Revert "block, dm: don't copy bios for request clones"
authorMike Snitzer <snitzer@redhat.com>
Fri, 26 Jun 2015 14:01:13 +0000 (10:01 -0400)
committerMike Snitzer <snitzer@redhat.com>
Fri, 26 Jun 2015 14:11:58 +0000 (10:11 -0400)
This reverts commit 5f1b670d0bef508a5554d92525f5f6d00d640b38.

Justification for revert as reported in this dm-devel post:
https://www.redhat.com/archives/dm-devel/2015-June/msg00160.html

this change should not be pushed to mainline yet.

Firstly, Christoph has a newer version of the patch that fixes silent
data corruption problem:
  https://www.redhat.com/archives/dm-devel/2015-May/msg00229.html

And the new version still depends on LLDDs to always complete requests
to the end when error happens, while block API doesn't enforce such a
requirement. If the assumption is ever broken, the inconsistency between
request and bio (e.g. rq->__sector and rq->bio) will cause silent data
corruption:
  https://www.redhat.com/archives/dm-devel/2015-June/msg00022.html

Reported-by: Junichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
block/blk-core.c
drivers/md/dm-table.c
drivers/md/dm.c
drivers/md/dm.h
include/linux/blk_types.h
include/linux/blkdev.h

index f6ab750060fe019f97d0ccfbca367b9e6cd3b426..706dd087022a961f4c2cfb0710957f1a33e096e5 100644 (file)
@@ -117,7 +117,7 @@ EXPORT_SYMBOL(blk_rq_init);
 static void req_bio_endio(struct request *rq, struct bio *bio,
                          unsigned int nbytes, int error)
 {
-       if (error && !(rq->cmd_flags & REQ_CLONE))
+       if (error)
                clear_bit(BIO_UPTODATE, &bio->bi_flags);
        else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
                error = -EIO;
@@ -128,8 +128,7 @@ static void req_bio_endio(struct request *rq, struct bio *bio,
        bio_advance(bio, nbytes);
 
        /* don't actually finish bio if it's part of flush sequence */
-       if (bio->bi_iter.bi_size == 0 &&
-           !(rq->cmd_flags & (REQ_FLUSH_SEQ|REQ_CLONE)))
+       if (bio->bi_iter.bi_size == 0 && !(rq->cmd_flags & REQ_FLUSH_SEQ))
                bio_endio(bio, error);
 }
 
@@ -2913,22 +2912,95 @@ int blk_lld_busy(struct request_queue *q)
 }
 EXPORT_SYMBOL_GPL(blk_lld_busy);
 
