Merge branch 'dev' of https://github.com/smartxworks/fio
[fio.git] / verify.c
index ee9160c673c3fa4efab1a48dea8eedbe7f939473..f79ab43aab6b97b0721f52485fad31053e8d6ca9 100644 (file)
--- a/verify.c
+++ b/verify.c
 #include "crc/sha512.h"
 #include "crc/sha1.h"
 #include "crc/xxhash.h"
+#include "crc/sha3.h"
 
 static void populate_hdr(struct thread_data *td, struct io_u *io_u,
                         struct verify_header *hdr, unsigned int header_num,
                         unsigned int header_len);
-static void fill_hdr(struct thread_data *td, struct io_u *io_u,
-                    struct verify_header *hdr, unsigned int header_num,
-                    unsigned int header_len, uint64_t rand_seed);
 static void __fill_hdr(struct thread_data *td, struct io_u *io_u,
                       struct verify_header *hdr, unsigned int header_num,
                       unsigned int header_len, uint64_t rand_seed);
@@ -41,13 +39,14 @@ void fill_buffer_pattern(struct thread_data *td, void *p, unsigned int len)
        (void)cpy_pattern(td->o.buffer_pattern, td->o.buffer_pattern_bytes, p, len);
 }
 
-void __fill_buffer(struct thread_options *o, unsigned long seed, void *p,
-                  unsigned int len)
+static void __fill_buffer(struct thread_options *o, uint64_t seed, void *p,
+                         unsigned int len)
 {
        __fill_random_buf_percentage(seed, p, o->compress_percentage, len, len, o->buffer_pattern, o->buffer_pattern_bytes);
 }
 
-unsigned long fill_buffer(struct thread_data *td, void *p, unsigned int len)
+static uint64_t fill_buffer(struct thread_data *td, void *p,
+                           unsigned int len)
 {
        struct frand_state *fs = &td->verify_state;
        struct thread_options *o = &td->o;
@@ -56,7 +55,7 @@ unsigned long fill_buffer(struct thread_data *td, void *p, unsigned int len)
 }
 
 void fill_verify_pattern(struct thread_data *td, void *p, unsigned int len,
-                        struct io_u *io_u, unsigned long seed, int use_seed)
+                        struct io_u *io_u, uint64_t seed, int use_seed)
 {
        struct thread_options *o = &td->o;
 
@@ -88,15 +87,20 @@ static unsigned int get_hdr_inc(struct thread_data *td, struct io_u *io_u)
 {
        unsigned int hdr_inc;
 
+       /*
+        * If we use bs_unaligned, buflen can be larger than the verify
+        * interval (which just defaults to the smallest blocksize possible).
+        */
        hdr_inc = io_u->buflen;
-       if (td->o.verify_interval && td->o.verify_interval <= io_u->buflen)
+       if (td->o.verify_interval && td->o.verify_interval <= io_u->buflen &&
+           !td->o.bs_unaligned)
                hdr_inc = td->o.verify_interval;
 
        return hdr_inc;
 }
 
 static void fill_pattern_headers(struct thread_data *td, struct io_u *io_u,
-                                unsigned long seed, int use_seed)
+                                uint64_t seed, int use_seed)
 {
        unsigned int hdr_inc, header_num;
        struct verify_header *hdr;
@@ -171,6 +175,18 @@ static inline unsigned int __hdr_size(int verify_type)
        case VERIFY_SHA512:
                len = sizeof(struct vhdr_sha512);
                break;
+       case VERIFY_SHA3_224:
+               len = sizeof(struct vhdr_sha3_224);
+               break;
+       case VERIFY_SHA3_256:
+               len = sizeof(struct vhdr_sha3_256);
+               break;
+       case VERIFY_SHA3_384:
+               len = sizeof(struct vhdr_sha3_384);
+               break;
+       case VERIFY_SHA3_512:
+               len = sizeof(struct vhdr_sha3_512);
+               break;
        case VERIFY_XXHASH:
                len = sizeof(struct vhdr_xxhash);
                break;
@@ -225,39 +241,27 @@ struct vcont {
 };
 
 #define DUMP_BUF_SZ    255
-static int dump_buf_warned;
 
 static void dump_buf(char *buf, unsigned int len, unsigned long long offset,
                     const char *type, struct fio_file *f)
 {
-       char *ptr, fname[DUMP_BUF_SZ];
-       size_t buf_left = DUMP_BUF_SZ;
+       char *ptr, *fname;
+       char sep[2] = { FIO_OS_PATH_SEPARATOR, 0 };
        int ret, fd;
 
        ptr = strdup(f->file_name);
 
