HMM profiler interface
Subsystems (3)
v2 cover HMM profiler interface
The ROCm profiler would expose the data from KFD profiling APIs to application developers to tune the applications based on how the address range attributes affect the behavior and performance.
Per process event log use the existing SMI (system management interface) event API. Each event log is one line of text with the event specific information.
v2:
* Keep existing events behaviour
* Use ktime_get_boottime_ns() as timestamp to correlate with other APIs
* Use compact message layout, stick with existing message convention
* Add unmap from GPU event
Philip Yang (8):
drm/amdkfd: Correct SMI event read size drm/amdkfd: Add KFD SMI event IDs and triggers drm/amdkfd: Enable per process SMI event drm/amdkfd: Add GPU recoverable fault SMI event drm/amdkfd: add migration SMI event
drm/amdkfd: Add user queue eviction restore SMI event drm/amdkfd: Add unmap from GPU SMI event drm/amdkfd: Bump KFD API version for SMI profiling event
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 7 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 11 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 67 ++++--- drivers/gpu/drm/amd/amdkfd/kfd_migrate.h | 5 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 37 +++- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 163 +++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 19 +- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 63 +++++-- include/uapi/linux/kfd_ioctl.h | 30 +++- 11 files changed, 343 insertions(+), 65 deletions(-)
Signature · 1 line
2.17.1
PATCH v2 2/8 drm/amdkfd: Add KFD SMI event IDs and triggers
Define new system management interface event IDs, migration triggers and user queue eviction triggers, those will be implemented in the following patches.
--- include/uapi/linux/kfd_ioctl.h | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+)
include/uapi/linux/kfd_ioctl.h +27 -0
@@ -459,10 +459,37 @@ enum kfd_smi_event {KFD_SMI_EVENT_THERMAL_THROTTLE =2 ,KFD_SMI_EVENT_GPU_PRE_RESET =3 ,KFD_SMI_EVENT_GPU_POST_RESET =4 , +KFD_SMI_EVENT_MIGRATION =5 , +KFD_SMI_EVENT_PAGE_FAULT_START =6 , +KFD_SMI_EVENT_PAGE_FAULT_END =7 , +KFD_SMI_EVENT_QUEUE_EVICTION =8 , +KFD_SMI_EVENT_QUEUE_EVICTION_RESTORE =9 , +KFD_SMI_EVENT_UNMAP_FROM_GPU =10 , + +/* +* max event number, as a flag bit to get events from all processes, +* this requires super user permission, otherwise will not be able to +* receive event from any process. Without this flag to receive events +* from same process. +*/ +KFD_SMI_EVENT_ALL_PROCESS =64 };#define KFD_SMI_EVENT_MASK_FROM_INDEX (i) (1ULL << ((i) - 1)) +enum KFD_MIGRATION_QUEUE_EVICTION_UNMAP_EVENT_TRIGGER { +MIGRATION_TRIGGER_PREFETCH =1 , +MIGRATION_TRIGGER_PAGEFAULT , +MIGRATION_TRIGGER_PAGEFAULT_CPU , +MIGRATION_TRIGGER_TTM_EVICTION , +SVM_RANGE_EVICTION , +SVM_RANGE_MIGRATION , +USERPTR_EVICTION , +TTM_EVICTION , +UNMAP_FROM_CPU , +SUSPEND_EVICTION +}; +struct kfd_ioctl_smi_events_args {__u32 gpuid;/* to KFD */ __u32 anon_fd;/* from KFD */
Signature · 1 line
2.17.1
PATCH v2 1/8 drm/amdkfd: Correct SMI event read size
In reply to Philip Yangsizeof(buf) is 8 bytes because it is defined as unsigned char *buf, each SMI event read only copy max 8 bytes to user buffer. Correct this by using the buf allocate size.
--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +3 -2
@@ -81,7 +81,8 @@ static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,struct kfd_smi_client *client = filep->private_data;unsigned char *buf; - buf =kmalloc_array (MAX_KFIFO_SIZE ,sizeof (*buf),GFP_KERNEL ); + size =min_t (size_t, size,MAX_KFIFO_SIZE ); + buf =kmalloc (size,GFP_KERNEL );if (!buf)return -ENOMEM ;
@@ -95,7 +96,7 @@ static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user, ret = -EAGAIN ; goto ret_err; } - to_copy =min3 (size,sizeof (buf), to_copy); + to_copy =min (size, to_copy); ret =kfifo_out (&client->fifo, buf, to_copy);spin_unlock (&client->lock);if (ret <=0 ) {
Signature · 1 line
2.17.1
PATCH v2 8/8 drm/amdkfd: Bump KFD API version for SMI profiling event
In reply to Philip YangIndicate SMI profiling events available,
--- include/uapi/linux/kfd_ioctl.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
include/uapi/linux/kfd_ioctl.h +2 -1
@@ -32,9 +32,10 @@* - 1.4 - Indicate new SRAM EDC bit in device properties * - 1.5 - Add SVM API * - 1.6 - Query clear flags in SVM get_attr API +* - 1.7 - Add SMI profiler event log */ #define KFD_IOCTL_MAJOR_VERSION 1 -#define KFD_IOCTL_MINOR_VERSION 6 +#define KFD_IOCTL_MINOR_VERSION 7struct kfd_ioctl_get_version_args {__u32 major_version;/* from KFD */
Signature · 1 line
2.17.1
PATCH v2 4/8 drm/amdkfd: Add GPU recoverable fault SMI event
In reply to Philip YangOutput timestamp when GPU recoverable fault starts, ends and duration to recover the fault, if migration happened or only GPU page table is updated, fault address, read or write fault.
--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 48 +++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 7 ++- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 17 ++++++-- 3 files changed, 67 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +48 -0
@@ -266,6 +266,54 @@ void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid)add_event_to_kfifo (0 , dev,KFD_SMI_EVENT_VMFAULT , fifo_in, len); } +static bool kfd_smi_event_duration (struct kfd_dev *dev,uint64_t ts, +uint64_t *duration) +{ +if (list_empty (&dev->smi_clients)) +return false; + + *duration =ktime_get_boottime_ns () - ts; +return true; +} + +void kfd_smi_event_page_fault_start (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool write_fault, +uint64_t ts) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%lx(%x) %c\n" , +KFD_SMI_EVENT_PAGE_FAULT_START , ts, pid, address, + dev->id, write_fault ?'W' :'R' ); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_PAGE_FAULT_START , fifo_in, + len); +} + +void kfd_smi_event_page_fault_end (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool migration, +uint64_t ts) +{ +char fifo_in[64 ]; +uint64_t duration; +int len; + +if (!kfd_smi_event_duration (dev, ts, &duration)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in), +"%x %lld(%lld) -%d @%lx(%x) %c\n" , +KFD_SMI_EVENT_PAGE_FAULT_END ,ktime_get_boottime_ns (), + duration, pid, address, dev->id, migration ?'M' :'m' ); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_PAGE_FAULT_END , fifo_in, + len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +6 -1
@@ -28,5 +28,10 @@ void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);void kfd_smi_event_update_thermal_throttling (struct kfd_dev *dev,uint64_t throttle_bitmask);void kfd_smi_event_update_gpu_reset (struct kfd_dev *dev,bool post_reset); - +void kfd_smi_event_page_fault_start (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool write_fault, +uint64_t ts); +void kfd_smi_event_page_fault_end (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool migration, +uint64_t ts);#endif
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +13 -4
@@ -32,6 +32,7 @@#include "kfd_priv.h" #include "kfd_svm.h" #include "kfd_migrate.h" +#include "kfd_smi_events.h" #ifdef dev_fmt#undef dev_fmt
@@ -1596,7 +1597,7 @@ static int svm_range_validate_and_map(struct mm_struct *mm,svm_range_unreserve_bos (&ctx);if (!r) - prange->validate_timestamp =ktime_to_us (ktime_get ()); + prange->validate_timestamp =ktime_get_boottime_ns ();return r; }
@@ -2665,11 +2666,12 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,struct svm_range_list *svms;struct svm_range *prange;struct kfd_process *p; -uint64_t timestamp; +uint64_t timestamp =ktime_get_boottime_ns ();int32_t best_loc;int32_t gpuidx =MAX_GPU_INSTANCE ;bool write_locked = false;struct vm_area_struct *vma; +bool migration = false;int r =0 ;if (!KFD_IS_SVM_API_SUPPORTED (adev->kfd.dev)) {
@@ -2745,9 +2747,9 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, goto out_unlock_range; } - timestamp =ktime_to_us (ktime_get ()) - prange->validate_timestamp;/* skip duplicate vm fault on different pages of same range */ -if (timestamp <AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING ) { +if (div_u64 (timestamp - prange->validate_timestamp,1000000 ) < +AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING ) {pr_debug ("svms 0x%p [0x%lx %lx] already restored\n" , svms, prange->start, prange->last); r =0 ;
@@ -2783,7 +2785,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, svms, prange->start, prange->last, best_loc, prange->actual_loc); +kfd_smi_event_page_fault_start (adev->kfd.dev, p->lead_thread->pid, addr, + write_fault, timestamp); +if (prange->actual_loc != best_loc) { + migration = true;if (best_loc) { r =svm_migrate_to_vram (prange, best_loc, mm);if (r) {
@@ -2812,6 +2818,9 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,pr_debug ("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n" , r, svms, prange->start, prange->last); +kfd_smi_event_page_fault_end (adev->kfd.dev, p->lead_thread->pid, addr, + migration, timestamp); + out_unlock_range:mutex_unlock (&prange->migrate_mutex); out_unlock_svms:
Signature · 1 line
2.17.1
PATCH v2 7/8 drm/amdkfd: Add unmap from GPU SMI event
In reply to Philip YangSVM range unmapped from GPUs when range is unmapped from CPU, or with xnack on when range is evicted or migrated.
