]> git.baikalelectronics.ru Git - kernel.git/commitdiff
IB/core: Add uverbs merge trees functionality
authorMatan Barak <matanb@mellanox.com>
Thu, 3 Aug 2017 13:07:00 +0000 (16:07 +0300)
committerDoug Ledford <dledford@redhat.com>
Thu, 31 Aug 2017 12:35:10 +0000 (08:35 -0400)
Different drivers support different features and even subset of the
common uverbs implementation. Currently, this is handled as bitmask
in every driver that represents which kind of methods it supports, but
doesn't go down to attributes granularity. Moreover, drivers might
want to add their specific types, methods and attributes to let
their user-space counter-parts be exposed to some more efficient
abstractions. It means that existence of different features is
validated syntactically via the parsing infrastructure rather than
using a complex in-handler logic.

In order to do that, we allow defining features and abstractions
as parsing trees. These per-feature parsing tree could be merged
to an efficient (perfect-hash based) parsing tree, which is later
used by the parsing infrastructure.

To sum it up, this makes a parse tree unique for a device and
represents only the features this particular device supports.
This is done by having a root specification tree per feature.
Before a device registers itself as an IB device, it merges
all these trees into one parsing tree. This parsing tree
is used to parse all user-space commands.

A future user-space application could read this parse tree. This
tree represents which objects, methods and attributes are
supported by this device.

This is based on the idea of
Jason Gunthorpe <jgunthorpe@obsidianresearch.com>

Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/core/Makefile
drivers/infiniband/core/uverbs_ioctl_merge.c [new file with mode: 0644]
include/rdma/uverbs_ioctl.h

index 746756dc98774e7e01bcc95fb446ecf7b0537ea7..b4df164f71a6af7c086e2a8df961e7c328e2909c 100644 (file)
@@ -32,4 +32,5 @@ ib_umad-y :=                  user_mad.o
 ib_ucm-y :=                    ucm.o
 
 ib_uverbs-y :=                 uverbs_main.o uverbs_cmd.o uverbs_marshall.o \
-                               rdma_core.o uverbs_std_types.o uverbs_ioctl.o
+                               rdma_core.o uverbs_std_types.o uverbs_ioctl.o \
+                               uverbs_ioctl_merge.o
diff --git a/drivers/infiniband/core/uverbs_ioctl_merge.c b/drivers/infiniband/core/uverbs_ioctl_merge.c
new file mode 100644 (file)
index 0000000..76ddb65
--- /dev/null
@@ -0,0 +1,665 @@
+/*
+ * Copyright (c) 2017, Mellanox Technologies inc.  All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <rdma/uverbs_ioctl.h>
+#include <rdma/rdma_user_ioctl.h>
+#include <linux/bitops.h>
+#include "uverbs.h"
+
+#define UVERBS_NUM_NS (UVERBS_ID_NS_MASK >> UVERBS_ID_NS_SHIFT)
+#define GET_NS_ID(idx) (((idx) & UVERBS_ID_NS_MASK) >> UVERBS_ID_NS_SHIFT)
+#define GET_ID(idx) ((idx) & ~UVERBS_ID_NS_MASK)
+
+#define _for_each_element(elem, tmpi, tmpj, hashes, num_buckets_offset,               \
+                         buckets_offset)                                      \
+       for (tmpj = 0,                                                         \
+            elem = (*(const void ***)((hashes)[tmpi] +                        \
+                                      (buckets_offset)))[0];                  \
+            tmpj < *(size_t *)((hashes)[tmpi] + (num_buckets_offset));        \
+            tmpj++)                                                           \
+               if ((elem = ((*(const void ***)(hashes[tmpi] +                 \
+                                               (buckets_offset)))[tmpj])))
+
+/*
+ * Iterate all elements of a few @hashes. The number of given hashes is
+ * indicated by @num_hashes. The offset of the number of buckets in the hash is
+ * represented by @num_buckets_offset, while the offset of the buckets array in
+ * the hash structure is represented by @buckets_offset. tmpi and tmpj are two
+ * short (or int) based indices that are given by the user. tmpi iterates over
+ * the different hashes. @elem points the current element in the hashes[tmpi]
+ * bucket we are looping on. To be honest, @hashes representation isn't exactly
+ * a hash, but more a collection of elements. These elements' ids are treated
+ * in a hash like manner, where the first upper bits are the bucket number.
