[PATCH] dm crypt: move io to workqueue
Milan Broz [Tue, 3 Oct 2006 08:15:39 +0000 (01:15 -0700)]
This patch is designed to help dm-crypt comply with the
new constraints imposed by the following patch in -mm:
  md-dm-reduce-stack-usage-with-stacked-block-devices.patch

Under low memory the existing implementation relies upon waiting for I/O
submitted recursively to generic_make_request() completing before the original
generic_make_request() call can return.

This patch moves the I/O submission to a workqueue so the original
generic_make_request() can return immediately.

Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

drivers/md/dm-crypt.c

index 946a9eb..c34433a 100644 (file)
@@ -35,6 +35,7 @@ struct crypt_io {
        struct work_struct work;
        atomic_t pending;
        int error;
+       int post_process;
 };
 
 /*
@@ -445,8 +446,7 @@ static void dec_pending(struct crypt_io *io, int error)
  * kcryptd:
  *
  * Needed because it would be very unwise to do decryption in an
- * interrupt context, so bios returning from read requests get
- * queued here.
+ * interrupt context.
  */
 static struct workqueue_struct *_kcryptd_workqueue;
 static void kcryptd_do_work(void *data);
@@ -470,12 +470,10 @@ static int crypt_endio(struct bio *clone, unsigned int done, int error)
        if (!read_io)
                crypt_free_buffer_pages(cc, clone, done);
 
+       /* keep going - not finished yet */
        if (unlikely(clone->bi_size))
                return 1;
 
-       /*
-        * successful reads are decrypted by the worker thread
-        */
        if (!read_io)
                goto out;
 
@@ -485,6 +483,7 @@ static int crypt_endio(struct bio *clone, unsigned int done, int error)
        }
 
        bio_put(clone);
+       io->post_process = 1;
        kcryptd_queue_io(io);
        return 0;
 
@@ -504,7 +503,7 @@ static void clone_init(struct crypt_io *io, struct bio *clone)
        clone->bi_rw      = io->base_bio->bi_rw;
 }
 
-static int process_read(struct crypt_io *io)
+static void process_read(struct crypt_io *io)
 {
        struct crypt_config *cc = io->target->private;
        struct bio *base_bio = io->base_bio;
@@ -521,7 +520,7 @@ static int process_read(struct crypt_io *io)
        clone = bio_alloc(GFP_NOIO, bio_segments(base_bio));
        if (unlikely(!clone)) {
                dec_pending(io, -ENOMEM);
-               return 0;
+               return;
        }
 
        clone_init(io, clone);
@@ -533,11 +532,9 @@ static int process_read(struct crypt_io *io)
               sizeof(struct bio_vec) * clone->bi_vcnt);
 
        generic_make_request(clone);
-
-       return 0;
 }
 
-static int process_write(struct crypt_io *io)
+static void process_write(struct crypt_io *io)
 {
        struct crypt_config *cc = io->target->private;
        struct bio *base_bio = io->base_bio;
@@ -558,15 +555,18 @@ static int process_write(struct crypt_io *io)
        while (remaining) {
                clone = crypt_alloc_buffer(cc, base_bio->bi_size,
                                           io->first_clone, &bvec_idx);
-               if (unlikely(!clone))
-                       goto cleanup;
+               if (unlikely(!clone)) {
+                       dec_pending(io, -ENOMEM);
+                       return;
+               }
 
                ctx.bio_out = clone;
 
                if (unlikely(crypt_convert(cc, &ctx) < 0)) {
                        crypt_free_buffer_pages(cc, clone, clone->bi_size);
                        bio_put(clone);
-                       goto cleanup;
+                       dec_pending(io, -EIO);
+                       return;
                }
 
                clone_init(io, clone);
@@ -582,31 +582,20 @@ static int process_write(struct crypt_io *io)
                        io->first_clone = clone;
                }
 
-               atomic_inc(&io->pending);
-
                remaining -= clone->bi_size;
                sector += bio_sectors(clone);
 
+               /* prevent bio_put of first_clone */
+               if (remaining)
+                       atomic_inc(&io->pending);
+
                generic_make_request(clone);
 
                /* out of memory -> run queues */
                if (remaining)
                        blk_congestion_wait(bio_data_dir(clone), HZ/100);
-       }
 
-       /* drop reference, clones could have returned before we reach this */
-       dec_pending(io, 0);
-       return 0;
-
-cleanup:
-       if (io->first_clone) {
-               dec_pending(io, -ENOMEM);
-               return 0;
        }
-
-        /* if no bio has been dispatched yet, we can directly return the error */
-       mempool_free(io, cc->io_pool);
-       return -ENOMEM;
 }
 
 static void process_read_endio(struct crypt_io *io)
@@ -624,7 +613,12 @@ static void kcryptd_do_work(void *data)
 {
        struct crypt_io *io = data;
 
-       process_read_endio(io);
+       if (io->post_process)
+               process_read_endio(io);
+       else if (bio_data_dir(io->base_bio) == READ)
+               process_read(io);
+       else
+               process_write(io);
 }
 
 /*
@@ -889,17 +883,14 @@ static int crypt_map(struct dm_target *ti, struct bio *bio,
        struct crypt_io *io;
 
        io = mempool_alloc(cc->io_pool, GFP_NOIO);
-
        io->target = ti;
        io->base_bio = bio;
        io->first_clone = NULL;
-       io->error = 0;
+       io->error = io->post_process = 0;
        atomic_set(&io->pending, 0);
+       kcryptd_queue_io(io);
 
-       if (bio_data_dir(bio) == WRITE)
-               return process_write(io);
-
-       return process_read(io);
+       return 0;
 }
 
 static int crypt_status(struct dm_target *ti, status_type_t type,
@@ -999,7 +990,7 @@ error:
 
 static struct target_type crypt_target = {
        .name   = "crypt",
-       .version= {1, 2, 0},
+       .version= {1, 3, 0},
        .module = THIS_MODULE,
        .ctr    = crypt_ctr,
        .dtr    = crypt_dtr,