-       memset(fname, 0, sizeof(fname));
-       if (aux_path)
-               sprintf(fname, "%s%s", aux_path, FIO_OS_PATH_SEPARATOR);
-
-       strncpy(fname + strlen(fname), basename(ptr), buf_left - 1);
-
-       buf_left -= strlen(fname);
-       if (buf_left <= 0) {
-               if (!dump_buf_warned) {
-                       log_err("fio: verify failure dump buffer too small\n");
-                       dump_buf_warned = 1;
-               }
-               free(ptr);
-               return;
+       if (asprintf(&fname, "%s%s%s.%llu.%s", aux_path ? : "",
+                    aux_path ? sep : "", basename(ptr), offset, type) < 0) {
+               if (!fio_did_warn(FIO_WARN_VERIFY_BUF))
+                       log_err("fio: not enough memory for dump buffer filename\n");
+               goto free_ptr;
        }
 
-       snprintf(fname + strlen(fname), buf_left, ".%llu.%s", offset, type);
-
        fd = open(fname, O_CREAT | O_TRUNC | O_WRONLY, 0644);
        if (fd < 0) {
                perror("open verify buf file");
-               return;
+               goto free_fname;
        }
 
        while (len) {
@@ -274,6 +278,11 @@ static void dump_buf(char *buf, unsigned int len, unsigned long long offset,
 
        close(fd);
        log_err("       %s data dumped as %s\n", type, fname);
+
+free_fname:
+       free(fname);
+
+free_ptr:
        free(ptr);
 }
 
@@ -336,8 +345,10 @@ static void log_verify_failure(struct verify_header *hdr, struct vcont *vc)
 
        offset = vc->io_u->offset;
        offset += vc->hdr_num * hdr->len;
-       log_err("%.8s: verify failed at file %s offset %llu, length %u\n",
-                       vc->name, vc->io_u->file->file_name, offset, hdr->len);
+       log_err("%.8s: verify failed at file %s offset %llu, length %u"
+                       " (requested block: offset=%llu, length=%llu)\n",
+                       vc->name, vc->io_u->file->file_name, offset, hdr->len,
+                       vc->io_u->offset, vc->io_u->buflen);
 
        if (vc->good_crc && vc->bad_crc) {
                log_err("       Expected CRC: ");
@@ -373,7 +384,7 @@ static int verify_io_u_pattern(struct verify_header *hdr, struct vcont *vc)
        (void)paste_format_inplace(pattern, pattern_size,
                                   td->o.verify_fmt, td->o.verify_fmt_sz, io_u);
 
-       buf = (void *) hdr + header_size;
+       buf = (char *) hdr + header_size;
        len = get_hdr_inc(td, io_u) - header_size;
        mod = (get_hdr_inc(td, io_u) * vc->hdr_num + header_size) % pattern_size;
 
@@ -392,7 +403,8 @@ static int verify_io_u_pattern(struct verify_header *hdr, struct vcont *vc)
                                (unsigned char)pattern[mod],
                                bits);
                        log_err("fio: bad pattern block offset %u\n", i);
-                       dump_verify_buffers(hdr, vc);
+                       vc->name = "pattern";
+                       log_verify_failure(hdr, vc);
                        return EILSEQ;
                }
                mod++;
@@ -429,6 +441,84 @@ static int verify_io_u_xxhash(struct verify_header *hdr, struct vcont *vc)
        return EILSEQ;
 }
 
