From: Ben Widawsky <benjamin.widawsky@xxxxxxxxx>
This patch continues on the idea from the previous patch. From here on,
in the steady state, PDEs are all pointing to the scratch page table (as
recommended in the spec). When an object is allocated in the VA range,
the code will determine if we need to allocate a page for the page
table. Similarly when the object is destroyed, we will remove, and free
the page table pointing the PDE back to the scratch page.
Following patches will work to unify the code a bit as we bring in GEN8
support. GEN6 and GEN8 are different enough that I had a hard time to
get to this point with as much common code as I do.
The aliasing PPGTT must pre-allocate all of the page tables. There are a
few reasons for this. Two trivial ones: aliasing ppgtt goes through the
ggtt paths, so it's hard to maintain, we currently do not restore the
default context (assuming the previous force reload is indeed
necessary). Most importantly though, the only way (it seems from
empirical evidence) to invalidate the CS TLBs on non-render ring is to
either use ring sync (which requires actually stopping the rings in
order to synchronize when the sync completes vs. where you are in
execution), or to reload DCLV. Since without full PPGTT we do not ever
reload the DCLV register, there is no good way to achieve this. The
simplest solution is just to not support dynamic page table
creation/destruction in the aliasing PPGTT.
We could always reload DCLV, but this seems like quite a bit of excess
overhead only to save at most 2MB-4k of memory for the aliasing PPGTT
page tables.
v2: Make the page table bitmap declared inside the function (Chris)
Simplify the way scratching address space works.
Move the alloc/teardown tracepoints up a level in the call stack so that
both all implementations get the trace.
v3: Updated trace event to spit out a name
v4: Aliasing ppgtt is now initialized differently (in setup global gtt)
v5: Rebase to latest code. Also removed unnecessary aliasing ppgtt check
for trace, as it is no longer possible after the PPGTT cleanup patch series
of a couple of months ago (Daniel).
v6: Implement changes from code review (Daniel):
- allocate/teardown_va_range calls added.
- Add a scratch page allocation helper (only need the address).
- Move trace events to a new patch.
- Use updated mark_tlbs_dirty.
- Moved pt preallocation for aliasing ppgtt into gen6_ppgtt_init.
Cc: Daniel Vetter <daniel@xxxxxxxx>
Signed-off-by: Ben Widawsky <ben@xxxxxxxxxxxx>
Signed-off-by: Michel Thierry <michel.thierry@xxxxxxxxx> (v4+)
---
drivers/gpu/drm/i915/i915_debugfs.c | 3 +-
drivers/gpu/drm/i915/i915_gem.c | 9 +++
drivers/gpu/drm/i915/i915_gem_gtt.c | 143 ++++++++++++++++++++++++++++++++----
drivers/gpu/drm/i915/i915_gem_gtt.h | 3 +
4 files changed, 142 insertions(+), 16 deletions(-)
diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index 60f91bc..0f63076 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -2149,6 +2149,8 @@ static void gen6_ppgtt_info(struct seq_file *m, struct drm_device *dev)
seq_printf(m, "PP_DIR_BASE_READ: 0x%08x\n", I915_READ(RING_PP_DIR_BASE_READ(ring)));
seq_printf(m, "PP_DIR_DCLV: 0x%08x\n", I915_READ(RING_PP_DIR_DCLV(ring)));
}
+ seq_printf(m, "ECOCHK: 0x%08x\n\n", I915_READ(GAM_ECOCHK));
+
if (dev_priv->mm.aliasing_ppgtt) {
struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
@@ -2165,7 +2167,6 @@ static void gen6_ppgtt_info(struct seq_file *m, struct drm_device *dev)
get_pid_task(file->pid, PIDTYPE_PID)->comm);
idr_for_each(&file_priv->context_idr, per_file_ctx, m);
}
- seq_printf(m, "ECOCHK: 0x%08x\n", I915_READ(GAM_ECOCHK));
}
static int i915_ppgtt_info(struct seq_file *m, void *data)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 6c40365..65e055c 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3585,6 +3585,15 @@ search_free:
if (ret)
goto err_remove_node;
+ /* allocate before insert / bind */
+ if (vma->vm->allocate_va_range) {
+ ret = vma->vm->allocate_va_range(vma->vm,
+ vma->node.start,
+ vma->node.size);
+ if (ret)
+ goto err_remove_node;
+ }
+
trace_i915_vma_bind(vma, flags);
ret = i915_vma_bind(vma, obj->cache_level,
flags & PIN_GLOBAL ? GLOBAL_BIND : 0);
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c b/drivers/gpu/drm/i915/i915_gem_gtt.c
index 62f492f..d37bd83 100644
--- a/drivers/gpu/drm/i915/i915_gem_gtt.c
+++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
@@ -316,6 +316,14 @@ static void unmap_and_free_pt(struct i915_page_table_entry *pt,
if (WARN_ON(!pt->page))
return;
+ if (!pt->scratch) {
+ const size_t count = INTEL_INFO(dev)->gen >= 8 ?
