Add LFSR generator
[fio.git] / io_u.c
diff --git a/io_u.c b/io_u.c
index 2deb5c7216aea5e02c66cf8237fb7648fc3a2dd2..2904ad40bd7183d1a4e9bab64fc434d3ebe4ff2b 100644 (file)
--- a/io_u.c
+++ b/io_u.c
 #include "verify.h"
 #include "trim.h"
 #include "lib/rand.h"
+#include "lib/bitmap.h"
 
 struct io_completion_data {
        int nr;                         /* input */
-       int account;                    /* input */
 
        int error;                      /* output */
-       unsigned long bytes_done[2];    /* output */
+       unsigned long bytes_done[DDIR_RWDIR_CNT];       /* output */
        struct timeval time;            /* output */
 };
 
 /*
- * The ->file_map[] contains a map of blocks we have or have not done io
+ * The ->io_bitmap contains a map of blocks we have or have not done io
  * to yet. Used to make sure we cover the entire range in a fair fashion.
  */
 static int random_map_free(struct fio_file *f, const unsigned long long block)
 {
-       unsigned int idx = RAND_MAP_IDX(f, block);
-       unsigned int bit = RAND_MAP_BIT(f, block);
-
-       dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
-
-       return (f->file_map[idx] & (1UL << bit)) == 0;
+       return !bitmap_isset(f->io_bitmap, block);
 }
 
 /*
@@ -42,61 +37,16 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
        unsigned int min_bs = td->o.rw_min_bs;
        struct fio_file *f = io_u->file;
        unsigned long long block;
-       unsigned int blocks, nr_blocks;
-       int busy_check;
+       unsigned int nr_blocks;
 
        block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
        nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
-       blocks = 0;
-       busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
-
-       while (nr_blocks) {
-               unsigned int idx, bit;
-               unsigned long mask, this_blocks;
-
-               /*
-                * If we have a mixed random workload, we may
-                * encounter blocks we already did IO to.
-                */
-               if (!busy_check) {
-                       blocks = nr_blocks;
-                       break;
-               }
-               if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
-                       break;
-
-               idx = RAND_MAP_IDX(f, block);
-               bit = RAND_MAP_BIT(f, block);
-
-               fio_assert(td, idx < f->num_maps);
-
-               this_blocks = nr_blocks;
-               if (this_blocks + bit > BLOCKS_PER_MAP)
-                       this_blocks = BLOCKS_PER_MAP - bit;
-
-               do {
-                       if (this_blocks == BLOCKS_PER_MAP)
-                               mask = -1UL;
-                       else
-                               mask = ((1UL << this_blocks) - 1) << bit;
-       
-                       if (!(f->file_map[idx] & mask))
-                               break;
 
-                       this_blocks--;
-               } while (this_blocks);
+       if (!(io_u->flags & IO_U_F_BUSY_OK))
+               nr_blocks = bitmap_set_nr(f->io_bitmap, block, nr_blocks);
 
-               if (!this_blocks)
-                       break;
-
-               f->file_map[idx] |= mask;
-               nr_blocks -= this_blocks;
-               blocks += this_blocks;
-               block += this_blocks;
-       }
-
-       if ((blocks * min_bs) < io_u->buflen)
-               io_u->buflen = blocks * min_bs;
+       if ((nr_blocks * min_bs) < io_u->buflen)
+               io_u->buflen = nr_blocks * min_bs;
 }
 
 static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
@@ -124,134 +74,116 @@ static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
        return max_blocks;
 }
 
-/*
- * Return the next free block in the map.
- */
-static int get_next_free_block(struct thread_data *td, struct fio_file *f,
-                              enum fio_ddir ddir, unsigned long long *b)
+static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+                                 enum fio_ddir ddir, unsigned long long *b)
 {
-       unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
-       int i;
-
-       lastb = last_block(td, f, ddir);
-       if (!lastb)
-               return 1;
+       unsigned long long r;
 
