drm/vgem: Pin our pages for dmabuf exports
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 22 Jun 2017 13:46:17 +0000 (14:46 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 23 Jun 2017 11:04:22 +0000 (13:04 +0200)
When the caller maps their dmabuf and we return an sg_table, the caller
doesn't expect the pages beneath that sg_table to vanish on a whim (i.e.
under mempressure). The contract is that the pages are pinned for the
duration of the mapping (from dma_buf_map_attachment() to
dma_buf_unmap_attachment). To comply, we need to introduce our own
vgem_object.pages_pin_count and elevate it across the mapping. However,
the drm_prime interface we use calls drv->prime_pin on dma_buf_attach
and drv->prime_unpin on dma_buf_detach, which while that does cover the
mapping is much broader than is desired -- but it will do for now.

v2: also hold the pin across prime_vmap/vunmap

Reported-by: Tomi Sarvela <tomi.p.sarvela@intel.com>
Testcase: igt/gem_concurrent_blit/*swap*vgem*
Fixes: 5ba6c9ff961a ("drm/vgem: Fix mmaping")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tomi Sarvela <tomi.p.sarvela@intel.com>
Cc: Laura Abbott <labbott@redhat.com>
Cc: Sean Paul <seanpaul@chromium.org>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: <stable@vger.kernel.org> # needs a backport
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20170622134617.17912-1-chris@chris-wilson.co.uk
drivers/gpu/drm/vgem/vgem_drv.c
drivers/gpu/drm/vgem/vgem_drv.h

index 18f401b442c2d0c334e1f879a8b9d9b0892d5306..c938af8c40cf54c41b8f87d6625ddd6a65d96d5c 100644 (file)
@@ -52,6 +52,7 @@ static void vgem_gem_free_object(struct drm_gem_object *obj)
        struct drm_vgem_gem_object *vgem_obj = to_vgem_bo(obj);
 
        kvfree(vgem_obj->pages);
+       mutex_destroy(&vgem_obj->pages_lock);
 
        if (obj->import_attach)
                drm_prime_gem_destroy(obj, vgem_obj->table);
@@ -76,11 +77,15 @@ static int vgem_gem_fault(struct vm_fault *vmf)
        if (page_offset > num_pages)
                return VM_FAULT_SIGBUS;
 
+       ret = -ENOENT;
+       mutex_lock(&obj->pages_lock);
        if (obj->pages) {
                get_page(obj->pages[page_offset]);
                vmf->page = obj->pages[page_offset];
                ret = 0;
-       } else {
+       }
+       mutex_unlock(&obj->pages_lock);
+       if (ret) {
                struct page *page;
 
                page = shmem_read_mapping_page(
@@ -161,6 +166,8 @@ static struct drm_vgem_gem_object *__vgem_gem_create(struct drm_device *dev,
                return ERR_PTR(ret);
        }
 
+       mutex_init(&obj->pages_lock);
+
        return obj;
 }
 
@@ -274,37 +281,66 @@ static const struct file_operations vgem_driver_fops = {
        .release        = drm_release,
 };
 
+static struct page **vgem_pin_pages(struct drm_vgem_gem_object *bo)
+{
+       mutex_lock(&bo->pages_lock);
+       if (bo->pages_pin_count++ == 0) {
+               struct page **pages;
+
+               pages = drm_gem_get_pages(&bo->base);
+               if (IS_ERR(pages)) {
+                       bo->pages_pin_count--;
+                       mutex_unlock(&bo->pages_lock);
+                       return pages;
+               }
+
+               bo->pages = pages;
+       }
+       mutex_unlock(&bo->pages_lock);
+
+       return bo->pages;
+}
+
+static void vgem_unpin_pages(struct drm_vgem_gem_object *bo)
+{
+       mutex_lock(&bo->pages_lock);
+       if (--bo->pages_pin_count == 0) {
+               drm_gem_put_pages(&bo->base, bo->pages, true, true);
+               bo->pages = NULL;
+       }
+       mutex_unlock(&bo->pages_lock);
+}
+
 static int vgem_prime_pin(struct drm_gem_object *obj)
 {
+       struct drm_vgem_gem_object *bo = to_vgem_bo(obj);
        long n_pages = obj->size >> PAGE_SHIFT;
        struct page **pages;
 