+ * These elements are later mapped into a perfect-hash.
+ */
+#define for_each_element(elem, tmpi, tmpj, hashes, num_hashes,                 \
+                        num_buckets_offset, buckets_offset)                   \
+       for (tmpi = 0; tmpi < (num_hashes); tmpi++)                            \
+               _for_each_element(elem, tmpi, tmpj, hashes, num_buckets_offset,\
+                                 buckets_offset)
+
+#define get_elements_iterators_entry_above(iters, num_elements, elements,     \
+                                         num_objects_fld, objects_fld, bucket,\
+                                         min_id)                              \
+       get_elements_above_id((const void **)iters, num_elements,       \
+                                    (const void **)(elements),                \
+                                    offsetof(typeof(**elements),              \
+                                             num_objects_fld),                \
+                                    offsetof(typeof(**elements), objects_fld),\
+                                    offsetof(typeof(***(*elements)->objects_fld), id),\
+                                    bucket, min_id)
+
+#define get_objects_above_id(iters, num_trees, trees, bucket, min_id)         \
+       get_elements_iterators_entry_above(iters, num_trees, trees,            \
+                                          num_objects, objects, bucket, min_id)
+
+#define get_methods_above_id(method_iters, num_iters, iters, bucket, min_id)\
+       get_elements_iterators_entry_above(method_iters, num_iters, iters,     \
+                                          num_methods, methods, bucket, min_id)
+
+#define get_attrs_above_id(attrs_iters, num_iters, iters, bucket, min_id)\
+       get_elements_iterators_entry_above(attrs_iters, num_iters, iters,      \
+                                          num_attrs, attrs, bucket, min_id)
+
+/*
+ * get_elements_above_id get a few hashes represented by @elements and
+ * @num_elements. The hashes fields are described by @num_offset, @data_offset
+ * and @id_offset in the same way as required by for_each_element. The function
+ * returns an array of @iters, represents an array of elements in the hashes
+ * buckets, which their ids are the smallest ids in all hashes but are all
+ * larger than the id given by min_id. Elements are only added to the iters
+ * array if their id belongs to the bucket @bucket. The number of elements in
+ * the returned array is returned by the function. @min_id is also updated to
+ * reflect the new min_id of all elements in iters.
+ */
+static size_t get_elements_above_id(const void **iters,
+                                   unsigned int num_elements,
+                                   const void **elements,
+                                   size_t num_offset,
+                                   size_t data_offset,
+                                   size_t id_offset,
+                                   u16 bucket,
+                                   short *min_id)
+{
+       size_t num_iters = 0;
+       short min = SHRT_MAX;
+       const void *elem;
+       int i, j, last_stored = -1;
+
+       for_each_element(elem, i, j, elements, num_elements, num_offset,
+                        data_offset) {
+               u16 id = *(u16 *)(elem + id_offset);
+
+               if (GET_NS_ID(id) != bucket)
+                       continue;
+
+               if (GET_ID(id) < *min_id ||
+                   (min != SHRT_MAX && GET_ID(id) > min))
+                       continue;
+
+               /*
+                * We first iterate all hashes represented by @elements. When
+                * we do, we try to find an element @elem in the bucket @bucket
+                * which its id is min. Since we can't ensure the user sorted
+                * the elements in increasing order, we override this hash's
+                * minimal id element we found, if a new element with a smaller
+                * id was just found.
+                */
+               iters[last_stored == i ? num_iters - 1 : num_iters++] = elem;
+               last_stored = i;
+               min = GET_ID(id);
+       }
+
+       /*
+        * We only insert to our iters array an element, if its id is smaller
+        * than all previous ids. Therefore, the final iters array is sorted so
+        * that smaller ids are in the end of the array.
+        * Therefore, we need to clean the beginning of the array to make sure
+        * all ids of final elements are equal to min.
