]> www.infradead.org Git - users/hch/misc.git/commitdiff
drm/amdgpu: Add mapping info option for GEM_OP ioctl
authorDavid Francis <David.Francis@amd.com>
Mon, 16 Jun 2025 13:49:33 +0000 (09:49 -0400)
committerAlex Deucher <alexander.deucher@amd.com>
Tue, 2 Sep 2025 19:53:33 +0000 (15:53 -0400)
Add new GEM_OP_IOCTL option GET_MAPPING_INFO, which
returns a list of mappings associated with a given bo, along with
their positions and offsets.

Userspace for this and the previous change can be found at:
https://github.com/checkpoint-restore/criu/pull/2613

Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
include/uapi/drm/amdgpu_drm.h

index bddb8d7475b2ed2303129b949289f257938147bd..e587e5ddeb000dbdecb5baa32ebb4df8664127c4 100644 (file)
@@ -955,17 +955,34 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
        struct drm_gem_object *gobj;
        struct amdgpu_vm_bo_base *base;
        struct amdgpu_bo *robj;
+       struct drm_exec exec;
+       struct amdgpu_fpriv *fpriv = filp->driver_priv;
        int r;
 
+       if (args->padding)
+               return -EINVAL;
+
        gobj = drm_gem_object_lookup(filp, args->handle);
        if (!gobj)
                return -ENOENT;
 
        robj = gem_to_amdgpu_bo(gobj);
 
-       r = amdgpu_bo_reserve(robj, false);
-       if (unlikely(r))
-               goto out;
+       drm_exec_init(&exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
+                         DRM_EXEC_IGNORE_DUPLICATES, 0);
+       drm_exec_until_all_locked(&exec) {
+               r = drm_exec_lock_obj(&exec, gobj);
+               drm_exec_retry_on_contention(&exec);
+               if (r)
+                       goto out_exec;
+
+               if (args->op == AMDGPU_GEM_OP_GET_MAPPING_INFO) {
+                       r = amdgpu_vm_lock_pd(&fpriv->vm, &exec, 0);
+                       drm_exec_retry_on_contention(&exec);
+                       if (r)
+                               goto out_exec;
+               }
+       }
 
        switch (args->op) {
        case AMDGPU_GEM_OP_GET_GEM_CREATE_INFO: {
@@ -976,7 +993,7 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
                info.alignment = robj->tbo.page_alignment << PAGE_SHIFT;
                info.domains = robj->preferred_domains;
                info.domain_flags = robj->flags;
-               amdgpu_bo_unreserve(robj);
+               drm_exec_fini(&exec);
                if (copy_to_user(out, &info, sizeof(info)))
                        r = -EFAULT;
                break;
@@ -985,20 +1002,17 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
                if (drm_gem_is_imported(&robj->tbo.base) &&
                    args->value & AMDGPU_GEM_DOMAIN_VRAM) {
                        r = -EINVAL;
-                       amdgpu_bo_unreserve(robj);
-                       break;
+                       goto out_exec;
                }
                if (amdgpu_ttm_tt_get_usermm(robj->tbo.ttm)) {
                        r = -EPERM;
-                       amdgpu_bo_unreserve(robj);
-                       break;
+                       goto out_exec;
                }
                for (base = robj->vm_bo; base; base = base->next)
                        if (amdgpu_xgmi_same_hive(amdgpu_ttm_adev(robj->tbo.bdev),
                                amdgpu_ttm_adev(base->vm->root.bo->tbo.bdev))) {
                                r = -EINVAL;
-                               amdgpu_bo_unreserve(robj);
-                               goto out;
+                               goto out_exec;
                        }
 
 
@@ -1011,15 +1025,63 @@ int amdgpu_gem_op_ioctl(struct drm_device *dev, void *data,
 
