dma-buf: generalize dma_fence unwrap & merging v3
authorChristian König <christian.koenig@amd.com>
Mon, 25 Apr 2022 12:22:12 +0000 (14:22 +0200)
committerChristian König <christian.koenig@amd.com>
Mon, 30 May 2022 12:24:04 +0000 (14:24 +0200)
Introduce a dma_fence_unwrap_merge() macro which allows to unwrap fences
which potentially can be containers as well and then merge them back
together into a flat dma_fence_array.

v2: rename the function, add some more comments about how the wrapper is
    used, move filtering of signaled fences into the unwrap iterator,
    add complex selftest which covers more cases.
v3: fix signaled fence filtering once more

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20220518135844.3338-5-christian.koenig@amd.com
drivers/dma-buf/dma-fence-unwrap.c
drivers/dma-buf/st-dma-fence-unwrap.c
drivers/dma-buf/sync_file.c
include/linux/dma-fence-unwrap.h

index 711be125428c1ad48a7e2f62952f8a11412aaa54..502a65ea6d44fd8546c0a6719867d995c8fc4cf8 100644 (file)
@@ -11,6 +11,7 @@
 #include <linux/dma-fence-array.h>
 #include <linux/dma-fence-chain.h>
 #include <linux/dma-fence-unwrap.h>
+#include <linux/slab.h>
 
 /* Internal helper to start new array iteration, don't use directly */
 static struct dma_fence *
@@ -57,3 +58,105 @@ struct dma_fence *dma_fence_unwrap_next(struct dma_fence_unwrap *cursor)
        return __dma_fence_unwrap_array(cursor);
 }
 EXPORT_SYMBOL_GPL(dma_fence_unwrap_next);
+
+/* Implementation for the dma_fence_merge() marco, don't use directly */
+struct dma_fence *__dma_fence_unwrap_merge(unsigned int num_fences,
+                                          struct dma_fence **fences,
+                                          struct dma_fence_unwrap *iter)
+{
+       struct dma_fence_array *result;
+       struct dma_fence *tmp, **array;
+       unsigned int i;
+       size_t count;
+
+       count = 0;
+       for (i = 0; i < num_fences; ++i) {
+               dma_fence_unwrap_for_each(tmp, &iter[i], fences[i])
+                       ++count;
+       }
+
+       if (count == 0)
+               return dma_fence_get_stub();
+
+       array = kmalloc_array(count, sizeof(*array), GFP_KERNEL);
+       if (!array)
+               return NULL;
+
+       /*
+        * This trashes the input fence array and uses it as position for the
+        * following merge loop. This works because the dma_fence_merge()
+        * wrapper macro is creating this temporary array on the stack together
+        * with the iterators.
+        */
+       for (i = 0; i < num_fences; ++i)
+               fences[i] = dma_fence_unwrap_first(fences[i], &iter[i]);
+
+       count = 0;
+       do {
+               unsigned int sel;
+
+restart:
+               tmp = NULL;
+               for (i = 0; i < num_fences; ++i) {
+                       struct dma_fence *next;
+
+                       while (fences[i] && dma_fence_is_signaled(fences[i]))
+                               fences[i] = dma_fence_unwrap_next(&iter[i]);
+
+                       next = fences[i];
+                       if (!next)
+                               continue;
+
+                       /*
+                        * We can't guarantee that inpute fences are ordered by
+                        * context, but it is still quite likely when this
+                        * function is used multiple times. So attempt to order
+                        * the fences by context as we pass over them and merge
+                        * fences with the same context.
+                        */
+                       if (!tmp || tmp->context > next->context) {
+                               tmp = next;
+                               sel = i;
+
+                       } else if (tmp->context < next->context) {
+                               continue;
+
+                       } else if (dma_fence_is_later(tmp, next)) {
+                               fences[i] = dma_fence_unwrap_next(&iter[i]);
+                               goto restart;
+                       } else {
+                               fences[sel] = dma_fence_unwrap_next(&iter[sel]);
+                               goto restart;
+                       }
+               }
+
+               if (tmp) {
+                       array[count++] = dma_fence_get(tmp);
+                       fences[sel] = dma_fence_unwrap_next(&iter[sel]);
+               }
+       } while (tmp);
+
+       if (count == 0) {
+               tmp = dma_fence_get_stub();
+               goto return_tmp;
+       }
+
+       if (count == 1) {
+               tmp = array[0];
+               goto return_tmp;
+       }
+
+       result = dma_fence_array_create(count, array,
+                                       dma_fence_context_alloc(1),
+                                       1, false);
+       if (!result) {
+               tmp = NULL;
+               goto return_tmp;
+       }
+       return &result->base;
+
+return_tmp:
+       kfree(array);
+       return tmp;
+}
+EXPORT_SYMBOL_GPL(__dma_fence_unwrap_merge);
index e20c5a7dcfe4b67baeac89ac42a216488a152488..4105d5ea8ddebee86eb5b40b7b329bf859f9d827 100644 (file)
@@ -238,6 +238,113 @@ static int unwrap_chain_array(void *arg)
        return err;
 }
 