-void blk_rq_prep_clone(struct request *dst, struct request *src)
+/**
+ * blk_rq_unprep_clone - Helper function to free all bios in a cloned request
+ * @rq: the clone request to be cleaned up
+ *
+ * Description:
+ *     Free all bios in @rq for a cloned request.
+ */
+void blk_rq_unprep_clone(struct request *rq)
+{
+       struct bio *bio;
+
+       while ((bio = rq->bio) != NULL) {
+               rq->bio = bio->bi_next;
+
+               bio_put(bio);
+       }
+}
+EXPORT_SYMBOL_GPL(blk_rq_unprep_clone);
+
+/*
+ * Copy attributes of the original request to the clone request.
+ * The actual data parts (e.g. ->cmd, ->sense) are not copied.
+ */
+static void __blk_rq_prep_clone(struct request *dst, struct request *src)
 {
        dst->cpu = src->cpu;
-       dst->cmd_flags |= (src->cmd_flags & REQ_CLONE_MASK);
-       dst->cmd_flags |= REQ_NOMERGE | REQ_CLONE;
+       dst->cmd_flags |= (src->cmd_flags & REQ_CLONE_MASK) | REQ_NOMERGE;
        dst->cmd_type = src->cmd_type;
        dst->__sector = blk_rq_pos(src);
        dst->__data_len = blk_rq_bytes(src);
        dst->nr_phys_segments = src->nr_phys_segments;
        dst->ioprio = src->ioprio;
        dst->extra_len = src->extra_len;
-       dst->bio = src->bio;
-       dst->biotail = src->biotail;
-       dst->cmd = src->cmd;
-       dst->cmd_len = src->cmd_len;
-       dst->sense = src->sense;
+}
+
+/**
+ * blk_rq_prep_clone - Helper function to setup clone request
+ * @rq: the request to be setup
+ * @rq_src: original request to be cloned
+ * @bs: bio_set that bios for clone are allocated from
+ * @gfp_mask: memory allocation mask for bio
+ * @bio_ctr: setup function to be called for each clone bio.
+ *           Returns %0 for success, non %0 for failure.
+ * @data: private data to be passed to @bio_ctr
+ *
+ * Description:
+ *     Clones bios in @rq_src to @rq, and copies attributes of @rq_src to @rq.
+ *     The actual data parts of @rq_src (e.g. ->cmd, ->sense)
+ *     are not copied, and copying such parts is the caller's responsibility.
+ *     Also, pages which the original bios are pointing to are not copied
+ *     and the cloned bios just point same pages.
+ *     So cloned bios must be completed before original bios, which means
+ *     the caller must complete @rq before @rq_src.
+ */
+int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
+                     struct bio_set *bs, gfp_t gfp_mask,
+                     int (*bio_ctr)(struct bio *, struct bio *, void *),
+                     void *data)
+{
+       struct bio *bio, *bio_src;
+
+       if (!bs)
+               bs = fs_bio_set;
+
+       __rq_for_each_bio(bio_src, rq_src) {
+               bio = bio_clone_fast(bio_src, gfp_mask, bs);
+               if (!bio)
+                       goto free_and_out;
+
+               if (bio_ctr && bio_ctr(bio, bio_src, data))
+                       goto free_and_out;
+
+               if (rq->bio) {
+                       rq->biotail->bi_next = bio;
+                       rq->biotail = bio;
+               } else
+                       rq->bio = rq->biotail = bio;
+       }
+
+       __blk_rq_prep_clone(rq, rq_src);
+
+       return 0;
+
+free_and_out:
+       if (bio)
+               bio_put(bio);
+       blk_rq_unprep_clone(rq);
+
+       return -ENOMEM;
 }
 EXPORT_SYMBOL_GPL(blk_rq_prep_clone);
 
index a5f94125ad01f6b3a3a43fd7ae78e9bad0747995..16ba55ad708992f7e942b2f6ce2048d12be5c1b6 100644 (file)
@@ -942,28 +942,21 @@ static int dm_table_alloc_md_mempools(struct dm_table *t, struct mapped_device *
 {
        unsigned type = dm_table_get_type(t);
        unsigned per_bio_data_size = 0;
+       struct dm_target *tgt;
        unsigned i;
 
-       switch (type) {
-       case DM_TYPE_BIO_BASED:
-               for (i = 0; i < t->num_targets; i++) {
-                       struct dm_target *tgt = t->targets + i;
-
-                       per_bio_data_size = max(per_bio_data_size,
-                                               tgt->per_bio_data_size);
-               }
-               t->mempools = dm_alloc_bio_mempools(t->integrity_supported,
-                                                   per_bio_data_size);
-               break;
-       case DM_TYPE_REQUEST_BASED:
-       case DM_TYPE_MQ_REQUEST_BASED:
-               t->mempools = dm_alloc_rq_mempools(md, type);
-               break;
-       default:
+       if (unlikely(type == DM_TYPE_NONE)) {
                DMWARN("no table type is set, can't allocate mempools");
                return -EINVAL;
        }
 