+static int verify_io_u_sha3(struct verify_header *hdr, struct vcont *vc,
+                           struct fio_sha3_ctx *sha3_ctx, uint8_t *sha,
+                           unsigned int sha_size, const char *name)
+{
+       void *p = io_u_verify_off(hdr, vc);
+
+       dprint(FD_VERIFY, "%s verify io_u %p, len %u\n", name, vc->io_u, hdr->len);
+
+       fio_sha3_update(sha3_ctx, p, hdr->len - hdr_size(vc->td, hdr));
+       fio_sha3_final(sha3_ctx);
+
+       if (!memcmp(sha, sha3_ctx->sha, sha_size))
+               return 0;
+
+       vc->name = name;
+       vc->good_crc = sha;
+       vc->bad_crc = sha3_ctx->sha;
+       vc->crc_len = sha_size;
+       log_verify_failure(hdr, vc);
+       return EILSEQ;
+}
+
+static int verify_io_u_sha3_224(struct verify_header *hdr, struct vcont *vc)
+{
+       struct vhdr_sha3_224 *vh = hdr_priv(hdr);
+       uint8_t sha[SHA3_224_DIGEST_SIZE];
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = sha,
+       };
+
+       fio_sha3_224_init(&sha3_ctx);
+
+       return verify_io_u_sha3(hdr, vc, &sha3_ctx, vh->sha,
+                               SHA3_224_DIGEST_SIZE, "sha3-224");
+}
+
+static int verify_io_u_sha3_256(struct verify_header *hdr, struct vcont *vc)
+{
+       struct vhdr_sha3_256 *vh = hdr_priv(hdr);
+       uint8_t sha[SHA3_256_DIGEST_SIZE];
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = sha,
+       };
+
+       fio_sha3_256_init(&sha3_ctx);
+
+       return verify_io_u_sha3(hdr, vc, &sha3_ctx, vh->sha,
+                               SHA3_256_DIGEST_SIZE, "sha3-256");
+}
+
+static int verify_io_u_sha3_384(struct verify_header *hdr, struct vcont *vc)
+{
+       struct vhdr_sha3_384 *vh = hdr_priv(hdr);
+       uint8_t sha[SHA3_384_DIGEST_SIZE];
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = sha,
+       };
+
+       fio_sha3_384_init(&sha3_ctx);
+
+       return verify_io_u_sha3(hdr, vc, &sha3_ctx, vh->sha,
+                               SHA3_384_DIGEST_SIZE, "sha3-384");
+}
+
+static int verify_io_u_sha3_512(struct verify_header *hdr, struct vcont *vc)
+{
+       struct vhdr_sha3_512 *vh = hdr_priv(hdr);
+       uint8_t sha[SHA3_512_DIGEST_SIZE];
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = sha,
+       };
+
+       fio_sha3_512_init(&sha3_ctx);
+
+       return verify_io_u_sha3(hdr, vc, &sha3_ctx, vh->sha,
+                               SHA3_512_DIGEST_SIZE, "sha3-512");
+}
+
 static int verify_io_u_sha512(struct verify_header *hdr, struct vcont *vc)
 {
        void *p = io_u_verify_off(hdr, vc);
@@ -651,48 +741,73 @@ int verify_io_u_async(struct thread_data *td, struct io_u **io_u_ptr)
 
        if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
                td->cur_depth--;
-               io_u_clear(io_u, IO_U_F_IN_CUR_DEPTH);
+               io_u_clear(td, io_u, IO_U_F_IN_CUR_DEPTH);
        }
        flist_add_tail(&io_u->verify_list, &td->verify_list);
        *io_u_ptr = NULL;
-       pthread_mutex_unlock(&td->io_u_lock);
 
        pthread_cond_signal(&td->verify_cond);
+       pthread_mutex_unlock(&td->io_u_lock);
        return 0;
 }
 
+/*
+ * Thanks Rusty, for spending the time so I don't have to.
+ *
+ * http://rusty.ozlabs.org/?p=560
+ */
+static int mem_is_zero(const void *data, size_t length)
+{
+       const unsigned char *p = data;
+       size_t len;
+
+       /* Check first 16 bytes manually */
+       for (len = 0; len < 16; len++) {
+               if (!length)
+                       return 1;
+               if (*p)
+                       return 0;
+               p++;
+               length--;
+       }
+
+       /* Now we know that's zero, memcmp with self. */
+       return memcmp(data, p, length) == 0;
+}
+
+static int mem_is_zero_slow(const void *data, size_t length, size_t *offset)
+{
+       const unsigned char *p = data;
+
+       *offset = 0;
+       while (length) {
+               if (*p)
+                       break;
+               (*offset)++;
+               length--;
+               p++;
+       }
+
+       return !length;
+}
+
 static int verify_trimmed_io_u(struct thread_data *td, struct io_u *io_u)
 {
-       static char zero_buf[1024];
-       unsigned int this_len, len;
-       int ret = 0;
-       void *p;
+       size_t offset;
 
        if (!td->o.trim_zero)
                return 0;
 
-       len = io_u->buflen;
-       p = io_u->buf;
-       do {
-               this_len = sizeof(zero_buf);
-               if (this_len > len)
-                       this_len = len;
-               if (memcmp(p, zero_buf, this_len)) {
-                       ret = EILSEQ;
-                       break;
-               }
-               len -= this_len;
-               p += this_len;
-       } while (len);
-
-       if (!ret)
+       if (mem_is_zero(io_u->buf, io_u->buflen))
                return 0;
 