--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 18 ++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 3 +++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 14 +++++++++----- 3 files changed, 30 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +18 -0
@@ -377,6 +377,24 @@ void kfd_smi_event_queue_eviction_restore(struct kfd_dev *dev, pid_t pid, fifo_in, len); } +void kfd_smi_event_unmap_from_gpu (struct kfd_dev *dev,pid_t pid, +unsigned long address,unsigned long last, +uint32_t trigger) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%lx(%lx) %x %d\n" , +KFD_SMI_EVENT_UNMAP_FROM_GPU ,ktime_get_boottime_ns (), + pid, address, last - address +1 , dev->id, trigger); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_UNMAP_FROM_GPU , fifo_in, + len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +3 -0
@@ -43,4 +43,7 @@ void kfd_smi_event_queue_eviction(struct kfd_dev *dev, pid_t pid, uint32_t trigger);void kfd_smi_event_queue_eviction_restore (struct kfd_dev *dev,pid_t pid,bool rescheduled); +void kfd_smi_event_unmap_from_gpu (struct kfd_dev *dev,pid_t pid, +unsigned long address,unsigned long last, +uint32_t trigger);#endif
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +9 -5
@@ -1191,7 +1191,7 @@ svm_range_unmap_from_gpu(struct amdgpu_device *adev, struct amdgpu_vm *vm,static int svm_range_unmap_from_gpus (struct svm_range *prange,unsigned long start, -unsigned long last) +unsigned long last,uint32_t trigger) {DECLARE_BITMAP (bitmap,MAX_GPU_INSTANCE );struct kfd_process_device *pdd;
@@ -1212,6 +1212,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,return -EINVAL ; } +kfd_smi_event_unmap_from_gpu (pdd->dev, p->lead_thread->pid, + start, last, trigger); + r =svm_range_unmap_from_gpu (pdd->dev->adev,drm_priv_to_vm (pdd->drm_priv), start, last, &fence);
@@ -1795,13 +1798,13 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm, s =max (start, pchild->start); l =min (last, pchild->last);if (l >= s) -svm_range_unmap_from_gpus (pchild, s, l); +svm_range_unmap_from_gpus (pchild, s, l, trigger);mutex_unlock (&pchild->lock); } s =max (start, prange->start); l =min (last, prange->last);if (l >= s) -svm_range_unmap_from_gpus (prange, s, l); +svm_range_unmap_from_gpus (prange, s, l, trigger); }return r;
@@ -2214,6 +2217,7 @@ static void svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, unsigned long start, unsigned long last) { + uint32_t trigger = UNMAP_FROM_CPU;struct svm_range_list *svms;struct svm_range *pchild;struct kfd_process *p;
@@ -2241,14 +2245,14 @@ svm_range_unmap_from_cpu(struct mm_struct *mm, struct svm_range *prange, s =max (start, pchild->start); l =min (last, pchild->last);if (l >= s) -svm_range_unmap_from_gpus (pchild, s, l); +svm_range_unmap_from_gpus (pchild, s, l, trigger);svm_range_unmap_split (mm, prange, pchild, start, last);mutex_unlock (&pchild->lock); } s =max (start, prange->start); l =min (last, prange->last);if (l >= s) -svm_range_unmap_from_gpus (prange, s, l); +svm_range_unmap_from_gpus (prange, s, l, trigger);svm_range_unmap_split (mm, prange, prange, start, last);if (unmap_parent)
Signature · 1 line
2.17.1
PATCH v2 5/8 drm/amdkfd: add migration SMI event
In reply to Philip YangAfter migration is finished, output timestamp when migration starts, duration of migration, svm range address and size, GPU id of migration source and destination and svm range attributes,
Migration trigger could be prefetch, CPU or GPU page fault and TTM eviction.
--- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 67 ++++++++++++++------- drivers/gpu/drm/amd/amdkfd/kfd_migrate.h | 5 +- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 29 +++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 16 +++-- 5 files changed, 91 insertions(+), 31 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +44 -23
@@ -32,6 +32,7 @@#include "kfd_priv.h" #include "kfd_svm.h" #include "kfd_migrate.h" +#include "kfd_smi_events.h" #ifdef dev_fmt#undef dev_fmt
@@ -402,10 +403,11 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,static long svm_migrate_vma_to_vram (struct amdgpu_device *adev,struct svm_range *prange,struct vm_area_struct *vma,uint64_t start, -uint64_t end) +uint64_t end,uint32_t trigger) {uint64_t npages = (end - start) >>PAGE_SHIFT ; -struct kfd_process_device *pdd; +struct kfd_process_device *pdd =NULL ; +uint64_t timestamp =ktime_get_boottime_ns ();struct dma_fence *mfence =NULL ;struct migrate_vma migrate;unsigned long cpages =0 ;
@@ -431,6 +433,10 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, migrate.dst = migrate.src + npages; scratch = (dma_addr_t *)(migrate.dst + npages); + pdd =svm_range_get_pdd_by_adev (prange, adev); +if (!pdd) + goto out_free; + r =migrate_vma_setup (&migrate);if (r) { dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r,
@@ -459,6 +465,11 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,svm_migrate_copy_done (adev, mfence);migrate_vma_finalize (&migrate); +kfd_smi_event_migration (adev->kfd.dev, pdd->process->pasid, + start >>PAGE_SHIFT , end >>PAGE_SHIFT , +0 , adev->kfd.dev->id, prange->prefetch_loc, + prange->preferred_loc, trigger, timestamp); +svm_range_dma_unmap (adev->dev, scratch,0 , npages);svm_range_free_dma_mappings (prange);
@@ -466,10 +477,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,kvfree (buf); out:if (!r && cpages) { - pdd =svm_range_get_pdd_by_adev (prange, adev); -if (pdd) -WRITE_ONCE (pdd->page_in, pdd->page_in + cpages); - +WRITE_ONCE (pdd->page_in, pdd->page_in + cpages);return cpages; }return r;
@@ -480,6 +488,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,* @prange: range structure * @best_loc: the device to migrate to * @mm: the process mm structure +* @trigger: reason of migration * * Context: Process context, caller hold mmap read lock, svms lock, prange lock *
@@ -488,7 +497,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,*/ static int svm_migrate_ram_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm) +struct mm_struct *mm,uint32_t trigger) {unsigned long addr, start, end;struct vm_area_struct *vma;
@@ -525,7 +534,7 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,break ; next =min (vma->vm_end, end); - r =svm_migrate_vma_to_vram (adev, prange, vma, addr, next); + r =svm_migrate_vma_to_vram (adev, prange, vma, addr, next, trigger);if (r <0 ) {pr_debug ("failed %ld to migrate\n" , r);break ;
@@ -641,12 +650,14 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,static long svm_migrate_vma_to_ram (struct amdgpu_device *adev,struct svm_range *prange, -struct vm_area_struct *vma,uint64_t start,uint64_t end) +struct vm_area_struct *vma,uint64_t start,uint64_t end, +uint32_t trigger) {uint64_t npages = (end - start) >>PAGE_SHIFT ; +uint64_t timestamp =ktime_get_boottime_ns ();unsigned long upages = npages;unsigned long cpages =0 ; -struct kfd_process_device *pdd; +struct kfd_process_device *pdd =NULL ;struct dma_fence *mfence =NULL ;struct migrate_vma migrate;dma_addr_t *scratch;
@@ -671,6 +682,10 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, migrate.dst = migrate.src + npages; scratch = (dma_addr_t *)(migrate.dst + npages); + pdd =svm_range_get_pdd_by_adev (prange, adev); +if (!pdd) + goto out_free; + r =migrate_vma_setup (&migrate);if (r) { dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r,
@@ -701,16 +716,19 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,svm_migrate_copy_done (adev, mfence);migrate_vma_finalize (&migrate); + +kfd_smi_event_migration (adev->kfd.dev, pdd->process->pasid, + start >>PAGE_SHIFT , end >>PAGE_SHIFT , + adev->kfd.dev->id,0 , prange->prefetch_loc, + prange->preferred_loc, trigger, timestamp); +svm_range_dma_unmap (adev->dev, scratch,0 , npages); out_free:kvfree (buf); out:if (!r && cpages) { - pdd =svm_range_get_pdd_by_adev (prange, adev); -if (pdd) -WRITE_ONCE (pdd->page_out, pdd->page_out + cpages); - +WRITE_ONCE (pdd->page_out, pdd->page_out + cpages);return upages; }return r ? r : upages;
@@ -720,13 +738,15 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,* svm_migrate_vram_to_ram - migrate svm range from device to system * @prange: range structure * @mm: process mm, use current->mm if NULL +* @trigger: reason of migration * * Context: Process context, caller hold mmap read lock, svms lock, prange lock * * Return: * 0 - OK, otherwise error code */ -int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm) +int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm, +uint32_t trigger) {struct amdgpu_device *adev;struct vm_area_struct *vma;
@@ -764,7 +784,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)break ; next =min (vma->vm_end, end); - r =svm_migrate_vma_to_ram (adev, prange, vma, addr, next); + r =svm_migrate_vma_to_ram (adev, prange, vma, addr, next, trigger);if (r <0 ) {pr_debug ("failed %ld to migrate\n" , r);break ;
