On 6/22/2021 12:36 PM, Christian König wrote:
Am 22.06.21 um 12:30 schrieb Das, Nirmoy:

On 6/22/2021 10:36 AM, Christian König wrote:

Am 22.06.21 um 09:39 schrieb Das, Nirmoy:

On 6/22/2021 9:03 AM, Christian König wrote:


Am 22.06.21 um 08:57 schrieb Nirmoy Das:
Cleanup code related to vm pasid by adding helper functions.

Signed-off-by: Nirmoy Das <nirmoy.das@amd.com>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 105 ++++++++++++-------------
  1 file changed, 50 insertions(+), 55 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 63975bda8e76..6e476b173cbb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -87,6 +87,46 @@ struct amdgpu_prt_cb {
      struct dma_fence_cb cb;
  };

+static int amdgpu_vm_pasid_alloc(struct amdgpu_device *adev,
+                 struct amdgpu_vm *vm,
+                 unsigned int pasid,
+                 unsigned int *vm_pasid)
+{
+    unsigned long flags;
+    int r;
+
+    if (!pasid)
+        return 0;
+
+    spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
+    r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
+              GFP_ATOMIC);
+ spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
+    if (r < 0)
+        return r;
+    if (vm_pasid)
+        *vm_pasid = pasid;
+

Ok the more I read from this patch the less it makes sense.

We don't allocate the passid here, we just set it up in the idr.

What we could do is to replace the idr with an xarray, that would certainly make more sense than this here.


xarray looks great, with that we don't need pasid_lock either.

You still need the lock to protect against VM destruction while looking things up, but you could switch to RCU for this instead.


xarray has xa_{lock|unloack}_irqsave() and adev->vm_manager.pasid_xa will exist till devices's lifetime.

That's just a wrapper around the lock.

So I am thinking something like:

amdgpu_vm_pasid_insert()

{

...

xa_lock_irqsave(adev->vm_manager.pasids, flags)
r = xa_store(&adev->vm_manager.pasids, pasid, vm, GFP_ATOMIC);
xa_unlock_irqsave(adev->vm_manager.pasids, flags)

It would be really nice if we could avoid the GFP_ATOMIC here, but not much of a problem since we had that before.


I think it is possible as I think  only amdgpu_vm_handle_fault() runs in interrupt context which tries to find a vm ptr not store it.



}

amdgpu_vm_pasid_remove()

{

....

xa_lock_irqsave(adev->vm_manager.pasids, flags)
xa_erase(&adev->vm_manager.pasids, pasid);
xa_unlock_irqsave(adev->vm_manager.pasids, flags)

}


xa_{lock|unloack}_irqsave() can be use while looking up vm ptr for a pasid.


Shouldn't this be enough ?


Yeah I think so.


Great.


Nirmoy


Christian.


Regards,

Nirmoy


Christian.



Thanks

Nirmoy



Christian.

+    return 0;
+}
+
+static void amdgpu_vm_pasid_remove(struct amdgpu_device *adev,
+                   unsigned int pasid,
+                   unsigned int *vm_pasid)
+{
+    unsigned long flags;
+
+    if (!pasid)
+        return;
+
+    spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
+    idr_remove(&adev->vm_manager.pasid_idr, pasid);
+ spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
+
+    if (vm_pasid)
+        *vm_pasid = 0;
+}
+
  /*
   * vm eviction_lock can be taken in MMU notifiers. Make sure no reclaim-FS
   * happens while holding this lock anywhere to prevent deadlocks when
@@ -2940,18 +2980,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, u32 pasid)

      amdgpu_bo_unreserve(vm->root.bo);

-    if (pasid) {
-        unsigned long flags;
-
- spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
-        r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
-                  GFP_ATOMIC);
- spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
-        if (r < 0)
-            goto error_free_root;
-
-        vm->pasid = pasid;
-    }
+    if (amdgpu_vm_pasid_alloc(adev, vm, pasid, &vm->pasid))
+        goto error_free_root;

      INIT_KFIFO(vm->faults);

@@ -3038,19 +3068,11 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
      r = amdgpu_vm_check_clean_reserved(adev, vm);
      if (r)
          goto unreserve_bo;
+    r = amdgpu_vm_pasid_alloc(adev, vm, pasid, NULL);
+    if (r ==  -ENOSPC)
+        goto unreserve_bo;

-    if (pasid) {
-        unsigned long flags;
-
- spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
-        r = idr_alloc(&adev->vm_manager.pasid_idr, vm, pasid, pasid + 1,
-                  GFP_ATOMIC);
- spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
-
-        if (r == -ENOSPC)
-            goto unreserve_bo;
-        r = 0;
-    }
+    r = 0;

      /* Check if PD needs to be reinitialized and do it before
       * changing any other state, in case it fails.
@@ -3089,35 +3111,23 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
      vm->is_compute_context = true;

      if (vm->pasid) {
-        unsigned long flags;
-
- spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
-        idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
- spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
-
          /* Free the original amdgpu allocated pasid
           * Will be replaced with kfd allocated pasid
           */
          amdgpu_pasid_free(vm->pasid);
-        vm->pasid = 0;
+        amdgpu_vm_pasid_remove(adev, vm->pasid, &vm->pasid);
      }

      /* Free the shadow bo for compute VM */
amdgpu_bo_unref(&to_amdgpu_bo_vm(vm->root.bo)->shadow);
-
      if (pasid)
          vm->pasid = pasid;

      goto unreserve_bo;

  free_idr:
-    if (pasid) {
-        unsigned long flags;
+    amdgpu_vm_pasid_remove(adev, pasid, NULL);

- spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
-        idr_remove(&adev->vm_manager.pasid_idr, pasid);
- spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
-    }
  unreserve_bo:
      amdgpu_bo_unreserve(vm->root.bo);
      return r;
@@ -3133,14 +3143,7 @@ int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm,
   */
  void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  {
-    if (vm->pasid) {
-        unsigned long flags;
-
- spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
-        idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
- spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
-    }
-    vm->pasid = 0;
+    amdgpu_vm_pasid_remove(adev, vm->pasid, &vm->pasid);
      vm->is_compute_context = false;
  }

@@ -3164,15 +3167,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)

      root = amdgpu_bo_ref(vm->root.bo);
      amdgpu_bo_reserve(root, true);
-    if (vm->pasid) {
-        unsigned long flags;
-
- spin_lock_irqsave(&adev->vm_manager.pasid_lock, flags);
-        idr_remove(&adev->vm_manager.pasid_idr, vm->pasid);
- spin_unlock_irqrestore(&adev->vm_manager.pasid_lock, flags);
-        vm->pasid = 0;
-    }
-
+    amdgpu_vm_pasid_remove(adev, vm->pasid, &vm->pasid);
      dma_fence_wait(vm->last_unlocked, false);
      dma_fence_put(vm->last_unlocked);

-- 
2.32.0


_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=04%7C01%7Cnirmoy.das%40amd.com%7C3285a973b5a4498f3b0608d93558d909%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C 637599478002028860%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=QLv4hbUpPF9H%2BVL4eOQlTeROWQA%2FG1LrPGFBzCQRt7o%3D&amp;reserved=0