dm crypt: prepare async callback fn
[deliverable/linux.git] / drivers / md / dm-crypt.c
index ac54f697c508e57adc95d545ca4ec15459118d84..c45bd0e59dcce63cb7316db9853ed42c5ecd61fa 100644 (file)
@@ -1,11 +1,12 @@
 /*
  * Copyright (C) 2003 Christophe Saout <christophe@saout.de>
  * Copyright (C) 2004 Clemens Fruhwirth <clemens@endorphin.org>
- * Copyright (C) 2006 Red Hat, Inc. All rights reserved.
+ * Copyright (C) 2006-2007 Red Hat, Inc. All rights reserved.
  *
  * This file is released under the GPL.
  */
 
+#include <linux/completion.h>
 #include <linux/err.h>
 #include <linux/module.h>
 #include <linux/init.h>
 #define DM_MSG_PREFIX "crypt"
 #define MESG_STR(x) x, sizeof(x)
 
-/*
- * per bio private data
- */
-struct dm_crypt_io {
-       struct dm_target *target;
-       struct bio *base_bio;
-       struct work_struct work;
-       atomic_t pending;
-       int error;
-};
-
 /*
  * context holding the current state of a multi-part conversion
  */
 struct convert_context {
+       struct completion restart;
        struct bio *bio_in;
        struct bio *bio_out;
        unsigned int offset_in;
@@ -49,7 +40,27 @@ struct convert_context {
        unsigned int idx_in;
        unsigned int idx_out;
        sector_t sector;
-       int write;
+       atomic_t pending;
+};
+
+/*
+ * per bio private data
+ */
+struct dm_crypt_io {
+       struct dm_target *target;
+       struct bio *base_bio;
+       struct work_struct work;
+
+       struct convert_context ctx;
+
+       atomic_t pending;
+       int error;
+       sector_t sector;
+};
+
+struct dm_crypt_request {
+       struct scatterlist sg_in;
+       struct scatterlist sg_out;
 };
 
 struct crypt_config;
@@ -72,10 +83,11 @@ struct crypt_config {
        sector_t start;
 
        /*
-        * pool for per bio private data and
-        * for encryption buffer pages
+        * pool for per bio private data, crypto requests and
+        * encryption requeusts/buffer pages
         */
        mempool_t *io_pool;
+       mempool_t *req_pool;
        mempool_t *page_pool;
        struct bio_set *bs;
 
@@ -93,6 +105,22 @@ struct crypt_config {
        sector_t iv_offset;
        unsigned int iv_size;
 
+       /*
+        * Layout of each crypto request:
+        *
+        *   struct ablkcipher_request
+        *      context
+        *      padding
+        *   struct dm_crypt_request
+        *      padding
+        *   IV
+        *
+        * The padding is added so that dm_crypt_request and the IV are
+        * correctly aligned.
+        */
+       unsigned int dmreq_start;
+       struct ablkcipher_request *req;
+
        char cipher[CRYPTO_MAX_ALG_NAME];
        char chainmode[CRYPTO_MAX_ALG_NAME];
        struct crypto_blkcipher *tfm;
@@ -108,6 +136,7 @@ struct crypt_config {
 static struct kmem_cache *_crypt_io_pool;
 
 static void clone_init(struct dm_crypt_io *, struct bio *);
+static void kcryptd_queue_crypt(struct dm_crypt_io *io);
 
 /*
  * Different IV generation algorithms:
@@ -168,7 +197,7 @@ static int crypt_iv_essiv_ctr(struct crypt_config *cc, struct dm_target *ti,
                return -ENOMEM;
        }
 
