drm/i915/guc: Move GuC wq_reserve_space to ring_reserve_space
Subsystems (2)
PATCH 1/1 drm/i915/guc: Move GuC wq_reserve_space to ring_reserve_space
From: Alex Dai <[email protected]>
Split GuC work queue space reserve and submission and move the space reserve to where ring space is reserved. The reason is that failure in intel_logical_ring_advance_and_submit won't be handled. In the case timeout happens, driver can return early in order to handle the error.
--- drivers/gpu/drm/i915/i915_guc_submission.c | 21 +++++++++------------ drivers/gpu/drm/i915/intel_guc.h | 1 + drivers/gpu/drm/i915/intel_lrc.c | 10 +++++++++- 3 files changed, 19 insertions(+), 13 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +9 -12
@@ -472,25 +472,21 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_reserve_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) {if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; - -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; -/* this will break the loop */ timeout_counter =0 ; ret =0 ;
@@ -512,11 +508,12 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,struct guc_wq_item *wqi;void *base;u32 tail, wq_len, wq_off =0 ; -int ret; - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; + wq_off = gc->wq_tail; + +/* advance the tail for next workqueue item */ + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither
drivers/gpu/drm/i915/intel_guc.h +1 -0
@@ -123,5 +123,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_reserve_space (struct i915_guc_client *client);#endif
drivers/gpu/drm/i915/intel_lrc.c +9 -1
@@ -844,6 +844,8 @@ int intel_logical_ring_begin(struct drm_i915_gem_request *req, int num_dwords)int intel_logical_ring_reserve_space (struct drm_i915_gem_request *request) { +int ret; +/* * The first call merely notes the reserve request and is common for * all back ends. The subsequent localised _begin() call actually
@@ -854,7 +856,13 @@ int intel_logical_ring_reserve_space(struct drm_i915_gem_request *request)*/ intel_ring_reserved_space_reserve (request->ringbuf,MIN_SPACE_FOR_ADD_REQUEST ); -return intel_logical_ring_begin (request,0 ); + ret =intel_logical_ring_begin (request,0 ); +if (ret) +return ret; + + ret =i915_guc_wq_reserve_space (request->i915->guc.execbuf_client); + +return ret; }/**
Signature · 6 lines
2.5.0 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 7 lines · On Tue, Dec 08, 2015 at 05:04:50PM -0800, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space reserve and submission and move the space reserve to where ring space is reserved. The reason is that failure in
intel_logical_ring_advance_and_submitwon't be handled. In the case timeout happens, driver can return early in order to handle the error.
Not here. You want the request_alloc_extras callback. -Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Quoted 10 lines · Chris Wilson
Quoted 7 lines · On Tue, Dec 08, 2015 at 05:04:50PM -0800, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space reserve and submission and move the space reserve to where ring space is reserved. The reason is that failure in
intel_logical_ring_advance_and_submitwon't be handled. In the case timeout happens, driver can return early in order to handle the error.Not here. You want the
request_alloc_extrascallback.
OK. That is even before touching the ring.
Thanks,
Alex
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 1/1 drm/i915/guc: Move GuC wq_reserve_space to alloc_request_extras
In reply to [email protected]From: Alex Dai <[email protected]>
Split GuC work queue space reserve from submission and move it to ring_alloc_request_extras. The reason is that failure in later i915_add_request() won't be handled. In the case timeout happens, driver can return early in order to handle the error.
