iommu: Add virtio-iommu driver
authorJean-Philippe Brucker <jean-philippe.brucker@arm.com>
Tue, 15 Jan 2019 12:19:57 +0000 (12:19 +0000)
committerMichael S. Tsirkin <mst@redhat.com>
Thu, 6 Jun 2019 21:32:13 +0000 (17:32 -0400)
The virtio IOMMU is a para-virtualized device, allowing to send IOMMU
requests such as map/unmap over virtio transport without emulating page
tables. This implementation handles ATTACH, DETACH, MAP and UNMAP
requests.

The bulk of the code transforms calls coming from the IOMMU API into
corresponding virtio requests. Mappings are kept in an interval tree
instead of page tables. A little more work is required for modular and x86
support, so for the moment the driver depends on CONFIG_VIRTIO=y and
CONFIG_ARM64.

Tested-by: Bharat Bhushan <bharat.bhushan@nxp.com>
Tested-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Jean-Philippe Brucker <jean-philippe.brucker@arm.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
MAINTAINERS
drivers/iommu/Kconfig
drivers/iommu/Makefile
drivers/iommu/virtio-iommu.c [new file with mode: 0644]
include/uapi/linux/virtio_ids.h
include/uapi/linux/virtio_iommu.h [new file with mode: 0644]

index 429c6c624861beb1c3d69f38131ca2a81c5dbd48..62bd1834d95ab1193ab915f41c05970a64b2b4e5 100644 (file)
@@ -16807,6 +16807,13 @@ S:     Maintained
 F:     drivers/virtio/virtio_input.c
 F:     include/uapi/linux/virtio_input.h
 
+VIRTIO IOMMU DRIVER
+M:     Jean-Philippe Brucker <jean-philippe.brucker@arm.com>
+L:     virtualization@lists.linux-foundation.org
+S:     Maintained
+F:     drivers/iommu/virtio-iommu.c
+F:     include/uapi/linux/virtio_iommu.h
+
 VIRTUAL BOX GUEST DEVICE DRIVER
 M:     Hans de Goede <hdegoede@redhat.com>
 M:     Arnd Bergmann <arnd@arndb.de>
index 83664db5221df02248676215b551087d28fc4727..e15cdcd8cb3c8c64597707aa54398444f6a1ac4c 100644 (file)
@@ -473,4 +473,15 @@ config HYPERV_IOMMU
          Stub IOMMU driver to handle IRQs as to allow Hyper-V Linux
          guests to run with x2APIC mode enabled.
 
+config VIRTIO_IOMMU
+       bool "Virtio IOMMU driver"
+       depends on VIRTIO=y
+       depends on ARM64
+       select IOMMU_API
+       select INTERVAL_TREE
+       help
+         Para-virtualised IOMMU driver with virtio.
+
+         Say Y here if you intend to run this kernel as a guest.
+
 endif # IOMMU_SUPPORT
index 8c71a15e986b5dc4a068956c8454eb26f7ee05a6..f13f36ae1af652836254ba07315f99303fba6f88 100644 (file)
@@ -33,3 +33,4 @@ obj-$(CONFIG_FSL_PAMU) += fsl_pamu.o fsl_pamu_domain.o
 obj-$(CONFIG_S390_IOMMU) += s390-iommu.o
 obj-$(CONFIG_QCOM_IOMMU) += qcom_iommu.o
 obj-$(CONFIG_HYPERV_IOMMU) += hyperv-iommu.o
+obj-$(CONFIG_VIRTIO_IOMMU) += virtio-iommu.o
diff --git a/drivers/iommu/virtio-iommu.c b/drivers/iommu/virtio-iommu.c
new file mode 100644 (file)
index 0000000..6fa012c
--- /dev/null
@@ -0,0 +1,916 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Virtio driver for the paravirtualized IOMMU
+ *
+ * Copyright (C) 2018 Arm Limited
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/amba/bus.h>
+#include <linux/delay.h>
+#include <linux/dma-iommu.h>
+#include <linux/freezer.h>
+#include <linux/interval_tree.h>
+#include <linux/iommu.h>
+#include <linux/module.h>
+#include <linux/of_iommu.h>
+#include <linux/of_platform.h>
+#include <linux/pci.h>
+#include <linux/platform_device.h>
+#include <linux/virtio.h>
+#include <linux/virtio_config.h>
+#include <linux/virtio_ids.h>
+#include <linux/wait.h>
+
+#include <uapi/linux/virtio_iommu.h>
+
+#define MSI_IOVA_BASE                  0x8000000
+#define MSI_IOVA_LENGTH                        0x100000
+
+#define VIOMMU_REQUEST_VQ              0
+#define VIOMMU_NR_VQS                  1
+
+struct viommu_dev {
+       struct iommu_device             iommu;
+       struct device                   *dev;
+       struct virtio_device            *vdev;
+
+       struct ida                      domain_ids;
+
+       struct virtqueue                *vqs[VIOMMU_NR_VQS];
+       spinlock_t                      request_lock;
+       struct list_head                requests;
+
+       /* Device configuration */
+       struct iommu_domain_geometry    geometry;