+        */
+       for (i = num_iters - 1; i >= 0 &&
+            GET_ID(*(u16 *)(iters[i] + id_offset)) == min; i--)
+               ;
+
+       num_iters -= i + 1;
+       memmove(iters, iters + i + 1, sizeof(*iters) * num_iters);
+
+       *min_id = min;
+       return num_iters;
+}
+
+#define find_max_element_entry_id(num_elements, elements, num_objects_fld, \
+                                 objects_fld, bucket)                     \
+       find_max_element_id(num_elements, (const void **)(elements),       \
+                           offsetof(typeof(**elements), num_objects_fld),    \
+                           offsetof(typeof(**elements), objects_fld),        \
+                           offsetof(typeof(***(*elements)->objects_fld), id),\
+                           bucket)
+
+static short find_max_element_ns_id(unsigned int num_elements,
+                                   const void **elements,
+                                   size_t num_offset,
+                                   size_t data_offset,
+                                   size_t id_offset)
+{
+       short max_ns = SHRT_MIN;
+       const void *elem;
+       int i, j;
+
+       for_each_element(elem, i, j, elements, num_elements, num_offset,
+                        data_offset) {
+               u16 id = *(u16 *)(elem + id_offset);
+
+               if (GET_NS_ID(id) > max_ns)
+                       max_ns = GET_NS_ID(id);
+       }
+
+       return max_ns;
+}
+
+static short find_max_element_id(unsigned int num_elements,
+                                const void **elements,
+                                size_t num_offset,
+                                size_t data_offset,
+                                size_t id_offset,
+                                u16 bucket)
+{
+       short max_id = SHRT_MIN;
+       const void *elem;
+       int i, j;
+
+       for_each_element(elem, i, j, elements, num_elements, num_offset,
+                        data_offset) {
+               u16 id = *(u16 *)(elem + id_offset);
+
+               if (GET_NS_ID(id) == bucket &&
+                   GET_ID(id) > max_id)
+                       max_id = GET_ID(id);
+       }
+       return max_id;
+}
+
+#define find_max_element_entry_id(num_elements, elements, num_objects_fld,   \
+                                 objects_fld, bucket)                        \
+       find_max_element_id(num_elements, (const void **)(elements),          \
+                           offsetof(typeof(**elements), num_objects_fld),    \
+                           offsetof(typeof(**elements), objects_fld),        \
+                           offsetof(typeof(***(*elements)->objects_fld), id),\
+                           bucket)
+
+#define find_max_element_ns_entry_id(num_elements, elements,               \
+                                    num_objects_fld, objects_fld)          \
+       find_max_element_ns_id(num_elements, (const void **)(elements),     \
+                             offsetof(typeof(**elements), num_objects_fld),\
+                             offsetof(typeof(**elements), objects_fld),    \
+                             offsetof(typeof(***(*elements)->objects_fld), id))
+
+/*
+ * find_max_xxxx_ns_id gets a few elements. Each element is described by an id
+ * which its upper bits represents a namespace. It finds the max namespace. This
+ * could be used in order to know how many buckets do we need to allocate. If no
+ * elements exist, SHRT_MIN is returned. Namespace represents here different
+ * buckets. The common example is "common bucket" and "driver bucket".
+ *
+ * find_max_xxxx_id gets a few elements and a bucket. Each element is described
+ * by an id which its upper bits represent a namespace. It returns the max id
+ * which is contained in the same namespace defined in @bucket. This could be
+ * used in order to know how many elements do we need to allocate in the bucket.
+ * If no elements exist, SHRT_MIN is returned.
