[PATCH] drm/amdgpu: Add notifier lock for KFD userptrs
Felix Kuehling
Felix.Kuehling at amd.com
Thu Nov 3 02:00:28 UTC 2022
Add a per-process MMU notifier lock for processing notifiers from
userptrs. Use that lock to properly synchronize page table updates with
MMU notifiers.
Signed-off-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 12 +-
.../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 202 +++++++++++-------
drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c | 12 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_mn.h | 3 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 18 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 4 +
6 files changed, 158 insertions(+), 93 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index f50e3ba4d7a5..1ca18a77818b 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -29,6 +29,7 @@
#include <linux/mm.h>
#include <linux/kthread.h>
#include <linux/workqueue.h>
+#include <linux/mmu_notifier.h>
#include <kgd_kfd_interface.h>
#include <drm/ttm/ttm_execbuf_util.h>
#include "amdgpu_sync.h"
@@ -75,7 +76,7 @@ struct kgd_mem {
uint32_t alloc_flags;
- atomic_t invalid;
+ uint32_t invalid;
struct amdkfd_process_info *process_info;
struct amdgpu_sync sync;
@@ -131,7 +132,8 @@ struct amdkfd_process_info {
struct amdgpu_amdkfd_fence *eviction_fence;
/* MMU-notifier related fields */
- atomic_t evicted_bos;
+ struct mutex notifier_lock;
+ uint32_t evicted_bos;
struct delayed_work restore_userptr_work;
struct pid *pid;
bool block_mmu_notifications;
@@ -180,7 +182,8 @@ int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data);
bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo);
-int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm);
+int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
+ unsigned long cur_seq, struct kgd_mem *mem);
#else
static inline
bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
@@ -201,7 +204,8 @@ int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
}
static inline
-int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
+int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
+ unsigned long cur_seq, struct kgd_mem *mem)
{
return 0;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
index 955fa8c8213b..5510b7c42ac7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
@@ -964,7 +964,9 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
* later stage when it is scheduled by another ioctl called by
* CRIU master process for the target pid for restore.
*/
- atomic_inc(&mem->invalid);
+ mutex_lock(&process_info->notifier_lock);
+ mem->invalid++;
+ mutex_unlock(&process_info->notifier_lock);
mutex_unlock(&process_info->lock);
return 0;
}
@@ -1301,6 +1303,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
return -ENOMEM;
mutex_init(&info->lock);
+ mutex_init(&info->notifier_lock);
INIT_LIST_HEAD(&info->vm_list_head);
INIT_LIST_HEAD(&info->kfd_bo_list);
INIT_LIST_HEAD(&info->userptr_valid_list);
@@ -1317,7 +1320,6 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
}
info->pid = get_task_pid(current->group_leader, PIDTYPE_PID);
- atomic_set(&info->evicted_bos, 0);
INIT_DELAYED_WORK(&info->restore_userptr_work,
amdgpu_amdkfd_restore_userptr_worker);
@@ -1372,6 +1374,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info,
put_pid(info->pid);
create_evict_fence_fail:
mutex_destroy(&info->lock);
+ mutex_destroy(&info->notifier_lock);
kfree(info);
}
return ret;
@@ -1496,6 +1499,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
cancel_delayed_work_sync(&process_info->restore_userptr_work);
put_pid(process_info->pid);
mutex_destroy(&process_info->lock);
+ mutex_destroy(&process_info->notifier_lock);
kfree(process_info);
}
}
@@ -1548,7 +1552,9 @@ int amdgpu_amdkfd_criu_resume(void *p)
mutex_lock(&pinfo->lock);
pr_debug("scheduling work\n");
- atomic_inc(&pinfo->evicted_bos);
+ mutex_lock(&pinfo->notifier_lock);
+ pinfo->evicted_bos++;
+ mutex_unlock(&pinfo->notifier_lock);
if (!READ_ONCE(pinfo->block_mmu_notifications)) {
ret = -EINVAL;
goto out_unlock;
@@ -1773,8 +1779,13 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
list_del(&bo_list_entry->head);
mutex_unlock(&process_info->lock);
- /* No more MMU notifiers */
- amdgpu_mn_unregister(mem->bo);
+ /* Cleanup user pages and MMU notifiers */
+ if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) {
