Various improvements around the GuC topic
Subsystems (2)
v2 cover Various improvements around the GuC topic
They have been discussed in various threads, but I am putting them together as a series for clarity. They now include Joonas' "Sanitize GuC client initialization" with a small tweak, plus an extra doorbell refactoring patch at the end, so they are self-contained and... ready to merge?.
Oscar Mateo (11):
drm/i915/guc: Sanitize GuC client initialization drm/i915/guc: Keep the ctx_pool_vaddr mapped, for easy access drm/i915/guc: Add onion teardown to the GuC setup drm/i915/guc: s/ads_vma/addon
drm/i915/guc: Break out the GuC log extras into their own "runtime"
struct
drm/i915/guc: Make intel_guc_send a function pointer drm/i915/guc: Improve the GuC documentation & comments about proxy
submissions
drm/i915/guc: Wait for doorbell to be inactive before deallocating drm/i915/guc: A little bit more of doorbell sanitization drm/i915/guc: Refactor the concept "GuC context descriptor" into "GuC
stage descriptor"
drm/i915/guc: Enable GuC loading & submission
drivers/gpu/drm/i915/i915_debugfs.c | 8 +- drivers/gpu/drm/i915/i915_drv.c | 6 +- drivers/gpu/drm/i915/i915_guc_submission.c | 733 +++++++++++++++++------------ drivers/gpu/drm/i915/i915_irq.c | 4 +- drivers/gpu/drm/i915/i915_params.c | 8 +- drivers/gpu/drm/i915/intel_guc_fwif.h | 46 +- drivers/gpu/drm/i915/intel_guc_loader.c | 54 ++- drivers/gpu/drm/i915/intel_guc_log.c | 384 ++++++++------- drivers/gpu/drm/i915/intel_uc.c | 17 +- drivers/gpu/drm/i915/intel_uc.h | 48 +- 10 files changed, 727 insertions(+), 581 deletions(-)
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v3 1/11 drm/i915/guc: Sanitize GuC client initialization
Started adding proper teardown to guc_client_alloc, ended up removing quite a few dead ends where errors communicating with the GuC were silently ignored. There also seemed to be quite a few erronous teardown actions performed in case of an error (ordering wrong).
v2:
- Increase function symmetry/proximity (Michal/Daniele)
- Fix __reserve_doorbell accounting for high priority (Daniele)
- Call __update_doorbell_desc! (Daniele)
- Isolate __guc_{,de}allocate_doorbell (Michal/Daniele)
v3:
- "Select" a cacheline is a more accurate verb than "reserve" (Daniele).
- We cannot update & create the doorbell without reserving it first, so move the whole doorbell creation for execbuf_client to the submission enable (Oscar).i
- Add a fixme for ignoring possible doorbell destroy errors.
--- drivers/gpu/drm/i915/i915_debugfs.c | 4 +- drivers/gpu/drm/i915/i915_guc_submission.c | 391 ++++++++++++++++------------- drivers/gpu/drm/i915/intel_guc_fwif.h | 4 +- drivers/gpu/drm/i915/intel_uc.h | 11 +- 4 files changed, 229 insertions(+), 181 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -2
@@ -2469,7 +2469,7 @@ static void i915_guc_client_info(struct seq_file *m,seq_printf (m,"\tPriority %d, GuC ctx index: %u, PD offset 0x%x\n" , client->priority, client->ctx_index, client->proc_desc_offset); -seq_printf (m,"\tDoorbell id %d, offset: 0x%x, cookie 0x%x\n" , +seq_printf (m,"\tDoorbell id %d, offset: 0x%lx, cookie 0x%x\n" , client->doorbell_id, client->doorbell_offset, client->doorbell_cookie);seq_printf (m,"\tWQ size %d, offset: 0x%x, tail %d\n" , client->wq_size, client->wq_offset, client->wq_tail);
@@ -2504,7 +2504,7 @@ static int i915_guc_info(struct seq_file *m, void *data) }seq_printf (m,"Doorbell map:\n" ); -seq_printf (m,"\t%*pb\n" ,GUC_MAX_DOORBELLS , guc->doorbell_bitmap); +seq_printf (m,"\t%*pb\n" ,GUC_NUM_DOORBELLS , guc->doorbell_bitmap);seq_printf (m,"Doorbell next cacheline: 0x%x\n\n" , guc->db_cacheline);seq_printf (m,"GuC total action count: %llu\n" , guc->action_count);
drivers/gpu/drm/i915/i915_guc_submission.c +220 -171
569 lines not on this page · raw message
drivers/gpu/drm/i915/intel_guc_fwif.h +2 -2
@@ -241,8 +241,8 @@ struct guc_doorbell_info {u64 value_qw; } __packed; -#define GUC_MAX_DOORBELLS 256 -#define GUC_INVALID_DOORBELL_ID (GUC_MAX_DOORBELLS) +#define GUC_NUM_DOORBELLS 256 +#define GUC_DOORBELL_INVALID (GUC_NUM_DOORBELLS)#define GUC_DB_SIZE (PAGE_SIZE)#define GUC_WQ_SIZE (PAGE_SIZE * 2)
drivers/gpu/drm/i915/intel_uc.h +5 -6
@@ -72,13 +72,12 @@ struct i915_guc_client {uint32_t engines;/* bitmap of (host) engine ids */ uint32_t priority; -uint32_t ctx_index; +u32 ctx_index;uint32_t proc_desc_offset; -uint32_t doorbell_offset; -uint32_t doorbell_cookie; -uint16_t doorbell_id; -uint16_t padding[3 ];/* Maintain alignment */ +u16 doorbell_id; +unsigned long doorbell_offset; +u32 doorbell_cookie;spinlock_t wq_lock;uint32_t wq_offset;
@@ -159,7 +158,7 @@ struct intel_guc {struct i915_guc_client *execbuf_client; -DECLARE_BITMAP (doorbell_bitmap,GUC_MAX_DOORBELLS ); +DECLARE_BITMAP (doorbell_bitmap,GUC_NUM_DOORBELLS );uint32_t db_cacheline;/* Cyclic counter mod pagesize */ /* Action status & statistics */
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [01/11 v3] drm/i915/guc: Sanitize GuC client initialization
Quoted 71 lines · Oscar Mateo
Started adding proper teardown to
guc_client_alloc, ended up removing quite a few dead ends where errors communicating with the GuC were silently ignored. There also seemed to be quite a few erronous teardown actions performed in case of an error (ordering wrong).v2:
- Increase function symmetry/proximity (Michal/Daniele)
- Fix__reserve_doorbellaccounting for high priority (Daniele)
- Call__update_doorbell_desc! (Daniele)
- Isolate __guc_{,de}allocate_doorbell (Michal/Daniele)v3:
- "Select" a cacheline is a more accurate verb than "reserve" (Daniele).
- We cannot update & create the doorbell without reserving it first, so move the whole doorbell creation forexecbuf_clientto the submission enable (Oscar).i
- Add a fixme for ignoring possible doorbell destroy errors.Signed-off-by Joonas Lahtinen <[email protected]>Cc Michal Wajdeczko <[email protected]>Cc Arkadiusz Hiler <[email protected]>Cc Tvrtko Ursulin <[email protected]>Cc Chris Wilson <[email protected]>Cc Daniele Ceraolo Spurio <[email protected]>Signed-off-by Oscar Mateo <[email protected]>--- drivers/gpu/drm/i915/i915_debugfs.c | 4 +- drivers/gpu/drm/i915/i915_guc_submission.c | 391 ++++++++++++++++------------- drivers/gpu/drm/i915/intel_guc_fwif.h | 4 +- drivers/gpu/drm/i915/intel_uc.h | 11 +- 4 files changed, 229 insertions(+), 181 deletions(-)drivers/gpu/drm/i915/i915_debugfs.c +2
-2@@ -2469,7 +2469,7 @@ static void i915_guc_client_info(struct seq_file *m,seq_printf (m,"\tPriority %d, GuC ctx index: %u, PD offset 0x%x\n" , client->priority, client->ctx_index, client->proc_desc_offset); -seq_printf (m,"\tDoorbell id %d, offset: 0x%x, cookie 0x%x\n" , +seq_printf (m,"\tDoorbell id %d, offset: 0x%lx, cookie 0x%x\n" , client->doorbell_id, client->doorbell_offset, client->doorbell_cookie);seq_printf (m,"\tWQ size %d, offset: 0x%x, tail %d\n" , client->wq_size, client->wq_offset, client->wq_tail);@@ -2504,7 +2504,7 @@ static int i915_guc_info(struct seq_file *m, void *data) }seq_printf (m,"Doorbell map:\n" ); -seq_printf (m,"\t%*pb\n" ,GUC_MAX_DOORBELLS , guc->doorbell_bitmap); +seq_printf (m,"\t%*pb\n" ,GUC_NUM_DOORBELLS , guc->doorbell_bitmap);seq_printf (m,"Doorbell next cacheline: 0x%x\n\n" , guc->db_cacheline);seq_printf (m,"GuC total action count: %llu\n" , guc->action_count);drivers/gpu/drm/i915/i915_guc_submission.c +9
-0@@ -62,27 +62,73 @@* */ +static inline bool is_high_priority (struct i915_guc_client * client) +{ +return client->priority <=GUC_CTX_PRIORITY_HIGH ; +} + +static int __reserve_doorbell (struct i915_guc_client *client) +{ +unsigned long offset; +unsigned long end;
drivers/gpu/drm/i915/i915_guc_submission.c line 83
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +10
-0+u16 id; + +GEM_BUG_ON (client->doorbell_id !=GUC_DOORBELL_INVALID ); + +/* +* The bitmap tracks which doorbell registers are currently in use. +* It is split into two halves; the first half is used for normal +* priority contexts, the second half for high-priority ones. +* Note that logically higher priorities are numerically less than +* normal ones, so the test below means "is it high-priority?"
The comment about the test doesn't apply anymore so we can either remove it or move it in the is_high_priority() function.
Quoted 6 lines
drivers/gpu/drm/i915/i915_guc_submission.c +6
-0+*/ + + offset =0 ; + end =GUC_NUM_DOORBELLS /2 ; +if (is_high_priority (client)) { + offset = end;
drivers/gpu/drm/i915/i915_guc_submission.c line 99
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +10
-0+ end += offset; + } + + id =find_next_zero_bit (client->guc->doorbell_bitmap, offset, end); +if (id == end) +return -ENOSPC ; + +__set_bit (id, client->guc->doorbell_bitmap); + client->doorbell_id = id; + DRM_DEBUG_DRIVER("client %u (high prio=%s) reserved doorbell %d: \n",
"%d:" -> ": %d" ?
