]> nv-tegra.nvidia Code Review - linux-2.6.git/blobdiff - drivers/md/dm-crypt.c
dm crypt: optionally support discard requests
[linux-2.6.git] / drivers / md / dm-crypt.c
index c7bee4f2eedbeedd77dadc7e42ea2fe61709e1bd..49da55c1528aa01137a61c98d6033c9b84dc2e05 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) 2006-2009 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>
 #include <linux/crypto.h>
 #include <linux/workqueue.h>
 #include <linux/backing-dev.h>
-#include <asm/atomic.h>
+#include <linux/percpu.h>
+#include <linux/atomic.h>
 #include <linux/scatterlist.h>
 #include <asm/page.h>
+#include <asm/unaligned.h>
+#include <crypto/hash.h>
+#include <crypto/md5.h>
+#include <crypto/algapi.h>
 
-#include "dm.h"
+#include <linux/device-mapper.h>
 
 #define DM_MSG_PREFIX "crypt"
-#define MESG_STR(x) x, sizeof(x)
-
-/*
- * per bio private data
- */
-struct crypt_io {
-       struct dm_target *target;
-       struct bio *base_bio;
-       struct bio *first_clone;
-       struct work_struct work;
-       atomic_t pending;
-       int error;
-       int post_process;
-};
 
 /*
  * 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;
@@ -50,17 +43,59 @@ 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_io *base_io;
+};
+
+struct dm_crypt_request {
+       struct convert_context *ctx;
+       struct scatterlist sg_in;
+       struct scatterlist sg_out;
+       sector_t iv_sector;
 };
 
 struct crypt_config;
 
 struct crypt_iv_operations {
        int (*ctr)(struct crypt_config *cc, struct dm_target *ti,
-                  const char *opts);
+                  const char *opts);
        void (*dtr)(struct crypt_config *cc);
-       const char *(*status)(struct crypt_config *cc);
-       int (*generator)(struct crypt_config *cc, u8 *iv, sector_t sector);
+       int (*init)(struct crypt_config *cc);
+       int (*wipe)(struct crypt_config *cc);
+       int (*generator)(struct crypt_config *cc, u8 *iv,
+                        struct dm_crypt_request *dmreq);
+       int (*post)(struct crypt_config *cc, u8 *iv,
+                   struct dm_crypt_request *dmreq);
+};
+
+struct iv_essiv_private {
+       struct crypto_hash *hash_tfm;
+       u8 *salt;
+};
+
+struct iv_benbi_private {
+       int shift;
+};
+
+#define LMK_SEED_SIZE 64 /* hash + 0 */
+struct iv_lmk_private {
+       struct crypto_shash *hash_tfm;
+       u8 *seed;
 };
 
 /*
@@ -68,32 +103,74 @@ struct crypt_iv_operations {
  * and encrypts / decrypts at the same time.
  */
 enum flags { DM_CRYPT_SUSPENDED, DM_CRYPT_KEY_VALID };
+
+/*
+ * Duplicated per-CPU state for cipher.
+ */
+struct crypt_cpu {
+       struct ablkcipher_request *req;
+       /* ESSIV: struct crypto_cipher *essiv_tfm */
+       void *iv_private;
+       struct crypto_ablkcipher *tfms[0];
+};
+
+/*
+ * The fields in here must be read only after initialization,
+ * changing state should be in crypt_cpu.
+ */
 struct crypt_config {
        struct dm_dev *dev;
        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;
 
-       /*
-        * crypto related data
-        */
+       struct workqueue_struct *io_queue;
+       struct workqueue_struct *crypt_queue;
+
+       char *cipher;
+       char *cipher_string;
+
        struct crypt_iv_operations *iv_gen_ops;
-       char *iv_mode;
-       struct crypto_cipher *iv_gen_private;
+       union {
+               struct iv_essiv_private essiv;
+               struct iv_benbi_private benbi;
+               struct iv_lmk_private lmk;
+       } iv_gen_private;
        sector_t iv_offset;
        unsigned int iv_size;
 
-       char cipher[CRYPTO_MAX_ALG_NAME];
-       char chainmode[CRYPTO_MAX_ALG_NAME];
-       struct crypto_blkcipher *tfm;
+       /*
+        * Duplicated per cpu state. Access through
+        * per_cpu_ptr() only.
+        */
+       struct crypt_cpu __percpu *cpu;
+       unsigned tfms_count;
+
+       /*
+        * 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;
+
        unsigned long flags;
        unsigned int key_size;
+       unsigned int key_parts;
        u8 key[0];
 };
 
@@ -103,282 +180,689 @@ 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);
+static u8 *iv_of_dmreq(struct crypt_config *cc, struct dm_crypt_request *dmreq);
+
+static struct crypt_cpu *this_crypt_config(struct crypt_config *cc)
+{
+       return this_cpu_ptr(cc->cpu);
+}
+
+/*
+ * Use this to access cipher attributes that are the same for each CPU.
+ */
+static struct crypto_ablkcipher *any_tfm(struct crypt_config *cc)
+{
+       return __this_cpu_ptr(cc->cpu)->tfms[0];
+}
+
 /*
  * Different IV generation algorithms:
  *
  * plain: the initial vector is the 32-bit little-endian version of the sector
- *        number, padded with zeros if neccessary.
+ *        number, padded with zeros if necessary.
+ *
+ * plain64: the initial vector is the 64-bit little-endian version of the sector
+ *        number, padded with zeros if necessary.
  *
  * essiv: "encrypted sector|salt initial vector", the sector number is
  *        encrypted with the bulk cipher using a salt as key. The salt
  *        should be derived from the bulk cipher's key via hashing.
  *
+ * benbi: the 64-bit "big-endian 'narrow block'-count", starting at 1
+ *        (needed for LRW-32-AES and possible other narrow block modes)
+ *
+ * null: the initial vector is always zero.  Provides compatibility with
+ *       obsolete loop_fish2 devices.  Do not use for new devices.
+ *
+ * lmk:  Compatible implementation of the block chaining mode used
+ *       by the Loop-AES block device encryption system
+ *       designed by Jari Ruusu. See http://loop-aes.sourceforge.net/
+ *       It operates on full 512 byte sectors and uses CBC
+ *       with an IV derived from the sector number, the data and
+ *       optionally extra IV seed.
+ *       This means that after decryption the first block
+ *       of sector must be tweaked according to decrypted data.
+ *       Loop-AES can use three encryption schemes:
+ *         version 1: is plain aes-cbc mode
+ *         version 2: uses 64 multikey scheme with lmk IV generator
+ *         version 3: the same as version 2 with additional IV seed
+ *                   (it uses 65 keys, last key is used as IV seed)
+ *
  * plumb: unimplemented, see:
  * http://article.gmane.org/gmane.linux.kernel.device-mapper.dm-crypt/454
  */
 
-static int crypt_iv_plain_gen(struct crypt_config *cc, u8 *iv, sector_t sector)
+static int crypt_iv_plain_gen(struct crypt_config *cc, u8 *iv,
+                             struct dm_crypt_request *dmreq)
 {
        memset(iv, 0, cc->iv_size);
-       *(u32 *)iv = cpu_to_le32(sector & 0xffffffff);
+       *(__le32 *)iv = cpu_to_le32(dmreq->iv_sector & 0xffffffff);
 
        return 0;
 }
 
-static int crypt_iv_essiv_ctr(struct crypt_config *cc, struct dm_target *ti,
-                             const char *opts)
+static int crypt_iv_plain64_gen(struct crypt_config *cc, u8 *iv,
+                               struct dm_crypt_request *dmreq)
 {
-       struct crypto_cipher *essiv_tfm;
-       struct crypto_hash *hash_tfm;
+       memset(iv, 0, cc->iv_size);
+       *(__le64 *)iv = cpu_to_le64(dmreq->iv_sector);
+
+       return 0;
+}
+
+/* Initialise ESSIV - compute salt but no local memory allocations */
+static int crypt_iv_essiv_init(struct crypt_config *cc)
+{
+       struct iv_essiv_private *essiv = &cc->iv_gen_private.essiv;
        struct hash_desc desc;
        struct scatterlist sg;
-       unsigned int saltsize;
-       u8 *salt;
-       int err;
+       struct crypto_cipher *essiv_tfm;
+       int err, cpu;
 
-       if (opts == NULL) {
-               ti->error = "Digest algorithm missing for ESSIV mode";
-               return -EINVAL;
-       }
+       sg_init_one(&sg, cc->key, cc->key_size);
+       desc.tfm = essiv->hash_tfm;
+       desc.flags = CRYPTO_TFM_REQ_MAY_SLEEP;
 
-       /* Hash the cipher key with the given hash algorithm */
-       hash_tfm = crypto_alloc_hash(opts, 0, CRYPTO_ALG_ASYNC);
-       if (IS_ERR(hash_tfm)) {
-               ti->error = "Error initializing ESSIV hash";
-               return PTR_ERR(hash_tfm);
-       }
+       err = crypto_hash_digest(&desc, &sg, cc->key_size, essiv->salt);
+       if (err)
+               return err;
 
-       saltsize = crypto_hash_digestsize(hash_tfm);
-       salt = kmalloc(saltsize, GFP_KERNEL);
-       if (salt == NULL) {
-               ti->error = "Error kmallocing salt storage in ESSIV";
-               crypto_free_hash(hash_tfm);
-               return -ENOMEM;
+       for_each_possible_cpu(cpu) {
+               essiv_tfm = per_cpu_ptr(cc->cpu, cpu)->iv_private,
+
+               err = crypto_cipher_setkey(essiv_tfm, essiv->salt,
+                                   crypto_hash_digestsize(essiv->hash_tfm));
+               if (err)
+                       return err;
        }
 