-       sg_set_buf(&sg, cc->key, cc->key_size);
+       sg_init_one(&sg, cc->key, cc->key_size);
        desc.tfm = hash_tfm;
        desc.flags = CRYPTO_TFM_REQ_MAY_SLEEP;
        err = crypto_hash_digest(&desc, &sg, cc->key_size, salt);
@@ -324,7 +353,7 @@ crypt_convert_scatterlist(struct crypt_config *cc, struct scatterlist *out,
 static void crypt_convert_init(struct crypt_config *cc,
                               struct convert_context *ctx,
                               struct bio *bio_out, struct bio *bio_in,
-                              sector_t sector, int write)
+                              sector_t sector)
 {
        ctx->bio_in = bio_in;
        ctx->bio_out = bio_out;
@@ -333,7 +362,61 @@ static void crypt_convert_init(struct crypt_config *cc,
        ctx->idx_in = bio_in ? bio_in->bi_idx : 0;
        ctx->idx_out = bio_out ? bio_out->bi_idx : 0;
        ctx->sector = sector + cc->iv_offset;
-       ctx->write = write;
+       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);
+}
+
+static int crypt_convert_block(struct crypt_config *cc,
+                              struct convert_context *ctx)
+{
+       struct bio_vec *bv_in = bio_iovec_idx(ctx->bio_in, ctx->idx_in);
+       struct bio_vec *bv_out = bio_iovec_idx(ctx->bio_out, ctx->idx_out);
+       struct dm_crypt_request dmreq;
+
+       sg_init_table(&dmreq.sg_in, 1);
+       sg_set_page(&dmreq.sg_in, bv_in->bv_page, 1 << SECTOR_SHIFT,
+                   bv_in->bv_offset + ctx->offset_in);
+
+       sg_init_table(&dmreq.sg_out, 1);
+       sg_set_page(&dmreq.sg_out, bv_out->bv_page, 1 << SECTOR_SHIFT,
+                   bv_out->bv_offset + ctx->offset_out);
+
+       ctx->offset_in += 1 << SECTOR_SHIFT;
+       if (ctx->offset_in >= bv_in->bv_len) {
+               ctx->offset_in = 0;
+               ctx->idx_in++;
+       }
+
+       ctx->offset_out += 1 << SECTOR_SHIFT;
+       if (ctx->offset_out >= bv_out->bv_len) {
+               ctx->offset_out = 0;
+               ctx->idx_out++;
+       }
+
+       return crypt_convert_scatterlist(cc, &dmreq.sg_out, &dmreq.sg_in,
+                                        dmreq.sg_in.length,
+                                        bio_data_dir(ctx->bio_in) == WRITE,
+                                        ctx->sector);
+}
+
+static void kcryptd_async_done(struct crypto_async_request *async_req,
+                              int error);
+static void crypt_alloc_req(struct crypt_config *cc,
+                           struct convert_context *ctx)
+{
+       if (!cc->req)
+               cc->req = mempool_alloc(cc->req_pool, GFP_NOIO);
+       ablkcipher_request_set_tfm(cc->req, cc->tfm);
+       ablkcipher_request_set_callback(cc->req, CRYPTO_TFM_REQ_MAY_BACKLOG |
+                                            CRYPTO_TFM_REQ_MAY_SLEEP,
+                                            kcryptd_async_done, ctx);
 }
 
 /*
@@ -346,40 +429,22 @@ static int crypt_convert(struct crypt_config *cc,
 
        while(ctx->idx_in < ctx->bio_in->bi_vcnt &&
              ctx->idx_out < ctx->bio_out->bi_vcnt) {
-               struct bio_vec *bv_in = bio_iovec_idx(ctx->bio_in, ctx->idx_in);
-               struct bio_vec *bv_out = bio_iovec_idx(ctx->bio_out, ctx->idx_out);
-               struct scatterlist sg_in, sg_out;
-
-               sg_init_table(&sg_in, 1);
-               sg_set_page(&sg_in, bv_in->bv_page);
-               sg_in.offset = bv_in->bv_offset + ctx->offset_in;
-               sg_in.length = 1 << SECTOR_SHIFT;
-
-               sg_init_table(&sg_out, 1);
-               sg_set_page(&sg_out, bv_out->bv_page);
-               sg_out.offset = bv_out->bv_offset + ctx->offset_out;
-               sg_out.length = 1 << SECTOR_SHIFT;
-
-               ctx->offset_in += sg_in.length;
-               if (ctx->offset_in >= bv_in->bv_len) {
-                       ctx->offset_in = 0;
-                       ctx->idx_in++;
-               }
-
-               ctx->offset_out += sg_out.length;
-               if (ctx->offset_out >= bv_out->bv_len) {
-                       ctx->offset_out = 0;
-                       ctx->idx_out++;
-               }
-
-               r = crypt_convert_scatterlist(cc, &sg_out, &sg_in, sg_in.length,
-                                             ctx->write, ctx->sector);
+               r = crypt_convert_block(cc, ctx);
                if (r < 0)
                        break;
 