-       log_err("trim: verify failed at file %s offset %llu, length %lu"
+       mem_is_zero_slow(io_u->buf, io_u->buflen, &offset);
+
+       log_err("trim: verify failed at file %s offset %llu, length %llu"
                ", block offset %lu\n",
                        io_u->file->file_name, io_u->offset, io_u->buflen,
-                       (unsigned long) (p - io_u->buf));
-       return ret;
+                       (unsigned long) offset);
+       return EILSEQ;
 }
 
 static int verify_header(struct io_u *io_u, struct thread_data *td,
@@ -733,7 +848,7 @@ static int verify_header(struct io_u *io_u, struct thread_data *td,
         * state of numberio, that would have been written to each block
         * in a previous run of fio, has been reached.
         */
-       if ((td_write(td) || td_rw(td)) && (td_min_bs(td) == td_max_bs(td)) &&
+       if (td_write(td) && (td_min_bs(td) == td_max_bs(td)) &&
            !td->o.time_based)
                if (!td->o.verify_only || td->o.loops == 0)
                        if (hdr->numberio != io_u->numberio) {
@@ -752,9 +867,16 @@ static int verify_header(struct io_u *io_u, struct thread_data *td,
        return 0;
 
 err:
-       log_err(" at file %s offset %llu, length %u\n",
+       log_err(" at file %s offset %llu, length %u"
+               " (requested block: offset=%llu, length=%llu)\n",
                io_u->file->file_name,
-               io_u->offset + hdr_num * hdr_len, hdr_len);
+               io_u->offset + hdr_num * hdr_len, hdr_len,
+               io_u->offset, io_u->buflen);
+
+       if (td->o.verify_dump)
+               dump_buf(p, hdr_len, io_u->offset + hdr_num * hdr_len,
+                               "hdr_fail", io_u->file);
+
        return EILSEQ;
 }
 
@@ -772,7 +894,7 @@ int verify_io_u(struct thread_data *td, struct io_u **io_u_ptr)
         * If the IO engine is faking IO (like null), then just pretend
         * we verified everything.
         */
-       if (td->io_ops->flags & FIO_FAKEIO)
+       if (td_ioengine_flagged(td, FIO_FAKEIO))
                return 0;
 
        if (io_u->flags & IO_U_F_TRIMMED) {
@@ -801,10 +923,9 @@ int verify_io_u(struct thread_data *td, struct io_u **io_u_ptr)
                hdr = p;
 
                /*
-                * Make rand_seed check pass when have verifysort or
-                * verify_backlog.
+                * Make rand_seed check pass when have verify_backlog.
                 */
-               if (td->o.verifysort || (td->flags & TD_F_VER_BACKLOG))
+               if (!td_rw(td) || (td->flags & TD_F_VER_BACKLOG))
                        io_u->rand_seed = hdr->rand_seed;
 
                if (td->o.verify != VERIFY_PATTERN_NO_HDR) {
@@ -850,6 +971,18 @@ int verify_io_u(struct thread_data *td, struct io_u **io_u_ptr)
                case VERIFY_SHA512:
                        ret = verify_io_u_sha512(hdr, &vc);
                        break;
+               case VERIFY_SHA3_224:
+                       ret = verify_io_u_sha3_224(hdr, &vc);
+                       break;
+               case VERIFY_SHA3_256:
+                       ret = verify_io_u_sha3_256(hdr, &vc);
+                       break;
+               case VERIFY_SHA3_384:
+                       ret = verify_io_u_sha3_384(hdr, &vc);
+                       break;
+               case VERIFY_SHA3_512:
+                       ret = verify_io_u_sha3_512(hdr, &vc);
+                       break;
                case VERIFY_XXHASH:
                        ret = verify_io_u_xxhash(hdr, &vc);
                        break;
@@ -887,6 +1020,56 @@ static void fill_xxhash(struct verify_header *hdr, void *p, unsigned int len)
        vh->hash = XXH32_digest(state);
 }
 