v1: Move wq_reserve_space to ring_reserve_space
v2: Move wq_reserve_space to alloc_request_extras (Chris Wilson)
--- drivers/gpu/drm/i915/i915_guc_submission.c | 21 +++++++++------------ drivers/gpu/drm/i915/intel_guc.h | 1 + drivers/gpu/drm/i915/intel_lrc.c | 6 ++++++ 3 files changed, 16 insertions(+), 12 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +9 -12
@@ -472,25 +472,21 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_reserve_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) {if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; - -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; -/* this will break the loop */ timeout_counter =0 ; ret =0 ;
@@ -512,11 +508,12 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,struct guc_wq_item *wqi;void *base;u32 tail, wq_len, wq_off =0 ; -int ret; - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; + wq_off = gc->wq_tail; + +/* advance the tail for next workqueue item */ + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither
drivers/gpu/drm/i915/intel_guc.h +1 -0
@@ -123,5 +123,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_reserve_space (struct i915_guc_client *client);#endif
drivers/gpu/drm/i915/intel_lrc.c +6 -0
@@ -667,6 +667,12 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret; } +/* Reserve GuC WQ space here (one request needs one WQ item) because +* the later i915_add_request() call can't fail. */ + ret =i915_guc_wq_reserve_space (request->i915->guc.execbuf_client); +if (ret) +return ret; +return 0 ; }
Signature · 6 lines
2.5.0 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v2] drm/i915/guc: Move GuC wq_reserve_space to alloc_request_extras
Quoted 22 lines · On 09/12/15 18:50, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space reserve from submission and move it to
ring_alloc_request_extras. The reason is that failure in lateri915_add_request()won't be handled. In the case timeout happens, driver can return early in order to handle the error.v1: Move
wq_reserve_spacetoring_reserve_space
v2: Movewq_reserve_spacetoalloc_request_extras(Chris Wilson)Signed-off-by Alex Dai <[email protected]>--- drivers/gpu/drm/i915/i915_guc_submission.c | 21 +++++++++------------ drivers/gpu/drm/i915/intel_guc.h | 1 + drivers/gpu/drm/i915/intel_lrc.c | 6 ++++++ 3 files changed, 16 insertions(+), 12 deletions(-)drivers/gpu/drm/i915/i915_guc_submission.c +0
-0
drivers/gpu/drm/i915/i915_guc_submission.c static void guc_fini_ctx_desc(struct intel_guc *guc, · line 475
Quoted 6 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1
-2@@ -472,25 +472,21 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_reserve_space (struct i915_guc_client *gc)
I think the name is misleading, because we don't actually reserve anything here, just check that there is some free space in the WQ.
(We certainly don't WANT to reserve anything, because that would be difficult to clean up in the event of submission failure for any other reason. So I think it's only the name needs changing. Although ...
Quoted 4 lines
drivers/gpu/drm/i915/i915_guc_submission.c +0
-0{struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );
drivers/gpu/drm/i915/i915_guc_submission.c static void guc_fini_ctx_desc(struct intel_guc *guc, · line 489
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3
-0int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) {if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) {
... as an alternative strategy, we could cache the calculated freespace in the client structure; then if we already know there's at least 1 slot free from last time we checked, we could then just decrement the cached value and avoid the kmap+spinwait overhead. Only when we reach 0 would we have to go through this code to refresh our view of desc->head and recalculate the actual current freespace. [NB: clear cached value on reset?]
Does that sound like a useful optimisation?
Quoted 13 lines
drivers/gpu/drm/i915/i915_guc_submission.c +0
-6- *offset = gc->wq_tail; - -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; -/* this will break the loop */ timeout_counter =0 ; ret =0 ;@@ -512,11 +508,12 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,struct guc_wq_item *wqi;void *base;u32 tail, wq_len, wq_off =0 ;
drivers/gpu/drm/i915/i915_guc_submission.c static int guc_add_workqueue_item(struct i915_guc_client *gc, · line 516
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +5
-4-int ret; - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; + wq_off = gc->wq_tail; + +/* advance the tail for next workqueue item */ + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;
I was a bit unhappy about this code just assuming that there *must* be space (because we KNOW we've checked above) -- unless someone violated the proper calling sequence (TDR?). OTOH, it would be too expensive to go through the map-and-calculate code all over again just to catch an unlikely scenario. But, if we cache the last-calculated value as above, then the check could be cheap :) For example, just add a pre_checked size field that's set by the pre-check and then checked and decremented on submission; there shouldn't be more than one submission in progress at a time, because dev->struct_mutex is held across the whole sequence (but it's not an error to see two pre-checks in a row, because a request can be abandoned partway).
