_list, &schan->ld_free);
+
+       spin_unlock_irqrestore(&schan->chan_lock, irq_flags);
+
+       return &first->async_tx;
+
+err_get_desc:
+       list_for_each_entry(new, &tx_list, node)
+               new->mark = DESC_IDLE;
+       list_splice(&tx_list, &schan->ld_free);
+
+       spin_unlock_irqrestore(&schan->chan_lock, irq_flags);
+
+       return NULL;
+}
+
+static struct dma_async_tx_descriptor *shdma_prep_memcpy(
+       struct dma_chan *chan, dma_addr_t dma_dest, dma_addr_t dma_src,
+       size_t len, unsigned long flags)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+       struct scatterlist sg;
+
+       if (!chan || !len)
+               return NULL;
+
+       BUG_ON(!schan->desc_num);
+
+       sg_init_table(&sg, 1);
+       sg_set_page(&sg, pfn_to_page(PFN_DOWN(dma_src)), len,
+                   offset_in_page(dma_src));
+       sg_dma_address(&sg) = dma_src;
+       sg_dma_len(&sg) = len;
+
+       return shdma_prep_sg(schan, &sg, 1, &dma_dest, DMA_MEM_TO_MEM, flags);
+}
+
+static struct dma_async_tx_descriptor *shdma_prep_slave_sg(
+       struct dma_chan *chan, struct scatterlist *sgl, unsigned int sg_len,
+       enum dma_transfer_direction direction, unsigned long flags, void *context)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+       struct shdma_dev *sdev = to_shdma_dev(schan->dma_chan.device);
+       const struct shdma_ops *ops = sdev->ops;
+       struct shdma_slave *slave = chan->private;
+       dma_addr_t slave_addr;
+
+       if (!chan)
+               return NULL;
+
+       BUG_ON(!schan->desc_num);
+
+       /* Someone calling slave DMA on a generic channel? */
+       if (!slave || !sg_len) {
+               dev_warn(schan->dev, "%s: bad parameter: %p, %d, %d\n",
+                        __func__, slave, sg_len, slave ? slave->slave_id : -1);
+               return NULL;
+       }
+
+       slave_addr = ops->slave_addr(schan);
+
+       return shdma_prep_sg(schan, sgl, sg_len, &slave_addr,
+                             direction, flags);
+}
+
+static int shdma_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
+                         unsigned long arg)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+       struct shdma_dev *sdev = to_shdma_dev(chan->device);
+       const struct shdma_ops *ops = sdev->ops;
+       unsigned long flags;
+
+       /* Only supports DMA_TERMINATE_ALL */
+       if (cmd != DMA_TERMINATE_ALL)
+               return -ENXIO;
+
+       if (!chan)
+               return -EINVAL;
+
+       spin_lock_irqsave(&schan->chan_lock, flags);
+
+       ops->halt_channel(schan);
+
+       spin_unlock_irqrestore(&schan->chan_lock, flags);
+
+       shdma_chan_ld_cleanup(schan, true);
+
+       return 0;
+}
+
+static void shdma_issue_pending(struct dma_chan *chan)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+
+       spin_lock_irq(&schan->chan_lock);
+       if (schan->pm_state == SHDMA_PM_ESTABLISHED)
+               shdma_chan_xfer_ld_queue(schan);
+       else
+               schan->pm_state = SHDMA_PM_PENDING;
+       spin_unlock_irq(&schan->chan_lock);
+}
+
+static enum dma_status shdma_tx_status(struct dma_chan *chan,
+                                       dma_cookie_t cookie,
+                                       struct dma_tx_state *txstate)
+{
+       struct shdma_chan *schan = to_shdma_chan(chan);
+       enum dma_status status;
+       unsigned long flags;
+
+       shdma_chan_ld_cleanup(schan, false);
+
+       spin_lock_irqsave(&schan->chan_lock, flags);
+
+       status = dma_cookie_status(chan, cookie, txstate);
+
+       /*
+        * If we don't find cookie on the queue, it has been aborted and we have
+        * to report error
+        */
+       if (status != DMA_SUCCESS) {