-       sg_set_buf(&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);
-       crypto_free_hash(hash_tfm);
+       return 0;
+}
 
-       if (err) {
-               ti->error = "Error calculating hash in ESSIV";
-               return err;
+/* Wipe salt and reset key derived from volume key */
+static int crypt_iv_essiv_wipe(struct crypt_config *cc)
+{
+       struct iv_essiv_private *essiv = &cc->iv_gen_private.essiv;
+       unsigned salt_size = crypto_hash_digestsize(essiv->hash_tfm);
+       struct crypto_cipher *essiv_tfm;
+       int cpu, r, err = 0;
+
+       memset(essiv->salt, 0, salt_size);
+
+       for_each_possible_cpu(cpu) {
+               essiv_tfm = per_cpu_ptr(cc->cpu, cpu)->iv_private;
+               r = crypto_cipher_setkey(essiv_tfm, essiv->salt, salt_size);
+               if (r)
+                       err = r;
        }
 
+       return err;
+}
+
+/* Set up per cpu cipher state */
+static struct crypto_cipher *setup_essiv_cpu(struct crypt_config *cc,
+                                            struct dm_target *ti,
+                                            u8 *salt, unsigned saltsize)
+{
+       struct crypto_cipher *essiv_tfm;
+       int err;
+
        /* Setup the essiv_tfm with the given salt */
        essiv_tfm = crypto_alloc_cipher(cc->cipher, 0, CRYPTO_ALG_ASYNC);
        if (IS_ERR(essiv_tfm)) {
                ti->error = "Error allocating crypto tfm for ESSIV";
-               kfree(salt);
-               return PTR_ERR(essiv_tfm);
+               return essiv_tfm;
        }
+
        if (crypto_cipher_blocksize(essiv_tfm) !=
-           crypto_blkcipher_ivsize(cc->tfm)) {
+           crypto_ablkcipher_ivsize(any_tfm(cc))) {
                ti->error = "Block size of ESSIV cipher does "
-                               "not match IV size of block cipher";
+                           "not match IV size of block cipher";
                crypto_free_cipher(essiv_tfm);
-               kfree(salt);
-               return -EINVAL;
+               return ERR_PTR(-EINVAL);
        }
+
        err = crypto_cipher_setkey(essiv_tfm, salt, saltsize);
        if (err) {
                ti->error = "Failed to set key for ESSIV cipher";
                crypto_free_cipher(essiv_tfm);
-               kfree(salt);
-               return err;
+               return ERR_PTR(err);
+       }
+
+       return essiv_tfm;
+}
+
+static void crypt_iv_essiv_dtr(struct crypt_config *cc)
+{
+       int cpu;
+       struct crypt_cpu *cpu_cc;
+       struct crypto_cipher *essiv_tfm;
+       struct iv_essiv_private *essiv = &cc->iv_gen_private.essiv;
+
+       crypto_free_hash(essiv->hash_tfm);
+       essiv->hash_tfm = NULL;
+
+       kzfree(essiv->salt);
+       essiv->salt = NULL;
+
+       for_each_possible_cpu(cpu) {
+               cpu_cc = per_cpu_ptr(cc->cpu, cpu);
+               essiv_tfm = cpu_cc->iv_private;
+
+               if (essiv_tfm)
+                       crypto_free_cipher(essiv_tfm);
+
+               cpu_cc->iv_private = NULL;
+       }
+}
+
+static int crypt_iv_essiv_ctr(struct crypt_config *cc, struct dm_target *ti,
+                             const char *opts)
+{
+       struct crypto_cipher *essiv_tfm = NULL;
+       struct crypto_hash *hash_tfm = NULL;
+       u8 *salt = NULL;
+       int err, cpu;
+
+       if (!opts) {
+               ti->error = "Digest algorithm missing for ESSIV mode";
+               return -EINVAL;
        }
+
+       /* Allocate hash algorithm */
+       hash_tfm = crypto_alloc_hash(opts, 0, CRYPTO_ALG_ASYNC);
+       if (IS_ERR(hash_tfm)) {
+               ti->error = "Error initializing ESSIV hash";
+               err = PTR_ERR(hash_tfm);
+               goto bad;
+       }
+
+       salt = kzalloc(crypto_hash_digestsize(hash_tfm), GFP_KERNEL);
+       if (!salt) {
+               ti->error = "Error kmallocing salt storage in ESSIV";
+               err = -ENOMEM;
+               goto bad;
+       }
+
+       cc->iv_gen_private.essiv.salt = salt;
+       cc->iv_gen_private.essiv.hash_tfm = hash_tfm;
+
+       for_each_possible_cpu(cpu) {
+               essiv_tfm = setup_essiv_cpu(cc, ti, salt,
+                                       crypto_hash_digestsize(hash_tfm));
+               if (IS_ERR(essiv_tfm)) {
+                       crypt_iv_essiv_dtr(cc);
+                       return PTR_ERR(essiv_tfm);
+               }
+               per_cpu_ptr(cc->cpu, cpu)->iv_private = essiv_tfm;
+       }
+
+       return 0;
+
+bad:
+       if (hash_tfm && !IS_ERR(hash_tfm))
+               crypto_free_hash(hash_tfm);
        kfree(salt);
+       return err;
+}
+
+static int crypt_iv_essiv_gen(struct crypt_config *cc, u8 *iv,
+                             struct dm_crypt_request *dmreq)
+{
+       struct crypto_cipher *essiv_tfm = this_crypt_config(cc)->iv_private;
+
+       memset(iv, 0, cc->iv_size);
+       *(__le64 *)iv = cpu_to_le64(dmreq->iv_sector);
+       crypto_cipher_encrypt_one(essiv_tfm, iv, iv);
 
-       cc->iv_gen_private = essiv_tfm;
        return 0;
 }
 
-static void crypt_iv_essiv_dtr(struct crypt_config *cc)
+static int crypt_iv_benbi_ctr(struct crypt_config *cc, struct dm_target *ti,
+                             const char *opts)
 {
-       crypto_free_cipher(cc->iv_gen_private);
-       cc->iv_gen_private = NULL;
+       unsigned bs = crypto_ablkcipher_blocksize(any_tfm(cc));
+       int log = ilog2(bs);
+
+       /* we need to calculate how far we must shift the sector count
+        * to get the cipher block count, we use this shift in _gen */
+
+       if (1 << log != bs) {
+               ti->error = "cypher blocksize is not a power of 2";
+               return -EINVAL;
+       }
+
+       if (log > 9) {
+               ti->error = "cypher blocksize is > 512";
+               return -EINVAL;
+       }
+
+       cc->iv_gen_private.benbi.shift = 9 - log;
+
+       return 0;
 }
 
-static int crypt_iv_essiv_gen(struct crypt_config *cc, u8 *iv, sector_t sector)
+static void crypt_iv_benbi_dtr(struct crypt_config *cc)
+{
+}
+
+static int crypt_iv_benbi_gen(struct crypt_config *cc, u8 *iv,
+                             struct dm_crypt_request *dmreq)
+{
+       __be64 val;
+
+       memset(iv, 0, cc->iv_size - sizeof(u64)); /* rest is cleared below */
+
+       val = cpu_to_be64(((u64)dmreq->iv_sector << cc->iv_gen_private.benbi.shift) + 1);
+       put_unaligned(val, (__be64 *)(iv + cc->iv_size - sizeof(u64)));
+
+       return 0;
+}
+
+static int crypt_iv_null_gen(struct crypt_config *cc, u8 *iv,
+                            struct dm_crypt_request *dmreq)
 {
        memset(iv, 0, cc->iv_size);
-       *(u64 *)iv = cpu_to_le64(sector);
-       crypto_cipher_encrypt_one(cc->iv_gen_private, iv, iv);
+
        return 0;
 }
 
