drm/i915: Synchronize pread/pwrite with wait_rendering
Subsystems (2)
PATCH 1/8 drm/i915: Synchronize pread/pwrite with wait_rendering
lifted from Daniel:
pread/pwrite isn't about the object's domain at all, but purely about synchronizing for outstanding rendering. Replacing the call to set_to_gtt_domain with a wait_rendering would imo improve code readability. Furthermore we could pimp pread to only block for outstanding writes and not for reads.
Since you're not the first one to trip over this: Can I volunteer you for a follow-up patch to fix this?
---
Notes:
v2: Switch the pwrite patch to use \!read_only. This was a typo in the original code. (Chris, Daniel)
drivers/gpu/drm/i915/i915_gem.c | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +9 -10
@@ -41,6 +41,9 @@ static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *ostatic void i915_gem_object_flush_cpu_write_domain (struct drm_i915_gem_object *obj,bool force);static __must_check int +i915_gem_object_wait_rendering (struct drm_i915_gem_object *obj, +bool readonly); +static __must_check int i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj, struct i915_address_space *vm, unsigned alignment,
@@ -430,11 +433,9 @@ i915_gem_shmem_pread(struct drm_device *dev,* optimizes for the case when the gpu will dirty the data * anyway again before the next pread happens. */ needs_clflush = !cpu_cache_is_coherent (dev, obj->cache_level); -if (i915_gem_obj_bound_any (obj)) { - ret =i915_gem_object_set_to_gtt_domain (obj, false); -if (ret) -return ret; - } + ret =i915_gem_object_wait_rendering (obj, false); +if (ret) +return ret; } ret =i915_gem_object_get_pages (obj);
@@ -746,11 +747,9 @@ i915_gem_shmem_pwrite(struct drm_device *dev,* optimizes for the case when the gpu will use the data * right away and we therefore have to clflush anyway. */ needs_clflush_after =cpu_write_needs_clflush (obj); -if (i915_gem_obj_bound_any (obj)) { - ret =i915_gem_object_set_to_gtt_domain (obj, true); -if (ret) -return ret; - } + ret =i915_gem_object_wait_rendering (obj, true); +if (ret) +return ret; }/* Same trick applies to invalidate partially written cachelines read * before writing. */
Signature · 1 line
1.8.4
PATCH 2/8 drm/i915: Extract vm specific part of eviction
As we'll see in the next patch, being able to evict for just 1 VM is handy.
--- drivers/gpu/drm/i915/i915_gem_evict.c | 28 ++++++++++++++++++++++------ 1 file changed, 22 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/i915_gem_evict.c +22 -6
@@ -155,12 +155,31 @@ found:return ret; } +static int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle) +{ +struct i915_vma *vma, *next; +int ret; + +if (do_idle) { + ret =i915_gpu_idle (vm->dev); +if (ret) +return ret; + +i915_gem_retire_requests (vm->dev); + } + +list_for_each_entry_safe (vma, next, &vm->inactive_list, mm_list) +if (vma->obj->pin_count ==0 ) +WARN_ON (i915_vma_unbind (vma)); + +return 0 ; +} +int i915_gem_evict_everything (struct drm_device *dev) {drm_i915_private_t *dev_priv = dev->dev_private;struct i915_address_space *vm; -struct i915_vma *vma, *next;bool lists_empty = true;int ret;
@@ -187,11 +206,8 @@ i915_gem_evict_everything(struct drm_device *dev)i915_gem_retire_requests (dev);/* Having flushed everything, unbind() should never raise an error */ -list_for_each_entry (vm, &dev_priv->vm_list, global_link) { -list_for_each_entry_safe (vma, next, &vm->inactive_list, mm_list) -if (vma->obj->pin_count ==0 ) -WARN_ON (i915_vma_unbind (vma)); - } +list_for_each_entry (vm, &dev_priv->vm_list, global_link) +WARN_ON (i915_gem_evict_vm (vm, false));return 0 ; }
Signature · 1 line
1.8.4
PATCH 3/8 drm/i915: evict VM instead of everything
In reply to Ben WidawskyWhen reserving objects during execbuf, it is possible to come across an object which will not fit given the current fragmentation of the address space. We do not have any defragment in drm_mm, so the strategy is to instead evict everything, and reallocate objects.
With the upcoming addition of multiple VMs, there is no point to evict everything since doing so is overkill for the specific case mentioned above.
--- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/i915_gem_evict.c | 17 ++++++++++++++++- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 8 +++++++- 3 files changed, 24 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +1 -0
@@ -2106,6 +2106,7 @@ int __must_check i915_gem_evict_something(struct drm_device *dev,unsigned cache_level, bool mappable, bool nonblock); +int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle);int i915_gem_evict_everything (struct drm_device *dev);/* i915_gem_stolen.c */
drivers/gpu/drm/i915/i915_gem_evict.c +16 -1
@@ -155,7 +155,22 @@ found:return ret; } -static int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle) +/** +* i915_gem_evict_vm - Try to free up VM space +* +* @vm: Address space to evict from +* @do_idle: Boolean directing whether to idle first. +* +* VM eviction is about freeing up virtual address space. If one wants fine +* grained eviction, they should see evict something for more details. In terms +* of freeing up actual system memory, this function may not accomplish the +* desired result. An object may be shared in multiple address space, and this +* function will not assert those objects be freed. +* +* Using do_idle will result in a more complete eviction because it retires, and +* inactivates current BOs. +*/ +int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle) {struct i915_vma *vma, *next;int ret;
drivers/gpu/drm/i915/i915_gem_execbuffer.c +7 -1
@@ -549,10 +549,16 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring, {struct drm_i915_gem_object *obj;struct i915_vma *vma; +struct i915_address_space *vm;struct list_head ordered_vmas;bool has_fenced_gpu_access =INTEL_INFO (ring->dev)->gen <4 ;int retry; +if (list_empty (vmas)) +return 0 ; + + vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm; +INIT_LIST_HEAD (&ordered_vmas);while (!list_empty (vmas)) {struct drm_i915_gem_exec_object2 *entry;
@@ -641,7 +647,7 @@ err: /* Decrement pin count for bound objects */if (ret != -ENOSPC || retry++)return ret; - ret =i915_gem_evict_everything (ring->dev); + ret =i915_gem_evict_vm (vm, true);if (ret)return ret; }while (1 );
Signature · 1 line
1.8.4
Re: [PATCH 3/8] drm/i915: evict VM instead of everything
Quoted 11 lines · Ben Widawsky
When reserving objects during execbuf, it is possible to come across an object which will not fit given the current fragmentation of the address space. We do not have any defragment in
drm_mm, so the strategy is to instead evict everything, and reallocate objects.With the upcoming addition of multiple VMs, there is no point to evict everything since doing so is overkill for the specific case mentioned above.
Recommended-by Daniel Vetter <[email protected]>Signed-off-by Ben Widawsky <[email protected]>
Merged the first three patches (with a tiny fixup for this one since you've forgotten to update one comment that needs to be updated). Leaves us with a short-time lack of tracepoint for evict_vm but Chris has imo a good point.
-Daniel
Quoted 83 lines
--- drivers/gpu/drm/i915/i915_drv.h | 1 + drivers/gpu/drm/i915/i915_gem_evict.c | 17 ++++++++++++++++- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 8 +++++++- 3 files changed, 24 insertions(+), 2 deletions(-)drivers/gpu/drm/i915/i915_drv.h +1
-0@@ -2106,6 +2106,7 @@ int __must_check i915_gem_evict_something(struct drm_device *dev,unsigned cache_level, bool mappable, bool nonblock); +int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle);int i915_gem_evict_everything (struct drm_device *dev);/* i915_gem_stolen.c */ drivers/gpu/drm/i915/i915_gem_evict.c +16
-1@@ -155,7 +155,22 @@ found:return ret; } -static int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle) +/** +* i915_gem_evict_vm - Try to free up VM space +* +* @vm: Address space to evict from +* @do_idle: Boolean directing whether to idle first. +* +* VM eviction is about freeing up virtual address space. If one wants fine +* grained eviction, they should see evict something for more details. In terms +* of freeing up actual system memory, this function may not accomplish the +* desired result. An object may be shared in multiple address space, and this +* function will not assert those objects be freed. +* +* Using do_idle will result in a more complete eviction because it retires, and +* inactivates current BOs. +*/ +int i915_gem_evict_vm (struct i915_address_space *vm,bool do_idle) {struct i915_vma *vma, *next;int ret;drivers/gpu/drm/i915/i915_gem_execbuffer.c +7
-1@@ -549,10 +549,16 @@ i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring, {struct drm_i915_gem_object *obj;struct i915_vma *vma; +struct i915_address_space *vm;struct list_head ordered_vmas;bool has_fenced_gpu_access =INTEL_INFO (ring->dev)->gen <4 ;int retry; +if (list_empty (vmas)) +return 0 ; + + vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm; +INIT_LIST_HEAD (&ordered_vmas);while (!list_empty (vmas)) {struct drm_i915_gem_exec_object2 *entry;@@ -641,7 +647,7 @@ err: /* Decrement pin count for bound objects */if (ret != -ENOSPC || retry++)return ret; - ret =i915_gem_evict_everything (ring->dev); + ret =i915_gem_evict_vm (vm, true);if (ret)return ret; }while (1 );Signature · 6 lines
1.8.4 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
PATCH 4/8 drm/i915: trace vm eviction instead of everything
In reply to Ben WidawskyTracing vm eviction is really the event we care about. For the cases we evict everything, we still will get the trace.
--- drivers/gpu/drm/i915/i915_gem_evict.c | 2 ++ drivers/gpu/drm/i915/i915_trace.h | 15 +++++++++++++++ 2 files changed, 17 insertions(+)
drivers/gpu/drm/i915/i915_gem_evict.c +2 -0
@@ -175,6 +175,8 @@ int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle)struct i915_vma *vma, *next;int ret; +trace_i915_gem_evict_vm (vm); +if (do_idle) { ret =i915_gpu_idle (vm->dev);if (ret)
drivers/gpu/drm/i915/i915_trace.h +15 -0
@@ -233,6 +233,21 @@ TRACE_EVENT(i915_gem_evict_everything,TP_printk ("dev=%d" , __entry->dev) ); +TRACE_EVENT (i915_gem_evict_vm, + TP_PROTO(struct i915_address_space *vm), +TP_ARGS (vm), + +TP_STRUCT__entry ( + __field(struct i915_address_space *, vm) + ), + +TP_fast_assign ( + __entry->vm = vm; + ), + +TP_printk ("vm=%p" , __entry->vm) +); + TRACE_EVENT(i915_gem_ring_dispatch, TP_PROTO(struct intel_ring_buffer *ring, u32 seqno, u32 flags), TP_ARGS(ring, seqno, flags),
Signature · 1 line
1.8.4
Re: [PATCH 4/8] drm/i915: trace vm eviction instead of everything
Quoted 34 lines · Ben Widawsky
Tracing vm eviction is really the event we care about. For the cases we evict everything, we still will get the trace.
