Index: head/sys/kern/vfs_acl.c =================================================================== --- head/sys/kern/vfs_acl.c (revision 363070) +++ head/sys/kern/vfs_acl.c (revision 363071) @@ -1,599 +1,599 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 1999-2006, 2016-2017 Robert N. M. Watson * All rights reserved. * * This software was developed by Robert Watson for the TrustedBSD Project. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Developed by the TrustedBSD Project. * * ACL system calls and other functions common across different ACL types. * Type-specific routines go into subr_acl_.c. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include CTASSERT(ACL_MAX_ENTRIES >= OLDACL_MAX_ENTRIES); MALLOC_DEFINE(M_ACL, "acl", "Access Control Lists"); static int kern___acl_aclcheck_path(struct thread *td, const char *path, acl_type_t type, struct acl *aclp, int follow); static int kern___acl_delete_path(struct thread *td, const char *path, acl_type_t type, int follow); static int kern___acl_get_path(struct thread *td, const char *path, acl_type_t type, struct acl *aclp, int follow); static int kern___acl_set_path(struct thread *td, const char *path, acl_type_t type, const struct acl *aclp, int follow); static int vacl_set_acl(struct thread *td, struct vnode *vp, acl_type_t type, const struct acl *aclp); static int vacl_get_acl(struct thread *td, struct vnode *vp, acl_type_t type, struct acl *aclp); static int vacl_aclcheck(struct thread *td, struct vnode *vp, acl_type_t type, const struct acl *aclp); int acl_copy_oldacl_into_acl(const struct oldacl *source, struct acl *dest) { int i; if (source->acl_cnt < 0 || source->acl_cnt > OLDACL_MAX_ENTRIES) return (EINVAL); - + bzero(dest, sizeof(*dest)); dest->acl_cnt = source->acl_cnt; dest->acl_maxcnt = ACL_MAX_ENTRIES; for (i = 0; i < dest->acl_cnt; i++) { dest->acl_entry[i].ae_tag = source->acl_entry[i].ae_tag; dest->acl_entry[i].ae_id = source->acl_entry[i].ae_id; dest->acl_entry[i].ae_perm = source->acl_entry[i].ae_perm; } return (0); } int acl_copy_acl_into_oldacl(const struct acl *source, struct oldacl *dest) { int i; if (source->acl_cnt > OLDACL_MAX_ENTRIES) return (EINVAL); bzero(dest, sizeof(*dest)); dest->acl_cnt = source->acl_cnt; for (i = 0; i < dest->acl_cnt; i++) { dest->acl_entry[i].ae_tag = source->acl_entry[i].ae_tag; dest->acl_entry[i].ae_id = source->acl_entry[i].ae_id; dest->acl_entry[i].ae_perm = source->acl_entry[i].ae_perm; } return (0); } /* * At one time, "struct ACL" was extended in order to add support for NFSv4 * ACLs. Instead of creating compatibility versions of all the ACL-related * syscalls, they were left intact. It's possible to find out what the code * calling these syscalls (libc) expects basing on "type" argument - if it's * either ACL_TYPE_ACCESS_OLD or ACL_TYPE_DEFAULT_OLD (which previously were * known as ACL_TYPE_ACCESS and ACL_TYPE_DEFAULT), then it's the "struct * oldacl". If it's something else, then it's the new "struct acl". In the * latter case, the routines below just copyin/copyout the contents. In the * former case, they copyin the "struct oldacl" and convert it to the new * format. */ static int acl_copyin(const void *user_acl, struct acl *kernel_acl, acl_type_t type) { int error; struct oldacl old; switch (type) { case ACL_TYPE_ACCESS_OLD: case ACL_TYPE_DEFAULT_OLD: error = copyin(user_acl, &old, sizeof(old)); if (error != 0) break; acl_copy_oldacl_into_acl(&old, kernel_acl); break; default: error = copyin(user_acl, kernel_acl, sizeof(*kernel_acl)); if (kernel_acl->acl_maxcnt != ACL_MAX_ENTRIES) return (EINVAL); } return (error); } static int acl_copyout(const struct acl *kernel_acl, void *user_acl, acl_type_t type) { uint32_t am; int error; struct oldacl old; switch (type) { case ACL_TYPE_ACCESS_OLD: case ACL_TYPE_DEFAULT_OLD: error = acl_copy_acl_into_oldacl(kernel_acl, &old); if (error != 0) break; error = copyout(&old, user_acl, sizeof(old)); break; default: error = fueword32((char *)user_acl + offsetof(struct acl, acl_maxcnt), &am); if (error == -1) return (EFAULT); if (am != ACL_MAX_ENTRIES) return (EINVAL); error = copyout(kernel_acl, user_acl, sizeof(*kernel_acl)); } return (error); } /* * Convert "old" type - ACL_TYPE_{ACCESS,DEFAULT}_OLD - into its "new" * counterpart. It's required for old (pre-NFSv4 ACLs) libc to work * with new kernel. Fixing 'type' for old binaries with new libc * is being done in lib/libc/posix1e/acl_support.c:_acl_type_unold(). */ static int acl_type_unold(int type) { switch (type) { case ACL_TYPE_ACCESS_OLD: return (ACL_TYPE_ACCESS); case ACL_TYPE_DEFAULT_OLD: return (ACL_TYPE_DEFAULT); default: return (type); } } /* * These calls wrap the real vnode operations, and are called by the syscall * code once the syscall has converted the path or file descriptor to a vnode * (unlocked). The aclp pointer is assumed still to point to userland, so * this should not be consumed within the kernel except by syscall code. * Other code should directly invoke VOP_{SET,GET}ACL. */ /* * Given a vnode, set its ACL. */ static int vacl_set_acl(struct thread *td, struct vnode *vp, acl_type_t type, const struct acl *aclp) { struct acl *inkernelacl; struct mount *mp; int error; AUDIT_ARG_VALUE(type); inkernelacl = acl_alloc(M_WAITOK); error = acl_copyin(aclp, inkernelacl, type); if (error != 0) goto out; error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error != 0) goto out; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); #ifdef MAC error = mac_vnode_check_setacl(td->td_ucred, vp, type, inkernelacl); if (error != 0) goto out_unlock; #endif error = VOP_SETACL(vp, acl_type_unold(type), inkernelacl, td->td_ucred, td); #ifdef MAC out_unlock: #endif VOP_UNLOCK(vp); vn_finished_write(mp); out: acl_free(inkernelacl); return (error); } /* * Given a vnode, get its ACL. */ static int vacl_get_acl(struct thread *td, struct vnode *vp, acl_type_t type, struct acl *aclp) { struct acl *inkernelacl; int error; AUDIT_ARG_VALUE(type); inkernelacl = acl_alloc(M_WAITOK | M_ZERO); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); #ifdef MAC error = mac_vnode_check_getacl(td->td_ucred, vp, type); if (error != 0) goto out; #endif error = VOP_GETACL(vp, acl_type_unold(type), inkernelacl, td->td_ucred, td); #ifdef MAC out: #endif VOP_UNLOCK(vp); if (error == 0) error = acl_copyout(inkernelacl, aclp, type); acl_free(inkernelacl); return (error); } /* * Given a vnode, delete its ACL. */ static int vacl_delete(struct thread *td, struct vnode *vp, acl_type_t type) { struct mount *mp; int error; AUDIT_ARG_VALUE(type); error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); #ifdef MAC error = mac_vnode_check_deleteacl(td->td_ucred, vp, type); if (error != 0) goto out; #endif error = VOP_SETACL(vp, acl_type_unold(type), 0, td->td_ucred, td); #ifdef MAC out: #endif VOP_UNLOCK(vp); vn_finished_write(mp); return (error); } /* * Given a vnode, check whether an ACL is appropriate for it * * XXXRW: No vnode lock held so can't audit vnode state...? */ static int vacl_aclcheck(struct thread *td, struct vnode *vp, acl_type_t type, const struct acl *aclp) { struct acl *inkernelacl; int error; inkernelacl = acl_alloc(M_WAITOK); error = acl_copyin(aclp, inkernelacl, type); if (error != 0) goto out; error = VOP_ACLCHECK(vp, acl_type_unold(type), inkernelacl, td->td_ucred, td); out: acl_free(inkernelacl); return (error); } /* * syscalls -- convert the path/fd to a vnode, and call vacl_whatever. Don't * need to lock, as the vacl_ code will get/release any locks required. */ /* * Given a file path, get an ACL for it */ int sys___acl_get_file(struct thread *td, struct __acl_get_file_args *uap) { return (kern___acl_get_path(td, uap->path, uap->type, uap->aclp, FOLLOW)); } /* * Given a file path, get an ACL for it; don't follow links. */ int sys___acl_get_link(struct thread *td, struct __acl_get_link_args *uap) { return(kern___acl_get_path(td, uap->path, uap->type, uap->aclp, NOFOLLOW)); } static int kern___acl_get_path(struct thread *td, const char *path, acl_type_t type, struct acl *aclp, int follow) { struct nameidata nd; int error; NDINIT(&nd, LOOKUP, follow | AUDITVNODE1, UIO_USERSPACE, path, td); error = namei(&nd); if (error == 0) { error = vacl_get_acl(td, nd.ni_vp, type, aclp); NDFREE(&nd, 0); } return (error); } /* * Given a file path, set an ACL for it. */ int sys___acl_set_file(struct thread *td, struct __acl_set_file_args *uap) { return(kern___acl_set_path(td, uap->path, uap->type, uap->aclp, FOLLOW)); } /* * Given a file path, set an ACL for it; don't follow links. */ int sys___acl_set_link(struct thread *td, struct __acl_set_link_args *uap) { return(kern___acl_set_path(td, uap->path, uap->type, uap->aclp, NOFOLLOW)); } static int kern___acl_set_path(struct thread *td, const char *path, acl_type_t type, const struct acl *aclp, int follow) { struct nameidata nd; int error; NDINIT(&nd, LOOKUP, follow | AUDITVNODE1, UIO_USERSPACE, path, td); error = namei(&nd); if (error == 0) { error = vacl_set_acl(td, nd.ni_vp, type, aclp); NDFREE(&nd, 0); } return (error); } /* * Given a file descriptor, get an ACL for it. */ int sys___acl_get_fd(struct thread *td, struct __acl_get_fd_args *uap) { struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(uap->filedes); error = getvnode(td, uap->filedes, cap_rights_init_one(&rights, CAP_ACL_GET), &fp); if (error == 0) { error = vacl_get_acl(td, fp->f_vnode, uap->type, uap->aclp); fdrop(fp, td); } return (error); } /* * Given a file descriptor, set an ACL for it. */ int sys___acl_set_fd(struct thread *td, struct __acl_set_fd_args *uap) { struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(uap->filedes); error = getvnode(td, uap->filedes, cap_rights_init_one(&rights, CAP_ACL_SET), &fp); if (error == 0) { error = vacl_set_acl(td, fp->f_vnode, uap->type, uap->aclp); fdrop(fp, td); } return (error); } /* * Given a file path, delete an ACL from it. */ int sys___acl_delete_file(struct thread *td, struct __acl_delete_file_args *uap) { return (kern___acl_delete_path(td, uap->path, uap->type, FOLLOW)); } /* * Given a file path, delete an ACL from it; don't follow links. */ int sys___acl_delete_link(struct thread *td, struct __acl_delete_link_args *uap) { return (kern___acl_delete_path(td, uap->path, uap->type, NOFOLLOW)); } static int kern___acl_delete_path(struct thread *td, const char *path, acl_type_t type, int follow) { struct nameidata nd; int error; NDINIT(&nd, LOOKUP, follow, UIO_USERSPACE, path, td); error = namei(&nd); if (error == 0) { error = vacl_delete(td, nd.ni_vp, type); NDFREE(&nd, 0); } return (error); } /* * Given a file path, delete an ACL from it. */ int sys___acl_delete_fd(struct thread *td, struct __acl_delete_fd_args *uap) { struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(uap->filedes); error = getvnode(td, uap->filedes, cap_rights_init_one(&rights, CAP_ACL_DELETE), &fp); if (error == 0) { error = vacl_delete(td, fp->f_vnode, uap->type); fdrop(fp, td); } return (error); } /* * Given a file path, check an ACL for it. */ int sys___acl_aclcheck_file(struct thread *td, struct __acl_aclcheck_file_args *uap) { return (kern___acl_aclcheck_path(td, uap->path, uap->type, uap->aclp, FOLLOW)); } /* * Given a file path, check an ACL for it; don't follow links. */ int sys___acl_aclcheck_link(struct thread *td, struct __acl_aclcheck_link_args *uap) { return (kern___acl_aclcheck_path(td, uap->path, uap->type, uap->aclp, NOFOLLOW)); } static int kern___acl_aclcheck_path(struct thread *td, const char *path, acl_type_t type, struct acl *aclp, int follow) { struct nameidata nd; int error; NDINIT(&nd, LOOKUP, follow, UIO_USERSPACE, path, td); error = namei(&nd); if (error == 0) { error = vacl_aclcheck(td, nd.ni_vp, type, aclp); NDFREE(&nd, 0); } return (error); } /* * Given a file descriptor, check an ACL for it. */ int sys___acl_aclcheck_fd(struct thread *td, struct __acl_aclcheck_fd_args *uap) { struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(uap->filedes); error = getvnode(td, uap->filedes, cap_rights_init_one(&rights, CAP_ACL_CHECK), &fp); if (error == 0) { error = vacl_aclcheck(td, fp->f_vnode, uap->type, uap->aclp); fdrop(fp, td); } return (error); } struct acl * acl_alloc(int flags) { struct acl *aclp; aclp = malloc(sizeof(*aclp), M_ACL, flags); if (aclp == NULL) return (NULL); aclp->acl_maxcnt = ACL_MAX_ENTRIES; return (aclp); } void acl_free(struct acl *aclp) { free(aclp, M_ACL); } Index: head/sys/kern/vfs_bio.c =================================================================== --- head/sys/kern/vfs_bio.c (revision 363070) +++ head/sys/kern/vfs_bio.c (revision 363071) @@ -1,5458 +1,5458 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2004 Poul-Henning Kamp * Copyright (c) 1994,1997 John S. Dyson * Copyright (c) 2013 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * this file contains a new buffer I/O scheme implementing a coherent * VM object and buffer cache scheme. Pains have been taken to make * sure that the performance degradation associated with schemes such * as this is not realized. * * Author: John S. Dyson * Significant help during the development and debugging phases * had been provided by David Greenman, also of the FreeBSD core team. * * see man buf(9) for more info. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_BIOBUF, "biobuf", "BIO buffer"); struct bio_ops bioops; /* I/O operation notification */ struct buf_ops buf_ops_bio = { .bop_name = "buf_ops_bio", .bop_write = bufwrite, .bop_strategy = bufstrategy, .bop_sync = bufsync, .bop_bdflush = bufbdflush, }; struct bufqueue { struct mtx_padalign bq_lock; TAILQ_HEAD(, buf) bq_queue; uint8_t bq_index; uint16_t bq_subqueue; int bq_len; } __aligned(CACHE_LINE_SIZE); #define BQ_LOCKPTR(bq) (&(bq)->bq_lock) #define BQ_LOCK(bq) mtx_lock(BQ_LOCKPTR((bq))) #define BQ_UNLOCK(bq) mtx_unlock(BQ_LOCKPTR((bq))) #define BQ_ASSERT_LOCKED(bq) mtx_assert(BQ_LOCKPTR((bq)), MA_OWNED) struct bufdomain { struct bufqueue bd_subq[MAXCPU + 1]; /* Per-cpu sub queues + global */ struct bufqueue bd_dirtyq; struct bufqueue *bd_cleanq; struct mtx_padalign bd_run_lock; /* Constants */ long bd_maxbufspace; long bd_hibufspace; long bd_lobufspace; long bd_bufspacethresh; int bd_hifreebuffers; int bd_lofreebuffers; int bd_hidirtybuffers; int bd_lodirtybuffers; int bd_dirtybufthresh; int bd_lim; /* atomics */ int bd_wanted; int __aligned(CACHE_LINE_SIZE) bd_numdirtybuffers; int __aligned(CACHE_LINE_SIZE) bd_running; long __aligned(CACHE_LINE_SIZE) bd_bufspace; int __aligned(CACHE_LINE_SIZE) bd_freebuffers; } __aligned(CACHE_LINE_SIZE); #define BD_LOCKPTR(bd) (&(bd)->bd_cleanq->bq_lock) #define BD_LOCK(bd) mtx_lock(BD_LOCKPTR((bd))) #define BD_UNLOCK(bd) mtx_unlock(BD_LOCKPTR((bd))) #define BD_ASSERT_LOCKED(bd) mtx_assert(BD_LOCKPTR((bd)), MA_OWNED) #define BD_RUN_LOCKPTR(bd) (&(bd)->bd_run_lock) #define BD_RUN_LOCK(bd) mtx_lock(BD_RUN_LOCKPTR((bd))) #define BD_RUN_UNLOCK(bd) mtx_unlock(BD_RUN_LOCKPTR((bd))) #define BD_DOMAIN(bd) (bd - bdomain) static struct buf *buf; /* buffer header pool */ extern struct buf *swbuf; /* Swap buffer header pool. */ caddr_t __read_mostly unmapped_buf; /* Used below and for softdep flushing threads in ufs/ffs/ffs_softdep.c */ struct proc *bufdaemonproc; static int inmem(struct vnode *vp, daddr_t blkno); static void vm_hold_free_pages(struct buf *bp, int newbsize); static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to); static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m); static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m); static void vfs_clean_pages_dirty_buf(struct buf *bp); static void vfs_setdirty_range(struct buf *bp); static void vfs_vmio_invalidate(struct buf *bp); static void vfs_vmio_truncate(struct buf *bp, int npages); static void vfs_vmio_extend(struct buf *bp, int npages, int size); static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno); static void breada(struct vnode *, daddr_t *, int *, int, struct ucred *, int, void (*)(struct buf *)); static int buf_flush(struct vnode *vp, struct bufdomain *, int); static int flushbufqueues(struct vnode *, struct bufdomain *, int, int); static void buf_daemon(void); static __inline void bd_wakeup(void); static int sysctl_runningspace(SYSCTL_HANDLER_ARGS); static void bufkva_reclaim(vmem_t *, int); static void bufkva_free(struct buf *); static int buf_import(void *, void **, int, int, int); static void buf_release(void *, void **, int); static void maxbcachebuf_adjust(void); static inline struct bufdomain *bufdomain(struct buf *); static void bq_remove(struct bufqueue *bq, struct buf *bp); static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock); static int buf_recycle(struct bufdomain *, bool kva); static void bq_init(struct bufqueue *bq, int qindex, int cpu, const char *lockname); static void bd_init(struct bufdomain *bd); static int bd_flushall(struct bufdomain *bd); static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS); static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS); static int sysctl_bufspace(SYSCTL_HANDLER_ARGS); int vmiodirenable = TRUE; SYSCTL_INT(_vfs, OID_AUTO, vmiodirenable, CTLFLAG_RW, &vmiodirenable, 0, "Use the VM system for directory writes"); long runningbufspace; SYSCTL_LONG(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0, "Amount of presently outstanding async buffer io"); SYSCTL_PROC(_vfs, OID_AUTO, bufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_bufspace, "L", "Physical memory used for buffers"); static counter_u64_t bufkvaspace; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufkvaspace, CTLFLAG_RD, &bufkvaspace, "Kernel virtual memory used for buffers"); static long maxbufspace; SYSCTL_PROC(_vfs, OID_AUTO, maxbufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &maxbufspace, __offsetof(struct bufdomain, bd_maxbufspace), sysctl_bufdomain_long, "L", "Maximum allowed value of bufspace (including metadata)"); static long bufmallocspace; SYSCTL_LONG(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0, "Amount of malloced memory for buffers"); static long maxbufmallocspace; SYSCTL_LONG(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RW, &maxbufmallocspace, 0, "Maximum amount of malloced memory for buffers"); static long lobufspace; SYSCTL_PROC(_vfs, OID_AUTO, lobufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &lobufspace, __offsetof(struct bufdomain, bd_lobufspace), sysctl_bufdomain_long, "L", "Minimum amount of buffers we want to have"); long hibufspace; SYSCTL_PROC(_vfs, OID_AUTO, hibufspace, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &hibufspace, __offsetof(struct bufdomain, bd_hibufspace), sysctl_bufdomain_long, "L", "Maximum allowed value of bufspace (excluding metadata)"); long bufspacethresh; SYSCTL_PROC(_vfs, OID_AUTO, bufspacethresh, CTLTYPE_LONG|CTLFLAG_MPSAFE|CTLFLAG_RW, &bufspacethresh, __offsetof(struct bufdomain, bd_bufspacethresh), sysctl_bufdomain_long, "L", "Bufspace consumed before waking the daemon to free some"); static counter_u64_t buffreekvacnt; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RW, &buffreekvacnt, "Number of times we have freed the KVA space from some buffer"); static counter_u64_t bufdefragcnt; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RW, &bufdefragcnt, "Number of times we have had to repeat buffer allocation to defragment"); static long lorunningspace; SYSCTL_PROC(_vfs, OID_AUTO, lorunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | CTLFLAG_RW, &lorunningspace, 0, sysctl_runningspace, "L", "Minimum preferred space used for in-progress I/O"); static long hirunningspace; SYSCTL_PROC(_vfs, OID_AUTO, hirunningspace, CTLTYPE_LONG | CTLFLAG_MPSAFE | CTLFLAG_RW, &hirunningspace, 0, sysctl_runningspace, "L", "Maximum amount of space to use for in-progress I/O"); int dirtybufferflushes; SYSCTL_INT(_vfs, OID_AUTO, dirtybufferflushes, CTLFLAG_RW, &dirtybufferflushes, 0, "Number of bdwrite to bawrite conversions to limit dirty buffers"); int bdwriteskip; SYSCTL_INT(_vfs, OID_AUTO, bdwriteskip, CTLFLAG_RW, &bdwriteskip, 0, "Number of buffers supplied to bdwrite with snapshot deadlock risk"); int altbufferflushes; SYSCTL_INT(_vfs, OID_AUTO, altbufferflushes, CTLFLAG_RW | CTLFLAG_STATS, &altbufferflushes, 0, "Number of fsync flushes to limit dirty buffers"); static int recursiveflushes; SYSCTL_INT(_vfs, OID_AUTO, recursiveflushes, CTLFLAG_RW | CTLFLAG_STATS, &recursiveflushes, 0, "Number of flushes skipped due to being recursive"); static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS); SYSCTL_PROC(_vfs, OID_AUTO, numdirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RD, NULL, 0, sysctl_numdirtybuffers, "I", "Number of buffers that are dirty (has unwritten changes) at the moment"); static int lodirtybuffers; SYSCTL_PROC(_vfs, OID_AUTO, lodirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lodirtybuffers, __offsetof(struct bufdomain, bd_lodirtybuffers), sysctl_bufdomain_int, "I", "How many buffers we want to have free before bufdaemon can sleep"); static int hidirtybuffers; SYSCTL_PROC(_vfs, OID_AUTO, hidirtybuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hidirtybuffers, __offsetof(struct bufdomain, bd_hidirtybuffers), sysctl_bufdomain_int, "I", "When the number of dirty buffers is considered severe"); int dirtybufthresh; SYSCTL_PROC(_vfs, OID_AUTO, dirtybufthresh, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &dirtybufthresh, __offsetof(struct bufdomain, bd_dirtybufthresh), sysctl_bufdomain_int, "I", "Number of bdwrite to bawrite conversions to clear dirty buffers"); static int numfreebuffers; SYSCTL_INT(_vfs, OID_AUTO, numfreebuffers, CTLFLAG_RD, &numfreebuffers, 0, "Number of free buffers"); static int lofreebuffers; SYSCTL_PROC(_vfs, OID_AUTO, lofreebuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &lofreebuffers, __offsetof(struct bufdomain, bd_lofreebuffers), sysctl_bufdomain_int, "I", "Target number of free buffers"); static int hifreebuffers; SYSCTL_PROC(_vfs, OID_AUTO, hifreebuffers, CTLTYPE_INT|CTLFLAG_MPSAFE|CTLFLAG_RW, &hifreebuffers, __offsetof(struct bufdomain, bd_hifreebuffers), sysctl_bufdomain_int, "I", "Threshold for clean buffer recycling"); static counter_u64_t getnewbufcalls; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, "Number of calls to getnewbuf"); static counter_u64_t getnewbufrestarts; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts, "Number of times getnewbuf has had to restart a buffer acquisition"); static counter_u64_t mappingrestarts; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, mappingrestarts, CTLFLAG_RD, &mappingrestarts, "Number of times getblk has had to restart a buffer mapping for " "unmapped buffer"); static counter_u64_t numbufallocfails; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, numbufallocfails, CTLFLAG_RW, &numbufallocfails, "Number of times buffer allocations failed"); static int flushbufqtarget = 100; SYSCTL_INT(_vfs, OID_AUTO, flushbufqtarget, CTLFLAG_RW, &flushbufqtarget, 0, "Amount of work to do in flushbufqueues when helping bufdaemon"); static counter_u64_t notbufdflushes; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, notbufdflushes, CTLFLAG_RD, ¬bufdflushes, "Number of dirty buffer flushes done by the bufdaemon helpers"); static long barrierwrites; SYSCTL_LONG(_vfs, OID_AUTO, barrierwrites, CTLFLAG_RW | CTLFLAG_STATS, &barrierwrites, 0, "Number of barrier writes"); SYSCTL_INT(_vfs, OID_AUTO, unmapped_buf_allowed, CTLFLAG_RD, &unmapped_buf_allowed, 0, "Permit the use of the unmapped i/o"); int maxbcachebuf = MAXBCACHEBUF; SYSCTL_INT(_vfs, OID_AUTO, maxbcachebuf, CTLFLAG_RDTUN, &maxbcachebuf, 0, "Maximum size of a buffer cache block"); /* * This lock synchronizes access to bd_request. */ static struct mtx_padalign __exclusive_cache_line bdlock; /* * This lock protects the runningbufreq and synchronizes runningbufwakeup and * waitrunningbufspace(). */ static struct mtx_padalign __exclusive_cache_line rbreqlock; /* * Lock that protects bdirtywait. */ static struct mtx_padalign __exclusive_cache_line bdirtylock; /* * Wakeup point for bufdaemon, as well as indicator of whether it is already * active. Set to 1 when the bufdaemon is already "on" the queue, 0 when it * is idling. */ static int bd_request; /* * Request for the buf daemon to write more buffers than is indicated by * lodirtybuf. This may be necessary to push out excess dependencies or * defragment the address space where a simple count of the number of dirty * buffers is insufficient to characterize the demand for flushing them. */ static int bd_speedupreq; /* * Synchronization (sleep/wakeup) variable for active buffer space requests. * Set when wait starts, cleared prior to wakeup(). * Used in runningbufwakeup() and waitrunningbufspace(). */ static int runningbufreq; /* * Synchronization for bwillwrite() waiters. */ static int bdirtywait; /* * Definitions for the buffer free lists. */ #define QUEUE_NONE 0 /* on no queue */ #define QUEUE_EMPTY 1 /* empty buffer headers */ #define QUEUE_DIRTY 2 /* B_DELWRI buffers */ #define QUEUE_CLEAN 3 /* non-B_DELWRI buffers */ #define QUEUE_SENTINEL 4 /* not an queue index, but mark for sentinel */ /* Maximum number of buffer domains. */ #define BUF_DOMAINS 8 struct bufdomainset bdlodirty; /* Domains > lodirty */ struct bufdomainset bdhidirty; /* Domains > hidirty */ /* Configured number of clean queues. */ static int __read_mostly buf_domains; BITSET_DEFINE(bufdomainset, BUF_DOMAINS); struct bufdomain __exclusive_cache_line bdomain[BUF_DOMAINS]; struct bufqueue __exclusive_cache_line bqempty; /* * per-cpu empty buffer cache. */ uma_zone_t buf_zone; /* * Single global constant for BUF_WMESG, to avoid getting multiple references. * buf_wmesg is referred from macros. */ const char *buf_wmesg = BUF_WMESG; static int sysctl_runningspace(SYSCTL_HANDLER_ARGS) { long value; int error; value = *(long *)arg1; error = sysctl_handle_long(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); mtx_lock(&rbreqlock); if (arg1 == &hirunningspace) { if (value < lorunningspace) error = EINVAL; else hirunningspace = value; } else { KASSERT(arg1 == &lorunningspace, ("%s: unknown arg1", __func__)); if (value > hirunningspace) error = EINVAL; else lorunningspace = value; } mtx_unlock(&rbreqlock); return (error); } static int sysctl_bufdomain_int(SYSCTL_HANDLER_ARGS) { int error; int value; int i; value = *(int *)arg1; error = sysctl_handle_int(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); *(int *)arg1 = value; for (i = 0; i < buf_domains; i++) *(int *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = value / buf_domains; return (error); } static int sysctl_bufdomain_long(SYSCTL_HANDLER_ARGS) { long value; int error; int i; value = *(long *)arg1; error = sysctl_handle_long(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); *(long *)arg1 = value; for (i = 0; i < buf_domains; i++) *(long *)(uintptr_t)(((uintptr_t)&bdomain[i]) + arg2) = value / buf_domains; return (error); } #if defined(COMPAT_FREEBSD4) || defined(COMPAT_FREEBSD5) || \ defined(COMPAT_FREEBSD6) || defined(COMPAT_FREEBSD7) static int sysctl_bufspace(SYSCTL_HANDLER_ARGS) { long lvalue; int ivalue; int i; lvalue = 0; for (i = 0; i < buf_domains; i++) lvalue += bdomain[i].bd_bufspace; if (sizeof(int) == sizeof(long) || req->oldlen >= sizeof(long)) return (sysctl_handle_long(oidp, &lvalue, 0, req)); if (lvalue > INT_MAX) /* On overflow, still write out a long to trigger ENOMEM. */ return (sysctl_handle_long(oidp, &lvalue, 0, req)); ivalue = lvalue; return (sysctl_handle_int(oidp, &ivalue, 0, req)); } #else static int sysctl_bufspace(SYSCTL_HANDLER_ARGS) { long lvalue; int i; lvalue = 0; for (i = 0; i < buf_domains; i++) lvalue += bdomain[i].bd_bufspace; return (sysctl_handle_long(oidp, &lvalue, 0, req)); } #endif static int sysctl_numdirtybuffers(SYSCTL_HANDLER_ARGS) { int value; int i; value = 0; for (i = 0; i < buf_domains; i++) value += bdomain[i].bd_numdirtybuffers; return (sysctl_handle_int(oidp, &value, 0, req)); } /* * bdirtywakeup: * * Wakeup any bwillwrite() waiters. */ static void bdirtywakeup(void) { mtx_lock(&bdirtylock); if (bdirtywait) { bdirtywait = 0; wakeup(&bdirtywait); } mtx_unlock(&bdirtylock); } /* * bd_clear: * * Clear a domain from the appropriate bitsets when dirtybuffers * is decremented. */ static void bd_clear(struct bufdomain *bd) { mtx_lock(&bdirtylock); if (bd->bd_numdirtybuffers <= bd->bd_lodirtybuffers) BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); if (bd->bd_numdirtybuffers <= bd->bd_hidirtybuffers) BIT_CLR(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); mtx_unlock(&bdirtylock); } /* * bd_set: * * Set a domain in the appropriate bitsets when dirtybuffers * is incremented. */ static void bd_set(struct bufdomain *bd) { mtx_lock(&bdirtylock); if (bd->bd_numdirtybuffers > bd->bd_lodirtybuffers) BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdlodirty); if (bd->bd_numdirtybuffers > bd->bd_hidirtybuffers) BIT_SET(BUF_DOMAINS, BD_DOMAIN(bd), &bdhidirty); mtx_unlock(&bdirtylock); } /* * bdirtysub: * * Decrement the numdirtybuffers count by one and wakeup any * threads blocked in bwillwrite(). */ static void bdirtysub(struct buf *bp) { struct bufdomain *bd; int num; bd = bufdomain(bp); num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, -1); if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) bdirtywakeup(); if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) bd_clear(bd); } /* * bdirtyadd: * * Increment the numdirtybuffers count by one and wakeup the buf * daemon if needed. */ static void bdirtyadd(struct buf *bp) { struct bufdomain *bd; int num; /* * Only do the wakeup once as we cross the boundary. The * buf daemon will keep running until the condition clears. */ bd = bufdomain(bp); num = atomic_fetchadd_int(&bd->bd_numdirtybuffers, 1); if (num == (bd->bd_lodirtybuffers + bd->bd_hidirtybuffers) / 2) bd_wakeup(); if (num == bd->bd_lodirtybuffers || num == bd->bd_hidirtybuffers) bd_set(bd); } /* * bufspace_daemon_wakeup: * * Wakeup the daemons responsible for freeing clean bufs. */ static void bufspace_daemon_wakeup(struct bufdomain *bd) { /* * avoid the lock if the daemon is running. */ if (atomic_fetchadd_int(&bd->bd_running, 1) == 0) { BD_RUN_LOCK(bd); atomic_store_int(&bd->bd_running, 1); wakeup(&bd->bd_running); BD_RUN_UNLOCK(bd); } } /* * bufspace_daemon_wait: * * Sleep until the domain falls below a limit or one second passes. */ static void bufspace_daemon_wait(struct bufdomain *bd) { /* * Re-check our limits and sleep. bd_running must be * cleared prior to checking the limits to avoid missed * wakeups. The waker will adjust one of bufspace or * freebuffers prior to checking bd_running. */ BD_RUN_LOCK(bd); atomic_store_int(&bd->bd_running, 0); if (bd->bd_bufspace < bd->bd_bufspacethresh && bd->bd_freebuffers > bd->bd_lofreebuffers) { msleep(&bd->bd_running, BD_RUN_LOCKPTR(bd), PRIBIO|PDROP, "-", hz); } else { /* Avoid spurious wakeups while running. */ atomic_store_int(&bd->bd_running, 1); BD_RUN_UNLOCK(bd); } } /* * bufspace_adjust: * * Adjust the reported bufspace for a KVA managed buffer, possibly * waking any waiters. */ static void bufspace_adjust(struct buf *bp, int bufsize) { struct bufdomain *bd; long space; int diff; KASSERT((bp->b_flags & B_MALLOC) == 0, ("bufspace_adjust: malloc buf %p", bp)); bd = bufdomain(bp); diff = bufsize - bp->b_bufsize; if (diff < 0) { atomic_subtract_long(&bd->bd_bufspace, -diff); } else if (diff > 0) { space = atomic_fetchadd_long(&bd->bd_bufspace, diff); /* Wake up the daemon on the transition. */ if (space < bd->bd_bufspacethresh && space + diff >= bd->bd_bufspacethresh) bufspace_daemon_wakeup(bd); } bp->b_bufsize = bufsize; } /* * bufspace_reserve: * * Reserve bufspace before calling allocbuf(). metadata has a * different space limit than data. */ static int bufspace_reserve(struct bufdomain *bd, int size, bool metadata) { long limit, new; long space; if (metadata) limit = bd->bd_maxbufspace; else limit = bd->bd_hibufspace; space = atomic_fetchadd_long(&bd->bd_bufspace, size); new = space + size; if (new > limit) { atomic_subtract_long(&bd->bd_bufspace, size); return (ENOSPC); } /* Wake up the daemon on the transition. */ if (space < bd->bd_bufspacethresh && new >= bd->bd_bufspacethresh) bufspace_daemon_wakeup(bd); return (0); } /* * bufspace_release: * * Release reserved bufspace after bufspace_adjust() has consumed it. */ static void bufspace_release(struct bufdomain *bd, int size) { atomic_subtract_long(&bd->bd_bufspace, size); } /* * bufspace_wait: * * Wait for bufspace, acting as the buf daemon if a locked vnode is * supplied. bd_wanted must be set prior to polling for space. The * operation must be re-tried on return. */ static void bufspace_wait(struct bufdomain *bd, struct vnode *vp, int gbflags, int slpflag, int slptimeo) { struct thread *td; int error, fl, norunbuf; if ((gbflags & GB_NOWAIT_BD) != 0) return; td = curthread; BD_LOCK(bd); while (bd->bd_wanted) { if (vp != NULL && vp->v_type != VCHR && (td->td_pflags & TDP_BUFNEED) == 0) { BD_UNLOCK(bd); /* * getblk() is called with a vnode locked, and * some majority of the dirty buffers may as * well belong to the vnode. Flushing the * buffers there would make a progress that * cannot be achieved by the buf_daemon, that * cannot lock the vnode. */ norunbuf = ~(TDP_BUFNEED | TDP_NORUNNINGBUF) | (td->td_pflags & TDP_NORUNNINGBUF); /* * Play bufdaemon. The getnewbuf() function * may be called while the thread owns lock * for another dirty buffer for the same * vnode, which makes it impossible to use * VOP_FSYNC() there, due to the buffer lock * recursion. */ td->td_pflags |= TDP_BUFNEED | TDP_NORUNNINGBUF; fl = buf_flush(vp, bd, flushbufqtarget); td->td_pflags &= norunbuf; BD_LOCK(bd); if (fl != 0) continue; if (bd->bd_wanted == 0) break; } error = msleep(&bd->bd_wanted, BD_LOCKPTR(bd), (PRIBIO + 4) | slpflag, "newbuf", slptimeo); if (error != 0) break; } BD_UNLOCK(bd); } /* * bufspace_daemon: * * buffer space management daemon. Tries to maintain some marginal * amount of free buffer space so that requesting processes neither * block nor work to reclaim buffers. */ static void bufspace_daemon(void *arg) { struct bufdomain *bd; EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, SHUTDOWN_PRI_LAST + 100); bd = arg; for (;;) { kthread_suspend_check(); /* * Free buffers from the clean queue until we meet our * targets. * * Theory of operation: The buffer cache is most efficient * when some free buffer headers and space are always * available to getnewbuf(). This daemon attempts to prevent * the excessive blocking and synchronization associated * with shortfall. It goes through three phases according * demand: * * 1) The daemon wakes up voluntarily once per-second * during idle periods when the counters are below * the wakeup thresholds (bufspacethresh, lofreebuffers). * * 2) The daemon wakes up as we cross the thresholds * ahead of any potential blocking. This may bounce * slightly according to the rate of consumption and * release. * * 3) The daemon and consumers are starved for working * clean buffers. This is the 'bufspace' sleep below * which will inefficiently trade bufs with bqrelse * until we return to condition 2. */ while (bd->bd_bufspace > bd->bd_lobufspace || bd->bd_freebuffers < bd->bd_hifreebuffers) { if (buf_recycle(bd, false) != 0) { if (bd_flushall(bd)) continue; /* * Speedup dirty if we've run out of clean * buffers. This is possible in particular * because softdep may held many bufs locked * pending writes to other bufs which are * marked for delayed write, exhausting * clean space until they are written. */ bd_speedup(); BD_LOCK(bd); if (bd->bd_wanted) { msleep(&bd->bd_wanted, BD_LOCKPTR(bd), PRIBIO|PDROP, "bufspace", hz/10); } else BD_UNLOCK(bd); } maybe_yield(); } bufspace_daemon_wait(bd); } } /* * bufmallocadjust: * * Adjust the reported bufspace for a malloc managed buffer, possibly * waking any waiters. */ static void bufmallocadjust(struct buf *bp, int bufsize) { int diff; KASSERT((bp->b_flags & B_MALLOC) != 0, ("bufmallocadjust: non-malloc buf %p", bp)); diff = bufsize - bp->b_bufsize; if (diff < 0) atomic_subtract_long(&bufmallocspace, -diff); else atomic_add_long(&bufmallocspace, diff); bp->b_bufsize = bufsize; } /* * runningwakeup: * * Wake up processes that are waiting on asynchronous writes to fall * below lorunningspace. */ static void runningwakeup(void) { mtx_lock(&rbreqlock); if (runningbufreq) { runningbufreq = 0; wakeup(&runningbufreq); } mtx_unlock(&rbreqlock); } /* * runningbufwakeup: * * Decrement the outstanding write count according. */ void runningbufwakeup(struct buf *bp) { long space, bspace; bspace = bp->b_runningbufspace; if (bspace == 0) return; space = atomic_fetchadd_long(&runningbufspace, -bspace); KASSERT(space >= bspace, ("runningbufspace underflow %ld %ld", space, bspace)); bp->b_runningbufspace = 0; /* * Only acquire the lock and wakeup on the transition from exceeding * the threshold to falling below it. */ if (space < lorunningspace) return; if (space - bspace > lorunningspace) return; runningwakeup(); } /* * waitrunningbufspace() * * runningbufspace is a measure of the amount of I/O currently * running. This