From: Philip Yang Philip.Yang@amd.com
Use HMM to get system memory pages address, which will be used to map to GPUs or migrate to vram.
Signed-off-by: Philip Yang Philip.Yang@amd.com Signed-off-by: Felix Kuehling Felix.Kuehling@amd.com --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 103 ++++++++++++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_svm.h | 4 ++ 2 files changed, 106 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index e57103a9025e..6024caf7373f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -28,6 +28,15 @@ #include "kfd_priv.h" #include "kfd_svm.h"
+static bool +svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, + const struct mmu_notifier_range *range, + unsigned long cur_seq); + +static const struct mmu_interval_notifier_ops svm_range_mn_ops = { + .invalidate = svm_range_cpu_invalidate_pagetables, +}; + /** * svm_range_unlink - unlink svm_range from lists and interval tree * @prange: svm range structure to be removed @@ -46,6 +55,18 @@ static void svm_range_unlink(struct svm_range *prange) interval_tree_remove(&prange->it_node, &prange->svms->objects); }
+static void +svm_range_add_notifier_locked(struct mm_struct *mm, struct svm_range *prange) +{ + pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, + prange, prange->start, prange->last); + + mmu_interval_notifier_insert_locked(&prange->notifier, mm, + prange->start << PAGE_SHIFT, + prange->npages << PAGE_SHIFT, + &svm_range_mn_ops); +} + /** * svm_range_add_to_svms - add svm range to svms * @prange: svm range structure to be added @@ -65,6 +86,18 @@ static void svm_range_add_to_svms(struct svm_range *prange) interval_tree_insert(&prange->it_node, &prange->svms->objects); }
+static void svm_range_remove_notifier(struct svm_range *prange) +{ + pr_debug("remove notifier svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", + prange->svms, prange, + prange->notifier.interval_tree.start >> PAGE_SHIFT, + prange->notifier.interval_tree.last >> PAGE_SHIFT); + + if (prange->notifier.interval_tree.start != 0 && + prange->notifier.interval_tree.last != 0) + mmu_interval_notifier_remove(&prange->notifier); +} + static void svm_range_free(struct svm_range *prange) { pr_debug("svms 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, @@ -112,6 +145,56 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, return prange; }
+/** + * svm_range_validate_ram - get system memory pages of svm range + * + * @mm: the mm_struct of process + * @prange: the range struct + * + * After mapping system memory to GPU, system memory maybe invalidated anytime + * during application running, we use HMM callback to sync GPU with CPU page + * table update, so we don't need use lock to prevent CPU invalidation and check + * hmm_range_get_pages_done return value. + * + * Return: + * 0 - OK, otherwise error code + */ +static int +svm_range_validate_ram(struct mm_struct *mm, struct svm_range *prange) +{ + int r; + + r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL, + prange->start << PAGE_SHIFT, + prange->npages, &prange->hmm_range, + false, true); + if (r) { + pr_debug("failed %d to get svm range pages\n", r); + return r; + } + + kvfree(prange->pages_addr); + prange->pages_addr = prange->hmm_range->hmm_pfns; + prange->hmm_range->hmm_pfns = NULL; + + amdgpu_hmm_range_get_pages_done(prange->hmm_range); + prange->hmm_range = NULL; + + return 0; +} + +static int +svm_range_validate(struct mm_struct *mm, struct svm_range *prange) +{ + int r = 0; + + pr_debug("actual loc 0x%x\n", prange->actual_loc); + + r = svm_range_validate_ram(mm, prange); + + return r; +} + static int svm_range_check_attr(struct kfd_process *p, uint32_t nattr, struct kfd_ioctl_svm_attribute *attrs) @@ -581,6 +664,18 @@ svm_range_handle_overlap(struct svm_range_list *svms, struct svm_range *new, return r; }
+/** + * svm_range_cpu_invalidate_pagetables - interval notifier callback + * + */ +static bool +svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni, + const struct mmu_notifier_range *range, + unsigned long cur_seq) +{ + return true; +} + void svm_range_list_fini(struct kfd_process *p) { pr_debug("pasid 0x%x svms 0x%p\n", p->pasid, &p->svms); @@ -732,6 +827,7 @@ svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size, /* Apply changes as a transaction */ list_for_each_entry_safe(prange, next, &insert_list, insert_list) { svm_range_add_to_svms(prange); + svm_range_add_notifier_locked(mm, prange); } list_for_each_entry(prange, &update_list, update_list) { svm_range_apply_attrs(p, prange, nattr, attrs); @@ -743,6 +839,7 @@ svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size, prange->svms, prange, prange->start, prange->last); svm_range_unlink(prange); + svm_range_remove_notifier(prange); svm_range_free(prange); }
@@ -753,7 +850,11 @@ svm_range_set_attr(struct kfd_process *p, uint64_t start, uint64_t size, * case because the rollback wouldn't be guaranteed to work either. */ list_for_each_entry(prange, &update_list, update_list) { - /* TODO */ + r = svm_range_validate(mm, prange); + if (r) { + pr_debug("failed %d to validate svm range\n", r); + break; + } }
svm_range_debug_dump(svms); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h index c46cdad9d1a3..f35a178c607f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h @@ -44,6 +44,7 @@ * @update_list:link list node used to add to update_list * @remove_list:link list node used to add to remove list * @insert_list:link list node used to add to insert list + * @hmm_range: hmm range structure used by hmm_range_fault to get system pages * @npages: number of pages * @pages_addr: list of system memory physical page address * @flags: flags defined as KFD_IOCTL_SVM_FLAG_* @@ -51,6 +52,7 @@ * @perfetch_loc: last prefetch location, 0 for CPU, or GPU id * @actual_loc: the actual location, 0 for CPU, or GPU id * @granularity:migration granularity, log2 num pages + * @notifier: register mmu interval notifier * @bitmap_access: index bitmap of GPUs which can access the range * @bitmap_aip: index bitmap of GPUs which can access the range in place * @@ -67,6 +69,7 @@ struct svm_range { struct list_head update_list; struct list_head remove_list; struct list_head insert_list; + struct hmm_range *hmm_range; uint64_t npages; unsigned long *pages_addr; uint32_t flags; @@ -74,6 +77,7 @@ struct svm_range { uint32_t prefetch_loc; uint32_t actual_loc; uint8_t granularity; + struct mmu_interval_notifier notifier; DECLARE_BITMAP(bitmap_access, MAX_GPU_INSTANCE); DECLARE_BITMAP(bitmap_aip, MAX_GPU_INSTANCE); };