drm/i915: Fallback to using CPU relocations for large batch buffers
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 14 Jan 2015 11:20:56 +0000 (11:20 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 20 Mar 2015 10:48:13 +0000 (11:48 +0100)
If the batch buffer is too large to fit into the aperture and we need a
GTT mapping for relocations, we currently fail. This only applies to a
subset of machines for a subset of environments, quite undesirable. We
can simply check after failing to insert the batch into the GTT as to
whether we only need a mappable binding for relocation and, if so, we can
revert to using a non-mappable binding and an alternate relocation
method. However, using relocate_entry_cpu() is excruciatingly slow for
large buffers on non-LLC as the entire buffer requires clflushing before
and after the relocation handling. Alternatively, we can implement a
third relocation method that only clflushes around the relocation entry.
This is still slower than updating through the GTT, so we prefer using
the GTT where possible, but is orders of magnitude faster as we
typically do not have to then clflush the entire buffer.

An alternative idea of using a temporary WC mapping of the backing store
is promising (it should be faster than using the GTT itself), but
requires fairly extensive arch/x86 support - along the lines of
kmap_atomic_prof_pfn() (which is not universally implemented even for
x86).

Testcase: igt/gem_exec_big #pnv,byt
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=88392
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add a WARN_ONCE for the impossible reloc case and explain in
a short comment why we want to avoid ping-pong.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_gem_execbuffer.c

index dc10bc43864e30497122ac98e746a3896573c14e..9838d11e43ee622ec770edb2e3ec719da48fe5f9 100644 (file)
@@ -251,7 +251,6 @@ static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
 {
        return (HAS_LLC(obj->base.dev) ||
                obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
-               !obj->map_and_fenceable ||
                obj->cache_level != I915_CACHE_NONE);
 }
 
@@ -337,6 +336,51 @@ relocate_entry_gtt(struct drm_i915_gem_object *obj,
        return 0;
 }
 
+static void
+clflush_write32(void *addr, uint32_t value)
+{
+       /* This is not a fast path, so KISS. */
+       drm_clflush_virt_range(addr, sizeof(uint32_t));
+       *(uint32_t *)addr = value;
+       drm_clflush_virt_range(addr, sizeof(uint32_t));
+}
+
+static int
+relocate_entry_clflush(struct drm_i915_gem_object *obj,
+                      struct drm_i915_gem_relocation_entry *reloc,
+                      uint64_t target_offset)
+{
+       struct drm_device *dev = obj->base.dev;
+       uint32_t page_offset = offset_in_page(reloc->offset);
+       uint64_t delta = (int)reloc->delta + target_offset;
+       char *vaddr;
+       int ret;
+
+       ret = i915_gem_object_set_to_gtt_domain(obj, true);
+       if (ret)
+               return ret;
+
+       vaddr = kmap_atomic(i915_gem_object_get_page(obj,
+                               reloc->offset >> PAGE_SHIFT));
+       clflush_write32(vaddr + page_offset, lower_32_bits(delta));
+
+       if (INTEL_INFO(dev)->gen >= 8) {
+               page_offset = offset_in_page(page_offset + sizeof(uint32_t));
+
+               if (page_offset == 0) {
+                       kunmap_atomic(vaddr);
+                       vaddr = kmap_atomic(i915_gem_object_get_page(obj,
+                           (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
+               }
+
+               clflush_write32(vaddr + page_offset, upper_32_bits(delta));
+       }
+
+       kunmap_atomic(vaddr);
+
+       return 0;
+}
+
 static int
 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
                                   struct eb_vmas *eb,
@@ -426,8 +470,14 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 
        if (use_cpu_reloc(obj))
                ret = relocate_entry_cpu(obj, reloc, target_offset);
-       else
+       else if (obj->map_and_fenceable)
                ret = relocate_entry_gtt(obj, reloc, target_offset);
+       else if (cpu_has_clflush)
+               ret = relocate_entry_clflush(obj, reloc, target_offset);
+       else {
+               WARN_ONCE(1, "Impossible case in relocation handling\n");
+               ret = -ENODEV;
+       }
 
        if (ret)
                return ret;
@@ -525,6 +575,12 @@ i915_gem_execbuffer_relocate(struct eb_vmas *eb)
        return ret;
 }
 
+static bool only_mappable_for_reloc(unsigned int flags)
+{
+       return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
+               __EXEC_OBJECT_NEEDS_MAP;
+}
+
 static int
 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
                                struct intel_engine_cs *ring,
@@ -536,14 +592,21 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
        int ret;
 
        flags = 0;
-       if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
-               flags |= PIN_GLOBAL | PIN_MAPPABLE;
-       if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
-               flags |= PIN_GLOBAL;
-       if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
-               flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
+       if (!drm_mm_node_allocated(&vma->node)) {
+               if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
+                       flags |= PIN_GLOBAL | PIN_MAPPABLE;
+               if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
+                       flags |= PIN_GLOBAL;
+               if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
+                       flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
+       }
 
        ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
+       if ((ret == -ENOSPC  || ret == -E2BIG) &&
+           only_mappable_for_reloc(entry->flags))
+               ret = i915_gem_object_pin(obj, vma->vm,
+                                         entry->alignment,
+                                         flags & ~(PIN_GLOBAL | PIN_MAPPABLE));
        if (ret)
                return ret;
 
@@ -605,13 +668,14 @@ eb_vma_misplaced(struct i915_vma *vma)
            vma->node.start & (entry->alignment - 1))
                return true;
 
-       if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
-               return true;
-
        if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
            vma->node.start < BATCH_OFFSET_BIAS)
                return true;
 
+       /* avoid costly ping-pong once a batch bo ended up non-mappable */
+       if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
+               return !only_mappable_for_reloc(entry->flags);
+
        return false;
 }