@@ -787,6 +807,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)* @prange: range structure * @best_loc: the device to migrate to * @mm: process mm, use current->mm if NULL +* @trigger: reason of migration * * Context: Process context, caller hold mmap read lock, svms lock, prange lock *
@@ -795,7 +816,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)*/ static int svm_migrate_vram_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm) +struct mm_struct *mm,uint32_t trigger) {int r, retries =3 ;
@@ -807,7 +828,7 @@ svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,pr_debug ("from gpu 0x%x to gpu 0x%x\n" , prange->actual_loc, best_loc);do { - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm, trigger);if (r)return r; }while (prange->actual_loc && --retries);
@@ -815,17 +836,17 @@ svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,if (prange->actual_loc)return -EDEADLK ; -return svm_migrate_ram_to_vram (prange, best_loc, mm); +return svm_migrate_ram_to_vram (prange, best_loc, mm, trigger); }int svm_migrate_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm) +struct mm_struct *mm,uint32_t trigger) {if (!prange->actual_loc) -return svm_migrate_ram_to_vram (prange, best_loc, mm); +return svm_migrate_ram_to_vram (prange, best_loc, mm, trigger);else -return svm_migrate_vram_to_vram (prange, best_loc, mm); +return svm_migrate_vram_to_vram (prange, best_loc, mm, trigger); }
@@ -894,7 +915,7 @@ static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) goto out_unlock_prange; } - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm,MIGRATION_TRIGGER_PAGEFAULT_CPU );if (r)pr_debug ("failed %d migrate 0x%p [0x%lx 0x%lx] to ram\n" , r, prange, prange->start, prange->last);
drivers/gpu/drm/amd/amdkfd/kfd_migrate.h +3 -2
@@ -41,8 +41,9 @@ enum MIGRATION_COPY_DIR { };int svm_migrate_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm); -int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm); +struct mm_struct *mm,uint32_t trigger); +int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm, +uint32_t trigger);unsigned long svm_migrate_addr_to_pfn (struct amdgpu_device *adev,unsigned long addr);
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +29 -0
@@ -314,6 +314,35 @@ void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, len); } +void kfd_smi_event_migration (struct kfd_dev *dev,uint16_t pasid, +unsigned long start,unsigned long end, +uint32_t from,uint32_t to, +uint32_t prefetch_loc,uint32_t preferred_loc, +uint32_t trigger,uint64_t ts) +{ +struct amdgpu_task_info task_info = {0 }; +char fifo_in[128 ]; +uint64_t duration; +pid_t pid; +int len; + +if (!kfd_smi_event_duration (dev, ts, &duration)) +return ; + +amdgpu_vm_get_task_info (dev->adev, pasid, &task_info); +if (!task_info.pid) +return ; + pid = task_info.pid; + + len =snprintf (fifo_in,sizeof (fifo_in), +"%x %lld(%lld) -%d @%lx(%lx) %x->%x %x:%x %d\n" , +KFD_SMI_EVENT_MIGRATION , ts, duration, pid, start, + end - start, from, to, prefetch_loc, preferred_loc, + trigger); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_MIGRATION , fifo_in, len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +5 -0
@@ -34,4 +34,9 @@ void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid,void kfd_smi_event_page_fault_end (struct kfd_dev *dev,pid_t pid,unsigned long address,bool migration,uint64_t ts); +void kfd_smi_event_migration (struct kfd_dev *dev,uint16_t pasid, +unsigned long start,unsigned long end, +uint32_t from,uint32_t to, +uint32_t prefetch_loc,uint32_t preferred_loc, +uint32_t trigger,uint64_t ts);#endif
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +10 -6
@@ -2791,7 +2791,8 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,if (prange->actual_loc != best_loc) { migration = true;if (best_loc) { - r =svm_migrate_to_vram (prange, best_loc, mm); + r =svm_migrate_to_vram (prange, best_loc, mm, +MIGRATION_TRIGGER_PAGEFAULT );if (r) {pr_debug ("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n" , r, addr);
@@ -2799,12 +2800,14 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,* VRAM failed */ if (prange->actual_loc) - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm, +MIGRATION_TRIGGER_PAGEFAULT );else r =0 ; } }else { - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm, +MIGRATION_TRIGGER_PAGEFAULT ); }if (r) { pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",
@@ -3126,12 +3129,12 @@ svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,return 0 ;if (!best_loc) { - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm,MIGRATION_TRIGGER_PREFETCH ); *migrated = !r;return r; } - r =svm_migrate_to_vram (prange, best_loc, mm); + r =svm_migrate_to_vram (prange, best_loc, mm,MIGRATION_TRIGGER_PREFETCH ); *migrated = !r;return r;
@@ -3188,7 +3191,8 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work)mutex_lock (&prange->migrate_mutex);do {svm_migrate_vram_to_ram (prange, - svm_bo->eviction_fence->mm); + svm_bo->eviction_fence->mm, +MIGRATION_TRIGGER_TTM_EVICTION ); }while (prange->actual_loc && --retries);WARN (prange->actual_loc,"Migration failed during eviction" );
Signature · 1 line
2.17.1
PATCH v2 6/8 drm/amdkfd: Add user queue eviction restore SMI event
In reply to Philip YangOutput user queue eviction and restore event. User queue eviction may be triggered by migration, MMU notifier, TTM eviction or device suspend.
User queue restore may be rescheduled if eviction happens again while restore.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 7 +++- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 11 ++++-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 37 +++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 34 +++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 4 ++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 16 ++++++-- 8 files changed, 101 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +6 -1
@@ -309,6 +309,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,*/ void amdgpu_amdkfd_release_notify (struct amdgpu_bo *bo);void amdgpu_amdkfd_reserve_system_mem (uint64_t size); +void kfd_process_smi_event_restore_rescheduled (struct mm_struct *mm);#else static inline void amdgpu_amdkfd_gpuvm_init_mem_limits (void )
@@ -325,9 +326,13 @@ static inlinevoid amdgpu_amdkfd_release_notify (struct amdgpu_bo *bo) { } + +static inline void kfd_process_smi_event_restore_rescheduled (struct mm_struct *mm) +{ +}#endif /* KGD2KFD callbacks */ -int kgd2kfd_quiesce_mm (struct mm_struct *mm); +int kgd2kfd_quiesce_mm (struct mm_struct *mm,uint32_t trigger);int kgd2kfd_resume_mm (struct mm_struct *mm);int kgd2kfd_schedule_evict_and_restore_process (struct mm_struct *mm,struct dma_fence *fence);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7 -4
@@ -2066,7 +2066,7 @@ int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, evicted_bos =atomic_inc_return (&process_info->evicted_bos);if (evicted_bos ==1 ) {/* First eviction, stop the queues */ - r =kgd2kfd_quiesce_mm (mm); + r =kgd2kfd_quiesce_mm (mm,USERPTR_EVICTION );if (r)pr_err ("Failed to quiesce KFD\n" ); schedule_delayed_work(&process_info->restore_userptr_work,
@@ -2340,13 +2340,16 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) unlock_out:mutex_unlock (&process_info->lock); -mmput (mm); -put_task_struct (usertask);/* If validation failed, reschedule another attempt */ -if (evicted_bos) +if (evicted_bos) {schedule_delayed_work (&process_info->restore_userptr_work,msecs_to_jiffies (AMDGPU_USERPTR_RESTORE_DELAY_MS )); + +kfd_process_smi_event_restore_rescheduled (mm); + } +mmput (mm); +put_task_struct (usertask); }/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given
drivers/gpu/drm/amd/amdkfd/kfd_device.c +2 -2
@@ -783,7 +783,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)spin_unlock_irqrestore (&kfd->interrupt_lock, flags); } -int kgd2kfd_quiesce_mm (struct mm_struct *mm) +int kgd2kfd_quiesce_mm (struct mm_struct *mm,uint32_t trigger) {struct kfd_process *p;int r;
@@ -797,7 +797,7 @@ int kgd2kfd_quiesce_mm(struct mm_struct *mm)return -ESRCH ;WARN (debug_evictions,"Evicting pid %d" , p->lead_thread->pid); - r =kfd_process_evict_queues (p); + r =kfd_process_evict_queues (p, trigger);kfd_unref_process (p);return r;
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -1
@@ -906,7 +906,7 @@ static inline struct kfd_process_device *kfd_process_device_from_gpuidx( }void kfd_unref_process (struct kfd_process *p); -int kfd_process_evict_queues (struct kfd_process *p); +int kfd_process_evict_queues (struct kfd_process *p,uint32_t trigger);int kfd_process_restore_queues (struct kfd_process *p);void kfd_suspend_all_processes (void );int kfd_resume_all_processes (void );
drivers/gpu/drm/amd/amdkfd/kfd_process.c +34 -3
@@ -43,6 +43,7 @@ struct mm_struct;#include "kfd_dbgmgr.h" #include "kfd_iommu.h" #include "kfd_svm.h" +#include "kfd_smi_events.h" /* * List of struct kfd_process (field kfd_process).
