drm/i915: Introduce an internal allocator for disposable private objects
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 12:58:30 +0000 (13:58 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 28 Oct 2016 19:53:44 +0000 (20:53 +0100)
Quite a few of our objects used for internal hardware programming do not
benefit from being swappable or from being zero initialised. As such
they do not benefit from using a shmemfs backing storage and since they
are internal and never directly exposed to the user, we do not need to
worry about providing a filp. For these we can use an
drm_i915_gem_object wrapper around a sg_table of plain struct page. They
are not swap backed and not automatically pinned. If they are reaped
by the shrinker, the pages are released and the contents discarded. For
the internal use case, this is fine as for example, ringbuffers are
pinned from being written by a request to be read by the hardware. Once
they are idle, they can be discarded entirely. As such they are a good
match for execlist ringbuffers and a small variety of other internal
objects.

In the first iteration, this is limited to the scratch batch buffers we
use (for command parsing and state initialisation).

v2: Allocate physically contiguous pages, where possible.
v3: Reduce maximum order on subsequent requests following an allocation
failure.
v4: Fix up mismatch between swiotlb segment size and page count (it
counts in 2k units, not 4k pages)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-7-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/Makefile
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_gem_batch_pool.c
drivers/gpu/drm/i915/i915_gem_internal.c [new file with mode: 0644]
drivers/gpu/drm/i915/i915_gem_render_state.c
drivers/gpu/drm/i915/intel_engine_cs.c
drivers/gpu/drm/i915/intel_ringbuffer.c

index 612340097f4ba2f083076b4ca3c06694b11835f1..7faa04c91e1af8d93d2bbadf7b52f6f40583c086 100644 (file)
@@ -35,6 +35,7 @@ i915-y += i915_cmd_parser.o \
          i915_gem_execbuffer.o \
          i915_gem_fence.o \
          i915_gem_gtt.o \
+         i915_gem_internal.o \
          i915_gem.o \
          i915_gem_render_state.o \
          i915_gem_request.o \
index edc59d08d0171a9170163d80c8b01b7294e566c3..96995d644da22be95bb314bf3d9a5d2785b0e21d 100644 (file)
@@ -3556,6 +3556,11 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
                                               u32 gtt_offset,
                                               u32 size);
 
+/* i915_gem_internal.c */
+struct drm_i915_gem_object *
+i915_gem_object_create_internal(struct drm_i915_private *dev_priv,
+                               unsigned int size);
+
 /* i915_gem_shrinker.c */
 unsigned long i915_gem_shrink(struct drm_i915_private *dev_priv,
                              unsigned long target,
index cb25cad3318c6d5db1744aa0367bc426974e6cc5..aa4e1e043b4e7de41d840a61704516a2c6d94b2b 100644 (file)
@@ -97,9 +97,9 @@ i915_gem_batch_pool_get(struct i915_gem_batch_pool *pool,
                        size_t size)
 {
        struct drm_i915_gem_object *obj = NULL;
-       struct drm_i915_gem_object *tmp, *next;
+       struct drm_i915_gem_object *tmp;
        struct list_head *list;
-       int n;
+       int n, ret;
 
        lockdep_assert_held(&pool->engine->i915->drm.struct_mutex);
 
@@ -112,19 +112,12 @@ i915_gem_batch_pool_get(struct i915_gem_batch_pool *pool,
                n = ARRAY_SIZE(pool->cache_list) - 1;
        list = &pool->cache_list[n];
 
-       list_for_each_entry_safe(tmp, next, list, batch_pool_link) {
+       list_for_each_entry(tmp, list, batch_pool_link) {
                /* The batches are strictly LRU ordered */
                if (!i915_gem_active_is_idle(&tmp->last_read[pool->engine->id],
                                             &tmp->base.dev->struct_mutex))
                        break;
 
-               /* While we're looping, do some clean up */
-               if (tmp->madv == __I915_MADV_PURGED) {
-                       list_del(&tmp->batch_pool_link);
-                       i915_gem_object_put(tmp);
-                       continue;
-               }
-
                if (tmp->base.size >= size) {
                        obj = tmp;
                        break;
@@ -132,19 +125,15 @@ i915_gem_batch_pool_get(struct i915_gem_batch_pool *pool,
        }
 
        if (obj == NULL) {
-               int ret;
-
-               obj = i915_gem_object_create(&pool->engine->i915->drm, size);
+               obj = i915_gem_object_create_internal(pool->engine->i915, size);
                if (IS_ERR(obj))
                        return obj;
-
-               ret = i915_gem_object_get_pages(obj);
-               if (ret)
-                       return ERR_PTR(ret);
-
-               obj->madv = I915_MADV_DONTNEED;
        }
 