routine is used in async-write situations to * prevent creating huge backups of pending writes to a device. * Only asynchronous writes are governed by this function. * * This does NOT turn an async write into a sync write. It waits * for earlier writes to complete and generally returns before the * caller's write has reached the device. */ void waitrunningbufspace(void) { mtx_lock(&rbreqlock); while (runningbufspace > hirunningspace) { runningbufreq = 1; msleep(&runningbufreq, &rbreqlock, PVM, "wdrain", 0); } mtx_unlock(&rbreqlock); } /* * vfs_buf_test_cache: * * Called when a buffer is extended. This function clears the B_CACHE * bit if the newly extended portion of the buffer does not contain * valid data. */ static __inline void vfs_buf_test_cache(struct buf *bp, vm_ooffset_t foff, vm_offset_t off, vm_offset_t size, vm_page_t m) { /* * This function and its results are protected by higher level * synchronization requiring vnode and buf locks to page in and * validate pages. */ if (bp->b_flags & B_CACHE) { int base = (foff + off) & PAGE_MASK; if (vm_page_is_valid(m, base, size) == 0) bp->b_flags &= ~B_CACHE; } } /* Wake up the buffer daemon if necessary */ static void bd_wakeup(void) { mtx_lock(&bdlock); if (bd_request == 0) { bd_request = 1; wakeup(&bd_request); } mtx_unlock(&bdlock); } /* * Adjust the maxbcachbuf tunable. */ static void maxbcachebuf_adjust(void) { int i; /* * maxbcachebuf must be a power of 2 >= MAXBSIZE. */ i = 2; while (i * 2 <= maxbcachebuf) i *= 2; maxbcachebuf = i; if (maxbcachebuf < MAXBSIZE) maxbcachebuf = MAXBSIZE; if (maxbcachebuf > MAXPHYS) maxbcachebuf = MAXPHYS; if (bootverbose != 0 && maxbcachebuf != MAXBCACHEBUF) printf("maxbcachebuf=%d\n", maxbcachebuf); } /* * bd_speedup - speedup the buffer cache flushing code */ void bd_speedup(void) { int needwake; mtx_lock(&bdlock); needwake = 0; if (bd_speedupreq == 0 || bd_request == 0) needwake = 1; bd_speedupreq = 1; bd_request = 1; if (needwake) wakeup(&bd_request); mtx_unlock(&bdlock); } #ifdef __i386__ #define TRANSIENT_DENOM 5 #else #define TRANSIENT_DENOM 10 #endif /* * Calculating buffer cache scaling values and reserve space for buffer * headers. This is called during low level kernel initialization and * may be called more then once. We CANNOT write to the memory area * being reserved at this time. */ caddr_t kern_vfs_bio_buffer_alloc(caddr_t v, long physmem_est) { int tuned_nbuf; long maxbuf, maxbuf_sz, buf_sz, biotmap_sz; /* * physmem_est is in pages. Convert it to kilobytes (assumes * PAGE_SIZE is >= 1K) */ physmem_est = physmem_est * (PAGE_SIZE / 1024); maxbcachebuf_adjust(); /* * The nominal buffer size (and minimum KVA allocation) is BKVASIZE. * For the first 64MB of ram nominally allocate sufficient buffers to * cover 1/4 of our ram. Beyond the first 64MB allocate additional * buffers to cover 1/10 of our ram over 64MB. When auto-sizing * the buffer cache we limit the eventual kva reservation to * maxbcache bytes. * * factor represents the 1/4 x ram conversion. */ if (nbuf == 0) { int factor = 4 * BKVASIZE / 1024; nbuf = 50; if (physmem_est > 4096) nbuf += min((physmem_est - 4096) / factor, 65536 / factor); if (physmem_est > 65536) nbuf += min((physmem_est - 65536) * 2 / (factor * 5), 32 * 1024 * 1024 / (factor * 5)); if (maxbcache && nbuf > maxbcache / BKVASIZE) nbuf = maxbcache / BKVASIZE; tuned_nbuf = 1; } else tuned_nbuf = 0; /* XXX Avoid unsigned long overflows later on with maxbufspace. */ maxbuf = (LONG_MAX / 3) / BKVASIZE; if (nbuf > maxbuf) { if (!tuned_nbuf) printf("Warning: nbufs lowered from %d to %ld\n", nbuf, maxbuf); nbuf = maxbuf; } /* * Ideal allocation size for the transient bio submap is 10% * of the maximal space buffer map. This roughly corresponds * to the amount of the buffer mapped for typical UFS load. * * Clip the buffer map to reserve space for the transient * BIOs, if its extent is bigger than 90% (80% on i386) of the * maximum buffer map extent on the platform. * * The fall-back to the maxbuf in case of maxbcache unset, * allows to not trim the buffer KVA for the architectures * with ample KVA space. */ if (bio_transient_maxcnt == 0 && unmapped_buf_allowed) { maxbuf_sz = maxbcache != 0 ? maxbcache : maxbuf * BKVASIZE; buf_sz = (long)nbuf * BKVASIZE; if (buf_sz < maxbuf_sz / TRANSIENT_DENOM * (TRANSIENT_DENOM - 1)) { /* * There is more KVA than memory. Do not * adjust buffer map size, and assign the rest * of maxbuf to transient map. */ biotmap_sz = maxbuf_sz - buf_sz; } else { /* * Buffer map spans all KVA we could afford on * this platform. Give 10% (20% on i386) of * the buffer map to the transient bio map. */ biotmap_sz = buf_sz / TRANSIENT_DENOM; buf_sz -= biotmap_sz; } if (biotmap_sz / INT_MAX > MAXPHYS) bio_transient_maxcnt = INT_MAX; else bio_transient_maxcnt = biotmap_sz / MAXPHYS; /* * Artificially limit to 1024 simultaneous in-flight I/Os * using the transient mapping. */ if (bio_transient_maxcnt > 1024) bio_transient_maxcnt = 1024; if (tuned_nbuf) nbuf = buf_sz / BKVASIZE; } if (nswbuf == 0) { nswbuf = min(nbuf / 4, 256); if (nswbuf < NSWBUF_MIN) nswbuf = NSWBUF_MIN; } /* * Reserve space for the buffer cache buffers */ buf = (void *)v; v = (caddr_t)(buf + nbuf); return(v); } /* Initialize the buffer subsystem. Called before use of any buffers. */ void bufinit(void) { struct buf *bp; int i; KASSERT(maxbcachebuf >= MAXBSIZE, ("maxbcachebuf (%d) must be >= MAXBSIZE (%d)\n", maxbcachebuf, MAXBSIZE)); bq_init(&bqempty, QUEUE_EMPTY, -1, "bufq empty lock"); mtx_init(&rbreqlock, "runningbufspace lock", NULL, MTX_DEF); mtx_init(&bdlock, "buffer daemon lock", NULL, MTX_DEF); mtx_init(&bdirtylock, "dirty buf lock", NULL, MTX_DEF); unmapped_buf = (caddr_t)kva_alloc(MAXPHYS); /* finally, initialize each buffer header and stick on empty q */ for (i = 0; i < nbuf; i++) { bp = &buf[i]; bzero(bp, sizeof *bp); bp->b_flags = B_INVAL; bp->b_rcred = NOCRED; bp->b_wcred = NOCRED; bp->b_qindex = QUEUE_NONE; bp->b_domain = -1; bp->b_subqueue = mp_maxid + 1; bp->b_xflags = 0; bp->b_data = bp->b_kvabase = unmapped_buf; LIST_INIT(&bp->b_dep); BUF_LOCKINIT(bp); bq_insert(&bqempty, bp, false); } /* * maxbufspace is the absolute maximum amount of buffer space we are * allowed to reserve in KVM and in real terms. The absolute maximum * is nominally used by metadata. hibufspace is the nominal maximum * used by most other requests. The differential is required to * ensure that metadata deadlocks don't occur. * * maxbufspace is based on BKVASIZE. Allocating buffers larger then * this may result in KVM fragmentation which is not handled optimally * by the system. XXX This is less true with vmem. We could use * PAGE_SIZE. */ maxbufspace = (long)nbuf * BKVASIZE; hibufspace = lmax(3 * maxbufspace / 4, maxbufspace - maxbcachebuf * 10); lobufspace = (hibufspace / 20) * 19; /* 95% */ bufspacethresh = lobufspace + (hibufspace - lobufspace) / 2; /* * Note: The 16 MiB upper limit for hirunningspace was chosen * arbitrarily and may need further tuning. It corresponds to * 128 outstanding write IO requests (if IO size is 128 KiB), * which fits with many RAID controllers' tagged queuing limits. * The lower 1 MiB limit is the historical upper limit for * hirunningspace. */ hirunningspace = lmax(lmin(roundup(hibufspace / 64, maxbcachebuf), 16 * 1024 * 1024), 1024 * 1024); lorunningspace = roundup((hirunningspace * 2) / 3, maxbcachebuf); /* * Limit the amount of malloc memory since it is wired permanently into * the kernel space. Even though this is accounted for in the buffer * allocation, we don't want the malloced region to grow uncontrolled. * The malloc scheme improves memory utilization significantly on * average (small) directories. */ maxbufmallocspace = hibufspace / 20; /* * Reduce the chance of a deadlock occurring by limiting the number * of delayed-write dirty buffers we allow to stack up. */ hidirtybuffers = nbuf / 4 + 20; dirtybufthresh = hidirtybuffers * 9 / 10; /* * To support extreme low-memory systems, make sure hidirtybuffers * cannot eat up all available buffer space. This occurs when our * minimum cannot be met. We try to size hidirtybuffers to 3/4 our * buffer space assuming BKVASIZE'd buffers. */ while ((long)hidirtybuffers * BKVASIZE > 3 * hibufspace / 4) { hidirtybuffers >>= 1; } lodirtybuffers = hidirtybuffers / 2; /* * lofreebuffers should be sufficient to avoid stalling waiting on * buf headers under heavy utilization. The bufs in per-cpu caches * are counted as free but will be unavailable to threads executing * on other cpus. * * hifreebuffers is the free target for the bufspace daemon. This * should be set appropriately to limit work per-iteration. */ lofreebuffers = MIN((nbuf / 25) + (20 * mp_ncpus), 128 * mp_ncpus); hifreebuffers = (3 * lofreebuffers) / 2; numfreebuffers = nbuf; /* Setup the kva and free list allocators. */ vmem_set_reclaim(buffer_arena, bufkva_reclaim); buf_zone = uma_zcache_create("buf free cache", sizeof(struct buf), NULL, NULL, NULL, NULL, buf_import, buf_release, NULL, 0); /* * Size the clean queue according to the amount of buffer space. * One queue per-256mb up to the max. More queues gives better * concurrency but less accurate LRU. */ buf_domains = MIN(howmany(maxbufspace, 256*1024*1024), BUF_DOMAINS); for (i = 0 ; i < buf_domains; i++) { struct bufdomain *bd; bd = &bdomain[i]; bd_init(bd); bd->bd_freebuffers = nbuf / buf_domains; bd->bd_hifreebuffers = hifreebuffers / buf_domains; bd->bd_lofreebuffers = lofreebuffers / buf_domains; bd->bd_bufspace = 0; bd->bd_maxbufspace = maxbufspace / buf_domains; bd->bd_hibufspace = hibufspace / buf_domains; bd->bd_lobufspace = lobufspace / buf_domains; bd->bd_bufspacethresh = bufspacethresh / buf_domains; bd->bd_numdirtybuffers = 0; bd->bd_hidirtybuffers = hidirtybuffers / buf_domains; bd->bd_lodirtybuffers = lodirtybuffers / buf_domains; bd->bd_dirtybufthresh = dirtybufthresh / buf_domains; /* Don't allow more than 2% of bufs in the per-cpu caches. */ bd->bd_lim = nbuf / buf_domains / 50 / mp_ncpus; } getnewbufcalls = counter_u64_alloc(M_WAITOK); getnewbufrestarts = counter_u64_alloc(M_WAITOK); mappingrestarts = counter_u64_alloc(M_WAITOK); numbufallocfails = counter_u64_alloc(M_WAITOK); notbufdflushes = counter_u64_alloc(M_WAITOK); buffreekvacnt = counter_u64_alloc(M_WAITOK); bufdefragcnt = counter_u64_alloc(M_WAITOK); bufkvaspace = counter_u64_alloc(M_WAITOK); } #ifdef INVARIANTS static inline void vfs_buf_check_mapped(struct buf *bp) { KASSERT(bp->b_kvabase != unmapped_buf, ("mapped buf: b_kvabase was not updated %p", bp)); KASSERT(bp->b_data != unmapped_buf, ("mapped buf: b_data was not updated %p", bp)); KASSERT(bp->b_data < unmapped_buf || bp->b_data >= unmapped_buf + MAXPHYS, ("b_data + b_offset unmapped %p", bp)); } static inline void vfs_buf_check_unmapped(struct buf *bp) { KASSERT(bp->b_data == unmapped_buf, ("unmapped buf: corrupted b_data %p", bp)); } #define BUF_CHECK_MAPPED(bp) vfs_buf_check_mapped(bp) #define BUF_CHECK_UNMAPPED(bp) vfs_buf_check_unmapped(bp) #else #define BUF_CHECK_MAPPED(bp) do {} while (0) #define BUF_CHECK_UNMAPPED(bp) do {} while (0) #endif static int isbufbusy(struct buf *bp) { if (((bp->b_flags & B_INVAL) == 0 && BUF_ISLOCKED(bp)) || ((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI)) return (1); return (0); } /* * Shutdown the system cleanly to prepare for reboot, halt, or power off. */ void bufshutdown(int show_busybufs) { static int first_buf_printf = 1; struct buf *bp; int iter, nbusy, pbusy; #ifndef PREEMPTION int subiter; #endif - /* + /* * Sync filesystems for shutdown */ wdog_kern_pat(WD_LASTVAL); kern_sync(curthread); /* * With soft updates, some buffers that are * written will be remarked as dirty until other * buffers are written. */ for (iter = pbusy = 0; iter < 20; iter++) { nbusy = 0; for (bp = &buf[nbuf]; --bp >= buf; ) if (isbufbusy(bp)) nbusy++; if (nbusy == 0) { if (first_buf_printf) printf("All buffers synced."); break; } if (first_buf_printf) { printf("Syncing disks, buffers remaining... "); first_buf_printf = 0; } printf("%d ", nbusy); if (nbusy < pbusy) iter = 0; pbusy = nbusy; wdog_kern_pat(WD_LASTVAL); kern_sync(curthread); #ifdef PREEMPTION /* * Spin for a while to allow interrupt threads to run. */ DELAY(50000 * iter); #else /* * Context switch several times to allow interrupt * threads to run. */ for (subiter = 0; subiter < 50 * iter; subiter++) { thread_lock(curthread); mi_switch(SW_VOL); DELAY(1000); } #endif } printf("\n"); /* * Count only busy local buffers to prevent forcing * a fsck if we're just a client of a wedged NFS server */ nbusy = 0; for (bp = &buf[nbuf]; --bp >= buf; ) { if (isbufbusy(bp)) { #if 0 /* XXX: This is bogus. We should probably have a BO_REMOTE flag instead */ if (bp->b_dev == NULL) { TAILQ_REMOVE(&mountlist, bp->b_vp->v_mount, mnt_list); continue; } #endif nbusy++; if (show_busybufs > 0) { printf( "%d: buf:%p, vnode:%p, flags:%0x, blkno:%jd, lblkno:%jd, buflock:", nbusy, bp, bp->b_vp, bp->b_flags, (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno); BUF_LOCKPRINTINFO(bp); if (show_busybufs > 1) vn_printf(bp->b_vp, "vnode content: "); } } } if (nbusy) { /* * Failed to sync all blocks. Indicate this and don't * unmount filesystems (thus forcing an fsck on reboot). */ printf("Giving up on %d buffers\n", nbusy); DELAY(5000000); /* 5 seconds */ } else { if (!first_buf_printf) printf("Final sync complete\n"); /* * Unmount filesystems */ if (!KERNEL_PANICKED()) vfs_unmountall(); } swapoff_all(); DELAY(100000); /* wait for console output to finish */ } static void bpmap_qenter(struct buf *bp) { BUF_CHECK_MAPPED(bp); /* * bp->b_data is relative to bp->b_offset, but * bp->b_offset may be offset into the first page. */ bp->b_data = (caddr_t)trunc_page((vm_offset_t)bp->b_data); pmap_qenter((vm_offset_t)bp->b_data, bp->b_pages, bp->b_npages); bp->b_data = (caddr_t)((vm_offset_t)bp->b_data | (vm_offset_t)(bp->b_offset & PAGE_MASK)); } static inline struct bufdomain * bufdomain(struct buf *bp) { return (&bdomain[bp->b_domain]); } static struct bufqueue * bufqueue(struct buf *bp) { switch (bp->b_qindex) { case QUEUE_NONE: /* FALLTHROUGH */ case QUEUE_SENTINEL: return (NULL); case QUEUE_EMPTY: return (&bqempty); case QUEUE_DIRTY: return (&bufdomain(bp)->bd_dirtyq); case QUEUE_CLEAN: return (&bufdomain(bp)->bd_subq[bp->b_subqueue]); default: break; } panic("bufqueue(%p): Unhandled type %d\n", bp, bp->b_qindex); } /* * Return the locked bufqueue that bp is a member of. */ static struct bufqueue * bufqueue_acquire(struct buf *bp) { struct bufqueue *bq, *nbq; /* * bp can be pushed from a per-cpu queue to the * cleanq while we're waiting on the lock. Retry * if the queues don't match. */ bq = bufqueue(bp); BQ_LOCK(bq); for (;;) { nbq = bufqueue(bp); if (bq == nbq) break; BQ_UNLOCK(bq); BQ_LOCK(nbq); bq = nbq; } return (bq); } /* * binsfree: * * Insert the buffer into the appropriate free list. Requires a * locked buffer on entry and buffer is unlocked before return. */ static void binsfree(struct buf *bp, int qindex) { struct bufdomain *bd; struct bufqueue *bq; KASSERT(qindex == QUEUE_CLEAN || qindex == QUEUE_DIRTY, ("binsfree: Invalid qindex %d", qindex)); BUF_ASSERT_XLOCKED(bp); /* * Handle delayed bremfree() processing. */ if (bp->b_flags & B_REMFREE) { if (bp->b_qindex == qindex) { bp->b_flags |= B_REUSE; bp->b_flags &= ~B_REMFREE; BUF_UNLOCK(bp); return; } bq = bufqueue_acquire(bp); bq_remove(bq, bp); BQ_UNLOCK(bq); } bd = bufdomain(bp); if (qindex == QUEUE_CLEAN) { if (bd->bd_lim != 0) bq = &bd->bd_subq[PCPU_GET(cpuid)]; else bq = bd->bd_cleanq; } else bq = &bd->bd_dirtyq; bq_insert(bq, bp, true); } /* * buf_free: * * Free a buffer to the buf zone once it no longer has valid contents. */ static void buf_free(struct buf *bp) { if (bp->b_flags & B_REMFREE) bremfreef(bp); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 1"); if (bp->b_rcred != NOCRED) { crfree(bp->b_rcred); bp->b_rcred = NOCRED; } if (bp->b_wcred != NOCRED) { crfree(bp->b_wcred); bp->b_wcred = NOCRED; } if (!LIST_EMPTY(&bp->b_dep)) buf_deallocate(bp); bufkva_free(bp); atomic_add_int(&bufdomain(bp)->bd_freebuffers, 1); BUF_UNLOCK(bp); uma_zfree(buf_zone, bp); } /* * buf_import: * * Import bufs into the uma cache from the buf list. The system still * expects a static array of bufs and much of the synchronization * around bufs assumes type stable storage. As a result, UMA is used * only as a per-cpu cache of bufs still maintained on a global list. */ static int buf_import(void *arg, void **store, int cnt, int domain, int flags) { struct buf *bp; int i; BQ_LOCK(&bqempty); for (i = 0; i < cnt; i++) { bp = TAILQ_FIRST(&bqempty.bq_queue); if (bp == NULL) break; bq_remove(&bqempty, bp); store[i] = bp; } BQ_UNLOCK(&bqempty); return (i); } /* * buf_release: * * Release bufs from the uma cache back to the buffer queues. */ static void buf_release(void *arg, void **store, int cnt) { struct bufqueue *bq; struct buf *bp; int i; bq = &bqempty; BQ_LOCK(bq); for (i = 0; i < cnt; i++) { bp = store[i]; /* Inline bq_insert() to batch locking. */ TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); bp->b_flags &= ~(B_AGE | B_REUSE); bq->bq_len++; bp->b_qindex = bq->bq_index; } BQ_UNLOCK(bq); } /* * buf_alloc: * * Allocate an empty buffer header. */ static struct buf * buf_alloc(struct bufdomain *bd) { struct buf *bp; int freebufs; /* * We can only run out of bufs in the buf zone if the average buf * is less than BKVASIZE. In this case the actual wait/block will * come from buf_reycle() failing to flush one of these small bufs. */ bp = NULL; freebufs = atomic_fetchadd_int(&bd->bd_freebuffers, -1); if (freebufs > 0) bp = uma_zalloc(buf_zone, M_NOWAIT); if (bp == NULL) { atomic_add_int(&bd->bd_freebuffers, 1); bufspace_daemon_wakeup(bd); counter_u64_add(numbufallocfails, 1); return (NULL); } /* * Wake-up the bufspace daemon on transition below threshold. */ if (freebufs == bd->bd_lofreebuffers) bufspace_daemon_wakeup(bd); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) panic("getnewbuf_empty: Locked buf %p on free queue.", bp); - + KASSERT(bp->b_vp == NULL, ("bp: %p still has vnode %p.", bp, bp->b_vp)); KASSERT((bp->b_flags & (B_DELWRI | B_NOREUSE)) == 0, ("invalid buffer %p flags %#x", bp, bp->b_flags)); KASSERT((bp->b_xflags & (BX_VNCLEAN|BX_VNDIRTY)) == 0, ("bp: %p still on a buffer list. xflags %X", bp, bp->b_xflags)); KASSERT(bp->b_npages == 0, ("bp: %p still has %d vm pages\n", bp, bp->b_npages)); KASSERT(bp->b_kvasize == 0, ("bp: %p still has kva\n", bp)); KASSERT(bp->b_bufsize == 0, ("bp: %p still has bufspace\n", bp)); bp->b_domain = BD_DOMAIN(bd); bp->b_flags = 0; bp->b_ioflags = 0; bp->b_xflags = 0; bp->b_vflags = 0; bp->b_vp = NULL; bp->b_blkno = bp->b_lblkno = 0; bp->b_offset = NOOFFSET; bp->b_iodone = 0; bp->b_error = 0; bp->b_resid = 0; bp->b_bcount = 0; bp->b_npages = 0; bp->b_dirtyoff = bp->b_dirtyend = 0; bp->b_bufobj = NULL; bp->b_data = bp->b_kvabase = unmapped_buf; bp->b_fsprivate1 = NULL; bp->b_fsprivate2 = NULL; bp->b_fsprivate3 = NULL; LIST_INIT(&bp->b_dep); return (bp); } /* * buf_recycle: * * Free a buffer from the given bufqueue. kva controls whether the * freed buf must own some kva resources. This is used for * defragmenting. */ static int buf_recycle(struct bufdomain *bd, bool kva) { struct bufqueue *bq; struct buf *bp, *nbp; if (kva) counter_u64_add(bufdefragcnt, 1); nbp = NULL; bq = bd->bd_cleanq; BQ_LOCK(bq); KASSERT(BQ_LOCKPTR(bq) == BD_LOCKPTR(bd), ("buf_recycle: Locks don't match")); nbp = TAILQ_FIRST(&bq->bq_queue); /* * Run scan, possibly freeing data and/or kva mappings on the fly * depending. */ while ((bp = nbp) != NULL) { /* * Calculate next bp (we can only use it if we do not * release the bqlock). */ nbp = TAILQ_NEXT(bp, b_freelist); /* * If we are defragging then we need a buffer with * some kva to reclaim. */ if (kva && bp->b_kvasize == 0) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) continue; /* * Implement a second chance algorithm for frequently * accessed buffers. */ if ((bp->b_flags & B_REUSE) != 0) { TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); bp->b_flags &= ~B_REUSE; BUF_UNLOCK(bp); continue; } /* * Skip buffers with background writes in progress. */ if ((bp->b_vflags & BV_BKGRDINPROG) != 0) { BUF_UNLOCK(bp); continue; } KASSERT(bp->b_qindex == QUEUE_CLEAN, ("buf_recycle: inconsistent queue %d bp %p", bp->b_qindex, bp)); KASSERT(bp->b_domain == BD_DOMAIN(bd), ("getnewbuf: queue domain %d doesn't match request %d", bp->b_domain, (int)BD_DOMAIN(bd))); /* * NOTE: nbp is now entirely invalid. We can only restart * the scan from this point on. */ bq_remove(bq, bp); BQ_UNLOCK(bq); /* * Requeue the background write buffer with error and * restart the scan. */ if ((bp->b_vflags & BV_BKGRDERR) != 0) { bqrelse(bp); BQ_LOCK(bq); nbp = TAILQ_FIRST(&bq->bq_queue); continue; } bp->b_flags |= B_INVAL; brelse(bp); return (0); } bd->bd_wanted = 1; BQ_UNLOCK(bq); return (ENOBUFS); } /* * bremfree: * * Mark the buffer for removal from the appropriate free list. - * + * */ void bremfree(struct buf *bp) { CTR3(KTR_BUF, "bremfree(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT((bp->b_flags & B_REMFREE) == 0, ("bremfree: buffer %p already marked for delayed removal.", bp)); KASSERT(bp->b_qindex != QUEUE_NONE, ("bremfree: buffer %p not on a queue.", bp)); BUF_ASSERT_XLOCKED(bp); bp->b_flags |= B_REMFREE; } /* * bremfreef: * * Force an immediate removal from a free list. Used only in nfs when * it abuses the b_freelist pointer. */ void bremfreef(struct buf *bp) { struct bufqueue *bq; bq = bufqueue_acquire(bp); bq_remove(bq, bp); BQ_UNLOCK(bq); } static void bq_init(struct bufqueue *bq, int qindex, int subqueue, const char *lockname) { mtx_init(&bq->bq_lock, lockname, NULL, MTX_DEF); TAILQ_INIT(&bq->bq_queue); bq->bq_len = 0; bq->bq_index = qindex; bq->bq_subqueue = subqueue; } static void bd_init(struct bufdomain *bd) { int i; bd->bd_cleanq = &bd->bd_subq[mp_maxid + 1]; bq_init(bd->bd_cleanq, QUEUE_CLEAN, mp_maxid + 1, "bufq clean lock"); bq_init(&bd->bd_dirtyq, QUEUE_DIRTY, -1, "bufq dirty lock"); for (i = 0; i <= mp_maxid; i++) bq_init(&bd->bd_subq[i], QUEUE_CLEAN, i, "bufq clean subqueue lock"); mtx_init(&bd->bd_run_lock, "bufspace daemon run lock", NULL, MTX_DEF); } /* * bq_remove: * * Removes a buffer from the free list, must be called with the * correct qlock held. */ static void bq_remove(struct bufqueue *bq, struct buf *bp) { CTR3(KTR_BUF, "bq_remove(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_qindex != QUEUE_NONE, ("bq_remove: buffer %p not on a queue.", bp)); KASSERT(bufqueue(bp) == bq, ("bq_remove: Remove buffer %p from wrong queue.", bp)); BQ_ASSERT_LOCKED(bq); if (bp->b_qindex != QUEUE_EMPTY) { BUF_ASSERT_XLOCKED(bp); } KASSERT(bq->bq_len >= 1, ("queue %d underflow", bp->b_qindex)); TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); bq->bq_len--; bp->b_qindex = QUEUE_NONE; bp->b_flags &= ~(B_REMFREE | B_REUSE); } static void bd_flush(struct bufdomain *bd, struct bufqueue *bq) { struct buf *bp; BQ_ASSERT_LOCKED(bq); if (bq != bd->bd_cleanq) { BD_LOCK(bd); while ((bp = TAILQ_FIRST(&bq->bq_queue)) != NULL) { TAILQ_REMOVE(&bq->bq_queue, bp, b_freelist); TAILQ_INSERT_TAIL(&bd->bd_cleanq->bq_queue, bp, b_freelist); bp->b_subqueue = bd->bd_cleanq->bq_subqueue; } bd->bd_cleanq->bq_len += bq->bq_len; bq->bq_len = 0; } if (bd->bd_wanted) { bd->bd_wanted = 0; wakeup(&bd->bd_wanted); } if (bq != bd->bd_cleanq) BD_UNLOCK(bd); } static int bd_flushall(struct bufdomain *bd) { struct bufqueue *bq; int flushed; int i; if (bd->bd_lim == 0) return (0); flushed = 0; for (i = 0; i <= mp_maxid; i++) { bq = &bd->bd_subq[i]; if (bq->bq_len == 0) continue; BQ_LOCK(bq); bd_flush(bd, bq); BQ_UNLOCK(bq); flushed++; } return (flushed); } static void bq_insert(struct bufqueue *bq, struct buf *bp, bool unlock) { struct bufdomain *bd; if (bp->b_qindex != QUEUE_NONE) panic("bq_insert: free buffer %p onto another queue?", bp); bd = bufdomain(bp); if (bp->b_flags & B_AGE) { /* Place this buf directly on the real queue. */ if (bq->bq_index == QUEUE_CLEAN) bq = bd->bd_cleanq; BQ_LOCK(bq); TAILQ_INSERT_HEAD(&bq->bq_queue, bp, b_freelist); } else { BQ_LOCK(bq); TAILQ_INSERT_TAIL(&bq->bq_queue, bp, b_freelist); } bp->b_flags &= ~(B_AGE | B_REUSE); bq->bq_len++; bp->b_qindex = bq->bq_index; bp->b_subqueue = bq->bq_subqueue; /* * Unlock before we notify so that we don't wakeup a waiter that * fails a trylock on the buf and sleeps again. */ if (unlock) BUF_UNLOCK(bp); if (bp->b_qindex == QUEUE_CLEAN) { /* * Flush the per-cpu queue and notify any waiters. */ if (bd->bd_wanted || (bq != bd->bd_cleanq && bq->bq_len >= bd->bd_lim)) bd_flush(bd, bq); } BQ_UNLOCK(bq); } /* * bufkva_free: * * Free the kva allocation for a buffer. * */ static void bufkva_free(struct buf *bp) { #ifdef INVARIANTS if (bp->b_kvasize == 0) { KASSERT(bp->b_kvabase == unmapped_buf && bp->b_data == unmapped_buf, ("Leaked KVA space on %p", bp)); } else if (buf_mapped(bp)) BUF_CHECK_MAPPED(bp); else BUF_CHECK_UNMAPPED(bp); #endif if (bp->b_kvasize == 0) return; vmem_free(buffer_arena, (vm_offset_t)bp->b_kvabase, bp->b_kvasize); counter_u64_add(bufkvaspace, -bp->b_kvasize); counter_u64_add(buffreekvacnt, 1); bp->b_data = bp->b_kvabase = unmapped_buf; bp->b_kvasize = 0; } /* * bufkva_alloc: * * Allocate the buffer KVA and set b_kvasize and b_kvabase. */ static int bufkva_alloc(struct buf *bp, int maxsize, int gbflags) { vm_offset_t addr; int error; KASSERT((gbflags & GB_UNMAPPED) == 0 || (gbflags & GB_KVAALLOC) != 0, ("Invalid gbflags 0x%x in %s", gbflags, __func__)); bufkva_free(bp); addr = 0; error = vmem_alloc(buffer_arena, maxsize, M_BESTFIT | M_NOWAIT, &addr); if (error != 0) { /* * Buffer map is too fragmented. Request the caller * to defragment the map. */ return (error); } bp->b_kvabase = (caddr_t)addr; bp->b_kvasize = maxsize; counter_u64_add(bufkvaspace, bp->b_kvasize); if ((gbflags & GB_UNMAPPED) != 0) { bp->b_data = unmapped_buf; BUF_CHECK_UNMAPPED(bp); } else { bp->b_data = bp->b_kvabase; BUF_CHECK_MAPPED(bp); } return (0); } /* * bufkva_reclaim: * * Reclaim buffer kva by freeing buffers holding kva. This is a vmem * callback that fires to avoid returning failure. */ static void bufkva_reclaim(vmem_t *vmem, int flags) { bool done; int q; int i; done = false; for (i = 0; i < 5; i++) { for (q = 0; q < buf_domains; q++) if (buf_recycle(&bdomain[q], true) != 0) done = true; if (done) break; } return; } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. We must * clear BIO_ERROR and B_INVAL prior to initiating I/O . If B_CACHE is set, * the buffer is valid and we do not have to do anything. */ static void breada(struct vnode * vp, daddr_t * rablkno, int * rabsize, int cnt, struct ucred * cred, int flags, void (*ckhashfunc)(struct buf *)) { struct buf *rabp; struct thread *td; int i; td = curthread; for (i = 0; i < cnt; i++, rablkno++, rabsize++) { if (inmem(vp, *rablkno)) continue; rabp = getblk(vp, *rablkno, *rabsize, 0, 0, 0); if ((rabp->b_flags & B_CACHE) != 0) { brelse(rabp); continue; } #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_buf(curproc, rabp, 0); PROC_UNLOCK(curproc); } #endif /* RACCT */ td->td_ru.ru_inblock++; rabp->b_flags |= B_ASYNC; rabp->b_flags &= ~B_INVAL; if ((flags & GB_CKHASH) != 0) { rabp->b_flags |= B_CKHASH; rabp->b_ckhashcalc = ckhashfunc; } rabp->b_ioflags &= ~BIO_ERROR; rabp->b_iocmd = BIO_READ; if (rabp->b_rcred == NOCRED && cred != NOCRED) rabp->b_rcred = crhold(cred); vfs_busy_pages(rabp, 0); BUF_KERNPROC(rabp); rabp->b_iooffset = dbtob(rabp->b_blkno); bstrategy(rabp); } } /* * Entry point for bread() and breadn() via #defines in sys/buf.h. * * Get a buffer with the specified data. Look in the cache first. We * must clear BIO_ERROR and B_INVAL prior to initiating I/O. If B_CACHE * is set, the buffer is valid and we do not have to do anything, see * getblk(). Also starts asynchronous I/O on read-ahead blocks. * * Always return a NULL buffer pointer (in bpp) when returning an error. * * The blkno parameter is the logical block being requested. Normally * the mapping of logical block number to disk block address is done * by calling VOP_BMAP(). However, if the mapping is already known, the * disk block address can be passed using the dblkno parameter. If the * disk block address is not known, then the same value should be passed * for blkno and dblkno. */ int breadn_flags(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, daddr_t *rablkno, int *rabsize, int cnt, struct ucred *cred, int flags, void (*ckhashfunc)(struct buf *), struct buf **bpp) { struct buf *bp; struct thread *td; int error, readwait, rv; CTR3(KTR_BUF, "breadn(%p, %jd, %d)", vp, blkno, size); td = curthread; /* * Can only return NULL if GB_LOCK_NOWAIT or GB_SPARSE flags * are specified. */ error = getblkx(vp, blkno, dblkno, size, 0, 0, flags, &bp); if (error != 0) { *bpp = NULL; return (error); } KASSERT(blkno == bp->b_lblkno, ("getblkx returned buffer for blkno %jd instead of blkno %jd", (intmax_t)bp->b_lblkno, (intmax_t)blkno)); flags &= ~GB_NOSPARSE; *bpp = bp; /* * If not found in cache, do some I/O */ readwait = 0; if ((bp->b_flags & B_CACHE) == 0) { #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, bp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; bp->b_iocmd = BIO_READ; bp->b_flags &= ~B_INVAL; if ((flags & GB_CKHASH) != 0) { bp->b_flags |= B_CKHASH; bp->b_ckhashcalc = ckhashfunc; } if ((flags & GB_CVTENXIO) != 0) bp->b_xflags |= BX_CVTENXIO; bp->b_ioflags &= ~BIO_ERROR; if (bp->b_rcred == NOCRED && cred != NOCRED) bp->b_rcred = crhold(cred); vfs_busy_pages(bp, 0); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); ++readwait; } /* * Attempt to initiate asynchronous I/O on read-ahead blocks. */ breada(vp, rablkno, rabsize, cnt, cred, flags, ckhashfunc); rv = 0; if (readwait) { rv = bufwait(bp); if (rv != 0) { brelse(bp); *bpp = NULL; } } return (rv); } /* * Write, release buffer on completion. (Done by iodone * if async). Do not bother writing anything if the buffer * is invalid. * * Note that we set B_CACHE here, indicating that buffer is * fully valid and thus cacheable. This is true even of NFS * now so we set it generally. This could be set either here * or in biodone() since the I/O is synchronous. We put it * here. */ int bufwrite(struct buf *bp) { int oldflags; struct vnode *vp; long space; int vp_md; CTR3(KTR_BUF, "bufwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if ((bp->b_bufobj->bo_flag & BO_DEAD) != 0) { bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_CACHE; brelse(bp); return (ENXIO); } if (bp->b_flags & B_INVAL) { brelse(bp); return (0); } if (bp->b_flags & B_BARRIER) atomic_add_long(&barrierwrites, 1); oldflags = bp->b_flags; KASSERT(!(bp->b_vflags & BV_BKGRDINPROG), ("FFS background buffer should not get here %p", bp)); vp = bp->b_vp; if (vp) vp_md = vp->v_vflag & VV_MD; else vp_md = 0; /* * Mark the buffer clean. Increment the bufobj write count * before bundirty() call, to prevent other thread from seeing * empty dirty list and zero counter for writes in progress, * falsely indicating that the bufobj is clean. */ bufobj_wref(bp->b_bufobj); bundirty(bp); bp->b_flags &= ~B_DONE; bp->b_ioflags &= ~BIO_ERROR; bp->b_flags |= B_CACHE; bp->b_iocmd = BIO_WRITE; vfs_busy_pages(bp, 1); /* * Normal bwrites pipeline writes */ bp->b_runningbufspace = bp->b_bufsize; space = atomic_fetchadd_long(&runningbufspace, bp->b_runningbufspace); #ifdef RACCT if (racct_enable) { PROC_LOCK(curproc); racct_add_buf(curproc, bp, 1); PROC_UNLOCK(curproc); } #endif /* RACCT */ curthread->td_ru.ru_oublock++; if (oldflags & B_ASYNC) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); buf_track(bp, __func__); bstrategy(bp); if ((oldflags & B_ASYNC) == 0) { int rtval = bufwait(bp); brelse(bp); return (rtval); } else if (space > hirunningspace) { /* * don't allow the async write to saturate the I/O * system. We will not deadlock here because * we are blocking waiting for I/O that is already in-progress * to complete. We do not block here if it is the update * or syncer daemon trying to clean up as that can lead * to deadlock. */ if ((curthread->td_pflags & TDP_NORUNNINGBUF) == 0 && !vp_md) waitrunningbufspace(); } return (0); } void bufbdflush(struct bufobj *bo, struct buf *bp) { struct buf *nbp; if (bo->bo_dirty.bv_cnt > dirtybufthresh + 10) { (void) VOP_FSYNC(bp->b_vp, MNT_NOWAIT, curthread); altbufferflushes++; } else if (bo->bo_dirty.bv_cnt > dirtybufthresh) { BO_LOCK(bo); /* * Try to find a buffer to flush. */ TAILQ_FOREACH(nbp, &bo->bo_dirty.bv_hd, b_bobufs) { if ((nbp->b_vflags & BV_BKGRDINPROG) || BUF_LOCK(nbp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) continue; if (bp == nbp) panic("bdwrite: found ourselves"); BO_UNLOCK(bo); /* Don't countdeps with the bo lock held. */ if (buf_countdeps(nbp, 0)) { BO_LOCK(bo); BUF_UNLOCK(nbp); continue; } if (nbp->b_flags & B_CLUSTEROK) { vfs_bio_awrite(nbp); } else { bremfree(nbp); bawrite(nbp); } dirtybufferflushes++; break; } if (nbp == NULL) BO_UNLOCK(bo); } } /* * Delayed write. (Buffer is marked dirty). Do not bother writing * anything if the buffer is marked invalid. * * Note that since the buffer must be completely valid, we can safely * set B_CACHE. In fact, we have to set B_CACHE here rather then in * biodone() in order to prevent getblk from writing the buffer * out synchronously. */ void bdwrite(struct buf *bp) { struct thread *td = curthread; struct vnode *vp; struct bufobj *bo; CTR3(KTR_BUF, "bdwrite(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT((bp->b_flags & B_BARRIER) == 0, ("Barrier request in delayed write %p", bp)); if (bp->b_flags & B_INVAL) { brelse(bp); return; } /* * If we have too many dirty buffers, don't create any more. * If we are wildly over our limit, then force a complete * cleanup. Otherwise, just keep the situation from getting * out of control. Note that we have to avoid a recursive * disaster and not try to clean up after our own cleanup! */ vp = bp->b_vp; bo = bp->b_bufobj; if ((td->td_pflags & (TDP_COWINPROGRESS|TDP_INBDFLUSH)) == 0) { td->td_pflags |= TDP_INBDFLUSH; BO_BDFLUSH(bo, bp); td->td_pflags &= ~TDP_INBDFLUSH; } else recursiveflushes++; bdirty(bp); /* * Set B_CACHE, indicating that the buffer is fully valid. This is * true even of NFS now. */ bp->b_flags |= B_CACHE; /* * This bmap keeps the system from needing to do the bmap later, * perhaps when the system is attempting to do a sync. Since it * is likely that the indirect block -- or whatever other datastructure * that the filesystem needs is still in memory now, it is a good * thing to do this. Note also, that if the pageout daemon is * requesting a sync -- there might not be enough memory to do * the bmap then... So, this is important to do. */ if (vp->v_type != VCHR && bp->b_lblkno == bp->b_blkno) { VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } buf_track(bp, __func__); /* * Set the *dirty* buffer range based upon the VM system dirty * pages. * * Mark the buffer pages as clean. We need to do this here to * satisfy the vnode_pager and the pageout daemon, so that it * thinks that the pages have been "cleaned". Note that since * the pages are in a delayed write buffer -- the VFS layer * "will" see that the pages get written out on the next sync, * or perhaps the cluster will be completed. */ vfs_clean_pages_dirty_buf(bp); bqrelse(bp); /* * note: we cannot initiate I/O from a bdwrite even if we wanted to, * due to the softdep code. */ } /* * bdirty: * * Turn buffer into delayed write request. We must clear BIO_READ and * B_RELBUF, and we must set B_DELWRI. We reassign the buffer to * itself to properly update it in the dirty/clean lists. We mark it * B_DONE to ensure that any asynchronization of the buffer properly * clears B_DONE ( else a panic will occur later ). * * bdirty() is kinda like bdwrite() - we have to clear B_INVAL which * might have been set pre-getblk(). Unlike bwrite/bdwrite, bdirty() * should only be called if the buffer is known-good. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. * * The buffer must be on QUEUE_NONE. */ void bdirty(struct buf *bp) { CTR3(KTR_BUF, "bdirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex)); bp->b_flags &= ~(B_RELBUF); bp->b_iocmd = BIO_WRITE; if ((bp->b_flags & B_DELWRI) == 0) { bp->b_flags |= /* XXX B_DONE | */ B_DELWRI; reassignbuf(bp); bdirtyadd(bp); } } /* * bundirty: * * Clear B_DELWRI for buffer. * * Since the buffer is not on a queue, we do not update the numfreebuffers * count. - * + * * The buffer must be on QUEUE_NONE. */ void bundirty(struct buf *bp) { CTR3(KTR_BUF, "bundirty(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); KASSERT(bp->b_flags & B_REMFREE || bp->b_qindex == QUEUE_NONE, ("bundirty: buffer %p still on queue %d", bp, bp->b_qindex)); if (bp->b_flags & B_DELWRI) { bp->b_flags &= ~B_DELWRI; reassignbuf(bp); bdirtysub(bp); } /* * Since it is now being written, we can clear its deferred write flag. */ bp->b_flags &= ~B_DEFERRED; } /* * bawrite: * * Asynchronous write. Start output on a buffer, but do not wait for * it to complete. The buffer is released when the output completes. * * bwrite() ( or the VOP routine anyway ) is responsible for handling * B_INVAL buffers. Not us. */ void bawrite(struct buf *bp) { bp->b_flags |= B_ASYNC; (void) bwrite(bp); } /* * babarrierwrite: * * Asynchronous barrier write. Start output on a buffer, but do not * wait for it to complete. Place a write barrier after this write so * that this buffer and all buffers written before it are committed to * the disk before any buffers written after this write are committed * to the disk. The buffer is released when the output completes. */ void babarrierwrite(struct buf *bp) { bp->b_flags |= B_ASYNC | B_BARRIER; (void) bwrite(bp); } /* * bbarrierwrite: * * Synchronous barrier write. Start output on a buffer and wait for * it to complete. Place a write barrier after this write so that * this buffer and all buffers written before it are committed to * the disk before any buffers written after this write are committed * to the disk. The buffer is released when the output completes. */ int bbarrierwrite(struct buf *bp) { bp->b_flags |= B_BARRIER; return (bwrite(bp)); } /* * bwillwrite: * * Called prior to the locking of any vnodes when we are expecting to * write. We do not want to starve the buffer cache with too many * dirty buffers so we block here. By blocking prior to the locking * of any vnodes we attempt to avoid the situation where a locked vnode * prevents the various system daemons from flushing related buffers. */ void bwillwrite(void) { if (buf_dirty_count_severe()) { mtx_lock(&bdirtylock); while (buf_dirty_count_severe()) { bdirtywait = 1; msleep(&bdirtywait, &bdirtylock, (PRIBIO + 4), "flswai", 0); } mtx_unlock(&bdirtylock); } } /* * Return true if we have too many dirty buffers. */ int buf_dirty_count_severe(void) { return (!BIT_EMPTY(BUF_DOMAINS, &bdhidirty)); } /* * brelse: * * Release a busy buffer and, if requested, free its resources. The * buffer will be stashed in the appropriate bufqueue[] allowing it * to be accessed later as a cache entity or reused for other purposes. */ void brelse(struct buf *bp) { struct mount *v_mnt; int qindex; /* * Many functions erroneously call brelse with a NULL bp under rare * error conditions. Simply return when called with a NULL bp. */ if (bp == NULL) return; CTR3(KTR_BUF, "brelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); KASSERT((bp->b_flags & B_VMIO) != 0 || (bp->b_flags & B_NOREUSE) == 0, ("brelse: non-VMIO buffer marked NOREUSE")); if (BUF_LOCKRECURSED(bp)) { /* * Do not process, in particular, do not handle the * B_INVAL/B_RELBUF and do not release to free list. */ BUF_UNLOCK(bp); return; } if (bp->b_flags & B_MANAGED) { bqrelse(bp); return; } if ((bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { BO_LOCK(bp->b_bufobj); bp->b_vflags &= ~BV_BKGRDERR; BO_UNLOCK(bp->b_bufobj); bdirty(bp); } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && (bp->b_flags & B_INVALONERR)) { /* * Forced invalidation of dirty buffer contents, to be used * after a failed write in the rare case that the loss of the * contents is acceptable. The buffer is invalidated and * freed. */ bp->b_flags |= B_INVAL | B_RELBUF | B_NOCACHE; bp->b_flags &= ~(B_ASYNC | B_CACHE); } if (bp->b_iocmd == BIO_WRITE && (bp->b_ioflags & BIO_ERROR) && (bp->b_error != ENXIO || !LIST_EMPTY(&bp->b_dep)) && !(bp->b_flags & B_INVAL)) { /* * Failed write, redirty. All errors except ENXIO (which * means the device is gone) are treated as being * transient. * * XXX Treating EIO as transient is not correct; the * contract with the local storage device drivers is that * they will only return EIO once the I/O is no longer * retriable. Network I/O also respects this through the * guarantees of TCP and/or the internal retries of NFS. * ENOMEM might be transient, but we also have no way of * knowing when its ok to retry/reschedule. In general, * this entire case should be made obsolete through better * error handling/recovery and resource scheduling. * * Do this also for buffers that failed with ENXIO, but have * non-empty dependencies - the soft updates code might need * to access the buffer to untangle them. * * Must clear BIO_ERROR to prevent pages from being scrapped. */ bp->b_ioflags &= ~BIO_ERROR; bdirty(bp); } else if ((bp->b_flags & (B_NOCACHE | B_INVAL)) || (bp->b_ioflags & BIO_ERROR) || (bp->b_bufsize <= 0)) { /* * Either a failed read I/O, or we were asked to free or not * cache the buffer, or we failed to write to a device that's * no longer present. */ bp->b_flags |= B_INVAL; if (!LIST_EMPTY(&bp->b_dep)) buf_deallocate(bp); if (bp->b_flags & B_DELWRI) bdirtysub(bp); bp->b_flags &= ~(B_DELWRI | B_CACHE); if ((bp->b_flags & B_VMIO) == 0) { allocbuf(bp, 0); if (bp->b_vp) brelvp(bp); } } /* * We must clear B_RELBUF if B_DELWRI is set. If vfs_vmio_truncate() * is called with B_DELWRI set, the underlying pages may wind up * getting freed causing a previous write (bdwrite()) to get 'lost' * because pages associated with a B_DELWRI bp are marked clean. * * We still allow the B_INVAL case to call vfs_vmio_truncate(), even * if B_DELWRI is set. */ if (bp->b_flags & B_DELWRI) bp->b_flags &= ~B_RELBUF; /* * VMIO buffer rundown. It is not very necessary to keep a VMIO buffer * constituted, not even NFS buffers now. Two flags effect this. If * B_INVAL, the struct buf is invalidated but the VM object is kept * around ( i.e. so it is trivial to reconstitute the buffer later ). * * If BIO_ERROR or B_NOCACHE is set, pages in the VM object will be * invalidated. BIO_ERROR cannot be set for a failed write unless the * buffer is also B_INVAL because it hits the re-dirtying code above. * * Normally we can do this whether a buffer is B_DELWRI or not. If * the buffer is an NFS buffer, it is tracking piecemeal writes or * the commit state and we cannot afford to lose the buffer. If the * buffer has a background write in progress, we need to keep it * around to prevent it from being reconstituted and starting a second * background write. */ v_mnt = bp->b_vp != NULL ? bp->b_vp->v_mount : NULL; if ((bp->b_flags & B_VMIO) && (bp->b_flags & B_NOCACHE || (bp->b_ioflags & BIO_ERROR && bp->b_iocmd == BIO_READ)) && (v_mnt == NULL || (v_mnt->mnt_vfc->vfc_flags & VFCF_NETWORK) == 0 || vn_isdisk(bp->b_vp, NULL) || (bp->b_flags & B_DELWRI) == 0)) { vfs_vmio_invalidate(bp); allocbuf(bp, 0); } if ((bp->b_flags & (B_INVAL | B_RELBUF)) != 0 || (bp->b_flags & (B_DELWRI | B_NOREUSE)) == B_NOREUSE) { allocbuf(bp, 0); bp->b_flags &= ~B_NOREUSE; if (bp->b_vp != NULL) brelvp(bp); } - + /* * If the buffer has junk contents signal it and eventually * clean up B_DELWRI and diassociate the vnode so that gbincore() * doesn't find it. */ if (bp->b_bufsize == 0 || (bp->b_ioflags & BIO_ERROR) != 0 || (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) != 0) bp->b_flags |= B_INVAL; if (bp->b_flags & B_INVAL) { if (bp->b_flags & B_DELWRI) bundirty(bp); if (bp->b_vp) brelvp(bp); } buf_track(bp, __func__); /* buffers with no memory */ if (bp->b_bufsize == 0) { buf_free(bp); return; } /* buffers with junk contents */ if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF) || (bp->b_ioflags & BIO_ERROR)) { bp->b_xflags &= ~(BX_BKGRDWRITE | BX_ALTDATA); if (bp->b_vflags & BV_BKGRDINPROG) panic("losing buffer 2"); qindex = QUEUE_CLEAN; bp->b_flags |= B_AGE; /* remaining buffers */ } else if (bp->b_flags & B_DELWRI) qindex = QUEUE_DIRTY; else qindex = QUEUE_CLEAN; if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("brelse: not dirty"); bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_RELBUF | B_DIRECT); bp->b_xflags &= ~(BX_CVTENXIO); /* binsfree unlocks bp. */ binsfree(bp, qindex); } /* * Release a buffer back to the appropriate queue but do not try to free * it. The buffer is expected to be used again soon. * * bqrelse() is used by bdwrite() to requeue a delayed write, and used by * biodone() to requeue an async I/O on completion. It is also used when * known good buffers need to be requeued but we think we may need the data * again soon. * * XXX we should be able to leave the B_RELBUF hint set on completion. */ void bqrelse(struct buf *bp) { int qindex; CTR3(KTR_BUF, "bqrelse(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp)); qindex = QUEUE_NONE; if (BUF_LOCKRECURSED(bp)) { /* do not release to free list */ BUF_UNLOCK(bp); return; } bp->b_flags &= ~(B_ASYNC | B_NOCACHE | B_AGE | B_RELBUF); bp->b_xflags &= ~(BX_CVTENXIO); if (bp->b_flags & B_MANAGED) { if (bp->b_flags & B_REMFREE) bremfreef(bp); goto out; } /* buffers with stale but valid contents */ if ((bp->b_flags & B_DELWRI) != 0 || (bp->b_vflags & (BV_BKGRDINPROG | BV_BKGRDERR)) == BV_BKGRDERR) { BO_LOCK(bp->b_bufobj); bp->b_vflags &= ~BV_BKGRDERR; BO_UNLOCK(bp->b_bufobj); qindex = QUEUE_DIRTY; } else { if ((bp->b_flags & B_DELWRI) == 0 && (bp->b_xflags & BX_VNDIRTY)) panic("bqrelse: not dirty"); if ((bp->b_flags & B_NOREUSE) != 0) { brelse(bp); return; } qindex = QUEUE_CLEAN; } buf_track(bp, __func__); /* binsfree unlocks bp. */ binsfree(bp, qindex); return; out: buf_track(bp, __func__); /* unlock */ BUF_UNLOCK(bp); } /* * Complete I/O to a VMIO backed page. Validate the pages as appropriate, * restore bogus pages. */ static void vfs_vmio_iodone(struct buf *bp) { vm_ooffset_t foff; vm_page_t m; vm_object_t obj; struct vnode *vp __unused; int i, iosize, resid; bool bogus; obj = bp->b_bufobj->bo_object; KASSERT(blockcount_read(&obj->paging_in_progress) >= bp->b_npages, ("vfs_vmio_iodone: paging in progress(%d) < b_npages(%d)", blockcount_read(&obj->paging_in_progress), bp->b_npages)); vp = bp->b_vp; VNPASS(vp->v_holdcnt > 0, vp); VNPASS(vp->v_object != NULL, vp); foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_vmio_iodone: bp %p has no buffer offset", bp)); bogus = false; iosize = bp->b_bcount - bp->b_resid; for (i = 0; i < bp->b_npages; i++) { resid = ((foff + PAGE_SIZE) & ~(off_t)PAGE_MASK) - foff; if (resid > iosize) resid = iosize; /* * cleanup bogus pages, restoring the originals */ m = bp->b_pages[i]; if (m == bogus_page) { bogus = true; m = vm_page_relookup(obj, OFF_TO_IDX(foff)); if (m == NULL) panic("biodone: page disappeared!"); bp->b_pages[i] = m; } else if ((bp->b_iocmd == BIO_READ) && resid > 0) { /* * In the write case, the valid and clean bits are * already changed correctly ( see bdwrite() ), so we * only need to do this here in the read case. */ KASSERT((m->dirty & vm_page_bits(foff & PAGE_MASK, resid)) == 0, ("vfs_vmio_iodone: page %p " "has unexpected dirty bits", m)); vfs_page_set_valid(bp, foff, m); } KASSERT(OFF_TO_IDX(foff) == m->pindex, ("vfs_vmio_iodone: foff(%jd)/pindex(%ju) mismatch", (intmax_t)foff, (uintmax_t)m->pindex)); vm_page_sunbusy(m); foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; iosize -= resid; } vm_object_pip_wakeupn(obj, bp->b_npages); if (bogus && buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * Perform page invalidation when a buffer is released. The fully invalid * pages will be reclaimed later in vfs_vmio_truncate(). */ static void vfs_vmio_invalidate(struct buf *bp) { vm_object_t obj; vm_page_t m; int flags, i, resid, poffset, presid; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qremove(trunc_page((vm_offset_t)bp->b_data), bp->b_npages); } else BUF_CHECK_UNMAPPED(bp); /* * Get the base offset and length of the buffer. Note that * in the VMIO case if the buffer block size is not * page-aligned then b_data pointer may not be page-aligned. * But our b_pages[] array *IS* page aligned. * * block sizes less then DEV_BSIZE (usually 512) are not * supported due to the page granularity bits (m->valid, * m->dirty, etc...). * * See man buf(9) for more information */ flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; obj = bp->b_bufobj->bo_object; resid = bp->b_bufsize; poffset = bp->b_offset & PAGE_MASK; VM_OBJECT_WLOCK(obj); for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) panic("vfs_vmio_invalidate: Unexpected bogus page."); bp->b_pages[i] = NULL; presid = resid > (PAGE_SIZE - poffset) ? (PAGE_SIZE - poffset) : resid; KASSERT(presid >= 0, ("brelse: extra page")); vm_page_busy_acquire(m, VM_ALLOC_SBUSY); if (pmap_page_wired_mappings(m) == 0) vm_page_set_invalid(m, poffset, presid); vm_page_sunbusy(m); vm_page_release_locked(m, flags); resid -= presid; poffset = 0; } VM_OBJECT_WUNLOCK(obj); bp->b_npages = 0; } /* * Page-granular truncation of an existing VMIO buffer. */ static void vfs_vmio_truncate(struct buf *bp, int desiredpages) { vm_object_t obj; vm_page_t m; int flags, i; if (bp->b_npages == desiredpages) return; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qremove((vm_offset_t)trunc_page((vm_offset_t)bp->b_data) + (desiredpages << PAGE_SHIFT), bp->b_npages - desiredpages); } else BUF_CHECK_UNMAPPED(bp); /* * The object lock is needed only if we will attempt to free pages. */ flags = (bp->b_flags & B_NOREUSE) != 0 ? VPR_NOREUSE : 0; if ((bp->b_flags & B_DIRECT) != 0) { flags |= VPR_TRYFREE; obj = bp->b_bufobj->bo_object; VM_OBJECT_WLOCK(obj); } else { obj = NULL; } for (i = desiredpages; i < bp->b_npages; i++) { m = bp->b_pages[i]; KASSERT(m != bogus_page, ("allocbuf: bogus page found")); bp->b_pages[i] = NULL; if (obj != NULL) vm_page_release_locked(m, flags); else vm_page_release(m, flags); } if (obj != NULL) VM_OBJECT_WUNLOCK(obj); bp->b_npages = desiredpages; } /* * Byte granular extension of VMIO buffers. */ static void vfs_vmio_extend(struct buf *bp, int desiredpages, int size) { /* * We are growing the buffer, possibly in a * byte-granular fashion. */ vm_object_t obj; vm_offset_t toff; vm_offset_t tinc; vm_page_t m; /* * Step 1, bring in the VM pages from the object, allocating * them if necessary. We must clear B_CACHE if these pages * are not valid for the range covered by the buffer. */ obj = bp->b_bufobj->bo_object; if (bp->b_npages < desiredpages) { /* * We must allocate system pages since blocking * here could interfere with paging I/O, no * matter which process we are. * * Only exclusive busy can be tested here. * Blocking on shared busy might lead to * deadlocks once allocbuf() is called after * pages are vfs_busy_pages(). */ (void)vm_page_grab_pages_unlocked(obj, OFF_TO_IDX(bp->b_offset) + bp->b_npages, VM_ALLOC_SYSTEM | VM_ALLOC_IGN_SBUSY | VM_ALLOC_NOBUSY | VM_ALLOC_WIRED, &bp->b_pages[bp->b_npages], desiredpages - bp->b_npages); bp->b_npages = desiredpages; } /* * Step 2. We've loaded the pages into the buffer, * we have to figure out if we can still have B_CACHE * set. Note that B_CACHE is set according to the * byte-granular range ( bcount and size ), not the * aligned range ( newbsize ). * * The VM test is against m->valid, which is DEV_BSIZE * aligned. Needless to say, the validity of the data * needs to also be DEV_BSIZE aligned. Note that this * fails with NFS if the server or some other client * extends the file's EOF. If our buffer is resized, * B_CACHE may remain set! XXX */ toff = bp->b_bcount; tinc = PAGE_SIZE - ((bp->b_offset + toff) & PAGE_MASK); while ((bp->b_flags & B_CACHE) && toff < size) { vm_pindex_t pi; if (tinc > (size - toff)) tinc = size - toff; pi = ((bp->b_offset & PAGE_MASK) + toff) >> PAGE_SHIFT; m = bp->b_pages[pi]; vfs_buf_test_cache(bp, bp->b_offset, toff, tinc, m); toff += tinc; tinc = PAGE_SIZE; } /* * Step 3, fixup the KVA pmap. */ if (buf_mapped(bp)) bpmap_qenter(bp); else BUF_CHECK_UNMAPPED(bp); } /* * Check to see if a block at a particular lbn is available for a clustered * write. */ static int vfs_bio_clcheck(struct vnode *vp, int size, daddr_t lblkno, daddr_t blkno) { struct buf *bpa; int match; match = 0; /* If the buf isn't in core skip it */ if ((bpa = gbincore(&vp->v_bufobj, lblkno)) == NULL) return (0); /* If the buf is busy we don't want to wait for it */ if (BUF_LOCK(bpa, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) return (0); /* Only cluster with valid clusterable delayed write buffers */ if ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) != (B_DELWRI | B_CLUSTEROK)) goto done; if (bpa->b_bufsize != size) goto done; /* * Check to see if it is in the expected place on disk and that the * block has been mapped. */ if ((bpa->b_blkno != bpa->b_lblkno) && (bpa->b_blkno == blkno)) match = 1; done: BUF_UNLOCK(bpa); return (match); } /* * vfs_bio_awrite: * * Implement clustered async writes for clearing out B_DELWRI buffers. * This is much better then the old way of writing only one buffer at * a time. Note that we may not be presented with the buffers in the * correct order, so we search for the cluster in both directions. */ int vfs_bio_awrite(struct buf *bp) { struct bufobj *bo; int i; int j; daddr_t lblkno = bp->b_lblkno; struct vnode *vp = bp->b_vp; int ncl; int nwritten; int size; int maxcl; int gbflags; bo = &vp->v_bufobj; gbflags = (bp->b_data == unmapped_buf) ? GB_UNMAPPED : 0; /* * right now we support clustered writing only to regular files. If * we find a clusterable block we could be in the middle of a cluster * rather then at the beginning. */ if ((vp->v_type == VREG) && (vp->v_mount != 0) && /* Only on nodes that have the size info */ (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) { size = vp->v_mount->mnt_stat.f_iosize; maxcl = MAXPHYS / size; BO_RLOCK(bo); for (i = 1; i < maxcl; i++) if (vfs_bio_clcheck(vp, size, lblkno + i, bp->b_blkno + ((i * size) >> DEV_BSHIFT)) == 0) break; for (j = 1; i + j <= maxcl && j <= lblkno; j++) if (vfs_bio_clcheck(vp, size, lblkno - j, bp->b_blkno - ((j * size) >> DEV_BSHIFT)) == 0) break; BO_RUNLOCK(bo); --j; ncl = i + j; /* * this is a possible cluster write */ if (ncl != 1) { BUF_UNLOCK(bp); nwritten = cluster_wbuild(vp, size, lblkno - j, ncl, gbflags); return (nwritten); } } bremfree(bp); bp->b_flags |= B_ASYNC; /* * default (old) behavior, writing out only one block * * XXX returns b_bufsize instead of b_bcount for nwritten? */ nwritten = bp->b_bufsize; (void) bwrite(bp); return (nwritten); } /* * getnewbuf_kva: * * Allocate KVA for an empty buf header according to gbflags. */ static int getnewbuf_kva(struct buf *bp, int gbflags, int maxsize) { if ((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_UNMAPPED) { /* * In order to keep fragmentation sane we only allocate kva * in BKVASIZE chunks. XXX with vmem we can do page size. */ maxsize = (maxsize + BKVAMASK) & ~BKVAMASK; if (maxsize != bp->b_kvasize && bufkva_alloc(bp, maxsize, gbflags)) return (ENOSPC); } return (0); } /* * getnewbuf: * * Find and initialize a new buffer header, freeing up existing buffers * in the bufqueues as necessary. The new buffer is returned locked. * * We block if: * We have insufficient buffer headers * We have insufficient buffer space * buffer_arena is too fragmented ( space reservation fails ) * If we have to flush dirty buffers ( but we try to avoid this ) * * The caller is responsible for releasing the reserved bufspace after * allocbuf() is called. */ static struct buf * getnewbuf(struct vnode *vp, int slpflag, int slptimeo, int maxsize, int gbflags) { struct bufdomain *bd; struct buf *bp; bool metadata, reserved; bp = NULL; KASSERT((gbflags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); if (!unmapped_buf_allowed) gbflags &= ~(GB_UNMAPPED | GB_KVAALLOC); if (vp == NULL || (vp->v_vflag & (VV_MD | VV_SYSTEM)) != 0 || vp->v_type == VCHR) metadata = true; else metadata = false; if (vp == NULL) bd = &bdomain[0]; else bd = &bdomain[vp->v_bufobj.bo_domain]; counter_u64_add(getnewbufcalls, 1); reserved = false; do { if (reserved == false && bufspace_reserve(bd, maxsize, metadata) != 0) { counter_u64_add(getnewbufrestarts, 1); continue; } reserved = true; if ((bp = buf_alloc(bd)) == NULL) { counter_u64_add(getnewbufrestarts, 1); continue; } if (getnewbuf_kva(bp, gbflags, maxsize) == 0) return (bp); break; } while (buf_recycle(bd, false) == 0); if (reserved) bufspace_release(bd, maxsize); if (bp != NULL) { bp->b_flags |= B_INVAL; brelse(bp); } bufspace_wait(bd, vp, gbflags, slpflag, slptimeo); return (NULL); } /* * buf_daemon: * * buffer flushing daemon. Buffers are normally flushed by the * update daemon but if it cannot keep up this process starts to * take the load in an attempt to prevent getnewbuf() from blocking. */ static struct kproc_desc buf_kp = { "bufdaemon", buf_daemon, &bufdaemonproc }; SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST, kproc_start, &buf_kp); static int buf_flush(struct vnode *vp, struct bufdomain *bd, int target) { int flushed; flushed = flushbufqueues(vp, bd, target, 0); if (flushed == 0) { /* * Could not find any buffers without rollback * dependencies, so just write the first one * in the hopes of eventually making progress. */ if (vp != NULL && target > 2) target /= 2; flushbufqueues(vp, bd, target, 1); } return (flushed); } static void buf_daemon() { struct bufdomain *bd; int speedupreq; int lodirty; int i; /* * This process needs to be suspended prior to shutdown sync. */ EVENTHANDLER_REGISTER(shutdown_pre_sync, kthread_shutdown, curthread, SHUTDOWN_PRI_LAST + 100); /* * Start the buf clean daemons as children threads. */ for (i = 0 ; i < buf_domains; i++) { int error; error = kthread_add((void (*)(void *))bufspace_daemon, &bdomain[i], curproc, NULL, 0, 0, "bufspacedaemon-%d", i); if (error) panic("error %d spawning bufspace daemon", error); } /* * This process is allowed to take the buffer cache to the limit */ curthread->td_pflags |= TDP_NORUNNINGBUF | TDP_BUFNEED; mtx_lock(&bdlock); for (;;) { bd_request = 0; mtx_unlock(&bdlock); kthread_suspend_check(); /* * Save speedupreq for this pass and reset to capture new * requests. */ speedupreq = bd_speedupreq; bd_speedupreq = 0; /* * Flush each domain sequentially according to its level and * the speedup request. */ for (i = 0; i < buf_domains; i++) { bd = &bdomain[i]; if (speedupreq) lodirty = bd->bd_numdirtybuffers / 2; else lodirty = bd->bd_lodirtybuffers; while (bd->bd_numdirtybuffers > lodirty) { if (buf_flush(NULL, bd, bd->bd_numdirtybuffers - lodirty) == 0) break; kern_yield(PRI_USER); } } /* * Only clear bd_request if we have reached our low water * mark. The buf_daemon normally waits 1 second and * then incrementally flushes any dirty buffers that have * built up, within reason. * * If we were unable to hit our low water mark and couldn't * find any flushable buffers, we sleep for a short period * to avoid endless loops on unlockable buffers. */ mtx_lock(&bdlock); if (!BIT_EMPTY(BUF_DOMAINS, &bdlodirty)) { /* * We reached our low water mark, reset the * request and sleep until we are needed again. * The sleep is just so the suspend code works. */ bd_request = 0; /* * Do an extra wakeup in case dirty threshold * changed via sysctl and the explicit transition * out of shortfall was missed. */ bdirtywakeup(); if (runningbufspace <= lorunningspace) runningwakeup(); msleep(&bd_request, &bdlock, PVM, "psleep", hz); } else { /* * We couldn't find any flushable dirty buffers but * still have too many dirty buffers, we * have to sleep and try again. (rare) */ msleep(&bd_request, &bdlock, PVM, "qsleep", hz / 10); } } } /* * flushbufqueues: * * Try to flush a buffer in the dirty queue. We must be careful to * free up B_INVAL buffers instead of write them, which NFS is * particularly sensitive to. */ static int flushwithdeps = 0; SYSCTL_INT(_vfs, OID_AUTO, flushwithdeps, CTLFLAG_RW | CTLFLAG_STATS, &flushwithdeps, 0, "Number of buffers flushed with dependecies that require rollbacks"); static int flushbufqueues(struct vnode *lvp, struct bufdomain *bd, int target, int flushdeps) { struct bufqueue *bq; struct buf *sentinel; struct vnode *vp; struct mount *mp; struct buf *bp; int hasdeps; int flushed; int error; bool unlock; flushed = 0; bq = &bd->bd_dirtyq; bp = NULL; sentinel = malloc(sizeof(struct buf), M_TEMP, M_WAITOK | M_ZERO); sentinel->b_qindex = QUEUE_SENTINEL; BQ_LOCK(bq); TAILQ_INSERT_HEAD(&bq->bq_queue, sentinel, b_freelist); BQ_UNLOCK(bq); while (flushed != target) { maybe_yield(); BQ_LOCK(bq); bp = TAILQ_NEXT(sentinel, b_freelist); if (bp != NULL) { TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); TAILQ_INSERT_AFTER(&bq->bq_queue, bp, sentinel, b_freelist); } else { BQ_UNLOCK(bq); break; } /* * Skip sentinels inserted by other invocations of the * flushbufqueues(), taking care to not reorder them. * * Only flush the buffers that belong to the * vnode locked by the curthread. */ if (bp->b_qindex == QUEUE_SENTINEL || (lvp != NULL && bp->b_vp != lvp)) { BQ_UNLOCK(bq); continue; } error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL); BQ_UNLOCK(bq); if (error != 0) continue; /* * BKGRDINPROG can only be set with the buf and bufobj * locks both held. We tolerate a race to clear it here. */ if ((bp->b_vflags & BV_BKGRDINPROG) != 0 || (bp->b_flags & B_DELWRI) == 0) { BUF_UNLOCK(bp); continue; } if (bp->b_flags & B_INVAL) { bremfreef(bp); brelse(bp); flushed++; continue; } if (!LIST_EMPTY(&bp->b_dep) && buf_countdeps(bp, 0)) { if (flushdeps == 0) { BUF_UNLOCK(bp); continue; } hasdeps = 1; } else hasdeps = 0; /* * We must hold the lock on a vnode before writing * one of its buffers. Otherwise we may confuse, or * in the case of a snapshot vnode, deadlock the * system. * * The lock order here is the reverse of the normal * of vnode followed by buf lock. This is ok because * the NOWAIT will prevent deadlock. */ vp = bp->b_vp; if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { BUF_UNLOCK(bp); continue; } if (lvp == NULL) { unlock = true; error = vn_lock(vp, LK_EXCLUSIVE | LK_NOWAIT); } else { ASSERT_VOP_LOCKED(vp, "getbuf"); unlock = false; error = VOP_ISLOCKED(vp) == LK_EXCLUSIVE ? 0 : vn_lock(vp, LK_TRYUPGRADE); } if (error == 0) { CTR3(KTR_BUF, "flushbufqueue(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); if (curproc == bufdaemonproc) { vfs_bio_awrite(bp); } else { bremfree(bp); bwrite(bp); counter_u64_add(notbufdflushes, 1); } vn_finished_write(mp); if (unlock) VOP_UNLOCK(vp); flushwithdeps += hasdeps; flushed++; /* * Sleeping on runningbufspace while holding * vnode lock leads to deadlock. */ if (curproc == bufdaemonproc && runningbufspace > hirunningspace) waitrunningbufspace(); continue; } vn_finished_write(mp); BUF_UNLOCK(bp); } BQ_LOCK(bq); TAILQ_REMOVE(&bq->bq_queue, sentinel, b_freelist); BQ_UNLOCK(bq); free(sentinel, M_TEMP); return (flushed); } /* * Check to see if a block is currently memory resident. */ struct buf * incore(struct bufobj *bo, daddr_t blkno) { struct buf *bp; BO_RLOCK(bo); bp = gbincore(bo, blkno); BO_RUNLOCK(bo); return (bp); } /* * Returns true if no I/O is needed to access the * associated VM object. This is like incore except * it also hunts around in the VM system for the data. */ static int inmem(struct vnode * vp, daddr_t blkno) { vm_object_t obj; vm_offset_t toff, tinc, size; vm_page_t m; vm_ooffset_t off; ASSERT_VOP_LOCKED(vp, "inmem"); if (incore(&vp->v_bufobj, blkno)) return 1; if (vp->v_mount == NULL) return 0; obj = vp->v_object; if (obj == NULL) return (0); size = PAGE_SIZE; if (size > vp->v_mount->mnt_stat.f_iosize) size = vp->v_mount->mnt_stat.f_iosize; off = (vm_ooffset_t)blkno * (vm_ooffset_t)vp->v_mount->mnt_stat.f_iosize; VM_OBJECT_RLOCK(obj); for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) { m = vm_page_lookup(obj, OFF_TO_IDX(off + toff)); if (!m) goto notinmem; tinc = size; if (tinc > PAGE_SIZE - ((toff + off) & PAGE_MASK)) tinc = PAGE_SIZE - ((toff + off) & PAGE_MASK); if (vm_page_is_valid(m, (vm_offset_t) ((toff + off) & PAGE_MASK), tinc) == 0) goto notinmem; } VM_OBJECT_RUNLOCK(obj); return 1; notinmem: VM_OBJECT_RUNLOCK(obj); return (0); } /* * Set the dirty range for a buffer based on the status of the dirty * bits in the pages comprising the buffer. The range is limited * to the size of the buffer. * * Tell the VM system that the pages associated with this buffer * are clean. This is used for delayed writes where the data is * going to go to disk eventually without additional VM intevention. * * Note that while we only really need to clean through to b_bcount, we * just go ahead and clean through to b_bufsize. */ static void vfs_clean_pages_dirty_buf(struct buf *bp) { vm_ooffset_t foff, noff, eoff; vm_page_t m; int i; if ((bp->b_flags & B_VMIO) == 0 || bp->b_bufsize == 0) return; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_clean_pages_dirty_buf: no buffer offset")); vfs_busy_pages_acquire(bp); vfs_setdirty_range(bp); for (i = 0; i < bp->b_npages; i++) { noff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; eoff = noff; if (eoff > bp->b_offset + bp->b_bufsize) eoff = bp->b_offset + bp->b_bufsize; m = bp->b_pages[i]; vfs_page_set_validclean(bp, foff, m); /* vm_page_clear_dirty(m, foff & PAGE_MASK, eoff - foff); */ foff = noff; } vfs_busy_pages_release(bp); } static void vfs_setdirty_range(struct buf *bp) { vm_offset_t boffset; vm_offset_t eoffset; int i; /* * test the pages to see if they have been modified directly * by users through the VM system. */ for (i = 0; i < bp->b_npages; i++) vm_page_test_dirty(bp->b_pages[i]); /* * Calculate the encompassing dirty range, boffset and eoffset, * (eoffset - boffset) bytes. */ for (i = 0; i < bp->b_npages; i++) { if (bp->b_pages[i]->dirty) break; } boffset = (i << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); for (i = bp->b_npages - 1; i >= 0; --i) { if (bp->b_pages[i]->dirty) { break; } } eoffset = ((i + 1) << PAGE_SHIFT) - (bp->b_offset & PAGE_MASK); /* * Fit it to the buffer. */ if (eoffset > bp->b_bcount) eoffset = bp->b_bcount; /* * If we have a good dirty range, merge with the existing * dirty range. */ if (boffset < eoffset) { if (bp->b_dirtyoff > boffset) bp->b_dirtyoff = boffset; if (bp->b_dirtyend < eoffset) bp->b_dirtyend = eoffset; } } /* * Allocate the KVA mapping for an existing buffer. * If an unmapped buffer is provided but a mapped buffer is requested, take * also care to properly setup mappings between pages and KVA. */ static void bp_unmapped_get_kva(struct buf *bp, daddr_t blkno, int size, int gbflags) { int bsize, maxsize, need_mapping, need_kva; off_t offset; need_mapping = bp->b_data == unmapped_buf && (gbflags & GB_UNMAPPED) == 0; need_kva = bp->b_kvabase == unmapped_buf && bp->b_data == unmapped_buf && (gbflags & GB_KVAALLOC) != 0; if (!need_mapping && !need_kva) return; BUF_CHECK_UNMAPPED(bp); if (need_mapping && bp->b_kvabase != unmapped_buf) { /* * Buffer is not mapped, but the KVA was already * reserved at the time of the instantiation. Use the * allocated space. */ goto has_addr; } /* * Calculate the amount of the address space we would reserve * if the buffer was mapped. */ bsize = vn_isdisk(bp->b_vp, NULL) ? DEV_BSIZE : bp->b_bufobj->bo_bsize; KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); offset = blkno * bsize; maxsize = size + (offset & PAGE_MASK); maxsize = imax(maxsize, bsize); while (bufkva_alloc(bp, maxsize, gbflags) != 0) { if ((gbflags & GB_NOWAIT_BD) != 0) { /* * XXXKIB: defragmentation cannot * succeed, not sure what else to do. */ panic("GB_NOWAIT_BD and GB_UNMAPPED %p", bp); } counter_u64_add(mappingrestarts, 1); bufspace_wait(bufdomain(bp), bp->b_vp, gbflags, 0, 0); } has_addr: if (need_mapping) { /* b_offset is handled by bpmap_qenter. */ bp->b_data = bp->b_kvabase; BUF_CHECK_MAPPED(bp); bpmap_qenter(bp); } } struct buf * getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo, int flags) { struct buf *bp; int error; error = getblkx(vp, blkno, blkno, size, slpflag, slptimeo, flags, &bp); if (error != 0) return (NULL); return (bp); } /* * getblkx: * * Get a block given a specified block and offset into a file/device. * The buffers B_DONE bit will be cleared on return, making it almost * ready for an I/O initiation. B_INVAL may or may not be set on * return. The caller should clear B_INVAL prior to initiating a * READ. * * For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for * an existing buffer. * * For a VMIO buffer, B_CACHE is modified according to the backing VM. * If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set * and then cleared based on the backing VM. If the previous buffer is * non-0-sized but invalid, B_CACHE will be cleared. * * If getblk() must create a new buffer, the new buffer is returned with * both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which * case it is returned with B_INVAL clear and B_CACHE set based on the * backing VM. * * getblk() also forces a bwrite() for any B_DELWRI buffer whose * B_CACHE bit is clear. - * + * * What this means, basically, is that the caller should use B_CACHE to * determine whether the buffer is fully valid or not and should clear * B_INVAL prior to issuing a read. If the caller intends to validate * the buffer by loading its data area with something, the caller needs * to clear B_INVAL. If the caller does this without issuing an I/O, * the caller should set B_CACHE ( as an optimization ), else the caller * should issue the I/O and biodone() will set B_CACHE if the I/O was * a write attempt or if it was a successful read. If the caller * intends to issue a READ, the caller must clear B_INVAL and BIO_ERROR * prior to issuing the READ. biodone() will *not* clear B_INVAL. * * The blkno parameter is the logical block being requested. Normally * the mapping of logical block number to disk block address is done * by calling VOP_BMAP(). However, if the mapping is already known, the * disk block address can be passed using the dblkno parameter. If the * disk block address is not known, then the same value should be passed * for blkno and dblkno. */ int getblkx(struct vnode *vp, daddr_t blkno, daddr_t dblkno, int size, int slpflag, int slptimeo, int flags, struct buf **bpp) { struct buf *bp; struct bufobj *bo; daddr_t d_blkno; int bsize, error, maxsize, vmio; off_t offset; CTR3(KTR_BUF, "getblk(%p, %ld, %d)", vp, (long)blkno, size); KASSERT((flags & (GB_UNMAPPED | GB_KVAALLOC)) != GB_KVAALLOC, ("GB_KVAALLOC only makes sense with GB_UNMAPPED")); ASSERT_VOP_LOCKED(vp, "getblk"); if (size > maxbcachebuf) panic("getblk: size(%d) > maxbcachebuf(%d)\n", size, maxbcachebuf); if (!unmapped_buf_allowed) flags &= ~(GB_UNMAPPED | GB_KVAALLOC); bo = &vp->v_bufobj; d_blkno = dblkno; loop: BO_RLOCK(bo); bp = gbincore(bo, blkno); if (bp != NULL) { int lockflags; /* * Buffer is in-core. If the buffer is not busy nor managed, * it must be on a queue. */ lockflags = LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK; if ((flags & GB_LOCK_NOWAIT) != 0) lockflags |= LK_NOWAIT; error = BUF_TIMELOCK(bp, lockflags, BO_LOCKPTR(bo), "getblk", slpflag, slptimeo); /* * If we slept and got the lock we have to restart in case * the buffer changed identities. */ if (error == ENOLCK) goto loop; /* We timed out or were interrupted. */ else if (error != 0) return (error); /* If recursed, assume caller knows the rules. */ else if (BUF_LOCKRECURSED(bp)) goto end; /* * The buffer is locked. B_CACHE is cleared if the buffer is * invalid. Otherwise, for a non-VMIO buffer, B_CACHE is set * and for a VMIO buffer B_CACHE is adjusted according to the * backing VM cache. */ if (bp->b_flags & B_INVAL) bp->b_flags &= ~B_CACHE; else if ((bp->b_flags & (B_VMIO | B_INVAL)) == 0) bp->b_flags |= B_CACHE; if (bp->b_flags & B_MANAGED) MPASS(bp->b_qindex == QUEUE_NONE); else bremfree(bp); /* * check for size inconsistencies for non-VMIO case. */ if (bp->b_bcount != size) { if ((bp->b_flags & B_VMIO) == 0 || (size > bp->b_kvasize)) { if (bp->b_flags & B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); } else { if (LIST_EMPTY(&bp->b_dep)) { bp->b_flags |= B_RELBUF; brelse(bp); } else { bp->b_flags |= B_NOCACHE; bwrite(bp); } } goto loop; } } /* * Handle the case of unmapped buffer which should * become mapped, or the buffer for which KVA * reservation is requested. */ bp_unmapped_get_kva(bp, blkno, size, flags); /* * If the size is inconsistent in the VMIO case, we can resize * the buffer. This might lead to B_CACHE getting set or * cleared. If the size has not changed, B_CACHE remains * unchanged from its previous state. */ allocbuf(bp, size); KASSERT(bp->b_offset != NOOFFSET, ("getblk: no buffer offset")); /* * A buffer with B_DELWRI set and B_CACHE clear must * be committed before we can return the buffer in * order to prevent the caller from issuing a read * ( due to B_CACHE not being set ) and overwriting * it. * * Most callers, including NFS and FFS, need this to * operate properly either because they assume they * can issue a read if B_CACHE is not set, or because * ( for example ) an uncached B_DELWRI might loop due * to softupdates re-dirtying the buffer. In the latter * case, B_CACHE is set after the first write completes, * preventing further loops. * NOTE! b*write() sets B_CACHE. If we cleared B_CACHE * above while extending the buffer, we cannot allow the * buffer to remain with B_CACHE set after the write * completes or it will represent a corrupt state. To * deal with this we set B_NOCACHE to scrap the buffer * after the write. * * We might be able to do something fancy, like setting * B_CACHE in bwrite() except if B_DELWRI is already set, * so the below call doesn't set B_CACHE, but that gets real * confusing. This is much easier. */ if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) { bp->b_flags |= B_NOCACHE; bwrite(bp); goto loop; } bp->b_flags &= ~B_DONE; } else { /* * Buffer is not in-core, create new buffer. The buffer * returned by