Signed-off-by Ben Widawsky <[email protected]>--- drivers/gpu/drm/i915/i915_gem_evict.c | 2 ++ drivers/gpu/drm/i915/i915_trace.h | 15 +++++++++++++++ 2 files changed, 17 insertions(+)drivers/gpu/drm/i915/i915_gem_evict.c +2
-0@@ -175,6 +175,8 @@ int i915_gem_evict_vm(struct i915_address_space *vm, bool do_idle)struct i915_vma *vma, *next;int ret; +trace_i915_gem_evict_vm (vm); +if (do_idle) { ret =i915_gpu_idle (vm->dev);if (ret)drivers/gpu/drm/i915/i915_trace.h +4
-0@@ -233,6 +233,21 @@ TRACE_EVENT(i915_gem_evict_everything,TP_printk ("dev=%d" , __entry->dev) ); +TRACE_EVENT(i915_gem_evict_vm, + TP_PROTO(struct i915_address_space *vm), + TP_ARGS(vm), +
drivers/gpu/drm/i915/i915_trace.h TRACE_EVENT(i915_gem_evict_everything, · line 249
Quoted 10 lines
drivers/gpu/drm/i915/i915_trace.h +10
-0+TP_STRUCT__entry ( +__field (struct i915_address_space *,vm ) + ), + +TP_fast_assign ( + __entry->vm = vm; + ), + +TP_printk ("vm=%p" , __entry->vm) +);
Jesse drilled into me that we must never assume that we are the solitary device on the system. Do we want to add dev here or remove all the others?
-Chris
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
PATCH 5/8 drm/i915: Convert active API to VMA
In reply to Ben WidawskyFrom: Ben Widawsky <[email protected]>
Even though we track object activity and not VMA, because we have the active_list be based on the VM, it makes the most sense to use VMAs in the APIs.
NOTE: Daniel intends to eventually rip out active/inactive LRUs, but for now, leave them be.
v2: Remove leftover hunk from the previous patch which didn't keep i915_gem_object_move_to_active. That patch had to rely on the ring to get the dev instead of the obj. (Chris)
--- drivers/gpu/drm/i915/i915_drv.h | 5 ++--- drivers/gpu/drm/i915/i915_gem.c | 9 ++++++++- drivers/gpu/drm/i915/i915_gem_context.c | 8 ++++---- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 3 +-- 4 files changed, 15 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +2 -3
@@ -1884,9 +1884,8 @@ static inline void i915_gem_object_unpin_pages(struct drm_i915_gem_object *obj)int __must_check i915_mutex_lock_interruptible (struct drm_device *dev);int i915_gem_object_sync (struct drm_i915_gem_object *obj,struct intel_ring_buffer *to); -void i915_gem_object_move_to_active (struct drm_i915_gem_object *obj, -struct intel_ring_buffer *ring); - +void i915_vma_move_to_active (struct i915_vma *vma, +struct intel_ring_buffer *ring);int i915_gem_dumb_create (struct drm_file *file_priv,struct drm_device *dev,struct drm_mode_create_dumb *args);
drivers/gpu/drm/i915/i915_gem.c +8 -1
@@ -1913,7 +1913,7 @@ i915_gem_object_get_pages(struct drm_i915_gem_object *obj)return 0 ; } -void +static void i915_gem_object_move_to_active (struct drm_i915_gem_object *obj,struct intel_ring_buffer *ring) {
@@ -1952,6 +1952,13 @@ i915_gem_object_move_to_active(struct drm_i915_gem_object *obj, } } +void i915_vma_move_to_active (struct i915_vma *vma, +struct intel_ring_buffer *ring) +{ +list_move_tail (&vma->mm_list, &vma->vm->active_list); +return i915_gem_object_move_to_active (vma->obj, ring); +} +static void i915_gem_object_move_to_inactive (struct drm_i915_gem_object *obj) {
drivers/gpu/drm/i915/i915_gem_context.c +4 -4
@@ -436,11 +436,11 @@ static int do_switch(struct i915_hw_context *to)* MI_SET_CONTEXT instead of when the next seqno has completed. */ if (from !=NULL ) { -struct drm_i915_private *dev_priv = from->obj->base.dev->dev_private; -struct i915_address_space *ggtt = &dev_priv->gtt.base; +struct drm_i915_private *dev_priv = ring->dev->dev_private; +struct i915_vma *vma = +i915_gem_obj_to_vma (from->obj, &dev_priv->gtt.base); from->obj->base.read_domains =I915_GEM_DOMAIN_INSTRUCTION ; -list_move_tail (&i915_gem_obj_to_vma (from->obj, ggtt)->mm_list, &ggtt->active_list); -i915_gem_object_move_to_active (from->obj, ring); +i915_vma_move_to_active (vma, ring);/* As long as MI_SET_CONTEXT is serializing, ie. it flushes the * whole damn pipeline, we don't need to explicitly mark the * object dirty. The only exception is that the context must be
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1 -2
@@ -872,8 +872,7 @@ i915_gem_execbuffer_move_to_active(struct list_head *vmas, obj->base.read_domains = obj->base.pending_read_domains; obj->fenced_gpu_access = obj->pending_fenced_gpu_access; -list_move_tail (&vma->mm_list, &vma->vm->active_list); -i915_gem_object_move_to_active (obj, ring); +i915_vma_move_to_active (vma, ring);if (obj->base.write_domain) { obj->dirty =1 ; obj->last_write_seqno =intel_ring_get_seqno (ring);
Signature · 1 line
1.8.4
Re: [PATCH 5/8] drm/i915: Convert active API to VMA
drivers/gpu/drm/i915/i915_gem_context.c static int do_switch(struct i915_hw_context *to) · line 443
Quoted 10 lines · Ben Widawsky
drivers/gpu/drm/i915/i915_gem_context.c +4
-4if (from !=NULL ) { -struct drm_i915_private *dev_priv = from->obj->base.dev->dev_private; -struct i915_address_space *ggtt = &dev_priv->gtt.base; +struct drm_i915_private *dev_priv = ring->dev->dev_private; +struct i915_vma *vma = +i915_gem_obj_to_vma (from->obj, &dev_priv->gtt.base); from->obj->base.read_domains =I915_GEM_DOMAIN_INSTRUCTION ; -list_move_tail (&i915_gem_obj_to_vma (from->obj, ggtt)->mm_list, &ggtt->active_list); -i915_gem_object_move_to_active (from->obj, ring); +i915_vma_move_to_active (vma, ring);
We do fairly often lookup the ggtt vma for an obj. Is this worth a speciallised function call?