+       ret = i915_gem_object_get_pages(obj);
+       if (ret)
+               return ERR_PTR(ret);
+
        list_move_tail(&obj->batch_pool_link, list);
        i915_gem_object_pin_pages(obj);
        return obj;
diff --git a/drivers/gpu/drm/i915/i915_gem_internal.c b/drivers/gpu/drm/i915/i915_gem_internal.c
new file mode 100644 (file)
index 0000000..02e66fa
--- /dev/null
@@ -0,0 +1,170 @@
+/*
+ * Copyright © 2014-2016 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ */
+
+#include <drm/drmP.h>
+#include <drm/i915_drm.h>
+#include "i915_drv.h"
+
+#define QUIET (__GFP_NORETRY | __GFP_NOWARN)
+
+/* convert swiotlb segment size into sensible units (pages)! */
+#define IO_TLB_SEGPAGES (IO_TLB_SEGSIZE << IO_TLB_SHIFT >> PAGE_SHIFT)
+
+static void internal_free_pages(struct sg_table *st)
+{
+       struct scatterlist *sg;
+
+       for (sg = st->sgl; sg; sg = __sg_next(sg))
+               __free_pages(sg_page(sg), get_order(sg->length));
+
+       sg_free_table(st);
+       kfree(st);
+}
+
+static int i915_gem_object_get_pages_internal(struct drm_i915_gem_object *obj)
+{
+       struct drm_i915_private *i915 = to_i915(obj->base.dev);
+       unsigned int npages = obj->base.size / PAGE_SIZE;
+       struct sg_table *st;
+       struct scatterlist *sg;
+       int max_order;
+       gfp_t gfp;
+
+       st = kmalloc(sizeof(*st), GFP_KERNEL);
+       if (!st)
+               return -ENOMEM;
+
+       if (sg_alloc_table(st, npages, GFP_KERNEL)) {
+               kfree(st);
+               return -ENOMEM;
+       }
+
+       sg = st->sgl;
+       st->nents = 0;
+
+       max_order = MAX_ORDER;
+#ifdef CONFIG_SWIOTLB
+       if (swiotlb_nr_tbl()) /* minimum max swiotlb size is IO_TLB_SEGSIZE */
+               max_order = min(max_order, ilog2(IO_TLB_SEGPAGES));
+#endif
+
+       gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_RECLAIMABLE;
+       if (IS_CRESTLINE(i915) || IS_BROADWATER(i915)) {
+               /* 965gm cannot relocate objects above 4GiB. */
+               gfp &= ~__GFP_HIGHMEM;
+               gfp |= __GFP_DMA32;
+       }
+
+       do {
+               int order = min(fls(npages) - 1, max_order);
+               struct page *page;
+
+               do {
+                       page = alloc_pages(gfp | (order ? QUIET : 0), order);
+                       if (page)
+                               break;
+                       if (!order--)
+                               goto err;
+
+                       /* Limit subsequent allocations as well */
+                       max_order = order;
+               } while (1);
+
+               sg_set_page(sg, page, PAGE_SIZE << order, 0);
+               st->nents++;
+
+               npages -= 1 << order;
+               if (!npages) {
+                       sg_mark_end(sg);
+                       break;
+               }
+
+               sg = __sg_next(sg);
+       } while (1);
+       obj->pages = st;
+
+       if (i915_gem_gtt_prepare_object(obj)) {
+               obj->pages = NULL;
+               goto err;
+       }
+
+       /* Mark the pages as dontneed whilst they are still pinned. As soon
+        * as they are unpinned they are allowed to be reaped by the shrinker,
+        * and the caller is expected to repopulate - the contents of this
+        * object are only valid whilst active and pinned.
+        */
+       obj->madv = I915_MADV_DONTNEED;
+       return 0;
+
+err:
+       sg_mark_end(sg);
+       internal_free_pages(st);
+       return -ENOMEM;
+}
+
+static void i915_gem_object_put_pages_internal(struct drm_i915_gem_object *obj)
+{
+       i915_gem_gtt_finish_object(obj);
+       internal_free_pages(obj->pages);
+
+       obj->dirty = 0;
+       obj->madv = I915_MADV_WILLNEED;
+}
+
+static const struct drm_i915_gem_object_ops i915_gem_object_internal_ops = {
+       .flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE,
+       .get_pages = i915_gem_object_get_pages_internal,
+       .put_pages = i915_gem_object_put_pages_internal,
+};
+
+/**
+ * Creates a new object that wraps some internal memory for private use.
+ * This object is not backed by swappable storage, and as such its contents
+ * are volatile and only valid whilst pinned. If the object is reaped by the
+ * shrinker, its pages and data will be discarded. Equally, it is not a full
+ * GEM object and so not valid for access from userspace. This makes it useful
+ * for hardware interfaces like ringbuffers (which are pinned from the time
+ * the request is written to the time the hardware stops accessing it), but
+ * not for contexts (which need to be preserved when not active for later
+ * reuse). Note that it is not cleared upon allocation.
+ */
+struct drm_i915_gem_object *
+i915_gem_object_create_internal(struct drm_i915_private *i915,
+                               unsigned int size)
+{
+       struct drm_i915_gem_object *obj;
+
+       obj = i915_gem_object_alloc(&i915->drm);
+       if (!obj)
+               return ERR_PTR(-ENOMEM);
+
+       drm_gem_private_object_init(&i915->drm, &obj->base, size);
+       i915_gem_object_init(obj, &i915_gem_object_internal_ops);
+
+       obj->base.write_domain = I915_GEM_DOMAIN_CPU;
+       obj->base.read_domains = I915_GEM_DOMAIN_CPU;
+       obj->cache_level = HAS_LLC(i915) ? I915_CACHE_LLC : I915_CACHE_NONE;
+
+       return obj;
+}
index e7c3dbcc6c8136c45c48742b50123ee7dafed7a5..217e0b58b9303db75e10b52e5567048610bcdf1f 100644 (file)
@@ -187,7 +187,7 @@ int i915_gem_render_state_init(struct drm_i915_gem_request *req)
        if (so.rodata->batch_items * 4 > 4096)
                return -EINVAL;
 