+ */
+
+#define find_max_object_id(num_trees, trees, bucket)                   \
+               find_max_element_entry_id(num_trees, trees, num_objects,\
+                                         objects, bucket)
+#define find_max_object_ns_id(num_trees, trees)                        \
+               find_max_element_ns_entry_id(num_trees, trees,          \
+                                            num_objects, objects)
+
+#define find_max_method_id(num_iters, iters, bucket)                   \
+               find_max_element_entry_id(num_iters, iters, num_methods,\
+                                         methods, bucket)
+#define find_max_method_ns_id(num_iters, iters)                        \
+               find_max_element_ns_entry_id(num_iters, iters,          \
+                                            num_methods, methods)
+
+#define find_max_attr_id(num_iters, iters, bucket)                     \
+               find_max_element_entry_id(num_iters, iters, num_attrs,  \
+                                         attrs, bucket)
+#define find_max_attr_ns_id(num_iters, iters)                          \
+               find_max_element_ns_entry_id(num_iters, iters,          \
+                                            num_attrs, attrs)
+
+static void free_method(struct uverbs_method_spec *method)
+{
+       unsigned int i;
+
+       if (!method)
+               return;
+
+       for (i = 0; i < method->num_buckets; i++)
+               kfree(method->attr_buckets[i]);
+
+       kfree(method);
+}
+
+#define IS_ATTR_OBJECT(attr) ((attr)->type == UVERBS_ATTR_TYPE_IDR || \
+                             (attr)->type == UVERBS_ATTR_TYPE_FD)
+
+/*
+ * This function gets array of size @num_method_defs which contains pointers to
+ * method definitions @method_defs. The function allocates an
+ * uverbs_method_spec structure and initializes its number of buckets and the
+ * elements in buckets to the correct attributes. While doing that, it
+ * validates that there aren't conflicts between attributes of different
+ * method_defs.
+ */
+static struct uverbs_method_spec *build_method_with_attrs(const struct uverbs_method_def **method_defs,
+                                                         size_t num_method_defs)
+{
+       int bucket_idx;
+       int max_attr_buckets = 0;
+       size_t num_attr_buckets = 0;
+       int res = 0;
+       struct uverbs_method_spec *method = NULL;
+       const struct uverbs_attr_def **attr_defs;
+       unsigned int num_of_singularities = 0;
+
+       max_attr_buckets = find_max_attr_ns_id(num_method_defs, method_defs);
+       if (max_attr_buckets >= 0)
+               num_attr_buckets = max_attr_buckets + 1;
+
+       method = kzalloc(sizeof(*method) +
+                        num_attr_buckets * sizeof(*method->attr_buckets),
+                        GFP_KERNEL);
+       if (!method)
+               return ERR_PTR(-ENOMEM);
+
+       method->num_buckets = num_attr_buckets;
+       attr_defs = kcalloc(num_method_defs, sizeof(*attr_defs), GFP_KERNEL);
+       if (!attr_defs) {
+               res = -ENOMEM;
+               goto free_method;
+       }
+       for (bucket_idx = 0; bucket_idx < method->num_buckets; bucket_idx++) {
+               short min_id = SHRT_MIN;
+               int attr_max_bucket = 0;
+               struct uverbs_attr_spec_hash *hash = NULL;
+
+               attr_max_bucket = find_max_attr_id(num_method_defs, method_defs,
+                                                  bucket_idx);
+               if (attr_max_bucket < 0)
+                       continue;
+
+               hash = kzalloc(sizeof(*hash) +
+                              ALIGN(sizeof(*hash->attrs) * (attr_max_bucket + 1),
+                                    sizeof(long)) +
+                              BITS_TO_LONGS(attr_max_bucket) * sizeof(long),
+                              GFP_KERNEL);
+               if (!hash) {
+                       res = -ENOMEM;
+                       goto free;
+               }
+               hash->num_attrs = attr_max_bucket + 1;
+               method->num_child_attrs += hash->num_attrs;
+               hash->mandatory_attrs_bitmask = (void *)(hash + 1) +
+                                                ALIGN(sizeof(*hash->attrs) *
+                                                      (attr_max_bucket + 1),
+                                                      sizeof(long));
+
+               method->attr_buckets[bucket_idx] = hash;
+
+               do {
+                       size_t                   num_attr_defs;
+                       struct uverbs_attr_spec *attr;
+                       bool attr_obj_with_special_access;
+
+                       num_attr_defs =
+                               get_attrs_above_id(attr_defs,
+                                                  num_method_defs,
+                                                  method_defs,
+                                                  bucket_idx,
+                                                  &min_id);
+                       /* Last attr in bucket */
+                       if (!num_attr_defs)
+                               break;
+
+                       if (num_attr_defs > 1) {
+                               /*
+                                * We don't allow two attribute definitions for
+                                * the same attribute. This is usually a
+                                * programmer error. If required, it's better to
+                                * just add a new attribute to capture the new
+                                * semantics.
