* [PATCH v2 1/3] drm/amdgpu: Allow TTM to evict svm bo from same process @ 2022-07-25 17:19 Philip Yang 2022-07-25 17:19 ` [PATCH v2 2/3] drm/amdkfd: Set svm range max pages Philip Yang 2022-07-25 17:19 ` [PATCH v2 3/3] drm/amdkfd: Split giant svm range Philip Yang 0 siblings, 2 replies; 6+ messages in thread From: Philip Yang @ 2022-07-25 17:19 UTC (permalink / raw) To: amd-gfx; +Cc: Philip Yang, Felix Kuehling To support SVM range VRAM overcommitment, TTM should be able to evict svm bo of same process to system memory, to get space to alloc new svm bo. Signed-off-by: Philip Yang <Philip.Yang@amd.com> Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com> --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c index 1d0dbff87d3f..469785d33791 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c @@ -159,11 +159,14 @@ static void amdkfd_fence_release(struct dma_fence *f) } /** - * amdkfd_fence_check_mm - Check if @mm is same as that of the fence @f - * if same return TRUE else return FALSE. + * amdkfd_fence_check_mm - Check whether to prevent eviction of @f by @mm * * @f: [IN] fence * @mm: [IN] mm that needs to be verified + * + * Check if @mm is same as that of the fence @f, if same return TRUE else + * return FALSE. + * For svm bo, which support vram overcommitment, always return FALSE. */ bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm) { @@ -171,7 +174,7 @@ bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm) if (!fence) return false; - else if (fence->mm == mm) + else if (fence->mm == mm && !fence->svm_bo) return true; return false; -- 2.35.1 ^ permalink raw reply related [flat|nested] 6+ messages in thread
* [PATCH v2 2/3] drm/amdkfd: Set svm range max pages 2022-07-25 17:19 [PATCH v2 1/3] drm/amdgpu: Allow TTM to evict svm bo from same process Philip Yang @ 2022-07-25 17:19 ` Philip Yang 2022-07-25 18:02 ` Felix Kuehling 2022-07-25 17:19 ` [PATCH v2 3/3] drm/amdkfd: Split giant svm range Philip Yang 1 sibling, 1 reply; 6+ messages in thread From: Philip Yang @ 2022-07-25 17:19 UTC (permalink / raw) To: amd-gfx; +Cc: Philip Yang, felix.kuehling This will be used to split giant svm range into smaller ranges, to support VRAM overcommitment by giant range and improve GPU retry fault recover on giant range. Signed-off-by: Philip Yang <Philip.Yang@amd.com> --- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 2 ++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 19 +++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 3 +++ 3 files changed, 24 insertions(+) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c index 9667015a6cbc..b1f87aa6138b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c @@ -1019,6 +1019,8 @@ int svm_migrate_init(struct amdgpu_device *adev) amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); + svm_range_set_max_pages(adev); + pr_info("HMM registered %ldMB device memory\n", size >> 20); return 0; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index b592aee6d9d6..098060147de6 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -46,6 +46,11 @@ */ #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC) +/* Giant svm range split into smaller ranges based on this, it is decided using + * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and align to 2MB. + */ +uint64_t max_svm_range_pages; + struct criu_svm_metadata { struct list_head list; struct kfd_criu_svm_range_priv_data data; @@ -1870,6 +1875,20 @@ static struct svm_range *svm_range_clone(struct svm_range *old) return new; } +void svm_range_set_max_pages(struct amdgpu_device *adev) +{ + uint64_t max_pages; + uint64_t pages; + + /* 1/32 VRAM size in pages */ + pages = adev->gmc.real_vram_size >> 17; + pages = clamp(pages, 1ULL << 9, 1ULL << 18); + max_pages = READ_ONCE(max_svm_range_pages); + max_pages = min_not_zero(max_pages, pages); + max_pages = ALIGN(max_pages, 1ULL << 9); + WRITE_ONCE(max_svm_range_pages, max_pages); +} + /** * svm_range_add - add svm range and handle overlap * @p: the range add to this process svms diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index eab7f6d3b13c..9156b041ef17 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -204,6 +204,9 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s #define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0) void svm_range_bo_unref_async(struct svm_range_bo *svm_bo); + +void svm_range_set_max_pages(struct amdgpu_device *adev); + #else struct kfd_process; -- 2.35.1 ^ permalink raw reply related [flat|nested] 6+ messages in thread
