X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=verify.c;h=f79ab43aab6b97b0721f52485fad31053e8d6ca9;hp=90cd093add1fbbb13ba2f9d713ce60dc9e4f49e9;hb=fc220349e45144360917db48010b503a9874930d;hpb=c4b6117b236370b174f3e3d2e299bf8b3733be97

diff --git a/verify.c b/verify.c
index 90cd093a..f79ab43a 100644
--- a/verify.c
+++ b/verify.c
@@ -13,6 +13,7 @@
 #include "trim.h"
 #include "lib/rand.h"
 #include "lib/hweight.h"
+#include "lib/pattern.h"
 
 #include "crc/md5.h"
 #include "crc/crc64.h"
@@ -23,67 +24,62 @@
 #include "crc/sha256.h"
 #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_pattern(struct thread_data *td, void *p, unsigned int len,
-			 char *pattern, unsigned int pattern_bytes)
+void fill_buffer_pattern(struct thread_data *td, void *p, unsigned int len)
 {
-	switch (pattern_bytes) {
-	case 0:
-		assert(0);
-		break;
-	case 1:
-		dprint(FD_VERIFY, "fill verify pattern b=0 len=%u\n", len);
-		memset(p, pattern[0], len);
-		break;
-	default: {
-		unsigned int i = 0, size = 0;
-		unsigned char *b = p;
-
-		dprint(FD_VERIFY, "fill verify pattern b=%d len=%u\n",
-					pattern_bytes, len);
-
-		while (i < len) {
-			size = pattern_bytes;
-			if (size > (len - i))
-				size = len - i;
-			memcpy(b+i, pattern, size);
-			i += size;
-		}
-		break;
-		}
-	}
+	(void)cpy_pattern(td->o.buffer_pattern, td->o.buffer_pattern_bytes, p, len);
 }
 
-void fill_buffer_pattern(struct thread_data *td, void *p, unsigned int len)
+static void __fill_buffer(struct thread_options *o, uint64_t seed, void *p,
+			  unsigned int len)
 {
-	fill_pattern(td, p, len, td->o.buffer_pattern, td->o.buffer_pattern_bytes);
+	__fill_random_buf_percentage(seed, p, o->compress_percentage, len, len, o->buffer_pattern, o->buffer_pattern_bytes);
+}
+
+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;
+
+	return fill_random_buf_percentage(fs, p, o->compress_percentage, len, len, o->buffer_pattern, o->buffer_pattern_bytes);
 }
 
 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)
 {
-	if (!td->o.verify_pattern_bytes) {
+	struct thread_options *o = &td->o;
+
+	if (!o->verify_pattern_bytes) {
 		dprint(FD_VERIFY, "fill random bytes len=%u\n", len);
 
 		if (use_seed)
-			__fill_random_buf(p, len, seed);
+			__fill_buffer(o, seed, p, len);
 		else
-			io_u->rand_seed = fill_random_buf(&td->__verify_state, p, len);
+			io_u->rand_seed = fill_buffer(td, p, len);
 		return;
 	}
 
-	if (io_u->buf_filled_len >= len) {
+	/* Skip if we were here and we do not need to patch pattern
+	 * with format */
+	if (!td->o.verify_fmt_sz && io_u->buf_filled_len >= len) {
 		dprint(FD_VERIFY, "using already filled verify pattern b=%d len=%u\n",
-			td->o.verify_pattern_bytes, len);
+			o->verify_pattern_bytes, len);
 		return;
 	}
 
-	fill_pattern(td, p, len, td->o.verify_pattern, td->o.verify_pattern_bytes);
-
+	(void)paste_format(td->o.verify_pattern, td->o.verify_pattern_bytes,
+			   td->o.verify_fmt, td->o.verify_fmt_sz,
+			   p, len, io_u);
 	io_u->buf_filled_len = len;
 }
 
@@ -91,15 +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;
@@ -146,7 +147,9 @@ static inline unsigned int __hdr_size(int verify_type)
 
 	switch (verify_type) {
 	case VERIFY_NONE:
+	case VERIFY_HDR_ONLY:
 	case VERIFY_NULL:
+	case VERIFY_PATTERN:
 		len = 0;
 		break;
 	case VERIFY_MD5:
@@ -172,15 +175,26 @@ static inline unsigned int __hdr_size(int verify_type)
 	case VERIFY_SHA512:
 		len = sizeof(struct vhdr_sha512);
 		break;
-	case VERIFY_META:
-		len = sizeof(struct vhdr_meta);
+	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;
 	case VERIFY_SHA1:
 		len = sizeof(struct vhdr_sha1);
 		break;
-	case VERIFY_PATTERN:
-		len = 0;
-		break;
+	case VERIFY_PATTERN_NO_HDR:
+		return 0;
 	default:
 		log_err("fio: unknown verify header!\n");
 		assert(0);
@@ -189,8 +203,12 @@ static inline unsigned int __hdr_size(int verify_type)
 	return len + sizeof(struct verify_header);
 }
 