-       i = f->last_free_lookup;
-       block = i * BLOCKS_PER_MAP;
-       while (block * min_bs < f->real_file_size &&
-               block * min_bs < f->io_size) {
-               if (f->file_map[i] != -1UL) {
-                       block += ffz(f->file_map[i]);
-                       if (block > lastb)
-                               break;
-                       f->last_free_lookup = i;
-                       *b = block;
-                       return 0;
-               }
-
-               block += BLOCKS_PER_MAP;
-               i++;
-       }
+       if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
+               unsigned long long rmax, lastb;
 
-       dprint(FD_IO, "failed finding a free block\n");
-       return 1;
-}
-
-static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
-                               enum fio_ddir ddir, unsigned long long *b)
-{
-       unsigned long long rmax, r, lastb;
-       int loops = 5;
-
-       lastb = last_block(td, f, ddir);
-       if (!lastb)
-               return 1;
+               lastb = last_block(td, f, ddir);
+               if (!lastb)
+                       return 1;
 
-       if (f->failed_rands >= 200)
-               goto ffz;
+               rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
 
-       rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
-       do {
-               if (td->o.use_os_rand)
+               if (td->o.use_os_rand) {
+                       rmax = OS_RAND_MAX;
                        r = os_random_long(&td->random_state);
-               else
+               } else {
+                       rmax = FRAND_MAX;
                        r = __rand(&td->__random_state);
-
-               *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
+               }
 
                dprint(FD_RANDOM, "off rand %llu\n", r);
 
+               *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
+       } else {
+               uint64_t off = 0;
 
-               /*
-                * if we are not maintaining a random map, we are done.
-                */
-               if (!file_randommap(td, f))
-                       goto ret_good;
-
-               /*
-                * calculate map offset and check if it's free
-                */
-               if (random_map_free(f, *b))
-                       goto ret_good;
-
-               dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
-                                                                       *b);
-       } while (--loops);
+               if (lfsr_next(&f->lfsr, &off))
+                       return 1;
 
-       if (!f->failed_rands++)
-               f->last_free_lookup = 0;
+               *b = off;
+       }
 
        /*
-        * we get here, if we didn't suceed in looking up a block. generate
-        * a random start offset into the filemap, and find the first free
-        * block from there.
+        * if we are not maintaining a random map, we are done.
         */
-       loops = 10;
-       do {
-               f->last_free_lookup = (f->num_maps - 1) *
-                                       (r / ((unsigned long long) rmax + 1.0));
-               if (!get_next_free_block(td, f, ddir, b))
-                       goto ret;
-
-               if (td->o.use_os_rand)
-                       r = os_random_long(&td->random_state);
-               else
-                       r = __rand(&td->__random_state);
-       } while (--loops);
+       if (!file_randommap(td, f))
+               goto ret;
 
        /*
-        * that didn't work either, try exhaustive search from the start
+        * calculate map offset and check if it's free
         */
-       f->last_free_lookup = 0;
-ffz:
-       if (!get_next_free_block(td, f, ddir, b))
-               return 0;
-       f->last_free_lookup = 0;
-       return get_next_free_block(td, f, ddir, b);
-ret_good:
-       f->failed_rands = 0;
+       if (random_map_free(f, *b))
+               goto ret;
+
+       dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", *b);
+
+       *b = bitmap_next_free(f->io_bitmap, *b);
+       if (*b == (uint64_t) -1ULL)
+               return 1;
 ret:
        return 0;
 }
 
+static int __get_next_rand_offset_zipf(struct thread_data *td,
+                                      struct fio_file *f, enum fio_ddir ddir,
+                                      unsigned long long *b)
+{
+       *b = zipf_next(&f->zipf);
+       return 0;
+}
+
+static int __get_next_rand_offset_pareto(struct thread_data *td,
+                                        struct fio_file *f, enum fio_ddir ddir,
+                                        unsigned long long *b)
+{
+       *b = pareto_next(&f->zipf);
+       return 0;
+}
+
+static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+                               enum fio_ddir ddir, unsigned long long *b)
+{
+       if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
+               return __get_next_rand_offset(td, f, ddir, b);
+       else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
+               return __get_next_rand_offset_zipf(td, f, ddir, b);
+       else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
+               return __get_next_rand_offset_pareto(td, f, ddir, b);
+
+       log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
+       return 1;
+}
+
 static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
                               enum fio_ddir ddir, unsigned long long *b)
 {
-       if (get_next_rand_offset(td, f, ddir, b)) {
-               dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
-                               f->file_name, f->last_pos, f->real_file_size);
-               return 1;
+       if (!get_next_rand_offset(td, f, ddir, b))
+               return 0;
+
+       if (td->o.time_based) {
+               fio_file_reset(f);
+               if (!get_next_rand_offset(td, f, ddir, b))
+                       return 0;
        }
 