+ amdgpu_mn_unregister(mem->bo);
+ mutex_lock(&process_info->notifier_lock);
+ amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm);
+ mutex_unlock(&process_info->notifier_lock);
+ }
ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx);
if (unlikely(ret))
@@ -1864,14 +1875,14 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
*/
mutex_lock(&mem->process_info->lock);
- /* Lock mmap-sem. If we find an invalid userptr BO, we can be
+ /* Lock notifier lock. If we find an invalid userptr BO, we can be
* sure that the MMU notifier is no longer running
* concurrently and the queues are actually stopped
*/
if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
- mmap_write_lock(current->mm);
- is_invalid_userptr = atomic_read(&mem->invalid);
- mmap_write_unlock(current->mm);
+ mutex_lock(&mem->process_info->notifier_lock);
+ is_invalid_userptr = !!mem->invalid;
+ mutex_unlock(&mem->process_info->notifier_lock);
}
mutex_lock(&mem->lock);
@@ -2251,34 +2262,38 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev,
*
* Runs in MMU notifier, may be in RECLAIM_FS context. This means it
* cannot do any memory allocations, and cannot take any locks that
- * are held elsewhere while allocating memory. Therefore this is as
- * simple as possible, using atomic counters.
+ * are held elsewhere while allocating memory.
*
* It doesn't do anything to the BO itself. The real work happens in
* restore, where we get updated page addresses. This function only
* ensures that GPU access to the BO is stopped.
*/
-int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem,
- struct mm_struct *mm)
+int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni,
+ unsigned long cur_seq, struct kgd_mem *mem)
{
struct amdkfd_process_info *process_info = mem->process_info;
- int evicted_bos;
int r = 0;
- /* Do not process MMU notifications until stage-4 IOCTL is received */
+ /* Do not process MMU notifications during CRIU restore until
+ * KFD_CRIU_OP_RESUME IOCTL is received
+ */
if (READ_ONCE(process_info->block_mmu_notifications))
return 0;
- atomic_inc(&mem->invalid);
- evicted_bos = atomic_inc_return(&process_info->evicted_bos);
- if (evicted_bos == 1) {
+ mutex_lock(&process_info->notifier_lock);
+ mmu_interval_set_seq(mni, cur_seq);
+
+ mem->invalid++;
+ if (++process_info->evicted_bos == 1) {
/* First eviction, stop the queues */
- r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
+ r = kgd2kfd_quiesce_mm(mni->mm,
+ KFD_QUEUE_EVICTION_TRIGGER_USERPTR);
if (r)
pr_err("Failed to quiesce KFD\n");
schedule_delayed_work(&process_info->restore_userptr_work,
msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS));
}
+ mutex_unlock(&process_info->notifier_lock);
return r;
}
@@ -2295,49 +2310,54 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
struct kgd_mem *mem, *tmp_mem;
struct amdgpu_bo *bo;
struct ttm_operation_ctx ctx = { false, false };
- int invalid, ret;
+ uint32_t invalid;
+ int ret = 0;
- /* Move all invalidated BOs to the userptr_inval_list and
- * release their user pages by migration to the CPU domain
- */
+ mutex_lock(&process_info->notifier_lock);
+
+ /* Move all invalidated BOs to the userptr_inval_list */
list_for_each_entry_safe(mem, tmp_mem,
&process_info->userptr_valid_list,
- validate_list.head) {
- if (!atomic_read(&mem->invalid))
- continue; /* BO is still valid */
-
- bo = mem->bo;
-
- if (amdgpu_bo_reserve(bo, true))
- return -EAGAIN;
- amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
- ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
- amdgpu_bo_unreserve(bo);
- if (ret) {
- pr_err("%s: Failed to invalidate userptr BO\n",
- __func__);
- return -EAGAIN;
- }
-
- list_move_tail(&mem->validate_list.head,
- &process_info->userptr_inval_list);
- }
-
- if (list_empty(&process_info->userptr_inval_list))
- return 0; /* All evicted userptr BOs were freed */
+ validate_list.head)
+ if (mem->invalid)
+ list_move_tail(&mem->validate_list.head,
+ &process_info->userptr_inval_list);
/* Go through userptr_inval_list and update any invalid user_pages */
list_for_each_entry(mem, &process_info->userptr_inval_list,
validate_list.head) {
- invalid = atomic_read(&mem->invalid);
+ invalid = mem->invalid;
if (!invalid)
/* BO hasn't been invalidated since the last
- * revalidation attempt. Keep its BO list.