-       /* Flush the object from the CPU cache so that importers can rely
-        * on coherent indirect access via the exported dma-address.
-        */
-       pages = drm_gem_get_pages(obj);
+       pages = vgem_pin_pages(bo);
        if (IS_ERR(pages))
                return PTR_ERR(pages);
 
+       /* Flush the object from the CPU cache so that importers can rely
+        * on coherent indirect access via the exported dma-address.
+        */
        drm_clflush_pages(pages, n_pages);
-       drm_gem_put_pages(obj, pages, true, false);
 
        return 0;
 }
 
-static struct sg_table *vgem_prime_get_sg_table(struct drm_gem_object *obj)
+static void vgem_prime_unpin(struct drm_gem_object *obj)
 {
-       struct sg_table *st;
-       struct page **pages;
+       struct drm_vgem_gem_object *bo = to_vgem_bo(obj);
 
-       pages = drm_gem_get_pages(obj);
-       if (IS_ERR(pages))
-               return ERR_CAST(pages);
+       vgem_unpin_pages(bo);
+}
 
-       st = drm_prime_pages_to_sg(pages, obj->size >> PAGE_SHIFT);
-       drm_gem_put_pages(obj, pages, false, false);
+static struct sg_table *vgem_prime_get_sg_table(struct drm_gem_object *obj)
+{
+       struct drm_vgem_gem_object *bo = to_vgem_bo(obj);
 
-       return st;
+       return drm_prime_pages_to_sg(bo->pages, bo->base.size >> PAGE_SHIFT);
 }
 
 static struct drm_gem_object* vgem_prime_import(struct drm_device *dev,
@@ -333,6 +369,8 @@ static struct drm_gem_object *vgem_prime_import_sg_table(struct drm_device *dev,
                __vgem_gem_destroy(obj);
                return ERR_PTR(-ENOMEM);
        }
+
+       obj->pages_pin_count++; /* perma-pinned */
        drm_prime_sg_to_page_addr_arrays(obj->table, obj->pages, NULL,
                                        npages);
        return &obj->base;
@@ -340,23 +378,23 @@ static struct drm_gem_object *vgem_prime_import_sg_table(struct drm_device *dev,
 
 static void *vgem_prime_vmap(struct drm_gem_object *obj)
 {
+       struct drm_vgem_gem_object *bo = to_vgem_bo(obj);
        long n_pages = obj->size >> PAGE_SHIFT;
        struct page **pages;
-       void *addr;
 
-       pages = drm_gem_get_pages(obj);
+       pages = vgem_pin_pages(bo);
        if (IS_ERR(pages))
                return NULL;
 
-       addr = vmap(pages, n_pages, 0, pgprot_writecombine(PAGE_KERNEL));
-       drm_gem_put_pages(obj, pages, false, false);
-
-       return addr;
+       return vmap(pages, n_pages, 0, pgprot_writecombine(PAGE_KERNEL));
 }
 
 static void vgem_prime_vunmap(struct drm_gem_object *obj, void *vaddr)
 {
+       struct drm_vgem_gem_object *bo = to_vgem_bo(obj);
+
        vunmap(vaddr);
+       vgem_unpin_pages(bo);
 }
 
 static int vgem_prime_mmap(struct drm_gem_object *obj,
@@ -409,6 +447,7 @@ static struct drm_driver vgem_driver = {
        .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
        .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
        .gem_prime_pin = vgem_prime_pin,
+       .gem_prime_unpin = vgem_prime_unpin,
        .gem_prime_import = vgem_prime_import,
        .gem_prime_export = drm_gem_prime_export,
        .gem_prime_import_sg_table = vgem_prime_import_sg_table,
index 1aae0141911239e9d1ff86aa24c55a13eab21af1..5c8f6d619ff382a096d585e82b5ddf4f282366a9 100644 (file)
@@ -43,7 +43,11 @@ struct vgem_file {
 #define to_vgem_bo(x) container_of(x, struct drm_vgem_gem_object, base)
 struct drm_vgem_gem_object {
        struct drm_gem_object base;
+
        struct page **pages;
+       unsigned int pages_pin_count;
+       struct mutex pages_lock;
+
        struct sg_table *table;
 };