@@ -1701,7 +1702,7 @@ struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)* Eviction is reference-counted per process-device. This means multiple * evictions from different sources can be nested safely. */ -int kfd_process_evict_queues (struct kfd_process *p) +int kfd_process_evict_queues (struct kfd_process *p,uint32_t trigger) {int r =0 ;int i;
@@ -1710,6 +1711,9 @@ int kfd_process_evict_queues(struct kfd_process *p)for (i =0 ; i < p->n_pdds; i++) {struct kfd_process_device *pdd = p->pdds[i]; +kfd_smi_event_queue_eviction (pdd->dev, p->lead_thread->pid, + trigger); + r = pdd->dev->dqm->ops.evict_process_queues (pdd->dev->dqm, &pdd->qpd);/* evict return -EIO if HWS is hang or asic is resetting, in this case
@@ -1734,6 +1738,11 @@ int kfd_process_evict_queues(struct kfd_process *p)if (n_evicted ==0 )break ; + +kfd_smi_event_queue_eviction_restore (pdd->dev, + p->lead_thread->pid, + false); +if (pdd->dev->dqm->ops.restore_process_queues (pdd->dev->dqm, &pdd->qpd))pr_err ("Failed to restore queues\n" );
@@ -1753,6 +1762,10 @@ int kfd_process_restore_queues(struct kfd_process *p)for (i =0 ; i < p->n_pdds; i++) {struct kfd_process_device *pdd = p->pdds[i]; +kfd_smi_event_queue_eviction_restore (pdd->dev, + p->lead_thread->pid, + false); + r = pdd->dev->dqm->ops.restore_process_queues (pdd->dev->dqm, &pdd->qpd);if (r) {
@@ -1765,6 +1778,24 @@ int kfd_process_restore_queues(struct kfd_process *p)return ret; } +void kfd_process_smi_event_restore_rescheduled (struct mm_struct *mm) +{ +struct kfd_process *p; +int i; + + p =kfd_lookup_process_by_mm (mm); +if (!p) +return ; + +for (i =0 ; i < p->n_pdds; i++) { +struct kfd_process_device *pdd = p->pdds[i]; + +kfd_smi_event_queue_eviction_restore (pdd->dev, + p->lead_thread->pid, true); + } +kfd_unref_process (p); +} +int kfd_process_gpuidx_from_gpuid (struct kfd_process *p,uint32_t gpu_id) {int i;
@@ -1814,7 +1845,7 @@ static void evict_process_worker(struct work_struct *work)flush_delayed_work (&p->restore_work);pr_debug ("Started evicting pasid 0x%x\n" , p->pasid); - ret =kfd_process_evict_queues (p); + ret =kfd_process_evict_queues (p,TTM_EVICTION );if (!ret) {dma_fence_signal (p->ef);dma_fence_put (p->ef);
@@ -1881,7 +1912,7 @@ void kfd_suspend_all_processes(void)cancel_delayed_work_sync (&p->eviction_work);cancel_delayed_work_sync (&p->restore_work); -if (kfd_process_evict_queues (p)) +if (kfd_process_evict_queues (p,SUSPEND_EVICTION ))pr_err ("Failed to suspend process 0x%x\n" , p->pasid);dma_fence_signal (p->ef);dma_fence_put (p->ef);
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +34 -0
@@ -343,6 +343,40 @@ void kfd_smi_event_migration(struct kfd_dev *dev, uint16_t pasid,add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_MIGRATION , fifo_in, len); } +void kfd_smi_event_queue_eviction (struct kfd_dev *dev,pid_t pid, +uint32_t trigger) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%x %d\n" , +KFD_SMI_EVENT_QUEUE_EVICTION ,ktime_get_boottime_ns (), + pid, dev->id, trigger); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_QUEUE_EVICTION , fifo_in, len); +} + +void kfd_smi_event_queue_eviction_restore (struct kfd_dev *dev,pid_t pid, +bool rescheduled) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%x %c\n" , +KFD_SMI_EVENT_QUEUE_EVICTION_RESTORE , +ktime_get_boottime_ns (), pid, dev->id, + rescheduled ?'r' :' ' ); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_QUEUE_EVICTION_RESTORE , + fifo_in, len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +4 -0
@@ -39,4 +39,8 @@ void kfd_smi_event_migration(struct kfd_dev *dev, uint16_t pasid, uint32_t from, uint32_t to, uint32_t prefetch_loc, uint32_t preferred_loc, uint32_t trigger, uint64_t ts); +void kfd_smi_event_queue_eviction (struct kfd_dev *dev,pid_t pid, +uint32_t trigger); +void kfd_smi_event_queue_eviction_restore (struct kfd_dev *dev,pid_t pid, +bool rescheduled);#endif
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +13 -3
@@ -1712,7 +1712,10 @@ static void svm_range_restore_work(struct work_struct *work)pr_debug ("reschedule to restore svm range\n" );schedule_delayed_work (&svms->restore_work,msecs_to_jiffies (AMDGPU_SVM_RANGE_RESTORE_DELAY_MS )); + +kfd_process_smi_event_restore_rescheduled (mm); } +mmput (mm); }/**
@@ -1732,15 +1735,22 @@ static void svm_range_restore_work(struct work_struct *work)*/ static int svm_range_evict (struct svm_range *prange,struct mm_struct *mm, -unsigned long start,unsigned long last) +unsigned long start,unsigned long last, +enum mmu_notifier_event event) {struct svm_range_list *svms = prange->svms;struct svm_range *pchild;struct kfd_process *p; +uint32_t trigger;int r =0 ; p = container_of(svms, struct kfd_process, svms); +if (event ==MMU_NOTIFY_MIGRATE ) + trigger =SVM_RANGE_MIGRATION ; +else + trigger =SVM_RANGE_EVICTION ; +pr_debug ("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n" , svms, prange->start, prange->last, start, last);
@@ -1768,7 +1778,7 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm, prange->svms, prange->start, prange->last);/* First eviction, stop the queues */ - r =kgd2kfd_quiesce_mm (mm); + r =kgd2kfd_quiesce_mm (mm, trigger);if (r)pr_debug ("failed to quiesce KFD\n" );
@@ -2303,7 +2313,7 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,svm_range_unmap_from_cpu (mni->mm, prange, start, last);break ;default : -svm_range_evict (prange, mni->mm, start, last); +svm_range_evict (prange, mni->mm, start, last, range->event);break ; }
Signature · 1 line
2.17.1
PATCH v2 3/8 drm/amdkfd: Enable per process SMI event
In reply to Philip YangProcess receive event from same process by default. Add a flag to be able to receive event from all processes, this requires super user permission.
Event with pid 0 send to all processes, to keep the default behavior of existing SMI events.
--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 29 ++++++++++++++++----- 1 file changed, 22 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +22 -7
@@ -37,6 +37,8 @@ struct kfd_smi_client {uint64_t events;struct kfd_dev *dev;spinlock_t lock; +pid_t pid; +bool suser; };#define MAX_KFIFO_SIZE 1024
@@ -150,16 +152,27 @@ static int kfd_smi_ev_release(struct inode *inode, struct file *filep)return 0 ; } -static void add_event_to_kfifo (struct kfd_dev *dev,unsigned int smi_event, -char *event_msg,int len) +static bool kfd_smi_ev_enabled (pid_t pid,struct kfd_smi_client *client, +unsigned int event) +{ +uint64_t all =KFD_SMI_EVENT_MASK_FROM_INDEX (KFD_SMI_EVENT_ALL_PROCESS ); +uint64_t events =READ_ONCE (client->events); + +if (pid && client->pid != pid && !(client->suser && (events & all))) +return false; + +return events &KFD_SMI_EVENT_MASK_FROM_INDEX (event); +} + +static void add_event_to_kfifo (pid_t pid,struct kfd_dev *dev, +unsigned int smi_event,char *event_msg,int len) {struct kfd_smi_client *client;rcu_read_lock ();list_for_each_entry_rcu (client, &dev->smi_clients, list) { -if (!(READ_ONCE (client->events) & -KFD_SMI_EVENT_MASK_FROM_INDEX (smi_event))) +if (!kfd_smi_ev_enabled (pid, client, smi_event))continue ;spin_lock (&client->lock);if (kfifo_avail (&client->fifo) >= len) {
@@ -202,7 +215,7 @@ void kfd_smi_event_update_gpu_reset(struct kfd_dev *dev, bool post_reset) len =snprintf (fifo_in,sizeof (fifo_in),"%x %x\n" , event, dev->reset_seq_num); -add_event_to_kfifo (dev, event, fifo_in, len); +add_event_to_kfifo (0 , dev, event, fifo_in, len); } void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev,
@@ -225,7 +238,7 @@ void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev,KFD_SMI_EVENT_THERMAL_THROTTLE , throttle_bitmask, amdgpu_dpm_get_thermal_throttling_counter(dev->adev)); -add_event_to_kfifo (dev,KFD_SMI_EVENT_THERMAL_THROTTLE , fifo_in, len); +add_event_to_kfifo (0 , dev,KFD_SMI_EVENT_THERMAL_THROTTLE , fifo_in, len); }void kfd_smi_event_update_vmfault (struct kfd_dev *dev,uint16_t pasid)
@@ -250,7 +263,7 @@ void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid) len =snprintf (fifo_in,sizeof (fifo_in),"%x %x:%s\n" ,KFD_SMI_EVENT_VMFAULT , task_info.pid, task_info.task_name); -add_event_to_kfifo (dev,KFD_SMI_EVENT_VMFAULT , fifo_in, len); +add_event_to_kfifo (0 , dev,KFD_SMI_EVENT_VMFAULT , fifo_in, len); }int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd)
@@ -282,6 +295,8 @@ int kfd_smi_event_open(struct kfd_dev *dev, uint32_t *fd)spin_lock_init (&client->lock); client->events =0 ; client->dev = dev; + client->pid = current->pid; + client->suser =capable (CAP_SYS_ADMIN );spin_lock (&dev->smi_lock);list_add_rcu (&client->list, &dev->smi_clients);
Signature · 1 line
2.17.1
Deucher, Alexander
maintainer[Public]
Please provide a link to the proposed userspace branch that makes use of this.
Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Philip Yang <[email protected]>
Sent: Thursday, January 20, 2022 6:13 PM
To: [email protected] <[email protected]>
Cc: Yang, Philip <[email protected]>; Kuehling, Felix <[email protected]>
Subject: [PATCH v2 0/8] HMM profiler interface
The ROCm profiler would expose the data from KFD profiling APIs to application developers to tune the applications based on how the address range attributes affect the behavior and performance.
Per process event log use the existing SMI (system management interface) event API. Each event log is one line of text with the event specific information.
v2:
* Keep existing events behaviour
* Use ktime_get_boottime_ns() as timestamp to correlate with other APIs
* Use compact message layout, stick with existing message convention
* Add unmap from GPU event
Philip Yang (8):
drm/amdkfd: Correct SMI event read size drm/amdkfd: Add KFD SMI event IDs and triggers drm/amdkfd: Enable per process SMI event drm/amdkfd: Add GPU recoverable fault SMI event drm/amdkfd: add migration SMI event
drm/amdkfd: Add user queue eviction restore SMI event drm/amdkfd: Add unmap from GPU SMI event drm/amdkfd: Bump KFD API version for SMI profiling event
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 7 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 11 +- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 67 ++++--- drivers/gpu/drm/amd/amdkfd/kfd_migrate.h | 5 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 37 +++- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 163 +++++++++++++++++- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 19 +- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 63 +++++-- include/uapi/linux/kfd_ioctl.h | 30 +++- 11 files changed, 343 insertions(+), 65 deletions(-)
--
2.17.1
Re: [PATCH v2 1/8] drm/amdkfd: Correct SMI event read size
In reply to Philip YangAm 2022-01-20 um 18:13 schrieb Philip Yang:
Quoted 5 lines
sizeof(buf) is 8 bytes because it is defined as unsigned char *buf, each SMI event read only copy max 8 bytes to user buffer. Correct this by using the buf allocate size.