-static inline unsigned int hdr_size(struct verify_header *hdr)
+static inline unsigned int hdr_size(struct thread_data *td,
+				    struct verify_header *hdr)
 {
+	if (td->o.verify == VERIFY_PATTERN_NO_HDR)
+		return 0;
+
 	return __hdr_size(hdr->verify_type);
 }
 
@@ -222,21 +240,28 @@ struct vcont {
 	unsigned int crc_len;
 };
 
+#define DUMP_BUF_SZ	255
+
 static void dump_buf(char *buf, unsigned int len, unsigned long long offset,
 		     const char *type, struct fio_file *f)
 {
-	char *ptr, fname[256];
+	char *ptr, *fname;
+	char sep[2] = { FIO_OS_PATH_SEPARATOR, 0 };
 	int ret, fd;
 
 	ptr = strdup(f->file_name);
-	strcpy(fname, basename(ptr));
 
-	sprintf(fname + strlen(fname), ".%llu.%s", offset, type);
+	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;
+	}
 
 	fd = open(fname, O_CREAT | O_TRUNC | O_WRONLY, 0644);
 	if (fd < 0) {
 		perror("open verify buf file");
-		return;
+		goto free_fname;
 	}
 
 	while (len) {
@@ -253,6 +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);
 }
 
@@ -260,7 +290,7 @@ static void dump_buf(char *buf, unsigned int len, unsigned long long offset,
  * Dump the contents of the read block and re-generate the correct data
  * and dump that too.
  */
-static void dump_verify_buffers(struct verify_header *hdr, struct vcont *vc)
+static void __dump_verify_buffers(struct verify_header *hdr, struct vcont *vc)
 {
 	struct thread_data *td = vc->td;
 	struct io_u *io_u = vc->io_u;
@@ -296,14 +326,29 @@ static void dump_verify_buffers(struct verify_header *hdr, struct vcont *vc)
 	free(buf);
 }
 
+static void dump_verify_buffers(struct verify_header *hdr, struct vcont *vc)
+{
+	struct thread_data *td = vc->td;
+	struct verify_header shdr;
+
+	if (td->o.verify == VERIFY_PATTERN_NO_HDR) {
+		__fill_hdr(td, vc->io_u, &shdr, 0, vc->io_u->buflen, 0);
+		hdr = &shdr;
+	}
+
+	__dump_verify_buffers(hdr, vc);
+}
+
 static void log_verify_failure(struct verify_header *hdr, struct vcont *vc)
 {
 	unsigned long long offset;
 
 	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: ");
@@ -320,7 +365,7 @@ static void log_verify_failure(struct verify_header *hdr, struct vcont *vc)
  */
 static inline void *io_u_verify_off(struct verify_header *hdr, struct vcont *vc)
 {
-	return vc->io_u->buf + vc->hdr_num * hdr->len + hdr_size(hdr);
+	return vc->io_u->buf + vc->hdr_num * hdr->len + hdr_size(vc->td, hdr);
 }
 
 static int verify_io_u_pattern(struct verify_header *hdr, struct vcont *vc)
@@ -329,35 +374,37 @@ static int verify_io_u_pattern(struct verify_header *hdr, struct vcont *vc)
 	struct io_u *io_u = vc->io_u;
 	char *buf, *pattern;
 	unsigned int header_size = __hdr_size(td->o.verify);
-	unsigned int len, mod, i, size, pattern_size;
+	unsigned int len, mod, i, pattern_size;
+	int rc;
 
 	pattern = td->o.verify_pattern;
 	pattern_size = td->o.verify_pattern_bytes;
-	if (pattern_size <= 1)
-		pattern_size = MAX_PATTERN_SIZE;
-	buf = (void *) hdr + header_size;
+	assert(pattern_size != 0);
+
+	(void)paste_format_inplace(pattern, pattern_size,
+				   td->o.verify_fmt, td->o.verify_fmt_sz, io_u);
+
+	buf = (char *) hdr + header_size;
 	len = get_hdr_inc(td, io_u) - header_size;
-	mod = header_size % pattern_size;
-
-	for (i = 0; i < len; i += size) {
-		size = pattern_size - mod;
-		if (size > (len - i))
-			size = len - i;
-		if (memcmp(buf + i, pattern + mod, size))
-			/* Let the slow compare find the first mismatch byte. */
-			break;
-		mod = 0;
-	}
+	mod = (get_hdr_inc(td, io_u) * vc->hdr_num + header_size) % pattern_size;
+
+	rc = cmp_pattern(pattern, pattern_size, mod, buf, len);
+	if (!rc)
+		return 0;
 
-	for (; i < len; i++) {
+	/* Slow path, compare each byte */
+	for (i = 0; i < len; i++) {
 		if (buf[i] != pattern[mod]) {
 			unsigned int bits;
 
 			bits = hweight8(buf[i] ^ pattern[mod]);
-			log_err("fio: got pattern %x, wanted %x. Bad bits %d\n",
-				buf[i], pattern[mod], bits);
+			log_err("fio: got pattern '%02x', wanted '%02x'. Bad bits %d\n",
+				(unsigned char)buf[i],
+				(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++;
@@ -365,43 +412,111 @@ static int verify_io_u_pattern(struct verify_header *hdr, struct vcont *vc)
 			mod = 0;
 	}
 