Quoted 80 lines
drivers/gpu/drm/i915/i915_guc_submission.c +22
-24+ client->ctx_index,yesno (is_high_priority (client)), + id); +return 0 ; +} + +static void __unreserve_doorbell (struct i915_guc_client *client) +{ +GEM_BUG_ON (client->doorbell_id ==GUC_DOORBELL_INVALID ); + +__clear_bit (client->doorbell_id, client->guc->doorbell_bitmap); + client->doorbell_id =GUC_DOORBELL_INVALID ; +} +/* * Tell the GuC to allocate or deallocate a specific doorbell */ -static int guc_allocate_doorbell (struct intel_guc *guc, -struct i915_guc_client *client) +static int __guc_allocate_doorbell (struct intel_guc *guc,u32 ctx_index) {u32 action[] = {INTEL_GUC_ACTION_ALLOCATE_DOORBELL , - client->ctx_index + ctx_index };return intel_guc_send (guc, action,ARRAY_SIZE (action)); } -static int guc_release_doorbell (struct intel_guc *guc, -struct i915_guc_client *client) +static int __guc_deallocate_doorbell (struct intel_guc *guc,u32 ctx_index) {u32 action[] = {INTEL_GUC_ACTION_DEALLOCATE_DOORBELL , - client->ctx_index + ctx_index };return intel_guc_send (guc, action,ARRAY_SIZE (action));@@ -95,104 +141,100 @@ static int guc_release_doorbell(struct intel_guc *guc,* client object which contains the page being used for the doorbell */ -static int guc_update_doorbell_id (struct intel_guc *guc, -struct i915_guc_client *client, -u16 new_id) +static int __update_doorbell_desc (struct i915_guc_client *client,u16 new_id) { -struct sg_table *sg = guc->ctx_pool_vma->pages; -void *doorbell_bitmap = guc->doorbell_bitmap; -struct guc_doorbell_info *doorbell; +struct sg_table *sg = client->guc->ctx_pool_vma->pages;struct guc_context_desc desc;size_t len; - doorbell = client->vaddr + client->doorbell_offset; - -if (client->doorbell_id !=GUC_INVALID_DOORBELL_ID && -test_bit (client->doorbell_id, doorbell_bitmap)) { -/* Deactivate the old doorbell */ - doorbell->db_status =GUC_DOORBELL_DISABLED ; - (void )guc_release_doorbell (guc, client); -__clear_bit (client->doorbell_id, doorbell_bitmap); - } -/* Update the GuC's idea of the doorbell ID */ len =sg_pcopy_to_buffer (sg->sgl, sg->nents, &desc,sizeof (desc), -sizeof (desc) * client->ctx_index); +sizeof (desc) * client->ctx_index);if (len !=sizeof (desc))return -EFAULT ; + desc.db_id = new_id; len =sg_pcopy_from_buffer (sg->sgl, sg->nents, &desc,sizeof (desc), -sizeof (desc) * client->ctx_index); +sizeof (desc) * client->ctx_index);if (len !=sizeof (desc))return -EFAULT ;
drivers/gpu/drm/i915/i915_guc_submission.c static int guc_release_doorbell(struct intel_guc *guc, · line 165
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3
-5- client->doorbell_id = new_id; -if (new_id ==GUC_INVALID_DOORBELL_ID ) -return 0 ; +return 0 ; +} -/* Activate the new doorbell */ -__set_bit (new_id, doorbell_bitmap); +static struct guc_doorbell_info *__get_doorbell (struct i915_guc_client *client)
from what I can see this could also be used from guc_ring_doorbell and possibly from guc_ctx_desc_init
drivers/gpu/drm/i915/i915_guc_submission.c static int guc_release_doorbell(struct intel_guc *guc, · line 167
Quoted 2 lines
drivers/gpu/drm/i915/i915_guc_submission.c +2
-0+{ +return client->vaddr + client->doorbell_offset;
Not introduced by this patch, but arithmetic on void pointers doesn't really look clean IMO. Can we cast the pointer to uintptr_t or char* or whatever else is appropriate before incrementing it? I'm happy for it to be done in a follow up.
Quoted 176 lines
drivers/gpu/drm/i915/i915_guc_submission.c +62
-71+} + +static bool has_doorbell (struct i915_guc_client *client) +{ +if (client->doorbell_id ==GUC_DOORBELL_INVALID ) +return false; + +return test_bit (client->doorbell_id, client->guc->doorbell_bitmap); +} + +static int __create_doorbell (struct i915_guc_client *client) +{ +struct guc_doorbell_info *doorbell; +int err; + + doorbell =__get_doorbell (client); doorbell->db_status =GUC_DOORBELL_ENABLED ; doorbell->cookie = client->doorbell_cookie; -return guc_allocate_doorbell (guc, client); + + err =__guc_allocate_doorbell (client->guc, client->ctx_index); +if (err) { + doorbell->db_status =GUC_DOORBELL_DISABLED ; + doorbell->cookie =0 ; + } +return err; } -static void guc_disable_doorbell (struct intel_guc *guc, -struct i915_guc_client *client) +static int __destroy_doorbell (struct i915_guc_client *client) { - (void )guc_update_doorbell_id (guc, client,GUC_INVALID_DOORBELL_ID ); +struct guc_doorbell_info *doorbell; -/* XXX: wait for any interrupts */ -/* XXX: wait for workqueue to drain */ + doorbell =__get_doorbell (client); + doorbell->db_status =GUC_DOORBELL_DISABLED ; + doorbell->cookie =0 ; + +return __guc_deallocate_doorbell (client->guc, client->ctx_index); } -static uint16_t -select_doorbell_register (struct intel_guc *guc,uint32_t priority) +static int destroy_doorbell (struct i915_guc_client *client) { -/* -* The bitmap tracks which doorbell registers are currently in use. -* It is split into two halves; the first half is used for normal -* priority contexts, the second half for high-priority ones. -* Note that logically higher priorities are numerically less than -* normal ones, so the test below means "is it high-priority?" -*/ -const bool hi_pri = (priority <=GUC_CTX_PRIORITY_HIGH ); -const uint16_t half =GUC_MAX_DOORBELLS /2 ; -const uint16_t start = hi_pri ? half :0 ; -const uint16_t end = start + half; -uint16_t id; +int err; - id =find_next_zero_bit (guc->doorbell_bitmap, end, start); -if (id == end) - id =GUC_INVALID_DOORBELL_ID ; +GEM_BUG_ON (!has_doorbell (client)); -DRM_DEBUG_DRIVER ("assigned %s priority doorbell id 0x%x\n" , - hi_pri ?"high" :"normal" , id); +/* XXX: wait for any interrupts */ +/* XXX: wait for workqueue to drain */ -return id; -} + err =__destroy_doorbell (client); +if (err) +return err; -/* -* Select, assign and relase doorbell cachelines -* -* These functions track which doorbell cachelines are in use. -* The data they manipulate is protected by the intel_guc_send lock. -*/ +__update_doorbell_desc (client,GUC_DOORBELL_INVALID ); + +__unreserve_doorbell (client); -static uint32_t select_doorbell_cacheline (struct intel_guc *guc) +return 0 ; +} + +static unsigned long __select_cacheline (struct intel_guc * guc) { -const uint32_t cacheline_size =cache_line_size (); -uint32_t offset; +unsigned long offset;/* Doorbell uses a single cache line within a page */ offset =offset_in_page (guc->db_cacheline);/* Moving to next cache line to reduce contention */ - guc->db_cacheline += cacheline_size; - -DRM_DEBUG_DRIVER ("selected doorbell cacheline 0x%x, next 0x%x, linesize %u\n" , - offset, guc->db_cacheline, cacheline_size); + guc->db_cacheline +=cache_line_size (); +DRM_DEBUG_DRIVER ("reserved cacheline 0x%lx, next 0x%x, linesize %u\n" , + offset, guc->db_cacheline,cache_line_size ());return offset; }@@ -594,93 +636,96 @@ struct i915_vma *intel_guc_allocate_vma(struct intel_guc *guc, u32 size)return vma; } -static void -guc_client_free (struct drm_i915_private *dev_priv, -struct i915_guc_client *client) +static void guc_client_free (struct i915_guc_client *client) { -struct intel_guc *guc = &dev_priv->guc; - -if (!client) -return ; -/* * XXX: wait for any outstanding submissions before freeing memory. * Be sure to drop any locks */ - -if (client->vaddr) { -/* -* If we got as far as setting up a doorbell, make sure we -* shut it down before unmapping & deallocating the memory. -*/ -guc_disable_doorbell (guc, client); - -i915_gem_object_unpin_map (client->vma->obj); - } - +guc_ctx_desc_fini (client->guc, client); +i915_gem_object_unpin_map (client->vma->obj);i915_vma_unpin_and_release (&client->vma); - -if (client->ctx_index !=GUC_INVALID_CTX_ID ) { -guc_ctx_desc_fini (guc, client); -ida_simple_remove (&guc->ctx_ids, client->ctx_index); - } - +ida_simple_remove (&client->guc->ctx_ids, client->ctx_index);kfree (client); }/* Check that a doorbell register is in the expected state */ -static bool guc_doorbell_check (struct intel_guc *guc,uint16_t db_id) +static bool doorbell_ok (struct intel_guc *guc,u16 db_id) {struct drm_i915_private *dev_priv =guc_to_i915 (guc); -i915_reg_t drbreg =GEN8_DRBREGL (db_id); -uint32_t value =I915_READ (drbreg); -bool enabled = (value &GUC_DOORBELL_ENABLED ) !=0 ; -bool expected =test_bit (db_id, guc->doorbell_bitmap); +u32 drbregl; +bool valid; -if (enabled == expected) +GEM_BUG_ON (db_id >=GUC_DOORBELL_INVALID ); + + drbregl =I915_READ (GEN8_DRBREGL (db_id)); + valid = drbregl &GEN8_DRB_VALID ; + +if (test_bit (db_id, guc->doorbell_bitmap) == valid)return true;
drivers/gpu/drm/i915/i915_guc_submission.c struct i915_vma *intel_guc_allocate_vma(struct intel_guc *guc, u32 size) · line 673
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +3
-3-DRM_DEBUG_DRIVER ("Doorbell %d (reg 0x%x) 0x%x, should be %s\n" , - db_id, drbreg.reg, value, - expected ?"active" :"inactive" ); +DRM_DEBUG_DRIVER ("Doorbell %d has unexpected state (0x%x): valid=%s\n" , + db_id, drbregl,yesno (valid));return false; } +static int __reset_doorbell (struct i915_guc_client * client,u16 db_id)
This reset only works if the GuC FW thinks that the doorbell is unassigned, otherwise the ALLOCATE_DOORBELL action will fail. We currently only use it to reset the doorbell HW status after resetting GuC and reloading the FW so it is ok, but we could use a comment to explain the limitation of this reset.
The patch looks functionally ok, so with the minor comments addressed:
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v3 2/11 drm/i915/guc: Keep the ctx_pool_vaddr mapped, for easy access
In reply to Oscar MateoThe GuC descriptor is big in size. If we use a local definition of guc_desc we have a chance to overflow stack, so avoid it.
