]> git.karo-electronics.de Git - linux-beck.git/commitdiff
drm/i915/cmdparser: Use cached vmappings
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 18 Aug 2016 16:17:12 +0000 (17:17 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 18 Aug 2016 21:36:59 +0000 (22:36 +0100)
The single largest factor in the overhead of parsing the commands is the
setup of the virtual mapping to provide a continuous block for the batch
buffer. If we keep those vmappings around (against the better judgement
of mm/vmalloc.c, which we offset by handwaving and looking suggestively
at the shrinker) we can dramatically improve the performance of the
parser for small batches (such as media workloads). Furthermore, we can
use the prepare shmem read/write functions to determine  how best we
need to clflush the range (rather than every page of the object).

The impact of caching both src/dst vmaps is +80% on ivb and +140% on byt
for the throughput on small batches. (Caching just the dst vmap and
iterating over the src, doing a page by page copy is roughly 5% slower
on both platforms. That may be an acceptable trade-off to eliminate one
cached vmapping, and we may be able to reduce the per-page copying overhead
further.) For *this* simple test case, the cmdparser is now within a
factor of 2 of ideal performance.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.william.auld@gmail.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20160818161718.27187-33-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_cmd_parser.c
drivers/gpu/drm/i915/i915_gem_execbuffer.c

index 8ebc0ce44a76f137b04ca1cbab996396c3d303c8..5d9ea163d1c812c99271f0bce98d34119e587f95 100644 (file)
@@ -937,98 +937,63 @@ find_reg_in_tables(const struct drm_i915_reg_table *tables,
        return NULL;
 }
 
-static u32 *vmap_batch(struct drm_i915_gem_object *obj,
-                      unsigned start, unsigned len)
-{
-       int i;
-       void *addr = NULL;
-       struct sg_page_iter sg_iter;
-       int first_page = start >> PAGE_SHIFT;
-       int last_page = (len + start + 4095) >> PAGE_SHIFT;
-       int npages = last_page - first_page;
-       struct page **pages;
-
-       pages = drm_malloc_ab(npages, sizeof(*pages));
-       if (pages == NULL) {
-               DRM_DEBUG_DRIVER("Failed to get space for pages\n");
-               goto finish;
-       }
-
-       i = 0;
-       for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, first_page) {
-               pages[i++] = sg_page_iter_page(&sg_iter);
-               if (i == npages)
-                       break;
-       }
-
-       addr = vmap(pages, i, 0, PAGE_KERNEL);
-       if (addr == NULL) {
-               DRM_DEBUG_DRIVER("Failed to vmap pages\n");
-               goto finish;
-       }
-
-finish:
-       if (pages)
-               drm_free_large(pages);
-       return (u32*)addr;
-}
-
-/* Returns a vmap'd pointer to dest_obj, which the caller must unmap */
-static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
+/* Returns a vmap'd pointer to dst_obj, which the caller must unmap */
+static u32 *copy_batch(struct drm_i915_gem_object *dst_obj,
                       struct drm_i915_gem_object *src_obj,
                       u32 batch_start_offset,
-                      u32 batch_len)
+                      u32 batch_len,
+                      bool *needs_clflush_after)
 {
-       unsigned int needs_clflush;
-       void *src_base, *src;
-       void *dst = NULL;
+       unsigned int src_needs_clflush;
+       unsigned int dst_needs_clflush;
+       void *src, *dst;
        int ret;
 
-       if (batch_len > dest_obj->base.size ||
-           batch_len + batch_start_offset > src_obj->base.size)
-               return ERR_PTR(-E2BIG);
-
-       if (WARN_ON(dest_obj->pages_pin_count == 0))
-               return ERR_PTR(-ENODEV);
-
-       ret = i915_gem_obj_prepare_shmem_read(src_obj, &needs_clflush);
-       if (ret) {
-               DRM_DEBUG_DRIVER("CMD: failed to prepare shadow batch\n");
+       ret = i915_gem_obj_prepare_shmem_read(src_obj, &src_needs_clflush);
+       if (ret)
                return ERR_PTR(ret);
-       }
 
-       src_base = vmap_batch(src_obj, batch_start_offset, batch_len);
-       if (!src_base) {
-               DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
-               ret = -ENOMEM;
+       ret = i915_gem_obj_prepare_shmem_write(dst_obj, &dst_needs_clflush);
+       if (ret) {
+               dst = ERR_PTR(ret);
                goto unpin_src;
        }
 
-       ret = i915_gem_object_set_to_cpu_domain(dest_obj, true);
-       if (ret) {
-               DRM_DEBUG_DRIVER("CMD: Failed to set shadow batch to CPU\n");
-               goto unmap_src;
+       src = i915_gem_object_pin_map(src_obj, I915_MAP_WB);
+       if (IS_ERR(src)) {
+               dst = src;
+               goto unpin_dst;
        }
 