-	return 0;
+	/* Unreachable line */
+	assert(0);
+	return EILSEQ;
 }
 
-static int verify_io_u_meta(struct verify_header *hdr, struct vcont *vc)
+static int verify_io_u_xxhash(struct verify_header *hdr, struct vcont *vc)
 {
-	struct thread_data *td = vc->td;
-	struct vhdr_meta *vh = hdr_priv(hdr);
-	struct io_u *io_u = vc->io_u;
-	int ret = EILSEQ;
+	void *p = io_u_verify_off(hdr, vc);
+	struct vhdr_xxhash *vh = hdr_priv(hdr);
+	uint32_t hash;
+	void *state;
 
-	dprint(FD_VERIFY, "meta verify io_u %p, len %u\n", io_u, hdr->len);
+	dprint(FD_VERIFY, "xxhash verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
-	if (vh->offset == io_u->offset + vc->hdr_num * td->o.verify_interval)
-		ret = 0;
+	state = XXH32_init(1);
+	XXH32_update(state, p, hdr->len - hdr_size(vc->td, hdr));
+	hash = XXH32_digest(state);
 
-	if (td->o.verify_pattern_bytes)
-		ret |= verify_io_u_pattern(hdr, vc);
+	if (vh->hash == hash)
+		return 0;
 
-	/*
-	 * For read-only workloads, the program cannot be certain of the
-	 * last numberio written to a block. Checking of numberio will be done
-	 * only for workloads that write data.
-	 * For verify_only, numberio will be checked in the last iteration when
-	 * the correct 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))
-		if (!td->o.verify_only || td->o.loops == 0)
-			if (vh->numberio != io_u->numberio)
-				ret = EILSEQ;
+	vc->name = "xxhash";
+	vc->good_crc = &vh->hash;
+	vc->bad_crc = &hash;
+	vc->crc_len = sizeof(hash);
+	log_verify_failure(hdr, 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 (!ret)
+	if (!memcmp(sha, sha3_ctx->sha, sha_size))
 		return 0;
 
-	vc->name = "meta";
+	vc->name = name;
+	vc->good_crc = sha;
+	vc->bad_crc = sha3_ctx->sha;
+	vc->crc_len = sha_size;
 	log_verify_failure(hdr, vc);
-	return ret;
+	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)
@@ -416,7 +531,7 @@ static int verify_io_u_sha512(struct verify_header *hdr, struct vcont *vc)
 	dprint(FD_VERIFY, "sha512 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
 	fio_sha512_init(&sha512_ctx);
-	fio_sha512_update(&sha512_ctx, p, hdr->len - hdr_size(hdr));
+	fio_sha512_update(&sha512_ctx, p, hdr->len - hdr_size(vc->td, hdr));
 
 	if (!memcmp(vh->sha512, sha512_ctx.buf, sizeof(sha512)))
 		return 0;
@@ -441,7 +556,8 @@ static int verify_io_u_sha256(struct verify_header *hdr, struct vcont *vc)
 	dprint(FD_VERIFY, "sha256 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
 	fio_sha256_init(&sha256_ctx);
-	fio_sha256_update(&sha256_ctx, p, hdr->len - hdr_size(hdr));
+	fio_sha256_update(&sha256_ctx, p, hdr->len - hdr_size(vc->td, hdr));
+	fio_sha256_final(&sha256_ctx);
 
 	if (!memcmp(vh->sha256, sha256_ctx.buf, sizeof(sha256)))
 		return 0;
@@ -466,7 +582,8 @@ static int verify_io_u_sha1(struct verify_header *hdr, struct vcont *vc)
 	dprint(FD_VERIFY, "sha1 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
 	fio_sha1_init(&sha1_ctx);
-	fio_sha1_update(&sha1_ctx, p, hdr->len - hdr_size(hdr));
+	fio_sha1_update(&sha1_ctx, p, hdr->len - hdr_size(vc->td, hdr));
+	fio_sha1_final(&sha1_ctx);
 
 	if (!memcmp(vh->sha1, sha1_ctx.H, sizeof(sha1)))
 		return 0;
@@ -487,7 +604,7 @@ static int verify_io_u_crc7(struct verify_header *hdr, struct vcont *vc)
 
 	dprint(FD_VERIFY, "crc7 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
-	c = fio_crc7(p, hdr->len - hdr_size(hdr));
+	c = fio_crc7(p, hdr->len - hdr_size(vc->td, hdr));
 
 	if (c == vh->crc7)
 		return 0;
@@ -508,7 +625,7 @@ static int verify_io_u_crc16(struct verify_header *hdr, struct vcont *vc)
 
 	dprint(FD_VERIFY, "crc16 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
-	c = fio_crc16(p, hdr->len - hdr_size(hdr));
+	c = fio_crc16(p, hdr->len - hdr_size(vc->td, hdr));
 