+ GEN8_PTES_PER_PAGE : I915_PPGTT_PT_ENTRIES;
+ WARN(!bitmap_empty(pt->used_ptes, count),
+ "Free page table with %d used pages\n",
+ bitmap_weight(pt->used_ptes, count));
+ }
+
i915_dma_unmap_single(pt, dev);
__free_page(pt->page);
kfree(pt->used_ptes);
@@ -359,6 +367,15 @@ fail_bitmap:
return ERR_PTR(ret);
}
+static inline struct i915_page_table_entry *alloc_pt_scratch(struct drm_device *dev)
+{
+ struct i915_page_table_entry *pt = alloc_pt_single(dev);
+ if (!IS_ERR(pt))
+ pt->scratch = 1;
+
+ return pt;
+}
+
/**
* alloc_pt_range() - Allocate a multiple page tables
* @pd: The page directory which will have at least @count entries
@@ -1123,10 +1140,46 @@ static inline void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
static int gen6_alloc_va_range(struct i915_address_space *vm,
uint64_t start, uint64_t length)
{
+ DECLARE_BITMAP(new_page_tables, GEN6_PPGTT_PD_ENTRIES);
+ struct drm_device *dev = vm->dev;
+ struct drm_i915_private *dev_priv = dev->dev_private;
struct i915_hw_ppgtt *ppgtt =
container_of(vm, struct i915_hw_ppgtt, base);
struct i915_page_table_entry *pt;
+ const uint32_t start_save = start, length_save = length;
uint32_t pde, temp;
+ int ret;
+
+ BUG_ON(upper_32_bits(start));
+
+ bitmap_zero(new_page_tables, GEN6_PPGTT_PD_ENTRIES);
+
+ /* The allocation is done in two stages so that we can bail out with
+ * minimal amount of pain. The first stage finds new page tables that
+ * need allocation. The second stage marks use ptes within the page
+ * tables.
