engines/libpmem: set file open/create mode always to RW
[fio.git] / engines / sg.c
index 7741f838be308f4949449e074dafcfb24ab5085b..0c2d2c8b861ad22fdc3080e51a605194450ffcc5 100644 (file)
 
 #ifdef FIO_HAVE_SGIO
 
+#ifndef SGV4_FLAG_HIPRI
+#define SGV4_FLAG_HIPRI 0x800
+#endif
+
 enum {
        FIO_SG_WRITE            = 1,
        FIO_SG_WRITE_VERIFY     = 2,
@@ -68,12 +72,22 @@ enum {
 
 struct sg_options {
        void *pad;
+       unsigned int hipri;
        unsigned int readfua;
        unsigned int writefua;
        unsigned int write_mode;
 };
 
 static struct fio_option options[] = {
+        {
+                .name   = "hipri",
+                .lname  = "High Priority",
+                .type   = FIO_OPT_STR_SET,
+                .off1   = offsetof(struct sg_options, hipri),
+                .help   = "Use polled IO completions",
+                .category = FIO_OPT_C_ENGINE,
+                .group  = FIO_OPT_G_SG,
+        },
        {
                .name   = "readfua",
                .lname  = "sg engine read fua flag support",
@@ -137,7 +151,7 @@ struct sgio_cmd {
 };
 
 struct sgio_trim {
-       char *unmap_param;
+       uint8_t *unmap_param;
        unsigned int unmap_range_count;
        struct io_u **trim_io_us;
 };
@@ -157,6 +171,37 @@ struct sgio_data {
 #endif
 };
 
+static inline uint32_t sgio_get_be32(uint8_t *buf)
+{
+       return be32_to_cpu(*((uint32_t *) buf));
+}
+
+static inline uint64_t sgio_get_be64(uint8_t *buf)
+{
+       return be64_to_cpu(*((uint64_t *) buf));
+}
+
+static inline void sgio_set_be16(uint16_t val, uint8_t *buf)
+{
+       uint16_t t = cpu_to_be16(val);
+
+       memcpy(buf, &t, sizeof(uint16_t));
+}
+
+static inline void sgio_set_be32(uint32_t val, uint8_t *buf)
+{
+       uint32_t t = cpu_to_be32(val);
+
+       memcpy(buf, &t, sizeof(uint32_t));
+}
+
+static inline void sgio_set_be64(uint64_t val, uint8_t *buf)
+{
+       uint64_t t = cpu_to_be64(val);
+
+       memcpy(buf, &t, sizeof(uint64_t));
+}
+
 static inline bool sgio_unbuffered(struct thread_data *td)
 {
        return (td->o.odirect || td->o.sync_io);
@@ -393,7 +438,8 @@ static enum fio_q_status fio_sgio_ioctl_doio(struct thread_data *td,
        return FIO_Q_COMPLETED;
 }
 
-static enum fio_q_status fio_sgio_rw_doio(struct fio_file *f,
+static enum fio_q_status fio_sgio_rw_doio(struct thread_data *td,
+                                         struct fio_file *f,
                                          struct io_u *io_u, int do_sync)
 {
        struct sg_io_hdr *hdr = &io_u->hdr;
@@ -404,13 +450,32 @@ static enum fio_q_status fio_sgio_rw_doio(struct fio_file *f,
                return ret;
 
        if (do_sync) {
-               ret = read(f->fd, hdr, sizeof(*hdr));
-               if (ret < 0)
-                       return ret;
+               /*
+                * We can't just read back the first command that completes
+                * and assume it's the one we need, it could be any command
+                * that is inflight.
+                */
+               do {
+                       struct io_u *__io_u;
 
-               /* record if an io error occurred */
-               if (hdr->info & SG_INFO_CHECK)
-                       io_u->error = EIO;
+                       ret = read(f->fd, hdr, sizeof(*hdr));
+                       if (ret < 0)
+                               return ret;
+
+                       __io_u = hdr->usr_ptr;
+
+                       /* record if an io error occurred */
+                       if (hdr->info & SG_INFO_CHECK)
+                               __io_u->error = EIO;
+
+                       if (__io_u == io_u)
+                               break;
+
+                       if (io_u_sync_complete(td, __io_u)) {
+                               ret = -1;
+                               break;
+                       }
+               } while (1);
 
                return FIO_Q_COMPLETED;
        }
@@ -426,10 +491,11 @@ static enum fio_q_status fio_sgio_doio(struct thread_data *td,
 
        if (f->filetype == FIO_TYPE_BLOCK) {
                ret = fio_sgio_ioctl_doio(td, f, io_u);
-               td_verror(td, io_u->error, __func__);
+               if (io_u->error)
+                       td_verror(td, io_u->error, __func__);
        } else {
-               ret = fio_sgio_rw_doio(f, io_u, do_sync);
-               if (do_sync)
+               ret = fio_sgio_rw_doio(td, f, io_u, do_sync);
+               if (io_u->error && do_sync)
                        td_verror(td, io_u->error, __func__);
        }
 
