]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915/gtt: Serialise both updates to PDE and our shadow
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 17 Jun 2019 14:04:26 +0000 (15:04 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Mon, 17 Jun 2019 19:53:03 +0000 (20:53 +0100)
Currently, we perform a locked update of the shadow entry when
allocating a page directory entry such that if two clients are
concurrently allocating neighbouring ranges we only insert one new entry
for the pair of them. However, we also need to serialise both clients
wrt to the actual entry in the HW table, or else we may allow one client
or even a third client to proceed ahead of the HW write. My handwave
before was that under the _pathological_ condition we would see the
scratch entry instead of the expected entry, causing a temporary
glitch. That starvation condition will eventually show up in practice, so
fix it.

The reason for the previous cheat was to avoid having to free the extra
allocation while under the spinlock. Now, we keep the extra entry
allocated until the end instead.

v2: Fix error paths for gen6

Fixes: 1d1b5490b91c ("drm/i915/gtt: Replace struct_mutex serialisation for allocation")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190617140426.7203-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_gtt.c

index ec00fccd0c6ff62b0fa079af61bff189f124c98c..8ab820145ea6cacd8b025ad8f8e2371088a91bc3 100644 (file)
@@ -1346,81 +1346,86 @@ static int gen8_ppgtt_alloc_pd(struct i915_address_space *vm,
                               struct i915_page_directory *pd,
                               u64 start, u64 length)
 {
-       struct i915_page_table *pt;
+       struct i915_page_table *pt, *alloc = NULL;
        u64 from = start;
        unsigned int pde;
+       int ret = 0;
 
        spin_lock(&pd->lock);
        gen8_for_each_pde(pt, pd, start, length, pde) {
                const int count = gen8_pte_count(start, length);
 
                if (pt == vm->scratch_pt) {
-                       struct i915_page_table *old;
-
                        spin_unlock(&pd->lock);
 
-                       pt = alloc_pt(vm);
-                       if (IS_ERR(pt))
+                       pt = fetch_and_zero(&alloc);
+                       if (!pt)
+                               pt = alloc_pt(vm);
+                       if (IS_ERR(pt)) {
+                               ret = PTR_ERR(pt);
                                goto unwind;
+                       }
 
                        if (count < GEN8_PTES || intel_vgpu_active(vm->i915))
                                gen8_initialize_pt(vm, pt);
 
-                       old = cmpxchg(&pd->entry[pde], vm->scratch_pt, pt);
-                       if (old == vm->scratch_pt) {
+                       spin_lock(&pd->lock);
+                       if (pd->entry[pde] == vm->scratch_pt) {
                                gen8_ppgtt_set_pde(vm, pd, pt, pde);
+                               pd->entry[pde] = pt;
                                atomic_inc(&pd->used);
                        } else {
-                               free_pt(vm, pt);
-                               pt = old;
+                               alloc = pt;
+                               pt = pd->entry[pde];
                        }
-
-                       spin_lock(&pd->lock);
                }
 
                atomic_add(count, &pt->used);
        }
        spin_unlock(&pd->lock);
-
-       return 0;
+       goto out;
 
 unwind:
        gen8_ppgtt_clear_pd(vm, pd, from, start - from);
-       return -ENOMEM;
+out:
+       if (alloc)
+               free_pt(vm, alloc);
+       return ret;
 }
 
 static int gen8_ppgtt_alloc_pdp(struct i915_address_space *vm,
                                struct i915_page_directory *pdp,
                                u64 start, u64 length)
 {
-       struct i915_page_directory *pd;
+       struct i915_page_directory *pd, *alloc = NULL;
        u64 from = start;
        unsigned int pdpe;
-       int ret;
+       int ret = 0;
 
        spin_lock(&pdp->lock);
        gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
                if (pd == vm->scratch_pd) {
-                       struct i915_page_directory *old;
-
                        spin_unlock(&pdp->lock);
 
-                       pd = alloc_pd(vm);
-                       if (IS_ERR(pd))
+                       pd = fetch_and_zero(&alloc);
+                       if (!pd)
+                               pd = alloc_pd(vm);
+                       if (IS_ERR(pd)) {
+                               ret = PTR_ERR(pd);
                                goto unwind;
+                       }
 
                        init_pd_with_page(vm, pd, vm->scratch_pt);
 
-                       old = cmpxchg(&pdp->entry[pdpe], vm->scratch_pd, pd);
-                       if (old == vm->scratch_pd) {
+                       spin_lock(&pdp->lock);
+                       if (pdp->entry[pdpe] == vm->scratch_pd) {
                                gen8_ppgtt_set_pdpe(pdp, pd, pdpe);
+                               pdp->entry[pdpe] = pd;
                                atomic_inc(&pdp->used);
                        } else {
-                               free_pd(vm, pd);
-                               pd = old;
+                               alloc = pd;
+                               pd = pdp->entry[pdpe];
                        }
-
-                       spin_lock(&pdp->lock);
                }
                atomic_inc(&pd->used);
                spin_unlock(&pdp->lock);
@@ -1433,8 +1438,7 @@ static int gen8_ppgtt_alloc_pdp(struct i915_address_space *vm,
                atomic_dec(&pd->used);
        }
        spin_unlock(&pdp->lock);
-
-       return 0;
+       goto out;
 