+       u64                             pgsize_bitmap;
+       u8                              domain_bits;
+};
+
+struct viommu_mapping {
+       phys_addr_t                     paddr;
+       struct interval_tree_node       iova;
+       u32                             flags;
+};
+
+struct viommu_domain {
+       struct iommu_domain             domain;
+       struct viommu_dev               *viommu;
+       struct mutex                    mutex; /* protects viommu pointer */
+       unsigned int                    id;
+
+       spinlock_t                      mappings_lock;
+       struct rb_root_cached           mappings;
+
+       unsigned long                   nr_endpoints;
+};
+
+struct viommu_endpoint {
+       struct viommu_dev               *viommu;
+       struct viommu_domain            *vdomain;
+};
+
+struct viommu_request {
+       struct list_head                list;
+       void                            *writeback;
+       unsigned int                    write_offset;
+       unsigned int                    len;
+       char                            buf[];
+};
+
+#define to_viommu_domain(domain)       \
+       container_of(domain, struct viommu_domain, domain)
+
+static int viommu_get_req_errno(void *buf, size_t len)
+{
+       struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
+
+       switch (tail->status) {
+       case VIRTIO_IOMMU_S_OK:
+               return 0;
+       case VIRTIO_IOMMU_S_UNSUPP:
+               return -ENOSYS;
+       case VIRTIO_IOMMU_S_INVAL:
+               return -EINVAL;
+       case VIRTIO_IOMMU_S_RANGE:
+               return -ERANGE;
+       case VIRTIO_IOMMU_S_NOENT:
+               return -ENOENT;
+       case VIRTIO_IOMMU_S_FAULT:
+               return -EFAULT;
+       case VIRTIO_IOMMU_S_IOERR:
+       case VIRTIO_IOMMU_S_DEVERR:
+       default:
+               return -EIO;
+       }
+}
+
+static void viommu_set_req_status(void *buf, size_t len, int status)
+{
+       struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
+
+       tail->status = status;
+}
+
+static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu,
+                                         struct virtio_iommu_req_head *req,
+                                         size_t len)
+{
+       size_t tail_size = sizeof(struct virtio_iommu_req_tail);
+
+       return len - tail_size;
+}
+
+/*
+ * __viommu_sync_req - Complete all in-flight requests
+ *
+ * Wait for all added requests to complete. When this function returns, all
+ * requests that were in-flight at the time of the call have completed.
+ */
+static int __viommu_sync_req(struct viommu_dev *viommu)
+{
+       int ret = 0;
+       unsigned int len;
+       size_t write_len;
+       struct viommu_request *req;
+       struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
+
+       assert_spin_locked(&viommu->request_lock);
+
+       virtqueue_kick(vq);
+
+       while (!list_empty(&viommu->requests)) {
+               len = 0;
+               req = virtqueue_get_buf(vq, &len);
+               if (!req)
+                       continue;
+
+               if (!len)
+                       viommu_set_req_status(req->buf, req->len,
+                                             VIRTIO_IOMMU_S_IOERR);
+
+               write_len = req->len - req->write_offset;
+               if (req->writeback && len == write_len)
+                       memcpy(req->writeback, req->buf + req->write_offset,
+                              write_len);
+
+               list_del(&req->list);
+               kfree(req);
+       }
+
+       return ret;
+}
+
+static int viommu_sync_req(struct viommu_dev *viommu)
+{
+       int ret;
+       unsigned long flags;
+
+       spin_lock_irqsave(&viommu->request_lock, flags);
+       ret = __viommu_sync_req(viommu);
+       if (ret)
+               dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
+       spin_unlock_irqrestore(&viommu->request_lock, flags);
+
+       return ret;
+}
+
+/*
+ * __viommu_add_request - Add one request to the queue
+ * @buf: pointer to the request buffer
+ * @len: length of the request buffer
+ * @writeback: copy data back to the buffer when the request completes.