+ * revalidation attempt. Keep its page list.
*/
continue;
bo = mem->bo;
+ amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm);
+
+ /* BO reservations and getting user pages (hmm_range_fault)
+ * must happen outside the notifier lock
+ */
+ mutex_unlock(&process_info->notifier_lock);
+
+ /* Move the BO to system (CPU) domain if necessary to unmap
+ * and free the SG table
+ */
+ if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) {
+ if (amdgpu_bo_reserve(bo, true))
+ return -EAGAIN;
+ amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU);
+ ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
+ amdgpu_bo_unreserve(bo);
+ if (ret) {
+ pr_err("%s: Failed to invalidate userptr BO\n",
+ __func__);
+ return -EAGAIN;
+ }
+ }
+
/* Get updated user pages */
ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages);
if (ret) {
@@ -2352,30 +2372,31 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info,
*/
if (ret != -EFAULT)
return ret;
- } else {
-
- /*
- * FIXME: Cannot ignore the return code, must hold
- * notifier_lock
- */
- amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
+ ret = 0;
}
+ mutex_lock(&process_info->notifier_lock);
+
/* Mark the BO as valid unless it was invalidated
* again concurrently.
*/
- if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid)
- return -EAGAIN;
+ if (mem->invalid != invalid) {
+ ret = -EAGAIN;
+ goto unlock_out;
+ }
+ mem->invalid = 0;
}
- return 0;
+unlock_out:
+ mutex_unlock(&process_info->notifier_lock);
+
+ return ret;
}
/* Validate invalid userptr BOs
*
- * Validates BOs on the userptr_inval_list, and moves them back to the
- * userptr_valid_list. Also updates GPUVM page tables with new page
- * addresses and waits for the page table updates to complete.
+ * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables
+ * with new page addresses and waits for the page table updates to complete.
*/
static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
{
@@ -2446,9 +2467,6 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
}
}
- list_move_tail(&mem->validate_list.head,
- &process_info->userptr_valid_list);
-
/* Update mapping. If the BO was not validated
* (because we couldn't get user pages), this will
* clear the page table entries, which will result in
@@ -2464,7 +2482,9 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
if (ret) {
pr_err("%s: update PTE failed\n", __func__);
/* make sure this gets validated again */
- atomic_inc(&mem->invalid);
+ mutex_lock(&process_info->notifier_lock);
+ mem->invalid++;
+ mutex_unlock(&process_info->notifier_lock);
goto unreserve_out;
}
}
@@ -2484,6 +2504,32 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info)
return ret;
}
+/* Confirm that all user pages are valid while holding the notifier lock
+ *
+ * Moves valid BOs from the userptr_inval_list back to userptr_val_list.
+ */
+static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info)
+{
+ struct kgd_mem *mem, *tmp_mem;
+ int ret = 0;
+
+ list_for_each_entry_safe(mem, tmp_mem,
+ &process_info->userptr_inval_list,
+ validate_list.head) {
+ if (!amdgpu_ttm_tt_get_user_pages_done(mem->bo->tbo.ttm)) {
+ WARN(!mem->invalid, "Invalid BO not marked invalid");
+ ret = -EAGAIN;
+ continue;
+ }
+ WARN(mem->invalid, "Valid BO is marked invalid");
+
+ list_move_tail(&mem->validate_list.head,
+ &process_info->userptr_valid_list);
+ }
+
+ return ret;
+}
+
/* Worker callback to restore evicted userptr BOs
*
* Tries to update and validate all userptr BOs. If successful and no
@@ -2498,9 +2544,11 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
restore_userptr_work);
struct task_struct *usertask;
struct mm_struct *mm;
- int evicted_bos;
+ uint32_t evicted_bos;
- evicted_bos = atomic_read(&process_info->evicted_bos);
+ mutex_lock(&process_info->notifier_lock);
+ evicted_bos = process_info->evicted_bos;
+ mutex_unlock(&process_info->notifier_lock);
if (!evicted_bos)
return;
@@ -2523,9 +2571,6 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
* and we can just restart the queues.