+       if (type == DM_TYPE_BIO_BASED)
+               for (i = 0; i < t->num_targets; i++) {
+                       tgt = t->targets + i;
+                       per_bio_data_size = max(per_bio_data_size, tgt->per_bio_data_size);
+               }
+
+       t->mempools = dm_alloc_md_mempools(md, type, t->integrity_supported, per_bio_data_size);
        if (!t->mempools)
                return -ENOMEM;
 
index 492181e16c6945cbfb463c99bd3af55c27f122b6..9d942aef0f750c89f76afe5742501a4cdfb347fb 100644 (file)
@@ -993,6 +993,57 @@ static void clone_endio(struct bio *bio, int error)
        dec_pending(io, error);
 }
 
+/*
+ * Partial completion handling for request-based dm
+ */
+static void end_clone_bio(struct bio *clone, int error)
+{
+       struct dm_rq_clone_bio_info *info =
+               container_of(clone, struct dm_rq_clone_bio_info, clone);
+       struct dm_rq_target_io *tio = info->tio;
+       struct bio *bio = info->orig;
+       unsigned int nr_bytes = info->orig->bi_iter.bi_size;
+
+       bio_put(clone);
+
+       if (tio->error)
+               /*
+                * An error has already been detected on the request.
+                * Once error occurred, just let clone->end_io() handle
+                * the remainder.
+                */
+               return;
+       else if (error) {
+               /*
+                * Don't notice the error to the upper layer yet.
+                * The error handling decision is made by the target driver,
+                * when the request is completed.
+                */
+               tio->error = error;
+               return;
+       }
+
+       /*
+        * I/O for the bio successfully completed.
+        * Notice the data completion to the upper layer.
+        */
+
+       /*
+        * bios are processed from the head of the list.
+        * So the completing bio should always be rq->bio.
+        * If it's not, something wrong is happening.
+        */
+       if (tio->orig->bio != bio)
+               DMERR("bio completion is going in the middle of the request");
+
+       /*
+        * Update the original request.
+        * Do not use blk_end_request() here, because it may complete
+        * the original request before the clone, and break the ordering.
+        */
+       blk_update_request(tio->orig, 0, nr_bytes);
+}
+
 static struct dm_rq_target_io *tio_from_request(struct request *rq)
 {
        return (rq->q->mq_ops ? blk_mq_rq_to_pdu(rq) : rq->special);
@@ -1050,6 +1101,8 @@ static void free_rq_clone(struct request *clone)
        struct dm_rq_target_io *tio = clone->end_io_data;
        struct mapped_device *md = tio->md;
 
+       blk_rq_unprep_clone(clone);
+
        if (md->type == DM_TYPE_MQ_REQUEST_BASED)
                /* stacked on blk-mq queue(s) */
                tio->ti->type->release_clone_rq(clone);
@@ -1784,13 +1837,39 @@ static void dm_dispatch_clone_request(struct request *clone, struct request *rq)
                dm_complete_request(rq, r);
 }
 
-static void setup_clone(struct request *clone, struct request *rq,
-                       struct dm_rq_target_io *tio)
+static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
+                                void *data)
 {
-       blk_rq_prep_clone(clone, rq);
+       struct dm_rq_target_io *tio = data;
+       struct dm_rq_clone_bio_info *info =
+               container_of(bio, struct dm_rq_clone_bio_info, clone);
+
+       info->orig = bio_orig;
+       info->tio = tio;
+       bio->bi_end_io = end_clone_bio;
+
+       return 0;
+}
+
+static int setup_clone(struct request *clone, struct request *rq,
+                      struct dm_rq_target_io *tio, gfp_t gfp_mask)
+{
+       int r;
+
+       r = blk_rq_prep_clone(clone, rq, tio->md->bs, gfp_mask,
+                             dm_rq_bio_constructor, tio);
+       if (r)
+               return r;
+
+       clone->cmd = rq->cmd;
+       clone->cmd_len = rq->cmd_len;
+       clone->sense = rq->sense;
        clone->end_io = end_clone_request;
        clone->end_io_data = tio;
+
        tio->clone = clone;
+
+       return 0;
 }
 
