FreeBSD: Add support for lockless lookup

Authored-by: mjg <mjg@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10657

include/os/freebsd/spl/sys/ccompat.h

@@ -55,6 +55,21 @@
 #else
 #define	getnewvnode_reserve_()	getnewvnode_reserve(1)
 #endif
+
+#if  __FreeBSD_version < 1300102
+#define	ASSERT_VOP_IN_SEQC(zp)
+#define	MNTK_FPLOOKUP 0
+#define	vn_seqc_write_begin(vp)
+#define	vn_seqc_write_end(vp)
+
+#ifndef VFS_SMR_DECLARE
+#define	VFS_SMR_DECLARE
+#endif
+#ifndef VFS_SMR_ZONE_SET
+#define	VFS_SMR_ZONE_SET(zone)
+#endif
+#endif
+
 struct hlist_node {
 	struct hlist_node *next, **pprev;
 };

include/os/freebsd/zfs/sys/zfs_znode_impl.h

@@ -99,6 +99,7 @@ extern minor_t zfsdev_minor_alloc(void);
 #define	ZTOV(ZP)	((ZP)->z_vnode)
 #define	ZTOI(ZP)	((ZP)->z_vnode)
 #define	VTOZ(VP)	((struct znode *)(VP)->v_data)
+#define	VTOZ_SMR(VP)	((znode_t *)vn_load_v_data_smr(VP))
 #define	ITOZ(VP)	((struct znode *)(VP)->v_data)
 #define	zhold(zp)	vhold(ZTOV((zp)))
 #define	zrele(zp)	vrele(ZTOV((zp)))

module/os/freebsd/spl/spl_vfs.c

@@ -161,6 +161,7 @@ mount_snapshot(kthread_t *td, vnode_t **vpp, const char *fstype, char *fspath,
 		vput(vp);
 		return (error);
 	}
+	vn_seqc_write_begin(vp);
 	VOP_UNLOCK1(vp);
 
 	/*
@@ -213,6 +214,7 @@ mount_snapshot(kthread_t *td, vnode_t **vpp, const char *fstype, char *fspath,
 		VI_LOCK(vp);
 		vp->v_iflag &= ~VI_MOUNT;
 		VI_UNLOCK(vp);
+		vn_seqc_write_end(vp);
 		vput(vp);
 		vfs_unbusy(mp);
 		vfs_freeopts(mp->mnt_optnew);
@@ -248,6 +250,7 @@ mount_snapshot(kthread_t *td, vnode_t **vpp, const char *fstype, char *fspath,
 	vfs_event_signal(NULL, VQ_MOUNT, 0);
 	if (VFS_ROOT(mp, LK_EXCLUSIVE, &mvp))
 		panic("mount: lost mount");
+	vn_seqc_write_end(vp);
 	VOP_UNLOCK1(vp);
 #if __FreeBSD_version >= 1300048
 	vfs_op_exit(mp);

module/os/freebsd/zfs/zfs_acl.c

@@ -1144,6 +1144,7 @@ zfs_acl_chown_setattr(znode_t *zp)
 	if (zp->z_zfsvfs->z_replay == B_FALSE)
 		ASSERT_VOP_ELOCKED(ZTOV(zp), __func__);
 	ASSERT(MUTEX_HELD(&zp->z_acl_lock));
+	ASSERT_VOP_IN_SEQC(ZTOV(zp));
 
 	if ((error = zfs_acl_node_read(zp, B_TRUE, &aclp, B_FALSE)) == 0)
 		zp->z_mode = zfs_mode_compute(zp->z_mode, aclp,
@@ -1171,6 +1172,8 @@ zfs_aclset_common(znode_t *zp, zfs_acl_t *aclp, cred_t *cr, dmu_tx_t *tx)
 	int			count = 0;
 	zfs_acl_phys_t		acl_phys;
 
+	ASSERT_VOP_IN_SEQC(ZTOV(zp));
+
 	mode = zp->z_mode;
 
 	mode = zfs_mode_compute(mode, aclp, &zp->z_pflags,

module/os/freebsd/zfs/zfs_vfsops.c

@@ -1192,6 +1192,9 @@ zfs_domount(vfs_t *vfsp, char *osname)
 	vfsp->mnt_kern_flag |= MNTK_NOMSYNC;
 	vfsp->mnt_kern_flag |= MNTK_VMSETSIZE_BUG;
 