Quoted 19 lines
drivers/gpu/drm/i915/i915_guc_submission.c +0
-0/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither drivers/gpu/drm/i915/intel_guc.h +1
-0@@ -123,5 +123,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_reserve_space (struct i915_guc_client *client);#endif drivers/gpu/drm/i915/intel_lrc.c +0
-0@@ -667,6 +667,12 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret;
drivers/gpu/drm/i915/intel_lrc.c int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request · line 677
Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +6
-0} +/* Reserve GuC WQ space here (one request needs one WQ item) because +* the later i915_add_request() call can't fail. */ + ret =i915_guc_wq_reserve_space (request->i915->guc.execbuf_client); +if (ret) +return ret; +return 0 ; }
Worth checking for GuC submission before that call? Maybe not ...
.Dave.
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v2] drm/i915/guc: Move GuC wq_reserve_space to alloc_request_extras
Quoted 33 lines · Dave Gordon
Quoted 22 lines · On 09/12/15 18:50, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space reserve from submission and move it to
ring_alloc_request_extras. The reason is that failure in lateri915_add_request()won't be handled. In the case timeout happens, driver can return early in order to handle the error.v1: Move
wq_reserve_spacetoring_reserve_space
v2: Movewq_reserve_spacetoalloc_request_extras(Chris Wilson)Signed-off-by Alex Dai <[email protected]>--- drivers/gpu/drm/i915/i915_guc_submission.c | 21 +++++++++------------ drivers/gpu/drm/i915/intel_guc.h | 1 + drivers/gpu/drm/i915/intel_lrc.c | 6 ++++++ 3 files changed, 16 insertions(+), 12 deletions(-)drivers/gpu/drm/i915/i915_guc_submission.c +0
-0drivers/gpu/drm/i915/i915_guc_submission.c static void guc_fini_ctx_desc(struct intel_guc *guc, · line 475
Quoted 6 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1
-2@@ -472,25 +472,21 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_reserve_space (struct i915_guc_client *gc)I think the name is misleading, because we don't actually reserve anything here, just check that there is some free space in the WQ.
Quoted 3 lines
(We certainly don't WANT to reserve anything, because that would be difficult to clean up in the event of submission failure for any other reason. So I think it's only the name needs changing. Although ...
I was trying to use similar name as ring_reserve_space. It follows the same pattern as filling ring buffer - reserve, emit and advance. The only difference is it reserves 4 dwords (or checks space) rather than various size of bytes for ring buffer case. Maybe using 'check' is better for this case because 4 dwords in wq is what we need for submission via GuC.
Quoted 22 lines
Quoted 4 lines
drivers/gpu/drm/i915/i915_guc_submission.c +0
-0{struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );drivers/gpu/drm/i915/i915_guc_submission.c static void guc_fini_ctx_desc(struct intel_guc *guc, · line 489
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3
-0int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) {if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) {... as an alternative strategy, we could cache the calculated freespace in the client structure; then if we already know there's at least 1 slot free from last time we checked, we could then just decrement the cached value and avoid the kmap+spinwait overhead. Only when we reach 0 would we have to go through this code to refresh our view of
desc->headand recalculate the actual current freespace. [NB: clear cached value on reset?]
Quoted 1 line
Does that sound like a useful optimisation?
I think it is a good idea.
Quoted 27 lines
Quoted 13 lines
drivers/gpu/drm/i915/i915_guc_submission.c +0
-6- *offset = gc->wq_tail; - -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; -/* this will break the loop */ timeout_counter =0 ; ret =0 ;@@ -512,11 +508,12 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,struct guc_wq_item *wqi;void *base;u32 tail, wq_len, wq_off =0 ;drivers/gpu/drm/i915/i915_guc_submission.c static int guc_add_workqueue_item(struct i915_guc_client *gc, · line 516
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +5
-4-int ret; - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; + wq_off = gc->wq_tail; + +/* advance the tail for next workqueue item */ + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;I was a bit unhappy about this code just assuming that there *must* be space (because we KNOW we've checked above) -- unless someone violated the proper calling sequence (TDR?). OTOH, it would be too expensive to
Quoted 8 lines
go through the map-and-calculate code all over again just to catch an unlikely scenario. But, if we cache the last-calculated value as above, then the check could be cheap :) For example, just add a
pre_checkedsize field that's set by the pre-check and then checked and decremented on submission; there shouldn't be more than one submission in progress at a time, becausedev->struct_mutexis held across the whole sequence (but it's not an error to see two pre-checks in a row, because a request can be abandoned partway).