-       dst = vmap_batch(dest_obj, 0, batch_len);
-       if (!dst) {
-               DRM_DEBUG_DRIVER("CMD: Failed to vmap shadow batch\n");
-               ret = -ENOMEM;
+       dst = i915_gem_object_pin_map(dst_obj, I915_MAP_WB);
+       if (IS_ERR(dst))
                goto unmap_src;
-       }
 
-       src = src_base + offset_in_page(batch_start_offset);
-       if (needs_clflush)
+       src += batch_start_offset;
+       if (src_needs_clflush)
                drm_clflush_virt_range(src, batch_len);
 
+       /* We can avoid clflushing partial cachelines before the write if we
+        * only every write full cache-lines. Since we know that both the
+        * source and destination are in multiples of PAGE_SIZE, we can simply
+        * round up to the next cacheline. We don't care about copying too much
+        * here as we only validate up to the end of the batch.
+        */
+       if (dst_needs_clflush & CLFLUSH_BEFORE)
+               batch_len = roundup(batch_len, boot_cpu_data.x86_clflush_size);
+
        memcpy(dst, src, batch_len);
 
+       /* dst_obj is returned with vmap pinned */
+       *needs_clflush_after = dst_needs_clflush & CLFLUSH_AFTER;
+
 unmap_src:
-       vunmap(src_base);
+       i915_gem_object_unpin_map(src_obj);
+unpin_dst:
+       i915_gem_obj_finish_shmem_access(dst_obj);
 unpin_src:
        i915_gem_obj_finish_shmem_access(src_obj);
-
-       return ret ? ERR_PTR(ret) : dst;
+       return dst;
 }
 
 /**
@@ -1206,16 +1171,18 @@ int intel_engine_cmd_parser(struct intel_engine_cs *engine,
                            u32 batch_len,
                            bool is_master)
 {
-       u32 *cmd, *batch_base, *batch_end;
+       u32 *cmd, *batch_end;
        struct drm_i915_cmd_descriptor default_desc = { 0 };
        bool oacontrol_set = false; /* OACONTROL tracking. See check_cmd() */
+       bool needs_clflush_after = false;
        int ret = 0;
 
-       batch_base = copy_batch(shadow_batch_obj, batch_obj,
-                               batch_start_offset, batch_len);
-       if (IS_ERR(batch_base)) {
+       cmd = copy_batch(shadow_batch_obj, batch_obj,
+                        batch_start_offset, batch_len,
+                        &needs_clflush_after);
+       if (IS_ERR(cmd)) {
                DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
-               return PTR_ERR(batch_base);
+               return PTR_ERR(cmd);
        }
 
        /*
@@ -1223,9 +1190,7 @@ int intel_engine_cmd_parser(struct intel_engine_cs *engine,
         * large or larger and copy_batch() will write MI_NOPs to the extra
         * space. Parsing should be faster in some cases this way.
         */
-       batch_end = batch_base + (batch_len / sizeof(*batch_end));
-
-       cmd = batch_base;
+       batch_end = cmd + (batch_len / sizeof(*batch_end));
        while (cmd < batch_end) {
                const struct drm_i915_cmd_descriptor *desc;
                u32 length;
@@ -1284,7 +1249,9 @@ int intel_engine_cmd_parser(struct intel_engine_cs *engine,
                ret = -EINVAL;
        }
 
-       vunmap(batch_base);
+       if (ret == 0 && needs_clflush_after)
+               drm_clflush_virt_range(shadow_batch_obj->mapping, batch_len);
+       i915_gem_object_unpin_map(shadow_batch_obj);
 
        return ret;
 }
index 907386630e264d8bc06b5ed4d6c02a267da4376a..4192066ff60e2de73530122608b7988e19536888 100644 (file)
@@ -1512,7 +1512,7 @@ execbuf_submit(struct i915_execbuffer_params *params,
                     params->args_batch_start_offset;
 
        if (exec_len == 0)
-               exec_len = params->batch->size;
+               exec_len = params->batch->size - params->args_batch_start_offset;
 
        ret = params->engine->emit_bb_start(params->request,
                                            exec_start, exec_len,
@@ -1738,6 +1738,12 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,
                ret = -EINVAL;
                goto err;
        }
+       if (args->batch_start_offset > params->batch->size ||
+           args->batch_len > params->batch->size - args->batch_start_offset) {
+               DRM_DEBUG("Attempting to use out-of-bounds batch\n");
+               ret = -EINVAL;
+               goto err;
+       }
 
        params->args_batch_start_offset = args->batch_start_offset;
        if (intel_engine_needs_cmd_parser(engine) && args->batch_len) {