-       return 0;
+       dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
+                       f->file_name, f->last_pos, f->real_file_size);
+       return 1;
 }
 
-static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
-                             enum fio_ddir ddir, unsigned long long *b)
+static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
+                              enum fio_ddir ddir, unsigned long long *offset)
 {
        assert(ddir_rw(ddir));
 
+       if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
+               f->last_pos = f->last_pos - f->io_size;
+
        if (f->last_pos < f->real_file_size) {
                unsigned long long pos;
 
@@ -262,7 +194,7 @@ static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
                if (pos)
                        pos += td->o.ddir_seq_add;
 
-               *b = pos / td->o.min_bs[ddir];
+               *offset = pos;
                return 0;
        }
 
@@ -270,31 +202,33 @@ static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
 }
 
 static int get_next_block(struct thread_data *td, struct io_u *io_u,
-                         enum fio_ddir ddir, int rw_seq, unsigned long long *b)
+                         enum fio_ddir ddir, int rw_seq)
 {
        struct fio_file *f = io_u->file;
+       unsigned long long b, offset;
        int ret;
 
        assert(ddir_rw(ddir));
 
+       b = offset = -1ULL;
+
        if (rw_seq) {
                if (td_random(td))
-                       ret = get_next_rand_block(td, f, ddir, b);
+                       ret = get_next_rand_block(td, f, ddir, &b);
                else
-                       ret = get_next_seq_block(td, f, ddir, b);
+                       ret = get_next_seq_offset(td, f, ddir, &offset);
        } else {
                io_u->flags |= IO_U_F_BUSY_OK;
 
                if (td->o.rw_seq == RW_SEQ_SEQ) {
-                       ret = get_next_seq_block(td, f, ddir, b);
+                       ret = get_next_seq_offset(td, f, ddir, &offset);
                        if (ret)
-                               ret = get_next_rand_block(td, f, ddir, b);
+                               ret = get_next_rand_block(td, f, ddir, &b);
                } else if (td->o.rw_seq == RW_SEQ_IDENT) {
                        if (f->last_start != -1ULL)
-                               *b = (f->last_start - f->file_offset)
-                                       / td->o.min_bs[ddir];
+                               offset = f->last_start - f->file_offset;
                        else
-                               *b = 0;
+                               offset = 0;
                        ret = 0;
                } else {
                        log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
@@ -302,6 +236,18 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
                }
        }
        
+       if (!ret) {
+               if (offset != -1ULL)
+                       io_u->offset = offset;
+               else if (b != -1ULL)
+                       io_u->offset = b * td->o.ba[ddir];
+               else {
+                       log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
+                                                               offset, b);
+                       ret = 1;
+               }
+       }
+
        return ret;
 }
 
@@ -313,7 +259,6 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
        struct fio_file *f = io_u->file;
-       unsigned long long b;
        enum fio_ddir ddir = io_u->ddir;
        int rw_seq_hit = 0;
 
@@ -324,10 +269,9 @@ static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
                td->ddir_seq_nr = td->o.ddir_seq_nr;
        }
 
-       if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
+       if (get_next_block(td, io_u, ddir, rw_seq_hit))
                return 1;
 
-       io_u->offset = b * td->o.ba[ddir];
        if (io_u->offset >= f->io_size) {
                dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
                                        io_u->offset, f->io_size);
@@ -359,13 +303,13 @@ static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
 {
        struct fio_file *f = io_u->file;
 
-       return io_u->offset + buflen <= f->io_size + td->o.start_offset;
+       return io_u->offset + buflen <= f->io_size + get_start_offset(td);
 }
 
