dm crypt: fix ctx pending
authorMilan Broz <mbroz@redhat.com>
Fri, 28 Mar 2008 21:16:07 +0000 (14:16 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 28 Mar 2008 21:45:22 +0000 (14:45 -0700)
Fix regression in dm-crypt introduced in commit
3a7f6c990ad04e6f576a159876c602d14d6f7fef ("dm crypt: use async crypto").

If write requests need to be split into pieces, the code must not process them
in parallel because the crypto context cannot be shared.  So there can be
parallel crypto operations on one part of the write, but only one write bio
can be processed at a time.

This is not optimal and the workqueue code needs to be optimized for parallel
processing, but for now it solves the problem without affecting the
performance of synchronous crypto operation (most of current dm-crypt users).

http://bugzilla.kernel.org/show_bug.cgi?id=10242
http://bugzilla.kernel.org/show_bug.cgi?id=10207

Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: "Rafael J. Wysocki" <rjw@sisk.pl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/md/dm-crypt.c

index b04f98df94ea310db726b91be82eacf090a71ae1..835def11419d36d7b56e0e216028714d384ad0f9 100644 (file)
@@ -1,7 +1,7 @@
 /*
  * Copyright (C) 2003 Christophe Saout <christophe@saout.de>
  * Copyright (C) 2004 Clemens Fruhwirth <clemens@endorphin.org>
- * Copyright (C) 2006-2007 Red Hat, Inc. All rights reserved.
+ * Copyright (C) 2006-2008 Red Hat, Inc. All rights reserved.
  *
  * This file is released under the GPL.
  */
@@ -93,6 +93,8 @@ struct crypt_config {
 
        struct workqueue_struct *io_queue;
        struct workqueue_struct *crypt_queue;
+       wait_queue_head_t writeq;
+
        /*
         * crypto related data
         */
@@ -331,14 +333,7 @@ static void crypt_convert_init(struct crypt_config *cc,
        ctx->idx_out = bio_out ? bio_out->bi_idx : 0;
        ctx->sector = sector + cc->iv_offset;
        init_completion(&ctx->restart);
-       /*
-        * Crypto operation can be asynchronous,
-        * ctx->pending is increased after request submission.
-        * We need to ensure that we don't call the crypt finish
-        * operation before pending got incremented
-        * (dependent on crypt submission return code).
-        */
-       atomic_set(&ctx->pending, 2);
+       atomic_set(&ctx->pending, 1);
 }
 
 static int crypt_convert_block(struct crypt_config *cc,
@@ -411,43 +406,42 @@ static void crypt_alloc_req(struct crypt_config *cc,
 static int crypt_convert(struct crypt_config *cc,
                         struct convert_context *ctx)
 {
-       int r = 0;
+       int r;
 
        while(ctx->idx_in < ctx->bio_in->bi_vcnt &&
              ctx->idx_out < ctx->bio_out->bi_vcnt) {
 
                crypt_alloc_req(cc, ctx);
 
+               atomic_inc(&ctx->pending);
+
                r = crypt_convert_block(cc, ctx, cc->req);
 
                switch (r) {
+               /* async */
                case -EBUSY:
                        wait_for_completion(&ctx->restart);
                        INIT_COMPLETION(ctx->restart);
                        /* fall through*/
                case -EINPROGRESS:
-                       atomic_inc(&ctx->pending);
                        cc->req = NULL;
-                       r = 0;
-                       /* fall through*/
+                       ctx->sector++;
+                       continue;
+
+               /* sync */
                case 0:
+                       atomic_dec(&ctx->pending);
                        ctx->sector++;
                        continue;
-               }
 
-               break;
+               /* error */
+               default:
+                       atomic_dec(&ctx->pending);
+                       return r;
+               }
        }
 
-       /*
-        * If there are pending crypto operation run async
-        * code. Otherwise process return code synchronously.
-        * The step of 2 ensures that async finish doesn't
-        * call crypto finish too early.
-        */
-       if (atomic_sub_return(2, &ctx->pending))
-               return -EINPROGRESS;
-
-       return r;
+       return 0;
 }
 
 static void dm_crypt_bio_destructor(struct bio *bio)
@@ -624,8 +618,10 @@ static void kcryptd_io_read(struct dm_crypt_io *io)
 static void kcryptd_io_write(struct dm_crypt_io *io)
 {
        struct bio *clone = io->ctx.bio_out;
+       struct crypt_config *cc = io->target->private;
 
        generic_make_request(clone);
+       wake_up(&cc->writeq);
 }
 
 static void kcryptd_io(struct work_struct *work)
@@ -698,7 +694,8 @@ static void kcryptd_crypt_write_convert_loop(struct dm_crypt_io *io)
 
                r = crypt_convert(cc, &io->ctx);
 
-               if (r != -EINPROGRESS) {
+               if (atomic_dec_and_test(&io->ctx.pending)) {
+                       /* processed, no running async crypto  */
                        kcryptd_crypt_write_io_submit(io, r, 0);
                        if (unlikely(r < 0))
                                return;
@@ -706,8 +703,12 @@ static void kcryptd_crypt_write_convert_loop(struct dm_crypt_io *io)
                        atomic_inc(&io->pending);
 
                /* out of memory -> run queues */
-               if (unlikely(remaining))
+               if (unlikely(remaining)) {
+                       /* wait for async crypto then reinitialize pending */
+                       wait_event(cc->writeq, !atomic_read(&io->ctx.pending));
+                       atomic_set(&io->ctx.pending, 1);
                        congestion_wait(WRITE, HZ/100);
+               }
        }
 }
 
@@ -746,7 +747,7 @@ static void kcryptd_crypt_read_convert(struct dm_crypt_io *io)
 
        r = crypt_convert(cc, &io->ctx);
 
-       if (r != -EINPROGRESS)
+       if (atomic_dec_and_test(&io->ctx.pending))
                kcryptd_crypt_read_done(io, r);
 
        crypt_dec_pending(io);
@@ -1047,6 +1048,7 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
                goto bad_crypt_queue;
        }
 
+       init_waitqueue_head(&cc->writeq);
        ti->private = cc;
        return 0;