+#if defined(_KERNEL) && !defined(KMEM_DEBUG)
+	vfsp->mnt_kern_flag |= MNTK_FPLOOKUP;
+#endif
 	/*
 	 * The fsid is 64 bits, composed of an 8-bit fs type, which
 	 * separates our fsid from any other filesystem types, and a

module/os/freebsd/zfs/zfs_vnops.c

@@ -39,6 +39,9 @@
 #include <sys/endian.h>
 #include <sys/vm.h>
 #include <sys/vnode.h>
+#if __FreeBSD_version >= 1300102
+#include <sys/smr.h>
+#endif
 #include <sys/dirent.h>
 #include <sys/file.h>
 #include <sys/stat.h>
@@ -91,6 +94,8 @@
 #define	VN_OPEN_INVFS	0x0
 #endif
 
+VFS_SMR_DECLARE;
+
 #if __FreeBSD_version >= 1300047
 #define	vm_page_wire_lock(pp)
 #define	vm_page_wire_unlock(pp)
@@ -3932,6 +3937,7 @@ zfs_rename_(vnode_t *sdvp, vnode_t **svpp, struct componentname *scnp,
 	char		*snm = scnp->cn_nameptr;
 	char		*tnm = tcnp->cn_nameptr;
 	int		error = 0;
+	bool	want_seqc_end __maybe_unused = false;
 
 	/* Reject renames across filesystems. */
 	if ((*svpp)->v_mount != tdvp->v_mount ||
@@ -4075,6 +4081,15 @@ zfs_rename_(vnode_t *sdvp, vnode_t **svpp, struct componentname *scnp,
 		}
 	}
 
+	vn_seqc_write_begin(*svpp);
+	vn_seqc_write_begin(sdvp);
+	if (*tvpp != NULL)
+		vn_seqc_write_begin(*tvpp);
+	if (tdvp != *tvpp)
+		vn_seqc_write_begin(tdvp);
+#if	__FreeBSD_version >= 1300102
+	want_seqc_end = true;
+#endif
 	vnevent_rename_src(*svpp, sdvp, scnp->cn_nameptr, ct);
 	if (tzp)
 		vnevent_rename_dest(*tvpp, tdvp, tnm, ct);
@@ -4161,10 +4176,20 @@ zfs_rename_(vnode_t *sdvp, vnode_t **svpp, struct componentname *scnp,
 
 unlockout:			/* all 4 vnodes are locked, ZFS_ENTER called */
 	ZFS_EXIT(zfsvfs);
+	if (want_seqc_end) {
+		vn_seqc_write_end(*svpp);
+		vn_seqc_write_end(sdvp);
+		if (*tvpp != NULL)
+			vn_seqc_write_end(*tvpp);
+		if (tdvp != *tvpp)
+			vn_seqc_write_end(tdvp);
+		want_seqc_end = false;
+	}
 	VOP_UNLOCK1(*svpp);
 	VOP_UNLOCK1(sdvp);
 
 out:				/* original two vnodes are locked */
+	MPASS(!want_seqc_end);
 	if (error == 0 && zfsvfs->z_os->os_sync == ZFS_SYNC_ALWAYS)
 		zil_commit(zilog, 0);
 
@@ -5196,6 +5221,33 @@ zfs_freebsd_write(struct vop_write_args *ap)
 	    ap->a_cred));
 }
 
+#if __FreeBSD_version >= 1300102
+/*
+ * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
+ * the comment above cache_fplookup for details.
+ */
+static int
+zfs_freebsd_fplookup_vexec(struct vop_fplookup_vexec_args *v)
+{
+	vnode_t *vp;
+	znode_t *zp;
+	uint64_t pflags;
+
+	vp = v->a_vp;
+	zp = VTOZ_SMR(vp);
+	if (__predict_false(zp == NULL))
+		return (EAGAIN);
+	pflags = atomic_load_64(&zp->z_pflags);
+	if (pflags & ZFS_AV_QUARANTINED)
+		return (EAGAIN);
+	if (pflags & ZFS_XATTR)
+		return (EAGAIN);
+	if ((pflags & ZFS_NO_EXECS_DENIED) == 0)
+		return (EAGAIN);
+	return (0);
+}
+#endif
+
 #ifndef _SYS_SYSPROTO_H_
 struct vop_access_args {
 	struct vnode *a_vp;
@@ -6483,6 +6535,9 @@ struct vop_vector zfs_vnodeops = {
 	.vop_need_inactive =	zfs_freebsd_need_inactive,
 #endif
 	.vop_reclaim =		zfs_freebsd_reclaim,
+#if __FreeBSD_version >= 1300102
+	.vop_fplookup_vexec = zfs_freebsd_fplookup_vexec,
+#endif
 	.vop_access =		zfs_freebsd_access,
 	.vop_allocate =		VOP_EINVAL,
 	.vop_lookup =		zfs_cache_lookup,
@@ -6537,6 +6592,9 @@ VFS_VOP_VECTOR_REGISTER(zfs_vnodeops);
 struct vop_vector zfs_fifoops = {
 	.vop_default =		&fifo_specops,
 	.vop_fsync =		zfs_freebsd_fsync,
+#if __FreeBSD_version >= 1300102
+	.vop_fplookup_vexec = zfs_freebsd_fplookup_vexec,
+#endif
 	.vop_access =		zfs_freebsd_access,
 	.vop_getattr =		zfs_freebsd_getattr,
 	.vop_inactive =		zfs_freebsd_inactive,