-       obj = i915_gem_object_create(&req->i915->drm, 4096);
+       obj = i915_gem_object_create_internal(req->i915, 4096);
        if (IS_ERR(obj))
                return PTR_ERR(obj);
 
index 8cceb345aa0f22ad71e3e1b93f6a232422a60d84..b2de371d2bf5922442b7a9dded87d7e40af415d6 100644 (file)
@@ -264,7 +264,7 @@ int intel_engine_create_scratch(struct intel_engine_cs *engine, int size)
 
        obj = i915_gem_object_create_stolen(&engine->i915->drm, size);
        if (!obj)
-               obj = i915_gem_object_create(&engine->i915->drm, size);
+               obj = i915_gem_object_create_internal(engine->i915, size);
        if (IS_ERR(obj)) {
                DRM_ERROR("Failed to allocate scratch page\n");
                return PTR_ERR(obj);
index 8eee9675d3bfcec570742659ca702b224fbd9849..a15b9b5f2924804cc4091206758df56895438d85 100644 (file)
@@ -1782,9 +1782,10 @@ static int init_status_page(struct intel_engine_cs *engine)
        struct drm_i915_gem_object *obj;
        struct i915_vma *vma;
        unsigned int flags;
+       void *vaddr;
        int ret;
 
-       obj = i915_gem_object_create(&engine->i915->drm, 4096);
+       obj = i915_gem_object_create_internal(engine->i915, 4096);
        if (IS_ERR(obj)) {
                DRM_ERROR("Failed to allocate status page\n");
                return PTR_ERR(obj);
@@ -1817,15 +1818,22 @@ static int init_status_page(struct intel_engine_cs *engine)
        if (ret)
                goto err;
 
+       vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
+       if (IS_ERR(vaddr)) {
+               ret = PTR_ERR(vaddr);
+               goto err_unpin;
+       }
+
        engine->status_page.vma = vma;
        engine->status_page.ggtt_offset = i915_ggtt_offset(vma);
-       engine->status_page.page_addr =
-               i915_gem_object_pin_map(obj, I915_MAP_WB);
+       engine->status_page.page_addr = memset(vaddr, 0, 4096);
 
        DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
                         engine->name, i915_ggtt_offset(vma));
        return 0;
 
+err_unpin:
+       i915_vma_unpin(vma);
 err:
        i915_gem_object_put(obj);
        return ret;