Signed-off-by Philip Yang <[email protected]>
Quoted 27 lines
--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +3
-2@@ -81,7 +81,8 @@ static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user,struct kfd_smi_client *client = filep->private_data;unsigned char *buf; - buf =kmalloc_array (MAX_KFIFO_SIZE ,sizeof (*buf),GFP_KERNEL ); + size =min_t (size_t, size,MAX_KFIFO_SIZE ); + buf =kmalloc (size,GFP_KERNEL );if (!buf)return -ENOMEM ;@@ -95,7 +96,7 @@ static ssize_t kfd_smi_ev_read(struct file *filep, char __user *user, ret = -EAGAIN ; goto ret_err; } - to_copy =min3 (size,sizeof (buf), to_copy); + to_copy =min (size, to_copy); ret =kfifo_out (&client->fifo, buf, to_copy);spin_unlock (&client->lock);if (ret <=0 ) {
Felix Kuehling
maintainerRe: [PATCH v2 2/8] drm/amdkfd: Add KFD SMI event IDs and triggers
In reply to Philip YangAm 2022-01-20 um 18:13 schrieb Philip Yang:
Quoted 28 lines
Define new system management interface event IDs, migration triggers and user queue eviction triggers, those will be implemented in the following patches.
Signed-off-by Philip Yang <[email protected]>--- include/uapi/linux/kfd_ioctl.h | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+)include/uapi/linux/kfd_ioctl.h +11
-0@@ -459,10 +459,37 @@ enum kfd_smi_event {KFD_SMI_EVENT_THERMAL_THROTTLE =2 ,KFD_SMI_EVENT_GPU_PRE_RESET =3 ,KFD_SMI_EVENT_GPU_POST_RESET =4 , +KFD_SMI_EVENT_MIGRATION =5 , +KFD_SMI_EVENT_PAGE_FAULT_START =6 , +KFD_SMI_EVENT_PAGE_FAULT_END =7 , +KFD_SMI_EVENT_QUEUE_EVICTION =8 , +KFD_SMI_EVENT_QUEUE_EVICTION_RESTORE =9 , +KFD_SMI_EVENT_UNMAP_FROM_GPU =10 , + +/* +* max event number, as a flag bit to get events from all processes, +* this requires super user permission, otherwise will not be able to +* receive event from any process. Without this flag to receive events
include/uapi/linux/kfd_ioctl.h enum kfd_smi_event { · line 482
Quoted 10 lines
include/uapi/linux/kfd_ioctl.h +6
-0+* from same process. +*/ +KFD_SMI_EVENT_ALL_PROCESS =64 };#define KFD_SMI_EVENT_MASK_FROM_INDEX (i) (1ULL << ((i) - 1)) +enum KFD_MIGRATION_QUEUE_EVICTION_UNMAP_EVENT_TRIGGER { +MIGRATION_TRIGGER_PREFETCH =1 , +MIGRATION_TRIGGER_PAGEFAULT ,
Call this KFD_MIGRATION_TRIGGER_PAGEFAULT_GPU for consistency. Add the KFD_ prefix to all the migration triggers.
include/uapi/linux/kfd_ioctl.h enum kfd_smi_event { · line 490
Quoted 8 lines
include/uapi/linux/kfd_ioctl.h +8
-0+MIGRATION_TRIGGER_PAGEFAULT_CPU , +MIGRATION_TRIGGER_TTM_EVICTION , +SVM_RANGE_EVICTION , +SVM_RANGE_MIGRATION , +USERPTR_EVICTION , +TTM_EVICTION , +UNMAP_FROM_CPU , +SUSPEND_EVICTION
Migration triggers and queue eviction triggers should be separate enums. The eviction triggers should have a consistent prefix: KFD_QUEUE_EVICTION_TRIGGER...
I'm not sure the distinction between SVM_RANGE_MIGRATION and SVM_RANGE_EVICTION is useful. It should be clear from the context, since migrations log their own events. If anything, the term SVM_RANGE_EVICTION is misleading. It's really any MMU notifier that's not MMU_NOTIFY_MIGRATE or MMU_NOTIFY_UNMAP.
I'd suggest the following names:
KFD_QUEUE_EVICTION_TRIGGER_SVMKFD_QUEUE_EVICTION_TRIGGER_USERPTRKFD_QUEUE_EVICTION_TRIGGER_TTMKFD_QUEUE_EVICTION_TRIGGER_SUSPEND
Then there are triggers for unmapping from GPU, which seems to have some overlap with queue eviction in svm_range_evict. But that becomes a bit easier if you remove the distinction between SVM_RANGE_MIGRATION and SVM_RANGE_EVICTION. Then you are left with these two unmap triggers:
KFD_SVM_UNMAP_TRIGGER_MMU_NOTIFIERKFD_SVM_UNMAP_TRIGGER_UNMAP_FROM_CPU
Regards,
Felix
Quoted 5 lines
include/uapi/linux/kfd_ioctl.h +2
-0+}; +struct kfd_ioctl_smi_events_args {__u32 gpuid;/* to KFD */ __u32 anon_fd;/* from KFD */
Re: [PATCH v2 3/8] drm/amdkfd: Enable per process SMI event
In reply to Philip YangAm 2022-01-20 um 18:13 schrieb Philip Yang:
Quoted 8 lines
Process receive event from same process by default. Add a flag to be able to receive event from all processes, this requires super user permission.
Event with pid 0 send to all processes, to keep the default behavior of existing SMI events.
Signed-off-by Philip Yang <[email protected]>
Quoted 85 lines
--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 29 ++++++++++++++++----- 1 file changed, 22 insertions(+), 7 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +22
-7@@ -37,6 +37,8 @@ struct kfd_smi_client {uint64_t events;struct kfd_dev *dev;spinlock_t lock; +pid_t pid; +bool suser; };#define MAX_KFIFO_SIZE 1024@@ -150,16 +152,27 @@ static int kfd_smi_ev_release(struct inode *inode, struct file *filep)return 0 ; } -static void add_event_to_kfifo (struct kfd_dev *dev,unsigned int smi_event, -char *event_msg,int len) +static bool kfd_smi_ev_enabled (pid_t pid,struct kfd_smi_client *client, +unsigned int event) +{ +uint64_t all =KFD_SMI_EVENT_MASK_FROM_INDEX (KFD_SMI_EVENT_ALL_PROCESS ); +uint64_t events =READ_ONCE (client->events); + +if (pid && client->pid != pid && !(client->suser && (events & all))) +return false; + +return events &KFD_SMI_EVENT_MASK_FROM_INDEX (event); +} + +static void add_event_to_kfifo (pid_t pid,struct kfd_dev *dev, +unsigned int smi_event,char *event_msg,int len) {struct kfd_smi_client *client;rcu_read_lock ();list_for_each_entry_rcu (client, &dev->smi_clients, list) { -if (!(READ_ONCE (client->events) & -KFD_SMI_EVENT_MASK_FROM_INDEX (smi_event))) +if (!kfd_smi_ev_enabled (pid, client, smi_event))continue ;spin_lock (&client->lock);if (kfifo_avail (&client->fifo) >= len) {@@ -202,7 +215,7 @@ void kfd_smi_event_update_gpu_reset(struct kfd_dev *dev, bool post_reset) len =snprintf (fifo_in,sizeof (fifo_in),"%x %x\n" , event, dev->reset_seq_num); -add_event_to_kfifo (dev, event, fifo_in, len); +add_event_to_kfifo (0 , dev, event, fifo_in, len); } void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev,@@ -225,7 +238,7 @@ void kfd_smi_event_update_thermal_throttling(struct kfd_dev *dev,KFD_SMI_EVENT_THERMAL_THROTTLE , throttle_bitmask, amdgpu_dpm_get_thermal_throttling_counter(dev->adev)); -add_event_to_kfifo (dev,KFD_SMI_EVENT_THERMAL_THROTTLE , fifo_in, len); +add_event_to_kfifo (0 , dev,KFD_SMI_EVENT_THERMAL_THROTTLE , fifo_in, len); }void kfd_smi_event_update_vmfault (struct kfd_dev *dev,uint16_t pasid)@@ -250,7 +263,7 @@ void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid) len =snprintf (fifo_in,sizeof (fifo_in),"%x %x:%s\n" ,KFD_SMI_EVENT_VMFAULT , task_info.pid, task_info.task_name); -add_event_to_kfifo (dev,KFD_SMI_EVENT_VMFAULT , fifo_in, len); +add_event_to_kfifo (0 , dev,KFD_SMI_EVENT_VMFAULT , fifo_in, len); }int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd)@@ -282,6 +295,8 @@ int kfd_smi_event_open(struct kfd_dev *dev, uint32_t *fd)spin_lock_init (&client->lock); client->events =0 ; client->dev = dev; + client->pid = current->pid; + client->suser =capable (CAP_SYS_ADMIN );spin_lock (&dev->smi_lock);list_add_rcu (&client->list, &dev->smi_clients);
Felix Kuehling
maintainerRe: [PATCH v2 4/8] drm/amdkfd: Add GPU recoverable fault SMI event
In reply to Philip YangAm 2022-01-20 um 18:13 schrieb Philip Yang:
Quoted 16 lines
Output timestamp when GPU recoverable fault starts, ends and duration to recover the fault, if migration happened or only GPU page table is updated, fault address, read or write fault.
Signed-off-by Philip Yang <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 48 +++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 7 ++- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 17 ++++++-- 3 files changed, 67 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +0
-0
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid) · line 273
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +5
-0@@ -266,6 +266,54 @@ void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid)add_event_to_kfifo (0 , dev,KFD_SMI_EVENT_VMFAULT , fifo_in, len); } +static bool kfd_smi_event_duration (struct kfd_dev *dev,uint64_t ts, +uint64_t *duration) +{ +if (list_empty (&dev->smi_clients)) +return false;
I'm not sure what this check has to do with the duration calculation. I think this should be checked in the caller.