*/
if (!list_empty(&process_info->userptr_inval_list)) {
- if (atomic_read(&process_info->evicted_bos) != evicted_bos)
- goto unlock_out; /* Concurrent eviction, try again */
-
if (validate_invalid_user_pages(process_info))
goto unlock_out;
}
@@ -2534,10 +2579,17 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
* be a first eviction that calls quiesce_mm. The eviction
* reference counting inside KFD will handle this case.
*/
- if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) !=
- evicted_bos)
- goto unlock_out;
- evicted_bos = 0;
+ mutex_lock(&process_info->notifier_lock);
+ if (process_info->evicted_bos != evicted_bos)
+ goto unlock_notifier_out;
+
+ if (confirm_valid_user_pages_locked(process_info)) {
+ WARN(1, "User pages unexpectedly invalid");
+ goto unlock_notifier_out;
+ }
+
+ process_info->evicted_bos = evicted_bos = 0;
+
if (kgd2kfd_resume_mm(mm)) {
pr_err("%s: Failed to resume KFD\n", __func__);
/* No recovery from this failure. Probably the CP is
@@ -2545,6 +2597,8 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work)
*/
}
+unlock_notifier_out:
+ mutex_unlock(&process_info->notifier_lock);
unlock_out:
mutex_unlock(&process_info->lock);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
index b86c0b8252a5..93b6a2a7d64c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
@@ -102,17 +102,11 @@ static bool amdgpu_mn_invalidate_hsa(struct mmu_interval_notifier *mni,
unsigned long cur_seq)
{
struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier);
- struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
if (!mmu_notifier_range_blockable(range))
return false;
- mutex_lock(&adev->notifier_lock);
-
- mmu_interval_set_seq(mni, cur_seq);
-
- amdgpu_amdkfd_evict_userptr(bo->kfd_bo, bo->notifier.mm);
- mutex_unlock(&adev->notifier_lock);
+ amdgpu_amdkfd_evict_userptr(mni, cur_seq, bo->kfd_bo);
return true;
}
@@ -231,9 +225,9 @@ int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
return r;
}
-int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range)
+bool amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range)
{
- int r;
+ bool r;
r = mmu_interval_read_retry(hmm_range->notifier,
hmm_range->notifier_seq);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.h
index 14a3c1864085..b37fcf99baf1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.h
@@ -29,13 +29,14 @@
#include <linux/rwsem.h>
#include <linux/workqueue.h>
#include <linux/interval_tree.h>
+#include <linux/mmu_notifier.h>
int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
struct mm_struct *mm, struct page **pages,
uint64_t start, uint64_t npages,
struct hmm_range **phmm_range, bool readonly,
bool mmap_locked, void *owner);
-int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range);
+bool amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range);
#if defined(CONFIG_HMM_MIRROR)
int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 585460ab8dfd..5f2b87dd5732 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -706,8 +706,20 @@ int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages)
return r;
}
+/* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations
+ */
+void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm)
+{
+ struct amdgpu_ttm_tt *gtt = (void *)ttm;
+
+ if (gtt && gtt->userptr && gtt->range) {
+ amdgpu_hmm_range_get_pages_done(gtt->range);
+ gtt->range = NULL;
+ }
+}
+
/*
- * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change
+ * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change
* Check if the pages backing this ttm range have been invalidated
*
* Returns: true if pages are still valid
@@ -727,10 +739,6 @@ bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
"No user pages to check\n");
if (gtt->range) {
- /*
- * FIXME: Must always hold notifier_lock for this, and must
- * not ignore the return code.
- */
r = amdgpu_hmm_range_get_pages_done(gtt->range);
gtt->range = NULL;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
index 9120ae80ef52..58b70d862437 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
@@ -158,6 +158,7 @@ uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type);
#if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR)
int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages);
+void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm);
bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm);
#else
static inline int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo,
@@ -165,6 +166,9 @@ static inline int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo,
{
return -EPERM;
}
+static void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm)
+{
+}
static inline bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm)
{
return false;
--
2.32.0
More information about the amd-gfx
mailing list