+               struct shdma_desc *sdesc;
+               status = DMA_ERROR;
+               list_for_each_entry(sdesc, &schan->ld_queue, node)
+                       if (sdesc->cookie == cookie) {
+                               status = DMA_IN_PROGRESS;
+                               break;
+                       }
+       }
+
+       spin_unlock_irqrestore(&schan->chan_lock, flags);
+
+       return status;
+}
+
+/* Called from error IRQ or NMI */
+bool shdma_reset(struct shdma_dev *sdev)
+{
+       const struct shdma_ops *ops = sdev->ops;
+       struct shdma_chan *schan;
+       unsigned int handled = 0;
+       int i;
+
+       /* Reset all channels */
+       shdma_for_each_chan(schan, sdev, i) {
+               struct shdma_desc *sdesc;
+               LIST_HEAD(dl);
+
+               if (!schan)
+                       continue;
+
+               spin_lock(&schan->chan_lock);
+
+               /* Stop the channel */
+               ops->halt_channel(schan);
+
+               list_splice_init(&schan->ld_queue, &dl);
+
+               if (!list_empty(&dl)) {
+                       dev_dbg(schan->dev, "Bring down channel %d\n", schan->id);
+                       pm_runtime_put(schan->dev);
+               }
+               schan->pm_state = SHDMA_PM_ESTABLISHED;
+
+               spin_unlock(&schan->chan_lock);
+
+               /* Complete all  */
+               list_for_each_entry(sdesc, &dl, node) {
+                       struct dma_async_tx_descriptor *tx = &sdesc->async_tx;
+                       sdesc->mark = DESC_IDLE;
+                       if (tx->callback)
+                               tx->callback(tx->callback_param);
+               }
+
+               spin_lock(&schan->chan_lock);
+               list_splice(&dl, &schan->ld_free);
+               spin_unlock(&schan->chan_lock);
+
+               handled++;
+       }
+
+       return !!handled;
+}
+EXPORT_SYMBOL(shdma_reset);
+
+static irqreturn_t chan_irq(int irq, void *dev)
+{
+       struct shdma_chan *schan = dev;
+       const struct shdma_ops *ops =
+               to_shdma_dev(schan->dma_chan.device)->ops;
+       irqreturn_t ret;
+
+       spin_lock(&schan->chan_lock);
+
+       ret = ops->chan_irq(schan, irq) ? IRQ_WAKE_THREAD : IRQ_NONE;
+
+       spin_unlock(&schan->chan_lock);
+
+       return ret;
+}
+
+static irqreturn_t chan_irqt(int irq, void *dev)
+{
+       struct shdma_chan *schan = dev;
+       const struct shdma_ops *ops =
+               to_shdma_dev(schan->dma_chan.device)->ops;
+       struct shdma_desc *sdesc;
+
+       spin_lock_irq(&schan->chan_lock);
+       list_for_each_entry(sdesc, &schan->ld_queue, node) {
+               if (sdesc->mark == DESC_SUBMITTED &&
+                   ops->desc_completed(schan, sdesc)) {
+                       dev_dbg(schan->dev, "done #%d@%p\n",
+                               sdesc->async_tx.cookie, &sdesc->async_tx);
+                       sdesc->mark = DESC_COMPLETED;
+                       break;
+               }
+       }
+       /* Next desc */
+       shdma_chan_xfer_ld_queue(schan);
+       spin_unlock_irq(&schan->chan_lock);
+
+       shdma_chan_ld_cleanup(schan, false);
+
+       return IRQ_HANDLED;
+}
+
+int shdma_request_irq(struct shdma_chan *schan, int irq,
+                          unsigned long flags, const char *name)
+{
+       int ret = request_threaded_irq(irq, chan_irq, chan_irqt,
+                                      flags, name, schan);
+
+       schan->irq = ret < 0 ? ret : irq;
+
+       return ret;
+}
+EXPORT_SYMBOL(shdma_request_irq);
+
+void shdma_free_irq(struct shdma_chan *schan)
+{
+       if (schan->irq >= 0)
+               free_irq(schan->irq, schan);
+}
+EXPORT_SYMBOL(shdma_free_irq);
+