+static void crypt_iv_lmk_dtr(struct crypt_config *cc)
+{
+       struct iv_lmk_private *lmk = &cc->iv_gen_private.lmk;
+
+       if (lmk->hash_tfm && !IS_ERR(lmk->hash_tfm))
+               crypto_free_shash(lmk->hash_tfm);
+       lmk->hash_tfm = NULL;
+
+       kzfree(lmk->seed);
+       lmk->seed = NULL;
+}
+
+static int crypt_iv_lmk_ctr(struct crypt_config *cc, struct dm_target *ti,
+                           const char *opts)
+{
+       struct iv_lmk_private *lmk = &cc->iv_gen_private.lmk;
+
+       lmk->hash_tfm = crypto_alloc_shash("md5", 0, 0);
+       if (IS_ERR(lmk->hash_tfm)) {
+               ti->error = "Error initializing LMK hash";
+               return PTR_ERR(lmk->hash_tfm);
+       }
+
+       /* No seed in LMK version 2 */
+       if (cc->key_parts == cc->tfms_count) {
+               lmk->seed = NULL;
+               return 0;
+       }
+
+       lmk->seed = kzalloc(LMK_SEED_SIZE, GFP_KERNEL);
+       if (!lmk->seed) {
+               crypt_iv_lmk_dtr(cc);
+               ti->error = "Error kmallocing seed storage in LMK";
+               return -ENOMEM;
+       }
+
+       return 0;
+}
+
+static int crypt_iv_lmk_init(struct crypt_config *cc)
+{
+       struct iv_lmk_private *lmk = &cc->iv_gen_private.lmk;
+       int subkey_size = cc->key_size / cc->key_parts;
+
+       /* LMK seed is on the position of LMK_KEYS + 1 key */
+       if (lmk->seed)
+               memcpy(lmk->seed, cc->key + (cc->tfms_count * subkey_size),
+                      crypto_shash_digestsize(lmk->hash_tfm));
+
+       return 0;
+}
+
+static int crypt_iv_lmk_wipe(struct crypt_config *cc)
+{
+       struct iv_lmk_private *lmk = &cc->iv_gen_private.lmk;
+
+       if (lmk->seed)
+               memset(lmk->seed, 0, LMK_SEED_SIZE);
+
+       return 0;
+}
+
+static int crypt_iv_lmk_one(struct crypt_config *cc, u8 *iv,
+                           struct dm_crypt_request *dmreq,
+                           u8 *data)
+{
+       struct iv_lmk_private *lmk = &cc->iv_gen_private.lmk;
+       struct {
+               struct shash_desc desc;
+               char ctx[crypto_shash_descsize(lmk->hash_tfm)];
+       } sdesc;
+       struct md5_state md5state;
+       u32 buf[4];
+       int i, r;
+
+       sdesc.desc.tfm = lmk->hash_tfm;
+       sdesc.desc.flags = CRYPTO_TFM_REQ_MAY_SLEEP;
+
+       r = crypto_shash_init(&sdesc.desc);
+       if (r)
+               return r;
+
+       if (lmk->seed) {
+               r = crypto_shash_update(&sdesc.desc, lmk->seed, LMK_SEED_SIZE);
+               if (r)
+                       return r;
+       }
+
+       /* Sector is always 512B, block size 16, add data of blocks 1-31 */
+       r = crypto_shash_update(&sdesc.desc, data + 16, 16 * 31);
+       if (r)
+               return r;
+
+       /* Sector is cropped to 56 bits here */
+       buf[0] = cpu_to_le32(dmreq->iv_sector & 0xFFFFFFFF);
+       buf[1] = cpu_to_le32((((u64)dmreq->iv_sector >> 32) & 0x00FFFFFF) | 0x80000000);
+       buf[2] = cpu_to_le32(4024);
+       buf[3] = 0;
+       r = crypto_shash_update(&sdesc.desc, (u8 *)buf, sizeof(buf));
+       if (r)
+               return r;
+
+       /* No MD5 padding here */
+       r = crypto_shash_export(&sdesc.desc, &md5state);
+       if (r)
+               return r;
+
+       for (i = 0; i < MD5_HASH_WORDS; i++)
+               __cpu_to_le32s(&md5state.hash[i]);
+       memcpy(iv, &md5state.hash, cc->iv_size);
+
+       return 0;
+}
+
+static int crypt_iv_lmk_gen(struct crypt_config *cc, u8 *iv,
+                           struct dm_crypt_request *dmreq)
+{
+       u8 *src;
+       int r = 0;
+
+       if (bio_data_dir(dmreq->ctx->bio_in) == WRITE) {
+               src = kmap_atomic(sg_page(&dmreq->sg_in), KM_USER0);
+               r = crypt_iv_lmk_one(cc, iv, dmreq, src + dmreq->sg_in.offset);
+               kunmap_atomic(src, KM_USER0);
+       } else
+               memset(iv, 0, cc->iv_size);
+
+       return r;
+}
+
+static int crypt_iv_lmk_post(struct crypt_config *cc, u8 *iv,
+                            struct dm_crypt_request *dmreq)
+{
+       u8 *dst;
+       int r;
+
+       if (bio_data_dir(dmreq->ctx->bio_in) == WRITE)
+               return 0;
+
+       dst = kmap_atomic(sg_page(&dmreq->sg_out), KM_USER0);
+       r = crypt_iv_lmk_one(cc, iv, dmreq, dst + dmreq->sg_out.offset);
+
+       /* Tweak the first block of plaintext sector */
+       if (!r)
+               crypto_xor(dst + dmreq->sg_out.offset, iv, cc->iv_size);
+
+       kunmap_atomic(dst, KM_USER0);
+       return r;
+}
+
 static struct crypt_iv_operations crypt_iv_plain_ops = {
        .generator = crypt_iv_plain_gen
 };
 
+static struct crypt_iv_operations crypt_iv_plain64_ops = {
+       .generator = crypt_iv_plain64_gen
+};
+
 static struct crypt_iv_operations crypt_iv_essiv_ops = {
        .ctr       = crypt_iv_essiv_ctr,
        .dtr       = crypt_iv_essiv_dtr,
+       .init      = crypt_iv_essiv_init,
+       .wipe      = crypt_iv_essiv_wipe,
        .generator = crypt_iv_essiv_gen
 };
 
+static struct crypt_iv_operations crypt_iv_benbi_ops = {
+       .ctr       = crypt_iv_benbi_ctr,
+       .dtr       = crypt_iv_benbi_dtr,
+       .generator = crypt_iv_benbi_gen
+};
 
-static int
-crypt_convert_scatterlist(struct crypt_config *cc, struct scatterlist *out,
-                          struct scatterlist *in, unsigned int length,
-                          int write, sector_t sector)
+static struct crypt_iv_operations crypt_iv_null_ops = {
+       .generator = crypt_iv_null_gen
+};
+
+static struct crypt_iv_operations crypt_iv_lmk_ops = {
+       .ctr       = crypt_iv_lmk_ctr,
+       .dtr       = crypt_iv_lmk_dtr,
+       .init      = crypt_iv_lmk_init,
+       .wipe      = crypt_iv_lmk_wipe,
+       .generator = crypt_iv_lmk_gen,
+       .post      = crypt_iv_lmk_post
+};
+
+static void crypt_convert_init(struct crypt_config *cc,
+                              struct convert_context *ctx,
+                              struct bio *bio_out, struct bio *bio_in,
+                              sector_t sector)
 {
-       u8 iv[cc->iv_size];
-       struct blkcipher_desc desc = {
-               .tfm = cc->tfm,
-               .info = iv,
-               .flags = CRYPTO_TFM_REQ_MAY_SLEEP,
-       };
-       int r;
+       ctx->bio_in = bio_in;
+       ctx->bio_out = bio_out;
+       ctx->offset_in = 0;
+       ctx->offset_out = 0;
+       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;
+       init_completion(&ctx->restart);
+}
+
+static struct dm_crypt_request *dmreq_of_req(struct crypt_config *cc,
+                                            struct ablkcipher_request *req)
+{
+       return (struct dm_crypt_request *)((char *)req + cc->dmreq_start);
+}
+
+static struct ablkcipher_request *req_of_dmreq(struct crypt_config *cc,
+                                              struct dm_crypt_request *dmreq)
+{
+       return (struct ablkcipher_request *)((char *)dmreq - cc->dmreq_start);
+}
+
+static u8 *iv_of_dmreq(struct crypt_config *cc,
+                      struct dm_crypt_request *dmreq)
+{
+       return (u8 *)ALIGN((unsigned long)(dmreq + 1),
+               crypto_ablkcipher_alignmask(any_tfm(cc)) + 1);
+}
+
+static int crypt_convert_block(struct crypt_config *cc,
+                              struct convert_context *ctx,
+                              struct ablkcipher_request *req)
+{
+       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;
+       u8 *iv;
+       int r = 0;
+
+       dmreq = dmreq_of_req(cc, req);
+       iv = iv_of_dmreq(cc, dmreq);
+
+       dmreq->iv_sector = ctx->sector;
+       dmreq->ctx = ctx;
+       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++;
+       }
 
        if (cc->iv_gen_ops) {
-               r = cc->iv_gen_ops->generator(cc, iv, sector);
+               r = cc->iv_gen_ops->generator(cc, iv, dmreq);
                if (r < 0)
                        return r;
-
-               if (write)
-                       r = crypto_blkcipher_encrypt_iv(&desc, out, in, length);
-               else
-                       r = crypto_blkcipher_decrypt_iv(&desc, out, in, length);
-       } else {
-               if (write)
-                       r = crypto_blkcipher_encrypt(&desc, out, in, length);
-               else
-                       r = crypto_blkcipher_decrypt(&desc, out, in, length);
        }
 
+       ablkcipher_request_set_crypt(req, &dmreq->sg_in, &dmreq->sg_out,
+                                    1 << SECTOR_SHIFT, iv);
+
+       if (bio_data_dir(ctx->bio_in) == WRITE)
+               r = crypto_ablkcipher_encrypt(req);
+       else
+               r = crypto_ablkcipher_decrypt(req);
+
+       if (!r && cc->iv_gen_ops && cc->iv_gen_ops->post)
+               r = cc->iv_gen_ops->post(cc, iv, dmreq);
+
        return r;
 }
 
-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)
+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)
 {
-       ctx->bio_in = bio_in;
-       ctx->bio_out = bio_out;
-       ctx->offset_in = 0;
-       ctx->offset_out = 0;
-       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;
+       struct crypt_cpu *this_cc = this_crypt_config(cc);
+       unsigned key_index = ctx->sector & (cc->tfms_count - 1);
+
+       if (!this_cc->req)
+               this_cc->req = mempool_alloc(cc->req_pool, GFP_NOIO);
+
+       ablkcipher_request_set_tfm(this_cc->req, this_cc->tfms[key_index]);
+       ablkcipher_request_set_callback(this_cc->req,
+           CRYPTO_TFM_REQ_MAY_BACKLOG | CRYPTO_TFM_REQ_MAY_SLEEP,
+           kcryptd_async_done, dmreq_of_req(cc, this_cc->req));
 }
 