+static void fill_sha3(struct fio_sha3_ctx *sha3_ctx, void *p, unsigned int len)
+{
+       fio_sha3_update(sha3_ctx, p, len);
+       fio_sha3_final(sha3_ctx);
+}
+
+static void fill_sha3_224(struct verify_header *hdr, void *p, unsigned int len)
+{
+       struct vhdr_sha3_224 *vh = hdr_priv(hdr);
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = vh->sha,
+       };
+
+       fio_sha3_224_init(&sha3_ctx);
+       fill_sha3(&sha3_ctx, p, len);
+}
+
+static void fill_sha3_256(struct verify_header *hdr, void *p, unsigned int len)
+{
+       struct vhdr_sha3_256 *vh = hdr_priv(hdr);
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = vh->sha,
+       };
+
+       fio_sha3_256_init(&sha3_ctx);
+       fill_sha3(&sha3_ctx, p, len);
+}
+
+static void fill_sha3_384(struct verify_header *hdr, void *p, unsigned int len)
+{
+       struct vhdr_sha3_384 *vh = hdr_priv(hdr);
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = vh->sha,
+       };
+
+       fio_sha3_384_init(&sha3_ctx);
+       fill_sha3(&sha3_ctx, p, len);
+}
+
+static void fill_sha3_512(struct verify_header *hdr, void *p, unsigned int len)
+{
+       struct vhdr_sha3_512 *vh = hdr_priv(hdr);
+       struct fio_sha3_ctx sha3_ctx = {
+               .sha = vh->sha,
+       };
+
+       fio_sha3_512_init(&sha3_ctx);
+       fill_sha3(&sha3_ctx, p, len);
+}
+
 static void fill_sha512(struct verify_header *hdr, void *p, unsigned int len)
 {
        struct vhdr_sha512 *vh = hdr_priv(hdr);
@@ -981,7 +1164,7 @@ static void __fill_hdr(struct thread_data *td, struct io_u *io_u,
        hdr->rand_seed = rand_seed;
        hdr->offset = io_u->offset + header_num * td->o.verify_interval;
        hdr->time_sec = io_u->start_time.tv_sec;
-       hdr->time_usec = io_u->start_time.tv_usec;
+       hdr->time_nsec = io_u->start_time.tv_nsec;
        hdr->thread = td->thread_number;
        hdr->numberio = io_u->numberio;
        hdr->crc32 = fio_crc32c(p, offsetof(struct verify_header, crc32));
@@ -992,7 +1175,6 @@ static void fill_hdr(struct thread_data *td, struct io_u *io_u,
                     struct verify_header *hdr, unsigned int header_num,
                     unsigned int header_len, uint64_t rand_seed)
 {
-
        if (td->o.verify != VERIFY_PATTERN_NO_HDR)
                __fill_hdr(td, io_u, hdr, header_num, header_len, rand_seed);
 }
@@ -1002,9 +1184,10 @@ static void populate_hdr(struct thread_data *td, struct io_u *io_u,
                         unsigned int header_len)
 {
        unsigned int data_len;
-       void *data, *p;
+       void *data;
+       char *p;
 
-       p = (void *) hdr;
+       p = (char *) hdr;
 
        fill_hdr(td, io_u, hdr, header_num, header_len, io_u->rand_seed);
 
@@ -1053,6 +1236,26 @@ static void populate_hdr(struct thread_data *td, struct io_u *io_u,
                                                io_u, hdr->len);
                fill_sha512(hdr, data, data_len);
                break;
+       case VERIFY_SHA3_224:
+               dprint(FD_VERIFY, "fill sha3-224 io_u %p, len %u\n",
+                                               io_u, hdr->len);
+               fill_sha3_224(hdr, data, data_len);
+               break;
+       case VERIFY_SHA3_256:
+               dprint(FD_VERIFY, "fill sha3-256 io_u %p, len %u\n",
+                                               io_u, hdr->len);
+               fill_sha3_256(hdr, data, data_len);
+               break;
+       case VERIFY_SHA3_384:
+               dprint(FD_VERIFY, "fill sha3-384 io_u %p, len %u\n",
+                                               io_u, hdr->len);
+               fill_sha3_384(hdr, data, data_len);
+               break;
+       case VERIFY_SHA3_512:
+               dprint(FD_VERIFY, "fill sha3-512 io_u %p, len %u\n",
+                                               io_u, hdr->len);
+               fill_sha3_512(hdr, data, data_len);
+               break;
        case VERIFY_XXHASH:
                dprint(FD_VERIFY, "fill xxhash io_u %p, len %u\n",
                                                io_u, hdr->len);
@@ -1102,7 +1305,7 @@ int get_next_verify(struct thread_data *td, struct io_u *io_u)
                return 0;
 
        if (!RB_EMPTY_ROOT(&td->io_hist_tree)) {
-               struct rb_node *n = rb_first(&td->io_hist_tree);
+               struct fio_rb_node *n = rb_first(&td->io_hist_tree);
 
                ipo = rb_entry(n, struct io_piece, rb_node);
 