+ */
+ gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
+ if (pt != ppgtt->scratch_pt) {
+ WARN_ON(bitmap_empty(pt->used_ptes, I915_PPGTT_PT_ENTRIES));
+ continue;
+ }
+
+ /* We've already allocated a page table */
+ WARN_ON(!bitmap_empty(pt->used_ptes, I915_PPGTT_PT_ENTRIES));
+
+ pt = alloc_pt_single(dev);
+ if (IS_ERR(pt)) {
+ ret = PTR_ERR(pt);
+ goto unwind_out;
+ }
+
+ ppgtt->pd.page_tables[pde] = pt;
+ set_bit(pde, new_page_tables);
+ }
+
+ start = start_save;
+ length = length_save;
gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
DECLARE_BITMAP(tmp_bitmap, I915_PPGTT_PT_ENTRIES);
@@ -1134,12 +1187,31 @@ static int gen6_alloc_va_range(struct i915_address_space *vm,
bitmap_set(tmp_bitmap, gen6_pte_index(start),
gen6_pte_count(start, length));
- bitmap_or(pt->used_ptes, pt->used_ptes, tmp_bitmap,
+ if (test_and_clear_bit(pde, new_page_tables))
+ gen6_write_pdes(&ppgtt->pd, pde, pt);
+
+ bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes,
I915_PPGTT_PT_ENTRIES);
}
+ WARN_ON(!bitmap_empty(new_page_tables, GEN6_PPGTT_PD_ENTRIES));
+
+ /* Make sure write is complete before other code can use this page
+ * table. Also require for WC mapped PTEs */
+ readl(dev_priv->gtt.gsm);
+
mark_tlbs_dirty(ppgtt);
return 0;
+
+unwind_out:
+ for_each_set_bit(pde, new_page_tables, GEN6_PPGTT_PD_ENTRIES) {
+ struct i915_page_table_entry *pt = ppgtt->pd.page_tables[pde];
+ ppgtt->pd.page_tables[pde] = NULL;
+ unmap_and_free_pt(pt, vm->dev);
+ }
+
+ mark_tlbs_dirty(ppgtt);
+ return ret;
}
static void gen6_teardown_va_range(struct i915_address_space *vm,
@@ -1151,8 +1223,19 @@ static void gen6_teardown_va_range(struct i915_address_space *vm,
uint32_t pde, temp;
gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) {
+
+ if (WARN(pt == ppgtt->scratch_pt,
+ "Tried to teardown scratch page vm %p. pde %u: %llx-%llx\n",
+ vm, pde, start, start + length))
+ continue;
+
bitmap_clear(pt->used_ptes, gen6_pte_index(start),
gen6_pte_count(start, length));
+
+ if (bitmap_empty(pt->used_ptes, I915_PPGTT_PT_ENTRIES)) {
+ gen6_write_pdes(&ppgtt->pd, pde, ppgtt->scratch_pt);
+ ppgtt->pd.page_tables[pde] = ppgtt->scratch_pt;
+ }
}
mark_tlbs_dirty(ppgtt);
@@ -1162,9 +1245,13 @@ static void gen6_ppgtt_free(struct i915_hw_ppgtt *ppgtt)
{
int i;
- for (i = 0; i < ppgtt->num_pd_entries; i++)
- unmap_and_free_pt(ppgtt->pd.page_tables[i], ppgtt->base.dev);
+ for (i = 0; i < ppgtt->num_pd_entries; i++) {
+ struct i915_page_table_entry *pt = ppgtt->pd.page_tables[i];
+ if (pt != ppgtt->scratch_pt)
+ unmap_and_free_pt(ppgtt->pd.page_tables[i], ppgtt->base.dev);
+ }
+ unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
unmap_and_free_pd(&ppgtt->pd);
}
@@ -1191,6 +1278,9 @@ static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt)
* size. We allocate at the top of the GTT to avoid fragmentation.