 /*
  * Encrypt / decrypt data from one bio to another one (can be the same one)
  */
 static int crypt_convert(struct crypt_config *cc,
-                         struct convert_context *ctx)
+                        struct convert_context *ctx)
 {
-       int r = 0;
+       struct crypt_cpu *this_cc = this_crypt_config(cc);
+       int r;
+
+       atomic_set(&ctx->pending, 1);
 
        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 = {
-                       .page = bv_in->bv_page,
-                       .offset = bv_in->bv_offset + ctx->offset_in,
-                       .length = 1 << SECTOR_SHIFT
-               };
-               struct scatterlist sg_out = {
-                       .page = bv_out->bv_page,
-                       .offset = bv_out->bv_offset + ctx->offset_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++;
+               crypt_alloc_req(cc, ctx);
+
+               atomic_inc(&ctx->pending);
+
+               r = crypt_convert_block(cc, ctx, this_cc->req);
+
+               switch (r) {
+               /* async */
+               case -EBUSY:
+                       wait_for_completion(&ctx->restart);
+                       INIT_COMPLETION(ctx->restart);
+                       /* fall through*/
+               case -EINPROGRESS:
+                       this_cc->req = NULL;
+                       ctx->sector++;
+                       continue;
+
+               /* sync */
+               case 0:
+                       atomic_dec(&ctx->pending);
+                       ctx->sector++;
+                       cond_resched();
+                       continue;
+
+               /* error */
+               default:
+                       atomic_dec(&ctx->pending);
+                       return r;
                }
-
-               r = crypt_convert_scatterlist(cc, &sg_out, &sg_in, sg_in.length,
-                                             ctx->write, ctx->sector);
-               if (r < 0)
-                       break;
-
-               ctx->sector++;
        }
 
-       return r;
+       return 0;
 }
 
- static void dm_crypt_bio_destructor(struct bio *bio)
- {
-       struct crypt_io *io = bio->bi_private;
+static void dm_crypt_bio_destructor(struct bio *bio)
+{
+       struct dm_crypt_io *io = bio->bi_private;
        struct crypt_config *cc = io->target->private;
 
        bio_free(bio, cc->bs);
- }
+}
 
 /*
  * Generate a new unfragmented bio with the given size
  * This should never violate the device limitations
- * May return a smaller bio when running out of pages
+ * May return a smaller bio when running out of pages, indicated by
+ * *out_of_pages set to 1.
  */
-static struct bio *
-crypt_alloc_buffer(struct crypt_config *cc, unsigned int size,
-                   struct bio *base_bio, unsigned int *bio_vec_idx)
+static struct bio *crypt_alloc_buffer(struct dm_crypt_io *io, unsigned size,
+                                     unsigned *out_of_pages)
 {
+       struct crypt_config *cc = io->target->private;
        struct bio *clone;
        unsigned int nr_iovecs = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
        gfp_t gfp_mask = GFP_NOIO | __GFP_HIGHMEM;
-       unsigned int i;
-
-       if (base_bio) {
-               clone = bio_alloc_bioset(GFP_NOIO, base_bio->bi_max_vecs, cc->bs);
-               __bio_clone(clone, base_bio);
-       } else
-               clone = bio_alloc_bioset(GFP_NOIO, nr_iovecs, cc->bs);
+       unsigned i, len;
+       struct page *page;
 
+       clone = bio_alloc_bioset(GFP_NOIO, nr_iovecs, cc->bs);
        if (!clone)
                return NULL;
 
-       clone->bi_destructor = dm_crypt_bio_destructor;
-
-       /* if the last bio was not complete, continue where that one ended */
-       clone->bi_idx = *bio_vec_idx;
-       clone->bi_vcnt = *bio_vec_idx;
-       clone->bi_size = 0;
-       clone->bi_flags &= ~(1 << BIO_SEG_VALID);
-
-       /* clone->bi_idx pages have already been allocated */
-       size -= clone->bi_idx * PAGE_SIZE;
-
-       for (i = clone->bi_idx; i < nr_iovecs; i++) {
-               struct bio_vec *bv = bio_iovec_idx(clone, i);
+       clone_init(io, clone);
+       *out_of_pages = 0;
 
-               bv->bv_page = mempool_alloc(cc->page_pool, gfp_mask);
-               if (!bv->bv_page)
+       for (i = 0; i < nr_iovecs; i++) {
+               page = mempool_alloc(cc->page_pool, gfp_mask);
+               if (!page) {
+                       *out_of_pages = 1;
                        break;
+               }
 
                /*
                 * if additional pages cannot be allocated without waiting,
                 * return a partially allocated bio, the caller will then try
                 * to allocate additional bios while submitting this partial bio
                 */
-               if ((i - clone->bi_idx) == (MIN_BIO_PAGES - 1))
+               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) {
@@ -386,42 +870,15 @@ crypt_alloc_buffer(struct crypt_config *cc, unsigned int size,
                return NULL;
        }
 
-       /*
-        * Remember the last bio_vec allocated to be able
-        * to correctly continue after the splitting.
-        */
-       *bio_vec_idx = clone->bi_vcnt;
-
        return clone;
 }
 
-static void crypt_free_buffer_pages(struct crypt_config *cc,
-                                    struct bio *clone, unsigned int bytes)
+static void crypt_free_buffer_pages(struct crypt_config *cc, struct bio *clone)
 {
-       unsigned int i, start, end;
+       unsigned int i;
        struct bio_vec *bv;
 
-       /*
-        * This is ugly, but Jens Axboe thinks that using bi_idx in the
-        * endio function is too dangerous at the moment, so I calculate the
-        * correct position using bi_vcnt and bi_size.
-        * The bv_offset and bv_len fields might already be modified but we
-        * know that we always allocated whole pages.
-        * A fix to the bi_idx issue in the kernel is in the works, so
-        * we will hopefully be able to revert to the cleaner solution soon.
-        */
-       i = clone->bi_vcnt - 1;
-       bv = bio_iovec_idx(clone, i);
-       end = (i << PAGE_SHIFT) + (bv->bv_offset + bv->bv_len) - clone->bi_size;
-       start = end - bytes;
-
-       start >>= PAGE_SHIFT;
-       if (!clone->bi_size)
-               end = clone->bi_vcnt;
-       else
-               end >>= PAGE_SHIFT;
-
-       for (i = start; i < end; i++) {
+       for (i = 0; i < clone->bi_vcnt; i++) {
                bv = bio_iovec_idx(clone, i);
                BUG_ON(!bv->bv_page);
                mempool_free(bv->bv_page, cc->page_pool);
@@ -429,80 +886,100 @@ static void crypt_free_buffer_pages(struct crypt_config *cc,
        }
 }
 
+static struct dm_crypt_io *crypt_io_alloc(struct dm_target *ti,
+                                         struct bio *bio, sector_t sector)
+{
+       struct crypt_config *cc = ti->private;
+       struct dm_crypt_io *io;
+
+       io = mempool_alloc(cc->io_pool, GFP_NOIO);
+       io->target = ti;
+       io->base_bio = bio;
+       io->sector = sector;
+       io->error = 0;
+       io->base_io = NULL;
+       atomic_set(&io->pending, 0);
+
+       return io;
+}
+
+static void crypt_inc_pending(struct dm_crypt_io *io)
+{
+       atomic_inc(&io->pending);
+}
+
 /*
  * One of the bios was finished. Check for completion of
  * the whole request and correctly clean up the buffer.
+ * If base_io is set, wait for the last fragment to complete.
  */
-static void dec_pending(struct 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;
+       struct bio *base_bio = io->base_bio;
+       struct dm_crypt_io *base_io = io->base_io;
+       int error = io->error;
 
        if (!atomic_dec_and_test(&io->pending))
                return;
 
-       if (io->first_clone)
-               bio_put(io->first_clone);
-
-       bio_endio(io->base_bio, io->base_bio->bi_size, io->error);
-
        mempool_free(io, cc->io_pool);
+
+       if (likely(!base_io))
+               bio_endio(base_bio, error);
+       else {
+               if (error && !base_io->error)
+                       base_io->error = error;
+               crypt_dec_pending(base_io);
+       }
 }
 
 /*
- * kcryptd:
+ * kcryptd/kcryptd_io:
  *
  * Needed because it would be very unwise to do decryption in an
  * interrupt context.
+ *
+ * kcryptd performs the actual encryption or decryption.
+ *
+ * kcryptd_io performs the IO submission.
+ *
+ * They must be separated as otherwise the final stages could be
+ * starved by new requests which can block in the first stages due
+ * to memory allocation.
+ *
+ * The work is done per CPU global for all dm-crypt instances.
+ * They should not depend on each other and do not block.
  */
-static struct workqueue_struct *_kcryptd_workqueue;
-static void kcryptd_do_work(struct work_struct *work);
-
-static void kcryptd_queue_io(struct crypt_io *io)
-{
-       INIT_WORK(&io->work, kcryptd_do_work);
-       queue_work(_kcryptd_workqueue, &io->work);
-}
-
-static int crypt_endio(struct bio *clone, unsigned int done, int error)
+static void crypt_endio(struct bio *clone, int error)
 {
-       struct crypt_io *io = clone->bi_private;
+       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, even if
-        * it's only a partially completed write
+        * free the processed pages
         */
-       if (!read_io)
-               crypt_free_buffer_pages(cc, clone, done);
+       if (rw == WRITE)
+               crypt_free_buffer_pages(cc, clone);
 