 	if (c == vh->crc16)
 		return 0;
@@ -529,7 +646,7 @@ static int verify_io_u_crc64(struct verify_header *hdr, struct vcont *vc)
 
 	dprint(FD_VERIFY, "crc64 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
-	c = fio_crc64(p, hdr->len - hdr_size(hdr));
+	c = fio_crc64(p, hdr->len - hdr_size(vc->td, hdr));
 
 	if (c == vh->crc64)
 		return 0;
@@ -550,7 +667,7 @@ static int verify_io_u_crc32(struct verify_header *hdr, struct vcont *vc)
 
 	dprint(FD_VERIFY, "crc32 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
-	c = fio_crc32(p, hdr->len - hdr_size(hdr));
+	c = fio_crc32(p, hdr->len - hdr_size(vc->td, hdr));
 
 	if (c == vh->crc32)
 		return 0;
@@ -571,7 +688,7 @@ static int verify_io_u_crc32c(struct verify_header *hdr, struct vcont *vc)
 
 	dprint(FD_VERIFY, "crc32c verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
-	c = fio_crc32c(p, hdr->len - hdr_size(hdr));
+	c = fio_crc32c(p, hdr->len - hdr_size(vc->td, hdr));
 
 	if (c == vh->crc32)
 		return 0;
@@ -596,7 +713,8 @@ static int verify_io_u_md5(struct verify_header *hdr, struct vcont *vc)
 	dprint(FD_VERIFY, "md5 verify io_u %p, len %u\n", vc->io_u, hdr->len);
 
 	fio_md5_init(&md5_ctx);
-	fio_md5_update(&md5_ctx, p, hdr->len - hdr_size(hdr));
+	fio_md5_update(&md5_ctx, p, hdr->len - hdr_size(vc->td, hdr));
+	fio_md5_final(&md5_ctx);
 
 	if (!memcmp(vh->md5_digest, md5_ctx.hash, sizeof(hash)))
 		return 0;
@@ -612,87 +730,173 @@ static int verify_io_u_md5(struct verify_header *hdr, struct vcont *vc)
 /*
  * Push IO verification to a separate thread
  */
-int verify_io_u_async(struct thread_data *td, struct io_u *io_u)
+int verify_io_u_async(struct thread_data *td, struct io_u **io_u_ptr)
 {
-	if (io_u->file)
-		put_file_log(td, io_u->file);
+	struct io_u *io_u = *io_u_ptr;
 
 	pthread_mutex_lock(&td->io_u_lock);
 
+	if (io_u->file)
+		put_file_log(td, io_u->file);
+
 	if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
 		td->cur_depth--;
-		io_u->flags &= ~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->flags |= IO_U_F_FREE_DEF;
-	pthread_mutex_unlock(&td->io_u_lock);
+	*io_u_ptr = NULL;
 
 	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 verify_header *hdr)
+static int verify_header(struct io_u *io_u, struct thread_data *td,
+			 struct verify_header *hdr, unsigned int hdr_num,
+			 unsigned int hdr_len)
 {
 	void *p = hdr;
 	uint32_t crc;
 
-	if (hdr->magic != FIO_HDR_MAGIC)
-		return 1;
-	if (hdr->len > io_u->buflen)
-		return 2;
-	if (hdr->rand_seed != io_u->rand_seed)
-		return 3;
+	if (hdr->magic != FIO_HDR_MAGIC) {
+		log_err("verify: bad magic header %x, wanted %x",
+			hdr->magic, FIO_HDR_MAGIC);
+		goto err;
+	}
+	if (hdr->len != hdr_len) {
+		log_err("verify: bad header length %u, wanted %u",
+			hdr->len, hdr_len);
+		goto err;
+	}
+	if (hdr->rand_seed != io_u->rand_seed) {
+		log_err("verify: bad header rand_seed %"PRIu64
+			", wanted %"PRIu64,
+			hdr->rand_seed, io_u->rand_seed);
+		goto err;
+	}
+	if (hdr->offset != io_u->offset + hdr_num * td->o.verify_interval) {
+		log_err("verify: bad header offset %"PRIu64
+			", wanted %llu",
+			hdr->offset, io_u->offset);
+		goto err;
+	}
+
+	/*
+	 * For read-only workloads, the program cannot be certain of the
+	 * last numberio written to a block. Checking of numberio will be
+	 * done only for workloads that write data.  For verify_only,
+	 * numberio will be checked in the last iteration when the correct
+	 * 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_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) {
+				log_err("verify: bad header numberio %"PRIu16
+					", wanted %"PRIu16,
+					hdr->numberio, io_u->numberio);
+				goto err;
+			}
 
 	crc = fio_crc32c(p, offsetof(struct verify_header, crc32));
-	if (crc == hdr->crc32)
-		return 0;
-	log_err("fio: verify header crc %x, calculated %x\n", hdr->crc32, crc);
-	return 4;
+	if (crc != hdr->crc32) {
+		log_err("verify: bad header crc %x, calculated %x",
+			hdr->crc32, crc);
+		goto err;
+	}
+	return 0;
+
+err:
+	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, 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;
 }
 