I don't understand the concerns here. As I said above, filling GuC WQ is same as filling ring buffer - reserve (check), emit and advance. These two lines are GuC version of intel_logical_ring_emit and intel_logical_ring_advance. Maybe if I move these two lines to end of guc_add_workqueue_item, people can understand it more easily.
Quoted 21 lines
Quoted 21 lines
drivers/gpu/drm/i915/i915_guc_submission.c +0
-0/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither drivers/gpu/drm/i915/intel_guc.h +1
-0@@ -123,5 +123,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_reserve_space (struct i915_guc_client *client);#endif drivers/gpu/drm/i915/intel_lrc.c +0
-0@@ -667,6 +667,12 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret; }
Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request · line 677
Quoted 8 lines
drivers/gpu/drm/i915/intel_lrc.c +6
-0+/* Reserve GuC WQ space here (one request needs one WQ item) because +* the later i915_add_request() call can't fail. */ + ret =i915_guc_wq_reserve_space (request->i915->guc.execbuf_client); +if (ret) +return ret; +return 0 ; }Worth checking for GuC submission before that call? Maybe not ...
It is purely code alignment issue if I add a if() block. The execbuf_client is valid only when GuC submission is enabled and the check is done inside that function call.
Alex
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v3 1/1 drm/i915/guc: Move GuC wq_check_space to alloc_request_extras
In reply to [email protected]From: Alex Dai <[email protected]>
Split GuC work queue space checking from submission and move it to ring_alloc_request_extras. The reason is that failure in later i915_add_request() won't be handled. In the case timeout happens, driver can return early in order to handle the error.
v1: Move wq_reserve_space to ring_reserve_space
v2: Move wq_reserve_space to alloc_request_extras (Chris Wilson)
v3: The work queue head pointer is cached by driver now. So we can
quickly return if space is available. s/reserve/check/g (Dave Gordon)
--- drivers/gpu/drm/i915/i915_guc_submission.c | 34 ++++++++++++++++++------------ drivers/gpu/drm/i915/intel_guc.h | 2 ++ drivers/gpu/drm/i915/intel_lrc.c | 6 ++++++ 3 files changed, 28 insertions(+), 14 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +20 -14
@@ -207,6 +207,9 @@ static int guc_ring_doorbell(struct i915_guc_client *gc)/* Update the tail so it is visible to GuC */ desc->tail = gc->wq_tail; +/* Cache the head where GuC is processing */ + gc->wq_head = desc->head; +/* current cookie */ db_cmp.db_status =GUC_DOORBELL_ENABLED ; db_cmp.cookie = gc->cookie;
@@ -472,28 +475,30 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_check_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + +/* Quickly return if wq space is available since last time we cache the +* head position. */ +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) { -if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; + gc->wq_head = desc->head; -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; - -/* this will break the loop */ - timeout_counter =0 ; +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) { ret =0 ; +break ; }if (timeout_counter)
@@ -512,11 +517,8 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,struct guc_wq_item *wqi;void *base;u32 tail, wq_len, wq_off =0 ; -int ret; - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; + wq_off = gc->wq_tail;/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither
@@ -551,6 +553,10 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,kunmap_atomic (base); +/* advance the tail for next workqueue item */ + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ; +return 0 ; }
drivers/gpu/drm/i915/intel_guc.h +2 -0
@@ -43,6 +43,7 @@ struct i915_guc_client {uint32_t wq_offset;uint32_t wq_size;uint32_t wq_tail; +uint32_t wq_head;/* GuC submission statistics & status */ uint64_t submissions[I915_NUM_RINGS ];
@@ -123,5 +124,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_check_space (struct i915_guc_client *client);#endif
drivers/gpu/drm/i915/intel_lrc.c +6 -0
@@ -667,6 +667,12 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret; } +/* Reserve GuC WQ space here (one request needs one WQ item) because +* the later i915_add_request() call can't fail. */ + ret =i915_guc_wq_check_space (request->i915->guc.execbuf_client); +if (ret) +return ret; +return 0 ; }
Signature · 6 lines
2.5.0 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v3] drm/i915/guc: Move GuC wq_check_space to alloc_request_extras
Quoted 25 lines · On 11/12/15 01:17, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space checking from submission and move it to