+ *
+ * Add a request to the queue. Only synchronize the queue if it's already full.
+ * Otherwise don't kick the queue nor wait for requests to complete.
+ *
+ * When @writeback is true, data written by the device, including the request
+ * status, is copied into @buf after the request completes. This is unsafe if
+ * the caller allocates @buf on stack and drops the lock between add_req() and
+ * sync_req().
+ *
+ * Return 0 if the request was successfully added to the queue.
+ */
+static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
+                           bool writeback)
+{
+       int ret;
+       off_t write_offset;
+       struct viommu_request *req;
+       struct scatterlist top_sg, bottom_sg;
+       struct scatterlist *sg[2] = { &top_sg, &bottom_sg };
+       struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
+
+       assert_spin_locked(&viommu->request_lock);
+
+       write_offset = viommu_get_write_desc_offset(viommu, buf, len);
+       if (write_offset <= 0)
+               return -EINVAL;
+
+       req = kzalloc(sizeof(*req) + len, GFP_ATOMIC);
+       if (!req)
+               return -ENOMEM;
+
+       req->len = len;
+       if (writeback) {
+               req->writeback = buf + write_offset;
+               req->write_offset = write_offset;
+       }
+       memcpy(&req->buf, buf, write_offset);
+
+       sg_init_one(&top_sg, req->buf, write_offset);
+       sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
+
+       ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
+       if (ret == -ENOSPC) {
+               /* If the queue is full, sync and retry */
+               if (!__viommu_sync_req(viommu))
+                       ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
+       }
+       if (ret)
+               goto err_free;
+
+       list_add_tail(&req->list, &viommu->requests);
+       return 0;
+
+err_free:
+       kfree(req);
+       return ret;
+}
+
+static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
+{
+       int ret;
+       unsigned long flags;
+
+       spin_lock_irqsave(&viommu->request_lock, flags);
+       ret = __viommu_add_req(viommu, buf, len, false);
+       if (ret)
+               dev_dbg(viommu->dev, "could not add request: %d\n", ret);
+       spin_unlock_irqrestore(&viommu->request_lock, flags);
+
+       return ret;
+}
+
+/*
+ * Send a request and wait for it to complete. Return the request status (as an
+ * errno)
+ */
+static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
+                               size_t len)
+{
+       int ret;
+       unsigned long flags;
+
+       spin_lock_irqsave(&viommu->request_lock, flags);
+
+       ret = __viommu_add_req(viommu, buf, len, true);
+       if (ret) {
+               dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
+               goto out_unlock;
+       }
+
+       ret = __viommu_sync_req(viommu);
+       if (ret) {
+               dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
+               /* Fall-through (get the actual request status) */
+       }
+
+       ret = viommu_get_req_errno(buf, len);
+out_unlock:
+       spin_unlock_irqrestore(&viommu->request_lock, flags);
+       return ret;
+}
+
+/*
+ * viommu_add_mapping - add a mapping to the internal tree
+ *
+ * On success, return the new mapping. Otherwise return NULL.
+ */
+static int viommu_add_mapping(struct viommu_domain *vdomain, unsigned long iova,
+                             phys_addr_t paddr, size_t size, u32 flags)
+{
+       unsigned long irqflags;
+       struct viommu_mapping *mapping;
+
+       mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
+       if (!mapping)
+               return -ENOMEM;
+
+       mapping->paddr          = paddr;
+       mapping->iova.start     = iova;
+       mapping->iova.last      = iova + size - 1;
+       mapping->flags          = flags;
+
+       spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
+       interval_tree_insert(&mapping->iova, &vdomain->mappings);
+       spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
+
+       return 0;
+}
+
+/*
+ * viommu_del_mappings - remove mappings from the internal tree
+ *
+ * @vdomain: the domain
+ * @iova: start of the range
+ * @size: size of the range. A size of 0 corresponds to the entire address
+ *     space.
+ *
+ * On success, returns the number of unmapped bytes (>= size)
+ */
+static size_t viommu_del_mappings(struct viommu_domain *vdomain,
+                                 unsigned long iova, size_t size)
+{
+       size_t unmapped = 0;
+       unsigned long flags;
+       unsigned long last = iova + size - 1;
+       struct viommu_mapping *mapping = NULL;
+       struct interval_tree_node *node, *next;
+
+       spin_lock_irqsave(&vdomain->mappings_lock, flags);
+       next = interval_tree_iter_first(&vdomain->mappings, iova, last);
+       while (next) {
+               node = next;
+               mapping = container_of(node, struct viommu_mapping, iova);
+               next = interval_tree_iter_next(node, iova, last);
+
+               /* Trying to split a mapping? */
+               if (mapping->iova.start < iova)
+                       break;
+
+               /*
+                * Virtio-iommu doesn't allow UNMAP to split a mapping created
+                * with a single MAP request, so remove the full mapping.