-int verify_io_u(struct thread_data *td, struct io_u *io_u)
+int verify_io_u(struct thread_data *td, struct io_u **io_u_ptr)
 {
 	struct verify_header *hdr;
+	struct io_u *io_u = *io_u_ptr;
 	unsigned int header_size, hdr_inc, hdr_num = 0;
 	void *p;
 	int ret;
 
 	if (td->o.verify == VERIFY_NULL || io_u->ddir != DDIR_READ)
 		return 0;
+	/*
+	 * If the IO engine is faking IO (like null), then just pretend
+	 * we verified everything.
+	 */
+	if (td_ioengine_flagged(td, FIO_FAKEIO))
+		return 0;
+
 	if (io_u->flags & IO_U_F_TRIMMED) {
 		ret = verify_trimmed_io_u(td, io_u);
 		goto done;
@@ -719,47 +923,15 @@ int verify_io_u(struct thread_data *td, struct io_u *io_u)
 		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;
 
-		ret = verify_header(io_u, hdr);
-		switch (ret) {
-		case 0:
-			break;
-		case 1:
-			log_err("verify: bad magic header %x, wanted %x at "
-				"file %s offset %llu, length %u\n",
-				hdr->magic, FIO_HDR_MAGIC,
-				io_u->file->file_name,
-				io_u->offset + hdr_num * hdr->len, hdr->len);
-			return EILSEQ;
-			break;
-		case 2:
-			log_err("fio: verify header exceeds buffer length (%u "
-				"> %lu)\n", hdr->len, io_u->buflen);
-			return EILSEQ;
-			break;
-		case 3:
-			log_err("verify: bad header rand_seed %"PRIu64
-				", wanted %"PRIu64" at file %s offset %llu, "
-				"length %u\n",
-				hdr->rand_seed, io_u->rand_seed,
-				io_u->file->file_name,
-				io_u->offset + hdr_num * hdr->len, hdr->len);
-			return EILSEQ;
-			break;
-		case 4:
-			return EILSEQ;
-			break;
-		default:
-			log_err("verify: unknown header error at file %s "
-			"offset %llu, length %u\n",
-			io_u->file->file_name,
-			io_u->offset + hdr_num * hdr->len, hdr->len);
-			return EILSEQ;
+		if (td->o.verify != VERIFY_PATTERN_NO_HDR) {
+			ret = verify_header(io_u, td, hdr, hdr_num, hdr_inc);
+			if (ret)
+				return ret;
 		}
 
 		if (td->o.verify != VERIFY_NONE)
@@ -768,6 +940,12 @@ int verify_io_u(struct thread_data *td, struct io_u *io_u)
 			verify_type = hdr->verify_type;
 
 		switch (verify_type) {
+		case VERIFY_HDR_ONLY:
+			/* Header is always verified, check if pattern is left
+			 * for verification. */
+			if (td->o.verify_pattern_bytes)
+				ret = verify_io_u_pattern(hdr, &vc);
+			break;
 		case VERIFY_MD5:
 			ret = verify_io_u_md5(hdr, &vc);
 			break;
@@ -793,13 +971,26 @@ int verify_io_u(struct thread_data *td, struct io_u *io_u)
 		case VERIFY_SHA512:
 			ret = verify_io_u_sha512(hdr, &vc);
 			break;
-		case VERIFY_META:
-			ret = verify_io_u_meta(hdr, &vc);
+		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;
 		case VERIFY_SHA1:
 			ret = verify_io_u_sha1(hdr, &vc);
 			break;
 		case VERIFY_PATTERN:
+		case VERIFY_PATTERN_NO_HDR:
 			ret = verify_io_u_pattern(hdr, &vc);
 			break;
 		default:
@@ -814,24 +1005,69 @@ int verify_io_u(struct thread_data *td, struct io_u *io_u)
 
 done:
 	if (ret && td->o.verify_fatal)
-		td->terminate = 1;
+		fio_mark_td_terminate(td);
 
 	return ret;
 }
 
-static void fill_meta(struct verify_header *hdr, struct thread_data *td,
-		      struct io_u *io_u, unsigned int header_num)
+static void fill_xxhash(struct verify_header *hdr, void *p, unsigned int len)
+{
+	struct vhdr_xxhash *vh = hdr_priv(hdr);
+	void *state;
+
+	state = XXH32_init(1);
+	XXH32_update(state, p, 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_meta *vh = hdr_priv(hdr);
+	struct vhdr_sha3_224 *vh = hdr_priv(hdr);
+	struct fio_sha3_ctx sha3_ctx = {
+		.sha = vh->sha,
+	};
 
-	vh->thread = td->thread_number;
+	fio_sha3_224_init(&sha3_ctx);
+	fill_sha3(&sha3_ctx, p, len);
+}
 
-	vh->time_sec = io_u->start_time.tv_sec;
-	vh->time_usec = io_u->start_time.tv_usec;
+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,
+	};
 
-	vh->numberio = io_u->numberio;
+	fio_sha3_256_init(&sha3_ctx);
+	fill_sha3(&sha3_ctx, p, len);
+}
 
