Context based TLB invalidations
Subsystems (2)
Matthew Brost
maintainerv2 cover Context based TLB invalidations
Add support for context based TLB invalidations.
Matt
Matthew Brost (12):
drm/xe: Add normalize_invalidation_range
drm/xe: Make usm.asid_to_vm allocation use GFP_NOWAIT drm/xe: Add xe_device_asid_to_vm helper
drm/xe: Add vm to exec queues association
drm/xe: Taint TLB invalidation seqno lock with GFP_KERNEL drm/xe: Do not forward invalid TLB invalidation seqnos to upper layers drm/xe: Rename send_tlb_inval_ppgtt to send_tlb_inval_asid_ppgtt drm/xe: Add send_tlb_inval_ppgtt helper
drm/xe: Add xe_tlb_inval_idle helper
drm/xe: Add exec queue active vfunc
drm/xe: Add context-based invalidation to GuC TLB invalidation backend drm/xe: Enable context TLB invalidations for CI
drivers/gpu/drm/xe/xe_device.c | 25 +++ drivers/gpu/drm/xe/xe_device.h | 11 +- drivers/gpu/drm/xe/xe_device_types.h | 9 + drivers/gpu/drm/xe/xe_exec_queue.c | 7 +- drivers/gpu/drm/xe/xe_exec_queue_types.h | 5 + drivers/gpu/drm/xe/xe_execlist.c | 7 + drivers/gpu/drm/xe/xe_guc_submit.c | 6 + drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 215 ++++++++++++++++++----- drivers/gpu/drm/xe/xe_pci.c | 2 + drivers/gpu/drm/xe/xe_pci_types.h | 1 + drivers/gpu/drm/xe/xe_tlb_inval.c | 33 ++++ drivers/gpu/drm/xe/xe_tlb_inval.h | 2 + drivers/gpu/drm/xe/xe_tlb_inval_types.h | 1 + drivers/gpu/drm/xe/xe_vm.c | 49 +++++- drivers/gpu/drm/xe/xe_vm.h | 3 + drivers/gpu/drm/xe/xe_vm_types.h | 13 ++ 16 files changed, 337 insertions(+), 52 deletions(-)
Signature · 1 line
2.34.1
Matthew Brost
maintainerPATCH v2 1/12 drm/xe: Add normalize_invalidation_range
Extract the code that determines the alignment of TLB invalidation into a helper function — normalize_invalidation_range. This will be useful when adding context-based invalidations to the GuC TLB invalidation backend.
--- drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 71 +++++++++++++-------------- 1 file changed, 35 insertions(+), 36 deletions(-)
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +35 -36
@@ -92,6 +92,38 @@ static int send_tlb_inval_ggtt(struct xe_tlb_inval *tlb_inval, u32 seqno)return -ECANCELED ; } +static u64 normalize_invalidation_range (struct xe_gt *gt,u64 *start,u64 *end) +{ +u64 orig_start = *start; +u64 length = *end - *start; +u64 align; + +if (length <SZ_4K ) + length =SZ_4K ; + + align =roundup_pow_of_two (length); + *start =ALIGN_DOWN (*start, align); + *end =ALIGN (*end, align); + length = align; +while (*start + length < *end) { + length <<=1 ; + *start =ALIGN_DOWN (orig_start, length); + } + +if (length >=SZ_2M ) { + length =max_t (u64,SZ_16M , length); + *start =ALIGN_DOWN (orig_start, length); + } + +xe_gt_assert (gt, length >=SZ_4K ); +xe_gt_assert (gt,is_power_of_2 (length)); +xe_gt_assert (gt, !(length &GENMASK (ilog2 (SZ_16M ) -1 , +ilog2 (SZ_2M ) +1 ))); +xe_gt_assert (gt,IS_ALIGNED (*start, length)); + +return length; +} +/* * Ensure that roundup_pow_of_two(length) doesn't overflow. * Note that roundup_pow_of_two() operates on unsigned long,
@@ -118,47 +150,14 @@ static int send_tlb_inval_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, length > MAX_RANGE_TLB_INVALIDATION_LENGTH) { action[len++] = MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL); } else { -u64 orig_start = start; -u64 align; - -if (length <SZ_4K ) - length =SZ_4K ; - -/* -* We need to invalidate a higher granularity if start address -* is not aligned to length. When start is not aligned with -* length we need to find the length large enough to create an -* address mask covering the required range. -*/ - align =roundup_pow_of_two (length); - start =ALIGN_DOWN (start, align); - end =ALIGN (end, align); - length = align; -while (start + length < end) { - length <<=1 ; - start =ALIGN_DOWN (orig_start, length); - } - -/* -* Minimum invalidation size for a 2MB page that the hardware -* expects is 16MB -*/ -if (length >=SZ_2M ) { - length =max_t (u64,SZ_16M , length); - start =ALIGN_DOWN (orig_start, length); - } - -xe_gt_assert (gt, length >=SZ_4K ); -xe_gt_assert (gt,is_power_of_2 (length)); -xe_gt_assert (gt, !(length &GENMASK (ilog2 (SZ_16M ) -1 , -ilog2 (SZ_2M ) +1 ))); -xe_gt_assert (gt,IS_ALIGNED (start, length)); +u64 normalize_len =normalize_invalidation_range (gt, &start, + &end); action[len++] =MAKE_INVAL_OP (XE_GUC_TLB_INVAL_PAGE_SELECTIVE ); action[len++] = asid; action[len++] =lower_32_bits (start); action[len++] =upper_32_bits (start); - action[len++] =ilog2 (length) -ilog2 (SZ_4K ); + action[len++] =ilog2 (normalize_len) -ilog2 (SZ_4K ); }xe_gt_assert (gt, len <=MAX_TLB_INVALIDATION_LEN );
Signature · 1 line
2.34.1
Re: [PATCH v2 01/12] drm/xe: Add normalize_invalidation_range
Quoted 8 lines · Matthew Brost
Extract the code that determines the alignment of TLB invalidation into
a helper function —normalize_invalidation_range. This will be useful when adding context-based invalidations to the GuC TLB invalidation backend.Signed-off-by Nirmoy Das <[email protected]>Signed-off-by Matthew Brost <[email protected]>
Quoted 112 lines
---
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 71 +++++++++++++------------ --1 file changed, 35 insertions(+), 36 deletions(-)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -92,6 +92,38 @@ static int send_tlb_inval_ggtt(struct xe_tlb_inval *tlb_inval, u32 seqno) return -ECANCELED; }
+static u64normalize_invalidation_range(struct xe_gt *gt, u64 *start, u64 *end)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +31
-0+{ + u64 orig_start = *start; + u64 length = *end - *start; + u64 align; + + if (length < SZ_4K) + length = SZ_4K; + + align = roundup_pow_of_two(length); + *start = ALIGN_DOWN(*start, align); + *end = ALIGN(*end, align); + length = align; + while (*start + length < *end) { + length <<= 1; + *start = ALIGN_DOWN(orig_start, length); + } + + if (length >= SZ_2M) { + length = max_t(u64, SZ_16M, length); + *start = ALIGN_DOWN(orig_start, length); + } + + xe_gt_assert(gt, length >= SZ_4K); + xe_gt_assert(gt, is_power_of_2(length)); + xe_gt_assert(gt, !(length & GENMASK(ilog2(SZ_16M) - 1, + ilog2(SZ_2M) + 1))); + xe_gt_assert(gt, IS_ALIGNED(*start, length)); + + return length; +} +/*
* Ensure thatroundup_pow_of_two(length)doesn't overflow.
* Note thatroundup_pow_of_two()operates on unsigned long,drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-8@@ -118,47 +150,14 @@ static int send_tlb_inval_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, length > MAX_RANGE_TLB_INVALIDATION_LENGTH) { action[len++] = MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL); } else { - u64 orig_start = start; - u64 align; - - if (length < SZ_4K) - length = SZ_4K; - - /* -* We need to invalidate a higher granularity if start address
- * is not aligned to length. When start is not aligned with
- * length we need to find the length large enough to create an
- * address mask covering the required range. - */
- align =roundup_pow_of_two(length);
- start =ALIGN_DOWN(start, align);
- end =ALIGN(end, align);
- length = align;
- while (start + length < end) {
- length <<= 1;
- start =ALIGN_DOWN(orig_start, length); - }
-
- /*
- * Minimum invalidation size for a 2MB page that the hardware
- * expects is 16MB
- */
- if (length >=SZ_2M) {
- length =max_t(u64, SZ_16M, length); - start =ALIGN_DOWN(orig_start, length); - }
-
- xe_gt_assert(gt, length >=SZ_4K);
- xe_gt_assert(gt,is_power_of_2(length)); - xe_gt_assert(gt, !(length & GENMASK(ilog2(SZ_16M) - 1,
-ilog2(SZ_2M)+ 1)));
- xe_gt_assert(gt,IS_ALIGNED(start, length)); + u64normalize_len= normalize_invalidation_range(gt, &start,
+
&end);
action[len++] =MAKE_INVAL_OP(XE_GUC_TLB_INVAL_PAGE_SELECTIVE); action[len++] = asid;
action[len++] =lower_32_bits(start); action[len++] =upper_32_bits(start);
- action[len++] = ilog2(length) -ilog2(SZ_4K); + action[len++] =ilog2(normalize_len)-ilog2(SZ_4K); }
xe_gt_assert(gt, len <=MAX_TLB_INVALIDATION_LEN);
Matthew Brost
maintainerPATCH v2 2/12 drm/xe: Make usm.asid_to_vm allocation use GFP_NOWAIT
In reply to Matthew BrostEnsure the asid_to_vm lookup is reclaim-safe so it can be performed during TLB invalidations, which is necessary for context-based TLB invalidation support.
--- drivers/gpu/drm/xe/xe_vm.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_vm.c +1 -1
@@ -1632,7 +1632,7 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags, struct xe_file *xef)down_write (&xe->usm.lock); err =xa_alloc_cyclic (&xe->usm.asid_to_vm, &asid, vm,XA_LIMIT (1 ,XE_MAX_ASID -1 ), - &xe->usm.next_asid,GFP_KERNEL ); + &xe->usm.next_asid,GFP_NOWAIT );up_write (&xe->usm.lock);if (err <0 ) goto err_close;
Signature · 1 line
2.34.1
Re: [PATCH v2 02/12] drm/xe: Make usm.asid_to_vm allocation use GFP_NOWAIT
Quoted 5 lines · Matthew Brost
Ensure the
asid_to_vmlookup is reclaim-safe so it can be performed during TLB invalidations, which is necessary for context-based TLB invalidation support.Signed-off-by Matthew Brost <[email protected]>
Although best to merge this with the full series of course...
Quoted 20 lines
---
drivers/gpu/drm/xe/xe_vm.c | 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/xe/xe_vm.c +1
-1@@ -1632,7 +1632,7 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags, struct xe_file *xef) down_write(&xe->usm.lock); err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, vm, XA_LIMIT(1, XE_MAX_ASID - 1), - &xe->usm.next_asid,GFP_KERNEL ); + &xe->usm.next_asid, GFP_NOWAIT);up_write(&xe->usm.lock);
if (err < 0)
gotoerr_close;
Re: [PATCH v2 02/12] drm/xe: Make usm.asid_to_vm allocation use GFP_NOWAIT
In reply to Matthew BrostQuoted 14 lines · Matthew Brost
Ensure the
asid_to_vmlookup is reclaim-safe so it can be performed during TLB invalidations, which is necessary for context-based TLB invalidation support.Signed-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_vm.c | 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/xe/xe_vm.c +0
-0
drivers/gpu/drm/xe/xe_vm.c struct xe_vm *xe_vm_create(struct xe_device · line 1638
Quoted 8 lines
drivers/gpu/drm/xe/xe_vm.c +1
-1@@ -1632,7 +1632,7 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags, struct xe_file *xef) down_write(&xe->usm.lock); err = xa_alloc_cyclic(&xe->usm.asid_to_vm, &asid, vm, XA_LIMIT(1, XE_MAX_ASID - 1), - &xe->usm.next_asid,GFP_KERNEL ); + &xe->usm.next_asid, GFP_NOWAIT);
Looks like there's also a for-ci usage of this in xe_device.c that attempts to force an asid_to_vm wrap. Should we make the same change there?
Thanks,
Stuart
Quoted 3 lines
up_write(&xe->usm.lock);
if (err < 0)
gotoerr_close;
Matthew Brost
maintainerPATCH v2 3/12 drm/xe: Add xe_device_asid_to_vm helper
In reply to Matthew BrostIntroduce the xe_device_asid_to_vm helper, which can be used throughout the driver to resolve the VM from a given ASID.
--- drivers/gpu/drm/xe/xe_device.c | 25 +++++++++++++++++++++++++ drivers/gpu/drm/xe/xe_device.h | 4 ++++ 2 files changed, 29 insertions(+)
drivers/gpu/drm/xe/xe_device.c +25 -0
@@ -1290,3 +1290,28 @@ void xe_device_declare_wedged(struct xe_device *xe)drm_dev_wedged_event (&xe->drm, xe->wedged.method,NULL ); } } + +/** +* xe_device_asid_to_vm() - Find VM from ASID +* @xe: the &xe_device +* @asid: Address space ID +* +* Find a VM from ASID and take a reference to VM which caller must drop. +* Reclaim safe. +* +* Return: VM on success, ERR_PTR on failure +*/ +struct xe_vm *xe_device_asid_to_vm (struct xe_device *xe,u32 asid) +{ +struct xe_vm *vm; + +down_read (&xe->usm.lock); + vm =xa_load (&xe->usm.asid_to_vm, asid); +if (vm) +xe_vm_get (vm); +else + vm =ERR_PTR (-EINVAL ); +up_read (&xe->usm.lock); + +return vm; +}
drivers/gpu/drm/xe/xe_device.h +4 -0
@@ -6,6 +6,8 @@#ifndef _XE_DEVICE_H_#define _XE_DEVICE_H_ +struct xe_vm ; +#include <drm/drm_util.h> #include "xe_device_types.h"
@@ -195,6 +197,8 @@ void xe_file_put(struct xe_file *xef);int xe_is_injection_active (void ); +struct xe_vm *xe_device_asid_to_vm (struct xe_device *xe,u32 asid); +/* * Occasionally it is seen that the G2H worker starts running after a delay of more than * a second even after being queued and activated by the Linux workqueue subsystem. This
Signature · 1 line
2.34.1
Re: [PATCH v2 03/12] drm/xe: Add xe_device_asid_to_vm helper
Quoted 4 lines · Matthew Brost
Introduce the
xe_device_asid_to_vmhelper, which can be used throughout the driver to resolve the VM from a given ASID.Signed-off-by Matthew Brost <[email protected]>
Quoted 63 lines
--- drivers/gpu/drm/xe/xe_device.c | 25 +++++++++++++++++++++++++ drivers/gpu/drm/xe/xe_device.h | 4 ++++ 2 files changed, 29 insertions(+)drivers/gpu/drm/xe/xe_device.c +25
-0@@ -1290,3 +1290,28 @@ void xe_device_declare_wedged(struct xe_device *xe)drm_dev_wedged_event (&xe->drm, xe->wedged.method,NULL ); } } + +/** +* xe_device_asid_to_vm() - Find VM from ASID +* @xe: the &xe_device +* @asid: Address space ID +* +* Find a VM from ASID and take a reference to VM which caller must drop. +* Reclaim safe. +* +* Return: VM on success, ERR_PTR on failure +*/ +struct xe_vm *xe_device_asid_to_vm (struct xe_device *xe,u32 asid) +{ +struct xe_vm *vm; + +down_read (&xe->usm.lock); + vm =xa_load (&xe->usm.asid_to_vm, asid); +if (vm) +xe_vm_get (vm); +else + vm =ERR_PTR (-EINVAL ); +up_read (&xe->usm.lock); + +return vm; +}drivers/gpu/drm/xe/xe_device.h +4
-0@@ -6,6 +6,8 @@#ifndef _XE_DEVICE_H_#define _XE_DEVICE_H_ +struct xe_vm ; +#include <drm/drm_util.h> #include "xe_device_types.h" @@ -195,6 +197,8 @@ void xe_file_put(struct xe_file *xef);int xe_is_injection_active (void ); +struct xe_vm *xe_device_asid_to_vm (struct xe_device *xe,u32 asid); +/* * Occasionally it is seen that the G2H worker starts running after a delay of more than * a second even after being queued and activated by the Linux workqueue subsystem. This Signature · 1 line
2.34.1
Matthew Brost
maintainerPATCH v2 4/12 drm/xe: Add vm to exec queues association
In reply to Matthew BrostMaintain a list of exec queues per vm which will be used by TLB invalidation code to do context-ID based tlb invalidations.
