On Thu, Jul 07, 2022 at 06:11:13AM -0700, Hellstrom, Thomas wrote:
On Fri, 2022-07-01 at 15:50 -0700, Niranjana Vishwanathapura wrote:For persistent (vm_bind) vmas of userptr BOs, handle the user page pinning by using the i915_gem_object_userptr_submit_init() /done() functions Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@xxxxxxxxx> --- .../gpu/drm/i915/gem/i915_gem_execbuffer3.c | 67 +++++++++++++++++++ .../drm/i915/gem/i915_gem_vm_bind_object.c | 16 +++++ drivers/gpu/drm/i915/gt/intel_gtt.c | 1 + drivers/gpu/drm/i915/gt/intel_gtt.h | 1 + 4 files changed, 85 insertions(+) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer3.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer3.c index 2079f5ca9010..bf13dd6d642e 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer3.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer3.c @@ -22,6 +22,7 @@ #include "i915_gem_vm_bind.h" #include "i915_trace.h" +#define __EXEC3_USERPTR_USED BIT_ULL(34) #define __EXEC3_HAS_PIN BIT_ULL(33) #define __EXEC3_ENGINE_PINNED BIT_ULL(32) #define __EXEC3_INTERNAL_FLAGS (~0ull << 32) @@ -147,10 +148,36 @@ static void eb_scoop_unbound_vmas(struct i915_address_space *vm) spin_unlock(&vm->vm_rebind_lock); } +static int eb_lookup_persistent_userptr_vmas(struct i915_execbuffer *eb) +{ + struct i915_address_space *vm = eb->context->vm; + struct i915_vma *last_vma = NULL; + struct i915_vma *vma; + int err; + + assert_vm_bind_held(vm); + + list_for_each_entry(vma, &vm->vm_bind_list, vm_bind_link) { + if (i915_gem_object_is_userptr(vma->obj)) { + err = i915_gem_object_userptr_submit_init(vma->obj); + if (err) + return err; + + last_vma = vma; + } + } +Don't we need to loop also over non-private userptr objects?
No, as explained in other thread, non-private BOs will also be there in vm_bind/bound_list.
+ if (last_vma) + eb->args->flags |= __EXEC3_USERPTR_USED; + + return 0; +} + static int eb_lookup_vmas(struct i915_execbuffer *eb) { unsigned int i, current_batch = 0; struct i915_vma *vma; + int err = 0; for (i = 0; i < eb->num_batches; i++) { vma = eb_find_vma(eb->context->vm, eb- >batch_addresses[i]); @@ -163,6 +190,10 @@ static int eb_lookup_vmas(struct i915_execbuffer *eb) eb_scoop_unbound_vmas(eb->context->vm); + err = eb_lookup_persistent_userptr_vmas(eb); + if (err) + return err; + return 0; } @@ -358,15 +389,51 @@ static void eb_persistent_vmas_move_to_active(struct i915_execbuffer *eb) static int eb_move_to_gpu(struct i915_execbuffer *eb) { + int err = 0, j; + assert_vm_bind_held(eb->context->vm); assert_vm_priv_held(eb->context->vm); eb_persistent_vmas_move_to_active(eb); +#ifdef CONFIG_MMU_NOTIFIER + if (!err && (eb->args->flags & __EXEC3_USERPTR_USED)) { + struct i915_vma *vma; + + assert_vm_bind_held(eb->context->vm); + assert_vm_priv_held(eb->context->vm); + + read_lock(&eb->i915->mm.notifier_lock); + list_for_each_entry(vma, &eb->context->vm- >vm_bind_list, + vm_bind_link) { + if (!i915_gem_object_is_userptr(vma->obj)) + continue; + + err = i915_gem_object_userptr_submit_done(vma->obj); + if (err) + break; + } + + read_unlock(&eb->i915->mm.notifier_lock); + }Since we don't loop over the vm_bound_list, there is a need to check whether the rebind_list is empty here under the notifier_lock in read mode, and in that case, restart from eb_lookup_vmas(). That might also eliminate the need for the __EXEC3_USERPTR_USED flag? That will also catch any objects that were evicted between eb_lookup_vmas() where the rebind_list was last checked, and i915_gem_vm_priv_lock(), which prohibits further eviction, but if we want to catch these earlier (which I think is a good idea), we could check that the rebind_list is indeed empty just after taking the vm_priv_lock(), and if not, restart from eb_lookup_vmas().