 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
        const int ddir = io_u->ddir;
-       unsigned int uninitialized_var(buflen);
+       unsigned int buflen = 0;
        unsigned int minbs, maxbs;
        unsigned long r, rand_max;
 
@@ -525,6 +469,8 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
        if (td_rw(td) && __should_check_rate(td, odir))
                td->rate_pending_usleep[odir] -= usec;
 
+       if (ddir_trim(ddir))
+               return ddir;
        return ddir;
 }
 
@@ -581,8 +527,10 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
                ddir = td->rwmix_ddir;
        } else if (td_read(td))
                ddir = DDIR_READ;
-       else
+       else if (td_write(td))
                ddir = DDIR_WRITE;
+       else
+               ddir = DDIR_TRIM;
 
        td->rwmix_ddir = rate_ddir(td, ddir);
        return td->rwmix_ddir;
@@ -1070,10 +1018,11 @@ again:
                assert(io_u->flags & IO_U_F_FREE);
                io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
                io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
+               io_u->flags &= ~IO_U_F_VER_LIST;
 
                io_u->error = 0;
                flist_del(&io_u->list);
-               flist_add(&io_u->list, &td->io_u_busylist);
+               flist_add_tail(&io_u->list, &td->io_u_busylist);
                td->cur_depth++;
                io_u->flags |= IO_U_F_IN_CUR_DEPTH;
        } else if (td->o.verify_async) {
@@ -1117,10 +1066,9 @@ static int check_get_verify(struct thread_data *td, struct io_u *io_u)
        if (td->o.verify_backlog && td->io_hist_len) {
                int get_verify = 0;
 
-               if (td->verify_batch) {
-                       td->verify_batch--;
+               if (td->verify_batch)
                        get_verify = 1;
-               else if (!(td->io_hist_len % td->o.verify_backlog) &&
+               else if (!(td->io_hist_len % td->o.verify_backlog) &&
                         td->last_ddir != DDIR_READ) {
                        td->verify_batch = td->o.verify_batch;
                        if (!td->verify_batch)
@@ -1128,8 +1076,10 @@ static int check_get_verify(struct thread_data *td, struct io_u *io_u)
                        get_verify = 1;
                }
 
-               if (get_verify && !get_next_verify(td, io_u))
+               if (get_verify && !get_next_verify(td, io_u)) {
+                       td->verify_batch--;
                        return 1;
+               }
        }
 
        return 0;
@@ -1225,13 +1175,15 @@ struct io_u *get_io_u(struct thread_data *td)
                f->last_pos = io_u->offset + io_u->buflen;
 
                if (io_u->ddir == DDIR_WRITE) {
-                       if (td->o.verify != VERIFY_NONE)
-                               populate_verify_io_u(td, io_u);
-                       else if (td->o.refill_buffers) {
+                       if (td->o.refill_buffers) {
                                io_u_fill_buffer(td, io_u,
                                        io_u->xfer_buflen, io_u->xfer_buflen);
                        } else if (td->o.scramble_buffers)
                                do_scramble = 1;
+                       if (td->o.verify != VERIFY_NONE) {
+                               populate_verify_io_u(td, io_u);
+                               do_scramble = 0;
+                       }
                } else if (io_u->ddir == DDIR_READ) {
                        /*
                         * Reset the buf_filled parameters so next time if the
@@ -1264,10 +1216,12 @@ err_put:
 
 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 {
+       enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
        const char *msg[] = { "read", "write", "sync", "datasync",
                                "sync_file_range", "wait", "trim" };
 
-
+       if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
+               return;
 
        log_err("fio: io_u error");
 