                ctx->sector++;
        }
 
+       /*
+        * 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;
 }
 
@@ -402,7 +467,8 @@ static struct bio *crypt_alloc_buffer(struct dm_crypt_io *io, unsigned size)
        struct bio *clone;
        unsigned int nr_iovecs = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
        gfp_t gfp_mask = GFP_NOIO | __GFP_HIGHMEM;
-       unsigned int i;
+       unsigned i, len;
+       struct page *page;
 
        clone = bio_alloc_bioset(GFP_NOIO, nr_iovecs, cc->bs);
        if (!clone)
@@ -411,10 +477,8 @@ static struct bio *crypt_alloc_buffer(struct dm_crypt_io *io, unsigned size)
        clone_init(io, clone);
 
        for (i = 0; i < nr_iovecs; i++) {
-               struct bio_vec *bv = bio_iovec_idx(clone, i);
-
-               bv->bv_page = mempool_alloc(cc->page_pool, gfp_mask);
-               if (!bv->bv_page)
+               page = mempool_alloc(cc->page_pool, gfp_mask);
+               if (!page)
                        break;
 
                /*
@@ -425,15 +489,14 @@ static struct bio *crypt_alloc_buffer(struct dm_crypt_io *io, unsigned size)
                if (i == (MIN_BIO_PAGES - 1))
                        gfp_mask = (gfp_mask | __GFP_NOWARN) & ~__GFP_WAIT;
 
-               bv->bv_offset = 0;
-               if (size > PAGE_SIZE)
-                       bv->bv_len = PAGE_SIZE;
-               else
-                       bv->bv_len = size;
+               len = (size > PAGE_SIZE) ? PAGE_SIZE : size;
 
-               clone->bi_size += bv->bv_len;
-               clone->bi_vcnt++;
-               size -= bv->bv_len;
+               if (!bio_add_page(clone, page, len, 0)) {
+                       mempool_free(page, cc->page_pool);
+                       break;
+               }
+
+               size -= len;
        }
 
        if (!clone->bi_size) {
@@ -461,18 +524,14 @@ static void crypt_free_buffer_pages(struct crypt_config *cc, struct bio *clone)
  * One of the bios was finished. Check for completion of
  * the whole request and correctly clean up the buffer.
  */
-static void crypt_dec_pending(struct dm_crypt_io *io, int error)
+static void crypt_dec_pending(struct dm_crypt_io *io)
 {
-       struct crypt_config *cc = (struct crypt_config *) io->target->private;
-
-       if (error < 0)
-               io->error = error;
+       struct crypt_config *cc = io->target->private;
 
        if (!atomic_dec_and_test(&io->pending))
                return;
 
        bio_endio(io->base_bio, io->error);
-
        mempool_free(io, cc->io_pool);
 }
 
@@ -490,51 +549,32 @@ static void crypt_dec_pending(struct dm_crypt_io *io, int error)
  * starved by new requests which can block in the first stages due
  * to memory allocation.
  */
-static void kcryptd_do_work(struct work_struct *work);
-static void kcryptd_do_crypt(struct work_struct *work);
-
-static void kcryptd_queue_io(struct dm_crypt_io *io)
-{
-       struct crypt_config *cc = io->target->private;
-
-       INIT_WORK(&io->work, kcryptd_do_work);
-       queue_work(cc->io_queue, &io->work);
-}
-
-static void kcryptd_queue_crypt(struct dm_crypt_io *io)
-{
-       struct crypt_config *cc = io->target->private;
-
-       INIT_WORK(&io->work, kcryptd_do_crypt);
-       queue_work(cc->crypt_queue, &io->work);
-}
-
 static void crypt_endio(struct bio *clone, int error)
 {
        struct dm_crypt_io *io = clone->bi_private;
        struct crypt_config *cc = io->target->private;
-       unsigned read_io = bio_data_dir(clone) == READ;
+       unsigned rw = bio_data_dir(clone);
+
+       if (unlikely(!bio_flagged(clone, BIO_UPTODATE) && !error))
+               error = -EIO;
 