getnewbuf() is locked. Note that the returned * buffer is also considered valid (not marked B_INVAL). */ BO_RUNLOCK(bo); /* * If the user does not want us to create the buffer, bail out * here. */ if (flags & GB_NOCREAT) return (EEXIST); bsize = vn_isdisk(vp, NULL) ? DEV_BSIZE : bo->bo_bsize; KASSERT(bsize != 0, ("bsize == 0, check bo->bo_bsize")); offset = blkno * bsize; vmio = vp->v_object != NULL; if (vmio) { maxsize = size + (offset & PAGE_MASK); } else { maxsize = size; /* Do not allow non-VMIO notmapped buffers. */ flags &= ~(GB_UNMAPPED | GB_KVAALLOC); } maxsize = imax(maxsize, bsize); if ((flags & GB_NOSPARSE) != 0 && vmio && !vn_isdisk(vp, NULL)) { error = VOP_BMAP(vp, blkno, NULL, &d_blkno, 0, 0); KASSERT(error != EOPNOTSUPP, ("GB_NOSPARSE from fs not supporting bmap, vp %p", vp)); if (error != 0) return (error); if (d_blkno == -1) return (EJUSTRETURN); } bp = getnewbuf(vp, slpflag, slptimeo, maxsize, flags); if (bp == NULL) { if (slpflag || slptimeo) return (ETIMEDOUT); /* * XXX This is here until the sleep path is diagnosed * enough to work under very low memory conditions. * * There's an issue on low memory, 4BSD+non-preempt * systems (eg MIPS routers with 32MB RAM) where buffer * exhaustion occurs without sleeping for buffer * reclaimation. This just sticks in a loop and * constantly attempts to allocate a buffer, which * hits exhaustion and tries to wakeup bufdaemon. * This never happens because we never yield. * * The real solution is to identify and fix these cases * so we aren't effectively busy-waiting in a loop * until the reclaimation path has cycles to run. */ kern_yield(PRI_USER); goto loop; } /* * This code is used to make sure that a buffer is not * created while the getnewbuf routine is blocked. * This can be a problem whether the vnode is locked or not. * If the buffer is created out from under us, we have to * throw away the one we just created. * * Note: this must occur before we associate the buffer * with the vp especially considering limitations in * the splay tree implementation when dealing with duplicate * lblkno's. */ BO_LOCK(bo); if (gbincore(bo, blkno)) { BO_UNLOCK(bo); bp->b_flags |= B_INVAL; bufspace_release(bufdomain(bp), maxsize); brelse(bp); goto loop; } /* * Insert the buffer into the hash, so that it can * be found by incore. */ bp->b_lblkno = blkno; bp->b_blkno = d_blkno; bp->b_offset = offset; bgetvp(vp, bp); BO_UNLOCK(bo); /* * set B_VMIO bit. allocbuf() the buffer bigger. Since the * buffer size starts out as 0, B_CACHE will be set by * allocbuf() for the VMIO case prior to it testing the * backing store for validity. */ if (vmio) { bp->b_flags |= B_VMIO; KASSERT(vp->v_object == bp->b_bufobj->bo_object, ("ARGH! different b_bufobj->bo_object %p %p %p\n", bp, vp->v_object, bp->b_bufobj->bo_object)); } else { bp->b_flags &= ~B_VMIO; KASSERT(bp->b_bufobj->bo_object == NULL, ("ARGH! has b_bufobj->bo_object %p %p\n", bp, bp->b_bufobj->bo_object)); BUF_CHECK_MAPPED(bp); } allocbuf(bp, size); bufspace_release(bufdomain(bp), maxsize); bp->b_flags &= ~B_DONE; } CTR4(KTR_BUF, "getblk(%p, %ld, %d) = %p", vp, (long)blkno, size, bp); end: buf_track(bp, __func__); KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); *bpp = bp; return (0); } /* * Get an empty, disassociated buffer of given size. The buffer is initially * set to B_INVAL. */ struct buf * geteblk(int size, int flags) { struct buf *bp; int maxsize; maxsize = (size + BKVAMASK) & ~BKVAMASK; while ((bp = getnewbuf(NULL, 0, 0, maxsize, flags)) == NULL) { if ((flags & GB_NOWAIT_BD) && (curthread->td_pflags & TDP_BUFNEED) != 0) return (NULL); } allocbuf(bp, size); bufspace_release(bufdomain(bp), maxsize); bp->b_flags |= B_INVAL; /* b_dep cleared by getnewbuf() */ return (bp); } /* * Truncate the backing store for a non-vmio buffer. */ static void vfs_nonvmio_truncate(struct buf *bp, int newbsize) { if (bp->b_flags & B_MALLOC) { /* * malloced buffers are not shrunk */ if (newbsize == 0) { bufmallocadjust(bp, 0); free(bp->b_data, M_BIOBUF); bp->b_data = bp->b_kvabase; bp->b_flags &= ~B_MALLOC; } return; } vm_hold_free_pages(bp, newbsize); bufspace_adjust(bp, newbsize); } /* * Extend the backing for a non-VMIO buffer. */ static void vfs_nonvmio_extend(struct buf *bp, int newbsize) { caddr_t origbuf; int origbufsize; /* * We only use malloced memory on the first allocation. * and revert to page-allocated memory when the buffer * grows. * * There is a potential smp race here that could lead * to bufmallocspace slightly passing the max. It * is probably extremely rare and not worth worrying * over. */ if (bp->b_bufsize == 0 && newbsize <= PAGE_SIZE/2 && bufmallocspace < maxbufmallocspace) { bp->b_data = malloc(newbsize, M_BIOBUF, M_WAITOK); bp->b_flags |= B_MALLOC; bufmallocadjust(bp, newbsize); return; } /* * If the buffer is growing on its other-than-first * allocation then we revert to the page-allocation * scheme. */ origbuf = NULL; origbufsize = 0; if (bp->b_flags & B_MALLOC) { origbuf = bp->b_data; origbufsize = bp->b_bufsize; bp->b_data = bp->b_kvabase; bufmallocadjust(bp, 0); bp->b_flags &= ~B_MALLOC; newbsize = round_page(newbsize); } vm_hold_load_pages(bp, (vm_offset_t) bp->b_data + bp->b_bufsize, (vm_offset_t) bp->b_data + newbsize); if (origbuf != NULL) { bcopy(origbuf, bp->b_data, origbufsize); free(origbuf, M_BIOBUF); } bufspace_adjust(bp, newbsize); } /* * This code constitutes the buffer memory from either anonymous system * memory (in the case of non-VMIO operations) or from an associated * VM object (in the case of VMIO operations). This code is able to * resize a buffer up or down. * * Note that this code is tricky, and has many complications to resolve * deadlock or inconsistent data situations. Tread lightly!!! * There are B_CACHE and B_DELWRI interactions that must be dealt with by * the caller. Calling this code willy nilly can result in the loss of data. * * allocbuf() only adjusts B_CACHE for VMIO buffers. getblk() deals with * B_CACHE for the non-VMIO case. */ int allocbuf(struct buf *bp, int size) { int newbsize; if (bp->b_bcount == size) return (1); if (bp->b_kvasize != 0 && bp->b_kvasize < size) panic("allocbuf: buffer too small"); newbsize = roundup2(size, DEV_BSIZE); if ((bp->b_flags & B_VMIO) == 0) { if ((bp->b_flags & B_MALLOC) == 0) newbsize = round_page(newbsize); /* * Just get anonymous memory from the kernel. Don't * mess with B_CACHE. */ if (newbsize < bp->b_bufsize) vfs_nonvmio_truncate(bp, newbsize); else if (newbsize > bp->b_bufsize) vfs_nonvmio_extend(bp, newbsize); } else { int desiredpages; desiredpages = (size == 0) ? 0 : num_pages((bp->b_offset & PAGE_MASK) + newbsize); if (bp->b_flags & B_MALLOC) panic("allocbuf: VMIO buffer can't be malloced"); /* * Set B_CACHE initially if buffer is 0 length or will become * 0-length. */ if (size == 0 || bp->b_bufsize == 0) bp->b_flags |= B_CACHE; if (newbsize < bp->b_bufsize) vfs_vmio_truncate(bp, desiredpages); /* XXX This looks as if it should be newbsize > b_bufsize */ else if (size > bp->b_bcount) vfs_vmio_extend(bp, desiredpages, size); bufspace_adjust(bp, newbsize); } bp->b_bcount = size; /* requested buffer size. */ return (1); } extern int inflight_transient_maps; static struct bio_queue nondump_bios; void biodone(struct bio *bp) { struct mtx *mtxp; void (*done)(struct bio *); vm_offset_t start, end; biotrack(bp, __func__); /* * Avoid completing I/O when dumping after a panic since that may * result in a deadlock in the filesystem or pager code. Note that * this doesn't affect dumps that were started manually since we aim * to keep the system usable after it has been resumed. */ if (__predict_false(dumping && SCHEDULER_STOPPED())) { TAILQ_INSERT_HEAD(&nondump_bios, bp, bio_queue); return; } if ((bp->bio_flags & BIO_TRANSIENT_MAPPING) != 0) { bp->bio_flags &= ~BIO_TRANSIENT_MAPPING; bp->bio_flags |= BIO_UNMAPPED; start = trunc_page((vm_offset_t)bp->bio_data); end = round_page((vm_offset_t)bp->bio_data + bp->bio_length); bp->bio_data = unmapped_buf; pmap_qremove(start, atop(end - start)); vmem_free(transient_arena, start, end - start); atomic_add_int(&inflight_transient_maps, -1); } done = bp->bio_done; if (done == NULL) { mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); bp->bio_flags |= BIO_DONE; wakeup(bp); mtx_unlock(mtxp); } else done(bp); } /* * Wait for a BIO to finish. */ int biowait(struct bio *bp, const char *wchan) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); while ((bp->bio_flags & BIO_DONE) == 0) msleep(bp, mtxp, PRIBIO, wchan, 0); mtx_unlock(mtxp); if (bp->bio_error != 0) return (bp->bio_error); if (!(bp->bio_flags & BIO_ERROR)) return (0); return (EIO); } void biofinish(struct bio *bp, struct devstat *stat, int error) { - + if (error) { bp->bio_error = error; bp->bio_flags |= BIO_ERROR; } if (stat != NULL) devstat_end_transaction_bio(stat, bp); biodone(bp); } #if defined(BUF_TRACKING) || defined(FULL_BUF_TRACKING) void biotrack_buf(struct bio *bp, const char *location) { buf_track(bp->bio_track_bp, location); } #endif /* * bufwait: * * Wait for buffer I/O completion, returning error status. The buffer * is left locked and B_DONE on return. B_EINTR is converted into an EINTR * error and cleared. */ int bufwait(struct buf *bp) { if (bp->b_iocmd == BIO_READ) bwait(bp, PRIBIO, "biord"); else bwait(bp, PRIBIO, "biowr"); if (bp->b_flags & B_EINTR) { bp->b_flags &= ~B_EINTR; return (EINTR); } if (bp->b_ioflags & BIO_ERROR) { return (bp->b_error ? bp->b_error : EIO); } else { return (0); } } /* * bufdone: * * Finish I/O on a buffer, optionally calling a completion function. * This is usually called from an interrupt so process blocking is * not allowed. * * biodone is also responsible for setting B_CACHE in a B_VMIO bp. * In a non-VMIO bp, B_CACHE will be set on the next getblk() * assuming B_INVAL is clear. * * For the VMIO case, we set B_CACHE if the op was a read and no * read error occurred, or if the op was a write. B_CACHE is never * set if the buffer is invalid or otherwise uncacheable. * * bufdone does not mess with B_INVAL, allowing the I/O routine or the * initiator to leave B_INVAL set to brelse the buffer out of existence * in the biodone routine. */ void bufdone(struct buf *bp) { struct bufobj *dropobj; void (*biodone)(struct buf *); buf_track(bp, __func__); CTR3(KTR_BUF, "bufdone(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); dropobj = NULL; KASSERT(!(bp->b_flags & B_DONE), ("biodone: bp %p already done", bp)); runningbufwakeup(bp); if (bp->b_iocmd == BIO_WRITE) dropobj = bp->b_bufobj; /* call optional completion function if requested */ if (bp->b_iodone != NULL) { biodone = bp->b_iodone; bp->b_iodone = NULL; (*biodone) (bp); if (dropobj) bufobj_wdrop(dropobj); return; } if (bp->b_flags & B_VMIO) { /* * Set B_CACHE if the op was a normal read and no error * occurred. B_CACHE is set for writes in the b*write() * routines. */ if (bp->b_iocmd == BIO_READ && !(bp->b_flags & (B_INVAL|B_NOCACHE)) && !(bp->b_ioflags & BIO_ERROR)) bp->b_flags |= B_CACHE; vfs_vmio_iodone(bp); } if (!LIST_EMPTY(&bp->b_dep)) buf_complete(bp); if ((bp->b_flags & B_CKHASH) != 0) { KASSERT(bp->b_iocmd == BIO_READ, ("bufdone: b_iocmd %d not BIO_READ", bp->b_iocmd)); KASSERT(buf_mapped(bp), ("bufdone: bp %p not mapped", bp)); (*bp->b_ckhashcalc)(bp); } /* * For asynchronous completions, release the buffer now. The brelse * will do a wakeup there if necessary - so no need to do a wakeup * here in the async case. The sync case always needs to do a wakeup. */ if (bp->b_flags & B_ASYNC) { if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_RELBUF)) || (bp->b_ioflags & BIO_ERROR)) brelse(bp); else bqrelse(bp); } else bdone(bp); if (dropobj) bufobj_wdrop(dropobj); } /* * This routine is called in lieu of iodone in the case of * incomplete I/O. This keeps the busy status for pages * consistent. */ void vfs_unbusy_pages(struct buf *bp) { int i; vm_object_t obj; vm_page_t m; runningbufwakeup(bp); if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; if (m == bogus_page) { m = vm_page_relookup(obj, OFF_TO_IDX(bp->b_offset) + i); if (!m) panic("vfs_unbusy_pages: page missing\n"); bp->b_pages[i] = m; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } else BUF_CHECK_UNMAPPED(bp); } vm_page_sunbusy(m); } vm_object_pip_wakeupn(obj, bp->b_npages); } /* * vfs_page_set_valid: * * Set the valid bits in a page based on the supplied offset. The * range is restricted to the buffer's size. * * This routine is typically called after a read completes. */ static void vfs_page_set_valid(struct buf *bp, vm_ooffset_t off, vm_page_t m) { vm_ooffset_t eoff; /* * Compute the end offset, eoff, such that [off, eoff) does not span a * page boundary and eoff is not greater than the end of the buffer. * The end of the buffer, in this case, is our file EOF, not the * allocation size of the buffer. */ eoff = (off + PAGE_SIZE) & ~(vm_ooffset_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > off) vm_page_set_valid_range(m, off & PAGE_MASK, eoff - off); } /* * vfs_page_set_validclean: * * Set the valid bits and clear the dirty bits in a page based on the * supplied offset. The range is restricted to the buffer's size. */ static void vfs_page_set_validclean(struct buf *bp, vm_ooffset_t off, vm_page_t m) { vm_ooffset_t soff, eoff; /* * Start and end offsets in buffer. eoff - soff may not cross a * page boundary or cross the end of the buffer. The end of the * buffer, in this case, is our file EOF, not the allocation size * of the buffer. */ soff = off; eoff = (off + PAGE_SIZE) & ~(off_t)PAGE_MASK; if (eoff > bp->b_offset + bp->b_bcount) eoff = bp->b_offset + bp->b_bcount; /* * Set valid range. This is typically the entire buffer and thus the * entire page. */ if (eoff > soff) { vm_page_set_validclean( m, (vm_offset_t) (soff & PAGE_MASK), (vm_offset_t) (eoff - soff) ); } } /* * Acquire a shared busy on all pages in the buf. */ void vfs_busy_pages_acquire(struct buf *bp) { int i; for (i = 0; i < bp->b_npages; i++) vm_page_busy_acquire(bp->b_pages[i], VM_ALLOC_SBUSY); } void vfs_busy_pages_release(struct buf *bp) { int i; for (i = 0; i < bp->b_npages; i++) vm_page_sunbusy(bp->b_pages[i]); } /* * This routine is called before a device strategy routine. * It is used to tell the VM system that paging I/O is in * progress, and treat the pages associated with the buffer * almost as being exclusive busy. Also the object paging_in_progress * flag is handled to make sure that the object doesn't become * inconsistent. * * Since I/O has not been initiated yet, certain buffer flags * such as BIO_ERROR or B_INVAL may be in an inconsistent state * and should be ignored. */ void vfs_busy_pages(struct buf *bp, int clear_modify) { vm_object_t obj; vm_ooffset_t foff; vm_page_t m; int i; bool bogus; if (!(bp->b_flags & B_VMIO)) return; obj = bp->b_bufobj->bo_object; foff = bp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("vfs_busy_pages: no buffer offset")); if ((bp->b_flags & B_CLUSTER) == 0) { vm_object_pip_add(obj, bp->b_npages); vfs_busy_pages_acquire(bp); } if (bp->b_bufsize != 0) vfs_setdirty_range(bp); bogus = false; for (i = 0; i < bp->b_npages; i++) { m = bp->b_pages[i]; vm_page_assert_sbusied(m); /* * When readying a buffer for a read ( i.e * clear_modify == 0 ), it is important to do * bogus_page replacement for valid pages in * partially instantiated buffers. Partially * instantiated buffers can, in turn, occur when * reconstituting a buffer from its VM backing store * base. We only have to do this if B_CACHE is * clear ( which causes the I/O to occur in the * first place ). The replacement prevents the read * I/O from overwriting potentially dirty VM-backed * pages. XXX bogus page replacement is, uh, bogus. * It may not work properly with small-block devices. * We need to find a better way. */ if (clear_modify) { pmap_remove_write(m); vfs_page_set_validclean(bp, foff, m); } else if (vm_page_all_valid(m) && (bp->b_flags & B_CACHE) == 0) { bp->b_pages[i] = bogus_page; bogus = true; } foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK; } if (bogus && buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); pmap_qenter(trunc_page((vm_offset_t)bp->b_data), bp->b_pages, bp->b_npages); } } /* * vfs_bio_set_valid: * * Set the range within the buffer to valid. The range is * relative to the beginning of the buffer, b_offset. Note that * b_offset itself may be offset from the beginning of the first * page. */ -void +void vfs_bio_set_valid(struct buf *bp, int base, int size) { int i, n; vm_page_t m; if (!(bp->b_flags & B_VMIO)) return; /* * Fixup base to be relative to beginning of first page. * Set initial n to be the maximum number of bytes in the * first page that can be validated. */ base += (bp->b_offset & PAGE_MASK); n = PAGE_SIZE - (base & PAGE_MASK); /* * Busy may not be strictly necessary here because the pages are * unlikely to be fully valid and the vnode lock will synchronize * their access via getpages. It is grabbed for consistency with * other page validation. */ vfs_busy_pages_acquire(bp); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; vm_page_set_valid_range(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } vfs_busy_pages_release(bp); } /* * vfs_bio_clrbuf: * * If the specified buffer is a non-VMIO buffer, clear the entire * buffer. If the specified buffer is a VMIO buffer, clear and * validate only the previously invalid portions of the buffer. * This routine essentially fakes an I/O, so we need to clear * BIO_ERROR and B_INVAL. * * Note that while we only theoretically need to clear through b_bcount, * we go ahead and clear through b_bufsize. */ void vfs_bio_clrbuf(struct buf *bp) { int i, j, mask, sa, ea, slide; if ((bp->b_flags & (B_VMIO | B_MALLOC)) != B_VMIO) { clrbuf(bp); return; } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; vfs_busy_pages_acquire(bp); sa = bp->b_offset & PAGE_MASK; slide = 0; for (i = 0; i < bp->b_npages; i++, sa = 0) { slide = imin(slide + PAGE_SIZE, bp->b_offset + bp->b_bufsize); ea = slide & PAGE_MASK; if (ea == 0) ea = PAGE_SIZE; if (bp->b_pages[i] == bogus_page) continue; j = sa / DEV_BSIZE; mask = ((1 << ((ea - sa) / DEV_BSIZE)) - 1) << j; if ((bp->b_pages[i]->valid & mask) == mask) continue; if ((bp->b_pages[i]->valid & mask) == 0) pmap_zero_page_area(bp->b_pages[i], sa, ea - sa); else { for (; sa < ea; sa += DEV_BSIZE, j++) { if ((bp->b_pages[i]->valid & (1 << j)) == 0) { pmap_zero_page_area(bp->b_pages[i], sa, DEV_BSIZE); } } } vm_page_set_valid_range(bp->b_pages[i], j * DEV_BSIZE, roundup2(ea - sa, DEV_BSIZE)); } vfs_busy_pages_release(bp); bp->b_resid = 0; } void vfs_bio_bzero_buf(struct buf *bp, int base, int size) { vm_page_t m; int i, n; if (buf_mapped(bp)) { BUF_CHECK_MAPPED(bp); bzero(bp->b_data + base, size); } else { BUF_CHECK_UNMAPPED(bp); n = PAGE_SIZE - (base & PAGE_MASK); for (i = base / PAGE_SIZE; size > 0 && i < bp->b_npages; ++i) { m = bp->b_pages[i]; if (n > size) n = size; pmap_zero_page_area(m, base & PAGE_MASK, n); base += n; size -= n; n = PAGE_SIZE; } } } /* * Update buffer flags based on I/O request parameters, optionally releasing the * buffer. If it's VMIO or direct I/O, the buffer pages are released to the VM, * where they may be placed on a page queue (VMIO) or freed immediately (direct * I/O). Otherwise the buffer is released to the cache. */ static void b_io_dismiss(struct buf *bp, int ioflag, bool release) { KASSERT((ioflag & IO_NOREUSE) == 0 || (ioflag & IO_VMIO) != 0, ("buf %p non-VMIO noreuse", bp)); if ((ioflag & IO_DIRECT) != 0) bp->b_flags |= B_DIRECT; if ((ioflag & IO_EXT) != 0) bp->b_xflags |= BX_ALTDATA; if ((ioflag & (IO_VMIO | IO_DIRECT)) != 0 && LIST_EMPTY(&bp->b_dep)) { bp->b_flags |= B_RELBUF; if ((ioflag & IO_NOREUSE) != 0) bp->b_flags |= B_NOREUSE; if (release) brelse(bp); } else if (release) bqrelse(bp); } void vfs_bio_brelse(struct buf *bp, int ioflag) { b_io_dismiss(bp, ioflag, true); } void vfs_bio_set_flags(struct buf *bp, int ioflag) { b_io_dismiss(bp, ioflag, false); } /* * vm_hold_load_pages and vm_hold_free_pages get pages into * a buffers address space. The pages are anonymous and are * not associated with a file object. */ static void vm_hold_load_pages(struct buf *bp, vm_offset_t from, vm_offset_t to) { vm_offset_t pg; vm_page_t p; int index; BUF_CHECK_MAPPED(bp); to = round_page(to); from = round_page(from); index = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; for (pg = from; pg < to; pg += PAGE_SIZE, index++) { /* * note: must allocate system pages since blocking here * could interfere with paging I/O, no matter which * process we are. */ p = vm_page_alloc(NULL, 0, VM_ALLOC_SYSTEM | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_COUNT((to - pg) >> PAGE_SHIFT) | VM_ALLOC_WAITOK); pmap_qenter(pg, &p, 1); bp->b_pages[index] = p; } bp->b_npages = index; } /* Return pages associated with this buf to the vm system */ static void vm_hold_free_pages(struct buf *bp, int newbsize) { vm_offset_t from; vm_page_t p; int index, newnpages; BUF_CHECK_MAPPED(bp); from = round_page((vm_offset_t)bp->b_data + newbsize); newnpages = (from - trunc_page((vm_offset_t)bp->b_data)) >> PAGE_SHIFT; if (bp->b_npages > newnpages) pmap_qremove(from, bp->b_npages - newnpages); for (index = newnpages; index < bp->b_npages; index++) { p = bp->b_pages[index]; bp->b_pages[index] = NULL; vm_page_unwire_noq(p); vm_page_free(p); } bp->b_npages = newnpages; } /* * Map an IO request into kernel virtual address space. * * All requests are (re)mapped into kernel VA space. * Notice that we use b_bufsize for the size of the buffer * to be mapped. b_bcount might be modified by the driver. * * Note that even if the caller determines that the address space should * be valid, a race or a smaller-file mapped into a larger space may * actually cause vmapbuf() to fail, so all callers of vmapbuf() MUST * check the return value. * * This function only works with pager buffers. */ int vmapbuf(struct buf *bp, int mapbuf) { vm_prot_t prot; int pidx; if (bp->b_bufsize < 0) return (-1); prot = VM_PROT_READ; if (bp->b_iocmd == BIO_READ) prot |= VM_PROT_WRITE; /* Less backwards than it looks */ if ((pidx = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map, (vm_offset_t)bp->b_data, bp->b_bufsize, prot, bp->b_pages, btoc(MAXPHYS))) < 0) return (-1); bp->b_npages = pidx; bp->b_offset = ((vm_offset_t)bp->b_data) & PAGE_MASK; if (mapbuf || !unmapped_buf_allowed) { pmap_qenter((vm_offset_t)bp->b_kvabase, bp->b_pages, pidx); bp->b_data = bp->b_kvabase + bp->b_offset; } else bp->b_data = unmapped_buf; return(0); } /* * Free the io map PTEs associated with this IO operation. * We also invalidate the TLB entries and restore the original b_addr. * * This function only works with pager buffers. */ void vunmapbuf(struct buf *bp) { int npages; npages = bp->b_npages; if (buf_mapped(bp)) pmap_qremove(trunc_page((vm_offset_t)bp->b_data), npages); vm_page_unhold_pages(bp->b_pages, npages); bp->b_data = unmapped_buf; } void bdone(struct buf *bp) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); bp->b_flags |= B_DONE; wakeup(bp); mtx_unlock(mtxp); } void bwait(struct buf *bp, u_char pri, const char *wchan) { struct mtx *mtxp; mtxp = mtx_pool_find(mtxpool_sleep, bp); mtx_lock(mtxp); while ((bp->b_flags & B_DONE) == 0) msleep(bp, mtxp, pri, wchan, 0); mtx_unlock(mtxp); } int bufsync(struct bufobj *bo, int waitfor) { return (VOP_FSYNC(bo2vnode(bo), waitfor, curthread)); } void bufstrategy(struct bufobj *bo, struct buf *bp) { int i __unused; struct vnode *vp; vp = bp->b_vp; KASSERT(vp == bo->bo_private, ("Inconsistent vnode bufstrategy")); KASSERT(vp->v_type != VCHR && vp->v_type != VBLK, ("Wrong vnode in bufstrategy(bp=%p, vp=%p)", bp, vp)); i = VOP_STRATEGY(vp, bp); KASSERT(i == 0, ("VOP_STRATEGY failed bp=%p vp=%p", bp, bp->b_vp)); } /* * Initialize a struct bufobj before use. Memory is assumed zero filled. */ void bufobj_init(struct bufobj *bo, void *private) { static volatile int bufobj_cleanq; bo->bo_domain = atomic_fetchadd_int(&bufobj_cleanq, 1) % buf_domains; rw_init(BO_LOCKPTR(bo), "bufobj interlock"); bo->bo_private = private; TAILQ_INIT(&bo->bo_clean.bv_hd); TAILQ_INIT(&bo->bo_dirty.bv_hd); } void bufobj_wrefl(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); ASSERT_BO_WLOCKED(bo); bo->bo_numoutput++; } void bufobj_wref(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wref")); BO_LOCK(bo); bo->bo_numoutput++; BO_UNLOCK(bo); } void bufobj_wdrop(struct bufobj *bo) { KASSERT(bo != NULL, ("NULL bo in bufobj_wdrop")); BO_LOCK(bo); KASSERT(bo->bo_numoutput > 0, ("bufobj_wdrop non-positive count")); if ((--bo->bo_numoutput == 0) && (bo->bo_flag & BO_WWAIT)) { bo->bo_flag &= ~BO_WWAIT; wakeup(&bo->bo_numoutput); } BO_UNLOCK(bo); } int bufobj_wwait(struct bufobj *bo, int slpflag, int timeo) { int error; KASSERT(bo != NULL, ("NULL bo in bufobj_wwait")); ASSERT_BO_WLOCKED(bo); error = 0; while (bo->bo_numoutput) { bo->bo_flag |= BO_WWAIT; error = msleep(&bo->bo_numoutput, BO_LOCKPTR(bo), slpflag | (PRIBIO + 1), "bo_wwait", timeo); if (error) break; } return (error); } /* * Set bio_data or bio_ma for struct bio from the struct buf. */ void bdata2bio(struct buf *bp, struct bio *bip) { if (!buf_mapped(bp)) { KASSERT(unmapped_buf_allowed, ("unmapped")); bip->bio_ma = bp->b_pages; bip->bio_ma_n = bp->b_npages; bip->bio_data = unmapped_buf; bip->bio_ma_offset = (vm_offset_t)bp->b_offset & PAGE_MASK; bip->bio_flags |= BIO_UNMAPPED; KASSERT(round_page(bip->bio_ma_offset + bip->bio_length) / PAGE_SIZE == bp->b_npages, ("Buffer %p too short: %d %lld %d", bp, bip->bio_ma_offset, (long long)bip->bio_length, bip->bio_ma_n)); } else { bip->bio_data = bp->b_data; bip->bio_ma = NULL; } } /* * The MIPS pmap code currently doesn't handle aliased pages. * The VIPT caches may not handle page aliasing themselves, leading * to data corruption. * * As such, this code makes a system extremely unhappy if said * system doesn't support unaliasing the above situation in hardware. * Some "recent" systems (eg some mips24k/mips74k cores) don't enable * this feature at build time, so it has to be handled in software. * * Once the MIPS pmap/cache code grows to support this function on * earlier chips, it should be flipped back off. */ #ifdef __mips__ static int buf_pager_relbuf = 1; #else static int buf_pager_relbuf = 0; #endif SYSCTL_INT(_vfs, OID_AUTO, buf_pager_relbuf, CTLFLAG_RWTUN, &buf_pager_relbuf, 0, "Make buffer pager release buffers after reading"); /* * The buffer pager. It uses buffer reads to validate pages. * * In contrast to the generic local pager from vm/vnode_pager.c, this * pager correctly and easily handles volumes where the underlying * device block size is greater than the machine page size. The * buffer cache transparently extends the requested page run to be * aligned at the block boundary, and does the necessary bogus page * replacements in the addends to avoid obliterating already valid * pages. * * The only non-trivial issue is that the exclusive busy state for * pages, which is assumed by the vm_pager_getpages() interface, is * incompatible with the VMIO buffer cache's desire to share-busy the * pages. This function performs a trivial downgrade of the pages' * state before reading buffers, and a less trivial upgrade from the * shared-busy to excl-busy state after the read. */ int vfs_bio_getpages(struct vnode *vp, vm_page_t *ma, int count, int *rbehind, int *rahead, vbg_get_lblkno_t get_lblkno, vbg_get_blksize_t get_blksize) { vm_page_t m; vm_object_t object; struct buf *bp; struct mount *mp; daddr_t lbn, lbnp; vm_ooffset_t la, lb, poff, poffe; long bsize; int bo_bs, br_flags, error, i, pgsin, pgsin_a, pgsin_b; bool redo, lpart; object = vp->v_object; mp = vp->v_mount; error = 0; la = IDX_TO_OFF(ma[count - 1]->pindex); if (la >= object->un_pager.vnp.vnp_size) return (VM_PAGER_BAD); /* * Change the meaning of la from where the last requested page starts * to where it ends, because that's the end of the requested region * and the start of the potential read-ahead region. */ la += PAGE_SIZE; lpart = la > object->un_pager.vnp.vnp_size; bo_bs = get_blksize(vp, get_lblkno(vp, IDX_TO_OFF(ma[0]->pindex))); /* * Calculate read-ahead, behind and total pages. */ pgsin = count; lb = IDX_TO_OFF(ma[0]->pindex); pgsin_b = OFF_TO_IDX(lb - rounddown2(lb, bo_bs)); pgsin += pgsin_b; if (rbehind != NULL) *rbehind = pgsin_b; pgsin_a = OFF_TO_IDX(roundup2(la, bo_bs) - la); if (la + IDX_TO_OFF(pgsin_a) >= object->un_pager.vnp.vnp_size) pgsin_a = OFF_TO_IDX(roundup2(object->un_pager.vnp.vnp_size, PAGE_SIZE) - la); pgsin += pgsin_a; if (rahead != NULL) *rahead = pgsin_a; VM_CNT_INC(v_vnodein); VM_CNT_ADD(v_vnodepgsin, pgsin); br_flags = (mp != NULL && (mp->mnt_kern_flag & MNTK_UNMAPPED_BUFS) != 0) ? GB_UNMAPPED : 0; again: for (i = 0; i < count; i++) { if (ma[i] != bogus_page) vm_page_busy_downgrade(ma[i]); } lbnp = -1; for (i = 0; i < count; i++) { m = ma[i]; if (m == bogus_page) continue; /* * Pages are shared busy and the object lock is not * owned, which together allow for the pages' * invalidation. The racy test for validity avoids * useless creation of the buffer for the most typical * case when invalidation is not used in redo or for * parallel read. The shared->excl upgrade loop at * the end of the function catches the race in a * reliable way (protected by the object lock). */ if (vm_page_all_valid(m)) continue; poff = IDX_TO_OFF(m->pindex); poffe = MIN(poff + PAGE_SIZE, object->un_pager.vnp.vnp_size); for (; poff < poffe; poff += bsize) { lbn = get_lblkno(vp, poff); if (lbn == lbnp) goto next_page; lbnp = lbn; bsize = get_blksize(vp, lbn); error = bread_gb(vp, lbn, bsize, curthread->td_ucred, br_flags, &bp); if (error != 0) goto end_pages; if (bp->b_rcred == curthread->td_ucred) { crfree(bp->b_rcred); bp->b_rcred = NOCRED; } if (LIST_EMPTY(&bp->b_dep)) { /* * Invalidation clears m->valid, but * may leave B_CACHE flag if the * buffer existed at the invalidation * time. In this case, recycle the * buffer to do real read on next * bread() after redo. * * Otherwise B_RELBUF is not strictly * necessary, enable to reduce buf * cache pressure. */ if (buf_pager_relbuf || !vm_page_all_valid(m)) bp->b_flags |= B_RELBUF; bp->b_flags &= ~B_NOCACHE; brelse(bp); } else { bqrelse(bp); } } KASSERT(1 /* racy, enable for debugging */ || vm_page_all_valid(m) || i == count - 1, ("buf %d %p invalid", i, m)); if (i == count - 1 && lpart) { if (!vm_page_none_valid(m) && !vm_page_all_valid(m)) vm_page_zero_invalid(m, TRUE); } next_page:; } end_pages: redo = false; for (i = 0; i < count; i++) { if (ma[i] == bogus_page) continue; if (vm_page_busy_tryupgrade(ma[i]) == 0) { vm_page_sunbusy(ma[i]); ma[i] = vm_page_grab_unlocked(object, ma[i]->pindex, VM_ALLOC_NORMAL); } /* * Since the pages were only sbusy while neither the * buffer nor the object lock was held by us, or * reallocated while vm_page_grab() slept for busy * relinguish, they could have been invalidated. * Recheck the valid bits and re-read as needed. * * Note that the last page is made fully valid in the * read loop, and partial validity for the page at * index count - 1 could mean that the page was * invalidated or removed, so we must restart for * safety as well. */ if (!vm_page_all_valid(ma[i])) redo = true; } if (redo && error == 0) goto again; return (error != 0 ? VM_PAGER_ERROR : VM_PAGER_OK); } #include "opt_ddb.h" #ifdef DDB #include /* DDB command to show buffer data */ DB_SHOW_COMMAND(buffer, db_show_buffer) { /* get args */ struct buf *bp = (struct buf *)addr; #ifdef FULL_BUF_TRACKING uint32_t i, j; #endif if (!have_addr) { db_printf("usage: show buffer \n"); return; } db_printf("buf at %p\n", bp); db_printf("b_flags = 0x%b, b_xflags=0x%b\n", (u_int)bp->b_flags, PRINT_BUF_FLAGS, (u_int)bp->b_xflags, PRINT_BUF_XFLAGS); db_printf("b_vflags=0x%b b_ioflags0x%b\n", (u_int)bp->b_vflags, PRINT_BUF_VFLAGS, (u_int)bp->b_ioflags, PRINT_BIO_FLAGS); db_printf( "b_error = %d, b_bufsize = %ld, b_bcount = %ld, b_resid = %ld\n" "b_bufobj = (%p), b_data = %p\n, b_blkno = %jd, b_lblkno = %jd, " "b_vp = %p, b_dep = %p\n", bp->b_error, bp->b_bufsize, bp->b_bcount, bp->b_resid, bp->b_bufobj, bp->b_data, (intmax_t)bp->b_blkno, (intmax_t)bp->b_lblkno, bp->b_vp, bp->b_dep.lh_first); db_printf("b_kvabase = %p, b_kvasize = %d\n", bp->b_kvabase, bp->b_kvasize); if (bp->b_npages) { int i; db_printf("b_npages = %d, pages(OBJ, IDX, PA): ", bp->b_npages); for (i = 0; i < bp->b_npages; i++) { vm_page_t m; m = bp->b_pages[i]; if (m != NULL) db_printf("(%p, 0x%lx, 0x%lx)", m->object, (u_long)m->pindex, (u_long)VM_PAGE_TO_PHYS(m)); else db_printf("( ??? )"); if ((i + 1) < bp->b_npages) db_printf(","); } db_printf("\n"); } BUF_LOCKPRINTINFO(bp); #if defined(FULL_BUF_TRACKING) db_printf("b_io_tracking: b_io_tcnt = %u\n", bp->b_io_tcnt); i = bp->b_io_tcnt % BUF_TRACKING_SIZE; for (j = 1; j <= BUF_TRACKING_SIZE; j++) { if (bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)] == NULL) continue; db_printf(" %2u: %s\n", j, bp->b_io_tracking[BUF_TRACKING_ENTRY(i - j)]); } #elif defined(BUF_TRACKING) db_printf("b_io_tracking: %s\n", bp->b_io_tracking); #endif db_printf(" "); } DB_SHOW_COMMAND(bufqueues, bufqueues) { struct bufdomain *bd; struct buf *bp; long total; int i, j, cnt; db_printf("bqempty: %d\n", bqempty.bq_len); for (i = 0; i < buf_domains; i++) { bd = &bdomain[i]; db_printf("Buf domain %d\n", i); db_printf("\tfreebufs\t%d\n", bd->bd_freebuffers); db_printf("\tlofreebufs\t%d\n", bd->bd_lofreebuffers); db_printf("\thifreebufs\t%d\n", bd->bd_hifreebuffers); db_printf("\n"); db_printf("\tbufspace\t%ld\n", bd->bd_bufspace); db_printf("\tmaxbufspace\t%ld\n", bd->bd_maxbufspace); db_printf("\thibufspace\t%ld\n", bd->bd_hibufspace); db_printf("\tlobufspace\t%ld\n", bd->bd_lobufspace); db_printf("\tbufspacethresh\t%ld\n", bd->bd_bufspacethresh); db_printf("\n"); db_printf("\tnumdirtybuffers\t%d\n", bd->bd_numdirtybuffers); db_printf("\tlodirtybuffers\t%d\n", bd->bd_lodirtybuffers); db_printf("\thidirtybuffers\t%d\n", bd->bd_hidirtybuffers); db_printf("\tdirtybufthresh\t%d\n", bd->bd_dirtybufthresh); db_printf("\n"); total = 0; TAILQ_FOREACH(bp, &bd->bd_cleanq->bq_queue, b_freelist) total += bp->b_bufsize; db_printf("\tcleanq count\t%d (%ld)\n", bd->bd_cleanq->bq_len, total); total = 0; TAILQ_FOREACH(bp, &bd->bd_dirtyq.bq_queue, b_freelist) total += bp->b_bufsize; db_printf("\tdirtyq count\t%d (%ld)\n", bd->bd_dirtyq.bq_len, total); db_printf("\twakeup\t\t%d\n", bd->bd_wanted); db_printf("\tlim\t\t%d\n", bd->bd_lim); db_printf("\tCPU "); for (j = 0; j <= mp_maxid; j++) db_printf("%d, ", bd->bd_subq[j].bq_len); db_printf("\n"); cnt = 0; total = 0; for (j = 0; j < nbuf; j++) if (buf[j].b_domain == i && BUF_ISLOCKED(&buf[j])) { cnt++; total += buf[j].b_bufsize; } db_printf("\tLocked buffers: %d space %ld\n", cnt, total); cnt = 0; total = 0; for (j = 0; j < nbuf; j++) if (buf[j].b_domain == i) { cnt++; total += buf[j].b_bufsize; } db_printf("\tTotal buffers: %d space %ld\n", cnt, total); } } DB_SHOW_COMMAND(lockedbufs, lockedbufs) { struct buf *bp; int i; for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (BUF_ISLOCKED(bp)) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); if (db_pager_quit) break; } } } DB_SHOW_COMMAND(vnodebufs, db_show_vnodebufs) { struct vnode *vp; struct buf *bp; if (!have_addr) { db_printf("usage: show vnodebufs \n"); return; } vp = (struct vnode *)addr; db_printf("Clean buffers:\n"); TAILQ_FOREACH(bp, &vp->v_bufobj.bo_clean.bv_hd, b_bobufs) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } db_printf("Dirty buffers:\n"); TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) { db_show_buffer((uintptr_t)bp, 1, 0, NULL); db_printf("\n"); } } DB_COMMAND(countfreebufs, db_coundfreebufs) { struct buf *bp; int i, used = 0, nfree = 0; if (have_addr) { db_printf("usage: countfreebufs\n"); return; } for (i = 0; i < nbuf; i++) { bp = &buf[i]; if (bp->b_qindex == QUEUE_EMPTY) nfree++; else used++; } db_printf("Counted %d free, %d used (%d tot)\n", nfree, used, nfree + used); db_printf("numfreebuffers is %d\n", numfreebuffers); } #endif /* DDB */ Index: head/sys/kern/vfs_cluster.c =================================================================== --- head/sys/kern/vfs_cluster.c (revision 363070) +++ head/sys/kern/vfs_cluster.c (revision 363071) @@ -1,1083 +1,1083 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1993 * The Regents of the University of California. All rights reserved. * Modifications/enhancements: * Copyright (c) 1995 John S. Dyson. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_debug_cluster.