But I'd actually prefer not to include the duration in the message because it's redundant. If you have both the start and end event in the log, the duration can be easily calculated from the time stamps by whatever tool is used to analyze or visualize the log. We do need a way to match the start event to the end event. I think the PID and address should be good enough for that.
Quoted 98 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +43
-0+ + *duration =ktime_get_boottime_ns () - ts; +return true; +} + +void kfd_smi_event_page_fault_start (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool write_fault, +uint64_t ts) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%lx(%x) %c\n" , +KFD_SMI_EVENT_PAGE_FAULT_START , ts, pid, address, + dev->id, write_fault ?'W' :'R' ); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_PAGE_FAULT_START , fifo_in, + len); +} + +void kfd_smi_event_page_fault_end (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool migration, +uint64_t ts) +{ +char fifo_in[64 ]; +uint64_t duration; +int len; + +if (!kfd_smi_event_duration (dev, ts, &duration)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in), +"%x %lld(%lld) -%d @%lx(%x) %c\n" , +KFD_SMI_EVENT_PAGE_FAULT_END ,ktime_get_boottime_ns (), + duration, pid, address, dev->id, migration ?'M' :'m' ); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_PAGE_FAULT_END , fifo_in, + len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +6
-1@@ -28,5 +28,10 @@ void kfd_smi_event_update_vmfault(struct kfd_dev *dev, uint16_t pasid);void kfd_smi_event_update_thermal_throttling (struct kfd_dev *dev,uint64_t throttle_bitmask);void kfd_smi_event_update_gpu_reset (struct kfd_dev *dev,bool post_reset); - +void kfd_smi_event_page_fault_start (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool write_fault, +uint64_t ts); +void kfd_smi_event_page_fault_end (struct kfd_dev *dev,pid_t pid, +unsigned long address,bool migration, +uint64_t ts);#endif drivers/gpu/drm/amd/amdkfd/kfd_svm.c +4
-2@@ -32,6 +32,7 @@#include "kfd_priv.h" #include "kfd_svm.h" #include "kfd_migrate.h" +#include "kfd_smi_events.h" #ifdef dev_fmt#undef dev_fmt@@ -1596,7 +1597,7 @@ static int svm_range_validate_and_map(struct mm_struct *mm,svm_range_unreserve_bos (&ctx);if (!r) - prange->validate_timestamp =ktime_to_us (ktime_get ()); + prange->validate_timestamp =ktime_get_boottime_ns ();return r; }@@ -2665,11 +2666,12 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,struct svm_range_list *svms;struct svm_range *prange;struct kfd_process *p; -uint64_t timestamp; +uint64_t timestamp =ktime_get_boottime_ns ();int32_t best_loc;int32_t gpuidx =MAX_GPU_INSTANCE ;bool write_locked = false;struct vm_area_struct *vma; +bool migration = false;int r =0 ;if (!KFD_IS_SVM_API_SUPPORTED (adev->kfd.dev)) {
drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, · line 2752
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +2
-2@@ -2745,9 +2747,9 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, goto out_unlock_range; } - timestamp =ktime_to_us (ktime_get ()) - prange->validate_timestamp;/* skip duplicate vm fault on different pages of same range */ -if (timestamp <AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING ) { +if (div_u64 (timestamp - prange->validate_timestamp,1000000 ) < +AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING ) {
You can probably avoid the division if you redefine AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING to be in nanoseconds.
Regards,
Felix
Quoted 25 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +7
-0pr_debug ("svms 0x%p [0x%lx %lx] already restored\n" , svms, prange->start, prange->last); r =0 ;@@ -2783,7 +2785,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid, svms, prange->start, prange->last, best_loc, prange->actual_loc); +kfd_smi_event_page_fault_start (adev->kfd.dev, p->lead_thread->pid, addr, + write_fault, timestamp); +if (prange->actual_loc != best_loc) { + migration = true;if (best_loc) { r =svm_migrate_to_vram (prange, best_loc, mm);if (r) {@@ -2812,6 +2818,9 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,pr_debug ("failed %d to map svms 0x%p [0x%lx 0x%lx] to gpus\n" , r, svms, prange->start, prange->last); +kfd_smi_event_page_fault_end (adev->kfd.dev, p->lead_thread->pid, addr, + migration, timestamp); + out_unlock_range:mutex_unlock (&prange->migrate_mutex); out_unlock_svms:
Felix Kuehling
maintainerRe: [PATCH v2 5/8] drm/amdkfd: add migration SMI event
In reply to Philip YangAm 2022-01-20 um 18:13 schrieb Philip Yang:
Quoted 54 lines
After migration is finished, output timestamp when migration starts, duration of migration, svm range address and size, GPU id of migration source and destination and svm range attributes,
Migration trigger could be prefetch, CPU or GPU page fault and TTM eviction.
Signed-off-by Philip Yang <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 67 ++++++++++++++------- drivers/gpu/drm/amd/amdkfd/kfd_migrate.h | 5 +- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 29 +++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 16 +++-- 5 files changed, 91 insertions(+), 31 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +8
-2@@ -32,6 +32,7 @@#include "kfd_priv.h" #include "kfd_svm.h" #include "kfd_migrate.h" +#include "kfd_smi_events.h" #ifdef dev_fmt#undef dev_fmt@@ -402,10 +403,11 @@ svm_migrate_copy_to_vram(struct amdgpu_device *adev, struct svm_range *prange,static long svm_migrate_vma_to_vram (struct amdgpu_device *adev,struct svm_range *prange,struct vm_area_struct *vma,uint64_t start, -uint64_t end) +uint64_t end,uint32_t trigger) {uint64_t npages = (end - start) >>PAGE_SHIFT ; -struct kfd_process_device *pdd; +struct kfd_process_device *pdd =NULL ; +uint64_t timestamp =ktime_get_boottime_ns ();struct dma_fence *mfence =NULL ;struct migrate_vma migrate;unsigned long cpages =0 ;@@ -431,6 +433,10 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, migrate.dst = migrate.src + npages; scratch = (dma_addr_t *)(migrate.dst + npages); + pdd =svm_range_get_pdd_by_adev (prange, adev); +if (!pdd) + goto out_free; + r =migrate_vma_setup (&migrate);if (r) { dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r,
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange, · line 468
Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +1
-0@@ -459,6 +465,11 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,svm_migrate_copy_done (adev, mfence);migrate_vma_finalize (&migrate); + kfd_smi_event_migration(adev->kfd.dev, pdd->process->pasid,
The pasid is only used by kfd_smi_event_migration to determine the PID. But you already know the PID here:
container_of(prange->svms, struct kfd_process, svms)->lead_thread->pid;
Quoted 219 lines
drivers/gpu/drm/amd/amdkfd/kfd_migrate.c +35
-21+ start >>PAGE_SHIFT , end >>PAGE_SHIFT , +0 , adev->kfd.dev->id, prange->prefetch_loc, + prange->preferred_loc, trigger, timestamp); +svm_range_dma_unmap (adev->dev, scratch,0 , npages);svm_range_free_dma_mappings (prange);@@ -466,10 +477,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,kvfree (buf); out:if (!r && cpages) { - pdd =svm_range_get_pdd_by_adev (prange, adev); -if (pdd) -WRITE_ONCE (pdd->page_in, pdd->page_in + cpages); - +WRITE_ONCE (pdd->page_in, pdd->page_in + cpages);return cpages; }return r;@@ -480,6 +488,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,* @prange: range structure * @best_loc: the device to migrate to * @mm: the process mm structure +* @trigger: reason of migration * * Context: Process context, caller hold mmap read lock, svms lock, prange lock * @@ -488,7 +497,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,*/ static int svm_migrate_ram_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm) +struct mm_struct *mm,uint32_t trigger) {unsigned long addr, start, end;struct vm_area_struct *vma;@@ -525,7 +534,7 @@ svm_migrate_ram_to_vram(struct svm_range *prange, uint32_t best_loc,break ; next =min (vma->vm_end, end); - r =svm_migrate_vma_to_vram (adev, prange, vma, addr, next); + r =svm_migrate_vma_to_vram (adev, prange, vma, addr, next, trigger);if (r <0 ) {pr_debug ("failed %ld to migrate\n" , r);break ;@@ -641,12 +650,14 @@ svm_migrate_copy_to_ram(struct amdgpu_device *adev, struct svm_range *prange,static long svm_migrate_vma_to_ram (struct amdgpu_device *adev,struct