+                */
+               unmapped += mapping->iova.last - mapping->iova.start + 1;
+
+               interval_tree_remove(node, &vdomain->mappings);
+               kfree(mapping);
+       }
+       spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
+
+       return unmapped;
+}
+
+/*
+ * viommu_replay_mappings - re-send MAP requests
+ *
+ * When reattaching a domain that was previously detached from all endpoints,
+ * mappings were deleted from the device. Re-create the mappings available in
+ * the internal tree.
+ */
+static int viommu_replay_mappings(struct viommu_domain *vdomain)
+{
+       int ret = 0;
+       unsigned long flags;
+       struct viommu_mapping *mapping;
+       struct interval_tree_node *node;
+       struct virtio_iommu_req_map map;
+
+       spin_lock_irqsave(&vdomain->mappings_lock, flags);
+       node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
+       while (node) {
+               mapping = container_of(node, struct viommu_mapping, iova);
+               map = (struct virtio_iommu_req_map) {
+                       .head.type      = VIRTIO_IOMMU_T_MAP,
+                       .domain         = cpu_to_le32(vdomain->id),
+                       .virt_start     = cpu_to_le64(mapping->iova.start),
+                       .virt_end       = cpu_to_le64(mapping->iova.last),
+                       .phys_start     = cpu_to_le64(mapping->paddr),
+                       .flags          = cpu_to_le32(mapping->flags),
+               };
+
+               ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
+               if (ret)
+                       break;
+
+               node = interval_tree_iter_next(node, 0, -1UL);
+       }
+       spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
+
+       return ret;
+}
+
+/* IOMMU API */
+
+static struct iommu_domain *viommu_domain_alloc(unsigned type)
+{
+       struct viommu_domain *vdomain;
+
+       if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
+               return NULL;
+
+       vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
+       if (!vdomain)
+               return NULL;
+
+       mutex_init(&vdomain->mutex);
+       spin_lock_init(&vdomain->mappings_lock);
+       vdomain->mappings = RB_ROOT_CACHED;
+
+       if (type == IOMMU_DOMAIN_DMA &&
+           iommu_get_dma_cookie(&vdomain->domain)) {
+               kfree(vdomain);
+               return NULL;
+       }
+
+       return &vdomain->domain;
+}
+
+static int viommu_domain_finalise(struct viommu_dev *viommu,
+                                 struct iommu_domain *domain)
+{
+       int ret;
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+       unsigned int max_domain = viommu->domain_bits > 31 ? ~0 :
+                                 (1U << viommu->domain_bits) - 1;
+
+       vdomain->viommu         = viommu;
+
+       domain->pgsize_bitmap   = viommu->pgsize_bitmap;
+       domain->geometry        = viommu->geometry;
+
+       ret = ida_alloc_max(&viommu->domain_ids, max_domain, GFP_KERNEL);
+       if (ret >= 0)
+               vdomain->id = (unsigned int)ret;
+
+       return ret > 0 ? 0 : ret;
+}
+
+static void viommu_domain_free(struct iommu_domain *domain)
+{
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+
+       iommu_put_dma_cookie(domain);
+
+       /* Free all remaining mappings (size 2^64) */
+       viommu_del_mappings(vdomain, 0, 0);
+
+       if (vdomain->viommu)
+               ida_free(&vdomain->viommu->domain_ids, vdomain->id);
+
+       kfree(vdomain);
+}
+
+static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
+{
+       int i;
+       int ret = 0;
+       struct virtio_iommu_req_attach req;
+       struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
+       struct viommu_endpoint *vdev = fwspec->iommu_priv;
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+
+       mutex_lock(&vdomain->mutex);
+       if (!vdomain->viommu) {
+               /*
+                * Properly initialize the domain now that we know which viommu
+                * owns it.
+                */
+               ret = viommu_domain_finalise(vdev->viommu, domain);
+       } else if (vdomain->viommu != vdev->viommu) {
+               dev_err(dev, "cannot attach to foreign vIOMMU\n");
+               ret = -EXDEV;
+       }
+       mutex_unlock(&vdomain->mutex);
+
+       if (ret)
+               return ret;
+
+       /*
+        * In the virtio-iommu device, when attaching the endpoint to a new
+        * domain, it is detached from the old one and, if as as a result the
+        * old domain isn't attached to any endpoint, all mappings are removed
+        * from the old domain and it is freed.