-	vh->offset = io_u->offset + header_num * td->o.verify_interval;
+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)
@@ -854,6 +1090,7 @@ static void fill_sha256(struct verify_header *hdr, void *p, unsigned int len)
 
 	fio_sha256_init(&sha256_ctx);
 	fio_sha256_update(&sha256_ctx, p, len);
+	fio_sha256_final(&sha256_ctx);
 }
 
 static void fill_sha1(struct verify_header *hdr, void *p, unsigned int len)
@@ -865,6 +1102,7 @@ static void fill_sha1(struct verify_header *hdr, void *p, unsigned int len)
 
 	fio_sha1_init(&sha1_ctx);
 	fio_sha1_update(&sha1_ctx, p, len);
+	fio_sha1_final(&sha1_ctx);
 }
 
 static void fill_crc7(struct verify_header *hdr, void *p, unsigned int len)
@@ -911,6 +1149,34 @@ static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
 
 	fio_md5_init(&md5_ctx);
 	fio_md5_update(&md5_ctx, p, len);
+	fio_md5_final(&md5_ctx);
+}
+
+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)
+{
+	void *p = hdr;
+
+	hdr->magic = FIO_HDR_MAGIC;
+	hdr->verify_type = td->o.verify;
+	hdr->len = header_len;
+	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_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));
+}
+
+
+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);
 }
 
 static void populate_hdr(struct thread_data *td, struct io_u *io_u,
@@ -918,19 +1184,16 @@ 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;
 
-	hdr->magic = FIO_HDR_MAGIC;
-	hdr->verify_type = td->o.verify;
-	hdr->len = header_len;
-	hdr->rand_seed = io_u->rand_seed;
-	hdr->crc32 = fio_crc32c(p, offsetof(struct verify_header, crc32));
+	fill_hdr(td, io_u, hdr, header_num, header_len, io_u->rand_seed);
 
-	data_len = header_len - hdr_size(hdr);
+	data_len = header_len - hdr_size(td, hdr);
 
-	data = p + hdr_size(hdr);
+	data = p + hdr_size(td, hdr);
 	switch (td->o.verify) {
 	case VERIFY_MD5:
 		dprint(FD_VERIFY, "fill md5 io_u %p, len %u\n",
@@ -973,25 +1236,48 @@ 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_META:
-		dprint(FD_VERIFY, "fill meta io_u %p, len %u\n",
+	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_meta(hdr, td, io_u, header_num);
+		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);
+		fill_xxhash(hdr, data, data_len);
 		break;
 	case VERIFY_SHA1:
 		dprint(FD_VERIFY, "fill sha1 io_u %p, len %u\n",
 						io_u, hdr->len);
 		fill_sha1(hdr, data, data_len);
 		break;
+	case VERIFY_HDR_ONLY:
 	case VERIFY_PATTERN:
+	case VERIFY_PATTERN_NO_HDR:
 		/* nothing to do here */
 		break;
 	default:
 		log_err("fio: bad verify type: %d\n", td->o.verify);
 		assert(0);
 	}
-	if (td->o.verify_offset)
-		memswp(p, p + td->o.verify_offset, hdr_size(hdr));
+
+	if (td->o.verify_offset && hdr_size(td, hdr))
+		memswp(p, p + td->o.verify_offset, hdr_size(td, hdr));
 }
 
 /*
@@ -1019,14 +1305,30 @@ 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);
+
+		/*
+		 * Ensure that the associated IO has completed
+		 */
+		read_barrier();
+		if (ipo->flags & IP_F_IN_FLIGHT)
+			goto nothing;
+
 		rb_erase(n, &td->io_hist_tree);
 		assert(ipo->flags & IP_F_ONRB);
 		ipo->flags &= ~IP_F_ONRB;
 	} else if (!flist_empty(&td->io_hist_list)) {
-		ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
+		ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
+
+		/*
+		 * Ensure that the associated IO has completed
+		 */
+		read_barrier();
+		if (ipo->flags & IP_F_IN_FLIGHT)
+			goto nothing;
+
 		flist_del(&ipo->list);
 		assert(ipo->flags & IP_F_ONLIST);
 		ipo->flags &= ~IP_F_ONLIST;
@@ -1039,10 +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->flags |= IO_U_F_VER_LIST;
+		io_u_set(td, io_u, IO_U_F_VER_LIST);
 
 		if (ipo->flags & IP_F_TRIMMED)
-			io_u->flags |= 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);
@@ -1065,13 +1367,14 @@ int get_next_verify(struct thread_data *td, struct io_u *io_u)
 		dprint(FD_VERIFY, "get_next_verify: ret io_u %p\n", io_u);
 
 		if (!td->o.verify_pattern_bytes) {
-			io_u->rand_seed = __rand(&td->__verify_state);
+			io_u->rand_seed = __rand(&td->verify_state);
 			if (sizeof(int) != sizeof(long *))
-				io_u->rand_seed *= __rand(&td->__verify_state);
+				io_u->rand_seed *= __rand(&td->verify_state);
 		}
 		return 0;
 	}
 