+static int unwrap_merge(void *arg)
+{
+       struct dma_fence *fence, *f1, *f2, *f3;
+       struct dma_fence_unwrap iter;
+       int err = 0;
+
+       f1 = mock_fence();
+       if (!f1)
+               return -ENOMEM;
+
+       f2 = mock_fence();
+       if (!f2) {
+               err = -ENOMEM;
+               goto error_put_f1;
+       }
+
+       f3 = dma_fence_unwrap_merge(f1, f2);
+       if (!f3) {
+               err = -ENOMEM;
+               goto error_put_f2;
+       }
+
+       dma_fence_unwrap_for_each(fence, &iter, f3) {
+               if (fence == f1) {
+                       dma_fence_put(f1);
+                       f1 = NULL;
+               } else if (fence == f2) {
+                       dma_fence_put(f2);
+                       f2 = NULL;
+               } else {
+                       pr_err("Unexpected fence!\n");
+                       err = -EINVAL;
+               }
+       }
+
+       if (f1 || f2) {
+               pr_err("Not all fences seen!\n");
+               err = -EINVAL;
+       }
+
+       dma_fence_put(f3);
+error_put_f2:
+       dma_fence_put(f2);
+error_put_f1:
+       dma_fence_put(f1);
+       return err;
+}
+
+static int unwrap_merge_complex(void *arg)
+{
+       struct dma_fence *fence, *f1, *f2, *f3, *f4, *f5;
+       struct dma_fence_unwrap iter;
+       int err = -ENOMEM;
+
+       f1 = mock_fence();
+       if (!f1)
+               return -ENOMEM;
+
+       f2 = mock_fence();
+       if (!f2)
+               goto error_put_f1;
+
+       f3 = dma_fence_unwrap_merge(f1, f2);
+       if (!f3)
+               goto error_put_f2;
+
+       /* The resulting array has the fences in reverse */
+       f4 = dma_fence_unwrap_merge(f2, f1);
+       if (!f4)
+               goto error_put_f3;
+
+       /* Signaled fences should be filtered, the two arrays merged. */
+       f5 = dma_fence_unwrap_merge(f3, f4, dma_fence_get_stub());
+       if (!f5)
+               goto error_put_f4;
+
+       err = 0;
+       dma_fence_unwrap_for_each(fence, &iter, f5) {
+               if (fence == f1) {
+                       dma_fence_put(f1);
+                       f1 = NULL;
+               } else if (fence == f2) {
+                       dma_fence_put(f2);
+                       f2 = NULL;
+               } else {
+                       pr_err("Unexpected fence!\n");
+                       err = -EINVAL;
+               }
+       }
+
+       if (f1 || f2) {
+               pr_err("Not all fences seen!\n");
+               err = -EINVAL;
+       }
+
+       dma_fence_put(f5);
+error_put_f4:
+       dma_fence_put(f4);
+error_put_f3:
+       dma_fence_put(f3);
+error_put_f2:
+       dma_fence_put(f2);
+error_put_f1:
+       dma_fence_put(f1);
+       return err;
+}
+
 int dma_fence_unwrap(void)
 {
        static const struct subtest tests[] = {
@@ -245,6 +352,8 @@ int dma_fence_unwrap(void)
                SUBTEST(unwrap_array),
                SUBTEST(unwrap_chain),
                SUBTEST(unwrap_chain_array),
+               SUBTEST(unwrap_merge),
+               SUBTEST(unwrap_merge_complex),
        };
 
        return subtests(tests, NULL);
index 0fe56453916673be4a06cb2b11e6023c845f288c..3ebec19a8e0296fb6c17ecc99424800ad5b8a05a 100644 (file)
@@ -146,50 +146,6 @@ char *sync_file_get_name(struct sync_file *sync_file, char *buf, int len)
        return buf;
 }
 
-static int sync_file_set_fence(struct sync_file *sync_file,
-                              struct dma_fence **fences, int num_fences)
-{
-       struct dma_fence_array *array;
-
-       /*
-        * The reference for the fences in the new sync_file and held
-        * in add_fence() during the merge procedure, so for num_fences == 1
-        * we already own a new reference to the fence. For num_fence > 1
-        * we own the reference of the dma_fence_array creation.
-        */
-
-       if (num_fences == 0) {
-               sync_file->fence = dma_fence_get_stub();
-               kfree(fences);
-
-       } else if (num_fences == 1) {
-               sync_file->fence = fences[0];
-               kfree(fences);
-
-       } else {
-               array = dma_fence_array_create(num_fences, fences,
-                                              dma_fence_context_alloc(1),
-                                              1, false);
-               if (!array)
-                       return -ENOMEM;
-
-               sync_file->fence = &array->base;
-       }
-
-       return 0;
-}
-
-static void add_fence(struct dma_fence **fences,
-                     int *i, struct dma_fence *fence)
-{
-       fences[*i] = fence;
-
-       if (!dma_fence_is_signaled(fence)) {
-               dma_fence_get(fence);
-               (*i)++;
-       }
-}
-
 /**
  * sync_file_merge() - merge two sync_files
  * @name:      name of new fence
@@ -203,84 +159,21 @@ static void add_fence(struct dma_fence **fences,
 static struct sync_file *sync_file_merge(const char *name, struct sync_file *a,
                                         struct sync_file *b)
 {
-       struct dma_fence *a_fence, *b_fence, **fences;
-       struct dma_fence_unwrap a_iter, b_iter;
-       unsigned int index, num_fences;
        struct sync_file *sync_file;
+       struct dma_fence *fence;
 
        sync_file = sync_file_alloc();
        if (!sync_file)
                return NULL;
 