svm_range *prange, -struct vm_area_struct *vma,uint64_t start,uint64_t end) +struct vm_area_struct *vma,uint64_t start,uint64_t end, +uint32_t trigger) {uint64_t npages = (end - start) >>PAGE_SHIFT ; +uint64_t timestamp =ktime_get_boottime_ns ();unsigned long upages = npages;unsigned long cpages =0 ; -struct kfd_process_device *pdd; +struct kfd_process_device *pdd =NULL ;struct dma_fence *mfence =NULL ;struct migrate_vma migrate;dma_addr_t *scratch;@@ -671,6 +682,10 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange, migrate.dst = migrate.src + npages; scratch = (dma_addr_t *)(migrate.dst + npages); + pdd =svm_range_get_pdd_by_adev (prange, adev); +if (!pdd) + goto out_free; + r =migrate_vma_setup (&migrate);if (r) { dev_err(adev->dev, "vma setup fail %d range [0x%lx 0x%lx]\n", r,@@ -701,16 +716,19 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,svm_migrate_copy_done (adev, mfence);migrate_vma_finalize (&migrate); + +kfd_smi_event_migration (adev->kfd.dev, pdd->process->pasid, + start >>PAGE_SHIFT , end >>PAGE_SHIFT , + adev->kfd.dev->id,0 , prange->prefetch_loc, + prange->preferred_loc, trigger, timestamp); +svm_range_dma_unmap (adev->dev, scratch,0 , npages); out_free:kvfree (buf); out:if (!r && cpages) { - pdd =svm_range_get_pdd_by_adev (prange, adev); -if (pdd) -WRITE_ONCE (pdd->page_out, pdd->page_out + cpages); - +WRITE_ONCE (pdd->page_out, pdd->page_out + cpages);return upages; }return r ? r : upages;@@ -720,13 +738,15 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,* svm_migrate_vram_to_ram - migrate svm range from device to system * @prange: range structure * @mm: process mm, use current->mm if NULL +* @trigger: reason of migration * * Context: Process context, caller hold mmap read lock, svms lock, prange lock * * Return: * 0 - OK, otherwise error code */ -int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm) +int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm, +uint32_t trigger) {struct amdgpu_device *adev;struct vm_area_struct *vma;@@ -764,7 +784,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)break ; next =min (vma->vm_end, end); - r =svm_migrate_vma_to_ram (adev, prange, vma, addr, next); + r =svm_migrate_vma_to_ram (adev, prange, vma, addr, next, trigger);if (r <0 ) {pr_debug ("failed %ld to migrate\n" , r);break ;@@ -787,6 +807,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)* @prange: range structure * @best_loc: the device to migrate to * @mm: process mm, use current->mm if NULL +* @trigger: reason of migration * * Context: Process context, caller hold mmap read lock, svms lock, prange lock * @@ -795,7 +816,7 @@ int svm_migrate_vram_to_ram(struct svm_range *prange, struct mm_struct *mm)*/ static int svm_migrate_vram_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm) +struct mm_struct *mm,uint32_t trigger) {int r, retries =3 ;@@ -807,7 +828,7 @@ svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,pr_debug ("from gpu 0x%x to gpu 0x%x\n" , prange->actual_loc, best_loc);do { - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm, trigger);if (r)return r; }while (prange->actual_loc && --retries);@@ -815,17 +836,17 @@ svm_migrate_vram_to_vram(struct svm_range *prange, uint32_t best_loc,if (prange->actual_loc)return -EDEADLK ; -return svm_migrate_ram_to_vram (prange, best_loc, mm); +return svm_migrate_ram_to_vram (prange, best_loc, mm, trigger); }int svm_migrate_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm) +struct mm_struct *mm,uint32_t trigger) {if (!prange->actual_loc) -return svm_migrate_ram_to_vram (prange, best_loc, mm); +return svm_migrate_ram_to_vram (prange, best_loc, mm, trigger);else -return svm_migrate_vram_to_vram (prange, best_loc, mm); +return svm_migrate_vram_to_vram (prange, best_loc, mm, trigger); }@@ -894,7 +915,7 @@ static vm_fault_t svm_migrate_to_ram(struct vm_fault *vmf) goto out_unlock_prange; } - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm,MIGRATION_TRIGGER_PAGEFAULT_CPU );if (r)pr_debug ("failed %d migrate 0x%p [0x%lx 0x%lx] to ram\n" , r, prange, prange->start, prange->last);drivers/gpu/drm/amd/amdkfd/kfd_migrate.h +3
-2@@ -41,8 +41,9 @@ enum MIGRATION_COPY_DIR { };int svm_migrate_to_vram (struct svm_range *prange,uint32_t best_loc, -struct mm_struct *mm); -int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm); +struct mm_struct *mm,uint32_t trigger); +int svm_migrate_vram_to_ram (struct svm_range *prange,struct mm_struct *mm, +uint32_t trigger);unsigned long svm_migrate_addr_to_pfn (struct amdgpu_device *adev,unsigned long addr);drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +15
-0@@ -314,6 +314,35 @@ void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, len); } +void kfd_smi_event_migration (struct kfd_dev *dev,uint16_t pasid, +unsigned long start,unsigned long end, +uint32_t from,uint32_t to, +uint32_t prefetch_loc,uint32_t preferred_loc, +uint32_t trigger,uint64_t ts) +{ +struct amdgpu_task_info task_info = {0 }; +char fifo_in[128 ]; +uint64_t duration; +pid_t pid; +int len; + +if (!kfd_smi_event_duration (dev, ts, &duration)) +return ; +
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c void kfd_smi_event_page_fault_end(struct kfd_dev *dev, pid_t pid, · line 341
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +10
-0+amdgpu_vm_get_task_info (dev->adev, pasid, &task_info); +if (!task_info.pid) +return ; + pid = task_info.pid; + + len =snprintf (fifo_in,sizeof (fifo_in), +"%x %lld(%lld) -%d @%lx(%lx) %x->%x %x:%x %d\n" , +KFD_SMI_EVENT_MIGRATION , ts, duration, pid, start, + end - start, from, to, prefetch_loc, preferred_loc, + trigger);
Would it make sense to have start and end events instead of a duration, for consistency with the page fault events?
Regards,
Felix
Quoted 77 lines
drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +4
-0+ +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_MIGRATION , fifo_in, len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +5
-0@@ -34,4 +34,9 @@ void kfd_smi_event_page_fault_start(struct kfd_dev *dev, pid_t pid,void kfd_smi_event_page_fault_end (struct kfd_dev *dev,pid_t pid,unsigned long address,bool migration,uint64_t ts); +void kfd_smi_event_migration (struct kfd_dev *dev,uint16_t pasid, +unsigned long start,unsigned long end, +uint32_t from,uint32_t to, +uint32_t prefetch_loc,uint32_t preferred_loc, +uint32_t trigger,uint64_t ts);#endif drivers/gpu/drm/amd/amdkfd/kfd_svm.c +10
-6@@ -2791,7 +2791,8 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,if (prange->actual_loc != best_loc) { migration = true;if (best_loc) { - r =svm_migrate_to_vram (prange, best_loc, mm); + r =svm_migrate_to_vram (prange, best_loc, mm, +MIGRATION_TRIGGER_PAGEFAULT );if (r) {pr_debug ("svm_migrate_to_vram failed (%d) at %llx, falling back to system memory\n" , r, addr);@@ -2799,12 +2800,14 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,* VRAM failed */ if (prange->actual_loc) - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm, +MIGRATION_TRIGGER_PAGEFAULT );else r =0 ; } }else { - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm, +MIGRATION_TRIGGER_PAGEFAULT ); }if (r) { pr_debug("failed %d to migrate svms %p [0x%lx 0x%lx]\n",@@ -3126,12 +3129,12 @@ svm_range_trigger_migration(struct mm_struct *mm, struct svm_range *prange,return 0 ;if (!best_loc) { - r =svm_migrate_vram_to_ram (prange, mm); + r =svm_migrate_vram_to_ram (prange, mm,MIGRATION_TRIGGER_PREFETCH ); *migrated = !r;return r; } - r =svm_migrate_to_vram (prange, best_loc, mm); + r =svm_migrate_to_vram (prange, best_loc, mm,MIGRATION_TRIGGER_PREFETCH ); *migrated = !r;return r;@@ -3188,7 +3191,8 @@ static void svm_range_evict_svm_bo_worker(struct work_struct *work)mutex_lock (&prange->migrate_mutex);do {svm_migrate_vram_to_ram (prange, - svm_bo->eviction_fence->mm); + svm_bo->eviction_fence->mm, +MIGRATION_TRIGGER_TTM_EVICTION ); }while (prange->actual_loc && --retries);WARN (prange->actual_loc,"Migration failed during eviction" );
Felix Kuehling
maintainerRe: [PATCH v2 6/8] drm/amdkfd: Add user queue eviction restore SMI event
In reply to Philip YangAm 2022-01-20 um 18:13 schrieb Philip Yang:
Quoted 135 lines
Output user queue eviction and restore event. User queue eviction may be triggered by migration, MMU notifier, TTM eviction or device suspend.
User queue restore may be rescheduled if eviction happens again while restore.