--- drivers/gpu/drm/xe/xe_device.h | 7 ---- drivers/gpu/drm/xe/xe_device_types.h | 7 ++++ drivers/gpu/drm/xe/xe_exec_queue.c | 7 +++- drivers/gpu/drm/xe/xe_exec_queue_types.h | 3 ++ drivers/gpu/drm/xe/xe_vm.c | 47 ++++++++++++++++++++++++ drivers/gpu/drm/xe/xe_vm.h | 3 ++ drivers/gpu/drm/xe/xe_vm_types.h | 13 +++++++ 7 files changed, 79 insertions(+), 8 deletions(-)
drivers/gpu/drm/xe/xe_device.h +0 -7
@@ -62,13 +62,6 @@ static inline struct xe_tile *xe_device_get_root_tile(struct xe_device *xe)return &xe->tiles[0 ]; } -/* -* Highest GT/tile count for any platform. Used only for memory allocation -* sizing. Any logic looping over GTs or mapping userspace GT IDs into GT -* structures should use the per-platform xe->info.max_gt_per_tile instead. -*/ -#define XE_MAX_GT_PER_TILE 2 -static inline struct xe_gt *xe_device_get_gt (struct xe_device *xe,u8 gt_id) {struct xe_tile *tile;
drivers/gpu/drm/xe/xe_device_types.h +7 -0
@@ -57,6 +57,13 @@ struct xe_vram_region;#define XE_GT1 1#define XE_MAX_TILES_PER_DEVICE (XE_GT1 + 1) +/* +* Highest GT/tile count for any platform. Used only for memory allocation +* sizing. Any logic looping over GTs or mapping userspace GT IDs into GT +* structures should use the per-platform xe->info.max_gt_per_tile instead. +*/ +#define XE_MAX_GT_PER_TILE 2 +#define XE_MAX_ASID (BIT(20))#define IS_PLATFORM_STEP (_xe, _platform, min_step, max_step) \
drivers/gpu/drm/xe/xe_exec_queue.c +6 -1
@@ -72,8 +72,10 @@ static void __xe_exec_queue_free(struct xe_exec_queue *q)if (xe_exec_queue_uses_pxp (q))xe_pxp_exec_queue_remove (gt_to_xe (q->gt)->pxp, q); -if (q->vm) +if (q->vm) { +xe_vm_remove_exec_queue (q->vm, q);xe_vm_put (q->vm); + }if (q->xef)xe_file_put (q->xef);
@@ -143,6 +145,7 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, q->ring_ops = gt->ring_ops[hwe->class]; q->ops = gt->exec_queue_ops;INIT_LIST_HEAD (&q->lr.link); +INIT_LIST_HEAD (&q->vm_exec_queue_link);INIT_LIST_HEAD (&q->multi_gt_link);INIT_LIST_HEAD (&q->hw_engine_group_link);INIT_LIST_HEAD (&q->pxp.link);
@@ -796,6 +799,8 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data, } q->xef =xe_file_get (xef); +if (eci[0 ].engine_class !=DRM_XE_ENGINE_CLASS_VM_BIND ) +xe_vm_add_exec_queue (vm, q);/* user id alloc must always be last in ioctl to prevent UAF */ err =xa_alloc (&xef->exec_queue.xa, &id, q, xa_limit_32b,GFP_KERNEL );
drivers/gpu/drm/xe/xe_exec_queue_types.h +3 -0
@@ -147,6 +147,9 @@ struct xe_exec_queue {struct xe_dep_scheduler *dep_scheduler; } tlb_inval[XE_EXEC_QUEUE_TLB_INVAL_COUNT ]; +/** @vm_exec_queue_link: Link to track exec queue within a VM's list of exec queues. */ +struct list_head vm_exec_queue_link; +/** @pxp: PXP info tracking */ struct {/** @pxp.type: PXP session type used by this queue */
drivers/gpu/drm/xe/xe_vm.c +47 -0
58 lines not on this page · raw message
drivers/gpu/drm/xe/xe_vm.h +3 -0
@@ -284,6 +284,9 @@ static inline struct dma_resv *xe_vm_resv(struct xe_vm *vm)void xe_vm_kill (struct xe_vm *vm,bool unlocked); +void xe_vm_add_exec_queue (struct xe_vm *vm,struct xe_exec_queue *q); +void xe_vm_remove_exec_queue (struct xe_vm *vm,struct xe_exec_queue *q); +/** * xe_vm_assert_held(vm) - Assert that the vm's reservation object is held. * @vm: The vm
drivers/gpu/drm/xe/xe_vm_types.h +13 -0
@@ -290,6 +290,19 @@ struct xe_vm {struct list_head pm_activate_link; } preempt; +/** @exec_queues: Manages list of exec queues attached to this VM, protected by lock. */ +struct { +/** @exec_queues.list: list of exec queues attached to this VM */ +struct list_head list; +/** +* @exec_queues.count: count of exec queues attached to this VM, +* per GT +*/ +int count[XE_MAX_TILES_PER_DEVICE *XE_MAX_GT_PER_TILE ]; +/** @exec_queues.lock: lock to protect exec_queues list */ +struct rw_semaphore lock; + } exec_queues; +/** @um: unified memory state */ struct {/** @asid: address space ID, unique to each VM */
Signature · 1 line
2.34.1
Re: [PATCH v2 04/12] drm/xe: Add vm to exec queues association
Quoted 2 lines · Matthew Brost
Maintain a list of exec queues per vm which will be used by TLB invalidation code to do context-ID based tlb invalidations.
I had mentioned this in a subsequent patch, but just to confirm here.. if we're going to do this in the exec queue layer, we need to make sure the actual usage of this ensures registration/deregistration compliance. The series as it stands today does not.
Thanks,
Stuart
Quoted 231 lines
Signed-off-by Nirmoy Das <[email protected]>Signed-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device.h | 7 ---- drivers/gpu/drm/xe/xe_device_types.h | 7 ++++ drivers/gpu/drm/xe/xe_exec_queue.c | 7 +++-drivers/gpu/drm/xe/xe_exec_queue_types.h| 3 ++ drivers/gpu/drm/xe/xe_vm.c | 47
++++++++++++++++++++++++
drivers/gpu/drm/xe/xe_vm.h | 3 ++ drivers/gpu/drm/xe/xe_vm_types.h | 13 +++++++7 files changed, 79 insertions(+), 8 deletions(-)drivers/gpu/drm/xe/xe_device.h +0
-0@@ -62,13 +62,6 @@ static inline struct xe_tile *xe_device_get_root_tile(struct xe_device *xe) return &xe->tiles[0]; }
-/*
- * Highest GT/tile count for any platform. Used only for memory allocation
- * sizing. Any logic looping over GTs or mapping userspace GT IDs into GT
- * structures should use the per-platformxe->info.max_gt_per_tileinstead.
- */drivers/gpu/drm/xe/xe_device.h +0
-2-#define XE_MAX_GT_PER_TILE 2 -static inline struct
xe_gt*xe_device_get_gt(structxe_device*xe, u8gt_id)
{
structxe_tile*tile;drivers/gpu/drm/xe/xe_device_types.h +0
-0@@ -57,6 +57,13 @@ struct xe_vram_region; #define XE_GT1 1 #define XE_MAX_TILES_PER_DEVICE (XE_GT1 + 1)
+/*
+ * Highest GT/tile count for any platform. Used only for memory allocation
+ * sizing. Any logic looping over GTs or mapping userspace GT IDs into GT
+ * structures should use the per-platformxe->info.max_gt_per_tileinstead.
+ */drivers/gpu/drm/xe/xe_device_types.h +2
-0+#define XE_MAX_GT_PER_TILE 2 +#define XE_MAX_ASID (BIT(20))
#defineIS_PLATFORM_STEP(_xe, _platform, min_step, max_step)\drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -72,8 +72,10 @@ static void __xe_exec_queue_free(struct xe_exec_queue *q)
if (xe_exec_queue_uses_pxp(q))xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q); - if (q->vm)
+ if (q->vm) {
+ xe_vm_remove_exec_queue(q->vm, q);xe_vm_put(q->vm);
+ }
if (q->xef)
xe_file_put(q->xef);drivers/gpu/drm/xe/xe_exec_queue.c +1
-0@@ -143,6 +145,7 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, q->ring_ops = gt->ring_ops[hwe->class]; q->ops = gt->exec_queue_ops; INIT_LIST_HEAD(&q->lr.link); + INIT_LIST_HEAD(&q->vm_exec_queue_link); INIT_LIST_HEAD(&q->multi_gt_link); INIT_LIST_HEAD(&q->hw_engine_group_link);INIT_LIST_HEAD(&q->pxp.link);
drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -796,6 +799,8 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data, }
q->xef =xe_file_get(xef);
+ if (eci[0].engine_class !=DRM_XE_ENGINE_CLASS_VM_BIND) + xe_vm_add_exec_queue(vm, q);
/* user id alloc must always be last in ioctl to prevent UAF */
err =xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL);drivers/gpu/drm/xe/xe_exec_queue_types.h +0
-0@@ -147,6 +147,9 @@ struct xe_exec_queue { struct xe_dep_scheduler *dep_scheduler; } tlb_inval[XE_EXEC_QUEUE_TLB_INVAL_COUNT];
+ /** @vm_exec_queue_link: Link to track exec queue within a VM's list of exec queues. */
+ structlist_headvm_exec_queue_link;
+
/** @pxp: PXP info tracking */
struct {
/** @pxp.type: PXP session type used by this queue */drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -1507,8 +1507,20 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags, struct xe_file *xef) INIT_WORK(&vm->destroy_work, vm_destroy_work_func);
INIT_LIST_HEAD(&vm->preempt.exec_queues);
+ INIT_LIST_HEAD(&vm->exec_queues.list);vm->preempt.min_run_period_ms= 10; /* FIXME: Wire up to uAPI */
+ init_rwsem(&vm->exec_queues.lock);
+ if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
+ fs_reclaim_acquire(GFP_KERNEL);
+ might_lock(&vm->exec_queues.lock);
+ fs_reclaim_release(GFP_KERNEL);drivers/gpu/drm/xe/xe_vm.c +6
-0+ + down_read(&vm->exec_queues.lock); + might_lock(&xe_root_mmio_gt(xe)->uc.guc.ct.lock); + up_read(&vm->exec_queues.lock); + } +for_each_tile(tile, xe, id)
xe_range_fence_tree_init(&vm->rftree[id]);drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -4387,3 +4399,38 @@ int xe_vm_alloc_cpu_addr_mirror_vma(struct
xe_vm*vm,uint64_tstart,uint64_tr
returnxe_vm_alloc_vma(vm, &map_req, false); }drivers/gpu/drm/xe/xe_vm.c +27
-0+ +/** +* xe_vm_add_exec_queue() - Add exec queue to VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_add_exec_queue (struct xe_vm *vm,struct xe_exec_queue *q) +{ + /* User VMs and queues only */ + xe_assert(vm->xe, !(q->flags & EXEC_QUEUE_FLAG_KERNEL)); + xe_assert(vm->xe, !(q->flags & EXEC_QUEUE_FLAG_PERMANENT)); + xe_assert(vm->xe, !(q->flags & EXEC_QUEUE_FLAG_VM)); + xe_assert(vm->xe, !(q->flags & EXEC_QUEUE_FLAG_MIGRATE)); + xe_assert(vm->xe, vm->xef); + + down_write(&vm->exec_queues.lock); + list_add(&q->vm_exec_queue_link, &vm->exec_queues.list); + ++vm->exec_queues.count[q->gt->info.id]; + up_write(&vm->exec_queues.lock); +} + +/** +* xe_vm_remove_exec_queue() - Remove exec queue from VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_remove_exec_queue(struct xe_vm *vm, struct xe_exec_queue*q)
drivers/gpu/drm/xe/xe_vm.c +8
-0+{ + down_write(&vm->exec_queues.lock); + if (!list_empty(&q->vm_exec_queue_link)) { + list_del(&q->vm_exec_queue_link); + --vm->exec_queues.count[q->gt->info.id]; + } + up_write(&vm->exec_queues.lock); +}drivers/gpu/drm/xe/xe_vm.h +0
-0@@ -284,6 +284,9 @@ static inline struct dma_resv *xe_vm_resv(struct xe_vm *vm)
voidxe_vm_kill(struct xe_vm *vm, bool unlocked);
+voidxe_vm_add_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+voidxe_vm_remove_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+
/**
*xe_vm_assert_held(vm)- Assert that the vm's reservation object is held.
* @vm: The vmdrivers/gpu/drm/xe/xe_vm_types.h +0
-0@@ -290,6 +290,19 @@ struct xe_vm { struct list_head pm_activate_link; } preempt;
+ /** @exec_queues: Manages list of exec queues attached to this VM, protected by lock. */
+ struct {
+ /** @exec_queues.list: list of exec queues attached to this VM */
+ structlist_headlist;
+ /**
+ * @exec_queues.count: count of exec queues attached to this VM,
+ * per GT
+ */
+ int count[XE_MAX_TILES_PER_DEVICE *XE_MAX_GT_PER_TILE];
+ /** @exec_queues.lock: lock to protectexec_queueslist */
+ structrw_semaphorelock;
+ }exec_queues;
+
/** @um: unified memory state */
struct {
/** @asid: address space ID, unique to each VM */
Re: [PATCH v2 04/12] drm/xe: Add vm to exec queues association
In reply to Matthew BrostQuoted 92 lines · Matthew Brost
Maintain a list of exec queues per vm which will be used by TLB invalidation code to do context-ID based tlb invalidations.
Signed-off-by Nirmoy Das <[email protected]>Signed-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device.h | 7 ---- drivers/gpu/drm/xe/xe_device_types.h | 7 ++++ drivers/gpu/drm/xe/xe_exec_queue.c | 7 +++-drivers/gpu/drm/xe/xe_exec_queue_types.h| 3 ++ drivers/gpu/drm/xe/xe_vm.c | 47
++++++++++++++++++++++++
drivers/gpu/drm/xe/xe_vm.h | 3 ++ drivers/gpu/drm/xe/xe_vm_types.h | 13 +++++++7 files changed, 79 insertions(+), 8 deletions(-)drivers/gpu/drm/xe/xe_device.h +0
-0@@ -62,13 +62,6 @@ static inline struct xe_tile *xe_device_get_root_tile(struct xe_device *xe) return &xe->tiles[0]; }
-/*
- * Highest GT/tile count for any platform. Used only for memory allocation
- * sizing. Any logic looping over GTs or mapping userspace GT IDs into GT
- * structures should use the per-platformxe->info.max_gt_per_tileinstead.
- */drivers/gpu/drm/xe/xe_device.h +0
-2-#define XE_MAX_GT_PER_TILE 2 -static inline struct
xe_gt*xe_device_get_gt(structxe_device*xe, u8gt_id)
{
structxe_tile*tile;drivers/gpu/drm/xe/xe_device_types.h +0
-0@@ -57,6 +57,13 @@ struct xe_vram_region; #define XE_GT1 1 #define XE_MAX_TILES_PER_DEVICE (XE_GT1 + 1)
+/*
+ * Highest GT/tile count for any platform. Used only for memory allocation
+ * sizing. Any logic looping over GTs or mapping userspace GT IDs into GT
+ * structures should use the per-platformxe->info.max_gt_per_tileinstead.
+ */drivers/gpu/drm/xe/xe_device_types.h +2
-0+#define XE_MAX_GT_PER_TILE 2 +#define XE_MAX_ASID (BIT(20))
#defineIS_PLATFORM_STEP(_xe, _platform, min_step, max_step)\drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -72,8 +72,10 @@ static void __xe_exec_queue_free(struct xe_exec_queue *q)
if (xe_exec_queue_uses_pxp(q))xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q); - if (q->vm)
+ if (q->vm) {
+ xe_vm_remove_exec_queue(q->vm, q);xe_vm_put(q->vm);
+ }
if (q->xef)
xe_file_put(q->xef);drivers/gpu/drm/xe/xe_exec_queue.c +1
-0@@ -143,6 +145,7 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, q->ring_ops = gt->ring_ops[hwe->class]; q->ops = gt->exec_queue_ops; INIT_LIST_HEAD(&q->lr.link); + INIT_LIST_HEAD(&q->vm_exec_queue_link); INIT_LIST_HEAD(&q->multi_gt_link); INIT_LIST_HEAD(&q->hw_engine_group_link);INIT_LIST_HEAD(&q->pxp.link);
drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -796,6 +799,8 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data, }
Quoted 4 lines
q->xef =xe_file_get(xef);
+ if (eci[0].engine_class !=DRM_XE_ENGINE_CLASS_VM_BIND) + xe_vm_add_exec_queue(vm, q);
Discussed this offline, but because of potential memory corruption in the register/deregister path, the plan is to continue for now in the queue create/destroy. This means we'll get errors from the GuC when we race for the register/deregister and submission which will be addressed in some later cleanup. Since this feature is blocked behind a feature flag anyway for testing, it shouldn't have any other adverse effects.
One other minor comment below...
Quoted 53 lines
/* user id alloc must always be last in ioctl to prevent UAF */
err =xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL);drivers/gpu/drm/xe/xe_exec_queue_types.h +0
-0@@ -147,6 +147,9 @@ struct xe_exec_queue { struct xe_dep_scheduler *dep_scheduler; } tlb_inval[XE_EXEC_QUEUE_TLB_INVAL_COUNT];
+ /** @vm_exec_queue_link: Link to track exec queue within a VM's list of exec queues. */
+ structlist_headvm_exec_queue_link;
+
/** @pxp: PXP info tracking */
struct {
/** @pxp.type: PXP session type used by this queue */drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -1507,8 +1507,20 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags, struct xe_file *xef) INIT_WORK(&vm->destroy_work, vm_destroy_work_func);
INIT_LIST_HEAD(&vm->preempt.exec_queues);
+ INIT_LIST_HEAD(&vm->exec_queues.list);vm->preempt.min_run_period_ms= 10; /* FIXME: Wire up to uAPI */
+ init_rwsem(&vm->exec_queues.lock);
+ if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
+ fs_reclaim_acquire(GFP_KERNEL);
+ might_lock(&vm->exec_queues.lock);
+ fs_reclaim_release(GFP_KERNEL);drivers/gpu/drm/xe/xe_vm.c +6
-0+ + down_read(&vm->exec_queues.lock); + might_lock(&xe_root_mmio_gt(xe)->uc.guc.ct.lock); + up_read(&vm->exec_queues.lock); + } +for_each_tile(tile, xe, id)
xe_range_fence_tree_init(&vm->rftree[id]);drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -4387,3 +4399,38 @@ int xe_vm_alloc_cpu_addr_mirror_vma(struct
xe_vm*vm,uint64_tstart,uint64_tr
returnxe_vm_alloc_vma(vm, &map_req, false); }
drivers/gpu/drm/xe/xe_vm.c int xe_vm_alloc_cpu_addr_mirror_vma(struct xe_vm *vm, uint64_t start, uint64_t r · line 4410
Quoted 9 lines
drivers/gpu/drm/xe/xe_vm.c +9
-0+ +/** +* xe_vm_add_exec_queue() - Add exec queue to VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_add_exec_queue (struct xe_vm *vm,struct xe_exec_queue *q) +{ + /* User VMs and queues only */
Why? So the expectation is we always do a full invalidation for kernel VMs? Do we want the potential performance impact of that, particularly for the migration queues?
I get that in the current version we are doing this as an explicit result of a user queue create, but I don't otherwise see why we need these restrictions.
Thanks,
Stuart
Quoted 75 lines
+ xe_assert(vm->xe, !(q->flags &
EXEC_QUEUE_FLAG_KERNEL)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_PERMANENT)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_VM)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_MIGRATE)); + xe_assert(vm->xe,vm->xef);drivers/gpu/drm/xe/xe_vm.c +13
-0+ + down_write(&vm->exec_queues.lock); + list_add(&q->vm_exec_queue_link, &vm->exec_queues.list); + ++vm->exec_queues.count[q->gt->info.id]; + up_write(&vm->exec_queues.lock); +} + +/** +* xe_vm_remove_exec_queue() - Remove exec queue from VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_remove_exec_queue(struct xe_vm *vm, struct xe_exec_queue*q)
drivers/gpu/drm/xe/xe_vm.c +8
-0+{ + down_write(&vm->exec_queues.lock); + if (!list_empty(&q->vm_exec_queue_link)) { + list_del(&q->vm_exec_queue_link); + --vm->exec_queues.count[q->gt->info.id]; + } + up_write(&vm->exec_queues.lock); +}drivers/gpu/drm/xe/xe_vm.h +0
-0@@ -284,6 +284,9 @@ static inline struct dma_resv *xe_vm_resv(struct xe_vm *vm)
voidxe_vm_kill(struct xe_vm *vm, bool unlocked);
+voidxe_vm_add_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+voidxe_vm_remove_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+
/**
*xe_vm_assert_held(vm)- Assert that the vm's reservation object is held.