@@ -1287,10 +1241,7 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
                                  struct io_completion_data *icd,
                                  const enum fio_ddir idx, unsigned int bytes)
 {
-       unsigned long uninitialized_var(lusec);
-
-       if (!icd->account)
-               return;
+       unsigned long lusec = 0;
 
        if (!td->o.disable_clat || !td->o.disable_bw)
                lusec = utime_since(&io_u->issue_time, &icd->time);
@@ -1300,6 +1251,13 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 
                tusec = utime_since(&io_u->start_time, &icd->time);
                add_lat_sample(td, idx, tusec, bytes);
+
+               if (td->o.max_latency && tusec > td->o.max_latency) {
+                       if (!td->error)
+                               log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
+                       td_verror(td, ETIMEDOUT, "max latency exceeded");
+                       icd->error = ETIMEDOUT;
+               }
        }
 
        if (!td->o.disable_clat) {
@@ -1326,11 +1284,6 @@ static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 static void io_completed(struct thread_data *td, struct io_u *io_u,
                         struct io_completion_data *icd)
 {
-       /*
-        * Older gcc's are too dumb to realize that usec is always used
-        * initialized, silence that warning.
-        */
-       unsigned long uninitialized_var(usec);
        struct fio_file *f;
 
        dprint_io_u(io_u, "io complete");
@@ -1362,7 +1315,9 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
                td->io_blocks[idx]++;
                td->this_io_blocks[idx]++;
                td->io_bytes[idx] += bytes;
-               td->this_io_bytes[idx] += bytes;
+
+               if (!(io_u->flags & IO_U_F_VER_LIST))
+                       td->this_io_bytes[idx] += bytes;
 
                if (idx == DDIR_WRITE) {
                        f = io_u->file;
@@ -1385,7 +1340,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
                                        (usec_for_io(td, idx) -
                                         utime_since_now(&td->start));
                        }
-                       if (__should_check_rate(td, odx))
+                       if (idx != DDIR_TRIM && __should_check_rate(td, odx))
                                td->rate_pending_usleep[odx] =
                                        (usec_for_io(td, odx) -
                                         utime_since_now(&td->start));
@@ -1407,8 +1362,10 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
                icd->error = io_u->error;
                io_u_log_error(td, io_u);
        }
-       if (icd->error && td_non_fatal_error(icd->error) &&
-           (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
+       if (icd->error) {
+               enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
+               if (!td_non_fatal_error(td, eb, icd->error))
+                       return;
                /*
                 * If there is a non_fatal error, then add to the error count
                 * and clear all the errors.
@@ -1423,14 +1380,15 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 static void init_icd(struct thread_data *td, struct io_completion_data *icd,
                     int nr)
 {
+       int ddir;
        if (!td->o.disable_clat || !td->o.disable_bw)
                fio_gettime(&icd->time, NULL);
 
        icd->nr = nr;
-       icd->account = 1;
 
        icd->error = 0;
-       icd->bytes_done[0] = icd->bytes_done[1] = 0;
+       for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+               icd->bytes_done[ddir] = 0;
 }
 
 static void ios_completed(struct thread_data *td,
@@ -1446,8 +1404,6 @@ static void ios_completed(struct thread_data *td,
 
                if (!(io_u->flags & IO_U_F_FREE_DEF))
                        put_io_u(td, io_u);
-
-               icd->account = 0;
        }
 }
 
@@ -1471,8 +1427,10 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
        }
 
        if (bytes) {
-               bytes[0] += icd.bytes_done[0];
-               bytes[1] += icd.bytes_done[1];
+               int ddir;
+
+               for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+                       bytes[ddir] += icd.bytes_done[ddir];
        }
 
        return 0;
@@ -1509,8 +1467,10 @@ int io_u_queued_complete(struct thread_data *td, int min_evts,
        }
 
        if (bytes) {
-               bytes[0] += icd.bytes_done[0];
-               bytes[1] += icd.bytes_done[1];
+               int ddir;
+
+               for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+                       bytes[ddir] += icd.bytes_done[ddir];
        }
 
        return 0;
@@ -1541,8 +1501,11 @@ void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
                unsigned int perc = td->o.compress_percentage;
 
                if (perc) {
+                       unsigned int seg = min_write;
+
+                       seg = min(min_write, td->o.compress_chunk);
                        fill_random_buf_percentage(&td->buf_state, io_u->buf,
-                                               perc, min_write, max_bs);
+                                               perc, seg, max_bs);
                } else
                        fill_random_buf(&td->buf_state, io_u->buf, max_bs);
        } else