+                                */
+                               res = -EEXIST;
+                               goto free;
+                       }
+
+                       attr = &hash->attrs[min_id];
+                       memcpy(attr, &attr_defs[0]->attr, sizeof(*attr));
+
+                       attr_obj_with_special_access = IS_ATTR_OBJECT(attr) &&
+                                  (attr->obj.access == UVERBS_ACCESS_NEW ||
+                                   attr->obj.access == UVERBS_ACCESS_DESTROY);
+                       num_of_singularities +=  !!attr_obj_with_special_access;
+                       if (WARN(num_of_singularities > 1,
+                                "ib_uverbs: Method contains more than one object attr (%d) with new/destroy access\n",
+                                min_id) ||
+                           WARN(attr_obj_with_special_access &&
+                                !(attr->flags & UVERBS_ATTR_SPEC_F_MANDATORY),
+                                "ib_uverbs: Tried to merge attr (%d) but it's an object with new/destroy aceess but isn't mandatory\n",
+                                min_id) ||
+                           WARN(IS_ATTR_OBJECT(attr) &&
+                                attr->flags & UVERBS_ATTR_SPEC_F_MIN_SZ,
+                                "ib_uverbs: Tried to merge attr (%d) but it's an object with min_sz flag\n",
+                                min_id)) {
+                               res = -EINVAL;
+                               goto free;
+                       }
+
+                       if (attr->flags & UVERBS_ATTR_SPEC_F_MANDATORY)
+                               set_bit(min_id, hash->mandatory_attrs_bitmask);
+                       min_id++;
+
+               } while (1);
+       }
+       kfree(attr_defs);
+       return method;
+
+free:
+       kfree(attr_defs);
+free_method:
+       free_method(method);
+       return ERR_PTR(res);
+}
+
+static void free_object(struct uverbs_object_spec *object)
+{
+       unsigned int i, j;
+
+       if (!object)
+               return;
+
+       for (i = 0; i < object->num_buckets; i++) {
+               struct uverbs_method_spec_hash  *method_buckets =
+                       object->method_buckets[i];
+
+               if (!method_buckets)
+                       continue;
+
+               for (j = 0; j < method_buckets->num_methods; j++)
+                       free_method(method_buckets->methods[j]);
+
+               kfree(method_buckets);
+       }
+
+       kfree(object);
+}
+
+/*
+ * This function gets array of size @num_object_defs which contains pointers to
+ * object definitions @object_defs. The function allocated an
+ * uverbs_object_spec structure and initialize its number of buckets and the
+ * elements in buckets to the correct methods. While doing that, it
+ * sorts out the correct relationship between conflicts in the same method.
+ */
+static struct uverbs_object_spec *build_object_with_methods(const struct uverbs_object_def **object_defs,
+                                                           size_t num_object_defs)
+{
+       u16 bucket_idx;
+       int max_method_buckets = 0;
+       u16 num_method_buckets = 0;
+       int res = 0;
+       struct uverbs_object_spec *object = NULL;
+       const struct uverbs_method_def **method_defs;
+
+       max_method_buckets = find_max_method_ns_id(num_object_defs, object_defs);
+       if (max_method_buckets >= 0)
+               num_method_buckets = max_method_buckets + 1;
+
+       object = kzalloc(sizeof(*object) +
+                        num_method_buckets *
+                        sizeof(*object->method_buckets), GFP_KERNEL);
+       if (!object)
+               return ERR_PTR(-ENOMEM);
+
+       object->num_buckets = num_method_buckets;
+       method_defs = kcalloc(num_object_defs, sizeof(*method_defs), GFP_KERNEL);
+       if (!method_defs) {
+               res = -ENOMEM;
+               goto free_object;
+       }
+
+       for (bucket_idx = 0; bucket_idx < object->num_buckets; bucket_idx++) {
+               short min_id = SHRT_MIN;
+               int methods_max_bucket = 0;
+               struct uverbs_method_spec_hash *hash = NULL;
+
+               methods_max_bucket = find_max_method_id(num_object_defs, object_defs,
+                                                       bucket_idx);
+               if (methods_max_bucket < 0)
+                       continue;
+
+               hash = kzalloc(sizeof(*hash) +
+                              sizeof(*hash->methods) * (methods_max_bucket + 1),
+                              GFP_KERNEL);
+               if (!hash) {
+                       res = -ENOMEM;
+                       goto free;
+               }
+
+               hash->num_methods = methods_max_bucket + 1;
+               object->method_buckets[bucket_idx] = hash;
+
+               do {
+                       size_t                          num_method_defs;
+                       struct uverbs_method_spec       *method;
+                       int i;
+
+                       num_method_defs =
+                               get_methods_above_id(method_defs,
+                                                    num_object_defs,
+                                                    object_defs,
+                                                    bucket_idx,
+                                                    &min_id);
+                       /* Last method in bucket */
+                       if (!num_method_defs)
+                               break;
+
+                       method = build_method_with_attrs(method_defs,
+                                                        num_method_defs);
+                       if (IS_ERR(method)) {
+                               res = PTR_ERR(method);
+                               goto free;
+                       }
+
+                       /*
+                        * The last tree which is given as an argument to the
+                        * merge overrides previous method handler.