* Re: [PATCH v2 2/3] drm/amdkfd: Set svm range max pages 2022-07-25 17:19 ` [PATCH v2 2/3] drm/amdkfd: Set svm range max pages Philip Yang @ 2022-07-25 18:02 ` Felix Kuehling 2022-07-25 21:14 ` philip yang 0 siblings, 1 reply; 6+ messages in thread From: Felix Kuehling @ 2022-07-25 18:02 UTC (permalink / raw) To: Philip Yang, amd-gfx Am 2022-07-25 um 13:19 schrieb Philip Yang: > This will be used to split giant svm range into smaller ranges, to > support VRAM overcommitment by giant range and improve GPU retry fault > recover on giant range. > > Signed-off-by: Philip Yang <Philip.Yang@amd.com> > --- > drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 2 ++ > drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 19 +++++++++++++++++++ > drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 3 +++ > 3 files changed, 24 insertions(+) > > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c > index 9667015a6cbc..b1f87aa6138b 100644 > --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c > @@ -1019,6 +1019,8 @@ int svm_migrate_init(struct amdgpu_device *adev) > > amdgpu_amdkfd_reserve_system_mem(SVM_HMM_PAGE_STRUCT_SIZE(size)); > > + svm_range_set_max_pages(adev); > + > pr_info("HMM registered %ldMB device memory\n", size >> 20); > > return 0; > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c > index b592aee6d9d6..098060147de6 100644 > --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c > @@ -46,6 +46,11 @@ > */ > #define AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING (2UL * NSEC_PER_MSEC) > > +/* Giant svm range split into smaller ranges based on this, it is decided using > + * minimum of all dGPU/APU 1/32 VRAM size, between 2MB to 1GB and align to 2MB. > + */ > +uint64_t max_svm_range_pages; > + > struct criu_svm_metadata { > struct list_head list; > struct kfd_criu_svm_range_priv_data data; > @@ -1870,6 +1875,20 @@ static struct svm_range *svm_range_clone(struct svm_range *old) > return new; > } > > +void svm_range_set_max_pages(struct amdgpu_device *adev) > +{ > + uint64_t max_pages; > + uint64_t pages; > + > + /* 1/32 VRAM size in pages */ > + pages = adev->gmc.real_vram_size >> 17; > + pages = clamp(pages, 1ULL << 9, 1ULL << 18); > + max_pages = READ_ONCE(max_svm_range_pages); > + max_pages = min_not_zero(max_pages, pages); > + max_pages = ALIGN(max_pages, 1ULL << 9); In the next patch you use max_svm_range_pages as alignment parameter in an ALIGN_DOWN macro. The ALIGN... macros assume that the alignment is a power of two. Just aligning it with 2MB is not enough. I also don't understand why you do the alignment after taking the min_not_zero. If you assume that max_svm_range_pages was correctly aligned before, you can just to the alignment to a power of two before the min_not_zero call. The READ_ONCE ... WRITE_ONCE is not atomic. It should work as long as this function can't be called on multiple threads concurrently. That is, it should work as long as GPU initialization or hotplug is somehow serialized. I'm not sure whether that's the case. If that is not assured, the proper way to do this is either with a global or static spinlock or with a cmpxchg in a retry loop. Regards, Felix > + WRITE_ONCE(max_svm_range_pages, max_pages); > +} > + > /** > * svm_range_add - add svm range and handle overlap > * @p: the range add to this process svms > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h > index eab7f6d3b13c..9156b041ef17 100644 > --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h > @@ -204,6 +204,9 @@ void svm_range_list_lock_and_flush_work(struct svm_range_list *svms, struct mm_s > #define KFD_IS_SVM_API_SUPPORTED(dev) ((dev)->pgmap.type != 0) > > void svm_range_bo_unref_async(struct svm_range_bo *svm_bo); > + > +void svm_range_set_max_pages(struct amdgpu_device *adev); > + > #else > > struct kfd_process; ^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v2 2/3] drm/amdkfd: Set svm range max pages 2022-07-25 18:02 ` Felix Kuehling @ 2022-07-25 21:14 ` philip yang 0 siblings, 0 replies; 6+ messages in thread From: philip yang @ 2022-07-25 21:14 UTC (permalink / raw) To: Felix Kuehling, Philip Yang, amd-gfx [-- Attachment #1: Type: text/html, Size: 6968 bytes --] ^ permalink raw reply [flat|nested] 6+ messages in thread
* [PATCH v2 3/3] drm/amdkfd: Split giant svm range 2022-07-25 17:19 [PATCH v2 1/3] drm/amdgpu: Allow TTM to evict svm bo from same process Philip Yang 2022-07-25 17:19 ` [PATCH v2 2/3] drm/amdkfd: Set svm range max pages Philip Yang @ 2022-07-25 17:19 ` Philip Yang 2022-07-25 18:04 ` Felix Kuehling 1 sibling, 1 reply; 6+ messages in thread From: Philip Yang @ 2022-07-25 17:19 UTC (permalink / raw) To: amd-gfx; +Cc: Philip Yang, felix.kuehling Giant svm range split to smaller ranges, align the range start address to max svm range pages to improve MMU TLB usage. Signed-off-by: Philip Yang <Philip.Yang@amd.com> --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 49 +++++++++++++++++++--------- 1 file changed, 33 insertions(+), 16 deletions(-) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 098060147de6..6529a40fb508 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1889,6 +1889,31 @@ void svm_range_set_max_pages(struct amdgpu_device *adev) WRITE_ONCE(max_svm_range_pages, max_pages); } +static int +svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last, + uint64_t max_pages, struct list_head *insert_list, + struct list_head *update_list) +{ + struct svm_range *prange; + uint64_t l; + + pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n", + max_pages, start, last); + + while (last >= start) { + l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1); + + prange = svm_range_new(svms, start, l); + if (!prange) + return -ENOMEM; + list_add(&prange->list, insert_list); + list_add(&prange->update_list, update_list); + + start = l + 1; + } + return 0; +} + /** * svm_range_add - add svm range and handle overlap * @p: the range add to this process svms @@ -1991,14 +2016,11 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, /* insert a new node if needed */ if (node->start > start) { - prange = svm_range_new(svms, start, node->start - 1); - if (!prange) { - r = -ENOMEM; + r = svm_range_split_new(svms, start, node->start - 1, + READ_ONCE(max_svm_range_pages), + insert_list, update_list); + if (r) goto out; - } - - list_add(&prange->list, insert_list); - list_add(&prange->update_list, update_list); } node = next; @@ -2006,15 +2028,10 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, } /* add a final range at the end if needed */ - if (start <= last) { - prange = svm_range_new(svms, start, last); - if (!prange) { - r = -ENOMEM; - goto out; - } - list_add(&prange->list, insert_list); - list_add(&prange->update_list, update_list); - } + if (start <= last) + r = svm_range_split_new(svms, start, last, + READ_ONCE(max_svm_range_pages), + insert_list, update_list); out: if (r) -- 2.35.1 ^ permalink raw reply related [flat|nested] 6+ messages in thread
* Re: [PATCH v2 3/3] drm/amdkfd: Split giant svm range 2022-07-25 17:19 ` [PATCH v2 3/3] drm/amdkfd: Split giant svm range Philip Yang @ 2022-07-25 18:04 ` Felix Kuehling 0 siblings, 0 replies; 6+ messages in thread From: Felix Kuehling @ 2022-07-25 18:04 UTC (permalink / raw) To: Philip Yang, amd-gfx Am 2022-07-25 um 13:19 schrieb Philip Yang: > Giant svm range split to smaller ranges, align the range start address > to max svm range pages to improve MMU TLB usage. > > Signed-off-by: Philip Yang <Philip.Yang@amd.com> Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com> > --- > drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 49 +++++++++++++++++++--------- > 1 file changed, 33 insertions(+), 16 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c > index 098060147de6..6529a40fb508 100644 > --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c > @@ -1889,6 +1889,31 @@ void svm_range_set_max_pages(struct amdgpu_device *adev) > WRITE_ONCE(max_svm_range_pages, max_pages); > } > > +static int > +svm_range_split_new(struct svm_range_list *svms, uint64_t start, uint64_t last, > + uint64_t max_pages, struct list_head *insert_list, > + struct list_head *update_list) > +{ > + struct svm_range *prange; > + uint64_t l; > + > + pr_debug("max_svm_range_pages 0x%llx adding [0x%llx 0x%llx]\n", > + max_pages, start, last); > + > + while (last >= start) { > + l = min(last, ALIGN_DOWN(start + max_pages, max_pages) - 1); > + > + prange = svm_range_new(svms, start, l); > + if (!prange) > + return -ENOMEM; > + list_add(&prange->list, insert_list); > + list_add(&prange->update_list, update_list); > + > + start = l + 1; > + } > + return 0; > +} > + > /** > * svm_range_add - add svm range and handle overlap > * @p: the range add to this process svms > @@ -1991,14 +2016,11 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, > > /* insert a new node if needed */ > if (node->start > start) { > - prange = svm_range_new(svms, start, node->start - 1); > - if (!prange) { > - r = -ENOMEM; > + r = svm_range_split_new(svms, start, node->start - 1, > + READ_ONCE(max_svm_range_pages), > + insert_list, update_list); > + if (r) > goto out; > - } > - > - list_add(&prange->list, insert_list); > - list_add(&prange->update_list, update_list); > } > > node = next; > @@ -2006,15 +2028,10 @@ svm_range_add(struct kfd_process *p, uint64_t start, uint64_t size, > } > > /* add a final range at the end if needed */ > - if (start <= last) { > - prange = svm_range_new(svms, start, last); > - if (!prange) { > - r = -ENOMEM; > - goto out; > - } > - list_add(&prange->list, insert_list); > - list_add(&prange->update_list, update_list); > - } > + if (start <= last) > + r = svm_range_split_new(svms, start, last, > + READ_ONCE(max_svm_range_pages), > + insert_list, update_list); > > out: > if (r) ^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2022-07-25 21:14 UTC | newest] Thread overview: 6+ messages (download: mbox.gz follow: Atom feed -- links below jump to the message on this page -- 2022-07-25 17:19 [PATCH v2 1/3] drm/amdgpu: Allow TTM to evict svm bo from same process Philip Yang 2022-07-25 17:19 ` [PATCH v2 2/3] drm/amdkfd: Set svm range max pages Philip Yang 2022-07-25 18:02 ` Felix Kuehling 2022-07-25 21:14 ` philip yang 2022-07-25 17:19 ` [PATCH v2 3/3] drm/amdkfd: Split giant svm range Philip Yang 2022-07-25 18:04 ` Felix Kuehling
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.