Also, Chris abhors scatterlists :)
v2: Rebased, helper function to retrieve the context descriptor, s/ctx_pool_vma/ctx_pool/
v3: Zero out guc_context_desc before initialization
--- drivers/gpu/drm/i915/i915_guc_submission.c | 95 +++++++++++++++--------------- drivers/gpu/drm/i915/intel_guc_loader.c | 2 +- drivers/gpu/drm/i915/intel_uc.h | 3 +- 3 files changed, 49 insertions(+), 51 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +46 -49
@@ -134,6 +134,12 @@ static int __guc_deallocate_doorbell(struct intel_guc *guc, u32 ctx_index)return intel_guc_send (guc, action,ARRAY_SIZE (action)); } +static struct guc_context_desc *__get_context_desc (struct i915_guc_client *client) +{ +return client->guc->ctx_pool_vaddr + +sizeof (struct guc_context_desc ) * client->ctx_index; +} +/* * Initialise, update, or clear doorbell data shared with the GuC *
@@ -143,21 +149,11 @@ static int __guc_deallocate_doorbell(struct intel_guc *guc, u32 ctx_index)static int __update_doorbell_desc (struct i915_guc_client *client,u16 new_id) { -struct sg_table *sg = client->guc->ctx_pool_vma->pages; -struct guc_context_desc desc; -size_t len; +struct guc_context_desc *desc;/* Update the GuC's idea of the doorbell ID */ - len =sg_pcopy_to_buffer (sg->sgl, sg->nents, &desc,sizeof (desc), -sizeof (desc) * client->ctx_index); -if (len !=sizeof (desc)) -return -EFAULT ; - - desc.db_id = new_id; - len =sg_pcopy_from_buffer (sg->sgl, sg->nents, &desc,sizeof (desc), -sizeof (desc) * client->ctx_index); -if (len !=sizeof (desc)) -return -EFAULT ; + desc =__get_context_desc (client); + desc->db_id = new_id;return 0 ; }
@@ -272,29 +268,28 @@ static void guc_proc_desc_init(struct intel_guc *guc,* data structures relating to this client (doorbell, process descriptor, * write queue, etc). */ -static void guc_ctx_desc_init (struct intel_guc *guc,struct i915_guc_client *client) {struct drm_i915_private *dev_priv =guc_to_i915 (guc);struct intel_engine_cs *engine;struct i915_gem_context *ctx = client->owner; -struct guc_context_desc desc; -struct sg_table *sg; +struct guc_context_desc *desc;unsigned int tmp;u32 gfx_addr; -memset (&desc,0 ,sizeof (desc)); + desc =__get_context_desc (client); +memset (desc,0 ,sizeof (*desc)); - desc.attribute =GUC_CTX_DESC_ATTR_ACTIVE |GUC_CTX_DESC_ATTR_KERNEL ; - desc.context_id = client->ctx_index; - desc.priority = client->priority; - desc.db_id = client->doorbell_id; + desc->attribute =GUC_CTX_DESC_ATTR_ACTIVE |GUC_CTX_DESC_ATTR_KERNEL ; + desc->context_id = client->ctx_index; + desc->priority = client->priority; + desc->db_id = client->doorbell_id;for_each_engine_masked (engine, dev_priv, client->engines, tmp) {struct intel_context *ce = &ctx->engine[engine->id];uint32_t guc_engine_id = engine->guc_id; -struct guc_execlist_context *lrc = &desc.lrc[guc_engine_id]; +struct guc_execlist_context *lrc = &desc->lrc[guc_engine_id];/* TODO: We have a design issue to be solved here. Only when we * receive the first batch, we know which engine is used by the
@@ -319,50 +314,41 @@ static void guc_ctx_desc_init(struct intel_guc *guc, lrc->ring_next_free_location = lrc->ring_begin; lrc->ring_current_tail_pointer_value =0 ; - desc.engines_used |= (1 << guc_engine_id); + desc->engines_used |= (1 << guc_engine_id); }DRM_DEBUG_DRIVER ("Host engines 0x%x => GuC engines used 0x%x\n" , - client->engines, desc.engines_used); -WARN_ON (desc.engines_used ==0 ); + client->engines, desc->engines_used); +WARN_ON (desc->engines_used ==0 );/* * The doorbell, process descriptor, and workqueue are all parts * of the client object, which the GuC will reference via the GGTT */ gfx_addr =guc_ggtt_offset (client->vma); - desc.db_trigger_phy =sg_dma_address (client->vma->pages->sgl) + + desc->db_trigger_phy =sg_dma_address (client->vma->pages->sgl) + client->doorbell_offset; - desc.db_trigger_cpu = - (uintptr_t )client->vaddr + client->doorbell_offset; - desc.db_trigger_uk = gfx_addr + client->doorbell_offset; - desc.process_desc = gfx_addr + client->proc_desc_offset; - desc.wq_addr = gfx_addr + client->wq_offset; - desc.wq_size = client->wq_size; + desc->db_trigger_cpu = (uintptr_t )client->vaddr + + client->doorbell_offset; + desc->db_trigger_uk = gfx_addr + client->doorbell_offset; + desc->process_desc = gfx_addr + client->proc_desc_offset; + desc->wq_addr = gfx_addr + client->wq_offset; + desc->wq_size = client->wq_size;/* * XXX: Take LRCs from an existing context if this is not an * IsKMDCreatedContext client */ - desc.desc_private = (uintptr_t )client; - -/* Pool context is pinned already */ - sg = guc->ctx_pool_vma->pages; -sg_pcopy_from_buffer (sg->sgl, sg->nents, &desc,sizeof (desc), -sizeof (desc) * client->ctx_index); + desc->desc_private = (uintptr_t )client; }static void guc_ctx_desc_fini (struct intel_guc *guc,struct i915_guc_client *client) { -struct guc_context_desc desc; -struct sg_table *sg; +struct guc_context_desc *desc; -memset (&desc,0 ,sizeof (desc)); - - sg = guc->ctx_pool_vma->pages; -sg_pcopy_from_buffer (sg->sgl, sg->nents, &desc,sizeof (desc), -sizeof (desc) * client->ctx_index); + desc =__get_context_desc (client); +memset (desc,0 ,sizeof (*desc)); }/**
@@ -929,6 +915,7 @@ int i915_guc_submission_init(struct drm_i915_private *dev_priv)const size_t gemsize =round_up (poolsize,PAGE_SIZE );struct intel_guc *guc = &dev_priv->guc;struct i915_vma *vma; +void *vaddr;if (!HAS_GUC_SCHED (dev_priv))return 0 ;
@@ -940,14 +927,21 @@ int i915_guc_submission_init(struct drm_i915_private *dev_priv)if (!i915.enable_guc_submission)return 0 ;/* not enabled */ -if (guc->ctx_pool_vma) +if (guc->ctx_pool)return 0 ;/* already allocated */ vma =intel_guc_allocate_vma (guc, gemsize);if (IS_ERR (vma))return PTR_ERR (vma); - guc->ctx_pool_vma = vma; + guc->ctx_pool = vma; + + vaddr =i915_gem_object_pin_map (vma->obj,I915_MAP_WB ); +if (IS_ERR (vaddr)) + goto err; + + guc->ctx_pool_vaddr = vaddr; +ida_init (&guc->ctx_ids);intel_guc_log_create (guc);guc_addon_create (guc);
@@ -1071,9 +1065,12 @@ void i915_guc_submission_fini(struct drm_i915_private *dev_priv)i915_vma_unpin_and_release (&guc->ads_vma);i915_vma_unpin_and_release (&guc->log.vma); -if (guc->ctx_pool_vma) +if (guc->ctx_pool_vaddr) {ida_destroy (&guc->ctx_ids); -i915_vma_unpin_and_release (&guc->ctx_pool_vma); +i915_gem_object_unpin_map (guc->ctx_pool->obj); + } + +i915_vma_unpin_and_release (&guc->ctx_pool); }/**
drivers/gpu/drm/i915/intel_guc_loader.c +1 -1
@@ -175,7 +175,7 @@ static void guc_params_init(struct drm_i915_private *dev_priv)/* If GuC submission is enabled, set up additional parameters here */ if (i915.enable_guc_submission) { -u32 pgs =guc_ggtt_offset (dev_priv->guc.ctx_pool_vma); +u32 pgs =guc_ggtt_offset (dev_priv->guc.ctx_pool);u32 ctx_in_16 =GUC_MAX_GPU_CONTEXTS /16 ; pgs >>=PAGE_SHIFT ;
drivers/gpu/drm/i915/intel_uc.h +2 -1
@@ -153,7 +153,8 @@ struct intel_guc {bool interrupts_enabled;struct i915_vma *ads_vma; -struct i915_vma *ctx_pool_vma; +struct i915_vma *ctx_pool; +void *ctx_pool_vaddr;struct ida ctx_ids;struct i915_guc_client *execbuf_client;
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v3 3/11 drm/i915/guc: Add onion teardown to the GuC setup
In reply to Oscar MateoStarting with intel_guc_loader, down to intel_guc_submission and finally to intel_guc_log.
v2:
- Null execbuf client outside guc_client_free (Daniele)
- Assert if things try to get allocated twice (Daniele/Joonas)
- Null guc->log.buf_addr when destroyed (Daniele)
- Newline between returning success and error labels (Joonas)
- Remove some unnecessary comments (Joonas)
- Keep guc_log_create_extras naming convention (Joonas)
- Helper function guc_log_has_extras (Joonas)
- No need for separate relay_channel create/destroy. It's just another extra.