        /*
         * free the processed pages
         */
-       if (!read_io) {
+       if (rw == WRITE)
                crypt_free_buffer_pages(cc, clone);
-               goto out;
-       }
 
-       if (unlikely(!bio_flagged(clone, BIO_UPTODATE))) {
-               error = -EIO;
-               goto out;
+       bio_put(clone);
+
+       if (rw == READ && !error) {
+               kcryptd_queue_crypt(io);
+               return;
        }
 
-       bio_put(clone);
-       kcryptd_queue_crypt(io);
-       return;
+       if (unlikely(error))
+               io->error = error;
 
-out:
-       bio_put(clone);
-       crypt_dec_pending(io, error);
+       crypt_dec_pending(io);
 }
 
 static void clone_init(struct dm_crypt_io *io, struct bio *clone)
@@ -548,12 +588,11 @@ static void clone_init(struct dm_crypt_io *io, struct bio *clone)
        clone->bi_destructor = dm_crypt_bio_destructor;
 }
 
-static void process_read(struct dm_crypt_io *io)
+static void kcryptd_io_read(struct dm_crypt_io *io)
 {
        struct crypt_config *cc = io->target->private;
        struct bio *base_bio = io->base_bio;
        struct bio *clone;
-       sector_t sector = base_bio->bi_sector - io->target->begin;
 
        atomic_inc(&io->pending);
 
@@ -564,7 +603,8 @@ static void process_read(struct dm_crypt_io *io)
         */
        clone = bio_alloc_bioset(GFP_NOIO, bio_segments(base_bio), cc->bs);
        if (unlikely(!clone)) {
-               crypt_dec_pending(io, -ENOMEM);
+               io->error = -ENOMEM;
+               crypt_dec_pending(io);
                return;
        }
 
@@ -572,25 +612,71 @@ static void process_read(struct dm_crypt_io *io)
        clone->bi_idx = 0;
        clone->bi_vcnt = bio_segments(base_bio);
        clone->bi_size = base_bio->bi_size;
-       clone->bi_sector = cc->start + sector;
+       clone->bi_sector = cc->start + io->sector;
        memcpy(clone->bi_io_vec, bio_iovec(base_bio),
               sizeof(struct bio_vec) * clone->bi_vcnt);
 
        generic_make_request(clone);
 }
 
-static void process_write(struct dm_crypt_io *io)
+static void kcryptd_io_write(struct dm_crypt_io *io)
+{
+       struct bio *clone = io->ctx.bio_out;
+
+       generic_make_request(clone);
+}
+
+static void kcryptd_io(struct work_struct *work)
+{
+       struct dm_crypt_io *io = container_of(work, struct dm_crypt_io, work);
+
+       if (bio_data_dir(io->base_bio) == READ)
+               kcryptd_io_read(io);
+       else
+               kcryptd_io_write(io);
+}
+
+static void kcryptd_queue_io(struct dm_crypt_io *io)
 {
        struct crypt_config *cc = io->target->private;
-       struct bio *base_bio = io->base_bio;
-       struct bio *clone;
-       struct convert_context ctx;
-       unsigned remaining = base_bio->bi_size;
-       sector_t sector = base_bio->bi_sector - io->target->begin;
 
-       atomic_inc(&io->pending);
+       INIT_WORK(&io->work, kcryptd_io);
+       queue_work(cc->io_queue, &io->work);
+}
 