module/os/freebsd/zfs/zfs_znode.c

@@ -97,7 +97,13 @@ SYSCTL_INT(_debug_sizeof, OID_AUTO, znode, CTLFLAG_RD,
  */
 krwlock_t zfsvfs_lock;
 
+#if defined(_KERNEL) && !defined(KMEM_DEBUG) && \
+    __FreeBSD_version >= 1300102
+#define	_ZFS_USE_SMR
+static uma_zone_t znode_uma_zone;
+#else
 static kmem_cache_t *znode_cache = NULL;
+#endif
 
 extern struct vop_vector zfs_vnodeops;
 extern struct vop_vector zfs_fifoops;
@@ -169,6 +175,53 @@ zfs_znode_cache_destructor(void *buf, void *arg)
 	ASSERT(zp->z_acl_cached == NULL);
 }
 
+
+#ifdef _ZFS_USE_SMR
+VFS_SMR_DECLARE;
+
+static int
+zfs_znode_cache_constructor_smr(void *mem, int size __unused, void *private,
+    int flags)
+{
+
+	return (zfs_znode_cache_constructor(mem, private, flags));
+}
+
+static void
+zfs_znode_cache_destructor_smr(void *mem, int size __unused, void *private)
+{
+
+	zfs_znode_cache_destructor(mem, private);
+}
+
+void
+zfs_znode_init(void)
+{
+	/*
+	 * Initialize zcache
+	 */
+	rw_init(&zfsvfs_lock, NULL, RW_DEFAULT, NULL);
+	ASSERT(znode_uma_zone == NULL);
+	znode_uma_zone = uma_zcreate("zfs_znode_cache",
+	    sizeof (znode_t), zfs_znode_cache_constructor_smr,
+	    zfs_znode_cache_destructor_smr, NULL, NULL, 0, 0);
+	VFS_SMR_ZONE_SET(znode_uma_zone);
+}
+
+static znode_t *
+zfs_znode_alloc_kmem(int flags)
+{
+
+	return (uma_zalloc_smr(znode_uma_zone, flags));
+}
+
+static void
+zfs_znode_free_kmem(znode_t *zp)
+{
+
+	uma_zfree_smr(znode_uma_zone, zp);
+}
+#else
 void
 zfs_znode_init(void)
 {
@@ -180,18 +233,40 @@ zfs_znode_init(void)
 	znode_cache = kmem_cache_create("zfs_znode_cache",
 	    sizeof (znode_t), 0, zfs_znode_cache_constructor,
 	    zfs_znode_cache_destructor, NULL, NULL, NULL, 0);
-	// kmem_cache_set_move(znode_cache, zfs_znode_move);
 }
 
+static znode_t *
+zfs_znode_alloc_kmem(int flags)
+{
+
+	return (kmem_cache_alloc(znode_cache, flags));
+}
+
+static void
+zfs_znode_free_kmem(znode_t *zp)
+{
+
+	kmem_cache_free(znode_cache, zp);
+}
+#endif
+
 void
 zfs_znode_fini(void)
 {
 	/*
 	 * Cleanup zcache
 	 */
-	if (znode_cache)
+#ifdef _ZFS_USE_SMR
+	if (znode_uma_zone) {
+		uma_zdestroy(znode_uma_zone);
+		znode_uma_zone = NULL;
+	}
+#else
+	if (znode_cache) {
 		kmem_cache_destroy(znode_cache);
-	znode_cache = NULL;
+		znode_cache = NULL;
+	}
+#endif
 	rw_destroy(&zfsvfs_lock);
 }
 