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(CLUSTERDEBUG) static int rcluster= 0; SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "Debug VFS clustering code"); #endif static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer"); static uma_zone_t cluster_pbuf_zone; static void cluster_init(void *); static struct cluster_save *cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags); static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, struct buf *fbp); static void cluster_callback(struct buf *); static int write_behind = 1; SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "Cluster write-behind; 0: disable, 1: enable, 2: backed off"); static int read_max = 64; SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0, "Cluster read-ahead max block count"); static int read_min = 1; SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0, "Cluster read min block count"); SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL); static void cluster_init(void *dummy) { cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2); } /* * Read data to a buf, including read-ahead if we find this to be beneficial. * cluster_read replaces bread. */ int cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size, struct ucred *cred, long totread, int seqcount, int gbflags, struct buf **bpp) { struct buf *bp, *rbp, *reqbp; struct bufobj *bo; struct thread *td; daddr_t blkno, origblkno; int maxra, racluster; int error, ncontig; int i; error = 0; td = curthread; bo = &vp->v_bufobj; if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; /* * Try to limit the amount of read-ahead by a few * ad-hoc parameters. This needs work!!! */ racluster = vp->v_mount->mnt_iosize_max / size; maxra = seqcount; maxra = min(read_max, maxra); maxra = min(nbuf/8, maxra); if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize) maxra = (filesize / size) - lblkno; /* * get the requested block */ error = getblkx(vp, lblkno, lblkno, size, 0, 0, gbflags, &bp); if (error != 0) { *bpp = NULL; return (error); } gbflags &= ~GB_NOSPARSE; origblkno = lblkno; *bpp = reqbp = bp; /* * if it is in the cache, then check to see if the reads have been * sequential. If they have, then try some read-ahead, otherwise * back-off on prospective read-aheads. */ if (bp->b_flags & B_CACHE) { if (!seqcount) { return 0; } else if ((bp->b_flags & B_RAM) == 0) { return 0; } else { bp->b_flags &= ~B_RAM; BO_RLOCK(bo); for (i = 1; i < maxra; i++) { /* * Stop if the buffer does not exist or it * is invalid (about to go away?) */ rbp = gbincore(&vp->v_bufobj, lblkno+i); if (rbp == NULL || (rbp->b_flags & B_INVAL)) break; /* * Set another read-ahead mark so we know * to check again. (If we can lock the * buffer without waiting) */ if ((((i % racluster) == (racluster - 1)) || (i == (maxra - 1))) && (0 == BUF_LOCK(rbp, LK_EXCLUSIVE | LK_NOWAIT, NULL))) { rbp->b_flags |= B_RAM; BUF_UNLOCK(rbp); - } + } } BO_RUNLOCK(bo); if (i >= maxra) { return 0; } lblkno += i; } reqbp = bp = NULL; /* * If it isn't in the cache, then get a chunk from * disk if sequential, otherwise just get the block. */ } else { off_t firstread = bp->b_offset; int nblks; long minread; KASSERT(bp->b_offset != NOOFFSET, ("cluster_read: no buffer offset")); ncontig = 0; /* * Adjust totread if needed */ minread = read_min * size; if (minread > totread) totread = minread; /* * Compute the total number of blocks that we should read * synchronously. */ if (firstread + totread > filesize) totread = filesize - firstread; nblks = howmany(totread, size); if (nblks > racluster) nblks = racluster; /* * Now compute the number of contiguous blocks. */ if (nblks > 1) { error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); /* * If this failed to map just do the original block. */ if (error || blkno == -1) ncontig = 0; } /* * If we have contiguous data available do a cluster * otherwise just read the requested block. */ if (ncontig) { /* Account for our first block. */ ncontig = min(ncontig + 1, nblks); if (ncontig < nblks) nblks = ncontig; bp = cluster_rbuild(vp, filesize, lblkno, blkno, size, nblks, gbflags, bp); lblkno += (bp->b_bufsize / size); } else { bp->b_flags |= B_RAM; bp->b_iocmd = BIO_READ; lblkno += 1; } } /* * handle the synchronous read so that it is available ASAP. */ if (bp) { if ((bp->b_flags & B_CLUSTER) == 0) { vfs_busy_pages(bp, 0); } bp->b_flags &= ~B_INVAL; bp->b_ioflags &= ~BIO_ERROR; if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL) BUF_KERNPROC(bp); bp->b_iooffset = dbtob(bp->b_blkno); bstrategy(bp); #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, bp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; } /* * If we have been doing sequential I/O, then do some read-ahead. */ while (lblkno < (origblkno + maxra)) { error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL); if (error) break; if (blkno == -1) break; /* * We could throttle ncontig here by maxra but we might as * well read the data if it is contiguous. We're throttled * by racluster anyway. */ if (ncontig) { ncontig = min(ncontig + 1, racluster); rbp = cluster_rbuild(vp, filesize, lblkno, blkno, size, ncontig, gbflags, NULL); lblkno += (rbp->b_bufsize / size); if (rbp->b_flags & B_DELWRI) { bqrelse(rbp); continue; } } else { rbp = getblk(vp, lblkno, size, 0, 0, gbflags); lblkno += 1; if (rbp->b_flags & B_DELWRI) { bqrelse(rbp); continue; } rbp->b_flags |= B_ASYNC | B_RAM; rbp->b_iocmd = BIO_READ; rbp->b_blkno = blkno; } if (rbp->b_flags & B_CACHE) { rbp->b_flags &= ~B_ASYNC; bqrelse(rbp); continue; } if ((rbp->b_flags & B_CLUSTER) == 0) { vfs_busy_pages(rbp, 0); } rbp->b_flags &= ~B_INVAL; rbp->b_ioflags &= ~BIO_ERROR; if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL) BUF_KERNPROC(rbp); rbp->b_iooffset = dbtob(rbp->b_blkno); bstrategy(rbp); #ifdef RACCT if (racct_enable) { PROC_LOCK(td->td_proc); racct_add_buf(td->td_proc, rbp, 0); PROC_UNLOCK(td->td_proc); } #endif /* RACCT */ td->td_ru.ru_inblock++; } if (reqbp) { /* * Like bread, always brelse() the buffer when * returning an error. */ error = bufwait(reqbp); if (error != 0) { brelse(reqbp); *bpp = NULL; } } return (error); } /* * If blocks are contiguous on disk, use this to provide clustered * read ahead. We will read as many blocks as possible sequentially * and then parcel them up into logical blocks in the buffer hash table. */ static struct buf * cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn, daddr_t blkno, long size, int run, int gbflags, struct buf *fbp) { struct buf *bp, *tbp; daddr_t bn; off_t off; long tinc, tsize; int i, inc, j, k, toff; KASSERT(size == vp->v_mount->mnt_stat.f_iosize, ("cluster_rbuild: size %ld != f_iosize %jd\n", size, (intmax_t)vp->v_mount->mnt_stat.f_iosize)); /* * avoid a division */ while ((u_quad_t) size * (lbn + run) > filesize) { --run; } if (fbp) { tbp = fbp; tbp->b_iocmd = BIO_READ; } else { tbp = getblk(vp, lbn, size, 0, 0, gbflags); if (tbp->b_flags & B_CACHE) return tbp; tbp->b_flags |= B_ASYNC | B_RAM; tbp->b_iocmd = BIO_READ; } tbp->b_blkno = blkno; if( (tbp->b_flags & B_MALLOC) || ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) ) return tbp; bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT); if (bp == NULL) return tbp; /* * We are synthesizing a buffer out of vm_page_t's, but * if the block size is not page aligned then the starting * address may not be either. Inherit the b_data offset * from the original buffer. */ bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO; if ((gbflags & GB_UNMAPPED) != 0) { bp->b_data = unmapped_buf; } else { bp->b_data = (char *)((vm_offset_t)bp->b_data | ((vm_offset_t)tbp->b_data & PAGE_MASK)); } bp->b_iocmd = BIO_READ; bp->b_iodone = cluster_callback; bp->b_blkno = blkno; bp->b_lblkno = lbn; bp->b_offset = tbp->b_offset; KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset")); pbgetvp(vp, bp); TAILQ_INIT(&bp->b_cluster.cluster_head); bp->b_bcount = 0; bp->b_bufsize = 0; bp->b_npages = 0; inc = btodb(size); for (bn = blkno, i = 0; i < run; ++i, bn += inc) { if (i == 0) { vm_object_pip_add(tbp->b_bufobj->bo_object, tbp->b_npages); vfs_busy_pages_acquire(tbp); } else { if ((bp->b_npages * PAGE_SIZE) + round_page(size) > vp->v_mount->mnt_iosize_max) { break; } tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT | (gbflags & GB_UNMAPPED)); /* Don't wait around for locked bufs. */ if (tbp == NULL) break; /* * Stop scanning if the buffer is fully valid * (marked B_CACHE), or locked (may be doing a * background write), or if the buffer is not * VMIO backed. The clustering code can only deal * with VMIO-backed buffers. The bo lock is not * required for the BKGRDINPROG check since it * can not be set without the buf lock. */ if ((tbp->b_vflags & BV_BKGRDINPROG) || (tbp->b_flags & B_CACHE) || (tbp->b_flags & B_VMIO) == 0) { bqrelse(tbp); break; } /* * The buffer must be completely invalid in order to * take part in the cluster. If it is partially valid * then we stop. */ off = tbp->b_offset; tsize = size; for (j = 0; tsize > 0; j++) { toff = off & PAGE_MASK; tinc = tsize; if (toff + tinc > PAGE_SIZE) tinc = PAGE_SIZE - toff; if (vm_page_trysbusy(tbp->b_pages[j]) == 0) break; if ((tbp->b_pages[j]->valid & vm_page_bits(toff, tinc)) != 0) { vm_page_sunbusy(tbp->b_pages[j]); break; } vm_object_pip_add(tbp->b_bufobj->bo_object, 1); off += tinc; tsize -= tinc; } if (tsize > 0) { clean_sbusy: vm_object_pip_wakeupn(tbp->b_bufobj->bo_object, j); for (k = 0; k < j; k++) vm_page_sunbusy(tbp->b_pages[k]); bqrelse(tbp); break; } /* * Set a read-ahead mark as appropriate */ if ((fbp && (i == 1)) || (i == (run - 1))) tbp->b_flags |= B_RAM; /* * Set the buffer up for an async read (XXX should * we do this only if we do not wind up brelse()ing?). * Set the block number if it isn't set, otherwise * if it is make sure it matches the block number we * expect. */ tbp->b_flags |= B_ASYNC; tbp->b_iocmd = BIO_READ; if (tbp->b_blkno == tbp->b_lblkno) { tbp->b_blkno = bn; } else if (tbp->b_blkno != bn) { goto clean_sbusy; } } /* * XXX fbp from caller may not be B_ASYNC, but we are going * to biodone() it in cluster_callback() anyway */ BUF_KERNPROC(tbp); TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, tbp, b_cluster.cluster_entry); for (j = 0; j < tbp->b_npages; j += 1) { vm_page_t m; m = tbp->b_pages[j]; if ((bp->b_npages == 0) || (bp->b_pages[bp->b_npages-1] != m)) { bp->b_pages[bp->b_npages] = m; bp->b_npages++; } if (vm_page_all_valid(m)) tbp->b_pages[j] = bogus_page; } /* * Don't inherit tbp->b_bufsize as it may be larger due to * a non-page-aligned size. Instead just aggregate using * 'size'. */ if (tbp->b_bcount != size) printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size); if (tbp->b_bufsize != size) printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size); bp->b_bcount += size; bp->b_bufsize += size; } /* * Fully valid pages in the cluster are already good and do not need * to be re-read from disk. Replace the page with bogus_page */ for (j = 0; j < bp->b_npages; j++) { if (vm_page_all_valid(bp->b_pages[j])) bp->b_pages[j] = bogus_page; } if (bp->b_bufsize > bp->b_kvasize) panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n", bp->b_bufsize, bp->b_kvasize); if (buf_mapped(bp)) { pmap_qenter(trunc_page((vm_offset_t) bp->b_data), (vm_page_t *)bp->b_pages, bp->b_npages); } return (bp); } /* * Cleanup after a clustered read or write. * This is complicated by the fact that any of the buffers might have * extra memory (if there were no empty buffer headers at allocbuf time) * that we will need to shift around. */ static void cluster_callback(struct buf *bp) { struct buf *nbp, *tbp; int error = 0; /* * Must propagate errors to all the components. */ if (bp->b_ioflags & BIO_ERROR) error = bp->b_error; if (buf_mapped(bp)) { pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages); } /* * Move memory from the large cluster buffer into the component * buffers and mark IO as done on these. */ for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head); tbp; tbp = nbp) { nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry); if (error) { tbp->b_ioflags |= BIO_ERROR; tbp->b_error = error; } else { tbp->b_dirtyoff = tbp->b_dirtyend = 0; tbp->b_flags &= ~B_INVAL; tbp->b_ioflags &= ~BIO_ERROR; /* * XXX the bdwrite()/bqrelse() issued during * cluster building clears B_RELBUF (see bqrelse() * comment). If direct I/O was specified, we have * to restore it here to allow the buffer and VM * to be freed. */ if (tbp->b_flags & B_DIRECT) tbp->b_flags |= B_RELBUF; } bufdone(tbp); } pbrelvp(bp); uma_zfree(cluster_pbuf_zone, bp); } /* * cluster_wbuild_wb: * * Implement modified write build for cluster. * * write_behind = 0 write behind disabled * write_behind = 1 write behind normal (default) * write_behind = 2 write behind backed-off */ static __inline int cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len, int gbflags) { int r = 0; switch (write_behind) { case 2: if (start_lbn < len) break; start_lbn -= len; /* FALLTHROUGH */ case 1: r = cluster_wbuild(vp, size, start_lbn, len, gbflags); /* FALLTHROUGH */ default: /* FALLTHROUGH */ break; } return(r); } /* * Do clustered write for FFS. * * Three cases: * 1. Write is not sequential (write asynchronously) * Write is sequential: * 2. beginning of cluster - begin cluster * 3. middle of a cluster - add to cluster * 4. end of a cluster - asynchronously write cluster */ void cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount, int gbflags) { daddr_t lbn; int maxclen, cursize; int lblocksize; int async; if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; if (vp->v_type == VREG) { async = DOINGASYNC(vp); lblocksize = vp->v_mount->mnt_stat.f_iosize; } else { async = 0; lblocksize = bp->b_bufsize; } lbn = bp->b_lblkno; KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset")); /* Initialize vnode to beginning of file. */ if (lbn == 0) vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 || (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) { maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1; if (vp->v_clen != 0) { /* * Next block is not sequential. * * If we are not writing at end of file, the process * seeked to another point in the file since its last * write, or we have reached our maximum cluster size, * then push the previous cluster. Otherwise try * reallocating to make it sequential. * * Change to algorithm: only push previous cluster if * it was sequential from the point of view of the * seqcount heuristic, otherwise leave the buffer * intact so we can potentially optimize the I/O * later on in the buf_daemon or update daemon * flush. */ cursize = vp->v_lastw - vp->v_cstart + 1; if (((u_quad_t) bp->b_offset + lblocksize) != filesize || lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) { if (!async && seqcount > 0) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, cursize, gbflags); } } else { struct buf **bpp, **endbp; struct cluster_save *buflist; buflist = cluster_collectbufs(vp, bp, gbflags); if (buflist == NULL) { /* * Cluster build failed so just write * it now. */ bawrite(bp); return; } endbp = &buflist->bs_children [buflist->bs_nchildren - 1]; if (VOP_REALLOCBLKS(vp, buflist)) { /* * Failed, push the previous cluster * if *really* writing sequentially * in the logical file (seqcount > 1), * otherwise delay it in the hopes that * the low level disk driver can * optimize the write ordering. */ for (bpp = buflist->bs_children; bpp < endbp; bpp++) brelse(*bpp); free(buflist, M_SEGMENT); if (seqcount > 1) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, cursize, gbflags); } } else { /* * Succeeded, keep building cluster. */ for (bpp = buflist->bs_children; bpp <= endbp; bpp++) bdwrite(*bpp); free(buflist, M_SEGMENT); vp->v_lastw = lbn; vp->v_lasta = bp->b_blkno; return; } } } /* * Consider beginning a cluster. If at end of file, make * cluster as large as possible, otherwise find size of * existing cluster. */ if ((vp->v_type == VREG) && ((u_quad_t) bp->b_offset + lblocksize) != filesize && (bp->b_blkno == bp->b_lblkno) && (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) || bp->b_blkno == -1)) { bawrite(bp); vp->v_clen = 0; vp->v_lasta = bp->b_blkno; vp->v_cstart = lbn + 1; vp->v_lastw = lbn; return; } vp->v_clen = maxclen; if (!async && maxclen == 0) { /* I/O not contiguous */ vp->v_cstart = lbn + 1; bawrite(bp); } else { /* Wait for rest of cluster */ vp->v_cstart = lbn; bdwrite(bp); } } else if (lbn == vp->v_cstart + vp->v_clen) { /* * At end of cluster, write it out if seqcount tells us we * are operating sequentially, otherwise let the buf or * update daemon handle it. */ bdwrite(bp); if (seqcount > 1) { cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1, gbflags); } vp->v_clen = 0; vp->v_cstart = lbn + 1; } else if (vm_page_count_severe()) { /* * We are low on memory, get it going NOW */ bawrite(bp); } else { /* * In the middle of a cluster, so just delay the I/O for now. */ bdwrite(bp); } vp->v_lastw = lbn; vp->v_lasta = bp->b_blkno; } /* * This is an awful lot like cluster_rbuild...wish they could be combined. * The last lbn argument is the current block on which I/O is being * performed. Check to see that it doesn't fall in the middle of * the current block (if last_bp == NULL). */ int cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len, int gbflags) { struct buf *bp, *tbp; struct bufobj *bo; int i, j; int totalwritten = 0; int dbsize = btodb(size); if (!unmapped_buf_allowed) gbflags &= ~GB_UNMAPPED; bo = &vp->v_bufobj; while (len > 0) { /* * If the buffer is not delayed-write (i.e. dirty), or it * is delayed-write but either locked or inval, it cannot * partake in the clustered write. */ BO_LOCK(bo); if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL || (tbp->b_vflags & BV_BKGRDINPROG)) { BO_UNLOCK(bo); ++start_lbn; --len; continue; } if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) { ++start_lbn; --len; continue; } if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) { BUF_UNLOCK(tbp); ++start_lbn; --len; continue; } bremfree(tbp); tbp->b_flags &= ~B_DONE; /* * Extra memory in the buffer, punt on this buffer. * XXX we could handle this in most cases, but we would * have to push the extra memory down to after our max * possible cluster size and then potentially pull it back * up if the cluster was terminated prematurely--too much * hassle. */ if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) != (B_CLUSTEROK | B_VMIO)) || (tbp->b_bcount != tbp->b_bufsize) || (tbp->b_bcount != size) || (len == 1) || ((bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT)) == NULL)) { totalwritten += tbp->b_bufsize; bawrite(tbp); ++start_lbn; --len; continue; } /* * We got a pbuf to make the cluster in. * so initialise it. */ TAILQ_INIT(&bp->b_cluster.cluster_head); bp->b_bcount = 0; bp->b_bufsize = 0; bp->b_npages = 0; if (tbp->b_wcred != NOCRED) bp->b_wcred = crhold(tbp->b_wcred); bp->b_blkno = tbp->b_blkno; bp->b_lblkno = tbp->b_lblkno; bp->b_offset = tbp->b_offset; /* * We are synthesizing a buffer out of vm_page_t's, but * if the block size is not page aligned then the starting * address may not be either. Inherit the b_data offset * from the original buffer. */ if ((gbflags & GB_UNMAPPED) == 0 || (tbp->b_flags & B_VMIO) == 0) { bp->b_data = (char *)((vm_offset_t)bp->b_data | ((vm_offset_t)tbp->b_data & PAGE_MASK)); } else { bp->b_data = unmapped_buf; } bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT)); bp->b_iodone = cluster_callback; pbgetvp(vp, bp); /* * From this location in the file, scan forward to see * if there are buffers with adjacent data that need to * be written as well. */ for (i = 0; i < len; ++i, ++start_lbn) { if (i != 0) { /* If not the first buffer */ /* * If the adjacent data is not even in core it * can't need to be written. */ BO_LOCK(bo); if ((tbp = gbincore(bo, start_lbn)) == NULL || (tbp->b_vflags & BV_BKGRDINPROG)) { BO_UNLOCK(bo); break; } /* * If it IS in core, but has different * characteristics, or is locked (which * means it could be undergoing a background * I/O or be in a weird state), then don't * cluster with it. */ if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) break; if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK | B_INVAL | B_DELWRI | B_NEEDCOMMIT)) != (B_DELWRI | B_CLUSTEROK | (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) || tbp->b_wcred != bp->b_wcred) { BUF_UNLOCK(tbp); break; } /* * Check that the combined cluster * would make sense with regard to pages * and would not be too large */ if ((tbp->b_bcount != size) || ((bp->b_blkno + (dbsize * i)) != tbp->b_blkno) || ((tbp->b_npages + bp->b_npages) > (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) { BUF_UNLOCK(tbp); break; } /* * Ok, it's passed all the tests, * so remove it from the free list * and mark it busy. We will use it. */ bremfree(tbp); tbp->b_flags &= ~B_DONE; } /* end of code for non-first buffers only */ /* * If the IO is via the VM then we do some * special VM hackery (yuck). Since the buffer's * block size may not be page-aligned it is possible * for a page to be shared between two buffers. We * have to get rid of the duplication when building * the cluster. */ if (tbp->b_flags & B_VMIO) { vm_page_t m; if (i == 0) { vfs_busy_pages_acquire(tbp); } else { /* if not first buffer */ for (j = 0; j < tbp->b_npages; j += 1) { m = tbp->b_pages[j]; if (vm_page_trysbusy(m) == 0) { for (j--; j >= 0; j--) vm_page_sunbusy( tbp->b_pages[j]); bqrelse(tbp); goto finishcluster; } } } vm_object_pip_add(tbp->b_bufobj->bo_object, tbp->b_npages); for (j = 0; j < tbp->b_npages; j += 1) { m = tbp->b_pages[j]; if ((bp->b_npages == 0) || (bp->b_pages[bp->b_npages - 1] != m)) { bp->b_pages[bp->b_npages] = m; bp->b_npages++; } } } bp->b_bcount += size; bp->b_bufsize += size; /* * If any of the clustered buffers have their * B_BARRIER flag set, transfer that request to * the cluster. */ bp->b_flags |= (tbp->b_flags & B_BARRIER); tbp->b_flags &= ~(B_DONE | B_BARRIER); tbp->b_flags |= B_ASYNC; tbp->b_ioflags &= ~BIO_ERROR; tbp->b_iocmd = BIO_WRITE; bundirty(tbp); reassignbuf(tbp); /* put on clean list */ bufobj_wref(tbp->b_bufobj); BUF_KERNPROC(tbp); buf_track(tbp, __func__); TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head, tbp, b_cluster.cluster_entry); } finishcluster: if (buf_mapped(bp)) { pmap_qenter(trunc_page((vm_offset_t) bp->b_data), (vm_page_t *)bp->b_pages, bp->b_npages); } if (bp->b_bufsize > bp->b_kvasize) panic( "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n", bp->b_bufsize, bp->b_kvasize); totalwritten += bp->b_bufsize; bp->b_dirtyoff = 0; bp->b_dirtyend = bp->b_bufsize; bawrite(bp); len -= i; } return totalwritten; } /* * Collect together all the buffers in a cluster. * Plus add one additional buffer. */ static struct cluster_save * cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags) { struct cluster_save *buflist; struct buf *bp; daddr_t lbn; int i, j, len, error; len = vp->v_lastw - vp->v_cstart + 1; buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist), M_SEGMENT, M_WAITOK); buflist->bs_nchildren = 0; buflist->bs_children = (struct buf **) (buflist + 1); for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) { error = bread_gb(vp, lbn, last_bp->b_bcount, NOCRED, gbflags, &bp); if (error != 0) { /* * If read fails, release collected buffers * and return failure. */ for (j = 0; j < i; j++) brelse(buflist->bs_children[j]); free(buflist, M_SEGMENT); return (NULL); } buflist->bs_children[i] = bp; if (bp->b_blkno == bp->b_lblkno) VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); } buflist->bs_children[i] = bp = last_bp; if (bp->b_blkno == bp->b_lblkno) VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL); buflist->bs_nchildren = i + 1; return (buflist); } Index: head/sys/kern/vfs_export.c =================================================================== --- head/sys/kern/vfs_export.c (revision 363070) +++ head/sys/kern/vfs_export.c (revision 363071) @@ -1,544 +1,544 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_NETADDR, "export_host", "Export host address structure"); #if defined(INET) || defined(INET6) static struct radix_node_head *vfs_create_addrlist_af( struct radix_node_head **prnh, int off); #endif static void vfs_free_addrlist(struct netexport *nep); static int vfs_free_netcred(struct radix_node *rn, void *w); static void vfs_free_addrlist_af(struct radix_node_head **prnh); static int vfs_hang_addrlist(struct mount *mp, struct netexport *nep, struct export_args *argp); static struct netcred *vfs_export_lookup(struct mount *, struct sockaddr *); /* * Network address lookup element */ struct netcred { struct radix_node netc_rnodes[2]; uint64_t netc_exflags; struct ucred *netc_anon; int netc_numsecflavors; int netc_secflavors[MAXSECFLAVORS]; }; /* * Network export information */ struct netexport { struct netcred ne_defexported; /* Default export */ struct radix_node_head *ne4; struct radix_node_head *ne6; }; /* * Build hash lists of net addresses and hang them off the mount point. * Called by vfs_export() to set up the lists of export addresses. */ static int vfs_hang_addrlist(struct mount *mp, struct netexport *nep, struct export_args *argp) { struct netcred *np; struct radix_node_head *rnh; int i; struct radix_node *rn; struct sockaddr *saddr, *smask = NULL; #if defined(INET6) || defined(INET) int off; #endif int error; KASSERT(argp->ex_numsecflavors > 0, ("%s: numsecflavors <= 0", __func__)); KASSERT(argp->ex_numsecflavors < MAXSECFLAVORS, ("%s: numsecflavors >= MAXSECFLAVORS", __func__)); /* * XXX: This routine converts from a uid plus gid list * to a `struct ucred' (np->netc_anon). This * operation is questionable; for example, what should be done * with fields like cr_uidinfo and cr_prison? Currently, this * routine does not touch them (leaves them as NULL). */ if (argp->ex_addrlen == 0) { if (mp->mnt_flag & MNT_DEFEXPORTED) { vfs_mount_error(mp, "MNT_DEFEXPORTED already set for mount %p", mp); return (EPERM); } np = &nep->ne_defexported; np->netc_exflags = argp->ex_flags; np->netc_anon = crget(); np->netc_anon->cr_uid = argp->ex_uid; crsetgroups(np->netc_anon, argp->ex_ngroups, argp->ex_groups); np->netc_anon->cr_prison = &prison0; prison_hold(np->netc_anon->cr_prison); np->netc_numsecflavors = argp->ex_numsecflavors; bcopy(argp->ex_secflavors, np->netc_secflavors, sizeof(np->netc_secflavors)); MNT_ILOCK(mp); mp->mnt_flag |= MNT_DEFEXPORTED; MNT_IUNLOCK(mp); return (0); } #if MSIZE <= 256 if (argp->ex_addrlen > MLEN) { vfs_mount_error(mp, "ex_addrlen %d is greater than %d", argp->ex_addrlen, MLEN); return (EINVAL); } #endif i = sizeof(struct netcred) + argp->ex_addrlen + argp->ex_masklen; np = (struct netcred *) malloc(i, M_NETADDR, M_WAITOK | M_ZERO); saddr = (struct sockaddr *) (np + 1); if ((error = copyin(argp->ex_addr, saddr, argp->ex_addrlen))) goto out; if (saddr->sa_family == AF_UNSPEC || saddr->sa_family > AF_MAX) { error = EINVAL; vfs_mount_error(mp, "Invalid saddr->sa_family: %d"); goto out; } if (saddr->sa_len > argp->ex_addrlen) saddr->sa_len = argp->ex_addrlen; if (argp->ex_masklen) { smask = (struct sockaddr *)((caddr_t)saddr + argp->ex_addrlen); error = copyin(argp->ex_mask, smask, argp->ex_masklen); if (error) goto out; if (smask->sa_len > argp->ex_masklen) smask->sa_len = argp->ex_masklen; } rnh = NULL; switch (saddr->sa_family) { #ifdef INET case AF_INET: if ((rnh = nep->ne4) == NULL) { off = offsetof(struct sockaddr_in, sin_addr) << 3; rnh = vfs_create_addrlist_af(&nep->ne4, off); } break; #endif #ifdef INET6 case AF_INET6: if ((rnh = nep->ne6) == NULL) { off = offsetof(struct sockaddr_in6, sin6_addr) << 3; rnh = vfs_create_addrlist_af(&nep->ne6, off); } break; #endif } if (rnh == NULL) { error = ENOBUFS; vfs_mount_error(mp, "%s %s %d", "Unable to initialize radix node head ", "for address family", saddr->sa_family); goto out; } RADIX_NODE_HEAD_LOCK(rnh); rn = (*rnh->rnh_addaddr)(saddr, smask, &rnh->rh, np->netc_rnodes); RADIX_NODE_HEAD_UNLOCK(rnh); if (rn == NULL || np != (struct netcred *)rn) { /* already exists */ error = EPERM; vfs_mount_error(mp, "netcred already exists for given addr/mask"); goto out; } np->netc_exflags = argp->ex_flags; np->netc_anon = crget(); np->netc_anon->cr_uid = argp->ex_uid; crsetgroups(np->netc_anon, argp->ex_ngroups, argp->ex_groups); np->netc_anon->cr_prison = &prison0; prison_hold(np->netc_anon->cr_prison); np->netc_numsecflavors = argp->ex_numsecflavors; bcopy(argp->ex_secflavors, np->netc_secflavors, sizeof(np->netc_secflavors)); return (0); out: free(np, M_NETADDR); return (error); } /* Helper for vfs_free_addrlist. */ /* ARGSUSED */ static int vfs_free_netcred(struct radix_node *rn, void *w) { struct radix_node_head *rnh = (struct radix_node_head *) w; struct ucred *cred; (*rnh->rnh_deladdr) (rn->rn_key, rn->rn_mask, &rnh->rh); cred = ((struct netcred *)rn)->netc_anon; if (cred != NULL) crfree(cred); free(rn, M_NETADDR); return (0); } #if defined(INET) || defined(INET6) static struct radix_node_head * vfs_create_addrlist_af(struct radix_node_head **prnh, int off) { if (rn_inithead((void **)prnh, off) == 0) return (NULL); RADIX_NODE_HEAD_LOCK_INIT(*prnh); return (*prnh); } #endif static void vfs_free_addrlist_af(struct radix_node_head **prnh) { struct radix_node_head *rnh; rnh = *prnh; RADIX_NODE_HEAD_LOCK(rnh); (*rnh->rnh_walktree)(&rnh->rh, vfs_free_netcred, rnh); RADIX_NODE_HEAD_UNLOCK(rnh); RADIX_NODE_HEAD_DESTROY(rnh); rn_detachhead((void **)prnh); prnh = NULL; } /* * Free the net address hash lists that are hanging off the mount points. */ static void vfs_free_addrlist(struct netexport *nep) { struct ucred *cred; if (nep->ne4 != NULL) vfs_free_addrlist_af(&nep->ne4); if (nep->ne6 != NULL) vfs_free_addrlist_af(&nep->ne6); cred = nep->ne_defexported.netc_anon; if (cred != NULL) crfree(cred); } /* * High level function to manipulate export options on a mount point * and the passed in netexport. * Struct export_args *argp is the variable used to twiddle options, * the structure is described in sys/mount.h */ int vfs_export(struct mount *mp, struct export_args *argp) { struct netexport *nep; int error; if ((argp->ex_flags & (MNT_DELEXPORT | MNT_EXPORTED)) == 0) return (EINVAL); if ((argp->ex_flags & MNT_EXPORTED) != 0 && (argp->ex_numsecflavors < 0 || argp->ex_numsecflavors >= MAXSECFLAVORS)) return (EINVAL); error = 0; lockmgr(&mp->mnt_explock, LK_EXCLUSIVE, NULL); nep = mp->mnt_export; if (argp->ex_flags & MNT_DELEXPORT) { if (nep == NULL) { error = ENOENT; goto out; } if (mp->mnt_flag & MNT_EXPUBLIC) { vfs_setpublicfs(NULL, NULL, NULL); MNT_ILOCK(mp); mp->mnt_flag &= ~MNT_EXPUBLIC; MNT_IUNLOCK(mp); } vfs_free_addrlist(nep); mp->mnt_export = NULL; free(nep, M_MOUNT); nep = NULL; MNT_ILOCK(mp); mp->mnt_flag &= ~(MNT_EXPORTED | MNT_DEFEXPORTED); MNT_IUNLOCK(mp); } if (argp->ex_flags & MNT_EXPORTED) { if (nep == NULL) { nep = malloc(sizeof(struct netexport), M_MOUNT, M_WAITOK | M_ZERO); mp->mnt_export = nep; } if (argp->ex_flags & MNT_EXPUBLIC) { if ((error = vfs_setpublicfs(mp, nep, argp)) != 0) goto out; MNT_ILOCK(mp); mp->mnt_flag |= MNT_EXPUBLIC; MNT_IUNLOCK(mp); } if (argp->ex_numsecflavors == 0) { argp->ex_numsecflavors = 1; argp->ex_secflavors[0] = AUTH_SYS; } if ((error = vfs_hang_addrlist(mp, nep, argp))) goto out; MNT_ILOCK(mp); mp->mnt_flag |= MNT_EXPORTED; MNT_IUNLOCK(mp); } out: lockmgr(&mp->mnt_explock, LK_RELEASE, NULL); /* * Once we have executed the vfs_export() command, we do * not want to keep the "export" option around in the * options list, since that will cause subsequent MNT_UPDATE * calls to fail. The export information is saved in * mp->mnt_export, so we can safely delete the "export" mount option * here. */ vfs_deleteopt(mp->mnt_optnew, "export"); vfs_deleteopt(mp->mnt_opt, "export"); return (error); } /* * Set the publicly exported filesystem (WebNFS). Currently, only * one public filesystem is possible in the spec (RFC 2054 and 2055) */ int vfs_setpublicfs(struct mount *mp, struct netexport *nep, struct export_args *argp) { int error; struct vnode *rvp; char *cp; /* * mp == NULL -> invalidate the current info, the FS is * no longer exported. May be called from either vfs_export * or unmount, so check if it hasn't already been done. */ if (mp == NULL) { if (nfs_pub.np_valid) { nfs_pub.np_valid = 0; if (nfs_pub.np_index != NULL) { free(nfs_pub.np_index, M_TEMP); nfs_pub.np_index = NULL; } } return (0); } /* * Only one allowed at a time. */ if (nfs_pub.np_valid != 0 && mp != nfs_pub.np_mount) return (EBUSY); /* * Get real filehandle for root of exported FS. */ bzero(&nfs_pub.np_handle, sizeof(nfs_pub.np_handle)); nfs_pub.np_handle.fh_fsid = mp->mnt_stat.f_fsid; if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rvp))) return (error); if ((error = VOP_VPTOFH(rvp, &nfs_pub.np_handle.fh_fid))) return (error); vput(rvp); /* * If an indexfile was specified, pull it in. */ if (argp->ex_indexfile != NULL) { if (nfs_pub.np_index == NULL) nfs_pub.np_index = malloc(MAXNAMLEN + 1, M_TEMP, M_WAITOK); error = copyinstr(argp->ex_indexfile, nfs_pub.np_index, MAXNAMLEN, (size_t *)0); if (!error) { /* * Check for illegal filenames. */ for (cp = nfs_pub.np_index; *cp; cp++) { if (*cp == '/') { error = EINVAL; break; } } } if (error) { free(nfs_pub.np_index, M_TEMP); nfs_pub.np_index = NULL; return (error); } } nfs_pub.np_mount = mp; nfs_pub.np_valid = 1; return (0); } /* * Used by the filesystems to determine if a given network address * (passed in 'nam') is present in their exports list, returns a pointer * to struct netcred so that the filesystem can examine it for * access rights (read/write/etc). */ static struct netcred * vfs_export_lookup(struct mount *mp, struct sockaddr *nam) { RADIX_NODE_HEAD_RLOCK_TRACKER; struct netexport *nep; struct netcred *np = NULL; struct radix_node_head *rnh; struct sockaddr *saddr; nep = mp->mnt_export; if (nep == NULL) return (NULL); if ((mp->mnt_flag & MNT_EXPORTED) == 0) return (NULL); /* * Lookup in the export list */ if (nam != NULL) { saddr = nam; rnh = NULL; switch (saddr->sa_family) { case AF_INET: rnh = nep->ne4; break; case AF_INET6: rnh = nep->ne6; break; } if (rnh != NULL) { RADIX_NODE_HEAD_RLOCK(rnh); np = (struct netcred *) (*rnh->rnh_matchaddr)(saddr, &rnh->rh); RADIX_NODE_HEAD_RUNLOCK(rnh); if (np != NULL && (np->netc_rnodes->rn_flags & RNF_ROOT) != 0) return (NULL); } } /* * If no address match, use the default if it exists. */ if (np == NULL && (mp->mnt_flag & MNT_DEFEXPORTED) != 0) return (&nep->ne_defexported); return (np); } /* * XXX: This comment comes from the deprecated ufs_check_export() * XXX: and may not entirely apply, but lacking something better: * This is the generic part of fhtovp called after the underlying * filesystem has validated the file handle. * * Verify that a host should have access to a filesystem. */ -int +int vfs_stdcheckexp(struct mount *mp, struct sockaddr *nam, uint64_t *extflagsp, struct ucred **credanonp, int *numsecflavors, int *secflavors) { struct netcred *np; lockmgr(&mp->mnt_explock, LK_SHARED, NULL); np = vfs_export_lookup(mp, nam); if (np == NULL) { lockmgr(&mp->mnt_explock, LK_RELEASE, NULL); *credanonp = NULL; return (EACCES); } *extflagsp = np->netc_exflags; if ((*credanonp = np->netc_anon) != NULL) crhold(*credanonp); if (numsecflavors) { *numsecflavors = np->netc_numsecflavors; KASSERT(*numsecflavors > 0, ("%s: numsecflavors <= 0", __func__)); KASSERT(*numsecflavors < MAXSECFLAVORS, ("%s: numsecflavors >= MAXSECFLAVORS", __func__)); } if (secflavors && np->netc_numsecflavors > 0) memcpy(secflavors, np->netc_secflavors, np->netc_numsecflavors * sizeof(int)); lockmgr(&mp->mnt_explock, LK_RELEASE, NULL); return (0); } Index: head/sys/kern/vfs_hash.c =================================================================== --- head/sys/kern/vfs_hash.c (revision 363070) +++ head/sys/kern/vfs_hash.c (revision 363071) @@ -1,236 +1,235 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2005 Poul-Henning Kamp * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include static MALLOC_DEFINE(M_VFS_HASH, "vfs_hash", "VFS hash table"); static LIST_HEAD(vfs_hash_head, vnode) *vfs_hash_tbl; static LIST_HEAD(,vnode) vfs_hash_side; static u_long vfs_hash_mask; static struct rwlock vfs_hash_lock; static void vfs_hashinit(void *dummy __unused) { vfs_hash_tbl = hashinit(desiredvnodes, M_VFS_HASH, &vfs_hash_mask); rw_init(&vfs_hash_lock, "vfs hash"); LIST_INIT(&vfs_hash_side); } /* Must be SI_ORDER_SECOND so desiredvnodes is available */ SYSINIT(vfs_hash, SI_SUB_VFS, SI_ORDER_SECOND, vfs_hashinit, NULL); u_int vfs_hash_index(struct vnode *vp) { return (vp->v_hash + vp->v_mount->mnt_hashseed); } static struct vfs_hash_head * vfs_hash_bucket(const struct mount *mp, u_int hash) { return (&vfs_hash_tbl[(hash + mp->mnt_hashseed) & vfs_hash_mask]); } int vfs_hash_get(const struct mount *mp, u_int hash, int flags, struct thread *td, struct vnode **vpp, vfs_hash_cmp_t *fn, void *arg) { struct vnode *vp; enum vgetstate vs; int error; while (1) { rw_rlock(&vfs_hash_lock); LIST_FOREACH(vp, vfs_hash_bucket(mp, hash), v_hashlist) { if (vp->v_hash != hash) continue; if (vp->v_mount != mp) continue; if (fn != NULL && fn(vp, arg)) continue; vs = vget_prep(vp); rw_runlock(&vfs_hash_lock); error = vget_finish(vp, flags, vs); if (error == ENOENT && (flags & LK_NOWAIT) == 0) break; if (error) return (error); *vpp = vp; return (0); } if (vp == NULL) { rw_runlock(&vfs_hash_lock); *vpp = NULL; return (0); } } } void vfs_hash_ref(const struct mount *mp, u_int hash, struct thread *td, struct vnode **vpp, vfs_hash_cmp_t *fn, void *arg) { struct vnode *vp; while (1) { rw_rlock(&vfs_hash_lock); LIST_FOREACH(vp, vfs_hash_bucket(mp, hash), v_hashlist) { if (vp->v_hash != hash) continue; if (vp->v_mount != mp) continue; if (fn != NULL && fn(vp, arg)) continue; vhold(vp); rw_runlock(&vfs_hash_lock); vref(vp); vdrop(vp); *vpp = vp; return; } if (vp == NULL) { rw_runlock(&vfs_hash_lock); *vpp = NULL; return; } } } void vfs_hash_remove(struct vnode *vp) { rw_wlock(&vfs_hash_lock); LIST_REMOVE(vp, v_hashlist); rw_wunlock(&vfs_hash_lock); } int vfs_hash_insert(struct vnode *vp, u_int hash, int flags, struct thread *td, struct vnode **vpp, vfs_hash_cmp_t *fn, void *arg) { struct vnode *vp2; enum vgetstate vs; int error; *vpp = NULL; while (1) { rw_wlock(&vfs_hash_lock); LIST_FOREACH(vp2, vfs_hash_bucket(vp->v_mount, hash), v_hashlist) { if (vp2->v_hash != hash) continue; if (vp2->v_mount != vp->v_mount) continue; if (fn != NULL && fn(vp2, arg)) continue; vs = vget_prep(vp2); rw_wunlock(&vfs_hash_lock); error = vget_finish(vp2, flags, vs); if (error == ENOENT && (flags & LK_NOWAIT) == 0) break; rw_wlock(&vfs_hash_lock); LIST_INSERT_HEAD(&vfs_hash_side, vp, v_hashlist); rw_wunlock(&vfs_hash_lock); vput(vp); if (!error) *vpp = vp2; return (error); } if (vp2 == NULL) break; - } vp->v_hash = hash; LIST_INSERT_HEAD(vfs_hash_bucket(vp->v_mount, hash), vp, v_hashlist); rw_wunlock(&vfs_hash_lock); return (0); } void vfs_hash_rehash(struct vnode *vp, u_int hash) { rw_wlock(&vfs_hash_lock); LIST_REMOVE(vp, v_hashlist); LIST_INSERT_HEAD(vfs_hash_bucket(vp->v_mount, hash), vp, v_hashlist); vp->v_hash = hash; rw_wunlock(&vfs_hash_lock); } void vfs_hash_changesize(u_long newmaxvnodes) { struct vfs_hash_head *vfs_hash_newtbl, *vfs_hash_oldtbl; u_long vfs_hash_newmask, vfs_hash_oldmask; struct vnode *vp; int i; vfs_hash_newtbl = hashinit(newmaxvnodes, M_VFS_HASH, &vfs_hash_newmask); /* If same hash table size, nothing to do */ if (vfs_hash_mask == vfs_hash_newmask) { free(vfs_hash_newtbl, M_VFS_HASH); return; } /* * Move everything from the old hash table to the new table. * None of the vnodes in the table can be recycled because to * do so, they have to be removed from the hash table. */ rw_wlock(&vfs_hash_lock); vfs_hash_oldtbl = vfs_hash_tbl; vfs_hash_oldmask = vfs_hash_mask; vfs_hash_tbl = vfs_hash_newtbl; vfs_hash_mask = vfs_hash_newmask; for (i = 0; i <= vfs_hash_oldmask; i++) { while ((vp = LIST_FIRST(&vfs_hash_oldtbl[i])) != NULL) { LIST_REMOVE(vp, v_hashlist); LIST_INSERT_HEAD( vfs_hash_bucket(vp->v_mount, vp->v_hash), vp, v_hashlist); } } rw_wunlock(&vfs_hash_lock); free(vfs_hash_oldtbl, M_VFS_HASH); } Index: head/sys/kern/vfs_init.c =================================================================== --- head/sys/kern/vfs_init.c (revision 363070) +++ head/sys/kern/vfs_init.c (revision 363071) @@ -1,593 +1,593 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed * to Berkeley by John Heidemann of the UCLA Ficus project. * * Source: * @(#)i405_init.c 2.10 92/04/27 UCLA Ficus project * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_init.c 8.3 (Berkeley) 1/4/94 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include static int vfs_register(struct vfsconf *); static int vfs_unregister(struct vfsconf *); MALLOC_DEFINE(M_VNODE, "vnodes", "Dynamically allocated vnodes"); /* * The highest defined VFS number. */ int maxvfsconf = VFS_GENERIC + 1; /* * Single-linked list of configured VFSes. * New entries are added/deleted by vfs_register()/vfs_unregister() */ struct vfsconfhead vfsconf = TAILQ_HEAD_INITIALIZER(vfsconf); struct sx vfsconf_sx; SX_SYSINIT(vfsconf, &vfsconf_sx, "vfsconf"); /* * Loader.conf variable vfs.typenumhash enables setting vfc_typenum using a hash * calculation on vfc_name, so that it doesn't change when file systems are * loaded in a different order. This will avoid the NFS server file handles from * changing for file systems that use vfc_typenum in their fsid. */ static int vfs_typenumhash = 1; SYSCTL_INT(_vfs, OID_AUTO, typenumhash, CTLFLAG_RDTUN, &vfs_typenumhash, 0, "Set vfc_typenum using a hash calculation on vfc_name, so that it does not" "change when file systems are loaded in a different order."); /* * A Zen vnode attribute structure. * * Initialized when the first filesystem registers by vfs_register(). */ struct vattr va_null; /* * vfs_init.c * * Allocate and fill in operations vectors. * * An undocumented feature of this approach to defining operations is that * there can be multiple entries in vfs_opv_descs for the same operations * vector. This allows third parties to extend the set of operations * supported by another layer in a binary compatibile way. For example, * assume that NFS needed to be modified to support Ficus. NFS has an entry * (probably nfs_vnopdeop_decls) declaring all the operations NFS supports by * default. Ficus could add another entry (ficus_nfs_vnodeop_decl_entensions) * listing those new operations Ficus adds to NFS, all without modifying the * NFS code. (Of couse, the OTW NFS protocol still needs to be munged, but * that is a(whole)nother story.) This is a feature. */ /* * Routines having to do with the management of the vnode table. */ static struct vfsconf * vfs_byname_locked(const char *name) { struct vfsconf *vfsp; sx_assert(&vfsconf_sx, SA_LOCKED); if (!strcmp(name, "ffs")) name = "ufs"; TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { if (!strcmp(name, vfsp->vfc_name)) return (vfsp); } return (NULL); } struct vfsconf * vfs_byname(const char *name) { struct vfsconf *vfsp; vfsconf_slock(); vfsp = vfs_byname_locked(name); vfsconf_sunlock(); return (vfsp); } struct vfsconf * vfs_byname_kld(const char *fstype, struct thread *td, int *error) { struct vfsconf *vfsp; int fileid, loaded; vfsp = vfs_byname(fstype); if (vfsp != NULL) return (vfsp); /* Try to load the respective module. */ *error = kern_kldload(td, fstype, &fileid); loaded = (*error == 0); if (*error == EEXIST) *error = 0; if (*error) return (NULL); /* Look up again to see if the VFS was loaded. */ vfsp = vfs_byname(fstype); if (vfsp == NULL) { if (loaded) (void)kern_kldunload(td, fileid, LINKER_UNLOAD_FORCE); *error = ENODEV; return (NULL); } return (vfsp); } static int vfs_mount_sigdefer(struct mount *mp) { int prev_stops, rc; TSRAW(curthread, TS_ENTER, "VFS_MOUNT", mp->mnt_vfc->vfc_name); prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_mount)(mp); sigallowstop(prev_stops); TSRAW(curthread, TS_EXIT, "VFS_MOUNT", mp->mnt_vfc->vfc_name); return (rc); } static int vfs_unmount_sigdefer(struct mount *mp, int mntflags) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_unmount)(mp, mntflags); sigallowstop(prev_stops); return (rc); } static int vfs_root_sigdefer(struct mount *mp, int flags, struct vnode **vpp) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_root)(mp, flags, vpp); sigallowstop(prev_stops); return (rc); } static int vfs_cachedroot_sigdefer(struct mount *mp, int flags, struct vnode **vpp) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_cachedroot)(mp, flags, vpp); sigallowstop(prev_stops); return (rc); } static int vfs_quotactl_sigdefer(struct mount *mp, int cmd, uid_t uid, void *arg) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_quotactl)(mp, cmd, uid, arg); sigallowstop(prev_stops); return (rc); } static int vfs_statfs_sigdefer(struct mount *mp, struct statfs *sbp) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_statfs)(mp, sbp); sigallowstop(prev_stops); return (rc); } static int vfs_sync_sigdefer(struct mount *mp, int waitfor) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_sync)(mp, waitfor); sigallowstop(prev_stops); return (rc); } static int vfs_vget_sigdefer(struct mount *mp, ino_t ino, int flags, struct vnode **vpp) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_vget)(mp, ino, flags, vpp); sigallowstop(prev_stops); return (rc); } static int vfs_fhtovp_sigdefer(struct mount *mp, struct fid *fidp, int flags, struct vnode **vpp) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_fhtovp)(mp, fidp, flags, vpp); sigallowstop(prev_stops); return (rc); } static int vfs_checkexp_sigdefer(struct mount *mp, struct sockaddr *nam, uint64_t *exflg, struct ucred **credp, int *numsecflavors, int *secflavors) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_checkexp)(mp, nam, exflg, credp, numsecflavors, secflavors); sigallowstop(prev_stops); return (rc); } static int vfs_extattrctl_sigdefer(struct mount *mp, int cmd, struct vnode *filename_vp, int attrnamespace, const char *attrname) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_extattrctl)(mp, cmd, filename_vp, attrnamespace, attrname); sigallowstop(prev_stops); return (rc); } static int vfs_sysctl_sigdefer(struct mount *mp, fsctlop_t op, struct sysctl_req *req) { int prev_stops, rc; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); rc = (*mp->mnt_vfc->vfc_vfsops_sd->vfs_sysctl)(mp, op, req); sigallowstop(prev_stops); return (rc); } static void vfs_susp_clean_sigdefer(struct mount *mp) { int prev_stops; if (*mp->mnt_vfc->vfc_vfsops_sd->vfs_susp_clean == NULL) return; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); (*mp->mnt_vfc->vfc_vfsops_sd->vfs_susp_clean)(mp); sigallowstop(prev_stops); } static void vfs_reclaim_lowervp_sigdefer(struct mount *mp, struct vnode *vp) { int prev_stops; if (*mp->mnt_vfc->vfc_vfsops_sd->vfs_reclaim_lowervp == NULL) return; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); (*mp->mnt_vfc->vfc_vfsops_sd->vfs_reclaim_lowervp)(mp, vp); sigallowstop(prev_stops); } static void vfs_unlink_lowervp_sigdefer(struct mount *mp, struct vnode *vp) { int prev_stops; if (*mp->mnt_vfc->vfc_vfsops_sd->vfs_unlink_lowervp == NULL) return; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); (*(mp)->mnt_vfc->vfc_vfsops_sd->vfs_unlink_lowervp)(mp, vp); sigallowstop(prev_stops); } static void vfs_purge_sigdefer(struct mount *mp) { int prev_stops; prev_stops = sigdeferstop(SIGDEFERSTOP_SILENT); (*mp->mnt_vfc->vfc_vfsops_sd->vfs_purge)(mp); sigallowstop(prev_stops); } static struct vfsops vfsops_sigdefer = { .vfs_mount = vfs_mount_sigdefer, .vfs_unmount = vfs_unmount_sigdefer, .vfs_root = vfs_root_sigdefer, .vfs_cachedroot = vfs_cachedroot_sigdefer, .vfs_quotactl = vfs_quotactl_sigdefer, .vfs_statfs = vfs_statfs_sigdefer, .vfs_sync = vfs_sync_sigdefer, .vfs_vget = vfs_vget_sigdefer, .vfs_fhtovp = vfs_fhtovp_sigdefer, .vfs_checkexp = vfs_checkexp_sigdefer, .vfs_extattrctl = vfs_extattrctl_sigdefer, .vfs_sysctl = vfs_sysctl_sigdefer, .vfs_susp_clean = vfs_susp_clean_sigdefer, .vfs_reclaim_lowervp = vfs_reclaim_lowervp_sigdefer, .vfs_unlink_lowervp = vfs_unlink_lowervp_sigdefer, .vfs_purge = vfs_purge_sigdefer, }; /* Register a new filesystem type in the global table */ static int vfs_register(struct vfsconf *vfc) { struct sysctl_oid *oidp; struct vfsops *vfsops; static int once; struct vfsconf *tvfc; uint32_t hashval; int secondpass; if (!once) { vattr_null(&va_null); once = 1; } - + if (vfc->vfc_version != VFS_VERSION) { printf("ERROR: filesystem %s, unsupported ABI version %x\n", vfc->vfc_name, vfc->vfc_version); return (EINVAL); } vfsconf_lock(); if (vfs_byname_locked(vfc->vfc_name) != NULL) { vfsconf_unlock(); return (EEXIST); } if (vfs_typenumhash != 0) { /* * Calculate a hash on vfc_name to use for vfc_typenum. Unless * all of 1<->255 are assigned, it is limited to 8bits since * that is what ZFS uses from vfc_typenum and is also the * preferred range for vfs_getnewfsid(). */ hashval = fnv_32_str(vfc->vfc_name, FNV1_32_INIT); hashval &= 0xff; secondpass = 0; do { /* Look for and fix any collision. */ TAILQ_FOREACH(tvfc, &vfsconf, vfc_list) { if (hashval == tvfc->vfc_typenum) { if (hashval == 255 && secondpass == 0) { hashval = 1; secondpass = 1; } else hashval++; break; } } } while (tvfc != NULL); vfc->vfc_typenum = hashval; if (vfc->vfc_typenum >= maxvfsconf) maxvfsconf = vfc->vfc_typenum + 1; } else vfc->vfc_typenum = maxvfsconf++; TAILQ_INSERT_TAIL(&vfsconf, vfc, vfc_list); /* * Initialise unused ``struct vfsops'' fields, to use * the vfs_std*() functions. Note, we need the mount * and unmount operations, at the least. The check * for vfsops available is just a debugging aid. */ KASSERT(vfc->vfc_vfsops != NULL, ("Filesystem %s has no vfsops", vfc->vfc_name)); /* * Check the mount and unmount operations. */ vfsops = vfc->vfc_vfsops; KASSERT(vfsops->vfs_mount != NULL, ("Filesystem %s has no mount op", vfc->vfc_name)); KASSERT(vfsops->vfs_unmount != NULL, ("Filesystem %s has no unmount op", vfc->vfc_name)); if (vfsops->vfs_root == NULL) /* return file system's root vnode */ vfsops->vfs_root = vfs_stdroot; if (vfsops->vfs_quotactl == NULL) /* quota control */ vfsops->vfs_quotactl = vfs_stdquotactl; if (vfsops->vfs_statfs == NULL) /* return file system's status */ vfsops->vfs_statfs = vfs_stdstatfs; if (vfsops->vfs_sync == NULL) /* * flush unwritten data (nosync) * file systems can use vfs_stdsync * explicitly by setting it in the * vfsop vector. */ vfsops->vfs_sync = vfs_stdnosync; if (vfsops->vfs_vget == NULL) /* convert an inode number to a vnode */ vfsops->vfs_vget = vfs_stdvget; if (vfsops->vfs_fhtovp == NULL) /* turn an NFS file handle into a vnode */ vfsops->vfs_fhtovp = vfs_stdfhtovp; if (vfsops->vfs_checkexp == NULL) /* check if file system is exported */ vfsops->vfs_checkexp = vfs_stdcheckexp; if (vfsops->vfs_init == NULL) /* file system specific initialisation */ vfsops->vfs_init = vfs_stdinit; if (vfsops->vfs_uninit == NULL) /* file system specific uninitialisation */ vfsops->vfs_uninit = vfs_stduninit; if (vfsops->vfs_extattrctl == NULL) /* extended attribute control */ vfsops->vfs_extattrctl = vfs_stdextattrctl; if (vfsops->vfs_sysctl == NULL) vfsops->vfs_sysctl = vfs_stdsysctl; if ((vfc->vfc_flags & VFCF_SBDRY) != 0) { vfc->vfc_vfsops_sd = vfc->vfc_vfsops; vfc->vfc_vfsops = &vfsops_sigdefer; } if (vfc->vfc_flags & VFCF_JAIL) prison_add_vfs(vfc); /* * Call init function for this VFS... */ if ((vfc->vfc_flags & VFCF_SBDRY) != 0) vfc->vfc_vfsops_sd->vfs_init(vfc); else vfc->vfc_vfsops->vfs_init(vfc); vfsconf_unlock(); /* * If this filesystem has a sysctl node under vfs * (i.e. vfs.xxfs), then change the oid number of that node to * match the filesystem's type number. This allows user code * which uses the type number to read sysctl variables defined * by the filesystem to continue working. Since the oids are * in a sorted list, we need to make sure the order is * preserved by re-registering the oid after modifying its * number. */ sysctl_wlock(); SLIST_FOREACH(oidp, SYSCTL_CHILDREN(&sysctl___vfs), oid_link) { if (strcmp(oidp->oid_name, vfc->vfc_name) == 0) { sysctl_unregister_oid(oidp); oidp->oid_number = vfc->vfc_typenum; sysctl_register_oid(oidp); break; } } sysctl_wunlock(); return (0); } /* Remove registration of a filesystem type */ static int vfs_unregister(struct vfsconf *vfc) { struct vfsconf *vfsp; int error, maxtypenum; vfsconf_lock(); vfsp = vfs_byname_locked(vfc->vfc_name); if (vfsp == NULL) { vfsconf_unlock(); return (EINVAL); } if (vfsp->vfc_refcount != 0) { vfsconf_unlock(); return (EBUSY); } error = 0; if ((vfc->vfc_flags & VFCF_SBDRY) != 0) { if (vfc->vfc_vfsops_sd->vfs_uninit != NULL) error = vfc->vfc_vfsops_sd->vfs_uninit(vfsp); } else { if (vfc->vfc_vfsops->vfs_uninit != NULL) { error = vfc->vfc_vfsops->vfs_uninit(vfsp); } if (error != 0) { vfsconf_unlock(); return (error); } } TAILQ_REMOVE(&vfsconf, vfsp, vfc_list); maxtypenum = VFS_GENERIC; TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) if (maxtypenum < vfsp->vfc_typenum) maxtypenum = vfsp->vfc_typenum; maxvfsconf = maxtypenum + 1; vfsconf_unlock(); return (0); } /* * Standard kernel module handling code for filesystem modules. * Referenced from VFS_SET(). */ int vfs_modevent(module_t mod, int type, void *data) { struct vfsconf *vfc; int error = 0; vfc = (struct vfsconf *)data; switch (type) { case MOD_LOAD: if (vfc) error = vfs_register(vfc); break; case MOD_UNLOAD: if (vfc) error = vfs_unregister(vfc); break; default: error = EOPNOTSUPP; break; } return (error); } Index: head/sys/kern/vfs_lookup.c =================================================================== --- head/sys/kern/vfs_lookup.c (revision 363070) +++ head/sys/kern/vfs_lookup.c (revision 363071) @@ -1,1501 +1,1501 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_lookup.c 8.4 (Berkeley) 2/16/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #define NAMEI_DIAGNOSTIC 1 #undef NAMEI_DIAGNOSTIC SDT_PROVIDER_DECLARE(vfs); SDT_PROBE_DEFINE3(vfs, namei, lookup, entry, "struct vnode *", "char *", "unsigned long"); SDT_PROBE_DEFINE2(vfs, namei, lookup, return, "int", "struct vnode *"); /* Allocation zone for namei. */ uma_zone_t namei_zone; /* Placeholder vnode for mp traversal. */ static struct vnode *vp_crossmp; static int crossmp_vop_islocked(struct vop_islocked_args *ap) { return (LK_SHARED); } static int crossmp_vop_lock1(struct vop_lock1_args *ap) { struct vnode *vp; struct lock *lk __unused; const char *file __unused; int flags, line __unused; vp = ap->a_vp; lk = vp->v_vnlock; flags = ap->a_flags; file = ap->a_file; line = ap->a_line; if ((flags & LK_SHARED) == 0) panic("invalid lock request for crossmp"); WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line, flags & LK_INTERLOCK ? &VI_MTX(vp)->lock_object : NULL); WITNESS_LOCK(&lk->lock_object, 0, file, line); if ((flags & LK_INTERLOCK) != 0) VI_UNLOCK(vp); LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, ap->a_file, line); return (0); } static int crossmp_vop_unlock(struct vop_unlock_args *ap) { struct vnode *vp; struct lock *lk __unused; vp = ap->a_vp; lk = vp->v_vnlock; WITNESS_UNLOCK(&lk->lock_object, 0, LOCK_FILE, LOCK_LINE); LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, LOCK_FILE, LOCK_LINE); return (0); } static struct vop_vector crossmp_vnodeops = { .vop_default = &default_vnodeops, .vop_islocked = crossmp_vop_islocked, .vop_lock1 = crossmp_vop_lock1, .vop_unlock = crossmp_vop_unlock, }; /* * VFS_VOP_VECTOR_REGISTER(crossmp_vnodeops) is not used here since the vnode * gets allocated early. See nameiinit for the direct call below. */ struct nameicap_tracker { struct vnode *dp; TAILQ_ENTRY(nameicap_tracker) nm_link; }; /* Zone for cap mode tracker elements used for dotdot capability checks. */ static uma_zone_t nt_zone; static void nameiinit(void *dummy __unused) { namei_zone = uma_zcreate("NAMEI", MAXPATHLEN, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); nt_zone = uma_zcreate("rentr", sizeof(struct nameicap_tracker), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); vfs_vector_op_register(&crossmp_vnodeops); getnewvnode("crossmp", NULL, &crossmp_vnodeops, &vp_crossmp); } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nameiinit, NULL); static int lookup_cap_dotdot = 1; SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot, CTLFLAG_RWTUN, &lookup_cap_dotdot, 0, "enables \"..\" components in path lookup in capability mode"); static int lookup_cap_dotdot_nonlocal = 1; SYSCTL_INT(_vfs, OID_AUTO, lookup_cap_dotdot_nonlocal, CTLFLAG_RWTUN, &lookup_cap_dotdot_nonlocal, 0, "enables \"..\" components in path lookup in capability mode " "on non-local mount"); static void nameicap_tracker_add(struct nameidata *ndp, struct vnode *dp) { struct nameicap_tracker *nt; if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp->v_type != VDIR) return; if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) { MPASS((ndp->ni_lcf & NI_LCF_LATCH) != 0); if (dp != ndp->ni_beneath_latch) return; ndp->ni_lcf |= NI_LCF_BENEATH_LATCHED; } nt = uma_zalloc(nt_zone, M_WAITOK); vhold(dp); nt->dp = dp; TAILQ_INSERT_TAIL(&ndp->ni_cap_tracker, nt, nm_link); } static void nameicap_cleanup(struct nameidata *ndp, bool clean_latch) { struct nameicap_tracker *nt, *nt1; KASSERT(TAILQ_EMPTY(&ndp->ni_cap_tracker) || (ndp->ni_lcf & NI_LCF_CAP_DOTDOT) != 0, ("not strictrelative")); TAILQ_FOREACH_SAFE(nt, &ndp->ni_cap_tracker, nm_link, nt1) { TAILQ_REMOVE(&ndp->ni_cap_tracker, nt, nm_link); vdrop(nt->dp); uma_zfree(nt_zone, nt); } if (clean_latch && (ndp->ni_lcf & NI_LCF_LATCH) != 0) { ndp->ni_lcf &= ~NI_LCF_LATCH; vrele(ndp->ni_beneath_latch); } } /* * For dotdot lookups in capability mode, only allow the component * lookup to succeed if the resulting directory was already traversed * during the operation. Also fail dotdot lookups for non-local * filesystems, where external agents might assist local lookups to * escape the compartment. */ static int nameicap_check_dotdot(struct nameidata *ndp, struct vnode *dp) { struct nameicap_tracker *nt; struct mount *mp; if ((ndp->ni_lcf & NI_LCF_CAP_DOTDOT) == 0 || dp == NULL || dp->v_type != VDIR) return (0); mp = dp->v_mount; if (lookup_cap_dotdot_nonlocal == 0 && mp != NULL && (mp->mnt_flag & MNT_LOCAL) == 0) return (ENOTCAPABLE); TAILQ_FOREACH_REVERSE(nt, &ndp->ni_cap_tracker, nameicap_tracker_head, nm_link) { if (dp == nt->dp) return (0); } if ((ndp->ni_lcf & NI_LCF_BENEATH_ABS) != 0) { ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED; nameicap_cleanup(ndp, false); return (0); } return (ENOTCAPABLE); } static void namei_cleanup_cnp(struct componentname *cnp) { uma_zfree(namei_zone, cnp->cn_pnbuf); #ifdef DIAGNOSTIC cnp->cn_pnbuf = NULL; cnp->cn_nameptr = NULL; #endif } static int namei_handle_root(struct nameidata *ndp, struct vnode **dpp) { struct componentname *cnp; cnp = &ndp->ni_cnd; if ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0) { #ifdef KTRACE if (KTRPOINT(curthread, KTR_CAPFAIL)) ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL); #endif return (ENOTCAPABLE); } if ((cnp->cn_flags & BENEATH) != 0) { ndp->ni_lcf |= NI_LCF_BENEATH_ABS; ndp->ni_lcf &= ~NI_LCF_BENEATH_LATCHED; nameicap_cleanup(ndp, false); } while (*(cnp->cn_nameptr) == '/') { cnp->cn_nameptr++; ndp->ni_pathlen--; } *dpp = ndp->ni_rootdir; vrefact(*dpp); return (0); } /* * Convert a pathname into a pointer to a locked vnode. * * The FOLLOW flag is set when symbolic links are to be followed * when they occur at the end of the name translation process. * Symbolic links are always followed for all other pathname * components other than the last. * * The segflg defines whether the name is to be copied from user * space or kernel space. * * Overall outline of namei: * * copy in name * get starting directory * while (!done && !error) { * call lookup to search path. * if symbolic link, massage name in buffer and continue * } */ int namei(struct nameidata *ndp) { char *cp; /* pointer into pathname argument */ struct vnode *dp; /* the directory we are searching */ struct iovec aiov; /* uio for reading symbolic links */ struct componentname *cnp; struct file *dfp; struct thread *td; struct proc *p; struct pwd *pwd; cap_rights_t rights; struct filecaps dirfd_caps; struct uio auio; int error, linklen, startdir_used; cnp = &ndp->ni_cnd; td = cnp->cn_thread; p = td->td_proc; ndp->ni_cnd.cn_cred = ndp->ni_cnd.cn_thread->td_ucred; KASSERT(cnp->cn_cred && p, ("namei: bad cred/proc")); KASSERT((cnp->cn_nameiop & (~OPMASK)) == 0, ("namei: nameiop contaminated with flags")); KASSERT((cnp->cn_flags & OPMASK) == 0, ("namei: flags contaminated with nameiops")); MPASS(ndp->ni_startdir == NULL || ndp->ni_startdir->v_type == VDIR || ndp->ni_startdir->v_type == VBAD); TAILQ_INIT(&ndp->ni_cap_tracker); ndp->ni_lcf = 0; /* We will set this ourselves if we need it. */ cnp->cn_flags &= ~TRAILINGSLASH; /* * Get a buffer for the name to be translated, and copy the * name into the buffer. */ if ((cnp->cn_flags & HASBUF) == 0) cnp->cn_pnbuf = uma_zalloc(namei_zone, M_WAITOK); if (ndp->ni_segflg == UIO_SYSSPACE) error = copystr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN, &ndp->ni_pathlen); else error = copyinstr(ndp->ni_dirp, cnp->cn_pnbuf, MAXPATHLEN, &ndp->ni_pathlen); /* * Don't allow empty pathnames. */ if (error == 0 && *cnp->cn_pnbuf == '\0') error = ENOENT; #ifdef CAPABILITY_MODE /* * In capability mode, lookups must be restricted to happen in * the subtree with the root specified by the file descriptor: * - The root must be real file descriptor, not the pseudo-descriptor * AT_FDCWD. * - The passed path must be relative and not absolute. * - If lookup_cap_dotdot is disabled, path must not contain the * '..' components. * - If lookup_cap_dotdot is enabled, we verify that all '..' * components lookups result in the directories which were * previously walked by us, which prevents an escape from * the relative root. */ if (error == 0 && IN_CAPABILITY_MODE(td) && (cnp->cn_flags & NOCAPCHECK) == 0) { ndp->ni_lcf |= NI_LCF_STRICTRELATIVE; if (ndp->ni_dirfd == AT_FDCWD) { #ifdef KTRACE if (KTRPOINT(td, KTR_CAPFAIL)) ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL); #endif error = ECAPMODE; } } #endif if (error != 0) { namei_cleanup_cnp(cnp); ndp->ni_vp = NULL; return (error); } ndp->ni_loopcnt = 0; #ifdef KTRACE if (KTRPOINT(td, KTR_NAMEI)) { KASSERT(cnp->cn_thread == curthread, ("namei not using curthread")); ktrnamei(cnp->cn_pnbuf); } #endif /* * Get starting point for the translation. */ pwd = pwd_hold(td); /* * The reference on ni_rootdir is acquired in the block below to avoid * back-to-back atomics for absolute lookups. */ ndp->ni_rootdir = pwd->pwd_rdir; ndp->ni_topdir = pwd->pwd_jdir; startdir_used = 0; dp = NULL; cnp->cn_nameptr = cnp->cn_pnbuf; if (cnp->cn_pnbuf[0] == '/') { ndp->ni_resflags |= NIRES_ABS; error = namei_handle_root(ndp, &dp); } else { if (ndp->ni_startdir != NULL) { dp = ndp->ni_startdir; startdir_used = 1; } else if (ndp->ni_dirfd == AT_FDCWD) { dp = pwd->pwd_cdir; vrefact(dp); } else { rights = ndp->ni_rightsneeded; cap_rights_set_one(&rights, CAP_LOOKUP); if (cnp->cn_flags & AUDITVNODE1) AUDIT_ARG_ATFD1(ndp->ni_dirfd); if (cnp->cn_flags & AUDITVNODE2) AUDIT_ARG_ATFD2(ndp->ni_dirfd); /* * Effectively inlined fgetvp_rights, because we need to * inspect the file as well as grabbing the vnode. */ error = fget_cap(td, ndp->ni_dirfd, &rights, &dfp, &ndp->ni_filecaps); if (error != 0) { /* * Preserve the error; it should either be EBADF * or capability-related, both of which can be * safely returned to the caller. */ } else { if (dfp->f_ops == &badfileops) { error = EBADF; } else if (dfp->f_vnode == NULL) { error = ENOTDIR; } else { dp = dfp->f_vnode; vrefact(dp); if ((dfp->f_flag & FSEARCH) != 0) cnp->cn_flags |= NOEXECCHECK; } fdrop(dfp, td); } #ifdef CAPABILITIES /* * If file descriptor doesn't have all rights, * all lookups relative to it must also be * strictly relative. */ CAP_ALL(&rights); if (!cap_rights_contains(&ndp->ni_filecaps.fc_rights, &rights) || ndp->ni_filecaps.fc_fcntls != CAP_FCNTL_ALL || ndp->ni_filecaps.fc_nioctls != -1) { ndp->ni_lcf |= NI_LCF_STRICTRELATIVE; } #endif } if (error == 0 && dp->v_type != VDIR) error = ENOTDIR; } if (error == 0 && (cnp->cn_flags & BENEATH) != 0) { if (ndp->ni_dirfd == AT_FDCWD) { ndp->ni_beneath_latch = pwd->pwd_cdir; vrefact(ndp->ni_beneath_latch); } else { rights = ndp->ni_rightsneeded; cap_rights_set_one(&rights, CAP_LOOKUP); error = fgetvp_rights(td, ndp->ni_dirfd, &rights, &dirfd_caps, &ndp->ni_beneath_latch); if (error == 0 && dp->v_type != VDIR) { vrele(ndp->ni_beneath_latch); error = ENOTDIR; } } if (error == 0) ndp->ni_lcf |= NI_LCF_LATCH; } /* * If we are auditing the kernel pathname, save the user pathname. */ if (cnp->cn_flags & AUDITVNODE1) AUDIT_ARG_UPATH1_VP(td, ndp->ni_rootdir, dp, cnp->cn_pnbuf); if (cnp->cn_flags & AUDITVNODE2) AUDIT_ARG_UPATH2_VP(td, ndp->ni_rootdir, dp, cnp->cn_pnbuf); if (ndp->ni_startdir != NULL && !startdir_used) vrele(ndp->ni_startdir); if (error != 0) { if (dp != NULL) vrele(dp); goto out; } MPASS((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_LATCH)) != NI_LCF_BENEATH_ABS); if (((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) != 0 && lookup_cap_dotdot != 0) || ((ndp->ni_lcf & NI_LCF_STRICTRELATIVE) == 0 && (cnp->cn_flags & BENEATH) != 0)) ndp->ni_lcf |= NI_LCF_CAP_DOTDOT; SDT_PROBE3(vfs, namei, lookup, entry, dp, cnp->cn_pnbuf, cnp->cn_flags); for (;;) { ndp->ni_startdir = dp; error = lookup(ndp); if (error != 0) goto out; /* * If not a symbolic link, we're done. */ if ((cnp->cn_flags & ISSYMLINK) == 0) { if ((cnp->cn_flags & (SAVENAME | SAVESTART)) == 0) { namei_cleanup_cnp(cnp); } else cnp->cn_flags |= HASBUF; if ((ndp->ni_lcf & (NI_LCF_BENEATH_ABS | NI_LCF_BENEATH_LATCHED)) == NI_LCF_BENEATH_ABS) { NDFREE(ndp, 0); error = ENOTCAPABLE; } nameicap_cleanup(ndp, true); SDT_PROBE2(vfs, namei, lookup, return, error, (error == 0 ? ndp->ni_vp : NULL)); pwd_drop(pwd); return (error); } if (ndp->ni_loopcnt++ >= MAXSYMLINKS) { error = ELOOP; break; } #ifdef MAC if ((cnp->cn_flags & NOMACCHECK) == 0) { error = mac_vnode_check_readlink(td->td_ucred, ndp->ni_vp); if (error != 0) break; } #endif if (ndp->ni_pathlen > 1) cp = uma_zalloc(namei_zone, M_WAITOK); else cp = cnp->cn_pnbuf; aiov.iov_base = cp; aiov.iov_len = MAXPATHLEN; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; auio.uio_offset = 0; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_resid = MAXPATHLEN; error = VOP_READLINK(ndp->ni_vp, &auio, cnp->cn_cred); if (error != 0) { if (ndp->ni_pathlen > 1) uma_zfree(namei_zone, cp); break; } linklen = MAXPATHLEN - auio.uio_resid; if (linklen == 0) { if (ndp->ni_pathlen > 1) uma_zfree(namei_zone, cp); error = ENOENT; break; } if (linklen + ndp->ni_pathlen > MAXPATHLEN) { if (ndp->ni_pathlen > 1) uma_zfree(namei_zone, cp); error = ENAMETOOLONG; break; } if (ndp->ni_pathlen > 1) { bcopy(ndp->ni_next, cp + linklen, ndp->ni_pathlen); uma_zfree(namei_zone, cnp->cn_pnbuf); cnp->cn_pnbuf = cp; } else cnp->cn_pnbuf[linklen] = '\0'; ndp->ni_pathlen += linklen; vput(ndp->ni_vp); dp = ndp->ni_dvp; /* * Check if root directory should replace current directory. */ cnp->cn_nameptr = cnp->cn_pnbuf; if (*(cnp->cn_nameptr) == '/') { vrele(dp); error = namei_handle_root(ndp, &dp); if (error != 0) goto out; } } vput(ndp->ni_vp); ndp->ni_vp = NULL; vrele(ndp->ni_dvp); out: MPASS(error != 0); namei_cleanup_cnp(cnp); nameicap_cleanup(ndp, true); SDT_PROBE2(vfs, namei, lookup, return, error, NULL); pwd_drop(pwd); return (error); } static int compute_cn_lkflags(struct mount *mp, int lkflags, int cnflags) { if (mp == NULL || ((lkflags & LK_SHARED) && (!(mp->mnt_kern_flag & MNTK_LOOKUP_SHARED) || ((cnflags & ISDOTDOT) && (mp->mnt_kern_flag & MNTK_LOOKUP_EXCL_DOTDOT))))) { lkflags &= ~LK_SHARED; lkflags |= LK_EXCLUSIVE; } lkflags |= LK_NODDLKTREAT; return (lkflags); } static __inline int needs_exclusive_leaf(struct mount *mp, int flags) { /* * Intermediate nodes can use shared locks, we only need to * force an exclusive lock for leaf nodes. */ if ((flags & (ISLASTCN | LOCKLEAF)) != (ISLASTCN | LOCKLEAF)) return (0); /* Always use exclusive locks if LOCKSHARED isn't set. */ if (!(flags & LOCKSHARED)) return (1); /* * For lookups during open(), if the mount point supports * extended shared operations, then use a shared lock for the * leaf node, otherwise use an exclusive lock. */ if ((flags & ISOPEN) != 0) return (!MNT_EXTENDED_SHARED(mp)); /* * Lookup requests outside of open() that specify LOCKSHARED * only need a shared lock on the leaf vnode. */ return (0); } /* * Search a pathname. * This is a very central and rather complicated routine. * * The pathname is pointed to by ni_ptr and is of length ni_pathlen. * The starting directory is taken from ni_startdir. The pathname is * descended until done, or a symbolic link is encountered. The variable * ni_more is clear if the path is completed; it is set to one if a * symbolic link needing interpretation is encountered. * * The flag argument is LOOKUP, CREATE, RENAME, or DELETE depending on * whether the name is to be looked up, created, renamed, or deleted. * When CREATE, RENAME, or DELETE is specified, information usable in * creating, renaming, or deleting a directory entry may be calculated. * If flag has LOCKPARENT or'ed into it, the parent directory is returned * locked. If flag has WANTPARENT or'ed into it, the parent directory is * returned unlocked. Otherwise the parent directory is not returned. If * the target of the pathname exists and LOCKLEAF is or'ed into the flag * the target is returned locked, otherwise it is returned unlocked. * When creating or renaming and LOCKPARENT is specified, the target may not * be ".". When deleting and LOCKPARENT is specified, the target may be ".". * * Overall outline of lookup: * * dirloop: * identify next component of name at ndp->ni_ptr * handle degenerate case where name is null string * if .. and crossing mount points and on mounted filesys, find parent * call VOP_LOOKUP routine for next component name * directory vnode returned in ni_dvp, unlocked unless LOCKPARENT set * component vnode returned in ni_vp (if it exists), locked. * if result vnode is mounted on and crossing mount points, * find mounted on vnode * if more components of name, do next level at dirloop * return the answer in ni_vp, locked if LOCKLEAF set * if LOCKPARENT set, return locked parent in ni_dvp * if WANTPARENT set, return unlocked parent in ni_dvp */ int lookup(struct nameidata *ndp) { char *cp; /* pointer into pathname argument */ char *prev_ni_next; /* saved ndp->ni_next */ struct vnode *dp = NULL; /* the directory we are searching */ struct vnode *tdp; /* saved dp */ struct mount *mp; /* mount table entry */ struct prison *pr; size_t prev_ni_pathlen; /* saved ndp->ni_pathlen */ int docache; /* == 0 do not cache last component */ int wantparent; /* 1 => wantparent or lockparent flag */ int rdonly; /* lookup read-only flag bit */ int error = 0; int dpunlocked = 0; /* dp has already been unlocked */ int relookup = 0; /* do not consume the path component */ struct componentname *cnp = &ndp->ni_cnd; int lkflags_save; int ni_dvp_unlocked; - + /* * Setup: break out flag bits into variables. */ ni_dvp_unlocked = 0; wantparent = cnp->cn_flags & (LOCKPARENT | WANTPARENT); KASSERT(cnp->cn_nameiop == LOOKUP || wantparent, ("CREATE, DELETE, RENAME require LOCKPARENT or WANTPARENT.")); docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE; if (cnp->cn_nameiop == DELETE || (wantparent && cnp->cn_nameiop != CREATE && cnp->cn_nameiop != LOOKUP)) docache = 0; rdonly = cnp->cn_flags & RDONLY; cnp->cn_flags &= ~ISSYMLINK; ndp->ni_dvp = NULL; /* * We use shared locks until we hit the parent of the last cn then * we adjust based on the requesting flags. */ cnp->cn_lkflags = LK_SHARED; dp = ndp->ni_startdir; ndp->ni_startdir = NULLVP; vn_lock(dp, compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY, cnp->cn_flags)); dirloop: /* * Search a new directory. * * The last component of the filename is left accessible via * cnp->cn_nameptr for callers that need the name. Callers needing * the name set the SAVENAME flag. When done, they assume * responsibility for freeing the pathname buffer. */ for (cp = cnp->cn_nameptr; *cp != 0 && *cp != '/'; cp++) continue; cnp->cn_namelen = cp - cnp->cn_nameptr; if (cnp->cn_namelen > NAME_MAX) { error = ENAMETOOLONG; goto bad; } #ifdef NAMEI_DIAGNOSTIC { char c = *cp; *cp = '\0'; printf("{%s}: ", cnp->cn_nameptr); *cp = c; } #endif prev_ni_pathlen = ndp->ni_pathlen; ndp->ni_pathlen -= cnp->cn_namelen; KASSERT(ndp->ni_pathlen <= PATH_MAX, ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen)); prev_ni_next = ndp->ni_next; ndp->ni_next = cp; /* * Replace multiple slashes by a single slash and trailing slashes * by a null. This must be done before VOP_LOOKUP() because some * fs's don't know about trailing slashes. Remember if there were * trailing slashes to handle symlinks, existing non-directories * and non-existing files that won't be directories specially later. */ while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) { cp++; ndp->ni_pathlen--; if (*cp == '\0') { *ndp->ni_next = '\0'; cnp->cn_flags |= TRAILINGSLASH; } } ndp->ni_next = cp; cnp->cn_flags |= MAKEENTRY; if (*cp == '\0' && docache == 0) cnp->cn_flags &= ~MAKEENTRY; if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.') cnp->cn_flags |= ISDOTDOT; else cnp->cn_flags &= ~ISDOTDOT; if (*ndp->ni_next == 0) cnp->cn_flags |= ISLASTCN; else cnp->cn_flags &= ~ISLASTCN; if ((cnp->cn_flags & ISLASTCN) != 0 && cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.' && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) { error = EINVAL; goto bad; } nameicap_tracker_add(ndp, dp); /* * Check for degenerate name (e.g. / or "") * which is a way of talking about a directory, * e.g. like "/." or ".". */ if (cnp->cn_nameptr[0] == '\0') { if (dp->v_type != VDIR) { error = ENOTDIR; goto bad; } if (cnp->cn_nameiop != LOOKUP) { error = EISDIR; goto bad; } if (wantparent) { ndp->ni_dvp = dp; VREF(dp); } ndp->ni_vp = dp; if (cnp->cn_flags & AUDITVNODE1) AUDIT_ARG_VNODE1(dp); else if (cnp->cn_flags & AUDITVNODE2) AUDIT_ARG_VNODE2(dp); if (!