i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->obj), ring);
struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
{
struct i915_vma *vma;
if (WARN_ON(list_empty(&obj->vma_list))
return NULL;
vma = list_first_entry(&obj->vma_list, typeof(*vma), vma_link);
if (WARN_ON(vma != &to_i915(obj->base.dev)->gtt.base))
return NULL;
return vma;
}-Chris
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
PATCH 6/8 drm/i915: Add bind/unbind object functions to VM
In reply to Ben WidawskyFrom: Ben Widawsky <[email protected]>
As we plumb the code with more VM information, it has become more obvious that the easiest way to deal with bind and unbind is to simply put the function pointers in the vm, and let those choose the correct way to handle the page table updates. This change allows many places in the code to simply be vm->bind, and not have to worry about distinguishing PPGTT vs GGTT.
Notice that this patch has no impact on functionality. I've decided to save the actual change until the next patch because I think it's easier to review that way. I'm happy to squash the two, or let Daniel do it on merge.
---
Notes:
v2:
Make ggtt handle the quirky aliasing ppgtt
Add flags to bind object to support above
Don't ever call bind/unbind directly for PPGTT until we have real, full PPGTT (use NULLs to assert this)
Make sure we rebind the ggtt if there already is a ggtt binding. This happens on set cache levels
Use VMA for bind/unbind (Daniel, Ben)
v3: Reorganize ggtt_vma_bind to be more concise and easier to read (Ville). Change logic in unbind to only unbind ggtt when there is a global mapping, and to remove a redundant check if the aliasing ppgtt exists.
drivers/gpu/drm/i915/i915_drv.h | 69 +++++++++++++----------- drivers/gpu/drm/i915/i915_gem_gtt.c | 102 ++++++++++++++++++++++++++++++++++++ 2 files changed, 141 insertions(+), 30 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +39 -30
@@ -465,6 +465,36 @@ enum i915_cache_level {typedef uint32_t gen6_gtt_pte_t ; +/** +* A VMA represents a GEM BO that is bound into an address space. Therefore, a +* VMA's presence cannot be guaranteed before binding, or after unbinding the +* object into/from the address space. +* +* To make things as simple as possible (ie. no refcounting), a VMA's lifetime +* will always be <= an objects lifetime. So object refcounting should cover us. +*/ +struct i915_vma { +struct drm_mm_node node; +struct drm_i915_gem_object *obj; +struct i915_address_space *vm; + +/** This object's place on the active/inactive lists */ +struct list_head mm_list; + +struct list_head vma_link;/* Link in the object's VMA list */ + +/** This vma's place in the batchbuffer or on the eviction list */ +struct list_head exec_list; + +/** +* Used for performing relocations during execbuffer insertion. +*/ +struct hlist_node exec_node; +unsigned long exec_handle; +struct drm_i915_gem_exec_object2 *exec_entry; + +}; +struct i915_address_space {struct drm_mm mm;struct drm_device *dev;
@@ -503,9 +533,18 @@ struct i915_address_space {/* FIXME: Need a more generic return type */ gen6_gtt_pte_t (*pte_encode)(dma_addr_t addr,enum i915_cache_level level); + +/** Unmap an object from an address space. This usually consists of +* setting the valid PTE entries to a reserved scratch page. */ +void (*unbind_vma)(struct i915_vma *vma);void (*clear_range)(struct i915_address_space *vm,unsigned int first_entry,unsigned int num_entries); +/* Map an object into an address space with the given cache flags. */ +#define GLOBAL_BIND (1<<0) +void (*bind_vma)(struct i915_vma *vma, +enum i915_cache_level cache_level, +u32 flags); void (*insert_entries)(struct i915_address_space *vm, struct sg_table *st, unsigned int first_entry,
@@ -552,36 +591,6 @@ struct i915_hw_ppgtt {int (*enable)(struct drm_device *dev); }; -/** -* A VMA represents a GEM BO that is bound into an address space. Therefore, a -* VMA's presence cannot be guaranteed before binding, or after unbinding the -* object into/from the address space. -* -* To make things as simple as possible (ie. no refcounting), a VMA's lifetime -* will always be <= an objects lifetime. So object refcounting should cover us. -*/ -struct i915_vma { -struct drm_mm_node node; -struct drm_i915_gem_object *obj; -struct i915_address_space *vm; - -/** This object's place on the active/inactive lists */ -struct list_head mm_list; - -struct list_head vma_link;/* Link in the object's VMA list */ - -/** This vma's place in the batchbuffer or on the eviction list */ -struct list_head exec_list; - -/** -* Used for performing relocations during execbuffer insertion. -*/ -struct hlist_node exec_node; -unsigned long exec_handle; -struct drm_i915_gem_exec_object2 *exec_entry; - -}; -struct i915_ctx_hang_stats {/* This context had batch pending when hang was declared */ unsigned batch_pending;
drivers/gpu/drm/i915/i915_gem_gtt.c +102 -0
159 lines not on this page · raw message
Signature · 1 line
1.8.4
Re: [PATCH 6/8] drm/i915: Add bind/unbind object functions to VM
Quoted 9 lines · Ben Widawsky
From: Ben Widawsky <[email protected]>
As we plumb the code with more VM information, it has become more obvious that the easiest way to deal with bind and unbind is to simply put the function pointers in the vm, and let those choose the correct way to handle the page table updates. This change allows many places in the code to simply be
vm->bind, and not have to worry about distinguishing PPGTT vs GGTT.