+        *
+        * In the driver the old domain still exists, and its mappings will be
+        * recreated if it gets reattached to an endpoint. Otherwise it will be
+        * freed explicitly.
+        *
+        * vdev->vdomain is protected by group->mutex
+        */
+       if (vdev->vdomain)
+               vdev->vdomain->nr_endpoints--;
+
+       req = (struct virtio_iommu_req_attach) {
+               .head.type      = VIRTIO_IOMMU_T_ATTACH,
+               .domain         = cpu_to_le32(vdomain->id),
+       };
+
+       for (i = 0; i < fwspec->num_ids; i++) {
+               req.endpoint = cpu_to_le32(fwspec->ids[i]);
+
+               ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
+               if (ret)
+                       return ret;
+       }
+
+       if (!vdomain->nr_endpoints) {
+               /*
+                * This endpoint is the first to be attached to the domain.
+                * Replay existing mappings (e.g. SW MSI).
+                */
+               ret = viommu_replay_mappings(vdomain);
+               if (ret)
+                       return ret;
+       }
+
+       vdomain->nr_endpoints++;
+       vdev->vdomain = vdomain;
+
+       return 0;
+}
+
+static int viommu_map(struct iommu_domain *domain, unsigned long iova,
+                     phys_addr_t paddr, size_t size, int prot)
+{
+       int ret;
+       int flags;
+       struct virtio_iommu_req_map map;
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+
+       flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
+               (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
+               (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
+
+       ret = viommu_add_mapping(vdomain, iova, paddr, size, flags);
+       if (ret)
+               return ret;
+
+       map = (struct virtio_iommu_req_map) {
+               .head.type      = VIRTIO_IOMMU_T_MAP,
+               .domain         = cpu_to_le32(vdomain->id),
+               .virt_start     = cpu_to_le64(iova),
+               .phys_start     = cpu_to_le64(paddr),
+               .virt_end       = cpu_to_le64(iova + size - 1),
+               .flags          = cpu_to_le32(flags),
+       };
+
+       if (!vdomain->nr_endpoints)
+               return 0;
+
+       ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
+       if (ret)
+               viommu_del_mappings(vdomain, iova, size);
+
+       return ret;
+}
+
+static size_t viommu_unmap(struct iommu_domain *domain, unsigned long iova,
+                          size_t size)
+{
+       int ret = 0;
+       size_t unmapped;
+       struct virtio_iommu_req_unmap unmap;
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+
+       unmapped = viommu_del_mappings(vdomain, iova, size);
+       if (unmapped < size)
+               return 0;
+
+       /* Device already removed all mappings after detach. */
+       if (!vdomain->nr_endpoints)
+               return unmapped;
+
+       unmap = (struct virtio_iommu_req_unmap) {
+               .head.type      = VIRTIO_IOMMU_T_UNMAP,
+               .domain         = cpu_to_le32(vdomain->id),
+               .virt_start     = cpu_to_le64(iova),
+               .virt_end       = cpu_to_le64(iova + unmapped - 1),
+       };
+
+       ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
+       return ret ? 0 : unmapped;
+}
+
+static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
+                                      dma_addr_t iova)
+{
+       u64 paddr = 0;
+       unsigned long flags;
+       struct viommu_mapping *mapping;
+       struct interval_tree_node *node;
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+
+       spin_lock_irqsave(&vdomain->mappings_lock, flags);
+       node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
+       if (node) {
+               mapping = container_of(node, struct viommu_mapping, iova);
+               paddr = mapping->paddr + (iova - mapping->iova.start);
+       }
+       spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
+
+       return paddr;
+}
+
+static void viommu_iotlb_sync(struct iommu_domain *domain)
+{
+       struct viommu_domain *vdomain = to_viommu_domain(domain);
+
+       viommu_sync_req(vdomain->viommu);
+}
+
+static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
+{
+       struct iommu_resv_region *region;
+       int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
+
+       region = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH, prot,
+                                        IOMMU_RESV_SW_MSI);
+       if (!region)
+               return;
+
+       list_add_tail(&region->list, head);
+       iommu_dma_get_resv_regions(dev, head);
+}
+
+static void viommu_put_resv_regions(struct device *dev, struct list_head *head)
+{
+       struct iommu_resv_region *entry, *next;
+
+       list_for_each_entry_safe(entry, next, head, list)
+               kfree(entry);
+}
+
+static struct iommu_ops viommu_ops;
+static struct virtio_driver virtio_iommu_drv;
+
+static int viommu_match_node(struct device *dev, void *data)
+{
+       return dev->parent->fwnode == data;
+}
+
+static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
+{
+       struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
+                                               fwnode, viommu_match_node);
+       put_device(dev);
+
+       return dev ? dev_to_virtio(dev)->priv : NULL;
+}
+
+static int viommu_add_device(struct device *dev)
+{
+       int ret;
+       struct iommu_group *group;
+       struct viommu_endpoint *vdev;
+       struct viommu_dev *viommu = NULL;
+       struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
+
+       if (!fwspec || fwspec->ops != &viommu_ops)
+               return -ENODEV;
+
+       viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
+       if (!viommu)
+               return -ENODEV;
+
+       vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
+       if (!vdev)
+               return -ENOMEM;
+
+       vdev->viommu = viommu;
+       fwspec->iommu_priv = vdev;
+
+       ret = iommu_device_link(&viommu->iommu, dev);
+       if (ret)
+               goto err_free_dev;
+
+       /*
+        * Last step creates a default domain and attaches to it. Everything
+        * must be ready.
