From mboxrd@z Thu Jan 1 00:00:00 1970 From: Daniel Vetter Subject: [PATCH 08/16] drm/i915: drop gtt slowpath Date: Sun, 25 Mar 2012 19:47:35 +0200 Message-ID: <1332697663-31256-8-git-send-email-daniel.vetter@ffwll.ch> References: <1332697663-31256-1-git-send-email-daniel.vetter@ffwll.ch> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: Received: from mail-we0-f177.google.com (mail-we0-f177.google.com [74.125.82.177]) by gabe.freedesktop.org (Postfix) with ESMTP id E8FCD9EFFA for ; Sun, 25 Mar 2012 11:51:42 -0700 (PDT) Received: by mail-we0-f177.google.com with SMTP id p11so4606235wer.36 for ; Sun, 25 Mar 2012 11:51:42 -0700 (PDT) In-Reply-To: <1332697663-31256-1-git-send-email-daniel.vetter@ffwll.ch> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: intel-gfx-bounces+gcfxdi-intel-gfx=m.gmane.org@lists.freedesktop.org Errors-To: intel-gfx-bounces+gcfxdi-intel-gfx=m.gmane.org@lists.freedesktop.org To: Intel Graphics Development Cc: Daniel Vetter List-Id: intel-gfx@lists.freedesktop.org With the proper prefault, it's extremely unlikely that we fall back to the gtt slowpath. So just kill it and use the shmem_pwrite path as fallback. To further clean up the code, move the preparatory gem calls into the respective pwrite functions. This way the gtt_fast->shmem fallback is much more obvious. Signed-off-by: Daniel Vetter --- drivers/gpu/drm/i915/i915_gem.c | 183 +++++++-------------------------------- 1 files changed, 30 insertions(+), 153 deletions(-) diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 117fda4..ddd1fa4 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -504,30 +504,6 @@ fast_user_write(struct io_mapping *mapping, return unwritten; } -/* Here's the write path which can sleep for - * page faults - */ - -static inline void -slow_kernel_write(struct io_mapping *mapping, - loff_t gtt_base, int gtt_offset, - struct page *user_page, int user_offset, - int length) -{ - char __iomem *dst_vaddr; - char *src_vaddr; - - dst_vaddr = io_mapping_map_wc(mapping, gtt_base); - src_vaddr = kmap(user_page); - - memcpy_toio(dst_vaddr + gtt_offset, - src_vaddr + user_offset, - length); - - kunmap(user_page); - io_mapping_unmap(dst_vaddr); -} - /** * This is the fast pwrite path, where we copy the data directly from the * user into the GTT, uncached. @@ -542,7 +518,19 @@ i915_gem_gtt_pwrite_fast(struct drm_device *dev, ssize_t remain; loff_t offset, page_base; char __user *user_data; - int page_offset, page_length; + int page_offset, page_length, ret; + + ret = i915_gem_object_pin(obj, 0, true); + if (ret) + goto out; + + ret = i915_gem_object_set_to_gtt_domain(obj, true); + if (ret) + goto out_unpin; + + ret = i915_gem_object_put_fence(obj); + if (ret) + goto out_unpin; user_data = (char __user *) (uintptr_t) args->data_ptr; remain = args->size; @@ -567,112 +555,19 @@ i915_gem_gtt_pwrite_fast(struct drm_device *dev, * retry in the slow path. */ if (fast_user_write(dev_priv->mm.gtt_mapping, page_base, - page_offset, user_data, page_length)) - return -EFAULT; + page_offset, user_data, page_length)) { + ret = -EFAULT; + goto out_unpin; + } remain -= page_length; user_data += page_length; offset += page_length; } - return 0; -} - -/** - * This is the fallback GTT pwrite path, which uses get_user_pages to pin - * the memory and maps it using kmap_atomic for copying. - * - * This code resulted in x11perf -rgb10text consuming about 10% more CPU - * than using i915_gem_gtt_pwrite_fast on a G45 (32-bit). - */ -static int -i915_gem_gtt_pwrite_slow(struct drm_device *dev, - struct drm_i915_gem_object *obj, - struct drm_i915_gem_pwrite *args, - struct drm_file *file) -{ - drm_i915_private_t *dev_priv = dev->dev_private; - ssize_t remain; - loff_t gtt_page_base, offset; - loff_t first_data_page, last_data_page, num_pages; - loff_t pinned_pages, i; - struct page **user_pages; - struct mm_struct *mm = current->mm; - int gtt_page_offset, data_page_offset, data_page_index, page_length; - int ret; - uint64_t data_ptr = args->data_ptr; - - remain = args->size; - - /* Pin the user pages containing the data. We can't fault while - * holding the struct mutex, and all of the pwrite implementations - * want to hold it while dereferencing the user data. - */ - first_data_page = data_ptr / PAGE_SIZE; - last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE; - num_pages = last_data_page - first_data_page + 1; - - user_pages = drm_malloc_ab(num_pages, sizeof(struct page *)); - if (user_pages == NULL) - return -ENOMEM; - - mutex_unlock(&dev->struct_mutex); - down_read(&mm->mmap_sem); - pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr, - num_pages, 0, 0, user_pages, NULL); - up_read(&mm->mmap_sem); - mutex_lock(&dev->struct_mutex); - if (pinned_pages < num_pages) { - ret = -EFAULT; - goto out_unpin_pages; - } - - ret = i915_gem_object_set_to_gtt_domain(obj, true); - if (ret) - goto out_unpin_pages; - - ret = i915_gem_object_put_fence(obj); - if (ret) - goto out_unpin_pages; - - offset = obj->gtt_offset + args->offset; - - while (remain > 0) { - /* Operation in this page - * - * gtt_page_base = page offset within aperture - * gtt_page_offset = offset within page in aperture - * data_page_index = page number in get_user_pages return - * data_page_offset = offset with data_page_index page. - * page_length = bytes to copy for this page - */ - gtt_page_base = offset & PAGE_MASK; - gtt_page_offset = offset_in_page(offset); - data_page_index = data_ptr / PAGE_SIZE - first_data_page; - data_page_offset = offset_in_page(data_ptr); - - page_length = remain; - if ((gtt_page_offset + page_length) > PAGE_SIZE) - page_length = PAGE_SIZE - gtt_page_offset; - if ((data_page_offset + page_length) > PAGE_SIZE) - page_length = PAGE_SIZE - data_page_offset; - - slow_kernel_write(dev_priv->mm.gtt_mapping, - gtt_page_base, gtt_page_offset, - user_pages[data_page_index], - data_page_offset, - page_length); - - remain -= page_length; - offset += page_length; - data_ptr += page_length; - } - -out_unpin_pages: - for (i = 0; i < pinned_pages; i++) - page_cache_release(user_pages[i]); - drm_free_large(user_pages); - +out_unpin: + i915_gem_object_unpin(obj); +out: return ret; } @@ -691,6 +586,10 @@ i915_gem_shmem_pwrite(struct drm_device *dev, int hit_slowpath = 0; int release_page; + ret = i915_gem_object_set_to_cpu_domain(obj, 1); + if (ret) + return ret; + user_data = (char __user *) (uintptr_t) args->data_ptr; remain = args->size; @@ -834,6 +733,7 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, trace_i915_gem_object_pwrite(obj, args->offset, args->size); + ret = -EFAULT; /* We can only do the GTT pwrite on untiled buffers, as otherwise * it would end up going through the fenced access, and we'll get * different detiling behavior between reading and writing. @@ -848,37 +748,14 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, if (obj->gtt_space && obj->cache_level == I915_CACHE_NONE && obj->base.write_domain != I915_GEM_DOMAIN_CPU) { - ret = i915_gem_object_pin(obj, 0, true); - if (ret) - goto out; - - ret = i915_gem_object_set_to_gtt_domain(obj, true); - if (ret) - goto out_unpin; - - ret = i915_gem_object_put_fence(obj); - if (ret) - goto out_unpin; - ret = i915_gem_gtt_pwrite_fast(dev, obj, args, file); - if (ret == -EFAULT) - ret = i915_gem_gtt_pwrite_slow(dev, obj, args, file); - -out_unpin: - i915_gem_object_unpin(obj); - - if (ret != -EFAULT) - goto out; - /* Fall through to the shmfs paths because the gtt paths might - * fail with non-page-backed user pointers (e.g. gtt mappings - * when moving data between textures). */ + /* Note that the gtt paths might fail with non-page-backed user + * pointers (e.g. gtt mappings when moving data between + * textures). Fallback to the shmem path in that case. */ } - ret = i915_gem_object_set_to_cpu_domain(obj, 1); - if (ret) - goto out; - - ret = i915_gem_shmem_pwrite(dev, obj, args, file); + if (ret == -EFAULT) + ret = i915_gem_shmem_pwrite(dev, obj, args, file); out: drm_gem_object_unreference(&obj->base); -- 1.7.7.6