The whole point of having dedicated worker thread for each leaf VDEV was to

avoid calling zio_interrupt() from geom_up thread context. It turns out that
when provider is forcibly removed from the system and we kill worker thread
there can still be some ZIOs pending. To complete pending ZIOs when there is
no worker thread anymore we still have to call zio_interrupt() from geom_up
context. To avoid this race just remove use of worker threads altogether.
This should be more or less fine, because I also thought that zio_interrupt()
does more work, but it only makes small UMA allocation with M_WAITOK.
It also saves one context switch per I/O request.

PR:		kern/145339
Reported by:	Alex Bakhtin <Alex.Bakhtin@gmail.com>
MFC after:	1 week

1 files changed, 22 insertions, 105 deletions

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 3bd245e..b831a9c 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
@@ -47,31 +47,6 @@ struct g_class zfs_vdev_class = {
 
 DECLARE_GEOM_CLASS(zfs_vdev_class, zfs_vdev);
 
-typedef struct vdev_geom_ctx {
-	struct g_consumer *gc_consumer;
-	int gc_state;
-	struct bio_queue_head gc_queue;
-	struct mtx gc_queue_mtx;
-} vdev_geom_ctx_t;
-
-static void
-vdev_geom_release(vdev_t *vd)
-{
-	vdev_geom_ctx_t *ctx;
-
-	ctx = vd->vdev_tsd;
-	vd->vdev_tsd = NULL;
-
-	mtx_lock(&ctx->gc_queue_mtx);
-	ctx->gc_state = 1;
-	wakeup_one(&ctx->gc_queue);
-	while (ctx->gc_state != 2)
-		msleep(&ctx->gc_state, &ctx->gc_queue_mtx, 0, "vgeom:w", 0);
-	mtx_unlock(&ctx->gc_queue_mtx);
-	mtx_destroy(&ctx->gc_queue_mtx);
-	kmem_free(ctx, sizeof(*ctx));
-}
-
 static void
 vdev_geom_orphan(struct g_consumer *cp)
 {
@@ -96,8 +71,7 @@ vdev_geom_orphan(struct g_consumer *cp)
 		ZFS_LOG(1, "Destroyed geom %s.", gp->name);
 		g_wither_geom(gp, error);
 	}
-	vdev_geom_release(vd);
-
+	vd->vdev_tsd = NULL;
 	vd->vdev_remove_wanted = B_TRUE;
 	spa_async_request(vd->vdev_spa, SPA_ASYNC_REMOVE);
 }
@@ -188,52 +162,6 @@ vdev_geom_detach(void *arg, int flag __unused)
 	}
 }
 
-static void
-vdev_geom_worker(void *arg)
-{
-	vdev_geom_ctx_t *ctx;
-	zio_t *zio;
-	struct bio *bp;
-
-	thread_lock(curthread);
-	sched_prio(curthread, PRIBIO);
-	thread_unlock(curthread);
-
-	ctx = arg;
-	for (;;) {
-		mtx_lock(&ctx->gc_queue_mtx);
-		bp = bioq_takefirst(&ctx->gc_queue);
-		if (bp == NULL) {
-			if (ctx->gc_state == 1) {
-				ctx->gc_state = 2;
-				wakeup_one(&ctx->gc_state);
-				mtx_unlock(&ctx->gc_queue_mtx);
-				kthread_exit();
-			}
-			msleep(&ctx->gc_queue, &ctx->gc_queue_mtx,
-			    PRIBIO | PDROP, "vgeom:io", 0);
-			continue;
-		}
-		mtx_unlock(&ctx->gc_queue_mtx);
-		zio = bp->bio_caller1;
-		zio->io_error = bp->bio_error;
-		if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
-			vdev_t *vd;
-
-			/*
-			 * If we get ENOTSUP, we know that no future
-			 * attempts will ever succeed.  In this case we
-			 * set a persistent bit so that we don't bother
-			 * with the ioctl in the future.
-			 */
-			vd = zio->io_vd;
-			vd->vdev_nowritecache = B_TRUE;
-		}
-		g_destroy_bio(bp);
-		zio_interrupt(zio);
-	}
-}
-
 static uint64_t
 nvlist_get_guid(nvlist_t *list)
 {
@@ -488,7 +416,6 @@ vdev_geom_open_by_path(vdev_t *vd, int check_guid)
 static int
 vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *ashift)
 {
-	vdev_geom_ctx_t *ctx;
 	struct g_provider *pp;
 	struct g_consumer *cp;
 	int error, owned;
@@ -557,19 +484,9 @@ vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *ashift)
 	}
 