Quoted 4 lines
Notice that this patch has no impact on functionality. I've decided to save the actual change until the next patch because I think it's easier to review that way. I'm happy to squash the two, or let Daniel do it on merge.
Oh, this is disappointing. The GLOBAL_BIND flag is a let down. Is there any way we can have the aliasing_ppgtt as a separate vm and so keep the purity?
-Chris
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
Re: [PATCH 6/8] drm/i915: Add bind/unbind object functions to VM
Quoted 10 lines · Chris Wilson
Quoted 9 lines · Ben Widawsky
From: Ben Widawsky <[email protected]>
As we plumb the code with more VM information, it has become more obvious that the easiest way to deal with bind and unbind is to simply put the function pointers in the vm, and let those choose the correct way to handle the page table updates. This change allows many places in the code to simply be
vm->bind, and not have to worry about distinguishing PPGTT vs GGTT.
Quoted 8 lines
Quoted 4 lines
Notice that this patch has no impact on functionality. I've decided to save the actual change until the next patch because I think it's easier to review that way. I'm happy to squash the two, or let Daniel do it on merge.
Oh, this is disappointing. The
GLOBAL_BINDflag is a let down. Is there any way we can have thealiasing_ppgttas a separate vm and so keep the purity?
The problem is that the vm manages the address space, and it's paramount on snb that we can line up objects into both the ggtt and ppgtt at will at the exact same offset. Otherwise at least pipe control w/a won't work and reading between bspec gives more promises of things that will break.
Adding complex logic to punt out such holes doesn't sound cleaner nor simpler than the magic flags we're using here. And I have to admit that I couldn't come up with another idea (besides flat-out killing aliasing ppgtt on snb). So ideas welcome ...
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation +41 (0) 79 365 57 48 - http://blog.ffwll.ch
PATCH 7/8 drm/i915: Use the new vm [un]bind functions
In reply to Ben WidawskyFrom: Ben Widawsky <[email protected]>
Building on the last patch which created the new function pointers in the VM for bind/unbind, here we actually put those new function pointers to use.
Split out as a separate patch to aid in review. I'm fine with squashing into the previous patch if people request it.
v2: Updated to address the smart ggtt which can do aliasing as needed Make sure we bind to global gtt when mappable and fenceable. I thought we could get away without this initialy, but we cannot.
--- drivers/gpu/drm/i915/i915_drv.h | 9 ------ drivers/gpu/drm/i915/i915_gem.c | 31 ++++++++----------- drivers/gpu/drm/i915/i915_gem_context.c | 8 +++-- drivers/gpu/drm/i915/i915_gem_execbuffer.c | 29 ++++++++---------- drivers/gpu/drm/i915/i915_gem_gtt.c | 48 ++---------------------------- 5 files changed, 34 insertions(+), 91 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +0 -9
@@ -2083,17 +2083,8 @@ int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,/* i915_gem_gtt.c */ void i915_gem_cleanup_aliasing_ppgtt (struct drm_device *dev); -void i915_ppgtt_bind_object (struct i915_hw_ppgtt *ppgtt, -struct drm_i915_gem_object *obj, -enum i915_cache_level cache_level); -void i915_ppgtt_unbind_object (struct i915_hw_ppgtt *ppgtt, -struct drm_i915_gem_object *obj); -void i915_gem_restore_gtt_mappings (struct drm_device *dev);int __must_check i915_gem_gtt_prepare_object (struct drm_i915_gem_object *obj); -void i915_gem_gtt_bind_object (struct drm_i915_gem_object *obj, -enum i915_cache_level cache_level); -void i915_gem_gtt_unbind_object (struct drm_i915_gem_object *obj);void i915_gem_gtt_finish_object (struct drm_i915_gem_object *obj);void i915_gem_init_global_gtt (struct drm_device *dev); void i915_gem_setup_global_gtt(struct drm_device *dev, unsigned long start,
drivers/gpu/drm/i915/i915_gem.c +13 -18
@@ -2696,12 +2696,8 @@ int i915_vma_unbind(struct i915_vma *vma)trace_i915_vma_unbind (vma); -if (obj->has_global_gtt_mapping) -i915_gem_gtt_unbind_object (obj); -if (obj->has_aliasing_ppgtt_mapping) { -i915_ppgtt_unbind_object (dev_priv->mm.aliasing_ppgtt, obj); - obj->has_aliasing_ppgtt_mapping =0 ; - } + vma->vm->unbind_vma (vma); +i915_gem_gtt_finish_object (obj);i915_gem_object_unpin_pages (obj);