-       /* keep going - not finished yet */
-       if (unlikely(clone->bi_size))
-               return 1;
-
-       if (!read_io)
-               goto out;
+       bio_put(clone);
 
-       if (unlikely(!bio_flagged(clone, BIO_UPTODATE))) {
-               error = -EIO;
-               goto out;
+       if (rw == READ && !error) {
+               kcryptd_queue_crypt(io);
+               return;
        }
 
-       bio_put(clone);
-       io->post_process = 1;
-       kcryptd_queue_io(io);
-       return 0;
+       if (unlikely(error))
+               io->error = error;
 
-out:
-       bio_put(clone);
-       dec_pending(io, error);
-       return error;
+       crypt_dec_pending(io);
 }
 
-static void clone_init(struct crypt_io *io, struct bio *clone)
+static void clone_init(struct dm_crypt_io *io, struct bio *clone)
 {
        struct crypt_config *cc = io->target->private;
 
@@ -510,124 +987,250 @@ static void clone_init(struct crypt_io *io, struct bio *clone)
        clone->bi_end_io  = crypt_endio;
        clone->bi_bdev    = cc->dev->bdev;
        clone->bi_rw      = io->base_bio->bi_rw;
+       clone->bi_destructor = dm_crypt_bio_destructor;
 }
 
-static void process_read(struct crypt_io *io)
+static int kcryptd_io_read(struct dm_crypt_io *io, gfp_t gfp)
 {
        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);
+       /*
+        * The block layer might modify the bvec array, so always
+        * copy the required bvecs because we need the original
+        * one in order to decrypt the whole bio data *afterwards*.
+        */
+       clone = bio_alloc_bioset(gfp, bio_segments(base_bio), cc->bs);
+       if (!clone)
+               return 1;
+
+       crypt_inc_pending(io);
+
+       clone_init(io, clone);
+       clone->bi_idx = 0;
+       clone->bi_vcnt = bio_segments(base_bio);
+       clone->bi_size = base_bio->bi_size;
+       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);
+       return 0;
+}
+
+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) {
+               crypt_inc_pending(io);
+               if (kcryptd_io_read(io, GFP_NOIO))
+                       io->error = -ENOMEM;
+               crypt_dec_pending(io);
+       } else
+               kcryptd_io_write(io);
+}
+
+static void kcryptd_queue_io(struct dm_crypt_io *io)
+{
+       struct crypt_config *cc = io->target->private;
+
+       INIT_WORK(&io->work, kcryptd_io);
+       queue_work(cc->io_queue, &io->work);
+}
+
+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;
 
-       /*
-        * The block layer might modify the bvec array, so always
-        * copy the required bvecs because we need the original
-        * one in order to decrypt the whole bio data *afterwards*.
-        */
-       clone = bio_alloc_bioset(GFP_NOIO, bio_segments(base_bio), cc->bs);
-       if (unlikely(!clone)) {
-               dec_pending(io, -ENOMEM);
+       if (unlikely(error < 0)) {
+               crypt_free_buffer_pages(cc, clone);
+               bio_put(clone);
+               io->error = -EIO;
+               crypt_dec_pending(io);
                return;
        }
 
-       clone_init(io, clone);
-       clone->bi_destructor = dm_crypt_bio_destructor;
-       clone->bi_idx = 0;
-       clone->bi_vcnt = bio_segments(base_bio);
-       clone->bi_size = base_bio->bi_size;
-       clone->bi_sector = cc->start + sector;
-       memcpy(clone->bi_io_vec, bio_iovec(base_bio),
-              sizeof(struct bio_vec) * clone->bi_vcnt);
+       /* crypt_convert should have filled the clone bio */
+       BUG_ON(io->ctx.idx_out < clone->bi_vcnt);
 
-       generic_make_request(clone);
+       clone->bi_sector = cc->start + io->sector;
+
+       if (async)
+               kcryptd_queue_io(io);
+       else
+               generic_make_request(clone);
 }
 
-static void process_write(struct crypt_io *io)
+static void kcryptd_crypt_write_convert(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;
-       unsigned bvec_idx = 0;
-
-       atomic_inc(&io->pending);
+       struct dm_crypt_io *new_io;
+       int crypt_finished;
+       unsigned out_of_pages = 0;
+       unsigned remaining = io->base_bio->bi_size;
+       sector_t sector = io->sector;
+       int r;
 
-       crypt_convert_init(cc, &ctx, NULL, base_bio, sector, 1);
+       /*
+        * Prevent io from disappearing until this function completes.
+        */
+       crypt_inc_pending(io);
+       crypt_convert_init(cc, &io->ctx, NULL, io->base_bio, sector);
 
        /*
         * The allocated buffers can be smaller than the whole bio,
         * so repeat the whole process until all the data can be handled.
         */
        while (remaining) {
-               clone = crypt_alloc_buffer(cc, base_bio->bi_size,
-                                          io->first_clone, &bvec_idx);
+               clone = crypt_alloc_buffer(io, remaining, &out_of_pages);
                if (unlikely(!clone)) {
-                       dec_pending(io, -ENOMEM);
-                       return;
+                       io->error = -ENOMEM;
+                       break;
                }
 
-               ctx.bio_out = clone;
+               io->ctx.bio_out = clone;
+               io->ctx.idx_out = 0;
+
+               remaining -= clone->bi_size;
+               sector += bio_sectors(clone);
+
+               crypt_inc_pending(io);
+               r = crypt_convert(cc, &io->ctx);
+               crypt_finished = atomic_dec_and_test(&io->ctx.pending);
+
+               /* Encryption was already finished, submit io now */
+               if (crypt_finished) {
+                       kcryptd_crypt_write_io_submit(io, r, 0);
+
+                       /*
+                        * If there was an error, do not try next fragments.
+                        * For async, error is processed in async handler.
+                        */
+                       if (unlikely(r < 0))
+                               break;
 
-               if (unlikely(crypt_convert(cc, &ctx) < 0)) {
-                       crypt_free_buffer_pages(cc, clone, clone->bi_size);
-                       bio_put(clone);
-                       dec_pending(io, -EIO);
-                       return;
+                       io->sector = sector;
                }
 
-               clone_init(io, clone);
-               clone->bi_sector = cc->start + sector;
+               /*
+                * Out of memory -> run queues
+                * But don't wait if split was due to the io size restriction
+                */
+               if (unlikely(out_of_pages))
+                       congestion_wait(BLK_RW_ASYNC, HZ/100);
+
+               /*
+                * With async crypto it is unsafe to share the crypto context
+                * between fragments, so switch to a new dm_crypt_io structure.
+                */
+               if (unlikely(!crypt_finished && remaining)) {
+                       new_io = crypt_io_alloc(io->target, io->base_bio,
+                                               sector);
+                       crypt_inc_pending(new_io);
+                       crypt_convert_init(cc, &new_io->ctx, NULL,
+                                          io->base_bio, sector);
+                       new_io->ctx.idx_in = io->ctx.idx_in;
+                       new_io->ctx.offset_in = io->ctx.offset_in;
 
-               if (!io->first_clone) {
                        /*
-                        * hold a reference to the first clone, because it
-                        * holds the bio_vec array and that can't be freed
-                        * before all other clones are released
+                        * Fragments after the first use the base_io
+                        * pending count.
                         */
-                       bio_get(clone);
-                       io->first_clone = clone;
+                       if (!io->base_io)
+                               new_io->base_io = io;
+                       else {
+                               new_io->base_io = io->base_io;
+                               crypt_inc_pending(io->base_io);
+                               crypt_dec_pending(io);
+                       }
+
+                       io = new_io;
                }
+       }
 
-               remaining -= clone->bi_size;
-               sector += bio_sectors(clone);
+       crypt_dec_pending(io);
+}
 
-               /* prevent bio_put of first_clone */
-               if (remaining)
-                       atomic_inc(&io->pending);
+static void kcryptd_crypt_read_done(struct dm_crypt_io *io, int error)
+{
+       if (unlikely(error < 0))
+               io->error = -EIO;
 
-               generic_make_request(clone);
+       crypt_dec_pending(io);
+}
 
-               /* out of memory -> run queues */
-               if (remaining)
-                       congestion_wait(bio_data_dir(clone), HZ/100);
-       }
+static void kcryptd_crypt_read_convert(struct dm_crypt_io *io)
+{
+       struct crypt_config *cc = io->target->private;
+       int r = 0;
+
+       crypt_inc_pending(io);
+
+       crypt_convert_init(cc, &io->ctx, io->base_bio, io->base_bio,
+                          io->sector);
+
+       r = crypt_convert(cc, &io->ctx);
+
+       if (atomic_dec_and_test(&io->ctx.pending))
+               kcryptd_crypt_read_done(io, r);
+
+       crypt_dec_pending(io);
 }
 
-static void process_read_endio(struct crypt_io *io)
+static void kcryptd_async_done(struct crypto_async_request *async_req,
+                              int error)
 {
+       struct dm_crypt_request *dmreq = async_req->data;
+       struct convert_context *ctx = dmreq->ctx;
+       struct dm_crypt_io *io = container_of(ctx, struct dm_crypt_io, ctx);
        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);
+       if (error == -EINPROGRESS) {
+               complete(&ctx->restart);
+               return;
+       }
+
+       if (!error && cc->iv_gen_ops && cc->iv_gen_ops->post)
+               error = cc->iv_gen_ops->post(cc, iv_of_dmreq(cc, dmreq), dmreq);
+
+       mempool_free(req_of_dmreq(cc, dmreq), cc->req_pool);
 