+void shdma_chan_probe(struct shdma_dev *sdev,
+                          struct shdma_chan *schan, int id)
+{
+       schan->pm_state = SHDMA_PM_ESTABLISHED;
+
+       /* reference struct dma_device */
+       schan->dma_chan.device = &sdev->dma_dev;
+       dma_cookie_init(&schan->dma_chan);
+
+       schan->dev = sdev->dma_dev.dev;
+       schan->id = id;
+
+       if (!schan->max_xfer_len)
+               schan->max_xfer_len = PAGE_SIZE;
+
+       spin_lock_init(&schan->chan_lock);
+
+       /* Init descripter manage list */
+       INIT_LIST_HEAD(&schan->ld_queue);
+       INIT_LIST_HEAD(&schan->ld_free);
+
+       /* Add the channel to DMA device channel list */
+       list_add_tail(&schan->dma_chan.device_node,
+                       &sdev->dma_dev.channels);
+       sdev->schan[sdev->dma_dev.chancnt++] = schan;
+}
+EXPORT_SYMBOL(shdma_chan_probe);
+
+void shdma_chan_remove(struct shdma_chan *schan)
+{
+       list_del(&schan->dma_chan.device_node);
+}
+EXPORT_SYMBOL(shdma_chan_remove);
+
+int shdma_init(struct device *dev, struct shdma_dev *sdev,
+                   int chan_num)
+{
+       struct dma_device *dma_dev = &sdev->dma_dev;
+
+       /*
+        * Require all call-backs for now, they can trivially be made optional
+        * later as required
+        */
+       if (!sdev->ops ||
+           !sdev->desc_size ||
+           !sdev->ops->embedded_desc ||
+           !sdev->ops->start_xfer ||
+           !sdev->ops->setup_xfer ||
+           !sdev->ops->set_slave ||
+           !sdev->ops->desc_setup ||
+           !sdev->ops->slave_addr ||
+           !sdev->ops->channel_busy ||
+           !sdev->ops->halt_channel ||
+           !sdev->ops->desc_completed)
+               return -EINVAL;
+
+       sdev->schan = kcalloc(chan_num, sizeof(*sdev->schan), GFP_KERNEL);
+       if (!sdev->schan)
+               return -ENOMEM;
+
+       INIT_LIST_HEAD(&dma_dev->channels);
+
+       /* Common and MEMCPY operations */
+       dma_dev->device_alloc_chan_resources
+               = shdma_alloc_chan_resources;
+       dma_dev->device_free_chan_resources = shdma_free_chan_resources;
+       dma_dev->device_prep_dma_memcpy = shdma_prep_memcpy;
+       dma_dev->device_tx_status = shdma_tx_status;
+       dma_dev->device_issue_pending = shdma_issue_pending;
+
+       /* Compulsory for DMA_SLAVE fields */
+       dma_dev->device_prep_slave_sg = shdma_prep_slave_sg;
+       dma_dev->device_control = shdma_control;
+
+       dma_dev->dev = dev;
+
+       return 0;
+}
+EXPORT_SYMBOL(shdma_init);
+
+void shdma_cleanup(struct shdma_dev *sdev)
+{
+       kfree(sdev->schan);
+}
+EXPORT_SYMBOL(shdma_cleanup);
+
+static int __init shdma_enter(void)
+{
+       shdma_slave_used = kzalloc(DIV_ROUND_UP(slave_num, BITS_PER_LONG) *
+                                   sizeof(long), GFP_KERNEL);
+       if (!shdma_slave_used)
+               return -ENOMEM;
+       return 0;
+}
+module_init(shdma_enter);
+
+static void __exit shdma_exit(void)
+{
+       kfree(shdma_slave_used);
+}
+module_exit(shdma_exit);
+
+MODULE_LICENSE("GPL v2");
+MODULE_DESCRIPTION("SH-DMA driver base library");
+MODULE_AUTHOR("Guennadi Liakhovetski <g.liakhovetski@gmx.de>");
diff --git a/include/linux/shdma-base.h b/include/linux/shdma-base.h
new file mode 100644 (file)
index 0000000..83efd13
--- /dev/null
@@ -0,0 +1,123 @@
+/*
+ * Dmaengine driver base library for DMA controllers, found on SH-based SoCs
+ *
+ * extracted from shdma.c and headers
+ *
+ * Copyright (C) 2011-2012 Guennadi Liakhovetski <g.liakhovetski@gmx.de>
+ * Copyright (C) 2009 Nobuhiro Iwamatsu <iwamatsu.nobuhiro@renesas.com>
+ * Copyright (C) 2009 Renesas Solutions, Inc. All rights reserved.
+ * Copyright (C) 2007 Freescale Semiconductor, Inc. All rights reserved.