@@ -211,7 +286,7 @@ zfs_create_share_dir(zfsvfs_t *zfsvfs, dmu_tx_t *tx)
 	vattr.va_uid = crgetuid(kcred);
 	vattr.va_gid = crgetgid(kcred);
 
-	sharezp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+	sharezp = zfs_znode_alloc_kmem(KM_SLEEP);
 	ASSERT(!POINTER_IS_VALID(sharezp->z_zfsvfs));
 	sharezp->z_moved = 0;
 	sharezp->z_unlinked = 0;
@@ -230,7 +305,7 @@ zfs_create_share_dir(zfsvfs_t *zfsvfs, dmu_tx_t *tx)
 
 	zfs_acl_ids_free(&acl_ids);
 	sa_handle_destroy(sharezp->z_sa_hdl);
-	kmem_cache_free(znode_cache, sharezp);
+	zfs_znode_free_kmem(sharezp);
 
 	return (error);
 }
@@ -349,7 +424,12 @@ zfs_znode_alloc(zfsvfs_t *zfsvfs, dmu_buf_t *db, int blksz,
 	int count = 0;
 	int error;
 
-	zp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+	zp = zfs_znode_alloc_kmem(KM_SLEEP);
+
+#ifndef _ZFS_USE_SMR
+	KASSERT((zfsvfs->z_parent->z_vfs->mnt_kern_flag & MNTK_FPLOOKUP) == 0,
+	    ("%s: fast path lookup enabled without smr", __func__));
+#endif
 
 #if __FreeBSD_version >= 1300076
 	KASSERT(curthread->td_vp_reserved != NULL,
@@ -360,7 +440,7 @@ zfs_znode_alloc(zfsvfs_t *zfsvfs, dmu_buf_t *db, int blksz,
 #endif
 	error = getnewvnode("zfs", zfsvfs->z_parent->z_vfs, &zfs_vnodeops, &vp);
 	if (error != 0) {
-		kmem_cache_free(znode_cache, zp);
+		zfs_znode_free_kmem(zp);
 		return (NULL);
 	}
 	zp->z_vnode = vp;
@@ -416,7 +496,7 @@ zfs_znode_alloc(zfsvfs_t *zfsvfs, dmu_buf_t *db, int blksz,
 			sa_handle_destroy(zp->z_sa_hdl);
 		zfs_vnode_forget(vp);
 		zp->z_vnode = NULL;
-		kmem_cache_free(znode_cache, zp);
+		zfs_znode_free_kmem(zp);
 		return (NULL);
 	}
 
@@ -763,6 +843,8 @@ zfs_xvattr_set(znode_t *zp, xvattr_t *xvap, dmu_tx_t *tx)
 	xoap = xva_getxoptattr(xvap);
 	ASSERT(xoap);
 
+	ASSERT_VOP_IN_SEQC(ZTOV(zp));
+
 	if (XVA_ISSET_REQ(xvap, XAT_CREATETIME)) {
 		uint64_t times[2];
 		ZFS_TIME_ENCODE(&xoap->xoa_createtime, times);
@@ -1191,8 +1273,7 @@ zfs_znode_free(znode_t *zp)
 		zp->z_acl_cached = NULL;
 	}
 
-	kmem_cache_free(znode_cache, zp);
+	zfs_znode_free_kmem(zp);
-
 }
 
 void
@@ -1628,7 +1709,7 @@ zfs_create_fs(objset_t *os, cred_t *cr, nvlist_t *zplprops, dmu_tx_t *tx)
 
 	zfsvfs = kmem_zalloc(sizeof (zfsvfs_t), KM_SLEEP);
 
-	rootzp = kmem_cache_alloc(znode_cache, KM_SLEEP);
+	rootzp = zfs_znode_alloc_kmem(KM_SLEEP);
 	ASSERT(!POINTER_IS_VALID(rootzp->z_zfsvfs));
 	rootzp->z_moved = 0;
 	rootzp->z_unlinked = 0;
@@ -1672,7 +1753,7 @@ zfs_create_fs(objset_t *os, cred_t *cr, nvlist_t *zplprops, dmu_tx_t *tx)
 	POINTER_INVALIDATE(&rootzp->z_zfsvfs);
 
 	sa_handle_destroy(rootzp->z_sa_hdl);
-	kmem_cache_free(znode_cache, rootzp);
+	zfs_znode_free_kmem(rootzp);
 
 	/*
 	 * Create shares directory