 unwind_pd:
        spin_lock(&pdp->lock);
@@ -1447,7 +1451,10 @@ unwind_pd:
        spin_unlock(&pdp->lock);
 unwind:
        gen8_ppgtt_clear_pdp(vm, pdp, from, start - from);
-       return -ENOMEM;
+out:
+       if (alloc)
+               free_pd(vm, alloc);
+       return ret;
 }
 
 static int gen8_ppgtt_alloc_3lvl(struct i915_address_space *vm,
@@ -1462,33 +1469,34 @@ static int gen8_ppgtt_alloc_4lvl(struct i915_address_space *vm,
 {
        struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
        struct i915_page_directory * const pml4 = ppgtt->pd;
-       struct i915_page_directory *pdp;
+       struct i915_page_directory *pdp, *alloc = NULL;
        u64 from = start;
+       int ret = 0;
        u32 pml4e;
-       int ret;
 
        spin_lock(&pml4->lock);
        gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
                if (pdp == vm->scratch_pdp) {
-                       struct i915_page_directory *old;
-
                        spin_unlock(&pml4->lock);
 
-                       pdp = alloc_pd(vm);
-                       if (IS_ERR(pdp))
+                       pdp = fetch_and_zero(&alloc);
+                       if (!pdp)
+                               pdp = alloc_pd(vm);
+                       if (IS_ERR(pdp)) {
+                               ret = PTR_ERR(pdp);
                                goto unwind;
+                       }
 
                        init_pd(vm, pdp, vm->scratch_pd);
 
-                       old = cmpxchg(&pml4->entry[pml4e], vm->scratch_pdp, pdp);
-                       if (old == vm->scratch_pdp) {
+                       spin_lock(&pml4->lock);
+                       if (pml4->entry[pml4e] == vm->scratch_pdp) {
                                gen8_ppgtt_set_pml4e(pml4, pdp, pml4e);
+                               pml4->entry[pml4e] = pdp;
                        } else {
-                               free_pd(vm, pdp);
-                               pdp = old;
+                               alloc = pdp;
+                               pdp = pml4->entry[pml4e];
                        }
-
-                       spin_lock(&pml4->lock);
                }
                atomic_inc(&pdp->used);
                spin_unlock(&pml4->lock);
@@ -1501,8 +1509,7 @@ static int gen8_ppgtt_alloc_4lvl(struct i915_address_space *vm,
                atomic_dec(&pdp->used);
        }
        spin_unlock(&pml4->lock);
-
-       return 0;
+       goto out;
 
 unwind_pdp:
        spin_lock(&pml4->lock);
@@ -1513,7 +1520,10 @@ unwind_pdp:
        spin_unlock(&pml4->lock);
 unwind:
        gen8_ppgtt_clear_4lvl(vm, from, start - from);
-       return -ENOMEM;
+out:
+       if (alloc)
+               free_pd(vm, alloc);
+       return ret;
 }
 
 static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt)
@@ -1792,11 +1802,12 @@ static int gen6_alloc_va_range(struct i915_address_space *vm,
 {
        struct gen6_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
        struct i915_page_directory * const pd = ppgtt->base.pd;
-       struct i915_page_table *pt;
+       struct i915_page_table *pt, *alloc = NULL;
        intel_wakeref_t wakeref;
        u64 from = start;
        unsigned int pde;
        bool flush = false;
+       int ret = 0;
 
        wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
 
@@ -1805,29 +1816,30 @@ static int gen6_alloc_va_range(struct i915_address_space *vm,
                const unsigned int count = gen6_pte_count(start, length);
 
                if (pt == vm->scratch_pt) {
-                       struct i915_page_table *old;
-
                        spin_unlock(&pd->lock);
 
-                       pt = alloc_pt(vm);
-                       if (IS_ERR(pt))
+                       pt = fetch_and_zero(&alloc);
+                       if (!pt)
+                               pt = alloc_pt(vm);
+                       if (IS_ERR(pt)) {
+                               ret = PTR_ERR(pt);
                                goto unwind_out;
+                       }
 
                        gen6_initialize_pt(vm, pt);
 
-                       old = cmpxchg(&pd->entry[pde], vm->scratch_pt, pt);
-                       if (old == vm->scratch_pt) {
+                       spin_lock(&pd->lock);
+                       if (pd->entry[pde] == vm->scratch_pt) {
+                               pd->entry[pde] = pt;
                                if (i915_vma_is_bound(ppgtt->vma,
                                                      I915_VMA_GLOBAL_BIND)) {
                                        gen6_write_pde(ppgtt, pde, pt);
                                        flush = true;
                                }
                        } else {
-                               free_pt(vm, pt);
-                               pt = old;
+                               alloc = pt;
+                               pt = pd->entry[pde];
                        }
-
-                       spin_lock(&pd->lock);
                }
 
                atomic_add(count, &pt->used);
@@ -1839,14 +1851,15 @@ static int gen6_alloc_va_range(struct i915_address_space *vm,
                gen6_ggtt_invalidate(vm->i915);
        }
 
-       intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
-
-       return 0;
+       goto out;
 
 unwind_out:
-       intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
        gen6_ppgtt_clear_range(vm, from, start - from);
-       return -ENOMEM;
+out:
+       if (alloc)
+               free_pt(vm, alloc);
+       intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
+       return ret;
 }
 
 static int gen6_ppgtt_init_scratch(struct gen6_ppgtt *ppgtt)