 static struct request *clone_rq(struct request *rq, struct mapped_device *md,
@@ -1811,7 +1890,12 @@ static struct request *clone_rq(struct request *rq, struct mapped_device *md,
                clone = tio->clone;
 
        blk_rq_init(NULL, clone);
-       setup_clone(clone, rq, tio);
+       if (setup_clone(clone, rq, tio, gfp_mask)) {
+               /* -ENOMEM */
+               if (alloc_clone)
+                       free_clone_request(md, clone);
+               return NULL;
+       }
 
        return clone;
 }
@@ -1905,7 +1989,11 @@ static int map_request(struct dm_rq_target_io *tio, struct request *rq,
                }
                if (r != DM_MAPIO_REMAPPED)
                        return r;
-               setup_clone(clone, rq, tio);
+               if (setup_clone(clone, rq, tio, GFP_ATOMIC)) {
+                       /* -ENOMEM */
+                       ti->type->release_clone_rq(clone);
+                       return DM_MAPIO_REQUEUE;
+               }
        }
 
        switch (r) {
@@ -2375,6 +2463,8 @@ static void __bind_mempools(struct mapped_device *md, struct dm_table *t)
                goto out;
        }
 
+       BUG_ON(!p || md->io_pool || md->rq_pool || md->bs);
+
        md->io_pool = p->io_pool;
        p->io_pool = NULL;
        md->rq_pool = p->rq_pool;
@@ -3481,23 +3571,48 @@ int dm_noflush_suspending(struct dm_target *ti)
 }
 EXPORT_SYMBOL_GPL(dm_noflush_suspending);
 
-struct dm_md_mempools *dm_alloc_bio_mempools(unsigned integrity,
-                                            unsigned per_bio_data_size)
+struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
+                                           unsigned integrity, unsigned per_bio_data_size)
 {
-       struct dm_md_mempools *pools;
-       unsigned int pool_size = dm_get_reserved_bio_based_ios();
+       struct dm_md_mempools *pools = kzalloc(sizeof(*pools), GFP_KERNEL);
+       struct kmem_cache *cachep = NULL;
+       unsigned int pool_size = 0;
        unsigned int front_pad;
 
-       pools = kzalloc(sizeof(*pools), GFP_KERNEL);
        if (!pools)
                return NULL;
 
-       front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) +
-               offsetof(struct dm_target_io, clone);
+       type = filter_md_type(type, md);
 
-       pools->io_pool = mempool_create_slab_pool(pool_size, _io_cache);
-       if (!pools->io_pool)
-               goto out;
+       switch (type) {
+       case DM_TYPE_BIO_BASED:
+               cachep = _io_cache;
+               pool_size = dm_get_reserved_bio_based_ios();
+               front_pad = roundup(per_bio_data_size, __alignof__(struct dm_target_io)) + offsetof(struct dm_target_io, clone);
+               break;
+       case DM_TYPE_REQUEST_BASED:
+               cachep = _rq_tio_cache;
+               pool_size = dm_get_reserved_rq_based_ios();
+               pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
+               if (!pools->rq_pool)
+                       goto out;
+               /* fall through to setup remaining rq-based pools */
+       case DM_TYPE_MQ_REQUEST_BASED:
+               if (!pool_size)
+                       pool_size = dm_get_reserved_rq_based_ios();
+               front_pad = offsetof(struct dm_rq_clone_bio_info, clone);
+               /* per_bio_data_size is not used. See __bind_mempools(). */
+               WARN_ON(per_bio_data_size != 0);
+               break;
+       default:
+               BUG();
+       }
+
+       if (cachep) {
+               pools->io_pool = mempool_create_slab_pool(pool_size, cachep);
+               if (!pools->io_pool)
+                       goto out;
+       }
 
        pools->bs = bioset_create_nobvec(pool_size, front_pad);
        if (!pools->bs)
@@ -3507,34 +3622,10 @@ struct dm_md_mempools *dm_alloc_bio_mempools(unsigned integrity,
                goto out;
 