- No need to nullify guc->log.flush_wq when destroyed (Joonas)
- Hoist the check for has_extras out of guc_log_create_extras (Joonas)
- Try to do i915_guc_log_register/unregister calls (kind of) symmetric (Daniele)
- Make sure initel_guc_fini is not called before init is ever called (Daniele)
v3:
- Remove unnecessary parenthesis (Joonas)
- Check for logs enabled on debugfs registration
- Rebase on top of Tvrtko's "Fix request re-submission after reset"
--- drivers/gpu/drm/i915/i915_drv.c | 7 +- drivers/gpu/drm/i915/i915_guc_submission.c | 94 ++++---- drivers/gpu/drm/i915/intel_guc_loader.c | 32 +-- drivers/gpu/drm/i915/intel_guc_log.c | 362 ++++++++++++++--------------- drivers/gpu/drm/i915/intel_uc.h | 8 +- 5 files changed, 261 insertions(+), 242 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +5 -2
@@ -549,6 +549,8 @@ static bool i915_switcheroo_can_switch(struct pci_dev *pdev)static void i915_gem_fini (struct drm_i915_private *dev_priv) {mutex_lock (&dev_priv->drm.struct_mutex); +if (i915.enable_guc_submission) +intel_guc_cleanup (dev_priv);i915_gem_cleanup_engines (dev_priv);i915_gem_context_fini (dev_priv);mutex_unlock (&dev_priv->drm.struct_mutex);
@@ -612,7 +614,7 @@ static int i915_load_modeset_init(struct drm_device *dev) ret =i915_gem_init (dev_priv);if (ret) - goto cleanup_irq; + goto cleanup_uc;intel_modeset_gem_init (dev);
@@ -634,9 +636,10 @@ static int i915_load_modeset_init(struct drm_device *dev)if (i915_gem_suspend (dev_priv))DRM_ERROR ("failed to idle hardware; continuing to unload!\n" );i915_gem_fini (dev_priv); -cleanup_irq: +cleanup_uc:intel_guc_fini (dev_priv);intel_huc_fini (dev_priv); +cleanup_irq:drm_irq_uninstall (dev);intel_teardown_gmbus (dev_priv); cleanup_csr:
drivers/gpu/drm/i915/i915_guc_submission.c +54 -40
164 lines not on this page · raw message
drivers/gpu/drm/i915/intel_guc_loader.c +17 -15
@@ -443,7 +443,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv) err =i915_guc_submission_init (dev_priv);if (err) - goto fail; + goto err_guc;/* * WaEnableuKernelHeaderValidFix:skl,bxt
@@ -458,7 +458,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)*/ err =guc_hw_reset (dev_priv);if (err) - goto fail; + goto err_submission;intel_huc_load (dev_priv); err =guc_ucode_xfer (dev_priv);
@@ -466,7 +466,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)break ;if (--retries ==0 ) - goto fail; + goto err_submission;DRM_INFO ("GuC fw load failed: %d; will reset and " "retry %d more time(s)\n" , err, retries);
@@ -482,7 +482,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv) err =i915_guc_submission_enable (dev_priv);if (err) - goto fail; + goto err_interrupts; } DRM_INFO("GuC %s (firmware %s [version %u.%u])\n",
@@ -492,15 +492,16 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)return 0 ; +err_interrupts: +gen9_disable_guc_interrupts (dev_priv); +err_submission: +i915_guc_submission_fini (dev_priv); +err_guc: +i915_ggtt_disable_guc (dev_priv); fail:if (guc_fw->load_status ==INTEL_UC_FIRMWARE_PENDING ) guc_fw->load_status =INTEL_UC_FIRMWARE_FAIL ; -guc_interrupts_release (dev_priv); -i915_guc_submission_disable (dev_priv); -i915_guc_submission_fini (dev_priv); -i915_ggtt_disable_guc (dev_priv); -/* * We've failed to load the firmware :( *
@@ -540,6 +541,13 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)return ret; } +void intel_guc_cleanup (struct drm_i915_private *dev_priv) +{ +gen9_disable_guc_interrupts (dev_priv); +i915_guc_submission_fini (dev_priv); +i915_ggtt_disable_guc (dev_priv); +} +void intel_uc_fw_fetch (struct drm_i915_private *dev_priv,struct intel_uc_fw *uc_fw) {
@@ -745,12 +753,6 @@ void intel_guc_fini(struct drm_i915_private *dev_priv)struct intel_uc_fw *guc_fw = &dev_priv->guc.fw;struct drm_i915_gem_object *obj; -mutex_lock (&dev_priv->drm.struct_mutex); -guc_interrupts_release (dev_priv); -i915_guc_submission_disable (dev_priv); -i915_guc_submission_fini (dev_priv); -mutex_unlock (&dev_priv->drm.struct_mutex); - obj =fetch_and_zero (&guc_fw->obj);if (obj)i915_gem_object_put (obj);
drivers/gpu/drm/i915/intel_guc_log.c +180 -182
459 lines not on this page · raw message
drivers/gpu/drm/i915/intel_uc.h +5 -3
@@ -190,8 +190,9 @@ struct intel_huc {/* intel_guc_loader.c */ extern void intel_guc_init (struct drm_i915_private *dev_priv); -extern int intel_guc_setup (struct drm_i915_private *dev_priv);extern void intel_guc_fini (struct drm_i915_private *dev_priv); +extern int intel_guc_setup (struct drm_i915_private *dev_priv); +extern void intel_guc_cleanup (struct drm_i915_private *dev_priv);extern const char *intel_uc_fw_status_repr (enum intel_uc_fw_status status);extern int intel_guc_suspend (struct drm_i915_private *dev_priv);extern int intel_guc_resume (struct drm_i915_private *dev_priv);
@@ -209,10 +210,11 @@ void intel_uc_fw_fetch(struct drm_i915_private *dev_priv,struct i915_vma *intel_guc_allocate_vma (struct intel_guc *guc,u32 size);/* intel_guc_log.c */ -void intel_guc_log_create (struct intel_guc *guc); +int intel_guc_log_create (struct intel_guc *guc); +void intel_guc_log_destroy (struct intel_guc *guc); +int i915_guc_log_control (struct drm_i915_private *dev_priv,u64 control_val);void i915_guc_log_register (struct drm_i915_private *dev_priv);void i915_guc_log_unregister (struct drm_i915_private *dev_priv); -int i915_guc_log_control (struct drm_i915_private *dev_priv,u64 control_val);static inline u32 guc_ggtt_offset (struct i915_vma *vma) {
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [03/11 v3] drm/i915/guc: Add onion teardown to the GuC setup
Quoted 33 lines · Oscar Mateo
Starting with
intel_guc_loader, down tointel_guc_submissionand finally tointel_guc_log.v2:
- Null execbuf client outsideguc_client_free(Daniele)
- Assert if things try to get allocated twice (Daniele/Joonas)
- Nullguc->log.buf_addrwhen destroyed (Daniele)
- Newline between returning success and error labels (Joonas)
- Remove some unnecessary comments (Joonas)
- Keepguc_log_create_extrasnaming convention (Joonas)
- Helper functionguc_log_has_extras(Joonas)
- No need for separaterelay_channelcreate/destroy. It's just another extra.
- No need to nullifyguc->log.flush_wqwhen destroyed (Joonas)
- Hoist the check forhas_extrasout ofguc_log_create_extras(Joonas)
- Try to do i915_guc_log_register/unregister calls (kind of) symmetric (Daniele)
- Make sureinitel_guc_finiis not called before init is ever called (Daniele)v3:
- Remove unnecessary parenthesis (Joonas)
- Check for logs enabled on debugfs registration
- Rebase on top of Tvrtko's "Fix request re-submission after reset"Cc Daniele Ceraolo Spurio <[email protected]>Cc Joonas Lahtinen <[email protected]>Signed-off-by Oscar Mateo <[email protected]>--- drivers/gpu/drm/i915/i915_drv.c | 7 +- drivers/gpu/drm/i915/i915_guc_submission.c | 94 ++++---- drivers/gpu/drm/i915/intel_guc_loader.c | 32 +-- drivers/gpu/drm/i915/intel_guc_log.c | 362 ++++++++++++++--------------- drivers/gpu/drm/i915/intel_uc.h | 8 +- 5 files changed, 261 insertions(+), 242 deletions(-)
<SNIP>
Quoted 41 lines
drivers/gpu/drm/i915/intel_guc_loader.c +4
-4@@ -443,7 +443,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv) err =i915_guc_submission_init (dev_priv);if (err) - goto fail; + goto err_guc;/* * WaEnableuKernelHeaderValidFix:skl,bxt @@ -458,7 +458,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)*/ err =guc_hw_reset (dev_priv);if (err) - goto fail; + goto err_submission;intel_huc_load (dev_priv); err =guc_ucode_xfer (dev_priv);@@ -466,7 +466,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)break ;if (--retries ==0 ) - goto fail; + goto err_submission;DRM_INFO ("GuC fw load failed: %d; will reset and " "retry %d more time(s)\n" , err, retries);@@ -482,7 +482,7 @@ int intel_guc_setup(struct drm_i915_private *dev_priv) err =i915_guc_submission_enable (dev_priv);if (err) - goto fail; + goto err_interrupts; } DRM_INFO("GuC %s (firmware %s [version %u.%u])\n",
drivers/gpu/drm/i915/intel_guc_loader.c int intel_guc_setup(struct drm_i915_private *dev_priv) · line 496
Quoted 5 lines
drivers/gpu/drm/i915/intel_guc_loader.c +2
-0@@ -492,15 +492,16 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)return 0 ; +err_interrupts: +gen9_disable_guc_interrupts (dev_priv);
It's no longer called from this place as of:
commit 7762ebb9a455db18eef5c366da5496fb38429d56
Author: Sagar Arun Kamble <[email protected]>
drm/i915/guc: Release GuC interrupts in i915_guc_submission_disable
Also, I did similar thing in my GuC series (for guc_loader.c only). It should be easy to rebase one on top of the other depending which will get there first.
Othere changes here LGTM
Signature · 2 lines
Cheers, Arek
Quoted 17 lines
drivers/gpu/drm/i915/intel_guc_loader.c +4
-5+err_submission: +i915_guc_submission_fini (dev_priv); +err_guc: +i915_ggtt_disable_guc (dev_priv); fail:if (guc_fw->load_status ==INTEL_UC_FIRMWARE_PENDING ) guc_fw->load_status =INTEL_UC_FIRMWARE_FAIL ; -guc_interrupts_release (dev_priv); -i915_guc_submission_disable (dev_priv); -i915_guc_submission_fini (dev_priv); -i915_ggtt_disable_guc (dev_priv); -/* * We've failed to load the firmware :( *
drivers/gpu/drm/i915/intel_guc_loader.c int intel_guc_setup(struct drm_i915_private *dev_priv) · line 543
Quoted 3 lines
drivers/gpu/drm/i915/intel_guc_loader.c +0
-0@@ -540,6 +541,13 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)return ret; }
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [03/11 v3] drm/i915/guc: Add onion teardown to the GuC setup
In reply to Oscar MateoQuoted 17 lines · Oscar Mateo
Starting with
intel_guc_loader, down tointel_guc_submissionand finally tointel_guc_log.v2:
- Null execbuf client outsideguc_client_free(Daniele)
- Assert if things try to get allocated twice (Daniele/Joonas)
- Nullguc->log.buf_addrwhen destroyed (Daniele)
- Newline between returning success and error labels (Joonas)
- Remove some unnecessary comments (Joonas)
- Keepguc_log_create_extrasnaming convention (Joonas)
- Helper functionguc_log_has_extras(Joonas)
- No need for separaterelay_channelcreate/destroy. It's just another extra.
- No need to nullifyguc->log.flush_wqwhen destroyed (Joonas)
- Hoist the check forhas_extrasout ofguc_log_create_extras(Joonas)
- Try to do i915_guc_log_register/unregister calls (kind of) symmetric (Daniele)
- Make sureinitel_guc_finiis not called before init is ever called (Daniele)
Quoted 8 lines
v3:
- Remove unnecessary parenthesis (Joonas)
- Check for logs enabled on debugfs registration
- Rebase on top of Tvrtko's "Fix request re-submission after reset"Cc Daniele Ceraolo Spurio <[email protected]>Cc Joonas Lahtinen <[email protected]>Signed-off-by Oscar Mateo <[email protected]>
Drop comments around the interrupt enabling/disabling that log is currently the only interrupt user.
Regards, Joonas
Signature · 7 lines
Joonas Lahtinen Open Source Technology Center Intel Corporation _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 4/11 drm/i915/guc: s/ads_vma/addon
In reply to Oscar MateoThis vma contains much more than just the additional data struct (ads) and since we were already using the word "addon" as an object in guc_addon_create, make it the preffered one. No need for the vma suffix either, as that dependency is given by the type.
While at it, add an explanation of what things go inside the addon object.
v2: Move the explanation to the header of the file (Joonas).