-       crypt_convert_init(cc, &ctx, NULL, base_bio, sector, 1);
+static void kcryptd_crypt_write_io_submit(struct dm_crypt_io *io,
+                                         int error, int async)
+{
+       struct bio *clone = io->ctx.bio_out;
+       struct crypt_config *cc = io->target->private;
+
+       if (unlikely(error < 0)) {
+               crypt_free_buffer_pages(cc, clone);
+               bio_put(clone);
+               io->error = -EIO;
+               return;
+       }
+
+       /* crypt_convert should have filled the clone bio */
+       BUG_ON(io->ctx.idx_out < clone->bi_vcnt);
+
+       clone->bi_sector = cc->start + io->sector;
+       io->sector += bio_sectors(clone);
+
+       if (async)
+               kcryptd_queue_io(io);
+       else {
+               atomic_inc(&io->pending);
+               generic_make_request(clone);
+       }
+}
+
+static void kcryptd_crypt_write_convert_loop(struct dm_crypt_io *io)
+{
+       struct crypt_config *cc = io->target->private;
+       struct bio *clone;
+       unsigned remaining = io->base_bio->bi_size;
+       int r;
 
        /*
         * The allocated buffers can be smaller than the whole bio,
@@ -599,70 +685,102 @@ static void process_write(struct dm_crypt_io *io)
        while (remaining) {
                clone = crypt_alloc_buffer(io, remaining);
                if (unlikely(!clone)) {
-                       crypt_dec_pending(io, -ENOMEM);
+                       io->error = -ENOMEM;
                        return;
                }
 
-               ctx.bio_out = clone;
-               ctx.idx_out = 0;
+               io->ctx.bio_out = clone;
+               io->ctx.idx_out = 0;
 
-               if (unlikely(crypt_convert(cc, &ctx) < 0)) {
-                       crypt_free_buffer_pages(cc, clone);
-                       bio_put(clone);
-                       crypt_dec_pending(io, -EIO);
-                       return;
-               }
-
-               /* crypt_convert should have filled the clone bio */
-               BUG_ON(ctx.idx_out < clone->bi_vcnt);
-
-               clone->bi_sector = cc->start + sector;
                remaining -= clone->bi_size;
-               sector += bio_sectors(clone);
-
-               /* Grab another reference to the io struct
-                * before we kick off the request */
-               if (remaining)
-                       atomic_inc(&io->pending);
 
-               generic_make_request(clone);
+               r = crypt_convert(cc, &io->ctx);
 
-               /* Do not reference clone after this - it
-                * may be gone already. */
+               kcryptd_crypt_write_io_submit(io, r, 0);
+               if (unlikely(r < 0))
+                       return;
 
                /* out of memory -> run queues */
-               if (remaining)
+               if (unlikely(remaining))
                        congestion_wait(WRITE, HZ/100);
        }
 }
 
-static void process_read_endio(struct dm_crypt_io *io)
+static void kcryptd_crypt_write_convert(struct dm_crypt_io *io)
 {
        struct crypt_config *cc = io->target->private;
-       struct convert_context ctx;
 
-       crypt_convert_init(cc, &ctx, io->base_bio, io->base_bio,
-                          io->base_bio->bi_sector - io->target->begin, 0);
+       /*
+        * Prevent io from disappearing until this function completes.
+        */
+       atomic_inc(&io->pending);
 
-       crypt_dec_pending(io, crypt_convert(cc, &ctx));
+       crypt_convert_init(cc, &io->ctx, NULL, io->base_bio, io->sector);
+       kcryptd_crypt_write_convert_loop(io);
+
+       crypt_dec_pending(io);
 }
 
