drm/xe: Handle -EDEADLK case in preempt worker

With multiple active VMs, under memory pressure, it is possible that
ttm_bo_validate() run into -EDEADLK in ttm_mem_evict_wait_busy() and
return -ENOMEM.

Until ttm properly handles locking in such scenarios, best thing the
driver can do is unwind the lock and retry.

Update preempt worker to retry validating BOs with a timeout upon
-ENOMEM.

v2: revert retry timeout upon -EAGAIN (Thomas)

Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>

diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
index 06b559ff80bf5689422249085c2e3756036b1746..d9579bf5002dbd0f43b7baf3f986d22298bc943f 100644 (file)
--- a/drivers/gpu/drm/xe/xe_vm.c
+++ b/drivers/gpu/drm/xe/xe_vm.c
@@ -10,6 +10,7 @@
 #include <drm/ttm/ttm_execbuf_util.h>
 #include <drm/ttm/ttm_tt.h>
 #include <drm/xe_drm.h>
+#include <linux/delay.h>
 #include <linux/kthread.h>
 #include <linux/mm.h>
 #include <linux/swap.h>
@@ -508,6 +509,8 @@ void xe_vm_unlock_dma_resv(struct xe_vm *vm,
                kvfree(tv);
 }
 
+#define XE_VM_REBIND_RETRY_TIMEOUT_MS 1000
+
 static void preempt_rebind_work_func(struct work_struct *w)
 {
        struct xe_vm *vm = container_of(w, struct xe_vm, preempt.rebind_work);
@@ -519,6 +522,7 @@ static void preempt_rebind_work_func(struct work_struct *w)
        struct dma_fence *rebind_fence;
        unsigned int fence_count = 0;
        LIST_HEAD(preempt_fences);
+       ktime_t end = 0;
        int err;
        long wait;
        int __maybe_unused tries = 0;
@@ -637,6 +641,24 @@ out_unlock_outer:
                trace_xe_vm_rebind_worker_retry(vm);
                goto retry;
        }
+
+       /*
+        * With multiple active VMs, under memory pressure, it is possible that
+        * ttm_bo_validate() run into -EDEADLK and in such case returns -ENOMEM.
+        * Until ttm properly handles locking in such scenarios, best thing the
+        * driver can do is retry with a timeout. Killing the VM or putting it
+        * in error state after timeout or other error scenarios is still TBD.
+        */
+       if (err == -ENOMEM) {
+               ktime_t cur = ktime_get();
+
+               end = end ? : ktime_add_ms(cur, XE_VM_REBIND_RETRY_TIMEOUT_MS);
+               if (ktime_before(cur, end)) {
+                       msleep(20);
+                       trace_xe_vm_rebind_worker_retry(vm);
+                       goto retry;
+               }
+       }
        up_write(&vm->lock);
 
        free_preempt_fences(&preempt_fences);