+                        * Therefore, we iterate backwards and search for the
+                        * first handler which != NULL. This also defines the
+                        * set of flags used for this handler.
+                        */
+                       for (i = num_object_defs - 1;
+                            i >= 0 && !method_defs[i]->handler; i--)
+                               ;
+                       hash->methods[min_id++] = method;
+                       /* NULL handler isn't allowed */
+                       if (WARN(i < 0,
+                                "ib_uverbs: tried to merge function id %d, but all handlers are NULL\n",
+                                min_id)) {
+                               res = -EINVAL;
+                               goto free;
+                       }
+                       method->handler = method_defs[i]->handler;
+                       method->flags = method_defs[i]->flags;
+
+               } while (1);
+       }
+       kfree(method_defs);
+       return object;
+
+free:
+       kfree(method_defs);
+free_object:
+       free_object(object);
+       return ERR_PTR(res);
+}
+
+void uverbs_free_spec_tree(struct uverbs_root_spec *root)
+{
+       unsigned int i, j;
+
+       if (!root)
+               return;
+
+       for (i = 0; i < root->num_buckets; i++) {
+               struct uverbs_object_spec_hash *object_hash =
+                       root->object_buckets[i];
+
+               if (!object_hash)
+                       continue;
+
+               for (j = 0; j < object_hash->num_objects; j++)
+                       free_object(object_hash->objects[j]);
+
+               kfree(object_hash);
+       }
+
+       kfree(root);
+}
+EXPORT_SYMBOL(uverbs_free_spec_tree);
+
+struct uverbs_root_spec *uverbs_alloc_spec_tree(unsigned int num_trees,
+                                               const struct uverbs_object_tree_def **trees)
+{
+       u16 bucket_idx;
+       short max_object_buckets = 0;
+       size_t num_objects_buckets = 0;
+       struct uverbs_root_spec *root_spec = NULL;
+       const struct uverbs_object_def **object_defs;
+       int i;
+       int res = 0;
+
+       max_object_buckets = find_max_object_ns_id(num_trees, trees);
+       /*
+        * Devices which don't want to support ib_uverbs, should just allocate
+        * an empty parsing tree. Every user-space command won't hit any valid
+        * entry in the parsing tree and thus will fail.
