From 96bafea5d19611688dc0491f4a327272731b39cd Mon Sep 17 00:00:00 2001 From: Taylor R Campbell Date: Sat, 29 Jan 2022 18:13:17 +0000 Subject: [PATCH 32/37] specfs: Let spec_node_lookup_by_dev wait for reclaim to finish. vdevgone relies on this to ensure that if there is a concurrent revoke in progress, it will wait for that revoke to finish -- that way, it can guarantee all I/O operations have completed and the device is closed. --- sys/coda/coda_vfsops.c | 2 +- sys/kern/vfs_mount.c | 3 ++- sys/kern/vfs_subr.c | 9 +++++++-- sys/kern/vfs_vnode.c | 3 ++- sys/miscfs/specfs/spec_vnops.c | 21 +++++++++++++++++++-- sys/miscfs/specfs/specdev.h | 2 +- 6 files changed, 32 insertions(+), 8 deletions(-) diff --git a/sys/coda/coda_vfsops.c b/sys/coda/coda_vfsops.c index 9e25704b6b2f..c0a1b98a9331 100644 --- a/sys/coda/coda_vfsops.c +++ b/sys/coda/coda_vfsops.c @@ -636,7 +636,7 @@ struct mount *devtomp(dev_t dev) struct mount *mp; struct vnode *vp; - if (spec_node_lookup_by_dev(VBLK, dev, &vp) == 0) { + if (spec_node_lookup_by_dev(VBLK, dev, VDEAD_NOWAIT, &vp) == 0) { mp = spec_node_getmountedfs(vp); vrele(vp); } else { diff --git a/sys/kern/vfs_mount.c b/sys/kern/vfs_mount.c index 2d4d6c28c292..58d3dff41d57 100644 --- a/sys/kern/vfs_mount.c +++ b/sys/kern/vfs_mount.c @@ -1372,7 +1372,8 @@ vfs_mountedon(vnode_t *vp) return ENOTBLK; if (spec_node_getmountedfs(vp) != NULL) return EBUSY; - if (spec_node_lookup_by_dev(vp->v_type, vp->v_rdev, &vq) == 0) { + if (spec_node_lookup_by_dev(vp->v_type, vp->v_rdev, VDEAD_NOWAIT, &vq) + == 0) { if (spec_node_getmountedfs(vq) != NULL) error = EBUSY; vrele(vq); diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c index 10cb4e57e129..a64cffb97054 100644 --- a/sys/kern/vfs_subr.c +++ b/sys/kern/vfs_subr.c @@ -495,7 +495,7 @@ int vfinddev(dev_t dev, enum vtype type, vnode_t **vpp) { - return (spec_node_lookup_by_dev(type, dev, vpp) == 0); + return (spec_node_lookup_by_dev(type, dev, VDEAD_NOWAIT, vpp) == 0); } /* @@ -511,7 +511,12 @@ vdevgone(int maj, int minl, int minh, enum vtype type) for (mn = minl; mn <= minh; mn++) { dev = makedev(maj, mn); - while (spec_node_lookup_by_dev(type, dev, &vp) == 0) { + /* + * Passing 0 as flags, instead of VDEAD_NOWAIT, means + * spec_node_lookup_by_dev will wait for vnodes it + * finds concurrently being revoked before returning. + */ + while (spec_node_lookup_by_dev(type, dev, 0, &vp) == 0) { VOP_REVOKE(vp, REVOKEALL); vrele(vp); } diff --git a/sys/kern/vfs_vnode.c b/sys/kern/vfs_vnode.c index 48065852a6ac..58db8e0e97e7 100644 --- a/sys/kern/vfs_vnode.c +++ b/sys/kern/vfs_vnode.c @@ -1234,7 +1234,8 @@ vrevoke(vnode_t *vp) type = vp->v_type; mutex_exit(vp->v_interlock); - while (spec_node_lookup_by_dev(type, dev, &vq) == 0) { + while (spec_node_lookup_by_dev(type, dev, VDEAD_NOWAIT, &vq) + == 0) { mp = vrevoke_suspend_next(mp, vq->v_mount); vgone(vq); } diff --git a/sys/miscfs/specfs/spec_vnops.c b/sys/miscfs/specfs/spec_vnops.c index 34b47f1ae1d4..3dce4cc7a83f 100644 --- a/sys/miscfs/specfs/spec_vnops.c +++ b/sys/miscfs/specfs/spec_vnops.c @@ -424,18 +424,35 @@ spec_node_init(vnode_t *vp, dev_t rdev) * Lookup a vnode by device number and return it referenced. */ int -spec_node_lookup_by_dev(enum vtype type, dev_t dev, vnode_t **vpp) +spec_node_lookup_by_dev(enum vtype type, dev_t dev, int flags, vnode_t **vpp) { int error; vnode_t *vp; - mutex_enter(&device_lock); +top: mutex_enter(&device_lock); for (vp = specfs_hash[SPECHASH(dev)]; vp; vp = vp->v_specnext) { if (type == vp->v_type && dev == vp->v_rdev) { mutex_enter(vp->v_interlock); /* If clean or being cleaned, then ignore it. */ if (vdead_check(vp, VDEAD_NOWAIT) == 0) break; + if ((flags & VDEAD_NOWAIT) == 0) { + mutex_exit(&device_lock); + /* + * It may be being revoked as we speak, + * and the caller wants to wait until + * all revocation has completed. Let + * vcache_vget wait for it to finish + * dying; as a side effect, vcache_vget + * releases vp->v_interlock. Note that + * vcache_vget cannot succeed at this + * point because vdead_check already + * failed. + */ + error = vcache_vget(vp); + KASSERT(error); + goto top; + } mutex_exit(vp->v_interlock); } } diff --git a/sys/miscfs/specfs/specdev.h b/sys/miscfs/specfs/specdev.h index 8b44cf4339f4..c095cf24d2d5 100644 --- a/sys/miscfs/specfs/specdev.h +++ b/sys/miscfs/specfs/specdev.h @@ -95,7 +95,7 @@ typedef struct specdev { */ void spec_node_init(vnode_t *, dev_t); void spec_node_destroy(vnode_t *); -int spec_node_lookup_by_dev(enum vtype, dev_t, vnode_t **); +int spec_node_lookup_by_dev(enum vtype, dev_t, int, vnode_t **); int spec_node_lookup_by_mount(struct mount *, vnode_t **); struct mount *spec_node_getmountedfs(vnode_t *); void spec_node_setmountedfs(vnode_t *, struct mount *);