-       dec_pending(io, crypt_convert(cc, &ctx));
+       if (!atomic_dec_and_test(&ctx->pending))
+               return;
+
+       if (bio_data_dir(io->base_bio) == READ)
+               kcryptd_crypt_read_done(io, error);
+       else
+               kcryptd_crypt_write_io_submit(io, error, 1);
 }
 
-static void kcryptd_do_work(struct work_struct *work)
+static void kcryptd_crypt(struct work_struct *work)
 {
-       struct crypt_io *io = container_of(work, struct crypt_io, work);
+       struct dm_crypt_io *io = container_of(work, struct dm_crypt_io, work);
 
-       if (io->post_process)
-               process_read_endio(io);
-       else if (bio_data_dir(io->base_bio) == READ)
-               process_read(io);
+       if (bio_data_dir(io->base_bio) == READ)
+               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);
 }
 
 /*
@@ -671,250 +1274,471 @@ static void crypt_encode_key(char *hex, u8 *key, unsigned int size)
        }
 }
 
+static void crypt_free_tfms(struct crypt_config *cc, int cpu)
+{
+       struct crypt_cpu *cpu_cc = per_cpu_ptr(cc->cpu, cpu);
+       unsigned i;
+
+       for (i = 0; i < cc->tfms_count; i++)
+               if (cpu_cc->tfms[i] && !IS_ERR(cpu_cc->tfms[i])) {
+                       crypto_free_ablkcipher(cpu_cc->tfms[i]);
+                       cpu_cc->tfms[i] = NULL;
+               }
+}
+
+static int crypt_alloc_tfms(struct crypt_config *cc, int cpu, char *ciphermode)
+{
+       struct crypt_cpu *cpu_cc = per_cpu_ptr(cc->cpu, cpu);
+       unsigned i;
+       int err;
+
+       for (i = 0; i < cc->tfms_count; i++) {
+               cpu_cc->tfms[i] = crypto_alloc_ablkcipher(ciphermode, 0, 0);
+               if (IS_ERR(cpu_cc->tfms[i])) {
+                       err = PTR_ERR(cpu_cc->tfms[i]);
+                       crypt_free_tfms(cc, cpu);
+                       return err;
+               }
+       }
+
+       return 0;
+}
+
+static int crypt_setkey_allcpus(struct crypt_config *cc)
+{
+       unsigned subkey_size = cc->key_size >> ilog2(cc->tfms_count);
+       int cpu, err = 0, i, r;
+
+       for_each_possible_cpu(cpu) {
+               for (i = 0; i < cc->tfms_count; i++) {
+                       r = crypto_ablkcipher_setkey(per_cpu_ptr(cc->cpu, cpu)->tfms[i],
+                                                    cc->key + (i * subkey_size), subkey_size);
+                       if (r)
+                               err = r;
+               }
+       }
+
+       return err;
+}
+
 static int crypt_set_key(struct crypt_config *cc, char *key)
 {
-       unsigned key_size = strlen(key) >> 1;
+       int r = -EINVAL;
+       int key_string_len = strlen(key);
 
-       if (cc->key_size && cc->key_size != key_size)
-               return -EINVAL;
+       /* The key size may not be changed. */
+       if (cc->key_size != (key_string_len >> 1))
+               goto out;
 
-       cc->key_size = key_size; /* initial settings */
+       /* Hyphen (which gives a key_size of zero) means there is no key. */
+       if (!cc->key_size && strcmp(key, "-"))
+               goto out;
 
-       if ((!key_size && strcmp(key, "-")) ||
-           (key_size && crypt_decode_key(cc->key, key, key_size) < 0))
-               return -EINVAL;
+       if (cc->key_size && crypt_decode_key(cc->key, key, cc->key_size) < 0)
+               goto out;
 
        set_bit(DM_CRYPT_KEY_VALID, &cc->flags);
 
-       return 0;
+       r = crypt_setkey_allcpus(cc);
+
+out:
+       /* Hex key string not needed after here, so wipe it. */
+       memset(key, '0', key_string_len);
+
+       return r;
 }
 
 static int crypt_wipe_key(struct crypt_config *cc)
 {
        clear_bit(DM_CRYPT_KEY_VALID, &cc->flags);
        memset(&cc->key, 0, cc->key_size * sizeof(u8));
-       return 0;
+
+       return crypt_setkey_allcpus(cc);
 }
 
-/*
- * Construct an encryption mapping:
- * <cipher> <key> <iv_offset> <dev_path> <start>
- */
-static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
+static void crypt_dtr(struct dm_target *ti)
 {
-       struct crypt_config *cc;
-       struct crypto_blkcipher *tfm;
-       char *tmp;
-       char *cipher;
-       char *chainmode;
-       char *ivmode;
-       char *ivopts;
-       unsigned int key_size;
-       unsigned long long tmpll;
+       struct crypt_config *cc = ti->private;
+       struct crypt_cpu *cpu_cc;
+       int cpu;
 
-       if (argc != 5) {
-               ti->error = "Not enough arguments";
+       ti->private = NULL;
+
+       if (!cc)
+               return;
+
+       if (cc->io_queue)
+               destroy_workqueue(cc->io_queue);
+       if (cc->crypt_queue)
+               destroy_workqueue(cc->crypt_queue);
+
+       if (cc->cpu)
+               for_each_possible_cpu(cpu) {
+                       cpu_cc = per_cpu_ptr(cc->cpu, cpu);
+                       if (cpu_cc->req)
+                               mempool_free(cpu_cc->req, cc->req_pool);
+                       crypt_free_tfms(cc, cpu);
+               }
+
+       if (cc->bs)
+               bioset_free(cc->bs);
+
+       if (cc->page_pool)
+               mempool_destroy(cc->page_pool);
+       if (cc->req_pool)
+               mempool_destroy(cc->req_pool);
+       if (cc->io_pool)
+               mempool_destroy(cc->io_pool);
+
+       if (cc->iv_gen_ops && cc->iv_gen_ops->dtr)
+               cc->iv_gen_ops->dtr(cc);
+
+       if (cc->dev)
+               dm_put_device(ti, cc->dev);
+
+       if (cc->cpu)
+               free_percpu(cc->cpu);
+
+       kzfree(cc->cipher);
+       kzfree(cc->cipher_string);
+
+       /* Must zero key material before freeing */
+       kzfree(cc);
+}
+
+static int crypt_ctr_cipher(struct dm_target *ti,
+                           char *cipher_in, char *key)
+{
+       struct crypt_config *cc = ti->private;
+       char *tmp, *cipher, *chainmode, *ivmode, *ivopts, *keycount;
+       char *cipher_api = NULL;
+       int cpu, ret = -EINVAL;
+
+       /* Convert to crypto api definition? */
+       if (strchr(cipher_in, '(')) {
+               ti->error = "Bad cipher specification";
+               return -EINVAL;
+       }
+
+       cc->cipher_string = kstrdup(cipher_in, GFP_KERNEL);
+       if (!cc->cipher_string)
+               goto bad_mem;
+
+       /*
+        * Legacy dm-crypt cipher specification
+        * cipher[:keycount]-mode-iv:ivopts
+        */
+       tmp = cipher_in;
+       keycount = strsep(&tmp, "-");
+       cipher = strsep(&keycount, ":");
+
+       if (!keycount)
+               cc->tfms_count = 1;
+       else if (sscanf(keycount, "%u", &cc->tfms_count) != 1 ||
+                !is_power_of_2(cc->tfms_count)) {
+               ti->error = "Bad cipher key count specification";
                return -EINVAL;
        }
+       cc->key_parts = cc->tfms_count;
+
+       cc->cipher = kstrdup(cipher, GFP_KERNEL);
+       if (!cc->cipher)
+               goto bad_mem;
 
-       tmp = argv[0];
-       cipher = strsep(&tmp, "-");
        chainmode = strsep(&tmp, "-");
        ivopts = strsep(&tmp, "-");
        ivmode = strsep(&ivopts, ":");
 
        if (tmp)
-               DMWARN("Unexpected additional cipher options");
-
-       key_size = strlen(argv[1]) >> 1;
-
-       cc = kzalloc(sizeof(*cc) + key_size * sizeof(u8), GFP_KERNEL);
-       if (cc == NULL) {
-               ti->error =
-                       "Cannot allocate transparent encryption context";
-               return -ENOMEM;
-       }
-
-       if (crypt_set_key(cc, argv[1])) {
-               ti->error = "Error decoding key";
-               goto bad1;
+               DMWARN("Ignoring unexpected additional cipher options");
+
+       cc->cpu = __alloc_percpu(sizeof(*(cc->cpu)) +
+                                cc->tfms_count * sizeof(*(cc->cpu->tfms)),
+                                __alignof__(struct crypt_cpu));
+       if (!cc->cpu) {
+               ti->error = "Cannot allocate per cpu state";
+               goto bad_mem;
        }
 
-       /* Compatiblity mode for old dm-crypt cipher strings */
-       if (!chainmode || (strcmp(chainmode, "plain") == 0 && !ivmode)) {
+       /*
+        * For compatibility with the original dm-crypt mapping format, if
+        * only the cipher name is supplied, use cbc-plain.
+        */
+       if (!chainmode || (!strcmp(chainmode, "plain") && !ivmode)) {
                chainmode = "cbc";
                ivmode = "plain";
        }
 
        if (strcmp(chainmode, "ecb") && !ivmode) {
-               ti->error = "This chaining mode requires an IV mechanism";
-               goto bad1;
+               ti->error = "IV mechanism required";
+               return -EINVAL;
        }
 