 	cp->private = vd;
-
-	ctx = kmem_zalloc(sizeof(*ctx), KM_SLEEP);
-	bioq_init(&ctx->gc_queue);
-	mtx_init(&ctx->gc_queue_mtx, "zfs:vdev:geom:queue", NULL, MTX_DEF);
-	ctx->gc_consumer = cp;
-	ctx->gc_state = 0;
-
-	vd->vdev_tsd = ctx;
+	vd->vdev_tsd = cp;
 	pp = cp->provider;
 
-	kproc_kthread_add(vdev_geom_worker, ctx, &zfsproc, NULL, 0, 0,
-	    "zfskern", "vdev %s", pp->name);
-
 	/*
 	 * Determine the actual size of the device.
 	 */
@@ -592,50 +509,49 @@ vdev_geom_open(vdev_t *vd, uint64_t *psize, uint64_t *ashift)
 static void
 vdev_geom_close(vdev_t *vd)
 {
-	vdev_geom_ctx_t *ctx;
 	struct g_consumer *cp;
 
-	if ((ctx = vd->vdev_tsd) == NULL)
-		return;
-	if ((cp = ctx->gc_consumer) == NULL)
+	cp = vd->vdev_tsd;
+	if (cp == NULL)
 		return;
-	vdev_geom_release(vd);
+	vd->vdev_tsd = NULL;
 	g_post_event(vdev_geom_detach, cp, M_WAITOK, NULL);
 }
 
 static void
 vdev_geom_io_intr(struct bio *bp)
 {
-	vdev_geom_ctx_t *ctx;
 	zio_t *zio;
 
 	zio = bp->bio_caller1;
-	ctx = zio->io_vd->vdev_tsd;
-
-	if ((zio->io_error = bp->bio_error) == 0 && bp->bio_resid != 0)
+	zio->io_error = bp->bio_error;
+	if (zio->io_error == 0 && bp->bio_resid != 0)
 		zio->io_error = EIO;
-
-	mtx_lock(&ctx->gc_queue_mtx);
-	bioq_insert_tail(&ctx->gc_queue, bp);
-	wakeup_one(&ctx->gc_queue);
-	mtx_unlock(&ctx->gc_queue_mtx);
+	if (bp->bio_cmd == BIO_FLUSH && bp->bio_error == ENOTSUP) {
+		vdev_t *vd;
+
+		/*
+		 * If we get ENOTSUP, we know that no future
+		 * attempts will ever succeed.  In this case we
+		 * set a persistent bit so that we don't bother
+		 * with the ioctl in the future.
+		 */
+		vd = zio->io_vd;
+		vd->vdev_nowritecache = B_TRUE;
+	}
+	g_destroy_bio(bp);
+	zio_interrupt(zio);
 }
 
 static int
 vdev_geom_io_start(zio_t *zio)
 {
 	vdev_t *vd;
-	vdev_geom_ctx_t *ctx;
 	struct g_consumer *cp;
 	struct bio *bp;
 	int error;
 
-	cp = NULL;
-
 	vd = zio->io_vd;
-	ctx = vd->vdev_tsd;
-	if (ctx != NULL)
-		cp = ctx->gc_consumer;
 
 	if (zio->io_type == ZIO_TYPE_IOCTL) {
 		/* XXPOLICY */
@@ -664,6 +580,7 @@ vdev_geom_io_start(zio_t *zio)
 		return (ZIO_PIPELINE_CONTINUE);
 	}
 sendreq:
+	cp = vd->vdev_tsd;
 	if (cp == NULL) {
 		zio->io_error = ENXIO;
 		return (ZIO_PIPELINE_CONTINUE);