+        */
+       group = iommu_group_get_for_dev(dev);
+       if (IS_ERR(group)) {
+               ret = PTR_ERR(group);
+               goto err_unlink_dev;
+       }
+
+       iommu_group_put(group);
+
+       return PTR_ERR_OR_ZERO(group);
+
+err_unlink_dev:
+       iommu_device_unlink(&viommu->iommu, dev);
+err_free_dev:
+       kfree(vdev);
+
+       return ret;
+}
+
+static void viommu_remove_device(struct device *dev)
+{
+       struct viommu_endpoint *vdev;
+       struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
+
+       if (!fwspec || fwspec->ops != &viommu_ops)
+               return;
+
+       vdev = fwspec->iommu_priv;
+
+       iommu_group_remove_device(dev);
+       iommu_device_unlink(&vdev->viommu->iommu, dev);
+       kfree(vdev);
+}
+
+static struct iommu_group *viommu_device_group(struct device *dev)
+{
+       if (dev_is_pci(dev))
+               return pci_device_group(dev);
+       else
+               return generic_device_group(dev);
+}
+
+static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
+{
+       return iommu_fwspec_add_ids(dev, args->args, 1);
+}
+
+static struct iommu_ops viommu_ops = {
+       .domain_alloc           = viommu_domain_alloc,
+       .domain_free            = viommu_domain_free,
+       .attach_dev             = viommu_attach_dev,
+       .map                    = viommu_map,
+       .unmap                  = viommu_unmap,
+       .iova_to_phys           = viommu_iova_to_phys,
+       .iotlb_sync             = viommu_iotlb_sync,
+       .add_device             = viommu_add_device,
+       .remove_device          = viommu_remove_device,
+       .device_group           = viommu_device_group,
+       .get_resv_regions       = viommu_get_resv_regions,
+       .put_resv_regions       = viommu_put_resv_regions,
+       .of_xlate               = viommu_of_xlate,
+};
+
+static int viommu_init_vqs(struct viommu_dev *viommu)
+{
+       struct virtio_device *vdev = dev_to_virtio(viommu->dev);
+       const char *name = "request";
+       void *ret;
+
+       ret = virtio_find_single_vq(vdev, NULL, name);
+       if (IS_ERR(ret)) {
+               dev_err(viommu->dev, "cannot find VQ\n");
+               return PTR_ERR(ret);
+       }
+
+       viommu->vqs[VIOMMU_REQUEST_VQ] = ret;
+
+       return 0;
+}
+
+static int viommu_probe(struct virtio_device *vdev)
+{
+       struct device *parent_dev = vdev->dev.parent;
+       struct viommu_dev *viommu = NULL;
+       struct device *dev = &vdev->dev;
+       u64 input_start = 0;
+       u64 input_end = -1UL;
+       int ret;
+
+       if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
+           !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
+               return -ENODEV;
+
+       viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
+       if (!viommu)
+               return -ENOMEM;
+
+       spin_lock_init(&viommu->request_lock);
+       ida_init(&viommu->domain_ids);
+       viommu->dev = dev;
+       viommu->vdev = vdev;
+       INIT_LIST_HEAD(&viommu->requests);
+
+       ret = viommu_init_vqs(viommu);
+       if (ret)
+               return ret;
+
+       virtio_cread(vdev, struct virtio_iommu_config, page_size_mask,
+                    &viommu->pgsize_bitmap);
+
+       if (!viommu->pgsize_bitmap) {
+               ret = -EINVAL;
+               goto err_free_vqs;
+       }
+
+       viommu->domain_bits = 32;
+
+       /* Optional features */
+       virtio_cread_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
+                            struct virtio_iommu_config, input_range.start,
+                            &input_start);
+
+       virtio_cread_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
+                            struct virtio_iommu_config, input_range.end,
+                            &input_end);
+
+       virtio_cread_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_BITS,