Yah, right, we need to check rebind_list here and if not empty, restart from lookup phase. It is bit tricky with userptr here as the unbind happens during submit_init() call after we scoop unbound vmas here, the vmas gets re-added to rebind_list :(. I think we need a separate 'invalidated_userptr_list' here and we iterate through it for submit_init() and submit_done() calls (yes, __EXEC3_USERPTR_USED flag won't be needed then). And, we call, eb_scoop_unbound_vmas() after calling eb_lookup_persistent_userptr_vmas(), so that we scoop all unbound vmas properly.
+#endif + + if (unlikely(err)) + goto err_skip; + /* Unconditionally flush any chipset caches (for streaming writes). */ intel_gt_chipset_flush(eb->gt); return 0; + +err_skip: + for_each_batch_create_order(eb, j) { + if (!eb->requests[j]) + break; + + i915_request_set_error_once(eb->requests[j], err); + } + return err; } static int eb_request_submit(struct i915_execbuffer *eb, diff --git a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c index 1a8efa83547f..cae282b91618 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c @@ -263,6 +263,12 @@ int i915_gem_vm_bind_obj(struct i915_address_space *vm, goto put_obj; } + if (i915_gem_object_is_userptr(obj)) { + ret = i915_gem_object_userptr_submit_init(obj); + if (ret) + goto put_obj; + } + ret = i915_gem_vm_bind_lock_interruptible(vm); if (ret) goto put_obj; @@ -295,6 +301,16 @@ int i915_gem_vm_bind_obj(struct i915_address_space *vm, /* Make it evictable */ __i915_vma_unpin(vma); +#ifdef CONFIG_MMU_NOTIFIER + if (i915_gem_object_is_userptr(obj)) { + write_lock(&vm->i915->mm.notifier_lock);Why do we need the lock in write mode here?
Looks like it was no intentional. Should switch to read_lock here. Niranjana
+ ret = i915_gem_object_userptr_submit_done(obj); + write_unlock(&vm->i915->mm.notifier_lock); + if (ret) + goto out_ww; + } +#endif + list_add_tail(&vma->vm_bind_link, &vm->vm_bound_list); i915_vm_bind_it_insert(vma, &vm->va); if (!obj->priv_root) diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.c b/drivers/gpu/drm/i915/gt/intel_gtt.c index 55d5389b2c6c..4ab3bda644ff 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.c +++ b/drivers/gpu/drm/i915/gt/intel_gtt.c @@ -295,6 +295,7 @@ void i915_address_space_init(struct i915_address_space *vm, int subclass) GEM_BUG_ON(IS_ERR(vm->root_obj)); INIT_LIST_HEAD(&vm->vm_rebind_list); spin_lock_init(&vm->vm_rebind_lock); + INIT_LIST_HEAD(&vm->invalidate_link); } void *__px_vaddr(struct drm_i915_gem_object *p) diff --git a/drivers/gpu/drm/i915/gt/intel_gtt.h b/drivers/gpu/drm/i915/gt/intel_gtt.h index fe5485c4a1cd..f9edf11c144f 100644 --- a/drivers/gpu/drm/i915/gt/intel_gtt.h +++ b/drivers/gpu/drm/i915/gt/intel_gtt.h @@ -267,6 +267,7 @@ struct i915_address_space { struct list_head vm_bound_list; struct list_head vm_rebind_list; spinlock_t vm_rebind_lock; /* Protects vm_rebind_list */ + struct list_head invalidate_link; /* va tree of persistent vmas */ struct rb_root_cached va; struct list_head non_priv_vm_bind_list;