-       if (snprintf(cc->cipher, CRYPTO_MAX_ALG_NAME, "%s(%s)", chainmode, 
-                    cipher) >= CRYPTO_MAX_ALG_NAME) {
-               ti->error = "Chain mode + cipher name is too long";
-               goto bad1;
+       cipher_api = kmalloc(CRYPTO_MAX_ALG_NAME, GFP_KERNEL);
+       if (!cipher_api)
+               goto bad_mem;
+
+       ret = snprintf(cipher_api, CRYPTO_MAX_ALG_NAME,
+                      "%s(%s)", chainmode, cipher);
+       if (ret < 0) {
+               kfree(cipher_api);
+               goto bad_mem;
        }
 
-       tfm = crypto_alloc_blkcipher(cc->cipher, 0, CRYPTO_ALG_ASYNC);
-       if (IS_ERR(tfm)) {
-               ti->error = "Error allocating crypto tfm";
-               goto bad1;
+       /* Allocate cipher */
+       for_each_possible_cpu(cpu) {
+               ret = crypt_alloc_tfms(cc, cpu, cipher_api);
+               if (ret < 0) {
+                       ti->error = "Error allocating crypto tfm";
+                       goto bad;
+               }
        }
 
-       strcpy(cc->cipher, cipher);
-       strcpy(cc->chainmode, chainmode);
-       cc->tfm = tfm;
+       /* Initialize and set key */
+       ret = crypt_set_key(cc, key);
+       if (ret < 0) {
+               ti->error = "Error decoding and setting key";
+               goto bad;
+       }
 
-       /*
-        * Choose ivmode. Valid modes: "plain", "essiv:<esshash>".
-        * See comments at iv code
-        */
+       /* Initialize IV */
+       cc->iv_size = crypto_ablkcipher_ivsize(any_tfm(cc));
+       if (cc->iv_size)
+               /* at least a 64 bit sector number should fit in our buffer */
+               cc->iv_size = max(cc->iv_size,
+                                 (unsigned int)(sizeof(u64) / sizeof(u8)));
+       else if (ivmode) {
+               DMWARN("Selected cipher does not support IVs");
+               ivmode = NULL;
+       }
 
+       /* Choose ivmode, see comments at iv code. */
        if (ivmode == NULL)
                cc->iv_gen_ops = NULL;
        else if (strcmp(ivmode, "plain") == 0)
                cc->iv_gen_ops = &crypt_iv_plain_ops;
+       else if (strcmp(ivmode, "plain64") == 0)
+               cc->iv_gen_ops = &crypt_iv_plain64_ops;
        else if (strcmp(ivmode, "essiv") == 0)
                cc->iv_gen_ops = &crypt_iv_essiv_ops;
-       else {
+       else if (strcmp(ivmode, "benbi") == 0)
+               cc->iv_gen_ops = &crypt_iv_benbi_ops;
+       else if (strcmp(ivmode, "null") == 0)
+               cc->iv_gen_ops = &crypt_iv_null_ops;
+       else if (strcmp(ivmode, "lmk") == 0) {
+               cc->iv_gen_ops = &crypt_iv_lmk_ops;
+               /* Version 2 and 3 is recognised according
+                * to length of provided multi-key string.
+                * If present (version 3), last key is used as IV seed.
+                */
+               if (cc->key_size % cc->key_parts)
+                       cc->key_parts++;
+       } else {
+               ret = -EINVAL;
                ti->error = "Invalid IV mode";
-               goto bad2;
+               goto bad;
        }
 
-       if (cc->iv_gen_ops && cc->iv_gen_ops->ctr &&
-           cc->iv_gen_ops->ctr(cc, ti, ivopts) < 0)
-               goto bad2;
+       /* Allocate IV */
+       if (cc->iv_gen_ops && cc->iv_gen_ops->ctr) {
+               ret = cc->iv_gen_ops->ctr(cc, ti, ivopts);
+               if (ret < 0) {
+                       ti->error = "Error creating IV";
+                       goto bad;
+               }
+       }
 
-       cc->iv_size = crypto_blkcipher_ivsize(tfm);
-       if (cc->iv_size)
-               /* at least a 64 bit sector number should fit in our buffer */
-               cc->iv_size = max(cc->iv_size,
-                                 (unsigned int)(sizeof(u64) / sizeof(u8)));
-       else {
-               if (cc->iv_gen_ops) {
-                       DMWARN("Selected cipher does not support IVs");
-                       if (cc->iv_gen_ops->dtr)
-                               cc->iv_gen_ops->dtr(cc);
-                       cc->iv_gen_ops = NULL;
+       /* Initialize IV (set keys for ESSIV etc) */
+       if (cc->iv_gen_ops && cc->iv_gen_ops->init) {
+               ret = cc->iv_gen_ops->init(cc);
+               if (ret < 0) {
+                       ti->error = "Error initialising IV";
+                       goto bad;
                }
        }
 
+       ret = 0;
+bad:
+       kfree(cipher_api);
+       return ret;
+
+bad_mem:
+       ti->error = "Cannot allocate cipher strings";
+       return -ENOMEM;
+}
+
+/*
+ * Construct an encryption mapping:
+ * <cipher> <key> <iv_offset> <dev_path> <start>
+ */
+static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
+{
+       struct crypt_config *cc;
+       unsigned int key_size, opt_params;
+       unsigned long long tmpll;
+       int ret;
+       struct dm_arg_set as;
+       const char *opt_string;
+
+       static struct dm_arg _args[] = {
+               {0, 1, "Invalid number of feature args"},
+       };
+
+       if (argc < 5) {
+               ti->error = "Not enough arguments";
+               return -EINVAL;
+       }
+
+       key_size = strlen(argv[1]) >> 1;
+
+       cc = kzalloc(sizeof(*cc) + key_size * sizeof(u8), GFP_KERNEL);
+       if (!cc) {
+               ti->error = "Cannot allocate encryption context";
+               return -ENOMEM;
+       }
+       cc->key_size = key_size;
+
+       ti->private = cc;
+       ret = crypt_ctr_cipher(ti, argv[0], argv[1]);
+       if (ret < 0)
+               goto bad;
+
+       ret = -ENOMEM;
        cc->io_pool = mempool_create_slab_pool(MIN_IOS, _crypt_io_pool);
        if (!cc->io_pool) {
                ti->error = "Cannot allocate crypt io mempool";
-               goto bad3;
+               goto bad;
+       }
+
+       cc->dmreq_start = sizeof(struct ablkcipher_request);
+       cc->dmreq_start += crypto_ablkcipher_reqsize(any_tfm(cc));
+       cc->dmreq_start = ALIGN(cc->dmreq_start, crypto_tfm_ctx_alignment());
+       cc->dmreq_start += crypto_ablkcipher_alignmask(any_tfm(cc)) &
+                          ~(crypto_tfm_ctx_alignment() - 1);
+
+       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;
        }
 
        cc->page_pool = mempool_create_page_pool(MIN_POOL_PAGES, 0);
        if (!cc->page_pool) {
                ti->error = "Cannot allocate page mempool";
-               goto bad4;
+               goto bad;
        }
 
-       cc->bs = bioset_create(MIN_IOS, MIN_IOS, 4);
+       cc->bs = bioset_create(MIN_IOS, 0);
        if (!cc->bs) {
                ti->error = "Cannot allocate crypt bioset";
-               goto bad_bs;
-       }
-
-       if (crypto_blkcipher_setkey(tfm, cc->key, key_size) < 0) {
-               ti->error = "Error setting key";
-               goto bad5;
+               goto bad;
        }
 
+       ret = -EINVAL;
        if (sscanf(argv[2], "%llu", &tmpll) != 1) {
                ti->error = "Invalid iv_offset sector";
-               goto bad5;
+               goto bad;
        }
        cc->iv_offset = tmpll;
 
+       if (dm_get_device(ti, argv[3], dm_table_get_mode(ti->table), &cc->dev)) {
+               ti->error = "Device lookup failed";
+               goto bad;
+       }
+
        if (sscanf(argv[4], "%llu", &tmpll) != 1) {
                ti->error = "Invalid device sector";
-               goto bad5;
+               goto bad;
        }
        cc->start = tmpll;
 
-       if (dm_get_device(ti, argv[3], cc->start, ti->len,
-                         dm_table_get_mode(ti->table), &cc->dev)) {
-               ti->error = "Device lookup failed";
-               goto bad5;
-       }
+       argv += 5;
+       argc -= 5;
 
-       if (ivmode && cc->iv_gen_ops) {
-               if (ivopts)
-                       *(ivopts - 1) = ':';
-               cc->iv_mode = kmalloc(strlen(ivmode) + 1, GFP_KERNEL);
-               if (!cc->iv_mode) {
-                       ti->error = "Error kmallocing iv_mode string";
-                       goto bad5;
-               }
-               strcpy(cc->iv_mode, ivmode);
-       } else
-               cc->iv_mode = NULL;
+       /* Optional parameters */
+       if (argc) {
+               as.argc = argc;
+               as.argv = argv;
 
-       ti->private = cc;
-       return 0;
+               ret = dm_read_arg_group(_args, &as, &opt_params, &ti->error);
+               if (ret)
+                       goto bad;
 