+nothing:
 	dprint(FD_VERIFY, "get_next_verify: empty\n");
 	return 1;
 }
@@ -1080,6 +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();
 	}
 }
@@ -1090,7 +1394,7 @@ static void *verify_async_thread(void *data)
 	struct io_u *io_u;
 	int ret = 0;
 
-	if (td->o.verify_cpumask_set &&
+	if (fio_option_is_set(&td->o, verify_cpumask) &&
 	    fio_setaffinity(td->pid, td->o.verify_cpumask)) {
 		log_err("fio: failed setting verify thread affinity\n");
 		goto done;
@@ -1122,10 +1426,12 @@ static void *verify_async_thread(void *data)
 			continue;
 
 		while (!flist_empty(&list)) {
-			io_u = flist_entry(list.next, struct io_u, verify_list);
-			flist_del(&io_u->verify_list);
+			io_u = flist_first_entry(&list, struct io_u, verify_list);
+			flist_del_init(&io_u->verify_list);
+
+			io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
+			ret = verify_io_u(td, &io_u);
 
-			ret = verify_io_u(td, io_u);
 			put_io_u(td, io_u);
 			if (!ret)
 				continue;
@@ -1140,15 +1446,15 @@ static void *verify_async_thread(void *data)
 	if (ret) {
 		td_verror(td, ret, "async_verify");
 		if (td->o.verify_fatal)
-			td->terminate = 1;
+			fio_mark_td_terminate(td);
 	}
 
 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;
 }
 
@@ -1158,7 +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;
 
@@ -1184,9 +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;
 	}
 
@@ -1195,12 +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);
 