+        */
+       if (max_object_buckets >= 0)
+               num_objects_buckets = max_object_buckets + 1;
+
+       root_spec = kzalloc(sizeof(*root_spec) +
+                           num_objects_buckets * sizeof(*root_spec->object_buckets),
+                           GFP_KERNEL);
+       if (!root_spec)
+               return ERR_PTR(-ENOMEM);
+       root_spec->num_buckets = num_objects_buckets;
+
+       object_defs = kcalloc(num_trees, sizeof(*object_defs),
+                             GFP_KERNEL);
+       if (!object_defs) {
+               res = -ENOMEM;
+               goto free_root;
+       }
+
+       for (bucket_idx = 0; bucket_idx < root_spec->num_buckets; bucket_idx++) {
+               short min_id = SHRT_MIN;
+               short objects_max_bucket;
+               struct uverbs_object_spec_hash *hash = NULL;
+
+               objects_max_bucket = find_max_object_id(num_trees, trees,
+                                                       bucket_idx);
+               if (objects_max_bucket < 0)
+                       continue;
+
+               hash = kzalloc(sizeof(*hash) +
+                              sizeof(*hash->objects) * (objects_max_bucket + 1),
+                              GFP_KERNEL);
+               if (!hash) {
+                       res = -ENOMEM;
+                       goto free;
+               }
+               hash->num_objects = objects_max_bucket + 1;
+               root_spec->object_buckets[bucket_idx] = hash;
+
+               do {
+                       size_t                          num_object_defs;
+                       struct uverbs_object_spec       *object;
+
+                       num_object_defs = get_objects_above_id(object_defs,
+                                                              num_trees,
+                                                              trees,
+                                                              bucket_idx,
+                                                              &min_id);
+                       /* Last object in bucket */
+                       if (!num_object_defs)
+                               break;
+
+                       object = build_object_with_methods(object_defs,
+                                                          num_object_defs);
+                       if (IS_ERR(object)) {
+                               res = PTR_ERR(object);
+                               goto free;
+                       }
+
+                       /*
+                        * The last tree which is given as an argument to the
+                        * merge overrides previous object's type_attrs.
+                        * Therefore, we iterate backwards and search for the
+                        * first type_attrs which != NULL.
+                        */
+                       for (i = num_object_defs - 1;
+                            i >= 0 && !object_defs[i]->type_attrs; i--)
+                               ;
+                       /*
+                        * NULL is a valid type_attrs. It means an object we
+                        * can't instantiate (like DEVICE).
+                        */
+                       object->type_attrs = i < 0 ? NULL :
+                               object_defs[i]->type_attrs;
+
+                       hash->objects[min_id++] = object;
+               } while (1);
+       }
+
+       kfree(object_defs);
+       return root_spec;
+
+free:
+       kfree(object_defs);
+free_root:
+       uverbs_free_spec_tree(root_spec);
+       return ERR_PTR(res);
+}
+EXPORT_SYMBOL(uverbs_alloc_spec_tree);
index 2e8925434d742c654006e637a82ff02c61405e80..cf5b238d2d8167c6de2c7185b254f4c5dfec983f 100644 (file)
@@ -235,5 +235,43 @@ static inline bool uverbs_attr_is_valid_in_hash(const struct uverbs_attr_bundle_
        return test_bit(idx, attrs_hash->valid_bitmap);
 }
 
-#endif
+/* =================================================
+ *      Definitions -> Specs infrastructure
+ * =================================================
+ */
+
+/*
+ * uverbs_alloc_spec_tree - Merges different common and driver specific feature
+ *     into one parsing tree that every uverbs command will be parsed upon.
+ *
+ * @num_trees: Number of trees in the array @trees.
+ * @trees: Array of pointers to tree root definitions to merge. Each such tree
+ *        possibly contains objects, methods and attributes definitions.
+ *
+ * Returns:
+ *     uverbs_root_spec *: The root of the merged parsing tree.
+ *     On error, we return an error code. Error is checked via IS_ERR.
+ *
+ * The following merges could take place:
+ * a. Two trees representing the same method with different handler
+ *     -> We take the handler of the tree that its handler != NULL
+ *        and its index in the trees array is greater. The incentive for that
+ *        is that developers are expected to first merge common trees and then
+ *        merge trees that gives specialized the behaviour.
+ * b. Two trees representing the same object with different
+ *    type_attrs (struct uverbs_obj_type):
+ *     -> We take the type_attrs of the tree that its type_attr != NULL
+ *        and its index in the trees array is greater. This could be used
+ *        in order to override the free function, allocation size, etc.
+ * c. Two trees representing the same method attribute (same id but possibly
+ *    different attributes):
+ *     -> ERROR (-ENOENT), we believe that's not the programmer's intent.
+ *
+ * An object without any methods is considered invalid and will abort the
+ * function with -ENOENT error.
+ */
+struct uverbs_root_spec *uverbs_alloc_spec_tree(unsigned int num_trees,
+                                               const struct uverbs_object_tree_def **trees);
+void uverbs_free_spec_tree(struct uverbs_root_spec *root);
 
+#endif