@@ -1138,10 +1341,10 @@ int get_next_verify(struct thread_data *td, struct io_u *io_u)
                io_u->buflen = ipo->len;
                io_u->numberio = ipo->numberio;
                io_u->file = ipo->file;
-               io_u_set(io_u, IO_U_F_VER_LIST);
+               io_u_set(td, io_u, IO_U_F_VER_LIST);
 
                if (ipo->flags & IP_F_TRIMMED)
-                       io_u_set(io_u, IO_U_F_TRIMMED);
+                       io_u_set(td, io_u, IO_U_F_TRIMMED);
 
                if (!fio_file_open(io_u->file)) {
                        int r = td_io_open_file(td, io_u->file);
@@ -1180,6 +1383,7 @@ void fio_verify_init(struct thread_data *td)
 {
        if (td->o.verify == VERIFY_CRC32C_INTEL ||
            td->o.verify == VERIFY_CRC32C) {
+               crc32c_arm64_probe();
                crc32c_intel_probe();
        }
 }
@@ -1225,7 +1429,7 @@ static void *verify_async_thread(void *data)
                        io_u = flist_first_entry(&list, struct io_u, verify_list);
                        flist_del_init(&io_u->verify_list);
 
-                       io_u_set(io_u, IO_U_F_NO_FILE_PUT);
+                       io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
                        ret = verify_io_u(td, &io_u);
 
                        put_io_u(td, io_u);
@@ -1248,9 +1452,9 @@ static void *verify_async_thread(void *data)
 done:
        pthread_mutex_lock(&td->io_u_lock);
        td->nr_verify_threads--;
+       pthread_cond_signal(&td->free_cond);
        pthread_mutex_unlock(&td->io_u_lock);
 
-       pthread_cond_signal(&td->free_cond);
        return NULL;
 }
 
@@ -1260,7 +1464,7 @@ int verify_async_init(struct thread_data *td)
        pthread_attr_t attr;
 
        pthread_attr_init(&attr);
-       pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
+       pthread_attr_setstacksize(&attr, 2 * PTHREAD_STACK_MIN);
 
        td->verify_thread_exit = 0;
 
@@ -1286,9 +1490,12 @@ int verify_async_init(struct thread_data *td)
 
        if (i != td->o.verify_async) {
                log_err("fio: only %d verify threads started, exiting\n", i);
+
+               pthread_mutex_lock(&td->io_u_lock);
                td->verify_thread_exit = 1;
-               write_barrier();
                pthread_cond_broadcast(&td->verify_cond);
+               pthread_mutex_unlock(&td->io_u_lock);
+
                return 1;
        }
 
@@ -1297,12 +1504,10 @@ int verify_async_init(struct thread_data *td)
 
 void verify_async_exit(struct thread_data *td)
 {
+       pthread_mutex_lock(&td->io_u_lock);
        td->verify_thread_exit = 1;
-       write_barrier();
        pthread_cond_broadcast(&td->verify_cond);
 
-       pthread_mutex_lock(&td->io_u_lock);
-
        while (td->nr_verify_threads)
                pthread_cond_wait(&td->free_cond, &td->io_u_lock);
 
@@ -1316,13 +1521,54 @@ int paste_blockoff(char *buf, unsigned int len, void *priv)
        struct io_u *io = priv;
        unsigned long long off;
 
-       typecheck(typeof(off), io->offset);
+       typecheck(__typeof__(off), io->offset);
        off = cpu_to_le64((uint64_t)io->offset);
        len = min(len, (unsigned int)sizeof(off));
        memcpy(buf, &off, len);
        return 0;
 }
 
+static int __fill_file_completions(struct thread_data *td,
+                                  struct thread_io_list *s,
+                                  struct fio_file *f, unsigned int *index)
+{
+       unsigned int comps;
+       int i, j;
+
+       if (!f->last_write_comp)
+               return 0;
+
+       if (td->io_blocks[DDIR_WRITE] < td->o.iodepth)
+               comps = td->io_blocks[DDIR_WRITE];
+       else
+               comps = td->o.iodepth;
+
+       j = f->last_write_idx - 1;
+       for (i = 0; i < comps; i++) {
+               if (j == -1)
+                       j = td->o.iodepth - 1;
+               s->comps[*index].fileno = __cpu_to_le64(f->fileno);
+               s->comps[*index].offset = cpu_to_le64(f->last_write_comp[j]);
+               (*index)++;
+               j--;
+       }
+
+       return comps;
+}
+
+static int fill_file_completions(struct thread_data *td,
+                                struct thread_io_list *s, unsigned int *index)
+{
+       struct fio_file *f;
+       unsigned int i;
+       int comps = 0;
+
+       for_each_file(td, f, i)
+               comps += __fill_file_completions(td, s, f, index);
+
+       return comps;
+}
+
 struct all_io_list *get_all_io_list(int save_mask, size_t *sz)
 {
        struct all_io_list *rep;
@@ -1344,7 +1590,7 @@ struct all_io_list *get_all_io_list(int save_mask, size_t *sz)
                        continue;
                td->stop_io = 1;
                td->flags |= TD_F_VSTATE_SAVED;
-               depth += td->o.iodepth;
+               depth += (td->o.iodepth * td->o.nr_files);
                nr++;
        }
 