-static void kcryptd_do_work(struct work_struct *work)
+static void kcryptd_crypt_read_done(struct dm_crypt_io *io, int error)
 {
-       struct dm_crypt_io *io = container_of(work, struct dm_crypt_io, work);
+       if (unlikely(error < 0))
+               io->error = -EIO;
+
+       crypt_dec_pending(io);
+}
+
+static void kcryptd_crypt_read_convert(struct dm_crypt_io *io)
+{
+       struct crypt_config *cc = io->target->private;
+       int r = 0;
+
+       crypt_convert_init(cc, &io->ctx, io->base_bio, io->base_bio,
+                          io->sector);
+
+       r = crypt_convert(cc, &io->ctx);
+
+       kcryptd_crypt_read_done(io, r);
+}
+
+static void kcryptd_async_done(struct crypto_async_request *async_req,
+                              int error)
+{
+       struct convert_context *ctx = async_req->data;
+       struct dm_crypt_io *io = container_of(ctx, struct dm_crypt_io, ctx);
+       struct crypt_config *cc = io->target->private;
+
+       if (error == -EINPROGRESS) {
+               complete(&ctx->restart);
+               return;
+       }
+
+       mempool_free(ablkcipher_request_cast(async_req), cc->req_pool);
+
+       if (!atomic_dec_and_test(&ctx->pending))
+               return;
 
        if (bio_data_dir(io->base_bio) == READ)
-               process_read(io);
+               kcryptd_crypt_read_done(io, error);
+       else
+               kcryptd_crypt_write_io_submit(io, error, 1);
 }
 
-static void kcryptd_do_crypt(struct work_struct *work)
+static void kcryptd_crypt(struct work_struct *work)
 {
        struct dm_crypt_io *io = container_of(work, struct dm_crypt_io, work);
 
        if (bio_data_dir(io->base_bio) == READ)
-               process_read_endio(io);
+               kcryptd_crypt_read_convert(io);
        else
-               process_write(io);
+               kcryptd_crypt_write_convert(io);
+}
+
+static void kcryptd_queue_crypt(struct dm_crypt_io *io)
+{
+       struct crypt_config *cc = io->target->private;
+
+       INIT_WORK(&io->work, kcryptd_crypt);
+       queue_work(cc->crypt_queue, &io->work);
 }
 
 /*
@@ -846,6 +964,17 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
                goto bad_slab_pool;
        }
 
+       cc->dmreq_start = sizeof(struct ablkcipher_request);
+       cc->dmreq_start = ALIGN(cc->dmreq_start, crypto_tfm_ctx_alignment());
+
+       cc->req_pool = mempool_create_kmalloc_pool(MIN_IOS, cc->dmreq_start +
+                       sizeof(struct dm_crypt_request) + cc->iv_size);
+       if (!cc->req_pool) {
+               ti->error = "Cannot allocate crypt request mempool";
+               goto bad_req_pool;
+       }
+       cc->req = NULL;
+
        cc->page_pool = mempool_create_page_pool(MIN_POOL_PAGES, 0);
        if (!cc->page_pool) {
                ti->error = "Cannot allocate page mempool";
@@ -919,6 +1048,8 @@ bad_device:
 bad_bs:
        mempool_destroy(cc->page_pool);
 bad_page_pool:
+       mempool_destroy(cc->req_pool);
+bad_req_pool:
        mempool_destroy(cc->io_pool);
 bad_slab_pool:
        if (cc->iv_gen_ops && cc->iv_gen_ops->dtr)
@@ -939,8 +1070,12 @@ static void crypt_dtr(struct dm_target *ti)
        destroy_workqueue(cc->io_queue);
        destroy_workqueue(cc->crypt_queue);
 
+       if (cc->req)
+               mempool_free(cc->req, cc->req_pool);
+
        bioset_free(cc->bs);
        mempool_destroy(cc->page_pool);
+       mempool_destroy(cc->req_pool);
        mempool_destroy(cc->io_pool);
 
        kfree(cc->iv_mode);
@@ -963,6 +1098,7 @@ static int crypt_map(struct dm_target *ti, struct bio *bio,
        io = mempool_alloc(cc->io_pool, GFP_NOIO);
        io->target = ti;
        io->base_bio = bio;
+       io->sector = bio->bi_sector - ti->begin;
        io->error = 0;
        atomic_set(&io->pending, 0);
 
This page took 0.034896 seconds and 5 git commands to generate.