+                            struct virtio_iommu_config, domain_bits,
+                            &viommu->domain_bits);
+
+       viommu->geometry = (struct iommu_domain_geometry) {
+               .aperture_start = input_start,
+               .aperture_end   = input_end,
+               .force_aperture = true,
+       };
+
+       viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
+
+       virtio_device_ready(vdev);
+
+       ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
+                                    virtio_bus_name(vdev));
+       if (ret)
+               goto err_free_vqs;
+
+       iommu_device_set_ops(&viommu->iommu, &viommu_ops);
+       iommu_device_set_fwnode(&viommu->iommu, parent_dev->fwnode);
+
+       iommu_device_register(&viommu->iommu);
+
+#ifdef CONFIG_PCI
+       if (pci_bus_type.iommu_ops != &viommu_ops) {
+               pci_request_acs();
+               ret = bus_set_iommu(&pci_bus_type, &viommu_ops);
+               if (ret)
+                       goto err_unregister;
+       }
+#endif
+#ifdef CONFIG_ARM_AMBA
+       if (amba_bustype.iommu_ops != &viommu_ops) {
+               ret = bus_set_iommu(&amba_bustype, &viommu_ops);
+               if (ret)
+                       goto err_unregister;
+       }
+#endif
+       if (platform_bus_type.iommu_ops != &viommu_ops) {
+               ret = bus_set_iommu(&platform_bus_type, &viommu_ops);
+               if (ret)
+                       goto err_unregister;
+       }
+
+       vdev->priv = viommu;
+
+       dev_info(dev, "input address: %u bits\n",
+                order_base_2(viommu->geometry.aperture_end));
+       dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
+
+       return 0;
+
+err_unregister:
+       iommu_device_sysfs_remove(&viommu->iommu);
+       iommu_device_unregister(&viommu->iommu);
+err_free_vqs:
+       vdev->config->del_vqs(vdev);
+
+       return ret;
+}
+
+static void viommu_remove(struct virtio_device *vdev)
+{
+       struct viommu_dev *viommu = vdev->priv;
+
+       iommu_device_sysfs_remove(&viommu->iommu);
+       iommu_device_unregister(&viommu->iommu);
+
+       /* Stop all virtqueues */
+       vdev->config->reset(vdev);
+       vdev->config->del_vqs(vdev);
+
+       dev_info(&vdev->dev, "device removed\n");
+}
+
+static void viommu_config_changed(struct virtio_device *vdev)
+{
+       dev_warn(&vdev->dev, "config changed\n");
+}
+
+static unsigned int features[] = {
+       VIRTIO_IOMMU_F_MAP_UNMAP,
+       VIRTIO_IOMMU_F_DOMAIN_BITS,
+       VIRTIO_IOMMU_F_INPUT_RANGE,
+};
+
+static struct virtio_device_id id_table[] = {
+       { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
+       { 0 },
+};
+
+static struct virtio_driver virtio_iommu_drv = {
+       .driver.name            = KBUILD_MODNAME,
+       .driver.owner           = THIS_MODULE,
+       .id_table               = id_table,
+       .feature_table          = features,
+       .feature_table_size     = ARRAY_SIZE(features),
+       .probe                  = viommu_probe,
+       .remove                 = viommu_remove,
+       .config_changed         = viommu_config_changed,
+};
+
+module_virtio_driver(virtio_iommu_drv);
+
+MODULE_DESCRIPTION("Virtio IOMMU driver");
+MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
+MODULE_LICENSE("GPL v2");
index 6d5c3b2d4f4d55368564e6e2ce003a1cb0b6d33b..cfe47c5d9a56f38f11af45bb699328046bbc2fd7 100644 (file)
@@ -43,5 +43,6 @@
 #define VIRTIO_ID_INPUT        18 /* virtio input */
 #define VIRTIO_ID_VSOCK        19 /* virtio vsock transport */
 #define VIRTIO_ID_CRYPTO       20 /* virtio crypto */
+#define VIRTIO_ID_IOMMU        23 /* virtio IOMMU */
 
 #endif /* _LINUX_VIRTIO_IDS_H */
diff --git a/include/uapi/linux/virtio_iommu.h b/include/uapi/linux/virtio_iommu.h
new file mode 100644 (file)
index 0000000..5e5fd62
--- /dev/null
@@ -0,0 +1,106 @@
+/* SPDX-License-Identifier: BSD-3-Clause */
+/*
+ * Virtio-iommu definition v0.9
+ *
+ * Copyright (C) 2018 Arm Ltd.