@@ -1353,39 +1599,25 @@ struct all_io_list *get_all_io_list(int save_mask, size_t *sz)
 
        *sz = sizeof(*rep);
        *sz += nr * sizeof(struct thread_io_list);
-       *sz += depth * sizeof(uint64_t);
+       *sz += depth * sizeof(struct file_comp);
        rep = malloc(*sz);
+       memset(rep, 0, *sz);
 
        rep->threads = cpu_to_le64((uint64_t) nr);
 
        next = &rep->state[0];
        for_each_td(td, i) {
                struct thread_io_list *s = next;
-               unsigned int comps;
+               unsigned int comps, index = 0;
 
                if (save_mask != IO_LIST_ALL && (i + 1) != save_mask)
                        continue;
 
-               if (td->last_write_comp) {
-                       int j, k;
-
-                       if (td->io_blocks[DDIR_WRITE] < td->o.iodepth)
-                               comps = td->io_blocks[DDIR_WRITE];
-                       else
-                               comps = td->o.iodepth;
-
-                       k = td->last_write_idx - 1;
-                       for (j = 0; j < comps; j++) {
-                               if (k == -1)
-                                       k = td->o.iodepth - 1;
-                               s->offsets[j] = cpu_to_le64(td->last_write_comp[k]);
-                               k--;
-                       }
-               } else
-                       comps = 0;
+               comps = fill_file_completions(td, s, &index);
 
                s->no_comps = cpu_to_le64((uint64_t) comps);
                s->depth = cpu_to_le64((uint64_t) td->o.iodepth);
+               s->nofiles = cpu_to_le64((uint64_t) td->o.nr_files);
                s->numberio = cpu_to_le64((uint64_t) td->io_issues[DDIR_WRITE]);
                s->index = cpu_to_le64((uint64_t) i);
                if (td->random_state.use64) {
@@ -1414,7 +1646,7 @@ struct all_io_list *get_all_io_list(int save_mask, size_t *sz)
 static int open_state_file(const char *name, const char *prefix, int num,
                           int for_write)
 {
-       char out[64];
+       char out[PATH_MAX];
        int flags;
        int fd;
 
@@ -1428,6 +1660,7 @@ static int open_state_file(const char *name, const char *prefix, int num,
        fd = open(out, flags, 0644);
        if (fd == -1) {
                perror("fio: open state file");
+               log_err("fio: state file: %s (for_write=%d)\n", out, for_write);
                return -1;
        }
 
@@ -1490,7 +1723,7 @@ void verify_save_state(int mask)
                char prefix[PATH_MAX];
 
                if (aux_path)
-                       sprintf(prefix, "%s%slocal", aux_path, FIO_OS_PATH_SEPARATOR);
+                       sprintf(prefix, "%s%clocal", aux_path, FIO_OS_PATH_SEPARATOR);
                else
                        strcpy(prefix, "local");
 
@@ -1505,72 +1738,34 @@ void verify_free_state(struct thread_data *td)
                free(td->vstate);
 }
 
-static struct thread_io_list *convert_v1_list(struct thread_io_list_v1 *s)
-{
-       struct thread_io_list *til;
-       int i;
-
-       til = malloc(__thread_io_list_sz(s->no_comps));
-       til->no_comps = s->no_comps;
-       til->depth = s->depth;
-       til->numberio = s->numberio;
-       til->index = s->index;
-       memcpy(til->name, s->name, sizeof(til->name));
-
-       til->rand.use64 = 0;
-       for (i = 0; i < 4; i++)
-               til->rand.state32.s[i] = s->rand.s[i];
-
-       for (i = 0; i < s->no_comps; i++)
-               til->offsets[i] = s->offsets[i];
-
-       return til;
-}
-
-void verify_convert_assign_state(struct thread_data *td, void *p, int version)
+void verify_assign_state(struct thread_data *td, void *p)
 {
-       struct thread_io_list *til;
+       struct thread_io_list *s = p;
        int i;
 
-       if (version == 1) {
-               struct thread_io_list_v1 *s = p;
-
-               s->no_comps = le64_to_cpu(s->no_comps);
-               s->depth = le64_to_cpu(s->depth);
-               s->numberio = le64_to_cpu(s->numberio);
-               for (i = 0; i < 4; i++)
-                       s->rand.s[i] = le32_to_cpu(s->rand.s[i]);
-               for (i = 0; i < s->no_comps; i++)
-                       s->offsets[i] = le64_to_cpu(s->offsets[i]);
+       s->no_comps = le64_to_cpu(s->no_comps);
+       s->depth = le32_to_cpu(s->depth);
+       s->nofiles = le32_to_cpu(s->nofiles);
+       s->numberio = le64_to_cpu(s->numberio);
+       s->rand.use64 = le64_to_cpu(s->rand.use64);
 