(cnp->cn_flags & (LOCKPARENT | LOCKLEAF))) VOP_UNLOCK(dp); /* XXX This should probably move to the top of function. */ if (cnp->cn_flags & SAVESTART) panic("lookup: SAVESTART"); goto success; } /* * Handle "..": five special cases. * 0. If doing a capability lookup and lookup_cap_dotdot is * disabled, return ENOTCAPABLE. * 1. Return an error if this is the last component of * the name and the operation is DELETE or RENAME. * 2. If at root directory (e.g. after chroot) * or at absolute root directory * then ignore it so can't get out. * 3. If this vnode is the root of a mounted * filesystem, then replace it with the * vnode which was mounted on so we take the * .. in the other filesystem. * 4. If the vnode is the top directory of * the jail or chroot, don't let them out. * 5. If doing a capability lookup and lookup_cap_dotdot is * enabled, return ENOTCAPABLE if the lookup would escape * from the initial file descriptor directory. Checks are * done by ensuring that namei() already traversed the * result of dotdot lookup. */ if (cnp->cn_flags & ISDOTDOT) { if ((ndp->ni_lcf & (NI_LCF_STRICTRELATIVE | NI_LCF_CAP_DOTDOT)) == NI_LCF_STRICTRELATIVE) { #ifdef KTRACE if (KTRPOINT(curthread, KTR_CAPFAIL)) ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL); #endif error = ENOTCAPABLE; goto bad; } if ((cnp->cn_flags & ISLASTCN) != 0 && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) { error = EINVAL; goto bad; } for (;;) { for (pr = cnp->cn_cred->cr_prison; pr != NULL; pr = pr->pr_parent) if (dp == pr->pr_root) break; if (dp == ndp->ni_rootdir || dp == ndp->ni_topdir || dp == rootvnode || pr != NULL || ((dp->v_vflag & VV_ROOT) != 0 && (cnp->cn_flags & NOCROSSMOUNT) != 0)) { ndp->ni_dvp = dp; ndp->ni_vp = dp; VREF(dp); goto nextname; } if ((dp->v_vflag & VV_ROOT) == 0) break; if (VN_IS_DOOMED(dp)) { /* forced unmount */ error = ENOENT; goto bad; } tdp = dp; dp = dp->v_mount->mnt_vnodecovered; VREF(dp); vput(tdp); vn_lock(dp, compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY, ISDOTDOT)); error = nameicap_check_dotdot(ndp, dp); if (error != 0) { #ifdef KTRACE if (KTRPOINT(curthread, KTR_CAPFAIL)) ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL); #endif goto bad; } } } /* * We now have a segment name to search for, and a directory to search. */ unionlookup: #ifdef MAC error = mac_vnode_check_lookup(cnp->cn_thread->td_ucred, dp, cnp); if (error) goto bad; #endif ndp->ni_dvp = dp; ndp->ni_vp = NULL; ASSERT_VOP_LOCKED(dp, "lookup"); /* * If we have a shared lock we may need to upgrade the lock for the * last operation. */ if ((cnp->cn_flags & LOCKPARENT) && (cnp->cn_flags & ISLASTCN) && dp != vp_crossmp && VOP_ISLOCKED(dp) == LK_SHARED) vn_lock(dp, LK_UPGRADE|LK_RETRY); if (VN_IS_DOOMED(dp)) { error = ENOENT; goto bad; } /* * If we're looking up the last component and we need an exclusive * lock, adjust our lkflags. */ if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags)) cnp->cn_lkflags = LK_EXCLUSIVE; #ifdef NAMEI_DIAGNOSTIC vn_printf(dp, "lookup in "); #endif lkflags_save = cnp->cn_lkflags; cnp->cn_lkflags = compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags, cnp->cn_flags); error = VOP_LOOKUP(dp, &ndp->ni_vp, cnp); cnp->cn_lkflags = lkflags_save; if (error != 0) { KASSERT(ndp->ni_vp == NULL, ("leaf should be empty")); #ifdef NAMEI_DIAGNOSTIC printf("not found\n"); #endif if ((error == ENOENT) && (dp->v_vflag & VV_ROOT) && (dp->v_mount != NULL) && (dp->v_mount->mnt_flag & MNT_UNION)) { tdp = dp; dp = dp->v_mount->mnt_vnodecovered; VREF(dp); vput(tdp); vn_lock(dp, compute_cn_lkflags(dp->v_mount, cnp->cn_lkflags | LK_RETRY, cnp->cn_flags)); nameicap_tracker_add(ndp, dp); goto unionlookup; } if (error == ERELOOKUP) { vref(dp); ndp->ni_vp = dp; error = 0; relookup = 1; goto good; } if (error != EJUSTRETURN) goto bad; /* * At this point, we know we're at the end of the * pathname. If creating / renaming, we can consider * allowing the file or directory to be created / renamed, * provided we're not on a read-only filesystem. */ if (rdonly) { error = EROFS; goto bad; } /* trailing slash only allowed for directories */ if ((cnp->cn_flags & TRAILINGSLASH) && !(cnp->cn_flags & WILLBEDIR)) { error = ENOENT; goto bad; } if ((cnp->cn_flags & LOCKPARENT) == 0) VOP_UNLOCK(dp); /* * We return with ni_vp NULL to indicate that the entry * doesn't currently exist, leaving a pointer to the * (possibly locked) directory vnode in ndp->ni_dvp. */ if (cnp->cn_flags & SAVESTART) { ndp->ni_startdir = ndp->ni_dvp; VREF(ndp->ni_startdir); } goto success; } good: #ifdef NAMEI_DIAGNOSTIC printf("found\n"); #endif dp = ndp->ni_vp; /* * Check to see if the vnode has been mounted on; * if so find the root of the mounted filesystem. */ while (dp->v_type == VDIR && (mp = dp->v_mountedhere) && (cnp->cn_flags & NOCROSSMOUNT) == 0) { if (vfs_busy(mp, 0)) continue; vput(dp); if (dp != ndp->ni_dvp) vput(ndp->ni_dvp); else vrele(ndp->ni_dvp); vrefact(vp_crossmp); ndp->ni_dvp = vp_crossmp; error = VFS_ROOT(mp, compute_cn_lkflags(mp, cnp->cn_lkflags, cnp->cn_flags), &tdp); vfs_unbusy(mp); if (vn_lock(vp_crossmp, LK_SHARED | LK_NOWAIT)) panic("vp_crossmp exclusively locked or reclaimed"); if (error) { dpunlocked = 1; goto bad2; } ndp->ni_vp = dp = tdp; } /* * Check for symbolic link */ if ((dp->v_type == VLNK) && ((cnp->cn_flags & FOLLOW) || (cnp->cn_flags & TRAILINGSLASH) || *ndp->ni_next == '/')) { cnp->cn_flags |= ISSYMLINK; if (VN_IS_DOOMED(dp)) { /* * We can't know whether the directory was mounted with * NOSYMFOLLOW, so we can't follow safely. */ error = ENOENT; goto bad2; } if (dp->v_mount->mnt_flag & MNT_NOSYMFOLLOW) { error = EACCES; goto bad2; } /* * Symlink code always expects an unlocked dvp. */ if (ndp->ni_dvp != ndp->ni_vp) { VOP_UNLOCK(ndp->ni_dvp); ni_dvp_unlocked = 1; } goto success; } nextname: /* * Not a symbolic link that we will follow. Continue with the * next component if there is any; otherwise, we're done. */ KASSERT((cnp->cn_flags & ISLASTCN) || *ndp->ni_next == '/', ("lookup: invalid path state.")); if (relookup) { relookup = 0; ndp->ni_pathlen = prev_ni_pathlen; ndp->ni_next = prev_ni_next; if (ndp->ni_dvp != dp) vput(ndp->ni_dvp); else vrele(ndp->ni_dvp); goto dirloop; } if (cnp->cn_flags & ISDOTDOT) { error = nameicap_check_dotdot(ndp, ndp->ni_vp); if (error != 0) { #ifdef KTRACE if (KTRPOINT(curthread, KTR_CAPFAIL)) ktrcapfail(CAPFAIL_LOOKUP, NULL, NULL); #endif goto bad2; } } if (*ndp->ni_next == '/') { cnp->cn_nameptr = ndp->ni_next; while (*cnp->cn_nameptr == '/') { cnp->cn_nameptr++; ndp->ni_pathlen--; } if (ndp->ni_dvp != dp) vput(ndp->ni_dvp); else vrele(ndp->ni_dvp); goto dirloop; } /* * If we're processing a path with a trailing slash, * check that the end result is a directory. */ if ((cnp->cn_flags & TRAILINGSLASH) && dp->v_type != VDIR) { error = ENOTDIR; goto bad2; } /* * Disallow directory write attempts on read-only filesystems. */ if (rdonly && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) { error = EROFS; goto bad2; } if (cnp->cn_flags & SAVESTART) { ndp->ni_startdir = ndp->ni_dvp; VREF(ndp->ni_startdir); } if (!wantparent) { ni_dvp_unlocked = 2; if (ndp->ni_dvp != dp) vput(ndp->ni_dvp); else vrele(ndp->ni_dvp); } else if ((cnp->cn_flags & LOCKPARENT) == 0 && ndp->ni_dvp != dp) { VOP_UNLOCK(ndp->ni_dvp); ni_dvp_unlocked = 1; } if (cnp->cn_flags & AUDITVNODE1) AUDIT_ARG_VNODE1(dp); else if (cnp->cn_flags & AUDITVNODE2) AUDIT_ARG_VNODE2(dp); if ((cnp->cn_flags & LOCKLEAF) == 0) VOP_UNLOCK(dp); success: /* * Because of shared lookup we may have the vnode shared locked, but * the caller may want it to be exclusively locked. */ if (needs_exclusive_leaf(dp->v_mount, cnp->cn_flags) && VOP_ISLOCKED(dp) != LK_EXCLUSIVE) { vn_lock(dp, LK_UPGRADE | LK_RETRY); if (VN_IS_DOOMED(dp)) { error = ENOENT; goto bad2; } } return (0); bad2: if (ni_dvp_unlocked != 2) { if (dp != ndp->ni_dvp && !ni_dvp_unlocked) vput(ndp->ni_dvp); else vrele(ndp->ni_dvp); } bad: if (!dpunlocked) vput(dp); ndp->ni_vp = NULL; return (error); } /* * relookup - lookup a path name component * Used by lookup to re-acquire things. */ int relookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp) { struct vnode *dp = NULL; /* the directory we are searching */ int wantparent; /* 1 => wantparent or lockparent flag */ int rdonly; /* lookup read-only flag bit */ int error = 0; KASSERT(cnp->cn_flags & ISLASTCN, ("relookup: Not given last component.")); /* * Setup: break out flag bits into variables. */ wantparent = cnp->cn_flags & (LOCKPARENT|WANTPARENT); KASSERT(wantparent, ("relookup: parent not wanted.")); rdonly = cnp->cn_flags & RDONLY; cnp->cn_flags &= ~ISSYMLINK; dp = dvp; cnp->cn_lkflags = LK_EXCLUSIVE; vn_lock(dp, LK_EXCLUSIVE | LK_RETRY); /* * Search a new directory. * * The last component of the filename is left accessible via * cnp->cn_nameptr for callers that need the name. Callers needing * the name set the SAVENAME flag. When done, they assume * responsibility for freeing the pathname buffer. */ #ifdef NAMEI_DIAGNOSTIC printf("{%s}: ", cnp->cn_nameptr); #endif /* * Check for "" which represents the root directory after slash * removal. */ if (cnp->cn_nameptr[0] == '\0') { /* * Support only LOOKUP for "/" because lookup() * can't succeed for CREATE, DELETE and RENAME. */ KASSERT(cnp->cn_nameiop == LOOKUP, ("nameiop must be LOOKUP")); KASSERT(dp->v_type == VDIR, ("dp is not a directory")); if (!(cnp->cn_flags & LOCKLEAF)) VOP_UNLOCK(dp); *vpp = dp; /* XXX This should probably move to the top of function. */ if (cnp->cn_flags & SAVESTART) panic("lookup: SAVESTART"); return (0); } if (cnp->cn_flags & ISDOTDOT) panic ("relookup: lookup on dot-dot"); /* * We now have a segment name to search for, and a directory to search. */ #ifdef NAMEI_DIAGNOSTIC vn_printf(dp, "search in "); #endif if ((error = VOP_LOOKUP(dp, vpp, cnp)) != 0) { KASSERT(*vpp == NULL, ("leaf should be empty")); if (error != EJUSTRETURN) goto bad; /* * If creating and at end of pathname, then can consider * allowing file to be created. */ if (rdonly) { error = EROFS; goto bad; } /* ASSERT(dvp == ndp->ni_startdir) */ if (cnp->cn_flags & SAVESTART) VREF(dvp); if ((cnp->cn_flags & LOCKPARENT) == 0) VOP_UNLOCK(dp); /* * We return with ni_vp NULL to indicate that the entry * doesn't currently exist, leaving a pointer to the * (possibly locked) directory vnode in ndp->ni_dvp. */ return (0); } dp = *vpp; /* * Disallow directory write attempts on read-only filesystems. */ if (rdonly && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) { if (dvp == dp) vrele(dvp); else vput(dvp); error = EROFS; goto bad; } /* * Set the parent lock/ref state to the requested state. */ if ((cnp->cn_flags & LOCKPARENT) == 0 && dvp != dp) { if (wantparent) VOP_UNLOCK(dvp); else vput(dvp); } else if (!wantparent) vrele(dvp); /* * Check for symbolic link */ KASSERT(dp->v_type != VLNK || !(cnp->cn_flags & FOLLOW), ("relookup: symlink found.\n")); /* ASSERT(dvp == ndp->ni_startdir) */ if (cnp->cn_flags & SAVESTART) VREF(dvp); - + if ((cnp->cn_flags & LOCKLEAF) == 0) VOP_UNLOCK(dp); return (0); bad: vput(dp); *vpp = NULL; return (error); } void NDINIT_ALL(struct nameidata *ndp, u_long op, u_long flags, enum uio_seg segflg, const char *namep, int dirfd, struct vnode *startdir, cap_rights_t *rightsp, struct thread *td) { ndp->ni_cnd.cn_nameiop = op; ndp->ni_cnd.cn_flags = flags; ndp->ni_segflg = segflg; ndp->ni_dirp = namep; ndp->ni_dirfd = dirfd; ndp->ni_startdir = startdir; ndp->ni_resflags = 0; filecaps_init(&ndp->ni_filecaps); ndp->ni_cnd.cn_thread = td; if (rightsp != NULL) ndp->ni_rightsneeded = *rightsp; else cap_rights_init_zero(&ndp->ni_rightsneeded); } /* * Free data allocated by namei(); see namei(9) for details. */ void NDFREE(struct nameidata *ndp, const u_int flags) { int unlock_dvp; int unlock_vp; unlock_dvp = 0; unlock_vp = 0; if (!(flags & NDF_NO_FREE_PNBUF) && (ndp->ni_cnd.cn_flags & HASBUF)) { uma_zfree(namei_zone, ndp->ni_cnd.cn_pnbuf); ndp->ni_cnd.cn_flags &= ~HASBUF; } if (!(flags & NDF_NO_VP_UNLOCK) && (ndp->ni_cnd.cn_flags & LOCKLEAF) && ndp->ni_vp) unlock_vp = 1; if (!(flags & NDF_NO_DVP_UNLOCK) && (ndp->ni_cnd.cn_flags & LOCKPARENT) && ndp->ni_dvp != ndp->ni_vp) unlock_dvp = 1; if (!(flags & NDF_NO_VP_RELE) && ndp->ni_vp) { if (unlock_vp) { vput(ndp->ni_vp); unlock_vp = 0; } else vrele(ndp->ni_vp); ndp->ni_vp = NULL; } if (unlock_vp) VOP_UNLOCK(ndp->ni_vp); if (!(flags & NDF_NO_DVP_RELE) && (ndp->ni_cnd.cn_flags & (LOCKPARENT|WANTPARENT))) { if (unlock_dvp) { vput(ndp->ni_dvp); unlock_dvp = 0; } else vrele(ndp->ni_dvp); ndp->ni_dvp = NULL; } if (unlock_dvp) VOP_UNLOCK(ndp->ni_dvp); if (!(flags & NDF_NO_STARTDIR_RELE) && (ndp->ni_cnd.cn_flags & SAVESTART)) { vrele(ndp->ni_startdir); ndp->ni_startdir = NULL; } } /* * Determine if there is a suitable alternate filename under the specified * prefix for the specified path. If the create flag is set, then the * alternate prefix will be used so long as the parent directory exists. * This is used by the various compatibility ABIs so that Linux binaries prefer * files under /compat/linux for example. The chosen path (whether under * the prefix or under /) is returned in a kernel malloc'd buffer pointed * to by pathbuf. The caller is responsible for free'ing the buffer from * the M_TEMP bucket if one is returned. */ int kern_alternate_path(struct thread *td, const char *prefix, const char *path, enum uio_seg pathseg, char **pathbuf, int create, int dirfd) { struct nameidata nd, ndroot; char *ptr, *buf, *cp; size_t len, sz; int error; buf = (char *) malloc(MAXPATHLEN, M_TEMP, M_WAITOK); *pathbuf = buf; /* Copy the prefix into the new pathname as a starting point. */ len = strlcpy(buf, prefix, MAXPATHLEN); if (len >= MAXPATHLEN) { *pathbuf = NULL; free(buf, M_TEMP); return (EINVAL); } sz = MAXPATHLEN - len; ptr = buf + len; /* Append the filename to the prefix. */ if (pathseg == UIO_SYSSPACE) error = copystr(path, ptr, sz, &len); else error = copyinstr(path, ptr, sz, &len); if (error) { *pathbuf = NULL; free(buf, M_TEMP); return (error); } /* Only use a prefix with absolute pathnames. */ if (*ptr != '/') { error = EINVAL; goto keeporig; } if (dirfd != AT_FDCWD) { /* * We want the original because the "prefix" is * included in the already opened dirfd. */ bcopy(ptr, buf, len); return (0); } /* * We know that there is a / somewhere in this pathname. * Search backwards for it, to find the file's parent dir * to see if it exists in the alternate tree. If it does, * and we want to create a file (cflag is set). We don't * need to worry about the root comparison in this case. */ if (create) { for (cp = &ptr[len] - 1; *cp != '/'; cp--); *cp = '\0'; NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td); error = namei(&nd); *cp = '/'; if (error != 0) goto keeporig; } else { NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, buf, td); error = namei(&nd); if (error != 0) goto keeporig; /* * We now compare the vnode of the prefix to the one * vnode asked. If they resolve to be the same, then we * ignore the match so that the real root gets used. * This avoids the problem of traversing "../.." to find the * root directory and never finding it, because "/" resolves * to the emulation root directory. This is expensive :-( */ NDINIT(&ndroot, LOOKUP, FOLLOW, UIO_SYSSPACE, prefix, td); /* We shouldn't ever get an error from this namei(). */ error = namei(&ndroot); if (error == 0) { if (nd.ni_vp == ndroot.ni_vp) error = ENOENT; NDFREE(&ndroot, NDF_ONLY_PNBUF); vrele(ndroot.ni_vp); } } NDFREE(&nd, NDF_ONLY_PNBUF); vrele(nd.ni_vp); keeporig: /* If there was an error, use the original path name. */ if (error) bcopy(ptr, buf, len); return (error); } Index: head/sys/kern/vfs_mountroot.c =================================================================== --- head/sys/kern/vfs_mountroot.c (revision 363070) +++ head/sys/kern/vfs_mountroot.c (revision 363071) @@ -1,1153 +1,1153 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2010 Marcel Moolenaar * Copyright (c) 1999-2004 Poul-Henning Kamp * Copyright (c) 1999 Michael Smith * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include "opt_rootdevname.h" #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * The root filesystem is detailed in the kernel environment variable * vfs.root.mountfrom, which is expected to be in the general format * * :[][ :[] ...] * vfsname := the name of a VFS known to the kernel and capable * of being mounted as root * path := disk device name or other data used by the filesystem * to locate its physical store * * If the environment variable vfs.root.mountfrom is a space separated list, * each list element is tried in turn and the root filesystem will be mounted * from the first one that succeeds. * * The environment variable vfs.root.mountfrom.options is a comma delimited * set of string mount options. These mount options must be parseable * by nmount() in the kernel. */ static int parse_mount(char **); static struct mntarg *parse_mountroot_options(struct mntarg *, const char *); static int sysctl_vfs_root_mount_hold(SYSCTL_HANDLER_ARGS); static void vfs_mountroot_wait(void); static int vfs_mountroot_wait_if_neccessary(const char *fs, const char *dev); /* * The vnode of the system's root (/ in the filesystem, without chroot * active.) */ struct vnode *rootvnode; /* * Mount of the system's /dev. */ struct mount *rootdevmp; char *rootdevnames[2] = {NULL, NULL}; struct mtx root_holds_mtx; MTX_SYSINIT(root_holds, &root_holds_mtx, "root_holds", MTX_DEF); static TAILQ_HEAD(, root_hold_token) root_holds = TAILQ_HEAD_INITIALIZER(root_holds); enum action { A_CONTINUE, A_PANIC, A_REBOOT, A_RETRY }; enum rh_flags { RH_FREE, RH_ALLOC, RH_ARG, }; static enum action root_mount_onfail = A_CONTINUE; static int root_mount_mddev; static int root_mount_complete; /* By default wait up to 3 seconds for devices to appear. */ static int root_mount_timeout = 3; TUNABLE_INT("vfs.mountroot.timeout", &root_mount_timeout); static int root_mount_always_wait = 0; SYSCTL_INT(_vfs, OID_AUTO, root_mount_always_wait, CTLFLAG_RDTUN, &root_mount_always_wait, 0, "Wait for root mount holds even if the root device already exists"); SYSCTL_PROC(_vfs, OID_AUTO, root_mount_hold, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_root_mount_hold, "A", "List of root mount hold tokens"); static int sysctl_vfs_root_mount_hold(SYSCTL_HANDLER_ARGS) { struct sbuf sb; struct root_hold_token *h; int error; sbuf_new(&sb, NULL, 256, SBUF_AUTOEXTEND | SBUF_INCLUDENUL); mtx_lock(&root_holds_mtx); TAILQ_FOREACH(h, &root_holds, list) { if (h != TAILQ_FIRST(&root_holds)) sbuf_putc(&sb, ' '); sbuf_printf(&sb, "%s", h->who); } mtx_unlock(&root_holds_mtx); error = sbuf_finish(&sb); if (error == 0) error = SYSCTL_OUT(req, sbuf_data(&sb), sbuf_len(&sb)); sbuf_delete(&sb); return (error); } struct root_hold_token * root_mount_hold(const char *identifier) { struct root_hold_token *h; h = malloc(sizeof *h, M_DEVBUF, M_ZERO | M_WAITOK); h->flags = RH_ALLOC; h->who = identifier; mtx_lock(&root_holds_mtx); TSHOLD("root mount"); TAILQ_INSERT_TAIL(&root_holds, h, list); mtx_unlock(&root_holds_mtx); return (h); } void root_mount_hold_token(const char *identifier, struct root_hold_token *h) { #ifdef INVARIANTS struct root_hold_token *t; #endif h->flags = RH_ARG; h->who = identifier; mtx_lock(&root_holds_mtx); #ifdef INVARIANTS TAILQ_FOREACH(t, &root_holds, list) { if (t == h) { panic("Duplicate mount hold by '%s' on %p", identifier, h); } } #endif TSHOLD("root mount"); TAILQ_INSERT_TAIL(&root_holds, h, list); mtx_unlock(&root_holds_mtx); } void root_mount_rel(struct root_hold_token *h) { if (h == NULL || h->flags == RH_FREE) return; mtx_lock(&root_holds_mtx); TAILQ_REMOVE(&root_holds, h, list); TSRELEASE("root mount"); wakeup(&root_holds); mtx_unlock(&root_holds_mtx); if (h->flags == RH_ALLOC) { free(h, M_DEVBUF); } else h->flags = RH_FREE; } int root_mounted(void) { /* No mutex is acquired here because int stores are atomic. */ return (root_mount_complete); } static void set_rootvnode(void) { if (VFS_ROOT(TAILQ_FIRST(&mountlist), LK_EXCLUSIVE, &rootvnode)) panic("set_rootvnode: Cannot find root vnode"); VOP_UNLOCK(rootvnode); pwd_set_rootvnode(); } static int vfs_mountroot_devfs(struct thread *td, struct mount **mpp) { struct vfsoptlist *opts; struct vfsconf *vfsp; struct mount *mp; int error; *mpp = NULL; if (rootdevmp != NULL) { /* * Already have /dev; this happens during rerooting. */ error = vfs_busy(rootdevmp, 0); if (error != 0) return (error); *mpp = rootdevmp; } else { vfsp = vfs_byname("devfs"); KASSERT(vfsp != NULL, ("Could not find devfs by name")); if (vfsp == NULL) return (ENOENT); mp = vfs_mount_alloc(NULLVP, vfsp, "/dev", td->td_ucred); error = VFS_MOUNT(mp); KASSERT(error == 0, ("VFS_MOUNT(devfs) failed %d", error)); if (error) return (error); error = VFS_STATFS(mp, &mp->mnt_stat); KASSERT(error == 0, ("VFS_STATFS(devfs) failed %d", error)); if (error) return (error); opts = malloc(sizeof(struct vfsoptlist), M_MOUNT, M_WAITOK); TAILQ_INIT(opts); mp->mnt_opt = opts; mtx_lock(&mountlist_mtx); TAILQ_INSERT_HEAD(&mountlist, mp, mnt_list); mtx_unlock(&mountlist_mtx); *mpp = mp; rootdevmp = mp; vfs_op_exit(mp); } set_rootvnode(); error = kern_symlinkat(td, "/", AT_FDCWD, "dev", UIO_SYSSPACE); if (error) printf("kern_symlink /dev -> / returns %d\n", error); return (error); } static void vfs_mountroot_shuffle(struct thread *td, struct mount *mpdevfs) { struct nameidata nd; struct mount *mporoot, *mpnroot; struct vnode *vp, *vporoot, *vpdevfs; char *fspath; int error; mpnroot = TAILQ_NEXT(mpdevfs, mnt_list); /* Shuffle the mountlist. */ mtx_lock(&mountlist_mtx); mporoot = TAILQ_FIRST(&mountlist); TAILQ_REMOVE(&mountlist, mpdevfs, mnt_list); if (mporoot != mpdevfs) { TAILQ_REMOVE(&mountlist, mpnroot, mnt_list); TAILQ_INSERT_HEAD(&mountlist, mpnroot, mnt_list); } TAILQ_INSERT_TAIL(&mountlist, mpdevfs, mnt_list); mtx_unlock(&mountlist_mtx); cache_purgevfs(mporoot, true); if (mporoot != mpdevfs) cache_purgevfs(mpdevfs, true); if (VFS_ROOT(mporoot, LK_EXCLUSIVE, &vporoot)) panic("vfs_mountroot_shuffle: Cannot find root vnode"); VI_LOCK(vporoot); vporoot->v_iflag &= ~VI_MOUNT; VI_UNLOCK(vporoot); vporoot->v_mountedhere = NULL; mporoot->mnt_flag &= ~MNT_ROOTFS; mporoot->mnt_vnodecovered = NULL; vput(vporoot); /* Set up the new rootvnode, and purge the cache */ mpnroot->mnt_vnodecovered = NULL; set_rootvnode(); cache_purgevfs(rootvnode->v_mount, true); if (mporoot != mpdevfs) { /* Remount old root under /.mount or /mnt */ fspath = "/.mount"; NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF, UIO_SYSSPACE, fspath, td); error = namei(&nd); if (error) { NDFREE(&nd, NDF_ONLY_PNBUF); fspath = "/mnt"; NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF, UIO_SYSSPACE, fspath, td); error = namei(&nd); } if (!error) { vp = nd.ni_vp; error = (vp->v_type == VDIR) ? 0 : ENOTDIR; if (!error) error = vinvalbuf(vp, V_SAVE, 0, 0); if (!error) { cache_purge(vp); mporoot->mnt_vnodecovered = vp; vp->v_mountedhere = mporoot; strlcpy(mporoot->mnt_stat.f_mntonname, fspath, MNAMELEN); VOP_UNLOCK(vp); } else vput(vp); } NDFREE(&nd, NDF_ONLY_PNBUF); if (error) printf("mountroot: unable to remount previous root " "under /.mount or /mnt (error %d)\n", error); } /* Remount devfs under /dev */ NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF, UIO_SYSSPACE, "/dev", td); error = namei(&nd); if (!error) { vp = nd.ni_vp; error = (vp->v_type == VDIR) ? 0 : ENOTDIR; if (!error) error = vinvalbuf(vp, V_SAVE, 0, 0); if (!error) { vpdevfs = mpdevfs->mnt_vnodecovered; if (vpdevfs != NULL) { cache_purge(vpdevfs); vpdevfs->v_mountedhere = NULL; vrele(vpdevfs); } mpdevfs->mnt_vnodecovered = vp; vp->v_mountedhere = mpdevfs; VOP_UNLOCK(vp); } else vput(vp); } if (error) printf("mountroot: unable to remount devfs under /dev " "(error %d)\n", error); NDFREE(&nd, NDF_ONLY_PNBUF); if (mporoot == mpdevfs) { vfs_unbusy(mpdevfs); /* Unlink the no longer needed /dev/dev -> / symlink */ error = kern_funlinkat(td, AT_FDCWD, "/dev/dev", FD_NONE, UIO_SYSSPACE, 0, 0); if (error) printf("mountroot: unable to unlink /dev/dev " "(error %d)\n", error); } } /* * Configuration parser. */ /* Parser character classes. */ #define CC_WHITESPACE -1 #define CC_NONWHITESPACE -2 /* Parse errors. */ #define PE_EOF -1 #define PE_EOL -2 static __inline int parse_peek(char **conf) { return (**conf); } static __inline void parse_poke(char **conf, int c) { **conf = c; } static __inline void parse_advance(char **conf) { (*conf)++; } static int parse_skipto(char **conf, int mc) { int c, match; while (1) { c = parse_peek(conf); if (c == 0) return (PE_EOF); switch (mc) { case CC_WHITESPACE: match = (c == ' ' || c == '\t' || c == '\n') ? 1 : 0; break; case CC_NONWHITESPACE: if (c == '\n') return (PE_EOL); match = (c != ' ' && c != '\t') ? 1 : 0; break; default: match = (c == mc) ? 1 : 0; break; } if (match) break; parse_advance(conf); } return (0); } static int parse_token(char **conf, char **tok) { char *p; size_t len; int error; *tok = NULL; error = parse_skipto(conf, CC_NONWHITESPACE); if (error) return (error); p = *conf; error = parse_skipto(conf, CC_WHITESPACE); len = *conf - p; *tok = malloc(len + 1, M_TEMP, M_WAITOK | M_ZERO); bcopy(p, *tok, len); return (0); } static void parse_dir_ask_printenv(const char *var) { char *val; val = kern_getenv(var); if (val != NULL) { printf(" %s=%s\n", var, val); freeenv(val); } } static int parse_dir_ask(char **conf) { char name[80]; char *mnt; int error; vfs_mountroot_wait(); printf("\nLoader variables:\n"); parse_dir_ask_printenv("vfs.root.mountfrom"); parse_dir_ask_printenv("vfs.root.mountfrom.options"); printf("\nManual root filesystem specification:\n"); printf(" : [options]\n"); printf(" Mount using filesystem \n"); printf(" and with the specified (optional) option list.\n"); printf("\n"); printf(" eg. ufs:/dev/da0s1a\n"); printf(" zfs:zroot/ROOT/default\n"); printf(" cd9660:/dev/cd0 ro\n"); printf(" (which is equivalent to: "); printf("mount -t cd9660 -o ro /dev/cd0 /)\n"); printf("\n"); printf(" ? List valid disk boot devices\n"); printf(" . Yield 1 second (for background tasks)\n"); printf(" Abort manual input\n"); do { error = EINVAL; printf("\nmountroot> "); cngets(name, sizeof(name), GETS_ECHO); if (name[0] == '\0') break; if (name[0] == '?' && name[1] == '\0') { printf("\nList of GEOM managed disk devices:\n "); g_dev_print(); continue; } if (name[0] == '.' && name[1] == '\0') { pause("rmask", hz); continue; } mnt = name; error = parse_mount(&mnt); if (error == -1) printf("Invalid file system specification.\n"); } while (error != 0); return (error); } static int parse_dir_md(char **conf) { struct stat sb; struct thread *td; struct md_ioctl *mdio; char *path, *tok; int error, fd, len; td = curthread; error = parse_token(conf, &tok); if (error) return (error); len = strlen(tok); mdio = malloc(sizeof(*mdio) + len + 1, M_TEMP, M_WAITOK | M_ZERO); path = (void *)(mdio + 1); bcopy(tok, path, len); free(tok, M_TEMP); /* Get file status. */ error = kern_statat(td, 0, AT_FDCWD, path, UIO_SYSSPACE, &sb, NULL); if (error) goto out; /* Open /dev/mdctl so that we can attach/detach. */ error = kern_openat(td, AT_FDCWD, "/dev/" MDCTL_NAME, UIO_SYSSPACE, O_RDWR, 0); if (error) goto out; fd = td->td_retval[0]; mdio->md_version = MDIOVERSION; mdio->md_type = MD_VNODE; if (root_mount_mddev != -1) { mdio->md_unit = root_mount_mddev; (void)kern_ioctl(td, fd, MDIOCDETACH, (void *)mdio); /* Ignore errors. We don't care. */ root_mount_mddev = -1; } mdio->md_file = (void *)(mdio + 1); mdio->md_options = MD_AUTOUNIT | MD_READONLY; mdio->md_mediasize = sb.st_size; mdio->md_unit = 0; error = kern_ioctl(td, fd, MDIOCATTACH, (void *)mdio); if (error) goto out; if (mdio->md_unit > 9) { printf("rootmount: too many md units\n"); mdio->md_file = NULL; mdio->md_options = 0; mdio->md_mediasize = 0; error = kern_ioctl(td, fd, MDIOCDETACH, (void *)mdio); /* Ignore errors. We don't care. */ error = ERANGE; goto out; } root_mount_mddev = mdio->md_unit; printf(MD_NAME "%u attached to %s\n", root_mount_mddev, mdio->md_file); error = kern_close(td, fd); out: free(mdio, M_TEMP); return (error); } static int parse_dir_onfail(char **conf) { char *action; int error; error = parse_token(conf, &action); if (error) return (error); if (!strcmp(action, "continue")) root_mount_onfail = A_CONTINUE; else if (!strcmp(action, "panic")) root_mount_onfail = A_PANIC; else if (!strcmp(action, "reboot")) root_mount_onfail = A_REBOOT; else if (!strcmp(action, "retry")) root_mount_onfail = A_RETRY; else { printf("rootmount: %s: unknown action\n", action); error = EINVAL; } free(action, M_TEMP); return (0); } static int parse_dir_timeout(char **conf) { char *tok, *endtok; long secs; int error; error = parse_token(conf, &tok); if (error) return (error); secs = strtol(tok, &endtok, 0); error = (secs < 0 || *endtok != '\0') ? EINVAL : 0; if (!error) root_mount_timeout = secs; free(tok, M_TEMP); return (error); } static int parse_directive(char **conf) { char *dir; int error; error = parse_token(conf, &dir); if (error) return (error); if (strcmp(dir, ".ask") == 0) error = parse_dir_ask(conf); else if (strcmp(dir, ".md") == 0) error = parse_dir_md(conf); else if (strcmp(dir, ".onfail") == 0) error = parse_dir_onfail(conf); else if (strcmp(dir, ".timeout") == 0) error = parse_dir_timeout(conf); else { printf("mountroot: invalid directive `%s'\n", dir); /* Ignore the rest of the line. */ (void)parse_skipto(conf, '\n'); error = EINVAL; } free(dir, M_TEMP); return (error); } static int parse_mount_dev_present(const char *dev) { struct nameidata nd; int error; NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF, UIO_SYSSPACE, dev, curthread); error = namei(&nd); if (!error) vput(nd.ni_vp); NDFREE(&nd, NDF_ONLY_PNBUF); return (error != 0) ? 0 : 1; } #define ERRMSGL 255 static int parse_mount(char **conf) { char *errmsg; struct mntarg *ma; char *dev, *fs, *opts, *tok; int delay, error, timeout; error = parse_token(conf, &tok); if (error) return (error); fs = tok; error = parse_skipto(&tok, ':'); if (error) { free(fs, M_TEMP); return (error); } parse_poke(&tok, '\0'); parse_advance(&tok); dev = tok; if (root_mount_mddev != -1) { /* Handle substitution for the md unit number. */ tok = strstr(dev, "md#"); if (tok != NULL) tok[2] = '0' + root_mount_mddev; } /* Parse options. */ error = parse_token(conf, &tok); opts = (error == 0) ? tok : NULL; printf("Trying to mount root from %s:%s [%s]...\n", fs, dev, (opts != NULL) ? opts : ""); errmsg = malloc(ERRMSGL, M_TEMP, M_WAITOK | M_ZERO); if (vfs_byname(fs) == NULL) { strlcpy(errmsg, "unknown file system", ERRMSGL); error = ENOENT; goto out; } error = vfs_mountroot_wait_if_neccessary(fs, dev); if (error != 0) goto out; delay = hz / 10; timeout = root_mount_timeout * hz; for (;;) { ma = NULL; ma = mount_arg(ma, "fstype", fs, -1); ma = mount_arg(ma, "fspath", "/", -1); ma = mount_arg(ma, "from", dev, -1); ma = mount_arg(ma, "errmsg", errmsg, ERRMSGL); ma = mount_arg(ma, "ro", NULL, 0); ma = parse_mountroot_options(ma, opts); error = kernel_mount(ma, MNT_ROOTFS); if (error == 0 || timeout <= 0) break; if (root_mount_timeout * hz == timeout || (bootverbose && timeout % hz == 0)) { printf("Mounting from %s:%s failed with error %d; " "retrying for %d more second%s\n", fs, dev, error, timeout / hz, (timeout / hz > 1) ? "s" : ""); } pause("rmretry", delay); timeout -= delay; } out: if (error) { printf("Mounting from %s:%s failed with error %d", fs, dev, error); if (errmsg[0] != '\0') printf(": %s", errmsg); printf(".\n"); } free(fs, M_TEMP); free(errmsg, M_TEMP); if (opts != NULL) free(opts, M_TEMP); /* kernel_mount can return -1 on error. */ return ((error < 0) ? EDOOFUS : error); } #undef ERRMSGL static int vfs_mountroot_parse(struct sbuf *sb, struct mount *mpdevfs) { struct mount *mp; char *conf; int error; root_mount_mddev = -1; retry: conf = sbuf_data(sb); mp = TAILQ_NEXT(mpdevfs, mnt_list); error = (mp == NULL) ? 0 : EDOOFUS; root_mount_onfail = A_CONTINUE; while (mp == NULL) { error = parse_skipto(&conf, CC_NONWHITESPACE); if (error == PE_EOL) { parse_advance(&conf); continue; } if (error < 0) break; switch (parse_peek(&conf)) { case '#': error = parse_skipto(&conf, '\n'); break; case '.': error = parse_directive(&conf); break; default: error = parse_mount(&conf); if (error == -1) { printf("mountroot: invalid file system " "specification.\n"); error = 0; } break; } if (error < 0) break; /* Ignore any trailing garbage on the line. */ if (parse_peek(&conf) != '\n') { printf("mountroot: advancing to next directive...\n"); (void)parse_skipto(&conf, '\n'); } mp = TAILQ_NEXT(mpdevfs, mnt_list); } if (mp != NULL) return (0); /* * We failed to mount (a new) root. */ switch (root_mount_onfail) { case A_CONTINUE: break; case A_PANIC: panic("mountroot: unable to (re-)mount root."); /* NOTREACHED */ case A_RETRY: goto retry; case A_REBOOT: kern_reboot(RB_NOSYNC); /* NOTREACHED */ } return (error); } static void vfs_mountroot_conf0(struct sbuf *sb) { char *s, *tok, *mnt, *opt; int error; sbuf_printf(sb, ".onfail panic\n"); sbuf_printf(sb, ".timeout %d\n", root_mount_timeout); if (boothowto & RB_ASKNAME) sbuf_printf(sb, ".ask\n"); #ifdef ROOTDEVNAME if (boothowto & RB_DFLTROOT) sbuf_printf(sb, "%s\n", ROOTDEVNAME); #endif if (boothowto & RB_CDROM) { sbuf_printf(sb, "cd9660:/dev/cd0 ro\n"); sbuf_printf(sb, ".timeout 0\n"); sbuf_printf(sb, "cd9660:/dev/cd1 ro\n"); sbuf_printf(sb, ".timeout %d\n", root_mount_timeout); } s = kern_getenv("vfs.root.mountfrom"); if (s != NULL) { opt = kern_getenv("vfs.root.mountfrom.options"); tok = s; error = parse_token(&tok, &mnt); while (!error) { sbuf_printf(sb, "%s %s\n", mnt, (opt != NULL) ? opt : ""); free(mnt, M_TEMP); error = parse_token(&tok, &mnt); } if (opt != NULL) freeenv(opt); freeenv(s); } if (rootdevnames[0] != NULL) sbuf_printf(sb, "%s\n", rootdevnames[0]); if (rootdevnames[1] != NULL) sbuf_printf(sb, "%s\n", rootdevnames[1]); #ifdef ROOTDEVNAME if (!(boothowto & RB_DFLTROOT)) sbuf_printf(sb, "%s\n", ROOTDEVNAME); #endif if (!(boothowto & RB_ASKNAME)) sbuf_printf(sb, ".ask\n"); } static int vfs_mountroot_readconf(struct thread *td, struct sbuf *sb) { static char buf[128]; struct nameidata nd; off_t ofs; ssize_t resid; int error, flags, len; NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, "/.mount.conf", td); flags = FREAD; error = vn_open(&nd, &flags, 0, NULL); if (error) return (error); NDFREE(&nd, NDF_ONLY_PNBUF); ofs = 0; len = sizeof(buf) - 1; while (1) { error = vn_rdwr(UIO_READ, nd.ni_vp, buf, len, ofs, UIO_SYSSPACE, IO_NODELOCKED, td->td_ucred, NOCRED, &resid, td); if (error) break; if (resid == len) break; buf[len - resid] = 0; sbuf_printf(sb, "%s", buf); ofs += len - resid; } VOP_UNLOCK(nd.ni_vp); vn_close(nd.ni_vp, FREAD, td->td_ucred, td); return (error); } static void vfs_mountroot_wait(void) { struct root_hold_token *h; struct timeval lastfail; int curfail; TSENTER(); curfail = 0; while (1) { g_waitidle(); mtx_lock(&root_holds_mtx); if (TAILQ_EMPTY(&root_holds)) { mtx_unlock(&root_holds_mtx); break; } if (ppsratecheck(&lastfail, &curfail, 1)) { printf("Root mount waiting for:"); TAILQ_FOREACH(h, &root_holds, list) printf(" %s", h->who); printf("\n"); } TSWAIT("root mount"); msleep(&root_holds, &root_holds_mtx, PZERO | PDROP, "roothold", hz); TSUNWAIT("root mount"); } TSEXIT(); } static int vfs_mountroot_wait_if_neccessary(const char *fs, const char *dev) { int delay, timeout; /* * In case of ZFS and NFS we don't have a way to wait for * specific device. Also do the wait if the user forced that * behaviour by setting vfs.root_mount_always_wait=1. */ if (strcmp(fs, "zfs") == 0 || strstr(fs, "nfs") != NULL || dev[0] == '\0' || root_mount_always_wait != 0) { vfs_mountroot_wait(); return (0); } /* * Otherwise, no point in waiting if the device is already there. * Note that we must wait for GEOM to finish reconfiguring itself, * eg for geom_part(4) to finish tasting. */ g_waitidle(); if (parse_mount_dev_present(dev)) return (0); /* * No luck. Let's wait. This code looks weird, but it's that way * to behave exactly as it used to work before. */ vfs_mountroot_wait(); printf("mountroot: waiting for device %s...\n", dev); delay = hz / 10; timeout = root_mount_timeout * hz; do { pause("rmdev", delay); timeout -= delay; } while (timeout > 0 && !parse_mount_dev_present(dev)); if (timeout <= 0) return (ENODEV); return (0); } void vfs_mountroot(void) { struct mount *mp; struct sbuf *sb; struct thread *td; time_t timebase; int error; - + mtx_assert(&Giant, MA_NOTOWNED); TSENTER(); td = curthread; sb = sbuf_new_auto(); vfs_mountroot_conf0(sb); sbuf_finish(sb); error = vfs_mountroot_devfs(td, &mp); while (!error) { error = vfs_mountroot_parse(sb, mp); if (!error) { vfs_mountroot_shuffle(td, mp); sbuf_clear(sb); error = vfs_mountroot_readconf(td, sb); sbuf_finish(sb); } } sbuf_delete(sb); /* * Iterate over all currently mounted file systems and use * the time stamp found to check and/or initialize the RTC. * Call inittodr() only once and pass it the largest of the * timestamps we encounter. */ timebase = 0; mtx_lock(&mountlist_mtx); mp = TAILQ_FIRST(&mountlist); while (mp != NULL) { if (mp->mnt_time > timebase) timebase = mp->mnt_time; mp = TAILQ_NEXT(mp, mnt_list); } mtx_unlock(&mountlist_mtx); inittodr(timebase); /* Keep prison0's root in sync with the global rootvnode. */ mtx_lock(&prison0.pr_mtx); prison0.pr_root = rootvnode; vref(prison0.pr_root); mtx_unlock(&prison0.pr_mtx); mtx_lock(&root_holds_mtx); atomic_store_rel_int(&root_mount_complete, 1); wakeup(&root_mount_complete); mtx_unlock(&root_holds_mtx); EVENTHANDLER_INVOKE(mountroot); TSEXIT(); } static struct mntarg * parse_mountroot_options(struct mntarg *ma, const char *options) { char *p; char *name, *name_arg; char *val, *val_arg; char *opts; if (options == NULL || options[0] == '\0') return (ma); p = opts = strdup(options, M_MOUNT); if (opts == NULL) { return (ma); } while((name = strsep(&p, ",")) != NULL) { if (name[0] == '\0') break; val = strchr(name, '='); if (val != NULL) { *val = '\0'; ++val; } if( strcmp(name, "rw") == 0 || strcmp(name, "noro") == 0) { /* * The first time we mount the root file system, * we need to mount 'ro', so We need to ignore * 'rw' and 'noro' mount options. */ continue; } name_arg = strdup(name, M_MOUNT); val_arg = NULL; if (val != NULL) val_arg = strdup(val, M_MOUNT); ma = mount_arg(ma, name_arg, val_arg, (val_arg != NULL ? -1 : 0)); } free(opts, M_MOUNT); return (ma); } Index: head/sys/kern/vfs_subr.c =================================================================== --- head/sys/kern/vfs_subr.c (revision 363070) +++ head/sys/kern/vfs_subr.c (revision 363071) @@ -1,6560 +1,6560 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95 */ /* * External virtual filesystem routines */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_watchdog.