@@ -3428,7 +3424,6 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj, enum i915_cache_level cache_level) {struct drm_device *dev = obj->base.dev; -drm_i915_private_t *dev_priv = dev->dev_private;struct i915_vma *vma;int ret;
@@ -3467,11 +3462,8 @@ int i915_gem_object_set_cache_level(struct drm_i915_gem_object *obj,return ret; } -if (obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (obj, cache_level); -if (obj->has_aliasing_ppgtt_mapping) -i915_ppgtt_bind_object (dev_priv->mm.aliasing_ppgtt, - obj, cache_level); +list_for_each_entry (vma, &obj->vma_list, vma_link) + vma->vm->bind_vma (vma, cache_level,0 ); }list_for_each_entry (vma, &obj->vma_list, vma_link)
@@ -3799,6 +3791,7 @@ i915_gem_object_pin(struct drm_i915_gem_object *obj, bool map_and_fenceable, bool nonblocking) { +const u32 flags = map_and_fenceable ?GLOBAL_BIND :0 ;struct i915_vma *vma;int ret;
@@ -3827,20 +3820,22 @@ i915_gem_object_pin(struct drm_i915_gem_object *obj, }if (!i915_gem_obj_bound (obj, vm)) { -struct drm_i915_private *dev_priv = obj->base.dev->dev_private; - ret =i915_gem_object_bind_to_vm (obj, vm, alignment, map_and_fenceable, nonblocking);if (ret)return ret; -if (!dev_priv->mm.aliasing_ppgtt) -i915_gem_gtt_bind_object (obj, obj->cache_level); - } + vma =i915_gem_obj_to_vma (obj, vm); + vm->bind_vma (vma, obj->cache_level, flags); + }else + vma =i915_gem_obj_to_vma (obj, vm); +/* Objects are created map and fenceable. If we bind an object +* the first time, and we had aliasing PPGTT (and didn't request +* GLOBAL), we'll need to do this on the second bind.*/ if (!obj->has_global_gtt_mapping && map_and_fenceable) -i915_gem_gtt_bind_object (obj, obj->cache_level); + vm->bind_vma (vma, obj->cache_level,GLOBAL_BIND ); obj->pin_count++; obj->pin_mappable |= map_and_fenceable;
drivers/gpu/drm/i915/i915_gem_context.c +6 -2
@@ -391,6 +391,7 @@ mi_set_context(struct intel_ring_buffer *ring,static int do_switch (struct i915_hw_context *to) {struct intel_ring_buffer *ring = to->ring; +struct drm_i915_private *dev_priv = ring->dev->dev_private;struct i915_hw_context *from = ring->last_context;u32 hw_flags =0 ;int ret;
@@ -415,8 +416,11 @@ static int do_switch(struct i915_hw_context *to)return ret; } -if (!to->obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (to->obj, to->obj->cache_level); +if (!to->obj->has_global_gtt_mapping) { +struct i915_vma *vma =i915_gem_obj_to_vma (to->obj, + &dev_priv->gtt.base); + vma->vm->bind_vma (vma, to->obj->cache_level,GLOBAL_BIND ); + }if (!to->is_initialized ||is_default_context (to)) hw_flags |=MI_RESTORE_INHIBIT ;
drivers/gpu/drm/i915/i915_gem_execbuffer.c +12 -17
@@ -286,8 +286,9 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,if (unlikely (IS_GEN6 (dev) && reloc->write_domain ==I915_GEM_DOMAIN_INSTRUCTION && !target_i915_obj->has_global_gtt_mapping)) { -i915_gem_gtt_bind_object (target_i915_obj, - target_i915_obj->cache_level); +struct i915_vma *vma =i915_gem_obj_to_vma (obj, vm); + vma->vm->bind_vma (vma, target_i915_obj->cache_level, +GLOBAL_BIND ); }/* Validate that the target is in a valid r/w GPU domain */
@@ -464,11 +465,12 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,struct intel_ring_buffer *ring, bool *need_reloc) { - struct drm_i915_private *dev_priv = ring->dev->dev_private; + struct drm_i915_gem_object *obj = vma->obj;struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;bool has_fenced_gpu_access =INTEL_INFO (ring->dev)->gen <4 ;bool need_fence, need_mappable; -struct drm_i915_gem_object *obj = vma->obj; +u32 flags = (entry->flags &EXEC_OBJECT_NEEDS_GTT ) && + !vma->obj->has_global_gtt_mapping ?GLOBAL_BIND :0 ;int ret; need_fence =
@@ -497,14 +499,6 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, } } -/* Ensure ppgtt mapping exists if needed */ -if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) { -i915_ppgtt_bind_object (dev_priv->mm.aliasing_ppgtt, - obj, obj->cache_level); - - obj->has_aliasing_ppgtt_mapping =1 ; - } -if (entry->offset != vma->node.start) { entry->offset = vma->node.start; *need_reloc = true;
@@ -515,9 +509,7 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, obj->base.pending_write_domain =I915_GEM_DOMAIN_RENDER ; } -if (entry->flags &EXEC_OBJECT_NEEDS_GTT && - !obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (obj, obj->cache_level); + vma->vm->bind_vma (vma, obj->cache_level, flags);return 0 ; }
