diff --git a/sys/cam/ata/ata_da.c b/sys/cam/ata/ata_da.c index 396802656cdf..8a279311b4a3 100644 --- a/sys/cam/ata/ata_da.c +++ b/sys/cam/ata/ata_da.c @@ -874,7 +874,8 @@ adastart(struct cam_periph *periph, union ccb *start_ccb) } bioq_remove(&softc->bio_queue, bp); - if ((softc->flags & ADA_FLAG_NEED_OTAG) != 0) { + if ((bp->bio_flags & BIO_ORDERED) != 0 + || (softc->flags & ADA_FLAG_NEED_OTAG) != 0) { softc->flags &= ~ADA_FLAG_NEED_OTAG; softc->ordered_tag_count++; tag_code = 0; diff --git a/sys/cam/scsi/scsi_da.c b/sys/cam/scsi/scsi_da.c index 8f643611c0cc..7729ecc5ef54 100644 --- a/sys/cam/scsi/scsi_da.c +++ b/sys/cam/scsi/scsi_da.c @@ -1354,7 +1354,8 @@ dastart(struct cam_periph *periph, union ccb *start_ccb) bioq_remove(&softc->bio_queue, bp); - if ((softc->flags & DA_FLAG_NEED_OTAG) != 0) { + if ((bp->bio_flags & BIO_ORDERED) != 0 + || (softc->flags & DA_FLAG_NEED_OTAG) != 0) { softc->flags &= ~DA_FLAG_NEED_OTAG; softc->ordered_tag_count++; tag_code = MSG_ORDERED_Q_TAG; @@ -1368,7 +1369,8 @@ dastart(struct cam_periph *periph, union ccb *start_ccb) /*retries*/da_retry_count, /*cbfcnp*/dadone, /*tag_action*/tag_code, - /*read_op*/bp->bio_cmd == BIO_READ, + /*read_op*/bp->bio_cmd + == BIO_READ, /*byte2*/0, softc->minimum_cmd_size, /*lba*/bp->bio_pblkno, @@ -1377,17 +1379,24 @@ dastart(struct cam_periph *periph, union ccb *start_ccb) /*data_ptr*/ bp->bio_data, /*dxfer_len*/ bp->bio_bcount, /*sense_len*/SSD_FULL_SIZE, - /*timeout*/da_default_timeout*1000); + da_default_timeout * 1000); break; case BIO_FLUSH: + /* + * BIO_FLUSH doesn't currently communicate + * range data, so we synchronize the cache + * over the whole disk. We also force + * ordered tag semantics the flush applies + * to all previously queued I/O. + */ scsi_synchronize_cache(&start_ccb->csio, /*retries*/1, /*cbfcnp*/dadone, - MSG_SIMPLE_Q_TAG, - /*begin_lba*/0,/* Cover the whole disk */ + MSG_ORDERED_Q_TAG, + /*begin_lba*/0, /*lb_count*/0, SSD_FULL_SIZE, - /*timeout*/da_default_timeout*1000); + da_default_timeout*1000); break; } start_ccb->ccb_h.ccb_state = DA_CCB_BUFFER_IO; diff --git a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c index 00817bfef24b..d0d5137913c8 100644 --- a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c +++ b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_geom.c @@ -598,6 +598,7 @@ vdev_geom_io_start(zio_t *zio) break; case ZIO_TYPE_IOCTL: bp->bio_cmd = BIO_FLUSH; + bp->bio_flags |= BIO_ORDERED; bp->bio_data = NULL; bp->bio_offset = cp->provider->mediasize; bp->bio_length = 0; diff --git a/sys/geom/geom_io.c b/sys/geom/geom_io.c index b145441cad75..b4044a75e33c 100644 --- a/sys/geom/geom_io.c +++ b/sys/geom/geom_io.c @@ -265,6 +265,7 @@ g_io_flush(struct g_consumer *cp) g_trace(G_T_BIO, "bio_flush(%s)", cp->provider->name); bp = g_alloc_bio(); bp->bio_cmd = BIO_FLUSH; + bp->bio_flags |= BIO_ORDERED; bp->bio_done = NULL; bp->bio_attribute = NULL; bp->bio_offset = cp->provider->mediasize; diff --git a/sys/geom/sched/subr_disk.c b/sys/geom/sched/subr_disk.c index 008eaab1c94b..db2a9ef11f8c 100644 --- a/sys/geom/sched/subr_disk.c +++ b/sys/geom/sched/subr_disk.c @@ -86,7 +86,7 @@ __FBSDID("$FreeBSD$"); * bioq_remove() remove a generic element from the queue, act as * bioq_takefirst() if invoked on the head of the queue. * - * The semantic of these methods is the same of the operations + * The semantic of these methods is the same as the operations * on the underlying TAILQ, but with additional guarantees on * subsequent bioq_disksort() calls. E.g. bioq_insert_tail() * can be useful for making sure that all previous ops are flushed @@ -115,10 +115,10 @@ void gs_bioq_remove(struct bio_queue_head *head, struct bio *bp) { - if (bp == TAILQ_FIRST(&head->queue)) - head->last_offset = bp->bio_offset + bp->bio_length; - - if (bp == head->insert_point) + if (head->insert_point == NULL) { + if (bp == TAILQ_FIRST(&head->queue)) + head->last_offset = bp->bio_offset + bp->bio_length; + } else if (bp == head->insert_point) head->insert_point = NULL; TAILQ_REMOVE(&head->queue, bp, bio_queue); @@ -137,7 +137,8 @@ void gs_bioq_insert_head(struct bio_queue_head *head, struct bio *bp) { - head->last_offset = bp->bio_offset; + if (head->insert_point == NULL) + head->last_offset = bp->bio_offset; TAILQ_INSERT_HEAD(&head->queue, bp, bio_queue); } @@ -147,6 +148,7 @@ gs_bioq_insert_tail(struct bio_queue_head *head, struct bio *bp) TAILQ_INSERT_TAIL(&head->queue, bp, bio_queue); head->insert_point = bp; + head->last_offset = bp->bio_offset; } struct bio * @@ -189,13 +191,28 @@ gs_bioq_bio_key(struct bio_queue_head *head, struct bio *bp) void gs_bioq_disksort(struct bio_queue_head *head, struct bio *bp) { - struct bio *cur, *prev = NULL; - uoff_t key = gs_bioq_bio_key(head, bp); + struct bio *cur, *prev; + uoff_t key; + if ((bp->bio_flags & BIO_ORDERED) != 0) { + /* + * Ordered transactions can only be dispatched + * after any currently queued transactions. They + * also have barrier semantics - no transactions + * queued in the future can pass them. + */ + gs_bioq_insert_tail(head, bp); + return; + } + + prev = NULL; + key = gs_bioq_bio_key(head, bp); cur = TAILQ_FIRST(&head->queue); - if (head->insert_point) - cur = head->insert_point; + if (head->insert_point) { + prev = head->insert_point; + cur = TAILQ_NEXT(head->insert_point, bio_queue); + } while (cur != NULL && key >= gs_bioq_bio_key(head, cur)) { prev = cur; diff --git a/sys/kern/subr_disk.c b/sys/kern/subr_disk.c index 12b015944c2a..23915402e30d 100644 --- a/sys/kern/subr_disk.c +++ b/sys/kern/subr_disk.c @@ -127,7 +127,7 @@ disk_err(struct bio *bp, const char *what, int blkdone, int nl) * bioq_remove() remove a generic element from the queue, act as * bioq_takefirst() if invoked on the head of the queue. * - * The semantic of these methods is the same of the operations + * The semantic of these methods is the same as the operations * on the underlying TAILQ, but with additional guarantees on * subsequent bioq_disksort() calls. E.g. bioq_insert_tail() * can be useful for making sure that all previous ops are flushed @@ -156,10 +156,10 @@ void bioq_remove(struct bio_queue_head *head, struct bio *bp) { - if (bp == TAILQ_FIRST(&head->queue)) - head->last_offset = bp->bio_offset + bp->bio_length; - - if (bp == head->insert_point) + if (head->insert_point == NULL) { + if (bp == TAILQ_FIRST(&head->queue)) + head->last_offset = bp->bio_offset + bp->bio_length; + } else if (bp == head->insert_point) head->insert_point = NULL; TAILQ_REMOVE(&head->queue, bp, bio_queue); @@ -178,7 +178,8 @@ void bioq_insert_head(struct bio_queue_head *head, struct bio *bp) { - head->last_offset = bp->bio_offset; + if (head->insert_point == NULL) + head->last_offset = bp->bio_offset; TAILQ_INSERT_HEAD(&head->queue, bp, bio_queue); } @@ -188,6 +189,7 @@ bioq_insert_tail(struct bio_queue_head *head, struct bio *bp) TAILQ_INSERT_TAIL(&head->queue, bp, bio_queue); head->insert_point = bp; + head->last_offset = bp->bio_offset; } struct bio * @@ -230,13 +232,28 @@ bioq_bio_key(struct bio_queue_head *head, struct bio *bp) void bioq_disksort(struct bio_queue_head *head, struct bio *bp) { - struct bio *cur, *prev = NULL; - uoff_t key = bioq_bio_key(head, bp); + struct bio *cur, *prev; + uoff_t key; + if ((bp->bio_flags & BIO_ORDERED) != 0) { + /* + * Ordered transactions can only be dispatched + * after any currently queued transactions. They + * also have barrier semantics - no transactions + * queued in the future can pass them. + */ + bioq_insert_tail(head, bp); + return; + } + + prev = NULL; + key = bioq_bio_key(head, bp); cur = TAILQ_FIRST(&head->queue); - if (head->insert_point) - cur = head->insert_point; + if (head->insert_point) { + prev = head->insert_point; + cur = TAILQ_NEXT(head->insert_point, bio_queue); + } while (cur != NULL && key >= bioq_bio_key(head, cur)) { prev = cur; diff --git a/sys/sys/bio.h b/sys/sys/bio.h index 8b6ff0ec696a..c016ee67b904 100644 --- a/sys/sys/bio.h +++ b/sys/sys/bio.h @@ -54,6 +54,7 @@ #define BIO_ERROR 0x01 #define BIO_DONE 0x02 #define BIO_ONQUEUE 0x04 +#define BIO_ORDERED 0x08 #ifdef _KERNEL struct disk;