@@ -440,25 +506,11 @@ static void fio_sgio_rw_lba(struct sg_io_hdr *hdr, unsigned long long lba,
                            unsigned long long nr_blocks)
 {
        if (lba < MAX_10B_LBA) {
-               hdr->cmdp[2] = (unsigned char) ((lba >> 24) & 0xff);
-               hdr->cmdp[3] = (unsigned char) ((lba >> 16) & 0xff);
-               hdr->cmdp[4] = (unsigned char) ((lba >>  8) & 0xff);
-               hdr->cmdp[5] = (unsigned char) (lba & 0xff);
-               hdr->cmdp[7] = (unsigned char) ((nr_blocks >> 8) & 0xff);
-               hdr->cmdp[8] = (unsigned char) (nr_blocks & 0xff);
+               sgio_set_be32((uint32_t) lba, &hdr->cmdp[2]);
+               sgio_set_be16((uint16_t) nr_blocks, &hdr->cmdp[7]);
        } else {
-               hdr->cmdp[2] = (unsigned char) ((lba >> 56) & 0xff);
-               hdr->cmdp[3] = (unsigned char) ((lba >> 48) & 0xff);
-               hdr->cmdp[4] = (unsigned char) ((lba >> 40) & 0xff);
-               hdr->cmdp[5] = (unsigned char) ((lba >> 32) & 0xff);
-               hdr->cmdp[6] = (unsigned char) ((lba >> 24) & 0xff);
-               hdr->cmdp[7] = (unsigned char) ((lba >> 16) & 0xff);
-               hdr->cmdp[8] = (unsigned char) ((lba >>  8) & 0xff);
-               hdr->cmdp[9] = (unsigned char) (lba & 0xff);
-               hdr->cmdp[10] = (unsigned char) ((nr_blocks >> 32) & 0xff);
-               hdr->cmdp[11] = (unsigned char) ((nr_blocks >> 16) & 0xff);
-               hdr->cmdp[12] = (unsigned char) ((nr_blocks >> 8) & 0xff);
-               hdr->cmdp[13] = (unsigned char) (nr_blocks & 0xff);
+               sgio_set_be64(lba, &hdr->cmdp[2]);
+               sgio_set_be32((uint32_t) nr_blocks, &hdr->cmdp[10]);
        }
 
        return;
@@ -489,6 +541,8 @@ static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
                else
                        hdr->cmdp[0] = 0x88; // read(16)
 
+               if (o->hipri)
+                       hdr->flags |= SGV4_FLAG_HIPRI;
                if (o->readfua)
                        hdr->cmdp[1] |= 0x08;
 
@@ -504,6 +558,8 @@ static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
                                hdr->cmdp[0] = 0x2a; // write(10)
                        else
                                hdr->cmdp[0] = 0x8a; // write(16)
+                       if (o->hipri)
+                               hdr->flags |= SGV4_FLAG_HIPRI;
                        if (o->writefua)
                                hdr->cmdp[1] |= 0x08;
                        break;
@@ -552,18 +608,8 @@ static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
 #endif
 
                offset = 8 + 16 * st->unmap_range_count;
-               st->unmap_param[offset] = (unsigned char) ((lba >> 56) & 0xff);
-               st->unmap_param[offset+1] = (unsigned char) ((lba >> 48) & 0xff);
-               st->unmap_param[offset+2] = (unsigned char) ((lba >> 40) & 0xff);
-               st->unmap_param[offset+3] = (unsigned char) ((lba >> 32) & 0xff);
-               st->unmap_param[offset+4] = (unsigned char) ((lba >> 24) & 0xff);
-               st->unmap_param[offset+5] = (unsigned char) ((lba >> 16) & 0xff);
-               st->unmap_param[offset+6] = (unsigned char) ((lba >>  8) & 0xff);
-               st->unmap_param[offset+7] = (unsigned char) (lba & 0xff);
-               st->unmap_param[offset+8] = (unsigned char) ((nr_blocks >> 32) & 0xff);
-               st->unmap_param[offset+9] = (unsigned char) ((nr_blocks >> 16) & 0xff);
-               st->unmap_param[offset+10] = (unsigned char) ((nr_blocks >> 8) & 0xff);
-               st->unmap_param[offset+11] = (unsigned char) (nr_blocks & 0xff);
+               sgio_set_be64(lba, &st->unmap_param[offset]);
+               sgio_set_be32((uint32_t) nr_blocks, &st->unmap_param[offset + 8]);
 
                st->unmap_range_count++;
 