@@ -1117,8 +1109,11 @@ i915_gem_do_execbuffer(struct drm_device *dev, void *data,* batch" bit. Hence we need to pin secure batches into the global gtt. * hsw should have this fixed, but let's be paranoid and do it * unconditionally for now. */ -if (flags &I915_DISPATCH_SECURE && !batch_obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (batch_obj, batch_obj->cache_level); +if (flags &I915_DISPATCH_SECURE && + !batch_obj->has_global_gtt_mapping) { +struct i915_vma *vma =i915_gem_obj_to_vma (batch_obj, vm); + vm->bind_vma (vma, batch_obj->cache_level,GLOBAL_BIND ); + } ret =i915_gem_execbuffer_move_to_gpu (ring, &eb->vmas);if (ret)
drivers/gpu/drm/i915/i915_gem_gtt.c +3 -45
73 lines not on this page · raw message
Signature · 1 line
1.8.4
Re: [PATCH 7/8] drm/i915: Use the new vm [un]bind functions
Quoted 31 lines · Ben Widawsky
drivers/gpu/drm/i915/i915_gem_execbuffer.c +3
-10@@ -464,11 +465,12 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,struct intel_ring_buffer *ring, bool *need_reloc) { - struct drm_i915_private *dev_priv = ring->dev->dev_private; + struct drm_i915_gem_object *obj = vma->obj;struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;bool has_fenced_gpu_access =INTEL_INFO (ring->dev)->gen <4 ;bool need_fence, need_mappable; -struct drm_i915_gem_object *obj = vma->obj; +u32 flags = (entry->flags &EXEC_OBJECT_NEEDS_GTT ) && + !vma->obj->has_global_gtt_mapping ?GLOBAL_BIND :0 ;int ret; need_fence =@@ -497,14 +499,6 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, } } -/* Ensure ppgtt mapping exists if needed */ -if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) { -i915_ppgtt_bind_object (dev_priv->mm.aliasing_ppgtt, - obj, obj->cache_level); - - obj->has_aliasing_ppgtt_mapping =1 ; - } -if (entry->offset != vma->node.start) { entry->offset = vma->node.start; *need_reloc = true;
drivers/gpu/drm/i915/i915_gem_execbuffer.c i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, · line 512
Quoted 7 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1
-3@@ -515,9 +509,7 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, obj->base.pending_write_domain =I915_GEM_DOMAIN_RENDER ; } -if (entry->flags &EXEC_OBJECT_NEEDS_GTT && - !obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (obj, obj->cache_level); + vma->vm->bind_vma (vma, obj->cache_level, flags);
This results in us rebinding into the vma unconditionally once/twice for every object on every execbuffer, right?
-Chris
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
Re: [PATCH 7/8] drm/i915: Use the new vm [un]bind functions
Quoted 35 lines · Chris Wilson
Quoted 34 lines · Ben Widawsky
drivers/gpu/drm/i915/i915_gem_execbuffer.c +3
-10@@ -464,11 +465,12 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,struct intel_ring_buffer *ring, bool *need_reloc) { - struct drm_i915_private *dev_priv = ring->dev->dev_private; + struct drm_i915_gem_object *obj = vma->obj;struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;bool has_fenced_gpu_access =INTEL_INFO (ring->dev)->gen <4 ;bool need_fence, need_mappable; -struct drm_i915_gem_object *obj = vma->obj; +u32 flags = (entry->flags &EXEC_OBJECT_NEEDS_GTT ) && + !vma->obj->has_global_gtt_mapping ?GLOBAL_BIND :0 ;int ret; need_fence =@@ -497,14 +499,6 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, } } -/* Ensure ppgtt mapping exists if needed */ -if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) { -i915_ppgtt_bind_object (dev_priv->mm.aliasing_ppgtt, - obj, obj->cache_level); - - obj->has_aliasing_ppgtt_mapping =1 ; - } -if (entry->offset != vma->node.start) { entry->offset = vma->node.start; *need_reloc = true;@@ -515,9 +509,7 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, obj->base.pending_write_domain =I915_GEM_DOMAIN_RENDER ; }
Quoted 9 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, · line 512
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1
-3-if (entry->flags &EXEC_OBJECT_NEEDS_GTT && - !obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (obj, obj->cache_level); + vma->vm->bind_vma (vma, obj->cache_level, flags);This results in us rebinding into the vma unconditionally once/twice for every object on every execbuffer, right?