Signed-off-by Philip Yang <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 7 +++- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 11 ++++-- drivers/gpu/drm/amd/amdkfd/kfd_device.c | 4 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +- drivers/gpu/drm/amd/amdkfd/kfd_process.c | 37 +++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c | 34 +++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h | 4 ++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 16 ++++++-- 8 files changed, 101 insertions(+), 14 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +6
-1@@ -309,6 +309,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,*/ void amdgpu_amdkfd_release_notify (struct amdgpu_bo *bo);void amdgpu_amdkfd_reserve_system_mem (uint64_t size); +void kfd_process_smi_event_restore_rescheduled (struct mm_struct *mm);#else static inline void amdgpu_amdkfd_gpuvm_init_mem_limits (void )@@ -325,9 +326,13 @@ static inlinevoid amdgpu_amdkfd_release_notify (struct amdgpu_bo *bo) { } + +static inline void kfd_process_smi_event_restore_rescheduled (struct mm_struct *mm) +{ +}#endif /* KGD2KFD callbacks */ -int kgd2kfd_quiesce_mm (struct mm_struct *mm); +int kgd2kfd_quiesce_mm (struct mm_struct *mm,uint32_t trigger);int kgd2kfd_resume_mm (struct mm_struct *mm);int kgd2kfd_schedule_evict_and_restore_process (struct mm_struct *mm,struct dma_fence *fence);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7
-4@@ -2066,7 +2066,7 @@ int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, evicted_bos =atomic_inc_return (&process_info->evicted_bos);if (evicted_bos ==1 ) {/* First eviction, stop the queues */ - r =kgd2kfd_quiesce_mm (mm); + r =kgd2kfd_quiesce_mm (mm,USERPTR_EVICTION );if (r)pr_err ("Failed to quiesce KFD\n" ); schedule_delayed_work(&process_info->restore_userptr_work,@@ -2340,13 +2340,16 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) unlock_out:mutex_unlock (&process_info->lock); -mmput (mm); -put_task_struct (usertask);/* If validation failed, reschedule another attempt */ -if (evicted_bos) +if (evicted_bos) {schedule_delayed_work (&process_info->restore_userptr_work,msecs_to_jiffies (AMDGPU_USERPTR_RESTORE_DELAY_MS )); + +kfd_process_smi_event_restore_rescheduled (mm); + } +mmput (mm); +put_task_struct (usertask); }/** amdgpu_amdkfd_gpuvm_restore_process_bos - Restore all BOs for the given drivers/gpu/drm/amd/amdkfd/kfd_device.c +2
-2@@ -783,7 +783,7 @@ void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)spin_unlock_irqrestore (&kfd->interrupt_lock, flags); } -int kgd2kfd_quiesce_mm (struct mm_struct *mm) +int kgd2kfd_quiesce_mm (struct mm_struct *mm,uint32_t trigger) {struct kfd_process *p;int r;@@ -797,7 +797,7 @@ int kgd2kfd_quiesce_mm(struct mm_struct *mm)return -ESRCH ;WARN (debug_evictions,"Evicting pid %d" , p->lead_thread->pid); - r =kfd_process_evict_queues (p); + r =kfd_process_evict_queues (p, trigger);kfd_unref_process (p);return r;drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1
-1@@ -906,7 +906,7 @@ static inline struct kfd_process_device *kfd_process_device_from_gpuidx( }void kfd_unref_process (struct kfd_process *p); -int kfd_process_evict_queues (struct kfd_process *p); +int kfd_process_evict_queues (struct kfd_process *p,uint32_t trigger);int kfd_process_restore_queues (struct kfd_process *p);void kfd_suspend_all_processes (void );int kfd_resume_all_processes (void );drivers/gpu/drm/amd/amdkfd/kfd_process.c +2
-1@@ -43,6 +43,7 @@ struct mm_struct;#include "kfd_dbgmgr.h" #include "kfd_iommu.h" #include "kfd_svm.h" +#include "kfd_smi_events.h" /* * List of struct kfd_process (field kfd_process). @@ -1701,7 +1702,7 @@ struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)* Eviction is reference-counted per process-device. This means multiple * evictions from different sources can be nested safely. */ -int kfd_process_evict_queues (struct kfd_process *p) +int kfd_process_evict_queues (struct kfd_process *p,uint32_t trigger) {int r =0 ;int i;
drivers/gpu/drm/amd/amdkfd/kfd_process.c int kfd_process_evict_queues(struct kfd_process *p) · line 1716
Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_process.c +3
-0@@ -1710,6 +1711,9 @@ int kfd_process_evict_queues(struct kfd_process *p)for (i =0 ; i < p->n_pdds; i++) {struct kfd_process_device *pdd = p->pdds[i]; +kfd_smi_event_queue_eviction (pdd->dev, p->lead_thread->pid, + trigger); +
If we're always evicting queues on all GPUs, we probably only need one event per process, not one event per GPU.
Quoted 34 lines
drivers/gpu/drm/amd/amdkfd/kfd_process.c +13
-0r = pdd->dev->dqm->ops.evict_process_queues (pdd->dev->dqm, &pdd->qpd);/* evict return -EIO if HWS is hang or asic is resetting, in this case @@ -1734,6 +1738,11 @@ int kfd_process_evict_queues(struct kfd_process *p)if (n_evicted ==0 )break ; + +kfd_smi_event_queue_eviction_restore (pdd->dev, + p->lead_thread->pid, + false); +if (pdd->dev->dqm->ops.restore_process_queues (pdd->dev->dqm, &pdd->qpd))pr_err ("Failed to restore queues\n" );@@ -1753,6 +1762,10 @@ int kfd_process_restore_queues(struct kfd_process *p)for (i =0 ; i < p->n_pdds; i++) {struct kfd_process_device *pdd = p->pdds[i]; +kfd_smi_event_queue_eviction_restore (pdd->dev, + p->lead_thread->pid, + false); + r = pdd->dev->dqm->ops.restore_process_queues (pdd->dev->dqm, &pdd->qpd);if (r) {@@ -1765,6 +1778,24 @@ int kfd_process_restore_queues(struct kfd_process *p)return ret; } +void kfd_process_smi_event_restore_rescheduled (struct mm_struct *mm) +{ +struct kfd_process *p; +int i;
drivers/gpu/drm/amd/amdkfd/kfd_process.c int kfd_process_restore_queues(struct kfd_process *p) · line 1794
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_process.c +10
-0+ + p =kfd_lookup_process_by_mm (mm); +if (!p) +return ; + +for (i =0 ; i < p->n_pdds; i++) { +struct kfd_process_device *pdd = p->pdds[i]; + +kfd_smi_event_queue_eviction_restore (pdd->dev, + p->lead_thread->pid, true);
Same as above. One event per process should be enough.
Quoted 88 lines
drivers/gpu/drm/amd/amdkfd/kfd_process.c +6
-2+ } +kfd_unref_process (p); +} +int kfd_process_gpuidx_from_gpuid (struct kfd_process *p,uint32_t gpu_id) {int i;@@ -1814,7 +1845,7 @@ static void evict_process_worker(struct work_struct *work)flush_delayed_work (&p->restore_work);pr_debug ("Started evicting pasid 0x%x\n" , p->pasid); - ret =kfd_process_evict_queues (p); + ret =kfd_process_evict_queues (p,TTM_EVICTION );if (!ret) {dma_fence_signal (p->ef);dma_fence_put (p->ef);@@ -1881,7 +1912,7 @@ void kfd_suspend_all_processes(void)cancel_delayed_work_sync (&p->eviction_work);cancel_delayed_work_sync (&p->restore_work); -if (kfd_process_evict_queues (p)) +if (kfd_process_evict_queues (p,SUSPEND_EVICTION ))pr_err ("Failed to suspend process 0x%x\n" , p->pasid);dma_fence_signal (p->ef);dma_fence_put (p->ef);drivers/gpu/drm/amd/amdkfd/kfd_smi_events.c +34
-0@@ -343,6 +343,40 @@ void kfd_smi_event_migration(struct kfd_dev *dev, uint16_t pasid,add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_MIGRATION , fifo_in, len); } +void kfd_smi_event_queue_eviction (struct kfd_dev *dev,pid_t pid, +uint32_t trigger) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%x %d\n" , +KFD_SMI_EVENT_QUEUE_EVICTION ,ktime_get_boottime_ns (), + pid, dev->id, trigger); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_QUEUE_EVICTION , fifo_in, len); +} + +void kfd_smi_event_queue_eviction_restore (struct kfd_dev *dev,pid_t pid, +bool rescheduled) +{ +char fifo_in[64 ]; +int len; + +if (list_empty (&dev->smi_clients)) +return ; + + len =snprintf (fifo_in,sizeof (fifo_in),"%x %lld -%d @%x %c\n" , +KFD_SMI_EVENT_QUEUE_EVICTION_RESTORE , +ktime_get_boottime_ns (), pid, dev->id, + rescheduled ?'r' :' ' ); + +add_event_to_kfifo (pid, dev,KFD_SMI_EVENT_QUEUE_EVICTION_RESTORE , + fifo_in, len); +} +int kfd_smi_event_open (struct kfd_dev *dev,uint32_t *fd) {struct kfd_smi_client *client;drivers/gpu/drm/amd/amdkfd/kfd_smi_events.h +4
-0@@ -39,4 +39,8 @@ void kfd_smi_event_migration(struct kfd_dev *dev, uint16_t pasid, uint32_t from, uint32_t to, uint32_t prefetch_loc, uint32_t preferred_loc, uint32_t trigger, uint64_t ts); +void kfd_smi_event_queue_eviction (struct kfd_dev *dev,pid_t pid, +uint32_t trigger); +void kfd_smi_event_queue_eviction_restore (struct kfd_dev *dev,pid_t pid, +bool rescheduled);#endif drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0
drivers/gpu/drm/amd/amdkfd/kfd_svm.c static void svm_range_restore_work(struct work_struct *work) · line 1718
Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-0@@ -1712,7 +1712,10 @@ static void svm_range_restore_work(struct work_struct *work)pr_debug ("reschedule to restore svm range\n" );schedule_delayed_work (&svms->restore_work,msecs_to_jiffies (AMDGPU_SVM_RANGE_RESTORE_DELAY_MS )); + +kfd_process_smi_event_restore_rescheduled (mm); } +mmput (mm);
Where is the matching mmget for this? If this is a fix for an old reference-leak bug in the code, it deserves its own patch.
Regards,
Felix
Quoted 45 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +10
-3}/** @@ -1732,15 +1735,22 @@ static void svm_range_restore_work(struct work_struct *work)*/ static int svm_range_evict (struct svm_range *prange,struct mm_struct *mm, -unsigned long start,unsigned long last) +unsigned long start,unsigned long last, +enum mmu_notifier_event event) {struct svm_range_list *svms = prange->svms;struct svm_range *pchild;struct kfd_process *p; +uint32_t trigger;int r =0 ; p = container_of(svms, struct kfd_process, svms); +if (event ==MMU_NOTIFY_MIGRATE ) + trigger =SVM_RANGE_MIGRATION ; +else + trigger =SVM_RANGE_EVICTION ; +pr_debug ("invalidate svms 0x%p prange [0x%lx 0x%lx] [0x%lx 0x%lx]\n" , svms, prange->start, prange->last, start, last);@@ -1768,7 +1778,7 @@ svm_range_evict(struct svm_range *prange, struct mm_struct *mm, prange->svms, prange->start, prange->last);/* First eviction, stop the queues */ - r =kgd2kfd_quiesce_mm (mm); + r =kgd2kfd_quiesce_mm (mm, trigger);if (r)pr_debug ("failed to quiesce KFD\n" );@@ -2303,7 +2313,7 @@ svm_range_cpu_invalidate_pagetables(struct mmu_interval_notifier *mni,svm_range_unmap_from_cpu (mni->mm, prange, start, last);break ;default : -svm_range_evict (prange, mni->mm, start, last); +svm_range_evict (prange, mni->mm, start, last, range->event);break ; }