--- drivers/gpu/drm/i915/i915_guc_submission.c | 14 ++++++++++---- drivers/gpu/drm/i915/intel_guc_loader.c | 4 ++-- drivers/gpu/drm/i915/intel_uc.h | 2 +- 3 files changed, 13 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +10 -4
@@ -60,6 +60,13 @@* ELSP context descriptor dword into Work Item. * See guc_wq_item_append() * +* Addon object: +* The additional data struct (ADS) has pointers for different buffers used by +* the GuC. One single gem object, the "addon" object, contains the ADS itself\ +* (guc_ads), the scheduling policies (guc_policies), a structure describing +* a collection of register sets (guc_mmio_reg_state) and some extra +* pages for the GuC to save its internal state for sleep. +* */ static inline bool is_high_priority (struct i915_guc_client * client)
@@ -846,9 +853,8 @@ static int guc_addon_create(struct intel_guc *guc)struct page *page;u32 size; -GEM_BUG_ON (guc->ads_vma); +GEM_BUG_ON (guc->addon); -/* The ads obj includes the struct itself and buffers passed to GuC */ size =sizeof (struct guc_ads ) +sizeof (struct guc_policies ) +sizeof (struct guc_mmio_reg_state ) +GUC_S3_SAVE_SPACE_PAGES *PAGE_SIZE ;
@@ -857,7 +863,7 @@ static int guc_addon_create(struct intel_guc *guc)if (IS_ERR (vma))return PTR_ERR (vma); - guc->ads_vma = vma; + guc->addon = vma; page =i915_vma_first_page (vma); ads =kmap (page);
@@ -906,7 +912,7 @@ static int guc_addon_create(struct intel_guc *guc)static void guc_addon_destroy (struct intel_guc *guc) { -i915_vma_unpin_and_release (&guc->ads_vma); +i915_vma_unpin_and_release (&guc->addon); }/*
drivers/gpu/drm/i915/intel_guc_loader.c +2 -2
@@ -167,8 +167,8 @@ static void guc_params_init(struct drm_i915_private *dev_priv) }else params[GUC_CTL_DEBUG ] =GUC_LOG_DISABLED ; -if (guc->ads_vma) { -u32 ads =guc_ggtt_offset (guc->ads_vma) >>PAGE_SHIFT ; +if (guc->addon) { +u32 ads =guc_ggtt_offset (guc->addon) >>PAGE_SHIFT ; params[GUC_CTL_DEBUG ] |= ads <<GUC_ADS_ADDR_SHIFT ; params[GUC_CTL_DEBUG ] |=GUC_ADS_ENABLED ; }
drivers/gpu/drm/i915/intel_uc.h +1 -1
@@ -152,7 +152,7 @@ struct intel_guc {/* intel_guc_recv interrupt related state */ bool interrupts_enabled; -struct i915_vma *ads_vma; +struct i915_vma *addon;struct i915_vma *ctx_pool;void *ctx_pool_vaddr;struct ida ctx_ids;
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 5/11 drm/i915/guc: Break out the GuC log extras into their own "runtime" struct
In reply to Oscar MateoWhen initializing the GuC log struct, there is an object we need to allocate always, since the GuC needs its address at fw load time. The rest is only needed during runtime, in the sense that we only create if we actually enable GuC logging. Make that distinction explicit by subdividing further the intel_guc_log struct.
v2: Call the new struct "runtime", instead of "extras" (Joonas).
--- drivers/gpu/drm/i915/i915_irq.c | 4 +-- drivers/gpu/drm/i915/intel_guc_log.c | 70 ++++++++++++++++++------------------ drivers/gpu/drm/i915/intel_uc.h | 12 ++++--- 3 files changed, 44 insertions(+), 42 deletions(-)
drivers/gpu/drm/i915/i915_irq.c +2 -2
@@ -1740,8 +1740,8 @@ static void gen9_guc_irq_handler(struct drm_i915_private *dev_priv, u32 gt_iir)I915_WRITE (SOFT_SCRATCH (15 ), msg & ~flush);/* Handle flush interrupt in bottom half */ -queue_work (dev_priv->guc.log.flush_wq, - &dev_priv->guc.log.flush_work); +queue_work (dev_priv->guc.log.runtime.flush_wq, + &dev_priv->guc.log.runtime.flush_work); dev_priv->guc.log.flush_interrupt_count++; }else {
drivers/gpu/drm/i915/intel_guc_log.c +35 -35
201 lines not on this page · raw message
drivers/gpu/drm/i915/intel_uc.h +7 -5
@@ -132,11 +132,13 @@ struct intel_uc_fw {struct intel_guc_log {uint32_t flags;struct i915_vma *vma; -void *buf_addr; -struct workqueue_struct *flush_wq; -struct work_struct flush_work; -struct rchan *relay_chan; - +/* The runtime stuff gets created only when GuC logging gets enabled */ +struct { +void *buf_addr; +struct workqueue_struct *flush_wq; +struct work_struct flush_work; +struct rchan *relay_chan; + } runtime;/* logging related stats */ u32 capture_miss_count;u32 flush_interrupt_count;
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [05/11 v2] drm/i915/guc: Break out the GuC log extras into their own "runtime" struct
Quoted 12 lines · Oscar Mateo
When initializing the GuC log struct, there is an object we need to allocate always, since the GuC needs its address at fw load time. The rest is only needed during runtime, in the sense that we only create if we actually enable GuC logging. Make that distinction explicit by subdividing further the
intel_guc_logstruct.v2: Call the new struct "runtime", instead of "extras" (Joonas).
Cc Daniele Ceraolo Spurio <[email protected]>Cc Joonas Lahtinen <[email protected]>Reviewed-by Michal Wajdeczko <[email protected]>Signed-off-by Oscar Mateo <[email protected]>
<SNIP>
drivers/gpu/drm/i915/intel_guc_log.c static int guc_log_create_extras(struct intel_guc *guc) · line 420
Quoted 7 lines
drivers/gpu/drm/i915/intel_guc_log.c +1
-1@@ -416,9 +416,9 @@ static int guc_log_create_extras(struct intel_guc *guc)* or scheduled later on resume. This way the handling of work * item can be kept same between system suspend & rpm suspend. */ - guc->log.flush_wq =alloc_ordered_workqueue ("i915-guc_log" , + guc->log.runtime.flush_wq =alloc_ordered_workqueue ("i915-guc_log" , WQ_HIGHPRI | WQ_FREEZABLE);
Check indent here.
I see this is as a major improvement in readability;
In the future, we might even refactor into guc_log_enable / guc_log_disable :)
Regards, Joonas
Signature · 7 lines
Joonas Lahtinen Open Source Technology Center Intel Corporation _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 7/11 drm/i915/guc: Improve the GuC documentation & comments about proxy submissions
In reply to Oscar MateoWhile at it, fix a typo (s/ring_lcra/ring_lrca) and improve the naming of one firware interface field (s/ring_tail/submit_element_info, since it can contain more than just the ring tail).
No change in functionality.
--- drivers/gpu/drm/i915/i915_guc_submission.c | 42 +++++++++++++++++++----------- drivers/gpu/drm/i915/intel_guc_fwif.h | 4 +-- 2 files changed, 29 insertions(+), 17 deletions(-)
drivers/gpu/drm/i915/i915_guc_submission.c +27 -15
@@ -28,16 +28,25 @@/** * DOC: GuC-based command submission * -* i915_guc_client: -* We use the term client to avoid confusion with contexts. A i915_guc_client is -* equivalent to GuC object guc_context_desc. This context descriptor is -* allocated from a pool of 1024 entries. Kernel driver will allocate doorbell -* and workqueue for it. Also the process descriptor (guc_process_desc), which -* is mapped to client space. So the client can write Work Item then ring the -* doorbell. +* GuC client: +* A i915_guc_client refers to a unique user of the GuC. Currently, there is +* only one of these (the execbuf_client) and this one is charged with all +* submissions to the GuC. This struct is the owner of a doorbell, a process +* descriptor and a workqueue (all of them inside a single gem object that +* contains all required pages for these elements). * -* To simplify the implementation, we allocate one gem object that contains all -* pages for doorbell, process descriptor and workqueue. +* GuC context descriptor: +* During initialization, the driver allocates a static pool of 1024 such +* descriptors, and shares them with the GuC. +* Currently, there exists a 1:1 mapping between a i915_guc_client and a +* guc_context_desc (via the client's context_index), so effectively only +* one guc_context_desc gets used. This context descriptor lets the GuC know +* about the doorbell, workqueue and process descriptor. Theoretically, it also +* lets the GuC know about our HW contexts (Context ID, etc...), but we actually +* employ a kind of submission where the GuC uses the LRCA sent via the work +* item instead (the single guc_context_desc associated to execbuf client +* contains information about the default kernel context only, but this is +* essentially unused). This is called a "proxy" submission. * * The Scratch registers: * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes
@@ -308,10 +317,17 @@ static void guc_ctx_desc_init(struct intel_guc *guc,if (!ce->state)break ;/* XXX: continue? */ +/* +* XXX: When this is a GUC_CTX_DESC_ATTR_KERNEL client (proxy +* submission or, in other words, not using a direct submission +* model) the KMD's LRCA is not used for any work submission. +* Instead, the GuC uses the LRCA of the user mode context (see +* guc_wq_item_append below). +*/ lrc->context_desc =lower_32_bits (ce->lrc_desc);/* The state page is after PPHWSP */ - lrc->ring_lcra = + lrc->ring_lrca =guc_ggtt_offset (ce->state) +LRC_STATE_PN *PAGE_SIZE ; lrc->context_id = (client->ctx_index <<GUC_ELC_CTXID_OFFSET ) | (guc_engine_id <<GUC_ELC_ENGINE_OFFSET );
@@ -342,10 +358,6 @@ static void guc_ctx_desc_init(struct intel_guc *guc, desc->wq_addr = gfx_addr + client->wq_offset; desc->wq_size = client->wq_size; -/* -* XXX: Take LRCs from an existing context if this is not an -* IsKMDCreatedContext client -*/ desc->desc_private = (uintptr_t )client; }
@@ -469,7 +481,7 @@ static void guc_wq_item_append(struct i915_guc_client *client,/* The GuC wants only the low-order word of the context descriptor */ wqi->context_desc = (u32 )intel_lr_context_descriptor (rq->ctx, engine); - wqi->ring_tail = tail <<WQ_RING_TAIL_SHIFT ; + wqi->submit_element_info = tail <<WQ_RING_TAIL_SHIFT ; wqi->fence_id = rq->global_seqno; }
drivers/gpu/drm/i915/intel_guc_fwif.h +2 -2
@@ -251,7 +251,7 @@ struct guc_doorbell_info {struct guc_wq_item {u32 header;u32 context_desc; -u32 ring_tail; +u32 submit_element_info;u32 fence_id; } __packed;
@@ -278,7 +278,7 @@ struct guc_execlist_context {u32 context_desc;u32 context_id;u32 ring_status; -u32 ring_lcra; +u32 ring_lrca;u32 ring_begin;u32 ring_end;u32 ring_next_free_location;
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [07/11] drm/i915/guc: Improve the GuC documentation & comments about proxy submissions
Quoted 19 lines · Oscar Mateo
While at it, fix a typo (s/ring_lcra/ring_lrca) and improve the naming of one firware interface field (s/ring_tail/submit_element_info, since it can contain more than just the ring tail).