ring_alloc_request_extras. The reason is that failure in lateri915_add_request()won't be handled. In the case timeout happens, driver can return early in order to handle the error.v1: Move
wq_reserve_spacetoring_reserve_space
v2: Movewq_reserve_spacetoalloc_request_extras(Chris Wilson)
v3: The work queue head pointer is cached by driver now. So we can
quickly return if space is available. s/reserve/check/g (Dave Gordon)Signed-off-by Alex Dai <[email protected]>--- drivers/gpu/drm/i915/i915_guc_submission.c | 34 ++++++++++++++++++------------ drivers/gpu/drm/i915/intel_guc.h | 2 ++ drivers/gpu/drm/i915/intel_lrc.c | 6 ++++++ 3 files changed, 28 insertions(+), 14 deletions(-)drivers/gpu/drm/i915/i915_guc_submission.c +0
-0
drivers/gpu/drm/i915/i915_guc_submission.c static int guc_ring_doorbell(struct i915_guc_client *gc) · line 212
Quoted 6 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3
-0@@ -207,6 +207,9 @@ static int guc_ring_doorbell(struct i915_guc_client *gc)/* Update the tail so it is visible to GuC */ desc->tail = gc->wq_tail; +/* Cache the head where GuC is processing */ + gc->wq_head = desc->head; +
This could be at the end of this function, thus:
/* update the cookie to newly read cookie from GuC */ db_cmp.cookie = db_ret.cookie; db_exc.cookie = db_ret.cookie + 1; if (db_exc.cookie == 0) db_exc.cookie = 1; }
+ /* Finally, update the cached copy of the GuC's WQ head */
+ gc->wq_head = desc->head;
+
kunmap_atomic(base); return ret; }
which might increase the chance that we will get an updated value and thus avoid having to spinwait on the next check().
Quoted 45 lines
drivers/gpu/drm/i915/i915_guc_submission.c +12
-10/* current cookie */ db_cmp.db_status =GUC_DOORBELL_ENABLED ; db_cmp.cookie = gc->cookie;@@ -472,28 +475,30 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_check_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + +/* Quickly return if wq space is available since last time we cache the +* head position. */ +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) { -if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; + gc->wq_head = desc->head; -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; - -/* this will break the loop */ - timeout_counter =0 ; +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) { ret =0 ; +break ; }if (timeout_counter)
drivers/gpu/drm/i915/i915_guc_submission.c static int guc_add_workqueue_item(struct i915_guc_client *gc, · line 521
Quoted 9 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1
-4@@ -512,11 +517,8 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,struct guc_wq_item *wqi;void *base;u32 tail, wq_len, wq_off =0 ; -int ret; - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; + wq_off = gc->wq_tail;
Conversely, I'd prefer the updates below to be here instead; also the check that I mentioned last time:
+ u32 space = CIRC_SPACE(gc->wq_tail, gc->wq_head, gc->wq_size);
+ if (WARN_ON(space < sizeof(struct guc_wq_item))
+ return -ENOSPC; /* shouldn't happen */
+
+ /* postincrement WQ tail for next time */
+ wq_off = gc->wq_tail;
+ gc->wq_tail += sizeof(struct guc_wq_item);
+ gc->wq_tail &= gc->wq_size - 1;
drivers/gpu/drm/i915/i915_guc_submission.c static int guc_add_workqueue_item(struct i915_guc_client *gc, · line 558
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3
-0/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither @@ -551,6 +553,10 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,kunmap_atomic (base); +/* advance the tail for next workqueue item */ + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;
(moved to above)
Quoted 29 lines
drivers/gpu/drm/i915/i915_guc_submission.c +1
-0+return 0 ; }drivers/gpu/drm/i915/intel_guc.h +2
-0@@ -43,6 +43,7 @@ struct i915_guc_client {uint32_t wq_offset;uint32_t wq_size;uint32_t wq_tail; +uint32_t wq_head;/* GuC submission statistics & status */ uint64_t submissions[I915_NUM_RINGS ];@@ -123,5 +124,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_check_space (struct i915_guc_client *client);#endif drivers/gpu/drm/i915/intel_lrc.c +0
-0@@ -667,6 +667,12 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret;
drivers/gpu/drm/i915/intel_lrc.c int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request · line 677
Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +6
-0} +/* Reserve GuC WQ space here (one request needs one WQ item) because +* the later i915_add_request() call can't fail. */ + ret =i915_guc_wq_check_space (request->i915->guc.execbuf_client); +if (ret) +return ret; +return 0 ; }
It looks odd to call a GuC-only function regardless of whether we're using the GuC, so:
+ if (i915.enable_guc_submission) {
+ /*
+ * Check that the GuC has space for the request before
+ * going any further, as the i915_add_request() call
+ * later on mustn't fail ...