-       num_fences = 0;
-       dma_fence_unwrap_for_each(a_fence, &a_iter, a->fence)
-               ++num_fences;
-       dma_fence_unwrap_for_each(b_fence, &b_iter, b->fence)
-               ++num_fences;
-
-       if (num_fences > INT_MAX)
-               goto err_free_sync_file;
-
-       fences = kcalloc(num_fences, sizeof(*fences), GFP_KERNEL);
-       if (!fences)
-               goto err_free_sync_file;
-
-       /*
-        * We can't guarantee that fences in both a and b are ordered, but it is
-        * still quite likely.
-        *
-        * So attempt to order the fences as we pass over them and merge fences
-        * with the same context.
-        */
-
-       index = 0;
-       for (a_fence = dma_fence_unwrap_first(a->fence, &a_iter),
-            b_fence = dma_fence_unwrap_first(b->fence, &b_iter);
-            a_fence || b_fence; ) {
-
-               if (!b_fence) {
-                       add_fence(fences, &index, a_fence);
-                       a_fence = dma_fence_unwrap_next(&a_iter);
-
-               } else if (!a_fence) {
-                       add_fence(fences, &index, b_fence);
-                       b_fence = dma_fence_unwrap_next(&b_iter);
-
-               } else if (a_fence->context < b_fence->context) {
-                       add_fence(fences, &index, a_fence);
-                       a_fence = dma_fence_unwrap_next(&a_iter);
-
-               } else if (b_fence->context < a_fence->context) {
-                       add_fence(fences, &index, b_fence);
-                       b_fence = dma_fence_unwrap_next(&b_iter);
-
-               } else if (__dma_fence_is_later(a_fence->seqno, b_fence->seqno,
-                                               a_fence->ops)) {
-                       add_fence(fences, &index, a_fence);
-                       a_fence = dma_fence_unwrap_next(&a_iter);
-                       b_fence = dma_fence_unwrap_next(&b_iter);
-
-               } else {
-                       add_fence(fences, &index, b_fence);
-                       a_fence = dma_fence_unwrap_next(&a_iter);
-                       b_fence = dma_fence_unwrap_next(&b_iter);
-               }
+       fence = dma_fence_unwrap_merge(a->fence, b->fence);
+       if (!fence) {
+               fput(sync_file->file);
+               return NULL;
        }
-
-       if (sync_file_set_fence(sync_file, fences, index) < 0)
-               goto err_put_fences;
-
+       sync_file->fence = fence;
        strlcpy(sync_file->user_name, name, sizeof(sync_file->user_name));
        return sync_file;
-
-err_put_fences:
-       while (index)
-               dma_fence_put(fences[--index]);
-       kfree(fences);
-
-err_free_sync_file:
-       fput(sync_file->file);
-       return NULL;
 }
 
 static int sync_file_release(struct inode *inode, struct file *file)
index a4d342fef8e0940b241fedefc8ded38abdff446c..390de1ee9d3539ee00bb1281869e1a6b6cd208d3 100644 (file)
@@ -52,4 +52,28 @@ struct dma_fence *dma_fence_unwrap_next(struct dma_fence_unwrap *cursor);
             fence = dma_fence_unwrap_next(cursor))                     \
                if (!dma_fence_is_signaled(fence))
 
+struct dma_fence *__dma_fence_unwrap_merge(unsigned int num_fences,
+                                          struct dma_fence **fences,
+                                          struct dma_fence_unwrap *cursors);
+
+/**
+ * dma_fence_unwrap_merge - unwrap and merge fences
+ *
+ * All fences given as parameters are unwrapped and merged back together as flat
+ * dma_fence_array. Useful if multiple containers need to be merged together.
+ *
+ * Implemented as a macro to allocate the necessary arrays on the stack and
+ * account the stack frame size to the caller.
+ *
+ * Returns NULL on memory allocation failure, a dma_fence object representing
+ * all the given fences otherwise.
+ */
+#define dma_fence_unwrap_merge(...)                                    \
+       ({                                                              \
+               struct dma_fence *__f[] = { __VA_ARGS__ };              \
+               struct dma_fence_unwrap __c[ARRAY_SIZE(__f)];           \
+                                                                       \
+               __dma_fence_unwrap_merge(ARRAY_SIZE(__f), __f, __c);    \
+       })
+
 #endif