No change in functionality.
Signed-off-by Oscar Mateo <[email protected]>--- drivers/gpu/drm/i915/i915_guc_submission.c | 42 +++++++++++++++++++----------- drivers/gpu/drm/i915/intel_guc_fwif.h | 4 +-- 2 files changed, 29 insertions(+), 17 deletions(-)drivers/gpu/drm/i915/i915_guc_submission.c +0
-0@@ -28,16 +28,25 @@/** * DOC: GuC-based command submission
drivers/gpu/drm/i915/i915_guc_submission.c line 32
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +2
-7* -* i915_guc_client: -* We use the term client to avoid confusion with contexts. A i915_guc_client is -* equivalent to GuC object guc_context_desc. This context descriptor is -* allocated from a pool of 1024 entries. Kernel driver will allocate doorbell -* and workqueue for it. Also the process descriptor (guc_process_desc), which -* is mapped to client space. So the client can write Work Item then ring the -* doorbell. +* GuC client: +* A i915_guc_client refers to a unique user of the GuC. Currently, there is
After re-reading this I think that matching a client with a "unique user" is slightly misleading because i915 can have multiple i915_guc_client objects but it still counts as a single user. Maybe we could define the client as a "submission path" through GuC (a set of work-queue, doorbell, priority) or something similar.
It might be also worth keeping the statement about the term "client" being introduced by us to reduce confusion.
Thanks,
Daniele
Quoted 76 lines
drivers/gpu/drm/i915/i915_guc_submission.c +25
-8+* only one of these (the execbuf_client) and this one is charged with all +* submissions to the GuC. This struct is the owner of a doorbell, a process +* descriptor and a workqueue (all of them inside a single gem object that +* contains all required pages for these elements). * -* To simplify the implementation, we allocate one gem object that contains all -* pages for doorbell, process descriptor and workqueue. +* GuC context descriptor: +* During initialization, the driver allocates a static pool of 1024 such +* descriptors, and shares them with the GuC. +* Currently, there exists a 1:1 mapping between a i915_guc_client and a +* guc_context_desc (via the client's context_index), so effectively only +* one guc_context_desc gets used. This context descriptor lets the GuC know +* about the doorbell, workqueue and process descriptor. Theoretically, it also +* lets the GuC know about our HW contexts (Context ID, etc...), but we actually +* employ a kind of submission where the GuC uses the LRCA sent via the work +* item instead (the single guc_context_desc associated to execbuf client +* contains information about the default kernel context only, but this is +* essentially unused). This is called a "proxy" submission. * * The Scratch registers: * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes @@ -308,10 +317,17 @@ static void guc_ctx_desc_init(struct intel_guc *guc,if (!ce->state)break ;/* XXX: continue? */ +/* +* XXX: When this is a GUC_CTX_DESC_ATTR_KERNEL client (proxy +* submission or, in other words, not using a direct submission +* model) the KMD's LRCA is not used for any work submission. +* Instead, the GuC uses the LRCA of the user mode context (see +* guc_wq_item_append below). +*/ lrc->context_desc =lower_32_bits (ce->lrc_desc);/* The state page is after PPHWSP */ - lrc->ring_lcra = + lrc->ring_lrca =guc_ggtt_offset (ce->state) +LRC_STATE_PN *PAGE_SIZE ; lrc->context_id = (client->ctx_index <<GUC_ELC_CTXID_OFFSET ) | (guc_engine_id <<GUC_ELC_ENGINE_OFFSET );@@ -342,10 +358,6 @@ static void guc_ctx_desc_init(struct intel_guc *guc, desc->wq_addr = gfx_addr + client->wq_offset; desc->wq_size = client->wq_size; -/* -* XXX: Take LRCs from an existing context if this is not an -* IsKMDCreatedContext client -*/ desc->desc_private = (uintptr_t )client; }@@ -469,7 +481,7 @@ static void guc_wq_item_append(struct i915_guc_client *client,/* The GuC wants only the low-order word of the context descriptor */ wqi->context_desc = (u32 )intel_lr_context_descriptor (rq->ctx, engine); - wqi->ring_tail = tail <<WQ_RING_TAIL_SHIFT ; + wqi->submit_element_info = tail <<WQ_RING_TAIL_SHIFT ; wqi->fence_id = rq->global_seqno; }drivers/gpu/drm/i915/intel_guc_fwif.h +1
-1@@ -251,7 +251,7 @@ struct guc_doorbell_info {struct guc_wq_item {u32 header;u32 context_desc; -u32 ring_tail; +u32 submit_element_info;u32 fence_id; } __packed;
drivers/gpu/drm/i915/intel_guc_fwif.h struct guc_execlist_context { · line 284
Quoted 8 lines
drivers/gpu/drm/i915/intel_guc_fwif.h +1
-1@@ -278,7 +278,7 @@ struct guc_execlist_context {u32 context_desc;u32 context_id;u32 ring_status; -u32 ring_lcra; +u32 ring_lrca;u32 ring_begin;u32 ring_end;u32 ring_next_free_location;
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 8/11 drm/i915/guc: Wait for doorbell to be inactive before deallocating
In reply to Oscar MateoDoorbell release flow requires that we wait for GEN8_DRB_VALID bit to go to zero after updating db_status before we call the GuC to release the doorbell.
Kudos to Daniele for finding this out.
--- drivers/gpu/drm/i915/i915_guc_submission.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
drivers/gpu/drm/i915/i915_guc_submission.c +10 -0
@@ -206,12 +206,22 @@ static int __create_doorbell(struct i915_guc_client *client)static int __destroy_doorbell (struct i915_guc_client *client) { +struct drm_i915_private *dev_priv =guc_to_i915 (client->guc);struct guc_doorbell_info *doorbell; +u16 db_id = client->doorbell_id; + +GEM_BUG_ON (db_id >=GUC_DOORBELL_INVALID ); doorbell =__get_doorbell (client); doorbell->db_status =GUC_DOORBELL_DISABLED ; doorbell->cookie =0 ; +/* Doorbell release flow requires that we wait for GEN8_DRB_VALID bit +* to go to zero after updating db_status before we call the GuC to +* release the doorbell */ +if (wait_for_us (!(I915_READ (GEN8_DRBREGL (db_id)) &GEN8_DRB_VALID ),10 )) +DRM_ERROR ("Doorbell never became invalid after disable\n" ); +return __guc_deallocate_doorbell (client->guc, client->ctx_index); }
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [08/11] drm/i915/guc: Wait for doorbell to be inactive before deallocating
Quoted 3 lines · Oscar Mateo
Doorbell release flow requires that we wait for
GEN8_DRB_VALIDbit to go to zero after updatingdb_statusbefore we call the GuC to release the doorbell.
Does this fix some of the '110' errors?
-Chris
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [08/11] drm/i915/guc: Wait for doorbell to be inactive before deallocating
Quoted 6 lines · Chris Wilson
Quoted 3 lines · Oscar Mateo
Doorbell release flow requires that we wait for
GEN8_DRB_VALIDbit to go to zero after updatingdb_statusbefore we call the GuC to release the doorbell.Does this fix some of the '110' errors?
-Chris
No, AFAICT this doesn't fix anything (but it's required nonetheless).
The 110 errors (ENTER_S_STATE and DEALLOCATE_DOORBELL failed calls to GuC) are cause by us resetting the GuC right before we ask it to do something. E.g.: in the suspend path, we do i915_gem_suspend()->i915_gem_sanitize()->intel_gpu_reset() and then intel_guc_suspend() inmediately after. After the sanitize, the GuC is in reset state and without a firmware loaded, so asking it to suspend is pretty useless. Same thing when trying to orderly shutdown the doorbells right after reset (the GuC has no idea of what doorbell we are talking about).
For now, the simple fix would be to keep doing the reset and not ask the GuC to do anything after. But sooner or later, we won't be able to get away with this. E.g.: if we enable direct submission, the GuC is the only one that will know about all the requests in-flight, and therefore the only one that can decide when to suspend. Also, surely loading the GuC firmware every time is slowing down the resume path by a lot?
-- Oscar
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [08/11] drm/i915/guc: Wait for doorbell to be inactive before deallocating
Quoted 22 lines · Oscar Mateo
Quoted 6 lines · Chris Wilson
Quoted 3 lines · Oscar Mateo
Doorbell release flow requires that we wait for
GEN8_DRB_VALIDbit to go to zero after updatingdb_statusbefore we call the GuC to release the doorbell.Does this fix some of the '110' errors?
-ChrisNo, AFAICT this doesn't fix anything (but it's required nonetheless).
The 110 errors (
ENTER_S_STATEandDEALLOCATE_DOORBELLfailed calls to GuC) are cause by us resetting the GuC right before we ask it to do something. E.g.: in the suspend path, we do i915_gem_suspend()->i915_gem_sanitize()->intel_gpu_reset() and thenintel_guc_suspend()inmediately after. After the sanitize, the GuC is in reset state and without a firmware loaded, so asking it to suspend is pretty useless. Same thing when trying to orderly shutdown the doorbells right after reset (the GuC has no idea of what doorbell we are talking about).
Quoted 8 lines
For now, the simple fix would be to keep doing the reset and not ask the GuC to do anything after. But sooner or later, we won't be able to get away with this. E.g.: if we enable direct submission, the GuC is the only one that will know about all the requests in-flight, and therefore the only one that can decide when to suspend. Also, surely loading the GuC firmware every time is slowing down the resume path by a lot?
-- Oscar
To add a bit of info, the wait is required because if there are pending rings on the doorbell the HW may require a slightly longer time to disable it after we update the status. If we call into GuC before the disabling is completed GuC will return an error because the doorbell will still be active from its point of view, even if it is in the process of shutting down. I think that with our current flow it should be impossible to hit this situation so it as Oscar said it doesn't fix anything, but it is still worth adding it to avoid issues in the future.
In regards to the reset, I agree that we should either reorder stuff to make sure that the H2G messages are sent before the reset or drop the messages entirely.
In the particular case of suspend, it is not guaranteed that the GuC will survive a suspend/resume cycle (it depends on the depth of the sleep state). That is why the ENTER_S_STATE message exists: this message tells GuC to save all the information it needs in the buffer we provided; we can then tell GuC to restore this state with the EXIT_S_STATE message, which can be used on a freshly loaded FW as well. With our current flow the GuC queues should be empty when we go to suspend so it shouldn't matter, but direct submission will need this to work.
Daniele
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Joonas Lahtinen
maintainerRe: [08/11] drm/i915/guc: Wait for doorbell to be inactive before deallocating
In reply to Oscar MateoQuoted 9 lines · Oscar Mateo
Doorbell release flow requires that we wait for
GEN8_DRB_VALIDbit to go to zero after updatingdb_statusbefore we call the GuC to release the doorbell.Kudos to Daniele for finding this out.