*/
BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm));
+ ppgtt->scratch_pt = alloc_pt_scratch(ppgtt->base.dev);
+ if (IS_ERR(ppgtt->scratch_pt))
+ return PTR_ERR(ppgtt->scratch_pt);
alloc:
ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm,
&ppgtt->node, GEN6_PD_SIZE,
@@ -1221,6 +1311,7 @@ alloc:
return 0;
err_out:
+ unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
return ret;
}
@@ -1232,18 +1323,20 @@ static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt)
if (ret)
return ret;
- ret = alloc_pt_range(&ppgtt->pd, 0, ppgtt->num_pd_entries,
- ppgtt->base.dev);
+ return 0;
+}
- if (ret) {
- drm_mm_remove_node(&ppgtt->node);
- return ret;
- }
+static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt,
+ uint64_t start, uint64_t length)
+{
+ struct i915_page_table_entry *unused;
+ uint32_t pde, temp;
- return 0;
+ gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde)
+ ppgtt->pd.page_tables[pde] = ppgtt->scratch_pt;
}
-static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
+static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt, bool aliasing)
{
struct drm_device *dev = ppgtt->base.dev;
struct drm_i915_private *dev_priv = dev->dev_private;
@@ -1263,6 +1356,18 @@ static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
if (ret)
return ret;
+ if (aliasing) {
+ /* preallocate all pts */
+ ret = alloc_pt_range(&ppgtt->pd, 0, ppgtt->num_pd_entries,
+ ppgtt->base.dev);
+
+ if (ret) {
+ unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev);
+ drm_mm_remove_node(&ppgtt->node);
+ return ret;
+ }
+ }
+
ppgtt->base.allocate_va_range = gen6_alloc_va_range;
ppgtt->base.teardown_va_range = gen6_teardown_va_range;
ppgtt->base.clear_range = gen6_ppgtt_clear_range;
@@ -1278,6 +1383,9 @@ static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
ppgtt->pd_addr = (gen6_gtt_pte_t __iomem *)dev_priv->gtt.gsm +
ppgtt->pd.pd_offset / sizeof(gen6_gtt_pte_t);
+ if (!aliasing)
+ gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total);
+
gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total);
DRM_DEBUG_DRIVER("Allocated pde space (%ldM) at GTT entry: %lx\n",
@@ -1290,7 +1398,8 @@ static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
return 0;
}
-static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
+static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt,
+ bool aliasing)
{
struct drm_i915_private *dev_priv = dev->dev_private;
@@ -1298,7 +1407,7 @@ static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
ppgtt->base.scratch = dev_priv->gtt.base.scratch;
if (INTEL_INFO(dev)->gen < 8)
- return gen6_ppgtt_init(ppgtt);
+ return gen6_ppgtt_init(ppgtt, aliasing);
else
return gen8_ppgtt_init(ppgtt, dev_priv->gtt.base.total);
}
@@ -1307,7 +1416,7 @@ int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt)
struct drm_i915_private *dev_priv = dev->dev_private;
int ret = 0;
- ret = __hw_ppgtt_init(dev, ppgtt);
+ ret = __hw_ppgtt_init(dev, ppgtt, false);
if (ret == 0) {
kref_init(&ppgtt->ref);
drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
@@ -1415,6 +1524,10 @@ static void ppgtt_unbind_vma(struct i915_vma *vma)
vma->node.start,
vma->obj->base.size,
true);
+ if (vma->vm->teardown_va_range) {
+ vma->vm->teardown_va_range(vma->vm,
+ vma->node.start, vma->node.size);
+ }
}
extern int intel_iommu_gfx_mapped;
@@ -1930,7 +2043,7 @@ static int i915_gem_setup_global_gtt(struct drm_device *dev,
if (!ppgtt)
return -ENOMEM;
- ret = __hw_ppgtt_init(dev, ppgtt);
+ ret = __hw_ppgtt_init(dev, ppgtt, true);
if (ret != 0)
return ret;
diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.h b/drivers/gpu/drm/i915/i915_gem_gtt.h
index 1ae70be..074b368 100644
--- a/drivers/gpu/drm/i915/i915_gem_gtt.h
+++ b/drivers/gpu/drm/i915/i915_gem_gtt.h
@@ -195,6 +195,7 @@ struct i915_page_table_entry {
dma_addr_t daddr;
unsigned long *used_ptes;
+ unsigned int scratch:1;
};
struct i915_page_directory_entry {
@@ -308,6 +309,8 @@ struct i915_hw_ppgtt {
struct i915_page_directory_entry pd;
};
+ struct i915_page_table_entry *scratch_pt;
+
struct drm_i915_file_private *file_priv;
gen6_gtt_pte_t __iomem *pd_addr;
--
2.1.1
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