@@ -582,14 +628,12 @@ static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
 
 static void fio_sgio_unmap_setup(struct sg_io_hdr *hdr, struct sgio_trim *st)
 {
-       hdr->dxfer_len = st->unmap_range_count * 16 + 8;
-       hdr->cmdp[7] = (unsigned char) (((st->unmap_range_count * 16 + 8) >> 8) & 0xff);
-       hdr->cmdp[8] = (unsigned char) ((st->unmap_range_count * 16 + 8) & 0xff);
+       uint16_t cnt = st->unmap_range_count * 16;
 
-       st->unmap_param[0] = (unsigned char) (((16 * st->unmap_range_count + 6) >> 8) & 0xff);
-       st->unmap_param[1] = (unsigned char)  ((16 * st->unmap_range_count + 6) & 0xff);
-       st->unmap_param[2] = (unsigned char) (((16 * st->unmap_range_count) >> 8) & 0xff);
-       st->unmap_param[3] = (unsigned char)  ((16 * st->unmap_range_count) & 0xff);
+       hdr->dxfer_len = cnt + 8;
+       sgio_set_be16(cnt + 8, &hdr->cmdp[7]);
+       sgio_set_be16(cnt + 6, st->unmap_param);
+       sgio_set_be16(cnt, &st->unmap_param[2]);
 
        return;
 }
@@ -673,38 +717,39 @@ static int fio_sgio_commit(struct thread_data *td)
 
        sd->current_queue = -1;
 
-       ret = fio_sgio_rw_doio(io_u->file, io_u, 0);
+       ret = fio_sgio_rw_doio(td, io_u->file, io_u, 0);
 
-       if (ret < 0)
-               for (i = 0; i < st->unmap_range_count; i++)
-                       st->trim_io_us[i]->error = errno;
-       else if (hdr->status)
-               for (i = 0; i < st->unmap_range_count; i++) {
-                       st->trim_io_us[i]->resid = hdr->resid;
-                       st->trim_io_us[i]->error = EIO;
+       if (ret < 0 || hdr->status) {
+               int error;
+
+               if (ret < 0)
+                       error = errno;
+               else {
+                       error = EIO;
+                       ret = -EIO;
                }
-       else {
-               if (fio_fill_issue_time(td)) {
-                       fio_gettime(&now, NULL);
-                       for (i = 0; i < st->unmap_range_count; i++) {
-                               struct io_u *io_u = st->trim_io_us[i];
-
-                               memcpy(&io_u->issue_time, &now, sizeof(now));
-                               io_u_queued(td, io_u);
-                       }
+
+               for (i = 0; i < st->unmap_range_count; i++) {
+                       st->trim_io_us[i]->error = error;
+                       clear_io_u(td, st->trim_io_us[i]);
+                       if (hdr->status)
+                               st->trim_io_us[i]->resid = hdr->resid;
                }
-               io_u_mark_submit(td, st->unmap_range_count);
+
+               td_verror(td, error, "xfer");
+               return ret;
        }
 
-       if (io_u->error) {
-               td_verror(td, io_u->error, "xfer");
-               return 0;
+       if (fio_fill_issue_time(td)) {
+               fio_gettime(&now, NULL);
+               for (i = 0; i < st->unmap_range_count; i++) {
+                       memcpy(&st->trim_io_us[i]->issue_time, &now, sizeof(now));
+                       io_u_queued(td, io_u);
+               }
        }
+       io_u_mark_submit(td, st->unmap_range_count);
 
-       if (ret == FIO_Q_QUEUED)
-               return 0;
-       else
-               return ret;
+       return 0;
 }
 
 static struct io_u *fio_sgio_event(struct thread_data *td, int event)
@@ -722,6 +767,8 @@ static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
         * io_u structures, which are not initialized until later.
         */
        struct sg_io_hdr hdr;
+       unsigned long long hlba;
+       unsigned int blksz = 0;
        unsigned char cmd[16];
        unsigned char sb[64];
        unsigned char buf[32];  // read capacity return
@@ -758,23 +805,23 @@ static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
                return ret;
        }
 