+ */
+ struct intel_guc *guc = &request->i915->guc;
+
+ ret = i915_guc_wq_check_space(guc->execbuf_client);
+ if (ret)
+ return ret;
+ }
.Dave.
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v4 1/1 drm/i915/guc: Move GuC wq_check_space to alloc_request_extras
In reply to [email protected]From: Alex Dai <[email protected]>
Split GuC work queue space checking from submission and move it to ring_alloc_request_extras. The reason is that failure in later i915_add_request() won't be handled. In the case timeout happens, driver can return early in order to handle the error.
v1: Move wq_reserve_space to ring_reserve_space
v2: Move wq_reserve_space to alloc_request_extras (Chris Wilson)
v3: The work queue head pointer is cached by driver now. So we can
quickly return if space is available. s/reserve/check/g (Dave Gordon)
v4: Update cached wq head after ring doorbell; check wq space before
ring doorbell in case unexpected error happens; call wq space check only when GuC submission is enabled. (Dave Gordon)
drivers/gpu/drm/i915/i915_guc_submission.c +24 -15
@@ -247,6 +247,9 @@ static int guc_ring_doorbell(struct i915_guc_client *gc) db_exc.cookie =1 ; } +/* Finally, update the cached copy of the GuC's WQ head */ + gc->wq_head = desc->head; +kunmap_atomic (base);return ret; }
@@ -472,28 +475,30 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_check_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + +/* Quickly return if wq space is available since last time we cache the +* head position. */ +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) { -if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; + gc->wq_head = desc->head; -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; - -/* this will break the loop */ - timeout_counter =0 ; +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) { ret =0 ; +break ; }if (timeout_counter)
@@ -511,12 +516,16 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,enum intel_ring_id ring_id = rq->ring->id;struct guc_wq_item *wqi;void *base; -u32 tail, wq_len, wq_off =0 ; -int ret; +u32 tail, wq_len, wq_off, space; + + space =CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size); +if (WARN_ON (space <sizeof (struct guc_wq_item ))) +return -ENOSPC ;/* shouldn't happen */ - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; +/* postincrement WQ tail for next time */ + wq_off = gc->wq_tail; + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither
drivers/gpu/drm/i915/intel_guc.h +2 -0
@@ -43,6 +43,7 @@ struct i915_guc_client {uint32_t wq_offset;uint32_t wq_size;uint32_t wq_tail; +uint32_t wq_head;/* GuC submission statistics & status */ uint64_t submissions[I915_NUM_RINGS ];
@@ -123,5 +124,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_check_space (struct i915_guc_client *client);#endif
drivers/gpu/drm/i915/intel_lrc.c +13 -0
@@ -667,6 +667,19 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret; } +if (i915.enable_guc_submission) { +/* +* Check that the GuC has space for the request before +* going any further, as the i915_add_request() call +* later on mustn't fail ... +*/ +struct intel_guc *guc = &request->i915->guc; + + ret =i915_guc_wq_check_space (guc->execbuf_client); +if (ret) +return ret; + } +return 0 ; }
Signature · 6 lines
2.5.0 _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [PATCH v4] drm/i915/guc: Move GuC wq_check_space to alloc_request_extras
Quoted 16 lines · On 16/12/15 19:45, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space checking from submission and move it to
ring_alloc_request_extras. The reason is that failure in lateri915_add_request()won't be handled. In the case timeout happens, driver can return early in order to handle the error.v1: Move
wq_reserve_spacetoring_reserve_space
v2: Movewq_reserve_spacetoalloc_request_extras(Chris Wilson)
v3: The work queue head pointer is cached by driver now. So we can
quickly return if space is available. s/reserve/check/g (Dave Gordon)
v4: Update cached wq head after ring doorbell; check wq space before
ring doorbell in case unexpected error happens; call wq space check only when GuC submission is enabled. (Dave Gordon)
Quoted 1 line
Signed-off-by Alex Dai <[email protected]>
LGTM.