+ *
+ * This is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ */
+
+#ifndef SHDMA_BASE_H
+#define SHDMA_BASE_H
+
+#include <linux/dmaengine.h>
+#include <linux/interrupt.h>
+#include <linux/list.h>
+#include <linux/types.h>
+
+/**
+ * shdma_pm_state - DMA channel PM state
+ * SHDMA_PM_ESTABLISHED:       either idle or during data transfer
+ * SHDMA_PM_BUSY:              during the transfer preparation, when we have to
+ *                             drop the lock temporarily
+ * SHDMA_PM_PENDING:   transfers pending
+ */
+enum shdma_pm_state {
+       SHDMA_PM_ESTABLISHED,
+       SHDMA_PM_BUSY,
+       SHDMA_PM_PENDING,
+};
+
+struct device;
+
+/*
+ * Drivers, using this library are expected to embed struct shdma_dev,
+ * struct shdma_chan, struct shdma_desc, and struct shdma_slave
+ * in their respective device, channel, descriptor and slave objects.
+ */
+
+struct shdma_slave {
+       unsigned int slave_id;
+};
+
+struct shdma_desc {
+       struct list_head node;
+       struct dma_async_tx_descriptor async_tx;
+       enum dma_transfer_direction direction;
+       dma_cookie_t cookie;
+       int chunks;
+       int mark;
+};
+
+struct shdma_chan {
+       spinlock_t chan_lock;           /* Channel operation lock */
+       struct list_head ld_queue;      /* Link descriptors queue */
+       struct list_head ld_free;       /* Free link descriptors */
+       struct dma_chan dma_chan;       /* DMA channel */
+       struct device *dev;             /* Channel device */
+       void *desc;                     /* buffer for descriptor array */
+       int desc_num;                   /* desc count */
+       size_t max_xfer_len;            /* max transfer length */
+       int id;                         /* Raw id of this channel */
+       int irq;                        /* Channel IRQ */
+       enum shdma_pm_state pm_state;
+};
+
+/**
+ * struct shdma_ops - simple DMA driver operations
+ * desc_completed:     return true, if this is the descriptor, that just has
+ *                     completed (atomic)
+ * halt_channel:       stop DMA channel operation (atomic)
+ * channel_busy:       return true, if the channel is busy (atomic)
+ * slave_addr:         return slave DMA address
+ * desc_setup:         set up the hardware specific descriptor portion (atomic)
+ * set_slave:          bind channel to a slave
+ * setup_xfer:         configure channel hardware for operation (atomic)
+ * start_xfer:         start the DMA transfer (atomic)
+ * embedded_desc:      return Nth struct shdma_desc pointer from the
+ *                     descriptor array
+ * chan_irq:           process channel IRQ, return true if a transfer has
+ *                     completed (atomic)
+ */
+struct shdma_ops {
+       bool (*desc_completed)(struct shdma_chan *, struct shdma_desc *);
+       void (*halt_channel)(struct shdma_chan *);
+       bool (*channel_busy)(struct shdma_chan *);
+       dma_addr_t (*slave_addr)(struct shdma_chan *);
+       int (*desc_setup)(struct shdma_chan *, struct shdma_desc *,
+                         dma_addr_t, dma_addr_t, size_t *);
+       int (*set_slave)(struct shdma_chan *, struct shdma_slave *);
+       void (*setup_xfer)(struct shdma_chan *, struct shdma_slave *);
+       void (*start_xfer)(struct shdma_chan *, struct shdma_desc *);
+       struct shdma_desc *(*embedded_desc)(void *, int);
+       bool (*chan_irq)(struct shdma_chan *, int);
+};
+
+struct shdma_dev {
+       struct dma_device dma_dev;
+       struct shdma_chan **schan;
+       const struct shdma_ops *ops;
+       size_t desc_size;
+};
+
+#define shdma_for_each_chan(c, d, i) for (i = 0, c = (d)->schan[0]; \
+                               i < (d)->dma_dev.chancnt; c = (d)->schan[++i])
+
+int shdma_request_irq(struct shdma_chan *, int,
+                          unsigned long, const char *);
+void shdma_free_irq(struct shdma_chan *);
+bool shdma_reset(struct shdma_dev *sdev);
+void shdma_chan_probe(struct shdma_dev *sdev,
+                          struct shdma_chan *schan, int id);
+void shdma_chan_remove(struct shdma_chan *schan);
+int shdma_init(struct device *dev, struct shdma_dev *sdev,
+                   int chan_num);
+void shdma_cleanup(struct shdma_dev *sdev);
+
+#endif