* @vm: The vmdrivers/gpu/drm/xe/xe_vm_types.h +0
-0@@ -290,6 +290,19 @@ struct xe_vm { struct list_head pm_activate_link; } preempt;
+ /** @exec_queues: Manages list of exec queues attached to this VM, protected by lock. */
+ struct {
+ /** @exec_queues.list: list of exec queues attached to this VM */
+ structlist_headlist;
+ /**
+ * @exec_queues.count: count of exec queues attached to this VM,
+ * per GT
+ */
+ int count[XE_MAX_TILES_PER_DEVICE *XE_MAX_GT_PER_TILE];
+ /** @exec_queues.lock: lock to protectexec_queueslist */
+ structrw_semaphorelock;
+ }exec_queues;
+
/** @um: unified memory state */
struct {
/** @asid: address space ID, unique to each VM */
Matthew Brost
maintainerRe: [PATCH v2 04/12] drm/xe: Add vm to exec queues association
Quoted 165 lines · Stuart
Quoted 92 lines · Matthew Brost
Maintain a list of exec queues per vm which will be used by TLB invalidation code to do context-ID based tlb invalidations.
Signed-off-by Nirmoy Das <[email protected]>Signed-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device.h | 7 ---- drivers/gpu/drm/xe/xe_device_types.h | 7 ++++ drivers/gpu/drm/xe/xe_exec_queue.c | 7 +++-drivers/gpu/drm/xe/xe_exec_queue_types.h| 3 ++ drivers/gpu/drm/xe/xe_vm.c | 47
++++++++++++++++++++++++
drivers/gpu/drm/xe/xe_vm.h | 3 ++ drivers/gpu/drm/xe/xe_vm_types.h | 13 +++++++7 files changed, 79 insertions(+), 8 deletions(-)drivers/gpu/drm/xe/xe_device.h +0
-0@@ -62,13 +62,6 @@ static inline struct xe_tile *xe_device_get_root_tile(struct xe_device *xe) return &xe->tiles[0]; }
-/*
- * Highest GT/tile count for any platform. Used only for memory allocation
- * sizing. Any logic looping over GTs or mapping userspace GT IDs into GT
- * structures should use the per-platformxe->info.max_gt_per_tileinstead.
- */drivers/gpu/drm/xe/xe_device.h +0
-2-#define XE_MAX_GT_PER_TILE 2 -static inline struct
xe_gt*xe_device_get_gt(structxe_device*xe, u8gt_id)
{
structxe_tile*tile;drivers/gpu/drm/xe/xe_device_types.h +0
-0@@ -57,6 +57,13 @@ struct xe_vram_region; #define XE_GT1 1 #define XE_MAX_TILES_PER_DEVICE (XE_GT1 + 1)
+/*
+ * Highest GT/tile count for any platform. Used only for memory allocation
+ * sizing. Any logic looping over GTs or mapping userspace GT IDs into GT
+ * structures should use the per-platformxe->info.max_gt_per_tileinstead.
+ */drivers/gpu/drm/xe/xe_device_types.h +2
-0+#define XE_MAX_GT_PER_TILE 2 +#define XE_MAX_ASID (BIT(20))
#defineIS_PLATFORM_STEP(_xe, _platform, min_step, max_step)\drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -72,8 +72,10 @@ static void __xe_exec_queue_free(struct xe_exec_queue *q)
if (xe_exec_queue_uses_pxp(q))xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q); - if (q->vm)
+ if (q->vm) {
+ xe_vm_remove_exec_queue(q->vm, q);xe_vm_put(q->vm);
+ }
if (q->xef)
xe_file_put(q->xef);drivers/gpu/drm/xe/xe_exec_queue.c +1
-0@@ -143,6 +145,7 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, q->ring_ops = gt->ring_ops[hwe->class]; q->ops = gt->exec_queue_ops; INIT_LIST_HEAD(&q->lr.link); + INIT_LIST_HEAD(&q->vm_exec_queue_link); INIT_LIST_HEAD(&q->multi_gt_link); INIT_LIST_HEAD(&q->hw_engine_group_link);INIT_LIST_HEAD(&q->pxp.link);
drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -796,6 +799,8 @@ int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data, }Quoted 4 lines
q->xef =xe_file_get(xef);
+ if (eci[0].engine_class !=DRM_XE_ENGINE_CLASS_VM_BIND) + xe_vm_add_exec_queue(vm, q);Discussed this offline, but because of potential memory corruption in the register/deregister path, the plan is to continue for now in the queue create/destroy. This means we'll get errors from the GuC when we race for the register/deregister and submission which will be addressed in some later cleanup. Since this feature is blocked behind a feature flag anyway for testing, it shouldn't have any other adverse effects.
One other minor comment below...
Quoted 58 lines
/* user id alloc must always be last in ioctl to prevent UAF */
err =xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL);drivers/gpu/drm/xe/xe_exec_queue_types.h +0
-0@@ -147,6 +147,9 @@ struct xe_exec_queue { struct xe_dep_scheduler *dep_scheduler; } tlb_inval[XE_EXEC_QUEUE_TLB_INVAL_COUNT];
+ /** @vm_exec_queue_link: Link to track exec queue within a VM's list of exec queues. */
+ structlist_headvm_exec_queue_link;
+
/** @pxp: PXP info tracking */
struct {
/** @pxp.type: PXP session type used by this queue */drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -1507,8 +1507,20 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32 flags, struct xe_file *xef) INIT_WORK(&vm->destroy_work, vm_destroy_work_func);
INIT_LIST_HEAD(&vm->preempt.exec_queues);
+ INIT_LIST_HEAD(&vm->exec_queues.list);vm->preempt.min_run_period_ms= 10; /* FIXME: Wire up to uAPI */
+ init_rwsem(&vm->exec_queues.lock);
+ if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
+ fs_reclaim_acquire(GFP_KERNEL);
+ might_lock(&vm->exec_queues.lock);
+ fs_reclaim_release(GFP_KERNEL);drivers/gpu/drm/xe/xe_vm.c +6
-0+ + down_read(&vm->exec_queues.lock); + might_lock(&xe_root_mmio_gt(xe)->uc.guc.ct.lock); + up_read(&vm->exec_queues.lock); + } +for_each_tile(tile, xe, id)
xe_range_fence_tree_init(&vm->rftree[id]);drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -4387,3 +4399,38 @@ int xe_vm_alloc_cpu_addr_mirror_vma(struct
xe_vm*vm,uint64_tstart,uint64_tr
returnxe_vm_alloc_vma(vm, &map_req, false); }drivers/gpu/drm/xe/xe_vm.c +5
-0+ +/** +* xe_vm_add_exec_queue() - Add exec queue to VM +* @vm: The VM. +* @q: The exec_queue
Quoted 9 lines
drivers/gpu/drm/xe/xe_vm.c int xe_vm_alloc_cpu_addr_mirror_vma(struct xe_vm *vm, uint64_t start, uint64_t r · line 4410
Quoted 4 lines
+ */
drivers/gpu/drm/xe/xe_vm.c +3
-0+void xe_vm_add_exec_queue (struct xe_vm *vm,struct xe_exec_queue *q) +{ + /* User VMs and queues only */Why? So the expectation is we always do a full invalidation for kernel VMs? Do we want the potential performance impact of that, particularly for the migration queues?
We don't issue TLB invalidations via H2G on kernel VMs; instead, we insert ring instructions to perform them. We fully control kernel VMs and queues, so we can trust that TLB invalidations are correctly inserted. TLB invalidations are only issued when we have a VMA and attempt to move the backing memory or perform an unbind. The migration VM doesn't have VMAs; rather, it has self-mapped page tables that are dynamically programmed to perform operations such as clears, copies, or updating another VM's page tables. PXP does have a single VMA, but the memory it backing it is pinned and never unbound.
Quoted 3 lines
I get that in the current version we are doing this as an explicit result of a user queue create, but I don't otherwise see why we need these restrictions.
It is about ensuring correct usage in the current code. If we need to add invalidations to a kernel VM, we can remove these, but any change like that would require a proper review. That said, we likely should have more asserts in the TLB invalidation layer(s) to ensure we are not issuing TLB invalidations on kernel VMs by accident.
Matt
Quoted 80 lines
Thanks,
StuartQuoted 75 lines
+ xe_assert(vm->xe, !(q->flags &
EXEC_QUEUE_FLAG_KERNEL)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_PERMANENT)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_VM)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_MIGRATE)); + xe_assert(vm->xe,vm->xef);drivers/gpu/drm/xe/xe_vm.c +13
-0+ + down_write(&vm->exec_queues.lock); + list_add(&q->vm_exec_queue_link, &vm->exec_queues.list); + ++vm->exec_queues.count[q->gt->info.id]; + up_write(&vm->exec_queues.lock); +} + +/** +* xe_vm_remove_exec_queue() - Remove exec queue from VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_remove_exec_queue(struct xe_vm *vm, struct xe_exec_queue*q)
drivers/gpu/drm/xe/xe_vm.c +8
-0+{ + down_write(&vm->exec_queues.lock); + if (!list_empty(&q->vm_exec_queue_link)) { + list_del(&q->vm_exec_queue_link); + --vm->exec_queues.count[q->gt->info.id]; + } + up_write(&vm->exec_queues.lock); +}drivers/gpu/drm/xe/xe_vm.h +0
-0@@ -284,6 +284,9 @@ static inline struct dma_resv *xe_vm_resv(struct xe_vm *vm)
voidxe_vm_kill(struct xe_vm *vm, bool unlocked);
+voidxe_vm_add_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+voidxe_vm_remove_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+
/**
*xe_vm_assert_held(vm)- Assert that the vm's reservation object is held.
* @vm: The vmdrivers/gpu/drm/xe/xe_vm_types.h +0
-0@@ -290,6 +290,19 @@ struct xe_vm { struct list_head pm_activate_link; } preempt;
+ /** @exec_queues: Manages list of exec queues attached to this VM, protected by lock. */
+ struct {
+ /** @exec_queues.list: list of exec queues attached to this VM */
+ structlist_headlist;
+ /**
+ * @exec_queues.count: count of exec queues attached to this VM,
+ * per GT
+ */
+ int count[XE_MAX_TILES_PER_DEVICE *XE_MAX_GT_PER_TILE];
+ /** @exec_queues.lock: lock to protectexec_queueslist */
+ structrw_semaphorelock;
+ }exec_queues;
+
/** @um: unified memory state */
struct {
/** @asid: address space ID, unique to each VM */
Summers, Stuart
Reviewed-byRe: [PATCH v2 04/12] drm/xe: Add vm to exec queues association
Quoted 216 lines · Matthew Brost
Quoted 187 lines · Stuart
Quoted 96 lines · Matthew Brost
Maintain a list of exec queues per vm which will be used by TLB invalidation code to do context-ID based tlb invalidations.
Signed-off-by Nirmoy Das <[email protected]>Signed-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device.h | 7 ---- drivers/gpu/drm/xe/xe_device_types.h | 7 ++++ drivers/gpu/drm/xe/xe_exec_queue.c | 7 +++-drivers/gpu/drm/xe/xe_exec_queue_types.h| 3 ++ drivers/gpu/drm/xe/xe_vm.c | 47
++++++++++++++++++++++++
drivers/gpu/drm/xe/xe_vm.h | 3 ++ drivers/gpu/drm/xe/xe_vm_types.h | 13 +++++++7 files changed, 79 insertions(+), 8 deletions(-)drivers/gpu/drm/xe/xe_device.h +0
-0@@ -62,13 +62,6 @@ static inline struct xe_tile *xe_device_get_root_tile(struct xe_device *xe) return &xe->tiles[0]; }
-/*
- * Highest GT/tile count for any platform. Used only for memory allocation
- * sizing. Any logic looping over GTs or mapping userspace GT IDs
into GT
- * structures should use the per-platform xe-Quoted 1 line
info.max_gt_per_tile
instead.
- */drivers/gpu/drm/xe/xe_device.h +0
-2-#define XE_MAX_GT_PER_TILE 2 -static inline struct
xe_gt*xe_device_get_gt(structxe_device*xe,
u8gt_id)
{
structxe_tile*tile;drivers/gpu/drm/xe/xe_device_types.h +0
-0@@ -57,6 +57,13 @@ struct xe_vram_region; #define XE_GT1 1 #define XE_MAX_TILES_PER_DEVICE (XE_GT1 + 1)
+/*
+ * Highest GT/tile count for any platform. Used only for memory allocation
+ * sizing. Any logic looping over GTs or mapping userspace GT IDs
into GT
+ * structures should use the per-platform xe-Quoted 1 line
info.max_gt_per_tile
instead.
+ */drivers/gpu/drm/xe/xe_device_types.h +2
-0+#define XE_MAX_GT_PER_TILE 2 +#define XE_MAX_ASID (BIT(20))
#defineIS_PLATFORM_STEP(_xe, _platform, min_step, max_step)\drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -72,8 +72,10 @@ static void __xe_exec_queue_free(struct xe_exec_queue *q)
if (xe_exec_queue_uses_pxp(q))xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q);
- if (q->vm)
+ if (q->vm) {
+ xe_vm_remove_exec_queue(q->vm, q);xe_vm_put(q->vm);
+ }
if (q->xef)
xe_file_put(q->xef);drivers/gpu/drm/xe/xe_exec_queue.c +1
-0@@ -143,6 +145,7 @@ static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe, q->ring_ops = gt->ring_ops[hwe->class]; q->ops = gt->exec_queue_ops; INIT_LIST_HEAD(&q->lr.link); + INIT_LIST_HEAD(&q->vm_exec_queue_link); INIT_LIST_HEAD(&q->multi_gt_link); INIT_LIST_HEAD(&q->hw_engine_group_link);INIT_LIST_HEAD(&q->pxp.link);
drivers/gpu/drm/xe/xe_exec_queue.c +0
-0@@ -796,6 +799,8 @@ int xe_exec_queue_create_ioctl(structQuoted 7 lines
drm_device
*dev, void *data,
}
q->xef =xe_file_get(xef);
+ if (eci[0].engine_class !=DRM_XE_ENGINE_CLASS_VM_BIND) + xe_vm_add_exec_queue(vm, q);Discussed this offline, but because of potential memory corruption in
the register/deregister path, the plan is to continue for now in the
queue create/destroy. This means we'll get errors from the GuC when we
race for the register/deregister and submission which will be addressed
in some later cleanup. Since this feature is blocked behind a feature
flag anyway for testing, it shouldn't have any other adverse effects.One other minor comment below...
Quoted 67 lines
/* user id alloc must always be last in ioctl to prevent UAF
*/
err =xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL);drivers/gpu/drm/xe/xe_exec_queue_types.h +0
-0@@ -147,6 +147,9 @@ struct xe_exec_queue { struct xe_dep_scheduler *dep_scheduler; } tlb_inval[XE_EXEC_QUEUE_TLB_INVAL_COUNT];
+ /** @vm_exec_queue_link: Link to track exec queue within a
VM's list of exec queues. */
+ structlist_headvm_exec_queue_link;
+
/** @pxp: PXP info tracking */
struct {
/** @pxp.type: PXP session type used by this queue */drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -1507,8 +1507,20 @@ struct xe_vm *xe_vm_create(struct
xe_device
*xe, u32 flags, structxe_file*xef)INIT_WORK(&vm->destroy_work, vm_destroy_work_func);
INIT_LIST_HEAD(&vm->preempt.exec_queues);
+ INIT_LIST_HEAD(&vm->exec_queues.list);vm->preempt.min_run_period_ms= 10; /* FIXME: Wire up to
uAPI */
+ init_rwsem(&vm->exec_queues.lock);
+ if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
+ fs_reclaim_acquire(GFP_KERNEL);
+ might_lock(&vm->exec_queues.lock);
+ fs_reclaim_release(GFP_KERNEL);drivers/gpu/drm/xe/xe_vm.c +6
-0+ + down_read(&vm->exec_queues.lock); + might_lock(&xe_root_mmio_gt(xe)->uc.guc.ct.lock); + up_read(&vm->exec_queues.lock); + } +for_each_tile(tile, xe, id)
xe_range_fence_tree_init(&vm->rftree[id]);drivers/gpu/drm/xe/xe_vm.c +0
-0@@ -4387,3 +4399,38 @@ int xe_vm_alloc_cpu_addr_mirror_vma(struct
xe_vm*vm,uint64_tstart,uint64_tr
returnxe_vm_alloc_vma(vm, &map_req, false); }drivers/gpu/drm/xe/xe_vm.c +7
-0+ +/** +* xe_vm_add_exec_queue() - Add exec queue to VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_add_exec_queue(struct xe_vm *vm, struct xe_exec_queue*q)
Quoted 9 lines
Quoted 2 lines
+{
+ /* User VMs and queues only */Why? So the expectation is we always do a full invalidation for kernel
VMs? Do we want the potential performance impact of that, particularly
for the migration queues?We don't issue TLB invalidations via H2G on kernel VMs; instead, we insert ring instructions to perform them. We fully control kernel VMs and queues, so we can trust that TLB invalidations are correctly inserted. TLB invalidations are only issued when we have a VMA and attempt to move the backing memory or perform an unbind. The migration
VM doesn't have VMAs; rather, it has self-mapped page tables that are dynamically programmed to perform operations such as clears, copies, or
updating another VM's page tables. PXP does have a single VMA, but the
memory it backing it is pinned and never unbound.Quoted 4 lines
I get that in the current version we are doing this as an explicit result of a user queue create, but I don't otherwise see why we need
these restrictions.
Quoted 6 lines
It is about ensuring correct usage in the current code. If we need to add invalidations to a kernel VM, we can remove these, but any change like that would require a proper review. That said, we likely should have more asserts in the TLB invalidation layer(s) to ensure we are not
issuing TLB invalidations on kernel VMs by accident.
Ok yeah this all makes sense and thanks for the detail there. I agree with what you said here about ensuring we aren't accidentally going to this path from the kernel VMs.
Quoted 92 lines
Matt
Quoted 89 lines
Thanks,
StuartQuoted 84 lines
+ xe_assert(vm->xe, !(q->flags &
EXEC_QUEUE_FLAG_KERNEL)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_PERMANENT));
+ xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_VM)); + xe_assert(vm->xe, !(q->flags &EXEC_QUEUE_FLAG_MIGRATE)); + xe_assert(vm->xe,vm->xef);drivers/gpu/drm/xe/xe_vm.c +13
-0+ + down_write(&vm->exec_queues.lock); + list_add(&q->vm_exec_queue_link, &vm->exec_queues.list); + ++vm->exec_queues.count[q->gt->info.id]; + up_write(&vm->exec_queues.lock); +} + +/** +* xe_vm_remove_exec_queue() - Remove exec queue from VM +* @vm: The VM. +* @q: The exec_queue +*/ +void xe_vm_remove_exec_queue(struct xe_vm *vm, struct
xe_exec_queue
*q)drivers/gpu/drm/xe/xe_vm.c +8
-0+{ + down_write(&vm->exec_queues.lock); + if (!list_empty(&q->vm_exec_queue_link)) { + list_del(&q->vm_exec_queue_link); + --vm->exec_queues.count[q->gt->info.id]; + } + up_write(&vm->exec_queues.lock); +}drivers/gpu/drm/xe/xe_vm.h +0
-0@@ -284,6 +284,9 @@ static inline struct dma_resv *xe_vm_resv (struct xe_vm *vm)
voidxe_vm_kill(struct xe_vm *vm, bool unlocked);
+voidxe_vm_add_exec_queue(struct xe_vm *vm, struct xe_exec_queue *q);
+void xe_vm_remove_exec_queue(structxe_vm*vm, structxe_exec_queue
*q);
+
/**
*xe_vm_assert_held(vm)- Assert that the vm's reservation object
is held.
* @vm: The vmdrivers/gpu/drm/xe/xe_vm_types.h +0
-0@@ -290,6 +290,19 @@ struct xe_vm { struct list_head pm_activate_link; } preempt;
+ /** @exec_queues: Manages list of exec queues attached to this VM, protected by lock. */
+ struct {
+ /** @exec_queues.list: list of exec queues attached
to this VM */
+ structlist_headlist;
+ /**
+ * @exec_queues.count: count of exec queues attached
to this VM,
+ * per GT
+ */
+ int count[XE_MAX_TILES_PER_DEVICE *XE_MAX_GT_PER_TILE];
+ /** @exec_queues.lock: lock to protectexec_queues
list */
+ structrw_semaphorelock;
+ }exec_queues;
+
/** @um: unified memory state */
struct {
/** @asid: address space ID, unique to each VM */
Matthew Brost
maintainerPATCH v2 5/12 drm/xe: Taint TLB invalidation seqno lock with GFP_KERNEL
In reply to Matthew BrostTaint TLB invalidation seqno lock with GFP_KERNEL as TLB invalidations can be in the path of reclaim (e.g., MMU notifiers).
--- drivers/gpu/drm/xe/xe_tlb_inval.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)
drivers/gpu/drm/xe/xe_tlb_inval.c +12 -0
@@ -114,6 +114,16 @@ static void tlb_inval_fini(struct drm_device *drm, void *arg)xe_tlb_inval_reset (tlb_inval); } +static void primelockdep (struct xe_tlb_inval *tlb_inval) +{ +if (!IS_ENABLED (CONFIG_LOCKDEP )) +return ; + +fs_reclaim_acquire (GFP_KERNEL ); +might_lock (&tlb_inval->seqno_lock); +fs_reclaim_release (GFP_KERNEL ); +} +/** * xe_gt_tlb_inval_init - Initialize TLB invalidation state * @gt: GT structure
@@ -140,6 +150,8 @@ int xe_gt_tlb_inval_init_early(struct xe_gt *gt)if (err)return err; +primelockdep (tlb_inval); + tlb_inval->job_wq =drmm_alloc_ordered_workqueue (&xe->drm,"gt-tbl-inval-job-wq" ,WQ_MEM_RECLAIM );
Signature · 1 line
2.34.1
Re: [PATCH v2 05/12] drm/xe: Taint TLB invalidation seqno lock with GFP_KERNEL
Quoted 4 lines · Matthew Brost
Taint TLB invalidation seqno lock with
GFP_KERNELas TLB invalidations can be in the path of reclaim (e.g., MMU notifiers).Signed-off-by Matthew Brost <[email protected]>
Quoted 37 lines
--- drivers/gpu/drm/xe/xe_tlb_inval.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)drivers/gpu/drm/xe/xe_tlb_inval.c +12
-0@@ -114,6 +114,16 @@ static void tlb_inval_fini(struct drm_device *drm, void *arg)xe_tlb_inval_reset (tlb_inval); } +static void primelockdep (struct xe_tlb_inval *tlb_inval) +{ +if (!IS_ENABLED (CONFIG_LOCKDEP )) +return ; + +fs_reclaim_acquire (GFP_KERNEL ); +might_lock (&tlb_inval->seqno_lock); +fs_reclaim_release (GFP_KERNEL ); +} +/** * xe_gt_tlb_inval_init - Initialize TLB invalidation state * @gt: GT structure @@ -140,6 +150,8 @@ int xe_gt_tlb_inval_init_early(struct xe_gt *gt)if (err)return err; +primelockdep (tlb_inval); + tlb_inval->job_wq =drmm_alloc_ordered_workqueue (&xe->drm,"gt-tbl-inval-job-wq" ,WQ_MEM_RECLAIM );Signature · 1 line
2.34.1
Matthew Brost
maintainerPATCH v2 6/12 drm/xe: Do not forward invalid TLB invalidation seqnos to upper layers
In reply to Matthew BrostContext-based TLB invalidations send multiple H2G messages per seqno, with only the final H2G containing a valid seqno — the others carry an invalid seqno. The G2H handler drops these invalid seqnos to avoid prematurely signaling a TLB invalidation fence.
--- drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 3 ++- drivers/gpu/drm/xe/xe_tlb_inval_types.h | 1 + 2 files changed, 3 insertions(+), 1 deletion(-)
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +2 -1
@@ -235,7 +235,8 @@ int xe_guc_tlb_inval_done_handler(struct xe_guc *guc, u32 *msg, u32 len)if (unlikely (len !=1 ))return -EPROTO ; -xe_tlb_inval_done_handler (>->tlb_inval, msg[0 ]); +if (msg[0 ] !=TLB_INVALIDATION_SEQNO_INVALID ) +xe_tlb_inval_done_handler (>->tlb_inval, msg[0 ]);return 0 ; }
drivers/gpu/drm/xe/xe_tlb_inval_types.h +1 -0
@@ -80,6 +80,7 @@ struct xe_tlb_inval {const struct xe_tlb_inval_ops *ops;/** @tlb_inval.seqno: TLB invalidation seqno, protected by CT lock */ #define TLB_INVALIDATION_SEQNO_MAX 0x100000 +#define TLB_INVALIDATION_SEQNO_INVALID TLB_INVALIDATION_SEQNO_MAXint seqno;/** @tlb_invalidation.seqno_lock: protects @tlb_invalidation.seqno */ struct mutex seqno_lock;
Signature · 1 line
2.34.1
Re: [PATCH v2 06/12] drm/xe: Do not forward invalid TLB invalidation seqnos to upper layers
Quoted 21 lines · Matthew Brost
Context-based TLB invalidations send multiple H2G messages per seqno, with only the final H2G containing a valid seqno — the others carry an
invalid seqno. The G2H handler drops these invalid seqnos to avoid prematurely signaling a TLB invalidation fence.Signed-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 3 ++-
drivers/gpu/drm/xe/xe_tlb_inval_types.h | 1 +2 files changed, 3 insertions(+), 1 deletion(-)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -235,7 +235,8 @@ int xe_guc_tlb_inval_done_handler(struct xe_guc *guc, u32 *msg, u32 len) if (unlikely(len != 1)) return -EPROTO;
Quoted 4 lines
- xe_tlb_inval_done_handler(>->tlb_inval, msg[0]); + if (msg[0] !=TLB_INVALIDATION_SEQNO_INVALID)
+ xe_tlb_inval_done_handler(>->tlb_inval, msg[0]);
Honestly I feel like this should be included as part of the patch that actually uses it. No problem with the actual contents though.
Thanks,
Stuart
Quoted 18 lines
return 0;
}drivers/gpu/drm/xe/xe_tlb_inval_types.h +0
-0@@ -80,6 +80,7 @@ struct xe_tlb_inval { const struct xe_tlb_inval_ops *ops; /** @tlb_inval.seqno: TLB invalidation seqno, protected by CTlock */
#define TLB_INVALIDATION_SEQNO_MAX 0x100000 +#define TLB_INVALIDATION_SEQNO_INVALID TLB_INVALIDATION_SEQNO_MAX int seqno;
/** @tlb_invalidation.seqno_lock: protects @tlb_invalidation.seqno */
struct mutexseqno_lock;
Matthew Brost
maintainerPATCH v2 7/12 drm/xe: Rename send_tlb_inval_ppgtt to send_tlb_inval_asid_ppgtt
In reply to Matthew BrostContext-based TLB invalidations have their own set of GuC TLB invalidation operations. Rename the current PPGTT invalidation function, which operates on ASIDs, to a more descriptive name that reflects its purpose.
--- drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +3 -3
@@ -131,8 +131,8 @@ static u64 normalize_invalidation_range(struct xe_gt *gt, u64 *start, u64 *end)*/ #define MAX_RANGE_TLB_INVALIDATION_LENGTH (rounddown_pow_of_two(ULONG_MAX)) -static int send_tlb_inval_ppgtt (struct xe_tlb_inval *tlb_inval,u32 seqno, -u64 start,u64 end,u32 asid) +static int send_tlb_inval_asid_ppgtt (struct xe_tlb_inval *tlb_inval,u32 seqno, +u64 start,u64 end,u32 asid) {#define MAX_TLB_INVALIDATION_LEN 7struct xe_guc *guc = tlb_inval->private;
@@ -195,7 +195,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval)static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .all = send_tlb_inval_all, .ggtt = send_tlb_inval_ggtt, - .ppgtt = send_tlb_inval_ppgtt, + .ppgtt = send_tlb_inval_asid_ppgtt, .initialized = tlb_inval_initialized, .flush = tlb_inval_flush, .timeout_delay = tlb_inval_timeout_delay,
Signature · 1 line
2.34.1
Re: [PATCH v2 07/12] drm/xe: Rename send_tlb_inval_ppgtt to send_tlb_inval_asid_ppgtt
Quoted 7 lines · Matthew Brost
Context-based TLB invalidations have their own set of GuC TLB invalidation operations. Rename the current PPGTT invalidation function,
which operates on ASIDs, to a more descriptive name that reflects its purpose.Signed-off-by Matthew Brost <[email protected]>
Quoted 34 lines
---
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 6 +++---1 file changed, 3 insertions(+), 3 deletions(-)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -131,8 +131,8 @@ static u64 normalize_invalidation_range(structxe_gt *gt, u64 *start, u64 *end) */ #define MAX_RANGE_TLB_INVALIDATION_LENGTH (rounddown_pow_of_two (ULONG_MAX ))
-static int send_tlb_inval_ppgtt(structxe_tlb_inval*tlb_inval, u32 seqno,
- u64 start, u64 end, u32 asid) +static int send_tlb_inval_asid_ppgtt(structxe_tlb_inval*tlb_inval, u32 seqno,
+ u64 start, u64 end, u32 asid) {
#define MAX_TLB_INVALIDATION_LEN 7
structxe_guc*guc =tlb_inval->private;drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-1@@ -195,7 +195,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval) static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .all = send_tlb_inval_all, .ggtt = send_tlb_inval_ggtt, - .ppgtt = send_tlb_inval_ppgtt, + .ppgtt = send_tlb_inval_asid_ppgtt, .initialized = tlb_inval_initialized, .flush = tlb_inval_flush,.timeout_delay =
tlb_inval_timeout_delay,
Matthew Brost
maintainerPATCH v2 8/12 drm/xe: Add send_tlb_inval_ppgtt helper
In reply to Matthew BrostExtract the common code that issues a TLB invalidation H2G for PPGTTs into a helper function. This helper can be reused for both ASID-based and context-based TLB invalidations.
--- drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 27 +++++++++++++++++++-------- 1 file changed, 19 insertions(+), 8 deletions(-)
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +19 -8
@@ -131,19 +131,15 @@ static u64 normalize_invalidation_range(struct xe_gt *gt, u64 *start, u64 *end)*/ #define MAX_RANGE_TLB_INVALIDATION_LENGTH (rounddown_pow_of_two(ULONG_MAX)) -static int send_tlb_inval_asid_ppgtt (struct xe_tlb_inval *tlb_inval,u32 seqno, -u64 start,u64 end,u32 asid) +static int send_tlb_inval_ppgtt (struct xe_guc *guc,u32 seqno,u64 start, +u64 end,u32 id,u32 type) {#define MAX_TLB_INVALIDATION_LEN 7 -struct xe_guc *guc = tlb_inval->private;struct xe_gt *gt =guc_to_gt (guc);u32 action[MAX_TLB_INVALIDATION_LEN ];u64 length = end - start;int len =0 ; -if (guc_to_xe (guc)->info.force_execlist) -return -ECANCELED ; - action[len++] =XE_GUC_ACTION_TLB_INVALIDATION ; action[len++] = seqno; if (!gt_to_xe(gt)->info.has_range_tlb_inval ||
@@ -153,18 +149,33 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno,u64 normalize_len =normalize_invalidation_range (gt, &start, &end); - action[len++] =MAKE_INVAL_OP (XE_GUC_TLB_INVAL_PAGE_SELECTIVE ); - action[len++] = asid; + action[len++] =MAKE_INVAL_OP (type); + action[len++] = id; action[len++] =lower_32_bits (start); action[len++] =upper_32_bits (start); action[len++] =ilog2 (normalize_len) -ilog2 (SZ_4K ); }xe_gt_assert (gt, len <=MAX_TLB_INVALIDATION_LEN ); +#undef MAX_TLB_INVALIDATION_LENreturn send_tlb_inval (guc, action, len); } +static int send_tlb_inval_asid_ppgtt (struct xe_tlb_inval *tlb_inval,u32 seqno, +u64 start,u64 end,u32 asid) +{ +struct xe_guc *guc = tlb_inval->private; + +lockdep_assert_held (&tlb_inval->seqno_lock); + +if (guc_to_xe (guc)->info.force_execlist) +return -ECANCELED ; + +return send_tlb_inval_ppgtt (guc, seqno, start, end, asid, +XE_GUC_TLB_INVAL_PAGE_SELECTIVE ); +} +static bool tlb_inval_initialized (struct xe_tlb_inval *tlb_inval) {struct xe_guc *guc = tlb_inval->private;
Signature · 1 line
2.34.1
Re: [PATCH v2 08/12] drm/xe: Add send_tlb_inval_ppgtt helper
Quoted 5 lines · Matthew Brost
Extract the common code that issues a TLB invalidation H2G for PPGTTs into a helper function. This helper can be reused for both ASID-based and context-based TLB invalidations.
Signed-off-by Matthew Brost <[email protected]>
Quoted 77 lines
---
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 27 +++++++++++++++++++------ --1 file changed, 19 insertions(+), 8 deletions(-)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -131,19 +131,15 @@ static u64 normalize_invalidation_range(structxe_gt *gt, u64 *start, u64 *end) */ #define MAX_RANGE_TLB_INVALIDATION_LENGTH (rounddown_pow_of_two (ULONG_MAX ))
-static int send_tlb_inval_asid_ppgtt(structxe_tlb_inval*tlb_inval, u32 seqno,
- u64 start, u64 end, u32 asid) +static int send_tlb_inval_ppgtt(structxe_guc*guc, u32 seqno, u64 start,
+ u64 end, u32 id, u32 type) {
#define MAX_TLB_INVALIDATION_LEN 7
- structxe_guc*guc =tlb_inval->private;
structxe_gt*gt =guc_to_gt(guc);
u32 action[MAX_TLB_INVALIDATION_LEN];
u64 length = end - start;
int len = 0;
- if (guc_to_xe(guc)->info.force_execlist)
- return-ECANCELED;
-
action[len++] =XE_GUC_ACTION_TLB_INVALIDATION; action[len++] = seqno;
if (!gt_to_xe(gt)->info.has_range_tlb_inval ||drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -153,18 +149,33 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, u64 normalize_len = normalize_invalidation_range(gt, &start,
&end);
- action[len++] =MAKE_INVAL_OP(XE_GUC_TLB_INVAL_PAGE_SELECTIVE);
- action[len++] = asid;
+ action[len++] =MAKE_INVAL_OP(type);
+ action[len++] = id;
action[len++] =lower_32_bits(start); action[len++] =upper_32_bits(start); action[len++] =ilog2(normalize_len)-ilog2(SZ_4K); }
xe_gt_assert(gt, len <=MAX_TLB_INVALIDATION_LEN); +#undefMAX_TLB_INVALIDATION_LEN
returnsend_tlb_inval(guc, action, len);
}
+static int send_tlb_inval_asid_ppgtt(structxe_tlb_inval*tlb_inval, u32 seqno,
+ u64 start, u64 end, u32 asid)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +12
-0+{ + struct xe_guc *guc = tlb_inval->private; + + lockdep_assert_held(&tlb_inval->seqno_lock); + + if (guc_to_xe(guc)->info.force_execlist) + return -ECANCELED; + + return send_tlb_inval_ppgtt(guc, seqno, start, end, asid, + XE_GUC_TLB_INVAL_PAGE_SELECTIVE); +} +static bool
tlb_inval_initialized(struct xe_tlb_inval *tlb_inval){
structxe_guc*guc =tlb_inval->private;
Matthew Brost
maintainerPATCH v2 9/12 drm/xe: Add xe_tlb_inval_idle helper
In reply to Matthew BrostIntroduce the xe_tlb_inval_idle helper to detect whether any TLB invalidations are currently in flight. This is used in context-based TLB invalidations to determine whether dummy TLB invalidations need to be sent to maintain proper TLB invalidation fence ordering..
v2:
- Implement xe_tlb_inval_idle based on pending list
--- drivers/gpu/drm/xe/xe_tlb_inval.c | 21 +++++++++++++++++++++ drivers/gpu/drm/xe/xe_tlb_inval.h | 2 ++ 2 files changed, 23 insertions(+)
drivers/gpu/drm/xe/xe_tlb_inval.c +21 -0
@@ -44,11 +44,14 @@ static void xe_tlb_inval_fence_fini(struct xe_tlb_inval_fence *fence)static void xe_tlb_inval_fence_signal (struct xe_tlb_inval_fence *fence) { +struct xe_tlb_inval *tlb_inval = fence->tlb_inval;bool stack =test_bit (FENCE_STACK_BIT , &fence->base.flags);lockdep_assert_held (&fence->tlb_inval->pending_lock);list_del (&fence->link); +if (list_empty (&tlb_inval->pending_fences)) +cancel_delayed_work (&tlb_inval->fence_tdr);trace_xe_tlb_inval_fence_signal (fence->tlb_inval->xe, fence);xe_tlb_inval_fence_fini (fence);dma_fence_signal (&fence->base);
@@ -443,3 +446,21 @@ void xe_tlb_inval_fence_init(struct xe_tlb_inval *tlb_inval,dma_fence_get (&fence->base); fence->tlb_inval = tlb_inval; } + +/** +* xe_tlb_inval_idle() - Initialize TLB invalidation is idle +* @tlb_inval: TLB invalidation client +* +* Check the TLB invalidation seqno to determine if it is idle (i.e., no TLB +* invalidations are in flight). Expected to be called in the backend after the +* fence has been added to the pending list, and takes this into account. +* +* Return: True if TLB invalidation client is idle, False otherwise +*/ +bool xe_tlb_inval_idle (struct xe_tlb_inval *tlb_inval) +{ +lockdep_assert_held (&tlb_inval->seqno_lock); + +guard (spinlock_irq)(&tlb_inval->pending_lock); +return list_is_singular (&tlb_inval->pending_fences); +}
drivers/gpu/drm/xe/xe_tlb_inval.h +2 -0
@@ -43,4 +43,6 @@ xe_tlb_inval_fence_wait(struct xe_tlb_inval_fence *fence)void xe_tlb_inval_done_handler (struct xe_tlb_inval *tlb_inval,int seqno); +bool xe_tlb_inval_idle (struct xe_tlb_inval *tlb_inval); +#endif /* _XE_TLB_INVAL_ */
Signature · 1 line
2.34.1
Re: [PATCH v2 09/12] drm/xe: Add xe_tlb_inval_idle helper
Quoted 10 lines · Matthew Brost
Introduce the
xe_tlb_inval_idlehelper to detect whether any TLB invalidations are currently in flight. This is used in context-based TLB
invalidations to determine whether dummy TLB invalidations need to be sent to maintain proper TLB invalidation fence ordering..v2:
- Implementxe_tlb_inval_idlebased on pending listSigned-off-by Matthew Brost <[email protected]>
I don't have any issue with this patch or the next one, but I want to hold on full review until we have a working solution for the registration/deregistration placement of this.
Thanks,
Stuart
Quoted 66 lines
---
drivers/gpu/drm/xe/xe_tlb_inval.c | 21 +++++++++++++++++++++drivers/gpu/drm/xe/xe_tlb_inval.h| 2 ++2 files changed, 23 insertions(+)drivers/gpu/drm/xe/xe_tlb_inval.c +0
-0@@ -44,11 +44,14 @@ static void xe_tlb_inval_fence_fini(struct xe_tlb_inval_fence *fence)static void
xe_tlb_inval_fence_signal(structxe_tlb_inval_fence*fence) {
+ structxe_tlb_inval*tlb_inval=fence->tlb_inval; bool stack =test_bit(FENCE_STACK_BIT, &fence->base.flags);
lockdep_assert_held(&fence->tlb_inval->pending_lock);
list_del(&fence->link);
+ if (list_empty(&tlb_inval->pending_fences))
+ cancel_delayed_work(&tlb_inval->fence_tdr);trace_xe_tlb_inval_fence_signal(fence->tlb_inval->xe, fence);xe_tlb_inval_fence_fini(fence);dma_fence_signal(&fence->base);drivers/gpu/drm/xe/xe_tlb_inval.c +0
-0@@ -443,3 +446,21 @@ void xe_tlb_inval_fence_init(struct xe_tlb_inval *tlb_inval, dma_fence_get(&fence->base); fence->tlb_inval = tlb_inval;}
drivers/gpu/drm/xe/xe_tlb_inval.c +6
-0+ +/** +* xe_tlb_inval_idle() - Initialize TLB invalidation is idle +* @tlb_inval: TLB invalidation client +* +* Check the TLB invalidation seqno to determine if it is idle (i.e., no TLB
+ * invalidations are in flight). Expected to be called in the backend after the
+ * fence has been added to the pending list, and takes this into account.
+ *
+ * Return: True if TLB invalidation client is idle, False otherwise
+ */drivers/gpu/drm/xe/xe_tlb_inval.c +7
-0+bool xe_tlb_inval_idle (struct xe_tlb_inval *tlb_inval) +{ + lockdep_assert_held(&tlb_inval->seqno_lock); + + guard(spinlock_irq)(&tlb_inval->pending_lock); + return list_is_singular(&tlb_inval->pending_fences); +}drivers/gpu/drm/xe/xe_tlb_inval.h +0
-0@@ -43,4 +43,6 @@ xe_tlb_inval_fence_wait(struct xe_tlb_inval_fence *fence)
voidxe_tlb_inval_done_handler(struct xe_tlb_inval *tlb_inval, int seqno);drivers/gpu/drm/xe/xe_tlb_inval.h +2
-0+bool xe_tlb_inval_idle (struct xe_tlb_inval *tlb_inval); +#endif /* _XE_TLB_INVAL_ */
Summers, Stuart
Reviewed-byRe: [PATCH v2 09/12] drm/xe: Add xe_tlb_inval_idle helper
Quoted 9 lines · Stuart
Quoted 8 lines · Matthew Brost
Introduce the
xe_tlb_inval_idlehelper to detect whether any TLB invalidations are currently in flight. This is used in context- based
TLB
invalidations to determine whether dummy TLB invalidations need to be
sent to maintain proper TLB invalidation fence ordering..
Quoted 8 lines
Quoted 4 lines
v2:
- Implementxe_tlb_inval_idlebased on pending listSigned-off-by Matthew Brost <[email protected]>I don't have any issue with this patch or the next one, but I want to hold on full review until we have a working solution for the registration/deregistration placement of this.
Quoted 74 lines
Thanks,
StuartQuoted 69 lines
---
drivers/gpu/drm/xe/xe_tlb_inval.c | 21 +++++++++++++++++++++drivers/gpu/drm/xe/xe_tlb_inval.h| 2 ++2 files changed, 23 insertions(+)drivers/gpu/drm/xe/xe_tlb_inval.c +0
-0@@ -44,11 +44,14 @@ static void xe_tlb_inval_fence_fini(struct xe_tlb_inval_fence *fence)static void
xe_tlb_inval_fence_signal(structxe_tlb_inval_fence*fence) {
+ structxe_tlb_inval*tlb_inval=fence->tlb_inval; bool stack =test_bit(FENCE_STACK_BIT, &fence->base.flags);
lockdep_assert_held(&fence->tlb_inval->pending_lock);
list_del(&fence->link);
+ if (list_empty(&tlb_inval->pending_fences))
+ cancel_delayed_work(&tlb_inval->fence_tdr);trace_xe_tlb_inval_fence_signal(fence->tlb_inval->xe, fence);
xe_tlb_inval_fence_fini(fence);dma_fence_signal(&fence->base);drivers/gpu/drm/xe/xe_tlb_inval.c +0
-0@@ -443,3 +446,21 @@ void xe_tlb_inval_fence_init(struct
xe_tlb_inval*tlb_inval,
dma_fence_get(&fence->base);
fence->tlb_inval =tlb_inval;
}drivers/gpu/drm/xe/xe_tlb_inval.c +6
-0+ +/** +* xe_tlb_inval_idle() - Initialize TLB invalidation is idle +* @tlb_inval: TLB invalidation client +* +* Check the TLB invalidation seqno to determine if it is idle (i.e., no TLB
+ * invalidations are in flight). Expected to be called in the backend after the
+ * fence has been added to the pending list, and takes this into account.
+ *
+ * Return: True if TLB invalidation client is idle, False otherwise
+ */drivers/gpu/drm/xe/xe_tlb_inval.c +7
-0+bool xe_tlb_inval_idle (struct xe_tlb_inval *tlb_inval) +{ + lockdep_assert_held(&tlb_inval->seqno_lock); + + guard(spinlock_irq)(&tlb_inval->pending_lock); + return list_is_singular(&tlb_inval->pending_fences); +}drivers/gpu/drm/xe/xe_tlb_inval.h +0
-0@@ -43,4 +43,6 @@ xe_tlb_inval_fence_wait(struct xe_tlb_inval_fence *fence)
voidxe_tlb_inval_done_handler(struct xe_tlb_inval *tlb_inval, int seqno);drivers/gpu/drm/xe/xe_tlb_inval.h +2
-0+bool xe_tlb_inval_idle (struct xe_tlb_inval *tlb_inval); +#endif /* _XE_TLB_INVAL_ */
Matthew Brost
maintainerPATCH v2 10/12 drm/xe: Add exec queue active vfunc
In reply to Matthew BrostIf an exec queue is inactive (e.g., not registered or scheduling is disabled), TLB invalidations are not issued for that queue. Add a virtual function to determine the active state, which TLB invalidation logic can hook into.
--- drivers/gpu/drm/xe/xe_exec_queue_types.h | 2 ++ drivers/gpu/drm/xe/xe_execlist.c | 7 +++++++ drivers/gpu/drm/xe/xe_guc_submit.c | 6 ++++++ 3 files changed, 15 insertions(+)
drivers/gpu/drm/xe/xe_exec_queue_types.h +2 -0
@@ -228,6 +228,8 @@ struct xe_exec_queue_ops {void (*resume)(struct xe_exec_queue *q);/** @reset_status: check exec queue reset status */ bool (*reset_status)(struct xe_exec_queue *q); +/** @active: check exec queue is active */ + bool (*active)(struct xe_exec_queue *q); };#endif
drivers/gpu/drm/xe/xe_execlist.c +7 -0
@@ -469,6 +469,12 @@ static bool execlist_exec_queue_reset_status(struct xe_exec_queue *q)return false; } +static bool execlist_exec_queue_active (struct xe_exec_queue *q) +{ +/* NIY */ +return false; +} +static const struct xe_exec_queue_ops execlist_exec_queue_ops = { .init = execlist_exec_queue_init, .kill = execlist_exec_queue_kill,
@@ -481,6 +487,7 @@ static const struct xe_exec_queue_ops execlist_exec_queue_ops = { .suspend_wait = execlist_exec_queue_suspend_wait, .resume = execlist_exec_queue_resume, .reset_status = execlist_exec_queue_reset_status, + .active = execlist_exec_queue_active, };int xe_execlist_init (struct xe_gt *gt)
drivers/gpu/drm/xe/xe_guc_submit.c +6 -0
@@ -1919,6 +1919,11 @@ static bool guc_exec_queue_reset_status(struct xe_exec_queue *q)return exec_queue_reset (q) ||exec_queue_killed_or_banned_or_wedged (q); } +static bool guc_exec_queue_active (struct xe_exec_queue *q) +{ +return exec_queue_enabled (q) && !exec_queue_pending_disable (q); +} +/* * All of these functions are an abstraction layer which other parts of Xe can * use to trap into the GuC backend. All of these functions, aside from init,
@@ -1937,6 +1942,7 @@ static const struct xe_exec_queue_ops guc_exec_queue_ops = { .suspend_wait = guc_exec_queue_suspend_wait, .resume = guc_exec_queue_resume, .reset_status = guc_exec_queue_reset_status, + .active = guc_exec_queue_active, };static void guc_exec_queue_stop (struct xe_guc *guc,struct xe_exec_queue *q)
Signature · 1 line
2.34.1
Matthew Brost
maintainerPATCH v2 11/12 drm/xe: Add context-based invalidation to GuC TLB invalidation backend
In reply to Matthew BrostIntroduce context-based invalidation support to the GuC TLB invalidation backend. This is implemented by iterating over each exec queue per GT within a VM, skipping inactive queues, and issuing a context-based (GuC ID) H2G TLB invalidation. All H2G messages, except the final one, are sent with an invalid seqno, which the G2H handler drops to ensure the TLB invalidation fence is only signaled once all H2G messages are completed.
A watermark mechanism is also added to switch between context-based TLB invalidations and full device-wide invalidations, as the return on investment for context-based invalidation diminishes when many exec queues are mapped.
v2:
- Fix checkpatch warnings
--- drivers/gpu/drm/xe/xe_device_types.h | 2 + drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 122 +++++++++++++++++++++++++- drivers/gpu/drm/xe/xe_pci.c | 1 + drivers/gpu/drm/xe/xe_pci_types.h | 1 + 4 files changed, 124 insertions(+), 2 deletions(-)
drivers/gpu/drm/xe/xe_device_types.h +2 -0
@@ -316,6 +316,8 @@ struct xe_device { u8 has_mem_copy_instr:1;/** @info.has_pxp: Device has PXP support */ u8 has_pxp:1; +/** @info.has_ctx_tlb_inval: Has context based TLB invalidations */ + u8 has_ctx_tlb_inval:1;/** @info.has_range_tlb_inval: Has range based TLB invalidations */ u8 has_range_tlb_inval:1;/** @info.has_sriov: Supports SR-IOV */
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +120 -2
171 lines not on this page · raw message
drivers/gpu/drm/xe/xe_pci.c +1 -0
@@ -863,6 +863,7 @@ static int xe_info_init(struct xe_device *xe, xe->info.has_device_atomics_on_smem =1 ; xe->info.has_range_tlb_inval = graphics_desc->has_range_tlb_inval; + xe->info.has_ctx_tlb_inval = graphics_desc->has_ctx_tlb_inval; xe->info.has_usm = graphics_desc->has_usm; xe->info.has_64bit_timestamp = graphics_desc->has_64bit_timestamp;
drivers/gpu/drm/xe/xe_pci_types.h +1 -0
@@ -63,6 +63,7 @@ struct xe_graphics_desc { u8 has_atomic_enable_pte_bit:1; u8 has_indirect_ring_state:1; u8 has_range_tlb_inval:1; + u8 has_ctx_tlb_inval:1; u8 has_usm:1; u8 has_64bit_timestamp:1; };
Signature · 1 line
2.34.1
Re: [PATCH v2 11/12] drm/xe: Add context-based invalidation to GuC TLB invalidation backend
Quoted 119 lines · Matthew Brost
Introduce context-based invalidation support to the GuC TLB invalidation
backend. This is implemented by iterating over each exec queue per GT within a VM, skipping inactive queues, and issuing a context-based (GuC
ID) H2G TLB invalidation. All H2G messages, except the final one, are sent with an invalid seqno, which the G2H handler drops to ensure the TLB invalidation fence is only signaled once all H2G messages are completed.A watermark mechanism is also added to switch between context-based TLB
invalidations and full device-wide invalidations, as the return on investment for context-based invalidation diminishes when many exec queues are mapped.v2:
- Fix checkpatch warningsSigned-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device_types.h | 2 +
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 122
+++++++++++++++++++++++++-
drivers/gpu/drm/xe/xe_pci.c | 1 +
drivers/gpu/drm/xe/xe_pci_types.h | 1 +4 files changed, 124 insertions(+), 2 deletions(-)drivers/gpu/drm/xe/xe_device_types.h +1
-0@@ -316,6 +316,8 @@ struct xe_device { u8 has_mem_copy_instr:1; /** @info.has_pxp: Device has PXP support */ u8 has_pxp:1; + /** @info.has_ctx_tlb_inval: Has context based TLBinvalidations */
+ u8 has_ctx_tlb_inval:1;
/** @info.has_range_tlb_inval: Has range based TLB invalidations */
u8 has_range_tlb_inval:1;
/** @info.has_sriov: Supports SR-IOV */drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -6,14 +6,17 @@ #include "abi/guc_actions_abi.h"
#include "xe_device.h"
+#include "xe_exec_queue_types.h"
#include "xe_gt_stats.h"
#include "xe_gt_types.h"
#include "xe_guc.h"
#include "xe_guc_ct.h"
+#include "xe_guc_exec_queue_types.h"
#include "xe_guc_tlb_inval.h"
#include "xe_force_wake.h"
#include "xe_mmio.h"
#include "xe_tlb_inval.h"
+#include "xe_vm.h"
#include "regs/xe_guc_regs.h"drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-0@@ -136,10 +139,16 @@ static int send_tlb_inval_ppgtt(struct xe_guc *guc, u32 seqno, u64 start, { #define MAX_TLB_INVALIDATION_LEN 7 struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = guc_to_xe(guc); u32 action[MAX_TLB_INVALIDATION_LEN]; u64 length = end - start; int len = 0;
+ xe_gt_assert(gt, (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE&& + !xe->info.has_ctx_tlb_inval) || + (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX&& +xe->info.has_ctx_tlb_inval));
+
action[len++] =XE_GUC_ACTION_TLB_INVALIDATION; action[len++] = seqno;
if (!gt_to_xe(gt)->info.has_range_tlb_inval ||drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -176,6 +185,100 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, XE_GUC_TLB_INVAL_PAGE_SELECTIVE); }
+static boolqueue_mapped_in_guc(struct xe_guc *guc, struct xe_exec_queue *q)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +5
-0+{ + return q->gt == guc_to_gt(guc); +} + +static int send_tlb_inval_ctx_ppgtt(struct xe_tlb_inval *tlb_inval,u32 seqno,
+ u64 start, u64 end, u32 asid)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +20
-0+{ + struct xe_guc *guc = tlb_inval->private; + struct xe_device *xe = guc_to_xe(guc); + struct xe_exec_queue *q, *last_q = NULL; + struct xe_vm *vm; + int err = 0; + + lockdep_assert_held(&tlb_inval->seqno_lock); + + if (xe->info.force_execlist) + return -ECANCELED; + + vm = xe_device_asid_to_vm(xe, asid); + if (IS_ERR(vm)) + return PTR_ERR(vm); + + down_read(&vm->exec_queues.lock); + + /* +* XXX: Randomly picking a threshold for now. This will need
Quoted 5 lines
to be
+ * tuned based on expected UMD queue counts and performance profiling.
+ */
+#define EXEC_QUEUE_COUNT_FULL_THRESHOLD 8
Does seem interesting for this to be configurable. Maybe different applications have different requirements here... But also we should have some kind of default. No issue with what you
Quoted 12 lines
+ if (vm->exec_queues.count[guc_to_gt(guc)->info.id] >= +
EXEC_QUEUE_COUNT_FULL_THRESHOLD) {
+ u32 action[] = {
+ XE_GUC_ACTION_TLB_INVALIDATION, + seqno,
+ MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL), + };
+
+ err =send_tlb_inval(guc, action, ARRAY_SIZE(action));
+ gotoerr_unlock;
+ }
drivers/gpu/drm/xe/xe_guc_tlb_inval.c static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, · line 232
Quoted 4 lines
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +4
-0+#undef EXEC_QUEUE_COUNT_FULL_THRESHOLD + + list_for_each_entry_reverse(q, &vm->exec_queues.list, + vm_exec_queue_link)
Braces here around the if() {}. Otherwise it's too easy to accidentally add lines below the if condition with unintended behavior.
Quoted 31 lines
+ if (
queue_mapped_in_guc(guc, q)&&q->ops->active(q)) {
+ last_q = q;
+ break;
+ }
+
+ if (!last_q) {
+ /*
+ * We can't break fence ordering for TLB invalidation jobs, if
+ * TLB invalidations are inflight issue a dummy invalidation to
+ * maintain ordering. Nor can we move safely theseqno_recvwhen
+ * returning-ECANCELEDif TLB invalidations are in flight. Use
+ * GGTT invalidation as dummy invalidation given ASID + * invalidations are unsupported here.
+ */
+ if (xe_tlb_inval_idle(tlb_inval))
+ err =-ECANCELED;
+ else
+ err =send_tlb_inval_ggtt(tlb_inval, seqno); + gotoerr_unlock;
+ }
+
+ list_for_each_entry(q,&vm->exec_queues.list,vm_exec_queue_link) {
+ int__seqno=last_q== q ? seqno :
+ TLB_INVALIDATION_SEQNO_INVALID;
+ u32 type =XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX;
Quoted 8 lines
+
+ /*
+ * XXX: Techincally we can race here and queue can become
+ * inactive, not ideal. The TLB invalidation will timeout in
+ * this case unless we get GuC support to convert to NOP...
We actually can't support this. It won't just time out, GuC will return an error (invalid parameter that the context isn't registered and we're trying to submit something via context invalidation - nothing actually gets sent to hardware) and we will issue a GT reset. We should make sure when we send this, the context is registered and any deregistration explicitly happens afterwards.
This is why generally we want to do this add/remove not at the queue level but at the guc registration/deregistration level to guarantee this. But I see you split the queue add and the submission based on this active flag. Right now I believe we're tracking "pending deregistration" with exec_queue_destroyed && exec_queue_registered, although it might be nice to have an explicit state there.
Alternatively we could also capture that error response from GuC, although there are a few different places we can get an error there and we might not want to overload the CT error handler in the event there is a legitimate state or hardware error that requires a guc reload/GT reset.
Thanks,
Stuart
Quoted 96 lines
+ */
+ if (!queue_mapped_in_guc(guc, q) || !q->ops-Quoted 1 line
active(q))
+ continue;
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +16
-0+ + xe_assert(xe, q->vm == vm); + + err = send_tlb_inval_ppgtt(guc, __seqno, start, end, + q->guc->id, type); + if (err) + goto err_unlock; + } + +err_unlock: + up_read(&vm->exec_queues.lock); + xe_vm_put(vm); + + return err; +} +static bool
tlb_inval_initialized(struct xe_tlb_inval *tlb_inval){
structxe_guc*guc =tlb_inval->private;drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -203,7 +306,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval) return hw_tlb_timeout + 2 * delay; }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-1-static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { +static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {.all =
send_tlb_inval_all,
.ggtt =send_tlb_inval_ggtt,
.ppgtt =send_tlb_inval_asid_ppgtt,drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -212,6 +315,15 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .timeout_delay = tlb_inval_timeout_delay, };drivers/gpu/drm/xe/xe_guc_tlb_inval.c +9
-0+static const struct xe_tlb_inval_ops guc_tlb_inval_ctx_ops = { + .ggtt = send_tlb_inval_ggtt, + .all = send_tlb_inval_all, + .ppgtt = send_tlb_inval_ctx_ppgtt, + .initialized = tlb_inval_initialized, + .flush = tlb_inval_flush, + .timeout_delay = tlb_inval_timeout_delay, +}; +/**
*xe_guc_tlb_inval_init_early()- Init GuC TLB invalidation early
* @guc: GuC objectdrivers/gpu/drm/xe/xe_guc_tlb_inval.c +7
-1@@ -223,8 +335,14 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { void xe_guc_tlb_inval_init_early(struct xe_guc *guc, struct xe_tlb_inval *tlb_inval) { + struct xe_device *xe = guc_to_xe(guc); + tlb_inval->private = guc; - tlb_inval->ops = &guc_tlb_inval_ops; + + if (xe->info.has_ctx_tlb_inval) + tlb_inval->ops = &guc_tlb_inval_ctx_ops; + else + tlb_inval->ops = &guc_tlb_inval_asid_ops; }
/**drivers/gpu/drm/xe/xe_pci.c +0
-0@@ -863,6 +863,7 @@ static int xe_info_init(struct xe_device *xe, xe->info.has_device_atomics_on_smem = 1;
xe->info.has_range_tlb_inval = graphics_desc-Quoted 1 line
has_range_tlb_inval;+ xe->info.has_ctx_tlb_inval = graphics_desc-
Quoted 1 line
has_ctx_tlb_inval;xe->info.has_usm =
graphics_desc->has_usm;xe->info.has_64bit_timestamp= graphics_desc-Quoted 1 line
has_64bit_timestamp;drivers/gpu/drm/xe/xe_pci_types.h +1
-0@@ -63,6 +63,7 @@ struct xe_graphics_desc { u8 has_atomic_enable_pte_bit:1; u8 has_indirect_ring_state:1; u8 has_range_tlb_inval:1; + u8 has_ctx_tlb_inval:1; u8 has_usm:1; u8 has_64bit_timestamp:1; };
Matthew Brost
maintainerRe: [PATCH v2 11/12] drm/xe: Add context-based invalidation to GuC TLB invalidation backend
Quoted 121 lines · Stuart
Quoted 120 lines · Matthew Brost
Introduce context-based invalidation support to the GuC TLB invalidation
backend. This is implemented by iterating over each exec queue per GT within a VM, skipping inactive queues, and issuing a context-based (GuC
ID) H2G TLB invalidation. All H2G messages, except the final one, are sent with an invalid seqno, which the G2H handler drops to ensure the TLB invalidation fence is only signaled once all H2G messages are completed.A watermark mechanism is also added to switch between context-based TLB
invalidations and full device-wide invalidations, as the return on investment for context-based invalidation diminishes when many exec queues are mapped.v2:
- Fix checkpatch warningsSigned-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device_types.h | 2 +
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 122
+++++++++++++++++++++++++-
drivers/gpu/drm/xe/xe_pci.c | 1 +
drivers/gpu/drm/xe/xe_pci_types.h | 1 +4 files changed, 124 insertions(+), 2 deletions(-)drivers/gpu/drm/xe/xe_device_types.h +1
-0@@ -316,6 +316,8 @@ struct xe_device { u8 has_mem_copy_instr:1; /** @info.has_pxp: Device has PXP support */ u8 has_pxp:1; + /** @info.has_ctx_tlb_inval: Has context based TLBinvalidations */
+ u8 has_ctx_tlb_inval:1;
/** @info.has_range_tlb_inval: Has range based TLB invalidations */
u8 has_range_tlb_inval:1;
/** @info.has_sriov: Supports SR-IOV */drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -6,14 +6,17 @@ #include "abi/guc_actions_abi.h"
#include "xe_device.h"
+#include "xe_exec_queue_types.h"
#include "xe_gt_stats.h"
#include "xe_gt_types.h"
#include "xe_guc.h"
#include "xe_guc_ct.h"
+#include "xe_guc_exec_queue_types.h"
#include "xe_guc_tlb_inval.h"
#include "xe_force_wake.h"
#include "xe_mmio.h"
#include "xe_tlb_inval.h"
+#include "xe_vm.h"
#include "regs/xe_guc_regs.h"drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-0@@ -136,10 +139,16 @@ static int send_tlb_inval_ppgtt(struct xe_guc *guc, u32 seqno, u64 start, { #define MAX_TLB_INVALIDATION_LEN 7 struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = guc_to_xe(guc); u32 action[MAX_TLB_INVALIDATION_LEN]; u64 length = end - start; int len = 0;
+ xe_gt_assert(gt, (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE&& + !xe->info.has_ctx_tlb_inval) || + (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX&& +xe->info.has_ctx_tlb_inval));
+
action[len++] =XE_GUC_ACTION_TLB_INVALIDATION; action[len++] = seqno;
if (!gt_to_xe(gt)->info.has_range_tlb_inval ||drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -176,6 +185,100 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, XE_GUC_TLB_INVAL_PAGE_SELECTIVE); }
+static boolqueue_mapped_in_guc(struct xe_guc *guc, struct xe_exec_queue *q)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +5
-0+{ + return q->gt == guc_to_gt(guc); +} + +static int send_tlb_inval_ctx_ppgtt(struct xe_tlb_inval *tlb_inval,u32 seqno,
+ u64 start, u64 end, u32 asid)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +20
-0+{ + struct xe_guc *guc = tlb_inval->private; + struct xe_device *xe = guc_to_xe(guc); + struct xe_exec_queue *q, *last_q = NULL; + struct xe_vm *vm; + int err = 0; + + lockdep_assert_held(&tlb_inval->seqno_lock); + + if (xe->info.force_execlist) + return -ECANCELED; + + vm = xe_device_asid_to_vm(xe, asid); + if (IS_ERR(vm)) + return PTR_ERR(vm); + + down_read(&vm->exec_queues.lock); + + /* +* XXX: Randomly picking a threshold for now. This will need to be
Quoted 9 lines
Quoted 4 lines
+ * tuned based on expected UMD queue counts and performance profiling.
+ */
+#define EXEC_QUEUE_COUNT_FULL_THRESHOLD 8Does seem interesting for this to be configurable. Maybe different applications have different requirements here... But also we should have some kind of default. No issue with what you
Yes, I figure we can make this tunable somehow.
Quoted 53 lines
Quoted 12 lines
+ if (vm->exec_queues.count[guc_to_gt(guc)->info.id] >= +
EXEC_QUEUE_COUNT_FULL_THRESHOLD) {
+ u32 action[] = {
+ XE_GUC_ACTION_TLB_INVALIDATION, + seqno,
+ MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL), + };
+
+ err =send_tlb_inval(guc, action, ARRAY_SIZE(action));
+ gotoerr_unlock;
+ }drivers/gpu/drm/xe/xe_guc_tlb_inval.c static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, · line 232
Quoted 4 lines
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +4
-0+#undef EXEC_QUEUE_COUNT_FULL_THRESHOLD + + list_for_each_entry_reverse(q, &vm->exec_queues.list, + vm_exec_queue_link)Braces here around the
if(){}. Otherwise it's too easy to accidentally add lines below the if condition with unintended behavior.Quoted 33 lines
+ if (
queue_mapped_in_guc(guc, q)&&q->ops->active(q)) {
+ last_q = q;
+ break;
+ }
+
+ if (!last_q) {
+ /*
+ * We can't break fence ordering for TLB invalidation jobs, if
+ * TLB invalidations are inflight issue a dummy invalidation to
+ * maintain ordering. Nor can we move safely theseqno_recvwhen
+ * returning-ECANCELEDif TLB invalidations are in flight. Use
+ * GGTT invalidation as dummy invalidation given ASID + * invalidations are unsupported here.
+ */
+ if (xe_tlb_inval_idle(tlb_inval))
+ err =-ECANCELED;
+ else
+ err =send_tlb_inval_ggtt(tlb_inval, seqno); + gotoerr_unlock;
+ }
+
+ list_for_each_entry(q,&vm->exec_queues.list,vm_exec_queue_link) {
+ int__seqno=last_q== q ? seqno :
+ TLB_INVALIDATION_SEQNO_INVALID;
+ u32 type =XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX; +
+ /*
Quoted 8 lines
Quoted 6 lines
+ * XXX: Techincally we can race here and queue can become
+ * inactive, not ideal. The TLB invalidation will timeout in
+ * this case unless we get GuC support to convert to NOP...We actually can't support this. It won't just time out, GuC will return
Yes, I agree. I think we'd receive a failure response to GUC_HXG_TYPE_FAST_REQUEST, but if I recall correctly, in my testing I only saw timeouts on TLB invalidations sent to unregistered GuC IDs. Let me follow on failure case (I hit this in case when then the multi-GT code wasn't quite right yet).
Quoted 3 lines
an error (invalid parameter that the context isn't registered and we're trying to submit something via context invalidation - nothing actually gets sent to hardware) and we will issue a GT reset. We should make
A failure response to GUC_HXG_TYPE_FAST_REQUEST is indeed a GT reset.
Quoted 9 lines
sure when we send this, the context is registered and any deregistration explicitly happens afterwards.
This is why generally we want to do this add/remove not at the queue level but at the guc registration/deregistration level to guarantee this. But I see you split the queue add and the submission based on this active flag. Right now I believe we're tracking "pending deregistration" with
exec_queue_destroyed&&exec_queue_registered, although it might be nice to have an explicit state there.
Locking is required to make this race-free, which just doesn't work.
- TLB invalidations must be issued if the context is enabled or even if a disable is pending—otherwise, we risk memory corruption, which could be catastrophic.
- The context disable "done" G2H is received under the CT lock; the subsequent deregister is also issued under the CT lock.
- The lock that iterates over a VM's exec queues (contexts) is the outer lock of the CT lock.
As far as I can tell, we can't resolve this race without inverting the locking order—at least not in any way I'd consider a reasonable locking scheme for the KMD.
Quoted 7 lines
Alternatively we could also capture that error response from GuC, although there are a few different places we can get an error there and we might not want to overload the CT error handler in the event there is a legitimate state or hardware error that requires a guc reload/GT reset.
I believe the solution lies in GuC support. We could introduce a flag in the H2G TLB invalidation that signals: "I understand the risks—this may race. If it does, discard the TLB invalidation and still return a valid G2H TLB done response."
We’d likely disable this in CI builds unless it causes noise. In production, if we hit this race (which should be rare, given CI passes with the race exposed), the worst-case outcome is an extra TLB invalidation—which is acceptable.
Matt
Quoted 100 lines
Thanks,
StuartQuoted 96 lines
+ */
+ if (!queue_mapped_in_guc(guc, q) || !q->ops-Quoted 1 line
active(q))
+ continue;
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +16
-0+ + xe_assert(xe, q->vm == vm); + + err = send_tlb_inval_ppgtt(guc, __seqno, start, end, + q->guc->id, type); + if (err) + goto err_unlock; + } + +err_unlock: + up_read(&vm->exec_queues.lock); + xe_vm_put(vm); + + return err; +} +static bool
tlb_inval_initialized(struct xe_tlb_inval *tlb_inval){
structxe_guc*guc =tlb_inval->private;drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -203,7 +306,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval) return hw_tlb_timeout + 2 * delay; }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-1-static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { +static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {.all =
send_tlb_inval_all,
.ggtt =send_tlb_inval_ggtt,
.ppgtt =send_tlb_inval_asid_ppgtt,drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -212,6 +315,15 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .timeout_delay = tlb_inval_timeout_delay, };drivers/gpu/drm/xe/xe_guc_tlb_inval.c +9
-0+static const struct xe_tlb_inval_ops guc_tlb_inval_ctx_ops = { + .ggtt = send_tlb_inval_ggtt, + .all = send_tlb_inval_all, + .ppgtt = send_tlb_inval_ctx_ppgtt, + .initialized = tlb_inval_initialized, + .flush = tlb_inval_flush, + .timeout_delay = tlb_inval_timeout_delay, +}; +/**
*xe_guc_tlb_inval_init_early()- Init GuC TLB invalidation early
* @guc: GuC objectdrivers/gpu/drm/xe/xe_guc_tlb_inval.c +7
-1@@ -223,8 +335,14 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { void xe_guc_tlb_inval_init_early(struct xe_guc *guc, struct xe_tlb_inval *tlb_inval) { + struct xe_device *xe = guc_to_xe(guc); + tlb_inval->private = guc; - tlb_inval->ops = &guc_tlb_inval_ops; + + if (xe->info.has_ctx_tlb_inval) + tlb_inval->ops = &guc_tlb_inval_ctx_ops; + else + tlb_inval->ops = &guc_tlb_inval_asid_ops; }
/**drivers/gpu/drm/xe/xe_pci.c +0
-0@@ -863,6 +863,7 @@ static int xe_info_init(struct xe_device *xe, xe->info.has_device_atomics_on_smem = 1;
xe->info.has_range_tlb_inval = graphics_desc-Quoted 1 line
has_range_tlb_inval;+ xe->info.has_ctx_tlb_inval = graphics_desc-
Quoted 1 line
has_ctx_tlb_inval;xe->info.has_usm =
graphics_desc->has_usm;xe->info.has_64bit_timestamp= graphics_desc-Quoted 1 line
has_64bit_timestamp;drivers/gpu/drm/xe/xe_pci_types.h +1
-0@@ -63,6 +63,7 @@ struct xe_graphics_desc { u8 has_atomic_enable_pte_bit:1; u8 has_indirect_ring_state:1; u8 has_range_tlb_inval:1; + u8 has_ctx_tlb_inval:1; u8 has_usm:1; u8 has_64bit_timestamp:1; };
Re: [PATCH v2 11/12] drm/xe: Add context-based invalidation to GuC TLB invalidation backend
Quoted 221 lines · Matthew Brost
Quoted 137 lines · Stuart
Quoted 136 lines · Matthew Brost
Introduce context-based invalidation support to the GuC TLB invalidation
backend. This is implemented by iterating over each exec queue per GT
within a VM, skipping inactive queues, and issuing a context- based
(GuC
ID) H2G TLB invalidation. All H2G messages, except the final one, are
sent with an invalid seqno, which the G2H handler drops to ensure the
TLB invalidation fence is only signaled once all H2G messages are completed.A watermark mechanism is also added to switch between context- based
TLB
invalidations and full device-wide invalidations, as the return on
investment for context-based invalidation diminishes when many exec
queues are mapped.v2:
- Fix checkpatch warningsSigned-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device_types.h | 2 +
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 122
+++++++++++++++++++++++++-
drivers/gpu/drm/xe/xe_pci.c | 1 +
drivers/gpu/drm/xe/xe_pci_types.h | 1 +4 files changed, 124 insertions(+), 2 deletions(-)drivers/gpu/drm/xe/xe_device_types.h +1
-0@@ -316,6 +316,8 @@ struct xe_device { u8 has_mem_copy_instr:1; /** @info.has_pxp: Device has PXP support */ u8 has_pxp:1; + /** @info.has_ctx_tlb_inval: Has context basedTLB
invalidations */
+ u8 has_ctx_tlb_inval:1;
/** @info.has_range_tlb_inval: Has range based TLB
invalidations */
u8 has_range_tlb_inval:1;
/** @info.has_sriov: Supports SR-IOV */drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -6,14 +6,17 @@ #include "abi/guc_actions_abi.h"
#include "xe_device.h"
+#include "xe_exec_queue_types.h"
#include "xe_gt_stats.h"
#include "xe_gt_types.h"
#include "xe_guc.h"
#include "xe_guc_ct.h"
+#include "xe_guc_exec_queue_types.h"
#include "xe_guc_tlb_inval.h"
#include "xe_force_wake.h"
#include "xe_mmio.h"
#include "xe_tlb_inval.h"
+#include "xe_vm.h"
#include "regs/xe_guc_regs.h"drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -136,10 +139,16 @@ static int send_tlb_inval_ppgtt(struct
xe_guc
*guc, u32 seqno, u64 start,
{
#define MAX_TLB_INVALIDATION_LEN 7
structxe_gt*gt =guc_to_gt(guc);
+ structxe_device*xe =guc_to_xe(guc);
u32 action[MAX_TLB_INVALIDATION_LEN];
u64 length = end - start;
int len = 0;
+ xe_gt_assert(gt, (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE&&
+ !xe->info.has_ctx_tlb_inval) || + (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX&&
+xe->info.has_ctx_tlb_inval));
+
action[len++] =XE_GUC_ACTION_TLB_INVALIDATION; action[len++] = seqno;
if (!gt_to_xe(gt)->info.has_range_tlb_inval ||drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -176,6 +185,100 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno,
XE_GUC_TLB_INVAL_PAGE_SELECTIVE);
}
+static boolqueue_mapped_in_guc(struct xe_guc *guc, struct xe_exec_queue *q)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +5
-0+{ + return q->gt == guc_to_gt(guc); +} + +static int send_tlb_inval_ctx_ppgtt(struct xe_tlb_inval
*tlb_inval,
u32 seqno,
+ u64 start, u64 end, u32 asid)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +20
-0+{ + struct xe_guc *guc = tlb_inval->private; + struct xe_device *xe = guc_to_xe(guc); + struct xe_exec_queue *q, *last_q = NULL; + struct xe_vm *vm; + int err = 0; + + lockdep_assert_held(&tlb_inval->seqno_lock); + + if (xe->info.force_execlist) + return -ECANCELED; + + vm = xe_device_asid_to_vm(xe, asid); + if (IS_ERR(vm)) + return PTR_ERR(vm); + + down_read(&vm->exec_queues.lock); + + /* +* XXX: Randomly picking a threshold for now. This will need
to be
+ * tuned based on expected UMD queue counts andQuoted 9 lines
Quoted 4 lines
performance
profiling.
+ */
+#define EXEC_QUEUE_COUNT_FULL_THRESHOLD 8Does seem interesting for this to be configurable. Maybe different applications have different requirements here... But also we should have some kind of default. No issue with what you
Yes, I figure we can make this tunable somehow.
Quoted 62 lines
Quoted 12 lines
+ if (vm->exec_queues.count[guc_to_gt(guc)->info.id] >= +
EXEC_QUEUE_COUNT_FULL_THRESHOLD) {
+ u32 action[] = {
+ XE_GUC_ACTION_TLB_INVALIDATION, + seqno,
+ MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL), + };
+
+ err =send_tlb_inval(guc, action, ARRAY_SIZE(action));
+ gotoerr_unlock;
+ }drivers/gpu/drm/xe/xe_guc_tlb_inval.c
Quoted 4 lines
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +4
-0+#undef EXEC_QUEUE_COUNT_FULL_THRESHOLD + + list_for_each_entry_reverse(q, &vm->exec_queues.list, + vm_exec_queue_link)Braces here around the
if(){}. Otherwise it's too easy to accidentally
add lines below the if condition with unintended behavior.Quoted 41 lines
+ if (
queue_mapped_in_guc(guc, q)&& q->ops-Quoted 1 line
active(q))
{
+ last_q = q;
+ break;
+ }
+
+ if (!last_q) {
+ /*
+ * We can't break fence ordering for TLB invalidation
jobs, if
+ * TLB invalidations are inflight issue a dummy invalidation to
+ * maintain ordering. Nor can we move safely theseqno_recvwhen
+ * returning-ECANCELEDif TLB invalidations are in
flight. Use
+ * GGTT invalidation as dummy invalidation given ASID
+ * invalidations are unsupported here. + */
+ if (xe_tlb_inval_idle(tlb_inval))
+ err =-ECANCELED;
+ else
+ err =send_tlb_inval_ggtt(tlb_inval, seqno);
+ gotoerr_unlock;
+ }
+
+ list_for_each_entry(q,&vm->exec_queues.list,vm_exec_queue_link) {
+ int__seqno=last_q== q ? seqno :
+ TLB_INVALIDATION_SEQNO_INVALID; + u32 type =XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX; +
+ /*
+ * XXX: Techincally we can race here and queue can
becomeQuoted 8 lines
Quoted 5 lines
+ * inactive, not ideal. The TLB invalidation will timeout in
+ * this case unless we get GuC support to convert to
NOP...We actually can't support this. It won't just time out, GuC will return
Quoted 6 lines
Yes, I agree. I think we'd receive a failure response to
GUC_HXG_TYPE_FAST_REQUEST, but if I recall correctly, in my testing I only saw timeouts on TLB invalidations sent to unregistered GuC IDs. Let
me follow on failure case (I hit this in case when then the multi-GT code wasn't quite right yet).
I can look up the exact cases, but I had done a lot of testing around this internally and any sufficiently large exec queue registration/deregistration count is going to likely hit this. It has a pretty large impact on CI.
Quoted 20 lines
Quoted 6 lines
an error (invalid parameter that the context isn't registered and we're
trying to submit something via context invalidation - nothing actually
gets sent to hardware) and we will issue a GT reset. We should makeA failure response to
GUC_HXG_TYPE_FAST_REQUESTis indeed a GT reset.Quoted 10 lines
sure when we send this, the context is registered and any deregistration explicitly happens afterwards.
This is why generally we want to do this add/remove not at the queue
level but at the guc registration/deregistration level to guarantee this. But I see you split the queue add and the submission based on this active flag. Right now I believe we're tracking "pending deregistration" withexec_queue_destroyed&&exec_queue_registered, although it might be nice to have an explicit state there.
Quoted 7 lines
Locking is required to make this race-free, which just doesn't work.
- TLB invalidations must be issued if the context is enabled or even if a
disable is pending—otherwise, we risk memory corruption, which could be
catastrophic.
Agreed.
Quoted 11 lines
- The context disable "done" G2H is received under the CT lock; the subsequent deregister is also issued under the CT lock.
- The lock that iterates over a VM's exec queues (contexts) is the outer
lock of the CT lock.As far as I can tell, we can't resolve this race without inverting the
locking order—at least not in any way I'd consider a reasonable locking
scheme for the KMD.
Can't we just tie this to register_exec_queue() (for the add), disable_scheduling_deregister(), __deregister_exec_queue(), and guc_exec_queue_stop() (for the remove) and avoid the in-flight issue around locking? So basically we ensure that the list is only populated after the "register" CT is sent and before the "deregister" CT is sent, therefore ensuring that any TLB inval sent to those queues are sequenced properly around the register and deregister CT messages.
So basically (see the removal after we mark the queue as "destroyed"):
(quoted diff) +0 -0
b/drivers/gpu/drm/xe/xe_guc_submit.c
index ec0d1d1b0249..7a6ef3811808 100644
drivers/gpu/drm/xe/xe_guc_submit.c +0 -0
@@ -708,6 +708,8 @@ static void register_exec_queue(struct
xe_exec_queue *q, int ctx_type)
else
__register_exec_queue(guc, &info);
init_policies(guc, q);+
+ add_to_vm_exec_queue_list();
}
static u32 wq_space_until_wrap(struct xe_exec_queue *q)
@@ -950,6 +952,7 @@ static void disable_scheduling_deregister(struct xe_guc *guc,
clear_exec_queue_enabled(q);
set_exec_queue_pending_disable(q);
set_exec_queue_destroyed(q);
+ remove_from_vm_exec_queue_list();
trace_xe_exec_queue_scheduling_disable(q);
/*
@@ -1219,6 +1222,7 @@ static void __deregister_exec_queue(struct xe_guc
*guc, struct xe_exec_queue *q)
xe_gt_assert(guc_to_gt(guc), !exec_queue_pending_disable(q));
set_exec_queue_destroyed(q);
+ remove_from_vm_exec_queue_list();
trace_xe_exec_queue_deregister(q);
xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action),@@ -1971,6 +1975,7 @@ static void guc_exec_queue_stop(struct xe_guc *guc, struct xe_exec_queue *q)
EXEC_QUEUE_STATE_KILLED | EXEC_QUEUE_STATE_DESTROYED
|
EXEC_QUEUE_STATE_SUSPENDED,
&q->guc->state);
+ remove_from_vm_exec_queue_list();
q->guc->resume_time = 0;
trace_xe_exec_queue_stop(q);Quoted 12 lines
Quoted 11 lines
Alternatively we could also capture that error response from GuC, although there are a few different places we can get an error there and
we might not want to overload the CT error handler in the event there
is a legitimate state or hardware error that requires a guc reload/GT
reset.
Quoted 7 lines
I believe the solution lies in GuC support. We could introduce a flag in
the H2G TLB invalidation that signals: "I understand the risks—this may
race. If it does, discard the TLB invalidation and still return a valid
G2H TLB done response."
Of course, it would be nice if we had a dedicated error G2H that would call this out explicitly. Right now we just get a general "invalid parameter". Even having a way to easily identify the sequence number in question (which I know is slightly different after your changes in this series).
Quoted 3 lines
We’d likely disable this in CI builds unless it causes noise. In production, if we hit this race (which should be rare, given CI
Not rare, see my comment above. I think this needs to be fixed - we can't merge this with GT resets happening after the exec queue submission and destroy count gets high enough.
Thanks,
Stuart
Quoted 109 lines
passes
with the race exposed), the worst-case outcome is an extra TLB invalidation—which is acceptable.Matt
Quoted 103 lines
Thanks,
StuartQuoted 99 lines
+ */
+ if (!queue_mapped_in_guc(guc, q) || !q->ops-Quoted 1 line
active(q))
+ continue;
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +4
-0+ + xe_assert(xe, q->vm == vm); + + err = send_tlb_inval_ppgtt(guc, __seqno, start,end,
+q->guc->id, type); + if (err)
+ gotoerr_unlock;
+ }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +8
-0+ +err_unlock: + up_read(&vm->exec_queues.lock); + xe_vm_put(vm); + + return err; +} +static bool
tlb_inval_initialized(struct xe_tlb_inval *tlb_inval)
{
structxe_guc*guc =tlb_inval->private;drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -203,7 +306,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval) return hw_tlb_timeout + 2 * delay; }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-1-static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { +static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {.all =
send_tlb_inval_all,
.ggtt =send_tlb_inval_ggtt,
.ppgtt =send_tlb_inval_asid_ppgtt,drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -212,6 +315,15 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .timeout_delay = tlb_inval_timeout_delay, };drivers/gpu/drm/xe/xe_guc_tlb_inval.c +9
-0+static const struct xe_tlb_inval_ops guc_tlb_inval_ctx_ops = { + .ggtt = send_tlb_inval_ggtt, + .all = send_tlb_inval_all, + .ppgtt = send_tlb_inval_ctx_ppgtt, + .initialized = tlb_inval_initialized, + .flush = tlb_inval_flush, + .timeout_delay = tlb_inval_timeout_delay, +}; +/**
*xe_guc_tlb_inval_init_early()- Init GuC TLB invalidation early
* @guc: GuC objectdrivers/gpu/drm/xe/xe_guc_tlb_inval.c +7
-1@@ -223,8 +335,14 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { void xe_guc_tlb_inval_init_early(struct xe_guc *guc, struct xe_tlb_inval *tlb_inval) { + struct xe_device *xe = guc_to_xe(guc); + tlb_inval->private = guc; - tlb_inval->ops = &guc_tlb_inval_ops; + + if (xe->info.has_ctx_tlb_inval) + tlb_inval->ops = &guc_tlb_inval_ctx_ops; + else + tlb_inval->ops = &guc_tlb_inval_asid_ops; }
/**drivers/gpu/drm/xe/xe_pci.c +0
-0@@ -863,6 +863,7 @@ static int xe_info_init(struct xe_device *xe, xe->info.has_device_atomics_on_smem = 1;
xe->info.has_range_tlb_inval = graphics_desc-Quoted 1 line
has_range_tlb_inval;+ xe->info.has_ctx_tlb_inval = graphics_desc-
Quoted 1 line
has_ctx_tlb_inval;xe->info.has_usm =
graphics_desc->has_usm;xe->info.has_64bit_timestamp= graphics_desc-Quoted 1 line
has_64bit_timestamp;drivers/gpu/drm/xe/xe_pci_types.h +1
-0@@ -63,6 +63,7 @@ struct xe_graphics_desc { u8 has_atomic_enable_pte_bit:1; u8 has_indirect_ring_state:1; u8 has_range_tlb_inval:1; + u8 has_ctx_tlb_inval:1; u8 has_usm:1; u8 has_64bit_timestamp:1; };
Matthew Brost
maintainerRe: [PATCH v2 11/12] drm/xe: Add context-based invalidation to GuC TLB invalidation backend
Quoted 301 lines · Stuart
Quoted 221 lines · Matthew Brost
Quoted 137 lines · Stuart
Quoted 136 lines · Matthew Brost
Introduce context-based invalidation support to the GuC TLB invalidation
backend. This is implemented by iterating over each exec queue per GT
within a VM, skipping inactive queues, and issuing a context- based
(GuC
ID) H2G TLB invalidation. All H2G messages, except the final one, are
sent with an invalid seqno, which the G2H handler drops to ensure the
TLB invalidation fence is only signaled once all H2G messages are completed.A watermark mechanism is also added to switch between context- based
TLB
invalidations and full device-wide invalidations, as the return on
investment for context-based invalidation diminishes when many exec
queues are mapped.v2:
- Fix checkpatch warningsSigned-off-by Matthew Brost <[email protected]>---
drivers/gpu/drm/xe/xe_device_types.h | 2 +
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 122
+++++++++++++++++++++++++-
drivers/gpu/drm/xe/xe_pci.c | 1 +
drivers/gpu/drm/xe/xe_pci_types.h | 1 +4 files changed, 124 insertions(+), 2 deletions(-)drivers/gpu/drm/xe/xe_device_types.h +1
-0@@ -316,6 +316,8 @@ struct xe_device { u8 has_mem_copy_instr:1; /** @info.has_pxp: Device has PXP support */ u8 has_pxp:1; + /** @info.has_ctx_tlb_inval: Has context basedTLB
invalidations */
+ u8 has_ctx_tlb_inval:1;
/** @info.has_range_tlb_inval: Has range based TLB
invalidations */
u8 has_range_tlb_inval:1;
/** @info.has_sriov: Supports SR-IOV */drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -6,14 +6,17 @@ #include "abi/guc_actions_abi.h"
#include "xe_device.h"
+#include "xe_exec_queue_types.h"
#include "xe_gt_stats.h"
#include "xe_gt_types.h"
#include "xe_guc.h"
#include "xe_guc_ct.h"
+#include "xe_guc_exec_queue_types.h"
#include "xe_guc_tlb_inval.h"
#include "xe_force_wake.h"
#include "xe_mmio.h"
#include "xe_tlb_inval.h"
+#include "xe_vm.h"
#include "regs/xe_guc_regs.h"drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -136,10 +139,16 @@ static int send_tlb_inval_ppgtt(struct
xe_guc
*guc, u32 seqno, u64 start,
{
#define MAX_TLB_INVALIDATION_LEN 7
structxe_gt*gt =guc_to_gt(guc);
+ structxe_device*xe =guc_to_xe(guc);
u32 action[MAX_TLB_INVALIDATION_LEN];
u64 length = end - start;
int len = 0;
+ xe_gt_assert(gt, (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE&&
+ !xe->info.has_ctx_tlb_inval) || + (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX&&
+xe->info.has_ctx_tlb_inval));
+
action[len++] =XE_GUC_ACTION_TLB_INVALIDATION; action[len++] = seqno;
if (!gt_to_xe(gt)->info.has_range_tlb_inval ||drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -176,6 +185,100 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno,
XE_GUC_TLB_INVAL_PAGE_SELECTIVE);
}
+static boolqueue_mapped_in_guc(struct xe_guc *guc, struct xe_exec_queue *q)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +5
-0+{ + return q->gt == guc_to_gt(guc); +} + +static int send_tlb_inval_ctx_ppgtt(struct xe_tlb_inval
*tlb_inval,
u32 seqno,
+ u64 start, u64 end, u32 asid)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +20
-0+{ + struct xe_guc *guc = tlb_inval->private; + struct xe_device *xe = guc_to_xe(guc); + struct xe_exec_queue *q, *last_q = NULL; + struct xe_vm *vm; + int err = 0; + + lockdep_assert_held(&tlb_inval->seqno_lock); + + if (xe->info.force_execlist) + return -ECANCELED; + + vm = xe_device_asid_to_vm(xe, asid); + if (IS_ERR(vm)) + return PTR_ERR(vm); + + down_read(&vm->exec_queues.lock); + + /* +* XXX: Randomly picking a threshold for now. This will need
to be
+ * tuned based on expected UMD queue counts andQuoted 9 lines
Quoted 4 lines
performance
profiling.
+ */
+#define EXEC_QUEUE_COUNT_FULL_THRESHOLD 8Does seem interesting for this to be configurable. Maybe different applications have different requirements here... But also we should have some kind of default. No issue with what you
Yes, I figure we can make this tunable somehow.
Quoted 62 lines
Quoted 12 lines
+ if (vm->exec_queues.count[guc_to_gt(guc)->info.id] >= +
EXEC_QUEUE_COUNT_FULL_THRESHOLD) {
+ u32 action[] = {
+ XE_GUC_ACTION_TLB_INVALIDATION, + seqno,
+ MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL), + };
+
+ err =send_tlb_inval(guc, action, ARRAY_SIZE(action));
+ gotoerr_unlock;
+ }drivers/gpu/drm/xe/xe_guc_tlb_inval.c
Quoted 4 lines
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +4
-0+#undef EXEC_QUEUE_COUNT_FULL_THRESHOLD + + list_for_each_entry_reverse(q, &vm->exec_queues.list, + vm_exec_queue_link)Braces here around the
if(){}. Otherwise it's too easy to accidentally
add lines below the if condition with unintended behavior.Quoted 41 lines
+ if (
queue_mapped_in_guc(guc, q)&& q->ops-Quoted 1 line
active(q))
{
+ last_q = q;
+ break;
+ }
+
+ if (!last_q) {
+ /*
+ * We can't break fence ordering for TLB invalidation
jobs, if
+ * TLB invalidations are inflight issue a dummy invalidation to
+ * maintain ordering. Nor can we move safely theseqno_recvwhen
+ * returning-ECANCELEDif TLB invalidations are in
flight. Use
+ * GGTT invalidation as dummy invalidation given ASID
+ * invalidations are unsupported here. + */
+ if (xe_tlb_inval_idle(tlb_inval))
+ err =-ECANCELED;
+ else
+ err =send_tlb_inval_ggtt(tlb_inval, seqno);
+ gotoerr_unlock;
+ }
+
+ list_for_each_entry(q,&vm->exec_queues.list,vm_exec_queue_link) {
+ int__seqno=last_q== q ? seqno :
+ TLB_INVALIDATION_SEQNO_INVALID; + u32 type =XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX; +
+ /*
+ * XXX: Techincally we can race here and queue can
becomeQuoted 8 lines
Quoted 5 lines
+ * inactive, not ideal. The TLB invalidation will timeout in
+ * this case unless we get GuC support to convert to
NOP...We actually can't support this. It won't just time out, GuC will return
Quoted 6 lines
Yes, I agree. I think we'd receive a failure response to
GUC_HXG_TYPE_FAST_REQUEST, but if I recall correctly, in my testing I only saw timeouts on TLB invalidations sent to unregistered GuC IDs. Let
me follow on failure case (I hit this in case when then the multi-GT code wasn't quite right yet).I can look up the exact cases, but I had done a lot of testing around this internally and any sufficiently large exec queue registration/deregistration count is going to likely hit this. It has a pretty large impact on CI.
Quoted 20 lines
Quoted 6 lines
an error (invalid parameter that the context isn't registered and we're
trying to submit something via context invalidation - nothing actually
gets sent to hardware) and we will issue a GT reset. We should makeA failure response to
GUC_HXG_TYPE_FAST_REQUESTis indeed a GT reset.Quoted 10 lines
sure when we send this, the context is registered and any deregistration explicitly happens afterwards.
This is why generally we want to do this add/remove not at the queue
level but at the guc registration/deregistration level to guarantee this. But I see you split the queue add and the submission based on this active flag. Right now I believe we're tracking "pending deregistration" withexec_queue_destroyed&&exec_queue_registered, although it might be nice to have an explicit state there.Quoted 7 lines
Locking is required to make this race-free, which just doesn't work.
- TLB invalidations must be issued if the context is enabled or even if a
disable is pending—otherwise, we risk memory corruption, which could be
catastrophic.Agreed.