-Chris
Hmm, how does this suit you (untested)? I think it is addressing your concern. Also, I suspect squashing this patch with the previous is probably a good thing to do.
static void gen6_ggtt_bind_vma(struct i915_vma *vma,
enum i915_cache_level cache_level, u32 flags)
{
struct drm_device *dev = vma->vm->dev;
struct drm_i915_private *dev_priv = dev->dev_private;
struct drm_i915_gem_object *obj = vma->obj;
const unsigned long entry = vma->node.start >> PAGE_SHIFT;
/* If there is an aliasing PPGTT, and the user didn't explicitly ask for
* the global, just use aliasing */
if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) {
/* If the object is unbound, or we're change the cache bits */
if (!obj->has_global_gtt_mapping ||
(cache_level != obj->cache_level)) {
gen6_ggtt_insert_entries(vma->vm, obj->pages, entry,
cache_level);
obj->has_global_gtt_mapping = 1;
}
}
/* If put the mapping in the aliasing PPGTT as well as Global if we have
* aliasing, but the user requested global. */
if (dev_priv->mm.aliasing_ppgtt &&
(!obj->has_aliasing_ppgtt_mapping ||
(cache_level != obj->cache_level))) {
gen6_ppgtt_insert_entries(&dev_priv->mm.aliasing_ppgtt->base,
vma->obj->pages, entry, cache_level);
vma->obj->has_aliasing_ppgtt_mapping = 1;
}
}Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
Re: [PATCH 7/8] drm/i915: Use the new vm [un]bind functions
Quoted 44 lines · Ben Widawsky
Quoted 35 lines · Chris Wilson
Quoted 34 lines · Ben Widawsky
drivers/gpu/drm/i915/i915_gem_execbuffer.c +3
-10@@ -464,11 +465,12 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,struct intel_ring_buffer *ring, bool *need_reloc) { - struct drm_i915_private *dev_priv = ring->dev->dev_private; + struct drm_i915_gem_object *obj = vma->obj;struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;bool has_fenced_gpu_access =INTEL_INFO (ring->dev)->gen <4 ;bool need_fence, need_mappable; -struct drm_i915_gem_object *obj = vma->obj; +u32 flags = (entry->flags &EXEC_OBJECT_NEEDS_GTT ) && + !vma->obj->has_global_gtt_mapping ?GLOBAL_BIND :0 ;int ret; need_fence =@@ -497,14 +499,6 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, } } -/* Ensure ppgtt mapping exists if needed */ -if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) { -i915_ppgtt_bind_object (dev_priv->mm.aliasing_ppgtt, - obj, obj->cache_level); - - obj->has_aliasing_ppgtt_mapping =1 ; - } -if (entry->offset != vma->node.start) { entry->offset = vma->node.start; *need_reloc = true;@@ -515,9 +509,7 @@ i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, obj->base.pending_write_domain =I915_GEM_DOMAIN_RENDER ; }Quoted 7 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c i915_gem_execbuffer_reserve_vma(struct i915_vma *vma, · line 512
Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_execbuffer.c +1
-3-if (entry->flags &EXEC_OBJECT_NEEDS_GTT && - !obj->has_global_gtt_mapping) -i915_gem_gtt_bind_object (obj, obj->cache_level); + vma->vm->bind_vma (vma, obj->cache_level, flags);This results in us rebinding into the vma unconditionally once/twice for every object on every execbuffer, right?
Quoted 3 lines
Hmm, how does this suit you (untested)? I think it is addressing your concern. Also, I suspect squashing this patch with the previous is probably a good thing to do.
That looks like it should do the trick. However, would it not be cleaner to move the vma->vm->bind_vma() into the pin()?
-Chris
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
PATCH 8/8 drm/i915: eliminate vm->insert_entries()
In reply to Ben WidawskyFrom: Ben Widawsky <[email protected]>
With bind/unbind function pointers in place, we no longer need insert_entries. We could, and want, to remove clear_range, however it's not totally easy at this point. Since it's used in a couple of place still that don't only deal in objects: setup, ppgtt init, and restore gtt mappings.
--- drivers/gpu/drm/i915/i915_gem_gtt.c | 17 +---------------- 1 file changed, 1 insertion(+), 16 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +1 -16
@@ -339,8 +339,8 @@ static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt) ppgtt->enable = gen6_ppgtt_enable; ppgtt->base.unbind_vma =NULL ; ppgtt->base.clear_range = gen6_ppgtt_clear_range; - ppgtt->base.bind_vma =NULL ; ppgtt->base.insert_entries = gen6_ppgtt_insert_entries; + ppgtt->base.bind_vma =NULL ; ppgtt->base.cleanup = gen6_ppgtt_cleanup; ppgtt->base.scratch = dev_priv->gtt.base.scratch; ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
@@ -591,19 +591,6 @@ static void gen6_ggtt_clear_range(struct i915_address_space *vm,readl (gtt_base); } - -static void i915_ggtt_insert_entries (struct i915_address_space *vm, -struct sg_table *st, -unsigned int pg_start, -enum i915_cache_level cache_level) -{ -unsigned int flags = (cache_level ==I915_CACHE_NONE ) ? -AGP_USER_MEMORY :AGP_USER_CACHED_MEMORY ; - -intel_gtt_insert_sg_entries (st, pg_start, flags); - -} -static void i915_ggtt_bind_vma (struct i915_vma *vma,enum i915_cache_level cache_level,u32 unused)
@@ -917,7 +904,6 @@ static int gen6_gmch_probe(struct drm_device *dev, dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range; dev_priv->gtt.base.unbind_vma = gen6_ggtt_unbind_vma; - dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries; dev_priv->gtt.base.bind_vma = gen6_ggtt_bind_vma;return ret;
@@ -951,7 +937,6 @@ static int i915_gmch_probe(struct drm_device *dev, dev_priv->gtt.do_idle_maps =needs_idle_maps (dev_priv->dev); dev_priv->gtt.base.clear_range = i915_ggtt_clear_range; dev_priv->gtt.base.unbind_vma = i915_ggtt_unbind_vma; - dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries; dev_priv->gtt.base.bind_vma = i915_ggtt_bind_vma;return 0 ;
Signature · 1 line
1.8.4