Quoted 111 lines
drivers/gpu/drm/i915/i915_guc_submission.c +24
-15@@ -247,6 +247,9 @@ static int guc_ring_doorbell(struct i915_guc_client *gc) db_exc.cookie =1 ; } +/* Finally, update the cached copy of the GuC's WQ head */ + gc->wq_head = desc->head; +kunmap_atomic (base);return ret; }@@ -472,28 +475,30 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_check_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + +/* Quickly return if wq space is available since last time we cache the +* head position. */ +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) { -if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; + gc->wq_head = desc->head; -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; - -/* this will break the loop */ - timeout_counter =0 ; +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) { ret =0 ; +break ; }if (timeout_counter)@@ -511,12 +516,16 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,enum intel_ring_id ring_id = rq->ring->id;struct guc_wq_item *wqi;void *base; -u32 tail, wq_len, wq_off =0 ; -int ret; +u32 tail, wq_len, wq_off, space; + + space =CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size); +if (WARN_ON (space <sizeof (struct guc_wq_item ))) +return -ENOSPC ;/* shouldn't happen */ - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; +/* postincrement WQ tail for next time */ + wq_off = gc->wq_tail; + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither drivers/gpu/drm/i915/intel_guc.h +2
-0@@ -43,6 +43,7 @@ struct i915_guc_client {uint32_t wq_offset;uint32_t wq_size;uint32_t wq_tail; +uint32_t wq_head;/* GuC submission statistics & status */ uint64_t submissions[I915_NUM_RINGS ];@@ -123,5 +124,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_check_space (struct i915_guc_client *client);#endif drivers/gpu/drm/i915/intel_lrc.c +6
-0@@ -667,6 +667,19 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret; } +if (i915.enable_guc_submission) { +/* +* Check that the GuC has space for the request before +* going any further, as the i915_add_request() call +* later on mustn't fail ... +*/
drivers/gpu/drm/i915/intel_lrc.c int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request · line 685
Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +7
-0+struct intel_guc *guc = &request->i915->guc; + + ret =i915_guc_wq_check_space (guc->execbuf_client); +if (ret) +return ret; + } +return 0 ; }
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Daniel Vetter
AppliedRe: [PATCH v4] drm/i915/guc: Move GuC wq_check_space to alloc_request_extras
Quoted 17 lines · Dave Gordon
Quoted 16 lines · On 16/12/15 19:45, [email protected] wrote:
From: Alex Dai <[email protected]>
Split GuC work queue space checking from submission and move it to
ring_alloc_request_extras. The reason is that failure in lateri915_add_request()won't be handled. In the case timeout happens, driver can return early in order to handle the error.v1: Move
wq_reserve_spacetoring_reserve_space
v2: Movewq_reserve_spacetoalloc_request_extras(Chris Wilson)
v3: The work queue head pointer is cached by driver now. So we can
quickly return if space is available. s/reserve/check/g (Dave Gordon)
v4: Update cached wq head after ring doorbell; check wq space before
ring doorbell in case unexpected error happens; call wq space check only when GuC submission is enabled. (Dave Gordon)
Quoted 4 lines
Quoted 1 line
Signed-off-by Alex Dai <[email protected]>LGTM.
Reviewed-by Dave Gordon <[email protected]>
Queued for -next, thanks for the patch.
-Daniel
Quoted 127 lines
Quoted 112 lines
drivers/gpu/drm/i915/i915_guc_submission.c +24
-15@@ -247,6 +247,9 @@ static int guc_ring_doorbell(struct i915_guc_client *gc) db_exc.cookie =1 ; } +/* Finally, update the cached copy of the GuC's WQ head */ + gc->wq_head = desc->head; +kunmap_atomic (base);return ret; }@@ -472,28 +475,30 @@ static void guc_fini_ctx_desc(struct intel_guc *guc, sizeof(desc) * client->ctx_index); } -/* Get valid workqueue item and return it back to offset */ -static int guc_get_workqueue_space (struct i915_guc_client *gc,u32 *offset) +int i915_guc_wq_check_space (struct i915_guc_client *gc) {struct guc_process_desc *desc;void *base;u32 size =sizeof (struct guc_wq_item );int ret = -ETIMEDOUT , timeout_counter =200 ; +if (!gc) +return 0 ; + +/* Quickly return if wq space is available since last time we cache the +* head position. */ +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) +return 0 ; + base =kmap_atomic (i915_gem_object_get_page (gc->client_obj,0 )); desc = base + gc->proc_desc_offset;while (timeout_counter-- >0 ) { -if (CIRC_SPACE (gc->wq_tail, desc->head, gc->wq_size) >= size) { - *offset = gc->wq_tail; + gc->wq_head = desc->head; -/* advance the tail for next workqueue item */ - gc->wq_tail += size; - gc->wq_tail &= gc->wq_size -1 ; - -/* this will break the loop */ - timeout_counter =0 ; +if (CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size) >= size) { ret =0 ; +break ; }if (timeout_counter)@@ -511,12 +516,16 @@ static int guc_add_workqueue_item(struct i915_guc_client *gc,enum intel_ring_id ring_id = rq->ring->id;struct guc_wq_item *wqi;void *base; -u32 tail, wq_len, wq_off =0 ; -int ret; +u32 tail, wq_len, wq_off, space; + + space =CIRC_SPACE (gc->wq_tail, gc->wq_head, gc->wq_size); +if (WARN_ON (space <sizeof (struct guc_wq_item ))) +return -ENOSPC ;/* shouldn't happen */ - ret =guc_get_workqueue_space (gc, &wq_off); -if (ret) -return ret; +/* postincrement WQ tail for next time */ + wq_off = gc->wq_tail; + gc->wq_tail +=sizeof (struct guc_wq_item ); + gc->wq_tail &= gc->wq_size -1 ;/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we * should not have the case where structure wqi is across page, neither drivers/gpu/drm/i915/intel_guc.h +2
-0@@ -43,6 +43,7 @@ struct i915_guc_client {uint32_t wq_offset;uint32_t wq_size;uint32_t wq_tail; +uint32_t wq_head;/* GuC submission statistics & status */ uint64_t submissions[I915_NUM_RINGS ];@@ -123,5 +124,6 @@ int i915_guc_submit(struct i915_guc_client *client, struct drm_i915_gem_request *rq);void i915_guc_submission_disable (struct drm_device *dev);void i915_guc_submission_fini (struct drm_device *dev); +int i915_guc_wq_check_space (struct i915_guc_client *client);#endif drivers/gpu/drm/i915/intel_lrc.c +6
-0@@ -667,6 +667,19 @@ int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *requestreturn ret; } +if (i915.enable_guc_submission) { +/* +* Check that the GuC has space for the request before +* going any further, as the i915_add_request() call +* later on mustn't fail ... +*/ drivers/gpu/drm/i915/intel_lrc.c int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request · line 685
Quoted 10 lines
drivers/gpu/drm/i915/intel_lrc.c +7
-0+struct intel_guc *guc = &request->i915->guc; + + ret =i915_guc_wq_check_space (guc->execbuf_client); +if (ret) +return ret; + } +return 0 ; }_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 7 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch _______________________________________________ Intel-gfx mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/intel-gfx