                if (robj->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
                        amdgpu_vm_bo_invalidate(robj, true);
+               drm_exec_fini(&exec);
+               break;
+       case AMDGPU_GEM_OP_GET_MAPPING_INFO: {
+               struct amdgpu_bo_va *bo_va = amdgpu_vm_bo_find(&fpriv->vm, robj);
+               struct drm_amdgpu_gem_vm_entry *vm_entries;
+               struct amdgpu_bo_va_mapping *mapping;
+               int num_mappings = 0;
+               /*
+                * num_entries is set as an input to the size of the user-allocated array of
+                * drm_amdgpu_gem_vm_entry stored at args->value.
+                * num_entries is sent back as output as the number of mappings the bo has.
+                * If that number is larger than the size of the array, the ioctl must
+                * be retried.
+                */
+               vm_entries = kvcalloc(args->num_entries, sizeof(*vm_entries), GFP_KERNEL);
+               if (!vm_entries)
+                       return -ENOMEM;
+
+               amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) {
+                       if (num_mappings < args->num_entries) {
+                               vm_entries[num_mappings].addr = mapping->start * AMDGPU_GPU_PAGE_SIZE;
+                               vm_entries[num_mappings].size = (mapping->last - mapping->start + 1) * AMDGPU_GPU_PAGE_SIZE;
+                               vm_entries[num_mappings].offset = mapping->offset;
+                               vm_entries[num_mappings].flags = mapping->flags;
+                       }
+                       num_mappings += 1;
+               }
+
+               amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) {
+                       if (num_mappings < args->num_entries) {
+                               vm_entries[num_mappings].addr = mapping->start * AMDGPU_GPU_PAGE_SIZE;
+                               vm_entries[num_mappings].size = (mapping->last - mapping->start + 1) * AMDGPU_GPU_PAGE_SIZE;
+                               vm_entries[num_mappings].offset = mapping->offset;
+                               vm_entries[num_mappings].flags = mapping->flags;
+                       }
+                       num_mappings += 1;
+               }
 
-               amdgpu_bo_unreserve(robj);
+               drm_exec_fini(&exec);
+
+               if (num_mappings > 0 && num_mappings <= args->num_entries)
+                       r = copy_to_user(u64_to_user_ptr(args->value), vm_entries, num_mappings * sizeof(*vm_entries));
+
+               args->num_entries = num_mappings;
+
+               kvfree(vm_entries);
                break;
+       }
        default:
-               amdgpu_bo_unreserve(robj);
+               drm_exec_fini(&exec);
                r = -EINVAL;
        }
 
-out:
+       drm_gem_object_put(gobj);
+       return r;
+out_exec:
+       drm_exec_fini(&exec);
        drm_gem_object_put(gobj);
        return r;
 }
index 3b4fa3246675fcc6264e28d46f9761b42f7b2461..e045c1590d78668e7d568c286b99507de4def690 100644 (file)
@@ -670,4 +670,9 @@ void amdgpu_vm_tlb_fence_create(struct amdgpu_device *adev,
 void amdgpu_vm_print_task_info(struct amdgpu_device *adev,
                               struct amdgpu_task_info *task_info);
 
+#define amdgpu_vm_bo_va_for_each_valid_mapping(bo_va, mapping) \
+               list_for_each_entry(mapping, &(bo_va)->valids, list)
+#define amdgpu_vm_bo_va_for_each_invalid_mapping(bo_va, mapping) \
+               list_for_each_entry(mapping, &(bo_va)->invalids, list)
+
 #endif
index 902e30263fcc70b4f3849be0ede69fb8e7ebc56c..9cebd072a042e302ceb4ad1309e42589a173a305 100644 (file)
@@ -802,6 +802,21 @@ union drm_amdgpu_wait_fences {
 
 #define AMDGPU_GEM_OP_GET_GEM_CREATE_INFO      0
 #define AMDGPU_GEM_OP_SET_PLACEMENT            1
+#define AMDGPU_GEM_OP_GET_MAPPING_INFO         2
+
+struct drm_amdgpu_gem_vm_entry {
+       /* Start of mapping (in bytes) */
+       __u64 addr;
+
+       /* Size of mapping (in bytes) */
+       __u64 size;
+
+       /* Mapping offset */
+       __u64 offset;
+
+       /* flags needed to recreate mapping */
+       __u64 flags;
+};
 
 /* Sets or returns a value associated with a buffer. */
 struct drm_amdgpu_gem_op {
@@ -809,8 +824,12 @@ struct drm_amdgpu_gem_op {
        __u32   handle;
        /** AMDGPU_GEM_OP_* */
        __u32   op;
-       /** Input or return value */
+       /** Input or return value. For MAPPING_INFO op: pointer to array of struct drm_amdgpu_gem_vm_entry */
        __u64   value;
+       /** For MAPPING_INFO op: number of mappings (in/out) */
+       __u32   num_entries;
+
+       __u32   padding;
 };
 
 #define AMDGPU_GEM_LIST_HANDLES_FLAG_IS_IMPORT (1 << 0)