Cc Joonas Lahtinen <[email protected]>Cc Daniele Ceraolo Spurio <[email protected]>Signed-off-by Oscar Mateo <[email protected]>
<SNIP>
Quoted 12 lines
static int
__destroy_doorbell(struct i915_guc_client *client){drivers/gpu/drm/i915/i915_guc_submission.c +1
-0+struct drm_i915_private *dev_priv =guc_to_i915 (client->guc);struct
guc_doorbell_info*doorbell;drivers/gpu/drm/i915/i915_guc_submission.c +3
-0+u16 db_id = client->doorbell_id; + +GEM_BUG_ON (db_id >=GUC_DOORBELL_INVALID );
doorbell =__get_doorbell(client);doorbell->db_status=GUC_DOORBELL_DISABLED;doorbell->cookie= 0;
drivers/gpu/drm/i915/i915_guc_submission.c static int __create_doorbell(struct i915_guc_client *client) · line 223
Quoted 5 lines
drivers/gpu/drm/i915/i915_guc_submission.c +5
-0+/* Doorbell release flow requires that we wait for GEN8_DRB_VALID bit +* to go to zero after updating db_status before we call the GuC to +* release the doorbell */ +if (wait_for_us (!(I915_READ (GEN8_DRBREGL (db_id)) &GEN8_DRB_VALID ),10 )) +DRM_ERROR ("Doorbell never became invalid after disable\n" );
Would the appropriate action be to return -ETIMEOUT and not proceed with deallocation? I would at least WARN here. If GuC submission gets enabled, we should yell if it's not working. So only the enable stage should bail out with DRM_ERRORs and without WARN.
Quoted 3 lines
+
return__guc_deallocate_doorbell(client->guc, client->ctx_index); }
Regards, Joonas
Signature · 7 lines
Joonas Lahtinen Open Source Technology Center Intel Corporation _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v3 6/11 drm/i915/guc: Make intel_guc_send a function pointer
In reply to Oscar MateoPrepare for an alternate GuC communication interface.
v2: Make a few functions static and name them correctly while we are at it (Oscar), but leave an intel_guc_send_mmio interface for users that require old-style communication.
v3: Send intel_uc_init_early back to the top (Michal).
--- drivers/gpu/drm/i915/intel_uc.c | 17 +++++++++++------ drivers/gpu/drm/i915/intel_uc.h | 10 +++++++++- 2 files changed, 20 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/intel_uc.c +11 -6
@@ -27,14 +27,17 @@void intel_uc_init_early (struct drm_i915_private *dev_priv) { -mutex_init (&dev_priv->guc.send_mutex); +struct intel_guc *guc = &dev_priv->guc; + +mutex_init (&guc->send_mutex); + guc->send = intel_guc_send_mmio; }/* * Read GuC command/status register (SOFT_SCRATCH_0) * Return true if it contains a response rather than a command */ -static bool intel_guc_recv (struct intel_guc *guc,u32 *status) +static bool guc_recv (struct intel_guc *guc,u32 *status) {struct drm_i915_private *dev_priv =guc_to_i915 (guc);
@@ -43,7 +46,10 @@ static bool intel_guc_recv(struct intel_guc *guc, u32 *status)return INTEL_GUC_RECV_IS_RESPONSE (val); } -int intel_guc_send (struct intel_guc *guc,const u32 *action,u32 len) +/* +* This function implements the MMIO based host to GuC interface. +*/ +int intel_guc_send_mmio (struct intel_guc *guc,const u32 *action,u32 len) {struct drm_i915_private *dev_priv =guc_to_i915 (guc);u32 status;
@@ -71,9 +77,9 @@ int intel_guc_send(struct intel_guc *guc, const u32 *action, u32 len)* up to that length of time, then switch to a slower sleep-wait loop. * No inte_guc_send command should ever take longer than 10ms. */ - ret =wait_for_us (intel_guc_recv (guc, &status),10 ); + ret =wait_for_us (guc_recv (guc, &status),10 );if (ret) - ret =wait_for (intel_guc_recv (guc, &status),10 ); + ret =wait_for (guc_recv (guc, &status),10 );if (status !=INTEL_GUC_STATUS_SUCCESS ) {/* * Either the GuC explicitly returned an error (which
@@ -113,4 +119,3 @@ int intel_guc_sample_forcewake(struct intel_guc *guc)return intel_guc_send (guc, action,ARRAY_SIZE (action)); } -
drivers/gpu/drm/i915/intel_uc.h +9 -1
@@ -176,6 +176,9 @@ struct intel_guc {/* To serialize the intel_guc_send actions */ struct mutex send_mutex; + +/* GuC's FW specific send function */ +int (*send)(struct intel_guc *guc,const u32 *data,u32 len); };struct intel_huc {
@@ -187,9 +190,14 @@ struct intel_huc {/* intel_uc.c */ void intel_uc_init_early (struct drm_i915_private *dev_priv); -int intel_guc_send (struct intel_guc *guc,const u32 *action,u32 len);int intel_guc_sample_forcewake (struct intel_guc *guc); +int intel_guc_send_mmio (struct intel_guc *guc,const u32 *action,u32 len); +static inline int intel_guc_send (struct intel_guc *guc,const u32 *action,u32 len) +{ +return guc->send (guc, action, len); +} +/* intel_guc_loader.c */ extern void intel_guc_init (struct drm_i915_private *dev_priv);extern void intel_guc_fini (struct drm_i915_private *dev_priv);
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [06/11 v3] drm/i915/guc: Make intel_guc_send a function pointer
Quoted 10 lines · Oscar Mateo
Prepare for an alternate GuC communication interface.
v2: Make a few functions static and name them correctly while we are at it (Oscar), but leave an
intel_guc_send_mmiointerface for users that require old-style communication.v3: Send
intel_uc_init_earlyback to the top (Michal).Signed-off-by Michel Thierry <[email protected]>Signed-off-by Michal Wajdeczko <[email protected]>Signed-off-by Oscar Mateo <[email protected]>
Thanks,
Daniele
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 9/11 drm/i915/guc: A little bit more of doorbell sanitization
In reply to Oscar MateoSome recent refactoring patches have left the doorbell creation outside the GuC client allocation, which does not make a lot of sense (a client without a doorbell is something useless). Move it back there, and refactor the init_doorbell_hw consequently.
Thanks to this, we can do some other improvements, like hoisting the check for GuC submission enabled out of the enable function.
--- drivers/gpu/drm/i915/i915_drv.c | 3 +- drivers/gpu/drm/i915/i915_guc_submission.c | 228 ++++++++++++++++------------- drivers/gpu/drm/i915/intel_guc_loader.c | 22 ++- 3 files changed, 143 insertions(+), 110 deletions(-)
drivers/gpu/drm/i915/i915_drv.c +1 -2
@@ -549,8 +549,7 @@ static bool i915_switcheroo_can_switch(struct pci_dev *pdev)static void i915_gem_fini (struct drm_i915_private *dev_priv) {mutex_lock (&dev_priv->drm.struct_mutex); -if (i915.enable_guc_submission) -intel_guc_cleanup (dev_priv); +intel_guc_cleanup (dev_priv);i915_gem_cleanup_engines (dev_priv);i915_gem_context_fini (dev_priv);mutex_unlock (&dev_priv->drm.struct_mutex);
drivers/gpu/drm/i915/i915_guc_submission.c +126 -102
367 lines not on this page · raw message
drivers/gpu/drm/i915/intel_guc_loader.c +16 -6
@@ -441,9 +441,13 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)intel_uc_fw_status_repr (guc_fw->fetch_status), intel_uc_fw_status_repr(guc_fw->load_status)); - err =i915_guc_submission_init (dev_priv); -if (err) - goto err_guc; +if (i915.enable_guc_submission) { +/* This is stuff we need to have available at fw load time +* if we are planning to enable submission later */ + err =i915_guc_submission_init (dev_priv); +if (err) + goto err_guc; + }/* * WaEnableuKernelHeaderValidFix:skl,bxt
@@ -495,7 +499,8 @@ int intel_guc_setup(struct drm_i915_private *dev_priv) err_interrupts:gen9_disable_guc_interrupts (dev_priv); err_submission: -i915_guc_submission_fini (dev_priv); +if (i915.enable_guc_submission) +i915_guc_submission_fini (dev_priv); err_guc:i915_ggtt_disable_guc (dev_priv); fail:
@@ -543,8 +548,13 @@ int intel_guc_setup(struct drm_i915_private *dev_priv)void intel_guc_cleanup (struct drm_i915_private *dev_priv) { -gen9_disable_guc_interrupts (dev_priv); -i915_guc_submission_fini (dev_priv); +if (i915.enable_guc_submission) { +guc_interrupts_release (dev_priv); +i915_guc_submission_disable (dev_priv); +gen9_disable_guc_interrupts (dev_priv); +i915_guc_submission_fini (dev_priv); + } +i915_ggtt_disable_guc (dev_priv); }
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [09/11] drm/i915/guc: A little bit more of doorbell sanitization
Quoted 8 lines · Oscar Mateo
Some recent refactoring patches have left the doorbell creation outside the GuC client allocation, which does not make a lot of sense (a client without a doorbell is something useless). Move it back there, and refactor the
init_doorbell_hwconsequently.Thanks to this, we can do some other improvements, like hoisting the check for GuC submission enabled out of the enable function.
Quoted 7 lines
Cc Joonas Lahtinen <[email protected]>Cc Michal Wajdeczko <[email protected]>Cc Arkadiusz Hiler <[email protected]>Cc Tvrtko Ursulin <[email protected]>Cc Chris Wilson <[email protected]>Cc Daniele Ceraolo Spurio <[email protected]>Signed-off-by Oscar Mateo <[email protected]>
Looks good.
Regards, Joonas
Signature · 7 lines
Joonas Lahtinen Open Source Technology Center Intel Corporation _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
PATCH v2 10/11 drm/i915/guc: Refactor the concept "GuC context descriptor" into "GuC stage descriptor"
In reply to Oscar MateoA GuC context and a HW context are in no way related, so the name "GuC context descriptor" is very unfortunate, because a new reader of the code gets overwhelmed very quickly with a lot of things called "context" that refer to different things. We can improve legibility a lot by simply renaming a few objects in the GuC code.