-bad5:
-       bioset_free(cc->bs);
-bad_bs:
-       mempool_destroy(cc->page_pool);
-bad4:
-       mempool_destroy(cc->io_pool);
-bad3:
-       if (cc->iv_gen_ops && cc->iv_gen_ops->dtr)
-               cc->iv_gen_ops->dtr(cc);
-bad2:
-       crypto_free_blkcipher(tfm);
-bad1:
-       /* Must zero key material before freeing */
-       memset(cc, 0, sizeof(*cc) + cc->key_size * sizeof(u8));
-       kfree(cc);
-       return -EINVAL;
-}
+               opt_string = dm_shift_arg(&as);
 
-static void crypt_dtr(struct dm_target *ti)
-{
-       struct crypt_config *cc = (struct crypt_config *) ti->private;
+               if (opt_params == 1 && opt_string &&
+                   !strcasecmp(opt_string, "allow_discards"))
+                       ti->num_discard_requests = 1;
+               else if (opt_params) {
+                       ret = -EINVAL;
+                       ti->error = "Invalid feature arguments";
+                       goto bad;
+               }
+       }
 
-       bioset_free(cc->bs);
-       mempool_destroy(cc->page_pool);
-       mempool_destroy(cc->io_pool);
+       ret = -ENOMEM;
+       cc->io_queue = alloc_workqueue("kcryptd_io",
+                                      WQ_NON_REENTRANT|
+                                      WQ_MEM_RECLAIM,
+                                      1);
+       if (!cc->io_queue) {
+               ti->error = "Couldn't create kcryptd io queue";
+               goto bad;
+       }
 
-       kfree(cc->iv_mode);
-       if (cc->iv_gen_ops && cc->iv_gen_ops->dtr)
-               cc->iv_gen_ops->dtr(cc);
-       crypto_free_blkcipher(cc->tfm);
-       dm_put_device(ti, cc->dev);
+       cc->crypt_queue = alloc_workqueue("kcryptd",
+                                         WQ_NON_REENTRANT|
+                                         WQ_CPU_INTENSIVE|
+                                         WQ_MEM_RECLAIM,
+                                         1);
+       if (!cc->crypt_queue) {
+               ti->error = "Couldn't create kcryptd queue";
+               goto bad;
+       }
 
-       /* Must zero key material before freeing */
-       memset(cc, 0, sizeof(*cc) + cc->key_size * sizeof(u8));
-       kfree(cc);
+       ti->num_flush_requests = 1;
+       return 0;
+
+bad:
+       crypt_dtr(ti);
+       return ret;
 }
 
 static int crypt_map(struct dm_target *ti, struct bio *bio,
                     union map_info *map_context)
 {
-       struct crypt_config *cc = ti->private;
-       struct crypt_io *io;
+       struct dm_crypt_io *io;
+       struct crypt_config *cc;
 
-       io = mempool_alloc(cc->io_pool, GFP_NOIO);
-       io->target = ti;
-       io->base_bio = bio;
-       io->first_clone = NULL;
-       io->error = io->post_process = 0;
-       atomic_set(&io->pending, 0);
-       kcryptd_queue_io(io);
+       /*
+        * If bio is REQ_FLUSH or REQ_DISCARD, just bypass crypt queues.
+        * - for REQ_FLUSH device-mapper core ensures that no IO is in-flight
+        * - for REQ_DISCARD caller must use flush if IO ordering matters
+        */
+       if (unlikely(bio->bi_rw & (REQ_FLUSH | REQ_DISCARD))) {
+               cc = ti->private;
+               bio->bi_bdev = cc->dev->bdev;
+               if (bio_sectors(bio))
+                       bio->bi_sector = cc->start + dm_target_offset(ti, bio->bi_sector);
+               return DM_MAPIO_REMAPPED;
+       }
 
-       return 0;
+       io = crypt_io_alloc(ti, bio, dm_target_offset(ti, bio->bi_sector));
+
+       if (bio_data_dir(io->base_bio) == READ) {
+               if (kcryptd_io_read(io, GFP_NOWAIT))
+                       kcryptd_queue_io(io);
+       } else
+               kcryptd_queue_crypt(io);
+
+       return DM_MAPIO_SUBMITTED;
 }
 
 static int crypt_status(struct dm_target *ti, status_type_t type,
                        char *result, unsigned int maxlen)
 {
-       struct crypt_config *cc = (struct crypt_config *) ti->private;
+       struct crypt_config *cc = ti->private;
        unsigned int sz = 0;
 
        switch (type) {
@@ -923,11 +1747,7 @@ static int crypt_status(struct dm_target *ti, status_type_t type,
                break;
 
        case STATUSTYPE_TABLE:
-               if (cc->iv_mode)
-                       DMEMIT("%s-%s-%s ", cc->cipher, cc->chainmode,
-                              cc->iv_mode);
-               else
-                       DMEMIT("%s-%s ", cc->cipher, cc->chainmode);
+               DMEMIT("%s ", cc->cipher_string);
 
                if (cc->key_size > 0) {
                        if ((maxlen - sz) < ((cc->key_size << 1) + 1))
@@ -943,6 +1763,10 @@ static int crypt_status(struct dm_target *ti, status_type_t type,
 
                DMEMIT(" %llu %s %llu", (unsigned long long)cc->iv_offset,
                                cc->dev->name, (unsigned long long)cc->start);
+
+               if (ti->num_discard_requests)
+                       DMEMIT(" 1 allow_discards");
+
                break;
        }
        return 0;
@@ -981,19 +1805,32 @@ static void crypt_resume(struct dm_target *ti)
 static int crypt_message(struct dm_target *ti, unsigned argc, char **argv)
 {
        struct crypt_config *cc = ti->private;
+       int ret = -EINVAL;
 
        if (argc < 2)
                goto error;
 
-       if (!strnicmp(argv[0], MESG_STR("key"))) {
+       if (!strcasecmp(argv[0], "key")) {
                if (!test_bit(DM_CRYPT_SUSPENDED, &cc->flags)) {
                        DMWARN("not suspended during key manipulation.");
                        return -EINVAL;
                }
-               if (argc == 3 && !strnicmp(argv[1], MESG_STR("set")))
-                       return crypt_set_key(cc, argv[2]);
-               if (argc == 2 && !strnicmp(argv[1], MESG_STR("wipe")))
+               if (argc == 3 && !strcasecmp(argv[1], "set")) {
+                       ret = crypt_set_key(cc, argv[2]);
+                       if (ret)
+                               return ret;
+                       if (cc->iv_gen_ops && cc->iv_gen_ops->init)
+                               ret = cc->iv_gen_ops->init(cc);
+                       return ret;
+               }
+               if (argc == 2 && !strcasecmp(argv[1], "wipe")) {
+                       if (cc->iv_gen_ops && cc->iv_gen_ops->wipe) {
+                               ret = cc->iv_gen_ops->wipe(cc);
+                               if (ret)
+                                       return ret;
+                       }
                        return crypt_wipe_key(cc);
+               }
        }
 
 error:
@@ -1001,9 +1838,32 @@ error:
        return -EINVAL;
 }
 
+static int crypt_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
+                      struct bio_vec *biovec, int max_size)
+{
+       struct crypt_config *cc = ti->private;
+       struct request_queue *q = bdev_get_queue(cc->dev->bdev);
+
+       if (!q->merge_bvec_fn)
+               return max_size;
+
+       bvm->bi_bdev = cc->dev->bdev;
+       bvm->bi_sector = cc->start + dm_target_offset(ti, bvm->bi_sector);
+
+       return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
+}
+
+static int crypt_iterate_devices(struct dm_target *ti,
+                                iterate_devices_callout_fn fn, void *data)
+{
+       struct crypt_config *cc = ti->private;
+
+       return fn(ti, cc->dev, cc->start, ti->len, data);
+}
+
 static struct target_type crypt_target = {
        .name   = "crypt",
-       .version= {1, 3, 0},
+       .version = {1, 11, 0},
        .module = THIS_MODULE,
        .ctr    = crypt_ctr,
        .dtr    = crypt_dtr,
@@ -1013,48 +1873,30 @@ static struct target_type crypt_target = {
        .preresume = crypt_preresume,
        .resume = crypt_resume,
        .message = crypt_message,
+       .merge  = crypt_merge,
+       .iterate_devices = crypt_iterate_devices,
 };
 
 static int __init dm_crypt_init(void)
 {
        int r;
 
-       _crypt_io_pool = kmem_cache_create("dm-crypt_io",
-                                          sizeof(struct crypt_io),
-                                          0, 0, NULL, NULL);
+       _crypt_io_pool = KMEM_CACHE(dm_crypt_io, 0);
        if (!_crypt_io_pool)
                return -ENOMEM;
 
-       _kcryptd_workqueue = create_workqueue("kcryptd");
-       if (!_kcryptd_workqueue) {
-               r = -ENOMEM;
-               DMERR("couldn't create kcryptd");
-               goto bad1;
-       }
-
        r = dm_register_target(&crypt_target);
        if (r < 0) {
                DMERR("register failed %d", r);
-               goto bad2;
+               kmem_cache_destroy(_crypt_io_pool);
        }
 
-       return 0;
-
-bad2:
-       destroy_workqueue(_kcryptd_workqueue);
-bad1:
-       kmem_cache_destroy(_crypt_io_pool);
        return r;
 }
 
 static void __exit dm_crypt_exit(void)
 {
-       int r = dm_unregister_target(&crypt_target);
-
-       if (r < 0)
-               DMERR("unregister failed %d", r);
-
-       destroy_workqueue(_kcryptd_workqueue);
+       dm_unregister_target(&crypt_target);
        kmem_cache_destroy(_crypt_io_pool);
 }