+ */
+#ifndef _UAPI_LINUX_VIRTIO_IOMMU_H
+#define _UAPI_LINUX_VIRTIO_IOMMU_H
+
+#include <linux/types.h>
+
+/* Feature bits */
+#define VIRTIO_IOMMU_F_INPUT_RANGE             0
+#define VIRTIO_IOMMU_F_DOMAIN_BITS             1
+#define VIRTIO_IOMMU_F_MAP_UNMAP               2
+#define VIRTIO_IOMMU_F_BYPASS                  3
+
+struct virtio_iommu_range {
+       __u64                                   start;
+       __u64                                   end;
+};
+
+struct virtio_iommu_config {
+       /* Supported page sizes */
+       __u64                                   page_size_mask;
+       /* Supported IOVA range */
+       struct virtio_iommu_range               input_range;
+       /* Max domain ID size */
+       __u8                                    domain_bits;
+       __u8                                    padding[3];
+       /* Probe buffer size */
+       __u32                                   probe_size;
+};
+
+/* Request types */
+#define VIRTIO_IOMMU_T_ATTACH                  0x01
+#define VIRTIO_IOMMU_T_DETACH                  0x02
+#define VIRTIO_IOMMU_T_MAP                     0x03
+#define VIRTIO_IOMMU_T_UNMAP                   0x04
+
+/* Status types */
+#define VIRTIO_IOMMU_S_OK                      0x00
+#define VIRTIO_IOMMU_S_IOERR                   0x01
+#define VIRTIO_IOMMU_S_UNSUPP                  0x02
+#define VIRTIO_IOMMU_S_DEVERR                  0x03
+#define VIRTIO_IOMMU_S_INVAL                   0x04
+#define VIRTIO_IOMMU_S_RANGE                   0x05
+#define VIRTIO_IOMMU_S_NOENT                   0x06
+#define VIRTIO_IOMMU_S_FAULT                   0x07
+
+struct virtio_iommu_req_head {
+       __u8                                    type;
+       __u8                                    reserved[3];
+};
+
+struct virtio_iommu_req_tail {
+       __u8                                    status;
+       __u8                                    reserved[3];
+};
+
+struct virtio_iommu_req_attach {
+       struct virtio_iommu_req_head            head;
+       __le32                                  domain;
+       __le32                                  endpoint;
+       __u8                                    reserved[8];
+       struct virtio_iommu_req_tail            tail;
+};
+
+struct virtio_iommu_req_detach {
+       struct virtio_iommu_req_head            head;
+       __le32                                  domain;
+       __le32                                  endpoint;
+       __u8                                    reserved[8];
+       struct virtio_iommu_req_tail            tail;
+};
+
+#define VIRTIO_IOMMU_MAP_F_READ                        (1 << 0)
+#define VIRTIO_IOMMU_MAP_F_WRITE               (1 << 1)
+#define VIRTIO_IOMMU_MAP_F_EXEC                        (1 << 2)
+#define VIRTIO_IOMMU_MAP_F_MMIO                        (1 << 3)
+
+#define VIRTIO_IOMMU_MAP_F_MASK                        (VIRTIO_IOMMU_MAP_F_READ |      \
+                                                VIRTIO_IOMMU_MAP_F_WRITE |     \
+                                                VIRTIO_IOMMU_MAP_F_EXEC |      \
+                                                VIRTIO_IOMMU_MAP_F_MMIO)
+
+struct virtio_iommu_req_map {
+       struct virtio_iommu_req_head            head;
+       __le32                                  domain;
+       __le64                                  virt_start;
+       __le64                                  virt_end;
+       __le64                                  phys_start;
+       __le32                                  flags;
+       struct virtio_iommu_req_tail            tail;
+};
+
+struct virtio_iommu_req_unmap {
+       struct virtio_iommu_req_head            head;
+       __le32                                  domain;
+       __le64                                  virt_start;
+       __le64                                  virt_end;
+       __u8                                    reserved[4];
+       struct virtio_iommu_req_tail            tail;
+};
+
+#endif