--- drivers/gpu/drm/i915/i915_debugfs.c | 4 +- drivers/gpu/drm/i915/i915_guc_submission.c | 103 +++++++++++++++-------------- drivers/gpu/drm/i915/intel_guc_fwif.h | 38 +++++------ drivers/gpu/drm/i915/intel_guc_loader.c | 2 +- drivers/gpu/drm/i915/intel_uc.h | 4 +- 5 files changed, 78 insertions(+), 73 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -2
@@ -2467,8 +2467,8 @@ static void i915_guc_client_info(struct seq_file *m,enum intel_engine_id id;uint64_t tot =0 ; -seq_printf (m,"\tPriority %d, GuC ctx index: %u, PD offset 0x%x\n" , - client->priority, client->ctx_index, client->proc_desc_offset); +seq_printf (m,"\tPriority %d, GuC stage index: %u, PD offset 0x%x\n" , + client->priority, client->stage_id, client->proc_desc_offset);seq_printf (m,"\tDoorbell id %d, offset: 0x%lx, cookie 0x%x\n" , client->doorbell_id, client->doorbell_offset, client->doorbell_cookie); seq_printf(m, "\tWQ size %d, offset: 0x%x, tail %d\n",
drivers/gpu/drm/i915/i915_guc_submission.c +54 -49
284 lines not on this page · raw message
drivers/gpu/drm/i915/intel_guc_fwif.h +19 -19
@@ -26,14 +26,14 @@#define GFXCORE_FAMILY_GEN9 12#define GFXCORE_FAMILY_UNKNOWN 0x7fffffff -#define GUC_CTX_PRIORITY_KMD_HIGH 0 -#define GUC_CTX_PRIORITY_HIGH 1 -#define GUC_CTX_PRIORITY_KMD_NORMAL 2 -#define GUC_CTX_PRIORITY_NORMAL 3 -#define GUC_CTX_PRIORITY_NUM 4 +#define GUC_CLIENT_PRIORITY_KMD_HIGH 0 +#define GUC_CLIENT_PRIORITY_HIGH 1 +#define GUC_CLIENT_PRIORITY_KMD_NORMAL 2 +#define GUC_CLIENT_PRIORITY_NORMAL 3 +#define GUC_CLIENT_PRIORITY_NUM 4 -#define GUC_MAX_GPU_CONTEXTS 1024 -#define GUC_INVALID_CTX_ID GUC_MAX_GPU_CONTEXTS +#define GUC_MAX_STAGE_DESCRIPTORS 1024 +#define GUC_INVALID_STAGE_ID GUC_MAX_STAGE_DESCRIPTORS#define GUC_RENDER_ENGINE 0#define GUC_VIDEO_ENGINE 1
@@ -68,14 +68,14 @@#define GUC_DOORBELL_ENABLED 1#define GUC_DOORBELL_DISABLED 0 -#define GUC_CTX_DESC_ATTR_ACTIVE (1 << 0) -#define GUC_CTX_DESC_ATTR_PENDING_DB (1 << 1) -#define GUC_CTX_DESC_ATTR_KERNEL (1 << 2) -#define GUC_CTX_DESC_ATTR_PREEMPT (1 << 3) -#define GUC_CTX_DESC_ATTR_RESET (1 << 4) -#define GUC_CTX_DESC_ATTR_WQLOCKED (1 << 5) -#define GUC_CTX_DESC_ATTR_PCH (1 << 6) -#define GUC_CTX_DESC_ATTR_TERMINATED (1 << 7) +#define GUC_STAGE_DESC_ATTR_ACTIVE (1 << 0) +#define GUC_STAGE_DESC_ATTR_PENDING_DB (1 << 1) +#define GUC_STAGE_DESC_ATTR_KERNEL (1 << 2) +#define GUC_STAGE_DESC_ATTR_PREEMPT (1 << 3) +#define GUC_STAGE_DESC_ATTR_RESET (1 << 4) +#define GUC_STAGE_DESC_ATTR_WQLOCKED (1 << 5) +#define GUC_STAGE_DESC_ATTR_PCH (1 << 6) +#define GUC_STAGE_DESC_ATTR_TERMINATED (1 << 7)/* The guc control data is 10 DWORDs */ #define GUC_CTL_CTXINFO 0
@@ -256,7 +256,7 @@ struct guc_wq_item { } __packed;struct guc_process_desc { -u32 context_id; +u32 stage_id;u64 db_base_addr;u32 head;u32 tail;
@@ -290,9 +290,9 @@ struct guc_execlist_context { } __packed;/*Context descriptor for communicating between uKernel and Driver*/ -struct guc_context_desc { +struct guc_stage_desc {u32 sched_common_area; -u32 context_id; +u32 stage_id;u32 pas_id;u8 engines_used;u64 db_trigger_cpu;
@@ -359,7 +359,7 @@ struct guc_policy { } __packed;struct guc_policies { -struct guc_policy policy[GUC_CTX_PRIORITY_NUM ][GUC_MAX_ENGINES_NUM ]; +struct guc_policy policy[GUC_CLIENT_PRIORITY_NUM ][GUC_MAX_ENGINES_NUM ];/* In micro seconds. How much time to allow before DPC processing is * called back via interrupt (to prevent DPC queue drain starving).
drivers/gpu/drm/i915/intel_guc_loader.c +1 -1
@@ -176,7 +176,7 @@ static void guc_params_init(struct drm_i915_private *dev_priv)/* If GuC submission is enabled, set up additional parameters here */ if (i915.enable_guc_submission) {u32 pgs =guc_ggtt_offset (dev_priv->guc.ctx_pool); -u32 ctx_in_16 =GUC_MAX_GPU_CONTEXTS /16 ; +u32 ctx_in_16 =GUC_MAX_STAGE_DESCRIPTORS /16 ; pgs >>=PAGE_SHIFT ; params[GUC_CTL_CTXINFO] = (pgs << GUC_CTL_BASE_ADDR_SHIFT) |
drivers/gpu/drm/i915/intel_uc.h +2 -2
@@ -72,7 +72,7 @@ struct i915_guc_client {uint32_t engines;/* bitmap of (host) engine ids */ uint32_t priority; -u32 ctx_index; +u32 stage_id;uint32_t proc_desc_offset;u16 doorbell_id;
@@ -157,7 +157,7 @@ struct intel_guc {struct i915_vma *addon;struct i915_vma *ctx_pool;void *ctx_pool_vaddr; -struct ida ctx_ids; +struct ida stage_ids;struct i915_guc_client *execbuf_client;
Signature · 6 lines
1.9.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [10/11] drm/i915/guc: Refactor the concept "GuC context descriptor" into "GuC stage descriptor"
Quoted 27 lines · Oscar Mateo
A GuC context and a HW context are in no way related, so the name "GuC context descriptor" is very unfortunate, because a new reader of the code gets overwhelmed very quickly with a lot of things called "context" that refer to different things. We can improve legibility a lot by simply renaming a few objects in the GuC code.
Signed-off-by Oscar Mateo <[email protected]>---
drivers/gpu/drm/i915/i915_guc_submission.c +5
-6@@ -35,16 +35,18 @@* descriptor and a workqueue (all of them inside a single gem object that * contains all required pages for these elements). * -* GuC context descriptor: -* During initialization, the driver allocates a static pool of 1024 such -* descriptors, and shares them with the GuC. +* GuC stage descriptor: +* Technically, this is known as a "GuC Context" descriptor in the specs, but +* we use the term "stage" to avoid confusion with all the other things already +* named "context" in the driver. During initialization, the driver allocates +* a static pool of 1024 such descriptors, and shares them with the GuC. * Currently, there exists a 1:1 mapping between a i915_guc_client and a -* guc_context_desc (via the client's context_index), so effectively only -* one guc_context_desc gets used. This context descriptor lets the GuC know -* about the doorbell, workqueue and process descriptor. Theoretically, it also
drivers/gpu/drm/i915/i915_guc_submission.c line 51
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +5
-2-* lets the GuC know about our HW contexts (Context ID, etc...), but we actually +* guc_stage_desc (via the client's stage_id), so effectively only one +* gets used. This stage descriptor lets the GuC know about the doorbell, +* workqueue and process descriptor. Theoretically, it also lets the GuC +* know about our HW contexts (context ID, etc...), but we actually * employ a kind of submission where the GuC uses the LRCA sent via the work -* item instead (the single guc_context_desc associated to execbuf client +* item instead (the single guc_stage_desc associated to execbuf client * contains information about the default kernel context only, but this is * essentially unused). This is called a "proxy" submission.
Some of the above (esp. "this is known as a "GuC Context" descriptor in the specs") would be good here. Some might say this is where we expect the kerneldoc for structs to be...
Quoted 11 lines
drivers/gpu/drm/i915/i915_guc_submission.c +2
-2/*Context descriptor for communicating between uKernel and Driver*/ -struct guc_context_desc { +struct guc_stage_desc {u32 sched_common_area; -u32 context_id; +u32 stage_id;u32 pas_id;u8 engines_used;u64 db_trigger_cpu;@@ -359,7 +359,7 @@ struct guc_policy { } __packed;
Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [10/11] drm/i915/guc: Refactor the concept "GuC context descriptor" into "GuC stage descriptor"
Quoted 42 lines · Chris Wilson
Quoted 27 lines · Oscar Mateo
A GuC context and a HW context are in no way related, so the name "GuC context descriptor" is very unfortunate, because a new reader of the code gets overwhelmed very quickly with a lot of things called "context" that refer to different things. We can improve legibility a lot by simply renaming a few objects in the GuC code.
Signed-off-by Oscar Mateo <[email protected]>---
drivers/gpu/drm/i915/i915_guc_submission.c +5
-6@@ -35,16 +35,18 @@* descriptor and a workqueue (all of them inside a single gem object that * contains all required pages for these elements). * -* GuC context descriptor: -* During initialization, the driver allocates a static pool of 1024 such -* descriptors, and shares them with the GuC. +* GuC stage descriptor: +* Technically, this is known as a "GuC Context" descriptor in the specs, but +* we use the term "stage" to avoid confusion with all the other things already +* named "context" in the driver. During initialization, the driver allocates +* a static pool of 1024 such descriptors, and shares them with the GuC. * Currently, there exists a 1:1 mapping between a i915_guc_client and a -* guc_context_desc (via the client's context_index), so effectively only -* one guc_context_desc gets used. This context descriptor lets the GuC know -* about the doorbell, workqueue and process descriptor. Theoretically, it also drivers/gpu/drm/i915/i915_guc_submission.c line 51
Quoted 10 lines
drivers/gpu/drm/i915/i915_guc_submission.c +5
-2-* lets the GuC know about our HW contexts (Context ID, etc...), but we actually +* guc_stage_desc (via the client's stage_id), so effectively only one +* gets used. This stage descriptor lets the GuC know about the doorbell, +* workqueue and process descriptor. Theoretically, it also lets the GuC +* know about our HW contexts (context ID, etc...), but we actually * employ a kind of submission where the GuC uses the LRCA sent via the work -* item instead (the single guc_context_desc associated to execbuf client +* item instead (the single guc_stage_desc associated to execbuf client * contains information about the default kernel context only, but this is * essentially unused). This is called a "proxy" submission. Some of the above (esp. "this is known as a "GuC Context" descriptor in the specs") would be good here. Some might say this is where we expect the kerneldoc for structs to be...
Quoted 12 lines
drivers/gpu/drm/i915/i915_guc_submission.c struct guc_policy { · line 360
Quoted 12 lines
drivers/gpu/drm/i915/i915_guc_submission.c +2
-2/*Context descriptor for communicating between uKernel and Driver*/ -struct guc_context_desc { +struct guc_stage_desc {u32 sched_common_area; -u32 context_id; +u32 stage_id;u32 pas_id;u8 engines_used;u64 db_trigger_cpu;@@ -359,7 +359,7 @@ struct guc_policy { } __packed;
Fair enough, I'll move it. Thanks,
-- Oscar
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx