Re: [PATCH] drm/amdkfd: Use partial mapping in GPU page fault recovery
On 10/19/2023 2:40 PM, Philip Yang wrote:
On 2023-10-19 12:20, Chen, Xiaogang wrote:
On 10/19/2023 11:08 AM, Philip Yang wrote:
On 2023-10-19 10:24, Xiaogang.Chen wrote:
From: Xiaogang Chen
After partial migration to recover GPU page fault this patch does
GPU vm
space mapping for same page range that got migrated instead of
mapping all
pages of svm range in which the page fault happened.
Signed-off-by: Xiaogang Chen
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 33
+---
1 file changed, 25 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 54af7a2b29f8..81dbcc8a4ccc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1619,6 +1619,7 @@ static void *kfd_svm_page_owner(struct
kfd_process *p, int32_t gpuidx)
* 5. Release page table (and SVM BO) reservation
*/
static int svm_range_validate_and_map(struct mm_struct *mm,
+ unsigned long map_start, unsigned long
map_last,
struct svm_range *prange, int32_t gpuidx,
bool intr, bool wait, bool flush_tlb)
{
@@ -1630,6 +1631,12 @@ static int svm_range_validate_and_map(struct
mm_struct *mm,
int32_t idx;
int r = 0;
+ if (map_start < prange->start || map_last > prange->last) {
This is not needed, as this case should never happen, and you also
use max/min to limit map_start, map_last below.
This was just a sanity check, I can remove it.
+ pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n",
+ map_start, map_last, prange->start, prange->last);
+ return -EFAULT;
+ }
+
ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
if (!ctx)
return -ENOMEM;
@@ -1747,9 +1754,16 @@ static int svm_range_validate_and_map(struct
mm_struct *mm,
r = -EAGAIN;
}
- if (!r)
- r = svm_range_map_to_gpus(prange, offset, npages,
readonly,
- ctx->bitmap, wait, flush_tlb);
+ if (!r) {
+ map_start = max(map_start, prange->start + offset);
+ map_last = min(map_last, prange->start + offset +
npages);
This should move forward to outside the for loop, otherwise
amdgpu_hmm_range_get_pages and svm_range_dma_map still work on the
entire prange, and then prange->vram_pages update logic should be
changed accordingly.
We need use hmm function to get all vram page number on whole range
as we did not know how many vram pages after partial migration, then
we know if can release vram bo.
ok,migrate to vram and migrate to ram have the vram pages updated
already, the is needed for the splite prange. We could update
prange->vram_pages when splitting prange, this can be done in another
patch.
map_last is inclusive,
+ map_last = min(map_last, prange->start + offset + npages
- 1);
ok, will update it.
Regards
Xiaogang
Regards,
Philip
Regards
Xiaogang
Regards,
Philip
+ if (map_start <= map_last) {
+ offset = map_start - prange->start;
+ npages = map_last - map_start + 1;
+ r = svm_range_map_to_gpus(prange, offset, npages,
readonly,
+ ctx->bitmap, wait, flush_tlb);
+ }
+ }
if (!r && next == end)
prange->mapped_to_gpu = true;
@@ -1855,8 +1869,8 @@ static void svm_range_restore_work(struct
work_struct *work)
*/
mutex_lock(&prange->migrate_mutex);
- r = svm_range_validate_and_map(mm, prange,
MAX_GPU_INSTANCE,
- false, true, false);
+ r = svm_range_validate_and_map(mm, prange->start,
prange->last, prange,
+ MAX_GPU_INSTANCE, false, true, false);
if (r)
pr_debug("failed %d to map 0x%lx to gpus\n", r,
prange->start);
@@ -3069,6 +3083,8 @@ svm_range_restore_pages(struct amdgpu_device
*adev, unsigned int pasid,
kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
write_fault, timestamp);
+ start = prange->start;
+ last = prange->last;
if (prange->actual_loc != 0 || best_loc != 0) {
migration = true;
/* Align migration range start and size to granularity
size */
@@ -3102,10 +3118,11 @@ svm_range_restore_pages(struct
amdgpu_device *adev, unsigned int pasid,
}
}
- r = svm_range_validate_and_map(mm, prange, gpuidx, false,
false, false);
+ r = svm_range_validate_and_map(mm, start, last, prange,
gpuidx, false,
+ false, false);
if (r)
pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to
gpus\n",
- r, svms, prange->start, prange->last);
+ r, svms, start, last);
kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
Re: [PATCH] drm/amdkfd: Use partial mapping in GPU page fault recovery
On 2023-10-19 12:20, Chen, Xiaogang
wrote:
On 10/19/2023 11:08 AM, Philip Yang wrote:
On 2023-10-19 10:24, Xiaogang.Chen wrote:
From: Xiaogang
Chen
After partial migration to recover GPU page fault this patch
does GPU vm
space mapping for same page range that got migrated instead of
mapping all
pages of svm range in which the page fault happened.
Signed-off-by: Xiaogang Chen
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 33
+---
1 file changed, 25 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 54af7a2b29f8..81dbcc8a4ccc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1619,6 +1619,7 @@ static void *kfd_svm_page_owner(struct
kfd_process *p, int32_t gpuidx)
* 5. Release page table (and SVM BO) reservation
*/
static int svm_range_validate_and_map(struct mm_struct *mm,
+ unsigned long map_start, unsigned long
map_last,
struct svm_range *prange, int32_t
gpuidx,
bool intr, bool wait, bool flush_tlb)
{
@@ -1630,6 +1631,12 @@ static int
svm_range_validate_and_map(struct mm_struct *mm,
int32_t idx;
int r = 0;
+ if (map_start < prange->start || map_last >
prange->last) {
This is not needed, as this case should never happen, and you
also use max/min to limit map_start, map_last below.
This was just a sanity check, I can remove it.
+ pr_debug("range [0x%lx 0x%lx]
out prange [0x%lx 0x%lx]\n",
+ map_start, map_last, prange->start,
prange->last);
+ return -EFAULT;
+ }
+
ctx = kzalloc(sizeof(struct svm_validate_context),
GFP_KERNEL);
if (!ctx)
return -ENOMEM;
@@ -1747,9 +1754,16 @@ static int
svm_range_validate_and_map(struct mm_struct *mm,
r = -EAGAIN;
}
- if (!r)
- r = svm_range_map_to_gpus(prange, offset, npages,
readonly,
- ctx->bitmap, wait, flush_tlb);
+ if (!r) {
+ map_start = max(map_start, prange->start +
offset);
+ map_last = min(map_last, prange->start +
offset + npages);
This should move forward to outside the for loop, otherwise
amdgpu_hmm_range_get_pages and svm_range_dma_map still work on
the entire prange, and then prange->vram_pages update logic
should be changed accordingly.
We need use hmm function to get all vram page number on whole
range as we did not know how many vram pages after partial
migration, then we know if can release vram bo.
ok,migrate to vram and migrate to ram have the vram pages updated
already, the is needed for the splite prange. We could update
prange->vram_pages when splitting prange, this can be done in
another patch.
map_last is inclusive,
+ map_last = min(map_last, prange->start + offset +
npages - 1);
Regards,
Philip
Regards
Xiaogang
Regards,
Philip
+ if (map_start <=
map_last) {
+ offset = map_start - prange->start;
+ npages = map_last - map_start + 1;
+ r = svm_range_map_to_gpus(prange, offset,
npages, readonly,
+ ctx->bitmap, wait,
flush_tlb);
+ }
+ }
if (!r && next == end)
prange->mapped_
Re: [PATCH] drm/amdkfd: Use partial mapping in GPU page fault recovery
On 10/19/2023 11:08 AM, Philip Yang wrote:
On 2023-10-19 10:24, Xiaogang.Chen wrote:
From: Xiaogang Chen
After partial migration to recover GPU page fault this patch does GPU vm
space mapping for same page range that got migrated instead of mapping all
pages of svm range in which the page fault happened.
Signed-off-by: Xiaogang Chen
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 33 +---
1 file changed, 25 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 54af7a2b29f8..81dbcc8a4ccc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1619,6 +1619,7 @@ static void *kfd_svm_page_owner(struct kfd_process *p,
int32_t gpuidx)
* 5. Release page table (and SVM BO) reservation
*/
static int svm_range_validate_and_map(struct mm_struct *mm,
+ unsigned long map_start, unsigned long
map_last,
struct svm_range *prange, int32_t gpuidx,
bool intr, bool wait, bool flush_tlb)
{
@@ -1630,6 +1631,12 @@ static int svm_range_validate_and_map(struct mm_struct
*mm,
int32_t idx;
int r = 0;
+ if (map_start < prange->start || map_last > prange->last) {
This is not needed, as this case should never happen, and you also use
max/min to limit map_start, map_last below.
This was just a sanity check, I can remove it.
+ pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n",
+map_start, map_last, prange->start,
prange->last);
+ return -EFAULT;
+ }
+
ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
if (!ctx)
return -ENOMEM;
@@ -1747,9 +1754,16 @@ static int svm_range_validate_and_map(struct mm_struct
*mm,
r = -EAGAIN;
}
- if (!r)
- r = svm_range_map_to_gpus(prange, offset, npages,
readonly,
- ctx->bitmap, wait, flush_tlb);
+ if (!r) {
+ map_start = max(map_start, prange->start + offset);
+ map_last = min(map_last, prange->start + offset +
npages);
This should move forward to outside the for loop, otherwise
amdgpu_hmm_range_get_pages and svm_range_dma_map still work on the
entire prange, and then prange->vram_pages update logic should be
changed accordingly.
We need use hmm function to get all vram page number on whole range as
we did not know how many vram pages after partial migration, then we
know if can release vram bo.
Regards
Xiaogang
Regards,
Philip
+ if (map_start <= map_last) {
+ offset = map_start - prange->start;
+ npages = map_last - map_start + 1;
+ r = svm_range_map_to_gpus(prange, offset,
npages, readonly,
+ ctx->bitmap, wait,
flush_tlb);
+ }
+ }
if (!r && next == end)
prange->mapped_to_gpu = true;
@@ -1855,8 +1869,8 @@ static void svm_range_restore_work(struct work_struct
*work)
*/
mutex_lock(&prange->migrate_mutex);
- r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
- false, true, false);
+ r = svm_range_validate_and_map(mm, prange->start, prange->last,
prange,
+ MAX_GPU_INSTANCE, false, true,
false);
if (r)
pr_debug("failed %d to map 0x%lx to gpus\n", r,
prange->start);
@@ -3069,6 +3083,8 @@ svm_range_restore_pages(struct amdgpu_device *adev,
unsigned int pasid,
kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
write_fault, timestamp);
+ start = prange->start;
+ last = prange->last;
if (prange->actual_loc != 0 || best_loc != 0) {
migration = true;
/* Align migration range start and size to granularity size */
@@ -3102,10 +3118,11 @@ svm_range_restore_pages(struct amdgpu_device *adev,
unsigned int pasid,
}
}
- r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
+ r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false,
+ false, false);
if (r)
pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
-r, svms, prange->start, prange->last);
+r, svms, start, last);
kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
migration)
Re: [PATCH] drm/amdkfd: Use partial mapping in GPU page fault recovery
On 2023-10-19 10:24, Xiaogang.Chen
wrote:
From: Xiaogang Chen
After partial migration to recover GPU page fault this patch does GPU vm
space mapping for same page range that got migrated instead of mapping all
pages of svm range in which the page fault happened.
Signed-off-by: Xiaogang Chen
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 33 +---
1 file changed, 25 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 54af7a2b29f8..81dbcc8a4ccc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1619,6 +1619,7 @@ static void *kfd_svm_page_owner(struct kfd_process *p, int32_t gpuidx)
* 5. Release page table (and SVM BO) reservation
*/
static int svm_range_validate_and_map(struct mm_struct *mm,
+ unsigned long map_start, unsigned long map_last,
struct svm_range *prange, int32_t gpuidx,
bool intr, bool wait, bool flush_tlb)
{
@@ -1630,6 +1631,12 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
int32_t idx;
int r = 0;
+ if (map_start < prange->start || map_last > prange->last) {
This is not needed, as this case should never happen, and you also
use max/min to limit map_start, map_last below.
+ pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n",
+ map_start, map_last, prange->start, prange->last);
+ return -EFAULT;
+ }
+
ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
if (!ctx)
return -ENOMEM;
@@ -1747,9 +1754,16 @@ static int svm_range_validate_and_map(struct mm_struct *mm,
r = -EAGAIN;
}
- if (!r)
- r = svm_range_map_to_gpus(prange, offset, npages, readonly,
- ctx->bitmap, wait, flush_tlb);
+ if (!r) {
+ map_start = max(map_start, prange->start + offset);
+ map_last = min(map_last, prange->start + offset + npages);
This should move forward to outside the for loop, otherwise
amdgpu_hmm_range_get_pages and svm_range_dma_map still work on the
entire prange, and then prange->vram_pages update logic should
be changed accordingly.
Regards,
Philip
+ if (map_start <= map_last) {
+offset = map_start - prange->start;
+npages = map_last - map_start + 1;
+r = svm_range_map_to_gpus(prange, offset, npages, readonly,
+ ctx->bitmap, wait, flush_tlb);
+ }
+ }
if (!r && next == end)
prange->mapped_to_gpu = true;
@@ -1855,8 +1869,8 @@ static void svm_range_restore_work(struct work_struct *work)
*/
mutex_lock(&prange->migrate_mutex);
- r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
- false, true, false);
+ r = svm_range_validate_and_map(mm, prange->start, prange->last, prange,
+ MAX_GPU_INSTANCE, false, true, false);
if (r)
pr_debug("failed %d to map 0x%lx to gpus\n", r,
prange->start);
@@ -3069,6 +3083,8 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
write_fault, timestamp);
+ start = prange->start;
+ last = prange->last;
if (prange->actual_loc != 0 || best_loc != 0) {
migration = true;
/* Align migration range start and size to granularity size */
@@ -3102,10 +3118,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
}
}
- r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
+ r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false,
+ false, false);
if (r)
pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
- r, svms, prange->start, prange->last);
+ r, svms, start, last);
kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
migration);
@@ -3650,7 +3667,7 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm,
flush_tlb = !migrated && update_mapping && prange->mapped_to_gpu;
- r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
+ r = svm_range_validate_and_map(mm, prange->start, prange->last, prange, MAX_GPU_INSTANCE,
true, true, flush_tlb);
if (r)
pr_debug("failed %d to map svm range\n", r);
[PATCH] drm/amdkfd: Use partial mapping in GPU page fault recovery
From: Xiaogang Chen
After partial migration to recover GPU page fault this patch does GPU vm
space mapping for same page range that got migrated instead of mapping all
pages of svm range in which the page fault happened.
Signed-off-by: Xiaogang Chen
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 33 +---
1 file changed, 25 insertions(+), 8 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index 54af7a2b29f8..81dbcc8a4ccc 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1619,6 +1619,7 @@ static void *kfd_svm_page_owner(struct kfd_process *p,
int32_t gpuidx)
* 5. Release page table (and SVM BO) reservation
*/
static int svm_range_validate_and_map(struct mm_struct *mm,
+ unsigned long map_start, unsigned long
map_last,
struct svm_range *prange, int32_t gpuidx,
bool intr, bool wait, bool flush_tlb)
{
@@ -1630,6 +1631,12 @@ static int svm_range_validate_and_map(struct mm_struct
*mm,
int32_t idx;
int r = 0;
+ if (map_start < prange->start || map_last > prange->last) {
+ pr_debug("range [0x%lx 0x%lx] out prange [0x%lx 0x%lx]\n",
+map_start, map_last, prange->start,
prange->last);
+ return -EFAULT;
+ }
+
ctx = kzalloc(sizeof(struct svm_validate_context), GFP_KERNEL);
if (!ctx)
return -ENOMEM;
@@ -1747,9 +1754,16 @@ static int svm_range_validate_and_map(struct mm_struct
*mm,
r = -EAGAIN;
}
- if (!r)
- r = svm_range_map_to_gpus(prange, offset, npages,
readonly,
- ctx->bitmap, wait, flush_tlb);
+ if (!r) {
+ map_start = max(map_start, prange->start + offset);
+ map_last = min(map_last, prange->start + offset +
npages);
+ if (map_start <= map_last) {
+ offset = map_start - prange->start;
+ npages = map_last - map_start + 1;
+ r = svm_range_map_to_gpus(prange, offset,
npages, readonly,
+ ctx->bitmap, wait,
flush_tlb);
+ }
+ }
if (!r && next == end)
prange->mapped_to_gpu = true;
@@ -1855,8 +1869,8 @@ static void svm_range_restore_work(struct work_struct
*work)
*/
mutex_lock(&prange->migrate_mutex);
- r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
- false, true, false);
+ r = svm_range_validate_and_map(mm, prange->start, prange->last,
prange,
+ MAX_GPU_INSTANCE, false, true,
false);
if (r)
pr_debug("failed %d to map 0x%lx to gpus\n", r,
prange->start);
@@ -3069,6 +3083,8 @@ svm_range_restore_pages(struct amdgpu_device *adev,
unsigned int pasid,
kfd_smi_event_page_fault_start(node, p->lead_thread->pid, addr,
write_fault, timestamp);
+ start = prange->start;
+ last = prange->last;
if (prange->actual_loc != 0 || best_loc != 0) {
migration = true;
/* Align migration range start and size to granularity size */
@@ -3102,10 +3118,11 @@ svm_range_restore_pages(struct amdgpu_device *adev,
unsigned int pasid,
}
}
- r = svm_range_validate_and_map(mm, prange, gpuidx, false, false, false);
+ r = svm_range_validate_and_map(mm, start, last, prange, gpuidx, false,
+ false, false);
if (r)
pr_debug("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n",
-r, svms, prange->start, prange->last);
+r, svms, start, last);
kfd_smi_event_page_fault_end(node, p->lead_thread->pid, addr,
migration);
@@ -3650,7 +3667,7 @@ svm_range_set_attr(struct kfd_process *p, struct
mm_struct *mm,
flush_tlb = !migrated && update_mapping &&
prange->mapped_to_gpu;
- r = svm_range_validate_and_map(mm, prange, MAX_GPU_INSTANCE,
+ r = svm_range_validate_and_map(mm, prange->start, prange->last,
prange, MAX_GPU_INSTANCE,
true, true, flush_tlb);
if (r)
pr_debug("failed %d to map svm range\n", r);
--
2.25.1