-               til = convert_v1_list(s);
-               free(s);
+       if (s->rand.use64) {
+               for (i = 0; i < 6; i++)
+                       s->rand.state64.s[i] = le64_to_cpu(s->rand.state64.s[i]);
        } else {
-               struct thread_io_list *s = p;
-
-               s->no_comps = le64_to_cpu(s->no_comps);
-               s->depth = le64_to_cpu(s->depth);
-               s->numberio = le64_to_cpu(s->numberio);
-               s->rand.use64 = le64_to_cpu(s->rand.use64);
-
-               if (s->rand.use64) {
-                       for (i = 0; i < 6; i++)
-                               s->rand.state64.s[i] = le64_to_cpu(s->rand.state64.s[i]);
-               } else {
-                       for (i = 0; i < 4; i++)
-                               s->rand.state32.s[i] = le32_to_cpu(s->rand.state32.s[i]);
-               }
-               for (i = 0; i < s->no_comps; i++)
-                       s->offsets[i] = le64_to_cpu(s->offsets[i]);
+               for (i = 0; i < 4; i++)
+                       s->rand.state32.s[i] = le32_to_cpu(s->rand.state32.s[i]);
+       }
 
-               til = p;
+       for (i = 0; i < s->no_comps; i++) {
+               s->comps[i].fileno = le64_to_cpu(s->comps[i].fileno);
+               s->comps[i].offset = le64_to_cpu(s->comps[i].offset);
        }
 
-       td->vstate = til;
+       td->vstate = p;
 }
 
-int verify_state_hdr(struct verify_state_hdr *hdr, struct thread_io_list *s,
-                    int *version)
+int verify_state_hdr(struct verify_state_hdr *hdr, struct thread_io_list *s)
 {
        uint64_t crc;
 
@@ -1578,15 +1773,13 @@ int verify_state_hdr(struct verify_state_hdr *hdr, struct thread_io_list *s,
        hdr->size = le64_to_cpu(hdr->size);
        hdr->crc = le64_to_cpu(hdr->crc);
 
-       if (hdr->version != VSTATE_HDR_VERSION ||
-           hdr->version != VSTATE_HDR_VERSION_V1)
+       if (hdr->version != VSTATE_HDR_VERSION)
                return 1;
 
        crc = fio_crc32c((void *)s, hdr->size);
        if (crc != hdr->crc)
                return 1;
 
-       *version = hdr->version;
        return 0;
 }
 
@@ -1617,9 +1810,9 @@ int verify_load_state(struct thread_data *td, const char *prefix)
        hdr.size = le64_to_cpu(hdr.size);
        hdr.crc = le64_to_cpu(hdr.crc);
 
-       if (hdr.version != VSTATE_HDR_VERSION &&
-           hdr.version != VSTATE_HDR_VERSION_V1) {
-               log_err("fio: bad version in verify state header\n");
+       if (hdr.version != VSTATE_HDR_VERSION) {
+               log_err("fio: unsupported (%d) version in verify state header\n",
+                               (unsigned int) hdr.version);
                goto err;
        }
 
@@ -1640,7 +1833,7 @@ int verify_load_state(struct thread_data *td, const char *prefix)
 
        close(fd);
 
-       verify_convert_assign_state(td, s, hdr.version);
+       verify_assign_state(td, s);
        return 0;
 err:
        if (s)
@@ -1655,9 +1848,10 @@ err:
 int verify_state_should_stop(struct thread_data *td, struct io_u *io_u)
 {
        struct thread_io_list *s = td->vstate;
+       struct fio_file *f = io_u->file;
        int i;
 
-       if (!s)
+       if (!s || !f)
                return 0;
 
        /*
@@ -1674,9 +1868,12 @@ int verify_state_should_stop(struct thread_data *td, struct io_u *io_u)
         * completed or not. If the IO was seen as completed, then
         * lets verify it.
         */
-       for (i = 0; i < s->no_comps; i++)
-               if (io_u->offset == s->offsets[i])
+       for (i = 0; i < s->no_comps; i++) {
+               if (s->comps[i].fileno != f->fileno)
+                       continue;
+               if (io_u->offset == s->comps[i].offset)
                        return 0;
+       }
 
        /*
         * Not found, we have to stop