        return pools;
-out:
-       dm_free_md_mempools(pools);
-       return NULL;
-}
-
-struct dm_md_mempools *dm_alloc_rq_mempools(struct mapped_device *md,
-                                           unsigned type)
-{
-       unsigned int pool_size = dm_get_reserved_rq_based_ios();
-       struct dm_md_mempools *pools;
-
-       pools = kzalloc(sizeof(*pools), GFP_KERNEL);
-       if (!pools)
-               return NULL;
 
-       if (filter_md_type(type, md) == DM_TYPE_REQUEST_BASED) {
-               pools->rq_pool = mempool_create_slab_pool(pool_size, _rq_cache);
-               if (!pools->rq_pool)
-                       goto out;
-       }
-
-       pools->io_pool = mempool_create_slab_pool(pool_size, _rq_tio_cache);
-       if (!pools->io_pool)
-               goto out;
-
-       return pools;
 out:
        dm_free_md_mempools(pools);
+
        return NULL;
 }
 
index e6e66d087b2696ae8671631978b2fffe3c80f49e..6123c2bf9150cb836c1ecd80ebfe51c9f9aa82fd 100644 (file)
@@ -222,9 +222,8 @@ void dm_kcopyd_exit(void);
 /*
  * Mempool operations
  */
-struct dm_md_mempools *dm_alloc_bio_mempools(unsigned integrity,
-                                            unsigned per_bio_data_size);
-struct dm_md_mempools *dm_alloc_rq_mempools(struct mapped_device *md, unsigned type);
+struct dm_md_mempools *dm_alloc_md_mempools(struct mapped_device *md, unsigned type,
+                                           unsigned integrity, unsigned per_bio_data_size);
 void dm_free_md_mempools(struct dm_md_mempools *pools);
 
 /*
index 6ab9d12d1f17762190870ab153129bac56b82176..7303b3405520b5f82011586887719b16611371be 100644 (file)
@@ -192,7 +192,6 @@ enum rq_flag_bits {
        __REQ_HASHED,           /* on IO scheduler merge hash */
        __REQ_MQ_INFLIGHT,      /* track inflight for MQ */
        __REQ_NO_TIMEOUT,       /* requests may never expire */
-       __REQ_CLONE,            /* cloned bios */
        __REQ_NR_BITS,          /* stops here */
 };
 
@@ -247,6 +246,5 @@ enum rq_flag_bits {
 #define REQ_HASHED             (1ULL << __REQ_HASHED)
 #define REQ_MQ_INFLIGHT                (1ULL << __REQ_MQ_INFLIGHT)
 #define REQ_NO_TIMEOUT         (1ULL << __REQ_NO_TIMEOUT)
-#define REQ_CLONE              (1ULL << __REQ_CLONE)
 
 #endif /* __LINUX_BLK_TYPES_H */
index 776d2ee43ba649749e470b4222169e6095bc6fc2..41c0fb573dff2dac8e051f5194ef774fe094ac98 100644 (file)
@@ -775,7 +775,11 @@ extern void blk_add_request_payload(struct request *rq, struct page *page,
                unsigned int len);
 extern int blk_rq_check_limits(struct request_queue *q, struct request *rq);
 extern int blk_lld_busy(struct request_queue *q);
-extern void blk_rq_prep_clone(struct request *rq, struct request *rq_src);
+extern int blk_rq_prep_clone(struct request *rq, struct request *rq_src,
+                            struct bio_set *bs, gfp_t gfp_mask,
+                            int (*bio_ctr)(struct bio *, struct bio *, void *),
+                            void *data);
+extern void blk_rq_unprep_clone(struct request *rq);
 extern int blk_insert_cloned_request(struct request_queue *q,
                                     struct request *rq);
 extern void blk_delay_queue(struct request_queue *, unsigned long);