@@ -1208,3 +1515,368 @@ void verify_async_exit(struct thread_data *td)
 	free(td->verify_threads);
 	td->verify_threads = NULL;
 }
+
+int paste_blockoff(char *buf, unsigned int len, void *priv)
+{
+	struct io_u *io = priv;
+	unsigned long long off;
+
+	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;
+	struct thread_data *td;
+	size_t depth;
+	void *next;
+	int i, nr;
+
+	compiletime_assert(sizeof(struct all_io_list) == 8, "all_io_list");
+
+	/*
+	 * Calculate reply space needed. We need one 'io_state' per thread,
+	 * and the size will vary depending on depth.
+	 */
+	depth = 0;
+	nr = 0;
+	for_each_td(td, i) {
+		if (save_mask != IO_LIST_ALL && (i + 1) != save_mask)
+			continue;
+		td->stop_io = 1;
+		td->flags |= TD_F_VSTATE_SAVED;
+		depth += (td->o.iodepth * td->o.nr_files);
+		nr++;
+	}
+
+	if (!nr)
+		return NULL;
+
+	*sz = sizeof(*rep);
+	*sz += nr * sizeof(struct thread_io_list);
+	*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, index = 0;
+
+		if (save_mask != IO_LIST_ALL && (i + 1) != save_mask)
+			continue;
+
+		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) {
+			s->rand.state64.s[0] = cpu_to_le64(td->random_state.state64.s1);
+			s->rand.state64.s[1] = cpu_to_le64(td->random_state.state64.s2);
+			s->rand.state64.s[2] = cpu_to_le64(td->random_state.state64.s3);
+			s->rand.state64.s[3] = cpu_to_le64(td->random_state.state64.s4);
+			s->rand.state64.s[4] = cpu_to_le64(td->random_state.state64.s5);
+			s->rand.state64.s[5] = 0;
+			s->rand.use64 = cpu_to_le64((uint64_t)1);
+		} else {
+			s->rand.state32.s[0] = cpu_to_le32(td->random_state.state32.s1);
+			s->rand.state32.s[1] = cpu_to_le32(td->random_state.state32.s2);
+			s->rand.state32.s[2] = cpu_to_le32(td->random_state.state32.s3);
+			s->rand.state32.s[3] = 0;
+			s->rand.use64 = 0;
+		}
+		s->name[sizeof(s->name) - 1] = '\0';
+		strncpy((char *) s->name, td->o.name, sizeof(s->name) - 1);
+		next = io_list_next(s);
+	}
+
+	return rep;
+}
+
+static int open_state_file(const char *name, const char *prefix, int num,
+			   int for_write)
+{
+	char out[PATH_MAX];
+	int flags;
+	int fd;
+
+	if (for_write)
+		flags = O_CREAT | O_TRUNC | O_WRONLY | O_SYNC;
+	else
+		flags = O_RDONLY;
+
+	verify_state_gen_name(out, sizeof(out), name, prefix, 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;
+	}
+
+	return fd;
+}
+
+static int write_thread_list_state(struct thread_io_list *s,
+				   const char *prefix)
+{
+	struct verify_state_hdr hdr;
+	uint64_t crc;
+	ssize_t ret;
+	int fd;
+
+	fd = open_state_file((const char *) s->name, prefix, s->index, 1);
+	if (fd == -1)
+		return 1;
+
+	crc = fio_crc32c((void *)s, thread_io_list_sz(s));
+
+	hdr.version = cpu_to_le64((uint64_t) VSTATE_HDR_VERSION);
+	hdr.size = cpu_to_le64((uint64_t) thread_io_list_sz(s));
+	hdr.crc = cpu_to_le64(crc);
+	ret = write(fd, &hdr, sizeof(hdr));
+	if (ret != sizeof(hdr))
+		goto write_fail;
+
+	ret = write(fd, s, thread_io_list_sz(s));
+	if (ret != thread_io_list_sz(s)) {
+write_fail:
+		if (ret < 0)
+			perror("fio: write state file");
+		log_err("fio: failed to write state file\n");
+		ret = 1;
+	} else
+		ret = 0;
+
+	close(fd);
+	return ret;
+}
+
+void __verify_save_state(struct all_io_list *state, const char *prefix)
+{
+	struct thread_io_list *s = &state->state[0];
+	unsigned int i;
+
+	for (i = 0; i < le64_to_cpu(state->threads); i++) {
+		write_thread_list_state(s,  prefix);
+		s = io_list_next(s);
+	}
+}
+
+void verify_save_state(int mask)
+{
+	struct all_io_list *state;
+	size_t sz;
+
+	state = get_all_io_list(mask, &sz);
+	if (state) {
+		char prefix[PATH_MAX];
+
+		if (aux_path)
+			sprintf(prefix, "%s%clocal", aux_path, FIO_OS_PATH_SEPARATOR);
+		else
+			strcpy(prefix, "local");
+
+		__verify_save_state(state, prefix);
+		free(state);
+	}
+}
+
+void verify_free_state(struct thread_data *td)
+{
+	if (td->vstate)
+		free(td->vstate);
+}
+
+void verify_assign_state(struct thread_data *td, void *p)
+{
+	struct thread_io_list *s = p;
+	int 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);
+
+	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->comps[i].fileno = le64_to_cpu(s->comps[i].fileno);
+		s->comps[i].offset = le64_to_cpu(s->comps[i].offset);
+	}
+
+	td->vstate = p;
+}
+
+int verify_state_hdr(struct verify_state_hdr *hdr, struct thread_io_list *s)
+{
+	uint64_t crc;
+
+	hdr->version = le64_to_cpu(hdr->version);
+	hdr->size = le64_to_cpu(hdr->size);
+	hdr->crc = le64_to_cpu(hdr->crc);
+
+	if (hdr->version != VSTATE_HDR_VERSION)
+		return 1;
+
+	crc = fio_crc32c((void *)s, hdr->size);
+	if (crc != hdr->crc)
+		return 1;
+
+	return 0;
+}
+
+int verify_load_state(struct thread_data *td, const char *prefix)
+{
+	struct verify_state_hdr hdr;
+	void *s = NULL;
+	uint64_t crc;
+	ssize_t ret;
+	int fd;
+
+	if (!td->o.verify_state)
+		return 0;
+
+	fd = open_state_file(td->o.name, prefix, td->thread_number - 1, 0);
+	if (fd == -1)
+		return 1;
+
+	ret = read(fd, &hdr, sizeof(hdr));
+	if (ret != sizeof(hdr)) {
+		if (ret < 0)
+			td_verror(td, errno, "read verify state hdr");
+		log_err("fio: failed reading verify state header\n");
+		goto err;
+	}
+
+	hdr.version = le64_to_cpu(hdr.version);
+	hdr.size = le64_to_cpu(hdr.size);
+	hdr.crc = le64_to_cpu(hdr.crc);
+
+	if (hdr.version != VSTATE_HDR_VERSION) {
+		log_err("fio: unsupported (%d) version in verify state header\n",
+				(unsigned int) hdr.version);
+		goto err;
+	}
+
+	s = malloc(hdr.size);
+	ret = read(fd, s, hdr.size);
+	if (ret != hdr.size) {
+		if (ret < 0)
+			td_verror(td, errno, "read verify state");
+		log_err("fio: failed reading verity state\n");
+		goto err;
+	}
+
+	crc = fio_crc32c(s, hdr.size);
+	if (crc != hdr.crc) {
+		log_err("fio: verify state is corrupt\n");
+		goto err;
+	}
+
+	close(fd);
+
+	verify_assign_state(td, s);
+	return 0;
+err:
+	if (s)
+		free(s);
+	close(fd);
+	return 1;
+}
+
+/*
+ * Use the loaded verify state to know when to stop doing verification
+ */
+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 || !f)
+		return 0;
+
+	/*
+	 * If we're not into the window of issues - depth yet, continue. If
+	 * issue is shorter than depth, do check.
+	 */
+	if ((td->io_blocks[DDIR_READ] < s->depth ||
+	    s->numberio - td->io_blocks[DDIR_READ] > s->depth) &&
+	    s->numberio > s->depth)
+		return 0;
+
+	/*
+	 * We're in the window of having to check if this io was
+	 * completed or not. If the IO was seen as completed, then
+	 * lets verify it.
+	 */
+	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
+	 */
+	return 1;
+}