-       *bs      = ((unsigned long) buf[4] << 24) | ((unsigned long) buf[5] << 16) |
-                  ((unsigned long) buf[6] << 8) | (unsigned long) buf[7];
-       *max_lba = ((unsigned long) buf[0] << 24) | ((unsigned long) buf[1] << 16) |
-                  ((unsigned long) buf[2] << 8) | (unsigned long) buf[3];
+       if (hdr.info & SG_INFO_CHECK) {
+               /* RCAP(10) might be unsupported by device. Force RCAP(16) */
+               hlba = MAX_10B_LBA;
+       } else {
+               blksz = sgio_get_be32(&buf[4]);
+               hlba = sgio_get_be32(buf);
+       }
 
        /*
         * If max lba masked by MAX_10B_LBA equals MAX_10B_LBA,
         * then need to retry with 16 byte Read Capacity command.
         */
-       if (*max_lba == MAX_10B_LBA) {
+       if (hlba == MAX_10B_LBA) {
                hdr.cmd_len = 16;
                hdr.cmdp[0] = 0x9e; // service action
                hdr.cmdp[1] = 0x10; // Read Capacity(16)
-               hdr.cmdp[10] = (unsigned char) ((sizeof(buf) >> 24) & 0xff);
-               hdr.cmdp[11] = (unsigned char) ((sizeof(buf) >> 16) & 0xff);
-               hdr.cmdp[12] = (unsigned char) ((sizeof(buf) >> 8) & 0xff);
-               hdr.cmdp[13] = (unsigned char) (sizeof(buf) & 0xff);
+               sgio_set_be32(sizeof(buf), &hdr.cmdp[10]);
 
                hdr.dxfer_direction = SG_DXFER_FROM_DEV;
                hdr.dxferp = buf;
@@ -790,19 +837,20 @@ static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
                if (hdr.info & SG_INFO_CHECK)
                        td_verror(td, EIO, "fio_sgio_read_capacity");
 
-               *bs = (buf[8] << 24) | (buf[9] << 16) | (buf[10] << 8) | buf[11];
-               *max_lba = ((unsigned long long)buf[0] << 56) |
-                               ((unsigned long long)buf[1] << 48) |
-                               ((unsigned long long)buf[2] << 40) |
-                               ((unsigned long long)buf[3] << 32) |
-                               ((unsigned long long)buf[4] << 24) |
-                               ((unsigned long long)buf[5] << 16) |
-                               ((unsigned long long)buf[6] << 8) |
-                               (unsigned long long)buf[7];
+               blksz = sgio_get_be32(&buf[8]);
+               hlba = sgio_get_be64(buf);
+       }
+
+       if (blksz) {
+               *bs = blksz;
+               *max_lba = hlba;
+               ret = 0;
+       } else {
+               ret = EIO;
        }
 
        close(fd);
-       return 0;
+       return ret;
 }
 
 static void fio_sgio_cleanup(struct thread_data *td)
@@ -835,6 +883,7 @@ static int fio_sgio_init(struct thread_data *td)
 {
        struct sgio_data *sd;
        struct sgio_trim *st;
+       struct sg_io_hdr *h3p;
        int i;
 
        sd = calloc(1, sizeof(*sd));
@@ -850,12 +899,13 @@ static int fio_sgio_init(struct thread_data *td)
 #ifdef FIO_SGIO_DEBUG
        sd->trim_queue_map = calloc(td->o.iodepth, sizeof(int));
 #endif
-       for (i = 0; i < td->o.iodepth; i++) {
+       for (i = 0, h3p = sd->sgbuf; i < td->o.iodepth; i++, ++h3p) {
                sd->trim_queues[i] = calloc(1, sizeof(struct sgio_trim));
                st = sd->trim_queues[i];
                st->unmap_param = calloc(td->o.iodepth + 1, sizeof(char[16]));
                st->unmap_range_count = 0;
                st->trim_io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
+               h3p->interface_id = 'S';
        }
 
        td->io_ops_data = sd;
@@ -1151,8 +1201,8 @@ static char *fio_sgio_errdetails(struct io_u *io_u)
        }
 
        if (!(hdr->info & SG_INFO_CHECK) && !strlen(msg))
-               strncpy(msg, "SG Driver did not report a Host, Driver or Device check",
-                       MAXERRDETAIL - 1);
+               snprintf(msg, MAXERRDETAIL, "%s",
+                        "SG Driver did not report a Host, Driver or Device check");
 
        return msg;
 }