Quoted 11 lines
- The context disable "done" G2H is received under the CT lock; the subsequent deregister is also issued under the CT lock.
- The lock that iterates over a VM's exec queues (contexts) is the outer
lock of the CT lock.As far as I can tell, we can't resolve this race without inverting the
locking order—at least not in any way I'd consider a reasonable locking
scheme for the KMD.Can't we just tie this to
register_exec_queue()(for the add),disable_scheduling_deregister(),__deregister_exec_queue(), andguc_exec_queue_stop()(for the remove) and avoid the in-flight issue around locking? So basically we ensure that the list is only populated after the "register" CT is sent and before the "deregister" CT is sent, therefore ensuring that any TLB inval sent to those queues are sequenced properly around the register and deregister CT messages.So basically (see the removal after we mark the queue as "destroyed"):
drivers/gpu/drm/xe/xe_guc_submit.c +2
-0@@ -708,6 +708,8 @@ static void register_exec_queue(struct xe_exec_queue *q, int ctx_type) else __register_exec_queue(guc, &info);init_policies (guc, q); + +add_to_vm_exec_queue_list (); }static u32
wq_space_until_wrap(struct xe_exec_queue *q)
drivers/gpu/drm/xe/xe_guc_submit.c static void disable_scheduling_deregister(struct · line 956
Quoted 5 lines
drivers/gpu/drm/xe/xe_guc_submit.c +1
-0@@ -950,6 +952,7 @@ static void disable_scheduling_deregister(structxe_guc *guc,clear_exec_queue_enabled (q );set_exec_queue_pending_disable (q);set_exec_queue_destroyed (q); +remove_from_vm_exec_queue_list ();
We can't avoid skipping the TLB invalidation here because the context is possibly running. So we skipping a TLB invalidation, the context touchs a page that should have been invalidated, and boom memory corruption.
drivers/gpu/drm/xe/xe_guc_submit.c static void __deregister_exec_queue(struct xe_guc · line 1226
Quoted 9 lines
drivers/gpu/drm/xe/xe_guc_submit.c +0
-0trace_xe_exec_queue_scheduling_disable (q);/*drivers/gpu/drm/xe/xe_guc_submit.c +1
-0@@ -1219,6 +1222,7 @@ static void __deregister_exec_queue(struct xe_guc *guc, struct xe_exec_queue *q)xe_gt_assert (guc_to_gt (guc), !exec_queue_pending_disable (q));set_exec_queue_destroyed (q); +remove_from_vm_exec_queue_list ();
This is actually fine.
But the 2nd half of disable_scheduling_deregister (the safe point to skip TLB invalidation), is done in deregister_exec_queue which is the in G2H handler under the CT lock, which is where we'd invert locking. Again, changing the locking model, is basically a non-starter for me.
Quoted 40 lines
drivers/gpu/drm/xe/xe_guc_submit.c +0
-0trace_xe_exec_queue_deregister (q);xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action),drivers/gpu/drm/xe/xe_guc_submit.c +0
-0@@ -1971,6 +1975,7 @@ static void guc_exec_queue_stop(struct xe_guc *guc, struct xe_exec_queue *q)EXEC_QUEUE_STATE_KILLED |EXEC_QUEUE_STATE_DESTROYED |
EXEC_QUEUE_STATE_SUSPENDED, &q->guc->state); + remove_from_vm_exec_queue_list(); q->guc->resume_time = 0; trace_xe_exec_queue_stop(q);Quoted 12 lines
Quoted 11 lines
Alternatively we could also capture that error response from GuC, although there are a few different places we can get an error there and
we might not want to overload the CT error handler in the event there
is a legitimate state or hardware error that requires a guc reload/GT
reset.Quoted 7 lines
I believe the solution lies in GuC support. We could introduce a flag in
the H2G TLB invalidation that signals: "I understand the risks—this may
race. If it does, discard the TLB invalidation and still return a valid
G2H TLB done response."Of course, it would be nice if we had a dedicated error G2H that would call this out explicitly. Right now we just get a general "invalid parameter". Even having a way to easily identify the sequence number in question (which I know is slightly different after your changes in this series).
Quoted 1 line
Quoted 7 lines
Quoted 2 lines
We’d likely disable this in CI builds unless it causes noise. In production, if we hit this race (which should be rare, given CI
Not rare, see my comment above. I think this needs to be fixed - we can't merge this with GT resets happening after the exec queue submission and destroy count gets high enough.
Ok, then I think we really need to take this up with arch - either we do something like I suggest above or change the GuC interface to remain ASID based (no KMD changes) and the GuC does a ASID -> context hash table lookup and figures out which contexts that actually need an invalidation (it has this information, racwe free, also it likely can be smart and skip ones which don't need an invalidations which should reduce the overall cost of TLB invalidations too).
Matt
Quoted 113 lines
Thanks,
StuartQuoted 109 lines
passes
with the race exposed), the worst-case outcome is an extra TLB invalidation—which is acceptable.Matt
Quoted 103 lines
Thanks,
StuartQuoted 99 lines
+ */
+ if (!queue_mapped_in_guc(guc, q) || !q->ops-Quoted 1 line
active(q))
+ continue;
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +4
-0+ + xe_assert(xe, q->vm == vm); + + err = send_tlb_inval_ppgtt(guc, __seqno, start,end,
+q->guc->id, type); + if (err)
+ gotoerr_unlock;
+ }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +8
-0+ +err_unlock: + up_read(&vm->exec_queues.lock); + xe_vm_put(vm); + + return err; +} +static bool
tlb_inval_initialized(struct xe_tlb_inval *tlb_inval)
{
structxe_guc*guc =tlb_inval->private;drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -203,7 +306,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval) return hw_tlb_timeout + 2 * delay; }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-1-static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { +static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {.all =
send_tlb_inval_all,
.ggtt =send_tlb_inval_ggtt,
.ppgtt =send_tlb_inval_asid_ppgtt,drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -212,6 +315,15 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .timeout_delay = tlb_inval_timeout_delay, };drivers/gpu/drm/xe/xe_guc_tlb_inval.c +9
-0+static const struct xe_tlb_inval_ops guc_tlb_inval_ctx_ops = { + .ggtt = send_tlb_inval_ggtt, + .all = send_tlb_inval_all, + .ppgtt = send_tlb_inval_ctx_ppgtt, + .initialized = tlb_inval_initialized, + .flush = tlb_inval_flush, + .timeout_delay = tlb_inval_timeout_delay, +}; +/**
*xe_guc_tlb_inval_init_early()- Init GuC TLB invalidation early
* @guc: GuC objectdrivers/gpu/drm/xe/xe_guc_tlb_inval.c +7
-1@@ -223,8 +335,14 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { void xe_guc_tlb_inval_init_early(struct xe_guc *guc, struct xe_tlb_inval *tlb_inval) { + struct xe_device *xe = guc_to_xe(guc); + tlb_inval->private = guc; - tlb_inval->ops = &guc_tlb_inval_ops; + + if (xe->info.has_ctx_tlb_inval) + tlb_inval->ops = &guc_tlb_inval_ctx_ops; + else + tlb_inval->ops = &guc_tlb_inval_asid_ops; }
/**drivers/gpu/drm/xe/xe_pci.c +0
-0@@ -863,6 +863,7 @@ static int xe_info_init(struct xe_device *xe, xe->info.has_device_atomics_on_smem = 1;
xe->info.has_range_tlb_inval = graphics_desc-Quoted 1 line
has_range_tlb_inval;+ xe->info.has_ctx_tlb_inval = graphics_desc-
Quoted 1 line
has_ctx_tlb_inval;xe->info.has_usm =
graphics_desc->has_usm;xe->info.has_64bit_timestamp= graphics_desc-Quoted 1 line
has_64bit_timestamp;drivers/gpu/drm/xe/xe_pci_types.h +1
-0@@ -63,6 +63,7 @@ struct xe_graphics_desc { u8 has_atomic_enable_pte_bit:1; u8 has_indirect_ring_state:1; u8 has_range_tlb_inval:1; + u8 has_ctx_tlb_inval:1; u8 has_usm:1; u8 has_64bit_timestamp:1; };
Re: [PATCH v2 11/12] drm/xe: Add context-based invalidation to GuC TLB invalidation backend
In reply to Matthew BrostQuoted 16 lines · Matthew Brost
Introduce context-based invalidation support to the GuC TLB invalidation
backend. This is implemented by iterating over each exec queue per GT within a VM, skipping inactive queues, and issuing a context-based (GuC
ID) H2G TLB invalidation. All H2G messages, except the final one, are sent with an invalid seqno, which the G2H handler drops to ensure the TLB invalidation fence is only signaled once all H2G messages are completed.A watermark mechanism is also added to switch between context-based TLB
invalidations and full device-wide invalidations, as the return on investment for context-based invalidation diminishes when many exec queues are mapped.
Quoted 4 lines
v2:
- Fix checkpatch warningsSigned-off-by Matthew Brost <[email protected]>
Mentioned in an earlier patch, but after offline discussion, let's go with what you have here. We will address this as a follow up either with the GuC team to rework how they handle the registration/deregistration requirements or rework the KMD sequence to ensure we aren't exposed to the mentioned memory corruption along that path.
Quoted 255 lines
---
drivers/gpu/drm/xe/xe_device_types.h | 2 +
drivers/gpu/drm/xe/xe_guc_tlb_inval.c | 122
+++++++++++++++++++++++++-
drivers/gpu/drm/xe/xe_pci.c | 1 +
drivers/gpu/drm/xe/xe_pci_types.h | 1 +4 files changed, 124 insertions(+), 2 deletions(-)drivers/gpu/drm/xe/xe_device_types.h +1
-0@@ -316,6 +316,8 @@ struct xe_device { u8 has_mem_copy_instr:1; /** @info.has_pxp: Device has PXP support */ u8 has_pxp:1; + /** @info.has_ctx_tlb_inval: Has context based TLBinvalidations */
+ u8 has_ctx_tlb_inval:1;
/** @info.has_range_tlb_inval: Has range based TLB invalidations */
u8 has_range_tlb_inval:1;
/** @info.has_sriov: Supports SR-IOV */drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -6,14 +6,17 @@ #include "abi/guc_actions_abi.h"
#include "xe_device.h"
+#include "xe_exec_queue_types.h"
#include "xe_gt_stats.h"
#include "xe_gt_types.h"
#include "xe_guc.h"
#include "xe_guc_ct.h"
+#include "xe_guc_exec_queue_types.h"
#include "xe_guc_tlb_inval.h"
#include "xe_force_wake.h"
#include "xe_mmio.h"
#include "xe_tlb_inval.h"
+#include "xe_vm.h"
#include "regs/xe_guc_regs.h"drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-0@@ -136,10 +139,16 @@ static int send_tlb_inval_ppgtt(struct xe_guc *guc, u32 seqno, u64 start, { #define MAX_TLB_INVALIDATION_LEN 7 struct xe_gt *gt = guc_to_gt(guc); + struct xe_device *xe = guc_to_xe(guc); u32 action[MAX_TLB_INVALIDATION_LEN]; u64 length = end - start; int len = 0;
+ xe_gt_assert(gt, (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE&& + !xe->info.has_ctx_tlb_inval) || + (type ==XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX&& +xe->info.has_ctx_tlb_inval));
+
action[len++] =XE_GUC_ACTION_TLB_INVALIDATION; action[len++] = seqno;
if (!gt_to_xe(gt)->info.has_range_tlb_inval ||drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -176,6 +185,100 @@ static int send_tlb_inval_asid_ppgtt(struct xe_tlb_inval *tlb_inval, u32 seqno, XE_GUC_TLB_INVAL_PAGE_SELECTIVE); }
+static boolqueue_mapped_in_guc(struct xe_guc *guc, struct xe_exec_queue *q)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +5
-0+{ + return q->gt == guc_to_gt(guc); +} + +static int send_tlb_inval_ctx_ppgtt(struct xe_tlb_inval *tlb_inval,u32 seqno,
+ u64 start, u64 end, u32 asid)drivers/gpu/drm/xe/xe_guc_tlb_inval.c +20
-0+{ + struct xe_guc *guc = tlb_inval->private; + struct xe_device *xe = guc_to_xe(guc); + struct xe_exec_queue *q, *last_q = NULL; + struct xe_vm *vm; + int err = 0; + + lockdep_assert_held(&tlb_inval->seqno_lock); + + if (xe->info.force_execlist) + return -ECANCELED; + + vm = xe_device_asid_to_vm(xe, asid); + if (IS_ERR(vm)) + return PTR_ERR(vm); + + down_read(&vm->exec_queues.lock); + + /* +* XXX: Randomly picking a threshold for now. This will need to be
+ * tuned based on expected UMD queue counts and performance profiling.
+ */drivers/gpu/drm/xe/xe_guc_tlb_inval.c +10
-0+#define EXEC_QUEUE_COUNT_FULL_THRESHOLD 8 + if (vm->exec_queues.count[guc_to_gt(guc)->info.id] >= + EXEC_QUEUE_COUNT_FULL_THRESHOLD) { + u32 action[] = { + XE_GUC_ACTION_TLB_INVALIDATION, + seqno, + MAKE_INVAL_OP(XE_GUC_TLB_INVAL_FULL), + }; + + err = send_tlb_inval(guc, action,
ARRAY_SIZE(action));
+ gotoerr_unlock;
+ }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +5
-0+#undef EXEC_QUEUE_COUNT_FULL_THRESHOLD + + list_for_each_entry_reverse(q, &vm->exec_queues.list, + vm_exec_queue_link) + if (queue_mapped_in_guc(guc, q) && q->ops->active(q)){
+ last_q = q;
+ break;
+ }
+
+ if (!last_q) {
+ /*
+ * We can't break fence ordering for TLB invalidation jobs, if
+ * TLB invalidations are inflight issue a dummy invalidation to
+ * maintain ordering. Nor can we move safely theseqno_recvwhen
+ * returning-ECANCELEDif TLB invalidations are in flight. Use
+ * GGTT invalidation as dummy invalidation given ASID + * invalidations are unsupported here.
+ */
+ if (xe_tlb_inval_idle(tlb_inval))
+ err =-ECANCELED;
+ else
+ err =send_tlb_inval_ggtt(tlb_inval, seqno); + gotoerr_unlock;
+ }
+
+ list_for_each_entry(q,&vm->exec_queues.list,vm_exec_queue_link) {
+ int__seqno=last_q== q ? seqno :
+ TLB_INVALIDATION_SEQNO_INVALID;
+ u32 type =XE_GUC_TLB_INVAL_PAGE_SELECTIVE_CTX; +
+ /*
+ * XXX: Techincally we can race here and queue can become
+ * inactive, not ideal. The TLB invalidation will timeout in
+ * this case unless we get GuC support to convert to NOP...
+ */
+ if (!queue_mapped_in_guc(guc, q) || !q->ops-Quoted 1 line
active(q))
+ continue;
drivers/gpu/drm/xe/xe_guc_tlb_inval.c +16
-0+ + xe_assert(xe, q->vm == vm); + + err = send_tlb_inval_ppgtt(guc, __seqno, start, end, + q->guc->id, type); + if (err) + goto err_unlock; + } + +err_unlock: + up_read(&vm->exec_queues.lock); + xe_vm_put(vm); + + return err; +} +static bool
tlb_inval_initialized(struct xe_tlb_inval *tlb_inval){
structxe_guc*guc =tlb_inval->private;drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -203,7 +306,7 @@ static long tlb_inval_timeout_delay(struct xe_tlb_inval *tlb_inval) return hw_tlb_timeout + 2 * delay; }drivers/gpu/drm/xe/xe_guc_tlb_inval.c +1
-1-static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { +static const struct xe_tlb_inval_ops guc_tlb_inval_asid_ops = {.all =
send_tlb_inval_all,
.ggtt =send_tlb_inval_ggtt,
.ppgtt =send_tlb_inval_asid_ppgtt,drivers/gpu/drm/xe/xe_guc_tlb_inval.c +0
-0@@ -212,6 +315,15 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { .timeout_delay = tlb_inval_timeout_delay, };drivers/gpu/drm/xe/xe_guc_tlb_inval.c +9
-0+static const struct xe_tlb_inval_ops guc_tlb_inval_ctx_ops = { + .ggtt = send_tlb_inval_ggtt, + .all = send_tlb_inval_all, + .ppgtt = send_tlb_inval_ctx_ppgtt, + .initialized = tlb_inval_initialized, + .flush = tlb_inval_flush, + .timeout_delay = tlb_inval_timeout_delay, +}; +/**
*xe_guc_tlb_inval_init_early()- Init GuC TLB invalidation early
* @guc: GuC objectdrivers/gpu/drm/xe/xe_guc_tlb_inval.c +7
-1@@ -223,8 +335,14 @@ static const struct xe_tlb_inval_ops guc_tlb_inval_ops = { void xe_guc_tlb_inval_init_early(struct xe_guc *guc, struct xe_tlb_inval *tlb_inval) { + struct xe_device *xe = guc_to_xe(guc); + tlb_inval->private = guc; - tlb_inval->ops = &guc_tlb_inval_ops; + + if (xe->info.has_ctx_tlb_inval) + tlb_inval->ops = &guc_tlb_inval_ctx_ops; + else + tlb_inval->ops = &guc_tlb_inval_asid_ops; }
/**drivers/gpu/drm/xe/xe_pci.c +0
-0@@ -863,6 +863,7 @@ static int xe_info_init(struct xe_device *xe, xe->info.has_device_atomics_on_smem = 1;
xe->info.has_range_tlb_inval = graphics_desc-Quoted 1 line
has_range_tlb_inval;+ xe->info.has_ctx_tlb_inval = graphics_desc-
Quoted 1 line
has_ctx_tlb_inval;xe->info.has_usm =
graphics_desc->has_usm;xe->info.has_64bit_timestamp= graphics_desc-Quoted 1 line
has_64bit_timestamp;drivers/gpu/drm/xe/xe_pci_types.h +1
-0@@ -63,6 +63,7 @@ struct xe_graphics_desc { u8 has_atomic_enable_pte_bit:1; u8 has_indirect_ring_state:1; u8 has_range_tlb_inval:1; + u8 has_ctx_tlb_inval:1; u8 has_usm:1; u8 has_64bit_timestamp:1; };
Matthew Brost
maintainerPATCH v2 12/12 drm/xe: Enable context TLB invalidations for CI
In reply to Matthew BrostDo not review, CI only.
--- drivers/gpu/drm/xe/xe_pci.c | 1 + 1 file changed, 1 insertion(+)
drivers/gpu/drm/xe/xe_pci.c +1 -0
@@ -95,6 +95,7 @@ static const struct xe_graphics_desc graphics_xelpg = { .has_asid = 1, \ .has_atomic_enable_pte_bit = 1, \ .has_range_tlb_inval = 1, \ + .has_ctx_tlb_inval = 1, \ .has_usm = 1, \ .has_64bit_timestamp = 1, \ .hw_engine_mask = \
Signature · 1 line
2.34.1