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif static void delmntque(struct vnode *vp); static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, int slptimeo); static void syncer_shutdown(void *arg, int howto); static int vtryrecycle(struct vnode *vp); static void v_init_counters(struct vnode *); static void v_incr_devcount(struct vnode *); static void v_decr_devcount(struct vnode *); static void vgonel(struct vnode *); static void vfs_knllock(void *arg); static void vfs_knlunlock(void *arg); static void vfs_knl_assert_locked(void *arg); static void vfs_knl_assert_unlocked(void *arg); static void destroy_vpollinfo(struct vpollinfo *vi); static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, daddr_t startlbn, daddr_t endlbn); static void vnlru_recalc(void); /* * These fences are intended for cases where some synchronization is * needed between access of v_iflags and lockless vnode refcount (v_holdcnt * and v_usecount) updates. Access to v_iflags is generally synchronized * by the interlock, but we have some internal assertions that check vnode * flags without acquiring the lock. Thus, these fences are INVARIANTS-only * for now. */ #ifdef INVARIANTS #define VNODE_REFCOUNT_FENCE_ACQ() atomic_thread_fence_acq() #define VNODE_REFCOUNT_FENCE_REL() atomic_thread_fence_rel() #else #define VNODE_REFCOUNT_FENCE_ACQ() #define VNODE_REFCOUNT_FENCE_REL() #endif /* * Number of vnodes in existence. Increased whenever getnewvnode() * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode. */ static u_long __exclusive_cache_line numvnodes; SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0, "Number of vnodes in existence"); static counter_u64_t vnodes_created; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created, "Number of vnodes created by getnewvnode"); /* * Conversion tables for conversion from vnode types to inode formats * and back. */ enum vtype iftovt_tab[16] = { VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON, VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON }; int vttoif_tab[10] = { 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK, S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT }; /* * List of allocates vnodes in the system. */ static TAILQ_HEAD(freelst, vnode) vnode_list; static struct vnode *vnode_list_free_marker; static struct vnode *vnode_list_reclaim_marker; /* * "Free" vnode target. Free vnodes are rarely completely free, but are * just ones that are cheap to recycle. Usually they are for files which * have been stat'd but not read; these usually have inode and namecache * data attached to them. This target is the preferred minimum size of a * sub-cache consisting mostly of such files. The system balances the size * of this sub-cache with its complement to try to prevent either from * thrashing while the other is relatively inactive. The targets express * a preference for the best balance. * * "Above" this target there are 2 further targets (watermarks) related * to recyling of free vnodes. In the best-operating case, the cache is * exactly full, the free list has size between vlowat and vhiwat above the * free target, and recycling from it and normal use maintains this state. * Sometimes the free list is below vlowat or even empty, but this state * is even better for immediate use provided the cache is not full. * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free * ones) to reach one of these states. The watermarks are currently hard- * coded as 4% and 9% of the available space higher. These and the default * of 25% for wantfreevnodes are too large if the memory size is large. * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim * whenever vnlru_proc() becomes active. */ static long wantfreevnodes; static long __exclusive_cache_line freevnodes; SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD, &freevnodes, 0, "Number of \"free\" vnodes"); static long freevnodes_old; static counter_u64_t recycles_count; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count, "Number of vnodes recycled to meet vnode cache targets"); static counter_u64_t recycles_free_count; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count, "Number of free vnodes recycled to meet vnode cache targets"); /* * Various variables used for debugging the new implementation of * reassignbuf(). * XXX these are probably of (very) limited utility now. */ static int reassignbufcalls; SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS, &reassignbufcalls, 0, "Number of calls to reassignbuf"); static counter_u64_t deferred_inact; SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact, "Number of times inactive processing was deferred"); /* To keep more than one thread at a time from running vfs_getnewfsid */ static struct mtx mntid_mtx; /* * Lock for any access to the following: * vnode_list * numvnodes * freevnodes */ static struct mtx __exclusive_cache_line vnode_list_mtx; /* Publicly exported FS */ struct nfs_public nfs_pub; static uma_zone_t buf_trie_zone; /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */ static uma_zone_t vnode_zone; static uma_zone_t vnodepoll_zone; __read_frequently smr_t vfs_smr; /* * The workitem queue. * * It is useful to delay writes of file data and filesystem metadata * for tens of seconds so that quickly created and deleted files need * not waste disk bandwidth being created and removed. To realize this, * we append vnodes to a "workitem" queue. When running with a soft * updates implementation, most pending metadata dependencies should * not wait for more than a few seconds. Thus, mounted on block devices * are delayed only about a half the time that file data is delayed. * Similarly, directory updates are more critical, so are only delayed * about a third the time that file data is delayed. Thus, there are * SYNCER_MAXDELAY queues that are processed round-robin at a rate of * one each second (driven off the filesystem syncer process). The * syncer_delayno variable indicates the next queue that is to be processed. * Items that need to be processed soon are placed in this queue: * * syncer_workitem_pending[syncer_delayno] * * A delay of fifteen seconds is done by placing the request fifteen * entries later in the queue: * * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask] * */ static int syncer_delayno; static long syncer_mask; LIST_HEAD(synclist, bufobj); static struct synclist *syncer_workitem_pending; /* * The sync_mtx protects: * bo->bo_synclist * sync_vnode_count * syncer_delayno * syncer_state * syncer_workitem_pending * syncer_worklist_len * rushjob */ static struct mtx sync_mtx; static struct cv sync_wakeup; #define SYNCER_MAXDELAY 32 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */ static int syncdelay = 30; /* max time to delay syncing data */ static int filedelay = 30; /* time to delay syncing files */ SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0, "Time to delay syncing files (in seconds)"); static int dirdelay = 29; /* time to delay syncing directories */ SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0, "Time to delay syncing directories (in seconds)"); static int metadelay = 28; /* time to delay syncing metadata */ SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0, "Time to delay syncing metadata (in seconds)"); static int rushjob; /* number of slots to run ASAP */ static int stat_rush_requests; /* number of times I/O speeded up */ SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0, "Number of times I/O speeded up (rush requests)"); #define VDBATCH_SIZE 8 struct vdbatch { u_int index; long freevnodes; struct mtx lock; struct vnode *tab[VDBATCH_SIZE]; }; DPCPU_DEFINE_STATIC(struct vdbatch, vd); static void vdbatch_dequeue(struct vnode *vp); /* * When shutting down the syncer, run it at four times normal speed. */ #define SYNCER_SHUTDOWN_SPEEDUP 4 static int sync_vnode_count; static int syncer_worklist_len; static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY } syncer_state; /* Target for maximum number of vnodes. */ u_long desiredvnodes; static u_long gapvnodes; /* gap between wanted and desired */ static u_long vhiwat; /* enough extras after expansion */ static u_long vlowat; /* minimal extras before expansion */ static u_long vstir; /* nonzero to stir non-free vnodes */ static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */ static u_long vnlru_read_freevnodes(void); /* * Note that no attempt is made to sanitize these parameters. */ static int sysctl_maxvnodes(SYSCTL_HANDLER_ARGS) { u_long val; int error; val = desiredvnodes; error = sysctl_handle_long(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val == desiredvnodes) return (0); mtx_lock(&vnode_list_mtx); desiredvnodes = val; wantfreevnodes = desiredvnodes / 4; vnlru_recalc(); mtx_unlock(&vnode_list_mtx); /* * XXX There is no protection against multiple threads changing * desiredvnodes at the same time. Locking above only helps vnlru and * getnewvnode. */ vfs_hash_changesize(desiredvnodes); cache_changesize(desiredvnodes); return (0); } SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes, "LU", "Target for maximum number of vnodes"); static int sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS) { u_long val; int error; val = wantfreevnodes; error = sysctl_handle_long(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val == wantfreevnodes) return (0); mtx_lock(&vnode_list_mtx); wantfreevnodes = val; vnlru_recalc(); mtx_unlock(&vnode_list_mtx); return (0); } SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes, CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes, "LU", "Target for minimum number of \"free\" vnodes"); SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW, &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)"); static int vnlru_nowhere; SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW, &vnlru_nowhere, 0, "Number of times the vnlru process ran without success"); static int sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS) { struct vnode *vp; struct nameidata nd; char *buf; unsigned long ndflags; int error; if (req->newptr == NULL) return (EINVAL); if (req->newlen >= PATH_MAX) return (E2BIG); buf = malloc(PATH_MAX, M_TEMP, M_WAITOK); error = SYSCTL_IN(req, buf, req->newlen); if (error != 0) goto out; buf[req->newlen] = '\0'; ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1 | NOCACHE | SAVENAME; NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf, curthread); if ((error = namei(&nd)) != 0) goto out; vp = nd.ni_vp; if (VN_IS_DOOMED(vp)) { /* * This vnode is being recycled. Return != 0 to let the caller * know that the sysctl had no effect. Return EAGAIN because a * subsequent call will likely succeed (since namei will create * a new vnode if necessary) */ error = EAGAIN; goto putvnode; } counter_u64_add(recycles_count, 1); vgone(vp); putvnode: NDFREE(&nd, 0); out: free(buf, M_TEMP); return (error); } static int sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS) { struct thread *td = curthread; struct vnode *vp; struct file *fp; int error; int fd; if (req->newptr == NULL) return (EBADF); error = sysctl_handle_int(oidp, &fd, 0, req); if (error != 0) return (error); error = getvnode(curthread, fd, &cap_fcntl_rights, &fp); if (error != 0) return (error); vp = fp->f_vnode; error = vn_lock(vp, LK_EXCLUSIVE); if (error != 0) goto drop; counter_u64_add(recycles_count, 1); vgone(vp); VOP_UNLOCK(vp); drop: fdrop(fp, td); return (error); } SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode, CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname"); SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode, CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, sysctl_ftry_reclaim_vnode, "I", "Try to reclaim a vnode by its file descriptor"); /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */ static int vnsz2log; /* * Support for the bufobj clean & dirty pctrie. */ static void * buf_trie_alloc(struct pctrie *ptree) { return uma_zalloc(buf_trie_zone, M_NOWAIT); } static void buf_trie_free(struct pctrie *ptree, void *node) { uma_zfree(buf_trie_zone, node); } PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free); /* * Initialize the vnode management data structures. * * Reevaluate the following cap on the number of vnodes after the physical * memory size exceeds 512GB. In the limit, as the physical memory size * grows, the ratio of the memory size in KB to vnodes approaches 64:1. */ #ifndef MAXVNODES_MAX #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */ #endif static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker"); static struct vnode * vn_alloc_marker(struct mount *mp) { struct vnode *vp; vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO); vp->v_type = VMARKER; vp->v_mount = mp; return (vp); } static void vn_free_marker(struct vnode *vp) { MPASS(vp->v_type == VMARKER); free(vp, M_VNODE_MARKER); } /* * Initialize a vnode as it first enters the zone. */ static int vnode_init(void *mem, int size, int flags) { struct vnode *vp; vp = mem; bzero(vp, size); /* * Setup locks. */ vp->v_vnlock = &vp->v_lock; mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF); /* * By default, don't allow shared locks unless filesystems opt-in. */ lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT, LK_NOSHARE | LK_IS_VNODE); /* * Initialize bufobj. */ bufobj_init(&vp->v_bufobj, vp); /* * Initialize namecache. */ LIST_INIT(&vp->v_cache_src); TAILQ_INIT(&vp->v_cache_dst); /* * Initialize rangelocks. */ rangelock_init(&vp->v_rl); vp->v_dbatchcpu = NOCPU; mtx_lock(&vnode_list_mtx); TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist); mtx_unlock(&vnode_list_mtx); return (0); } /* * Free a vnode when it is cleared from the zone. */ static void vnode_fini(void *mem, int size) { struct vnode *vp; struct bufobj *bo; vp = mem; vdbatch_dequeue(vp); mtx_lock(&vnode_list_mtx); TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); mtx_unlock(&vnode_list_mtx); rangelock_destroy(&vp->v_rl); lockdestroy(vp->v_vnlock); mtx_destroy(&vp->v_interlock); bo = &vp->v_bufobj; rw_destroy(BO_LOCKPTR(bo)); } /* * Provide the size of NFS nclnode and NFS fh for calculation of the * vnode memory consumption. The size is specified directly to * eliminate dependency on NFS-private header. * * Other filesystems may use bigger or smaller (like UFS and ZFS) * private inode data, but the NFS-based estimation is ample enough. * Still, we care about differences in the size between 64- and 32-bit * platforms. * * Namecache structure size is heuristically * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1. */ #ifdef _LP64 #define NFS_NCLNODE_SZ (528 + 64) #define NC_SZ 148 #else #define NFS_NCLNODE_SZ (360 + 32) #define NC_SZ 92 #endif static void vntblinit(void *dummy __unused) { struct vdbatch *vd; int cpu, physvnodes, virtvnodes; u_int i; /* * Desiredvnodes is a function of the physical memory size and the * kernel's heap size. Generally speaking, it scales with the * physical memory size. The ratio of desiredvnodes to the physical * memory size is 1:16 until desiredvnodes exceeds 98,304. * Thereafter, the * marginal ratio of desiredvnodes to the physical memory size is * 1:64. However, desiredvnodes is limited by the kernel's heap * size. The memory required by desiredvnodes vnodes and vm objects * must not exceed 1/10th of the kernel's heap size. */ physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 + 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64; virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) + sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ)); desiredvnodes = min(physvnodes, virtvnodes); if (desiredvnodes > MAXVNODES_MAX) { if (bootverbose) printf("Reducing kern.maxvnodes %lu -> %lu\n", desiredvnodes, MAXVNODES_MAX); desiredvnodes = MAXVNODES_MAX; } wantfreevnodes = desiredvnodes / 4; mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF); TAILQ_INIT(&vnode_list); mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF); /* * The lock is taken to appease WITNESS. */ mtx_lock(&vnode_list_mtx); vnlru_recalc(); mtx_unlock(&vnode_list_mtx); vnode_list_free_marker = vn_alloc_marker(NULL); TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist); vnode_list_reclaim_marker = vn_alloc_marker(NULL); TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist); vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL, vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_SMR); vfs_smr = uma_zone_get_smr(vnode_zone); vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); /* * Preallocate enough nodes to support one-per buf so that * we can not fail an insert. reassignbuf() callers can not * tolerate the insertion failure. */ buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(), NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); uma_prealloc(buf_trie_zone, nbuf); vnodes_created = counter_u64_alloc(M_WAITOK); recycles_count = counter_u64_alloc(M_WAITOK); recycles_free_count = counter_u64_alloc(M_WAITOK); deferred_inact = counter_u64_alloc(M_WAITOK); /* * Initialize the filesystem syncer. */ syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE, &syncer_mask); syncer_maxdelay = syncer_mask + 1; mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF); cv_init(&sync_wakeup, "syncer"); for (i = 1; i <= sizeof(struct vnode); i <<= 1) vnsz2log++; vnsz2log--; CPU_FOREACH(cpu) { vd = DPCPU_ID_PTR((cpu), vd); bzero(vd, sizeof(*vd)); mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF); } } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL); /* * Mark a mount point as busy. Used to synchronize access and to delay * unmounting. Eventually, mountlist_mtx is not released on failure. * * vfs_busy() is a custom lock, it can block the caller. * vfs_busy() only sleeps if the unmount is active on the mount point. * For a mountpoint mp, vfs_busy-enforced lock is before lock of any * vnode belonging to mp. * * Lookup uses vfs_busy() to traverse mount points. * root fs var fs * / vnode lock A / vnode lock (/var) D * /var vnode lock B /log vnode lock(/var/log) E * vfs_busy lock C vfs_busy lock F * * Within each file system, the lock order is C->A->B and F->D->E. * * When traversing across mounts, the system follows that lock order: * * C->A->B * | * +->F->D->E * * The lookup() process for namei("/var") illustrates the process: * VOP_LOOKUP() obtains B while A is held * vfs_busy() obtains a shared lock on F while A and B are held * vput() releases lock on B * vput() releases lock on A * VFS_ROOT() obtains lock on D while shared lock on F is held * vfs_unbusy() releases shared lock on F * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A. * Attempt to lock A (instead of vp_crossmp) while D is held would * violate the global order, causing deadlocks. * * dounmount() locks B while F is drained. */ int vfs_busy(struct mount *mp, int flags) { MPASS((flags & ~MBF_MASK) == 0); CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags); if (vfs_op_thread_enter(mp)) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0); MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0); vfs_mp_count_add_pcpu(mp, ref, 1); vfs_mp_count_add_pcpu(mp, lockref, 1); vfs_op_thread_exit(mp); if (flags & MBF_MNTLSTLOCK) mtx_unlock(&mountlist_mtx); return (0); } MNT_ILOCK(mp); vfs_assert_mount_counters(mp); MNT_REF(mp); /* * If mount point is currently being unmounted, sleep until the * mount point fate is decided. If thread doing the unmounting fails, * it will clear MNTK_UNMOUNT flag before waking us up, indicating * that this mount point has survived the unmount attempt and vfs_busy * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE * flag in addition to MNTK_UNMOUNT, indicating that mount point is * about to be really destroyed. vfs_busy needs to release its * reference on the mount point in this case and return with ENOENT, * telling the caller that mount mount it tried to busy is no longer * valid. */ while (mp->mnt_kern_flag & MNTK_UNMOUNT) { if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) { MNT_REL(mp); MNT_IUNLOCK(mp); CTR1(KTR_VFS, "%s: failed busying before sleeping", __func__); return (ENOENT); } if (flags & MBF_MNTLSTLOCK) mtx_unlock(&mountlist_mtx); mp->mnt_kern_flag |= MNTK_MWAIT; msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0); if (flags & MBF_MNTLSTLOCK) mtx_lock(&mountlist_mtx); MNT_ILOCK(mp); } if (flags & MBF_MNTLSTLOCK) mtx_unlock(&mountlist_mtx); mp->mnt_lockref++; MNT_IUNLOCK(mp); return (0); } /* * Free a busy filesystem. */ void vfs_unbusy(struct mount *mp) { int c; CTR2(KTR_VFS, "%s: mp %p", __func__, mp); if (vfs_op_thread_enter(mp)) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); vfs_mp_count_sub_pcpu(mp, lockref, 1); vfs_mp_count_sub_pcpu(mp, ref, 1); vfs_op_thread_exit(mp); return; } MNT_ILOCK(mp); vfs_assert_mount_counters(mp); MNT_REL(mp); c = --mp->mnt_lockref; if (mp->mnt_vfs_ops == 0) { MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0); MNT_IUNLOCK(mp); return; } if (c < 0) vfs_dump_mount_counters(mp); if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) { MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT); CTR1(KTR_VFS, "%s: waking up waiters", __func__); mp->mnt_kern_flag &= ~MNTK_DRAINING; wakeup(&mp->mnt_lockref); } MNT_IUNLOCK(mp); } /* * Lookup a mount point by filesystem identifier. */ struct mount * vfs_getvfs(fsid_t *fsid) { struct mount *mp; CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { vfs_ref(mp); mtx_unlock(&mountlist_mtx); return (mp); } } mtx_unlock(&mountlist_mtx); CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); return ((struct mount *) 0); } /* * Lookup a mount point by filesystem identifier, busying it before * returning. * * To avoid congestion on mountlist_mtx, implement simple direct-mapped * cache for popular filesystem identifiers. The cache is lockess, using * the fact that struct mount's are never freed. In worst case we may * get pointer to unmounted or even different filesystem, so we have to * check what we got, and go slow way if so. */ struct mount * vfs_busyfs(fsid_t *fsid) { #define FSID_CACHE_SIZE 256 typedef struct mount * volatile vmp_t; static vmp_t cache[FSID_CACHE_SIZE]; struct mount *mp; int error; uint32_t hash; CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid); hash = fsid->val[0] ^ fsid->val[1]; hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1); mp = cache[hash]; if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0) goto slow; if (vfs_busy(mp, 0) != 0) { cache[hash] = NULL; goto slow; } if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) return (mp); else vfs_unbusy(mp); slow: mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) { error = vfs_busy(mp, MBF_MNTLSTLOCK); if (error) { cache[hash] = NULL; mtx_unlock(&mountlist_mtx); return (NULL); } cache[hash] = mp; return (mp); } } CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid); mtx_unlock(&mountlist_mtx); return ((struct mount *) 0); } /* * Check if a user can access privileged mount options. */ int vfs_suser(struct mount *mp, struct thread *td) { int error; if (jailed(td->td_ucred)) { /* * If the jail of the calling thread lacks permission for * this type of file system, deny immediately. */ if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag)) return (EPERM); /* * If the file system was mounted outside the jail of the * calling thread, deny immediately. */ if (prison_check(td->td_ucred, mp->mnt_cred) != 0) return (EPERM); } /* * If file system supports delegated administration, we don't check * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified * by the file system itself. * If this is not the user that did original mount, we check for * the PRIV_VFS_MOUNT_OWNER privilege. */ if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) && mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) { if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0) return (error); } return (0); } /* * Get a new unique fsid. Try to make its val[0] unique, since this value * will be used to create fake device numbers for stat(). Also try (but * not so hard) make its val[0] unique mod 2^16, since some emulators only * support 16-bit device numbers. We end up with unique val[0]'s for the * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls. * * Keep in mind that several mounts may be running in parallel. Starting * the search one past where the previous search terminated is both a * micro-optimization and a defense against returning the same fsid to * different mounts. */ void vfs_getnewfsid(struct mount *mp) { static uint16_t mntid_base; struct mount *nmp; fsid_t tfsid; int mtype; CTR2(KTR_VFS, "%s: mp %p", __func__, mp); mtx_lock(&mntid_mtx); mtype = mp->mnt_vfc->vfc_typenum; tfsid.val[1] = mtype; mtype = (mtype & 0xFF) << 24; for (;;) { tfsid.val[0] = makedev(255, mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF)); mntid_base++; if ((nmp = vfs_getvfs(&tfsid)) == NULL) break; vfs_rel(nmp); } mp->mnt_stat.f_fsid.val[0] = tfsid.val[0]; mp->mnt_stat.f_fsid.val[1] = tfsid.val[1]; mtx_unlock(&mntid_mtx); } /* * Knob to control the precision of file timestamps: * * 0 = seconds only; nanoseconds zeroed. * 1 = seconds and nanoseconds, accurate within 1/HZ. * 2 = seconds and nanoseconds, truncated to microseconds. * >=3 = seconds and nanoseconds, maximum precision. */ enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC }; static int timestamp_precision = TSP_USEC; SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW, ×tamp_precision, 0, "File timestamp precision (0: seconds, " "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, " "3+: sec + ns (max. precision))"); /* * Get a current timestamp. */ void vfs_timestamp(struct timespec *tsp) { struct timeval tv; switch (timestamp_precision) { case TSP_SEC: tsp->tv_sec = time_second; tsp->tv_nsec = 0; break; case TSP_HZ: getnanotime(tsp); break; case TSP_USEC: microtime(&tv); TIMEVAL_TO_TIMESPEC(&tv, tsp); break; case TSP_NSEC: default: nanotime(tsp); break; } } /* * Set vnode attributes to VNOVAL */ void vattr_null(struct vattr *vap) { vap->va_type = VNON; vap->va_size = VNOVAL; vap->va_bytes = VNOVAL; vap->va_mode = VNOVAL; vap->va_nlink = VNOVAL; vap->va_uid = VNOVAL; vap->va_gid = VNOVAL; vap->va_fsid = VNOVAL; vap->va_fileid = VNOVAL; vap->va_blocksize = VNOVAL; vap->va_rdev = VNOVAL; vap->va_atime.tv_sec = VNOVAL; vap->va_atime.tv_nsec = VNOVAL; vap->va_mtime.tv_sec = VNOVAL; vap->va_mtime.tv_nsec = VNOVAL; vap->va_ctime.tv_sec = VNOVAL; vap->va_ctime.tv_nsec = VNOVAL; vap->va_birthtime.tv_sec = VNOVAL; vap->va_birthtime.tv_nsec = VNOVAL; vap->va_flags = VNOVAL; vap->va_gen = VNOVAL; vap->va_vaflags = 0; } /* * Try to reduce the total number of vnodes. * * This routine (and its user) are buggy in at least the following ways: * - all parameters were picked years ago when RAM sizes were significantly * smaller * - it can pick vnodes based on pages used by the vm object, but filesystems * like ZFS don't use it making the pick broken * - since ZFS has its own aging policy it gets partially combated by this one * - a dedicated method should be provided for filesystems to let them decide * whether the vnode should be recycled * * This routine is called when we have too many vnodes. It attempts * to free vnodes and will potentially free vnodes that still * have VM backing store (VM backing store is typically the cause * of a vnode blowout so we want to do this). Therefore, this operation * is not considered cheap. * * A number of conditions may prevent a vnode from being reclaimed. * the buffer cache may have references on the vnode, a directory * vnode may still have references due to the namei cache representing * underlying files, or the vnode may be in active use. It is not * desirable to reuse such vnodes. These conditions may cause the * number of vnodes to reach some minimum value regardless of what * you set kern.maxvnodes to. Do not set kern.maxvnodes too low. * * @param reclaim_nc_src Only reclaim directories with outgoing namecache * entries if this argument is strue * @param trigger Only reclaim vnodes with fewer than this many resident * pages. * @param target How many vnodes to reclaim. * @return The number of vnodes that were reclaimed. */ static int vlrureclaim(bool reclaim_nc_src, int trigger, u_long target) { struct vnode *vp, *mvp; struct mount *mp; struct vm_object *object; u_long done; bool retried; mtx_assert(&vnode_list_mtx, MA_OWNED); retried = false; done = 0; mvp = vnode_list_reclaim_marker; restart: vp = mvp; while (done < target) { vp = TAILQ_NEXT(vp, v_vnodelist); if (__predict_false(vp == NULL)) break; if (__predict_false(vp->v_type == VMARKER)) continue; /* * If it's been deconstructed already, it's still * referenced, or it exceeds the trigger, skip it. * Also skip free vnodes. We are trying to make space * to expand the free list, not reduce it. */ if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src))) goto next_iter; if (vp->v_type == VBAD || vp->v_type == VNON) goto next_iter; if (!VI_TRYLOCK(vp)) goto next_iter; if (vp->v_usecount > 0 || vp->v_holdcnt == 0 || (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || VN_IS_DOOMED(vp) || vp->v_type == VNON) { VI_UNLOCK(vp); goto next_iter; } object = atomic_load_ptr(&vp->v_object); if (object == NULL || object->resident_page_count > trigger) { VI_UNLOCK(vp); goto next_iter; } vholdl(vp); VI_UNLOCK(vp); TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { vdrop(vp); goto next_iter_unlocked; } if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) { vdrop(vp); vn_finished_write(mp); goto next_iter_unlocked; } VI_LOCK(vp); if (vp->v_usecount > 0 || (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) || (vp->v_object != NULL && vp->v_object->resident_page_count > trigger)) { VOP_UNLOCK(vp); vdropl(vp); vn_finished_write(mp); goto next_iter_unlocked; } counter_u64_add(recycles_count, 1); vgonel(vp); VOP_UNLOCK(vp); vdropl(vp); vn_finished_write(mp); done++; next_iter_unlocked: if (should_yield()) kern_yield(PRI_USER); mtx_lock(&vnode_list_mtx); goto restart; next_iter: MPASS(vp->v_type != VMARKER); if (!should_yield()) continue; TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); mtx_unlock(&vnode_list_mtx); kern_yield(PRI_USER); mtx_lock(&vnode_list_mtx); goto restart; } if (done == 0 && !retried) { TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist); retried = true; goto restart; } return (done); } static int max_vnlru_free = 10000; /* limit on vnode free requests per call */ SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free, 0, "limit on vnode free requests per call to the vnlru_free routine"); /* * Attempt to reduce the free list by the requested amount. */ static int vnlru_free_locked(int count, struct vfsops *mnt_op) { struct vnode *vp, *mvp; struct mount *mp; int ocount; mtx_assert(&vnode_list_mtx, MA_OWNED); if (count > max_vnlru_free) count = max_vnlru_free; ocount = count; mvp = vnode_list_free_marker; restart: vp = mvp; while (count > 0) { vp = TAILQ_NEXT(vp, v_vnodelist); if (__predict_false(vp == NULL)) { TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist); break; } if (__predict_false(vp->v_type == VMARKER)) continue; /* * Don't recycle if our vnode is from different type * of mount point. Note that mp is type-safe, the * check does not reach unmapped address even if * vnode is reclaimed. * Don't recycle if we can't get the interlock without * blocking. */ if (vp->v_holdcnt > 0 || (mnt_op != NULL && (mp = vp->v_mount) != NULL && mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) { continue; } TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist); TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist); if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) { VI_UNLOCK(vp); continue; } vholdl(vp); count--; mtx_unlock(&vnode_list_mtx); VI_UNLOCK(vp); vtryrecycle(vp); vdrop(vp); mtx_lock(&vnode_list_mtx); goto restart; } return (ocount - count); } void vnlru_free(int count, struct vfsops *mnt_op) { mtx_lock(&vnode_list_mtx); vnlru_free_locked(count, mnt_op); mtx_unlock(&vnode_list_mtx); } static void vnlru_recalc(void) { mtx_assert(&vnode_list_mtx, MA_OWNED); gapvnodes = imax(desiredvnodes - wantfreevnodes, 100); vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */ vlowat = vhiwat / 2; } /* * Attempt to recycle vnodes in a context that is always safe to block. * Calling vlrurecycle() from the bowels of filesystem code has some * interesting deadlock problems. */ static struct proc *vnlruproc; static int vnlruproc_sig; /* * The main freevnodes counter is only updated when threads requeue their vnode * batches. CPUs are conditionally walked to compute a more accurate total. * * Limit how much of a slop are we willing to tolerate. Note: the actual value * at any given moment can still exceed slop, but it should not be by significant * margin in practice. */ #define VNLRU_FREEVNODES_SLOP 128 static u_long vnlru_read_freevnodes(void) { struct vdbatch *vd; long slop; int cpu; mtx_assert(&vnode_list_mtx, MA_OWNED); if (freevnodes > freevnodes_old) slop = freevnodes - freevnodes_old; else slop = freevnodes_old - freevnodes; if (slop < VNLRU_FREEVNODES_SLOP) return (freevnodes >= 0 ? freevnodes : 0); freevnodes_old = freevnodes; CPU_FOREACH(cpu) { vd = DPCPU_ID_PTR((cpu), vd); freevnodes_old += vd->freevnodes; } return (freevnodes_old >= 0 ? freevnodes_old : 0); } static bool vnlru_under(u_long rnumvnodes, u_long limit) { u_long rfreevnodes, space; if (__predict_false(rnumvnodes > desiredvnodes)) return (true); space = desiredvnodes - rnumvnodes; if (space < limit) { rfreevnodes = vnlru_read_freevnodes(); if (rfreevnodes > wantfreevnodes) space += rfreevnodes - wantfreevnodes; } return (space < limit); } static bool vnlru_under_unlocked(u_long rnumvnodes, u_long limit) { long rfreevnodes, space; if (__predict_false(rnumvnodes > desiredvnodes)) return (true); space = desiredvnodes - rnumvnodes; if (space < limit) { rfreevnodes = atomic_load_long(&freevnodes); if (rfreevnodes > wantfreevnodes) space += rfreevnodes - wantfreevnodes; } return (space < limit); } static void vnlru_kick(void) { mtx_assert(&vnode_list_mtx, MA_OWNED); if (vnlruproc_sig == 0) { vnlruproc_sig = 1; wakeup(vnlruproc); } } static void vnlru_proc(void) { u_long rnumvnodes, rfreevnodes, target; unsigned long onumvnodes; int done, force, trigger, usevnodes; bool reclaim_nc_src, want_reread; EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc, SHUTDOWN_PRI_FIRST); force = 0; want_reread = false; for (;;) { kproc_suspend_check(vnlruproc); mtx_lock(&vnode_list_mtx); rnumvnodes = atomic_load_long(&numvnodes); if (want_reread) { force = vnlru_under(numvnodes, vhiwat) ? 1 : 0; want_reread = false; } /* * If numvnodes is too large (due to desiredvnodes being * adjusted using its sysctl, or emergency growth), first * try to reduce it by discarding from the free list. */ if (rnumvnodes > desiredvnodes) { vnlru_free_locked(rnumvnodes - desiredvnodes, NULL); rnumvnodes = atomic_load_long(&numvnodes); } /* * Sleep if the vnode cache is in a good state. This is * when it is not over-full and has space for about a 4% * or 9% expansion (by growing its size or inexcessively * reducing its free list). Otherwise, try to reclaim * space for a 10% expansion. */ if (vstir && force == 0) { force = 1; vstir = 0; } if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) { vnlruproc_sig = 0; wakeup(&vnlruproc_sig); msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz); continue; } rfreevnodes = vnlru_read_freevnodes(); onumvnodes = rnumvnodes; /* * Calculate parameters for recycling. These are the same * throughout the loop to give some semblance of fairness. * The trigger point is to avoid recycling vnodes with lots * of resident pages. We aren't trying to free memory; we * are trying to recycle or at least free vnodes. */ if (rnumvnodes <= desiredvnodes) usevnodes = rnumvnodes - rfreevnodes; else usevnodes = rnumvnodes; if (usevnodes <= 0) usevnodes = 1; /* * The trigger value is is chosen to give a conservatively * large value to ensure that it alone doesn't prevent * making progress. The value can easily be so large that * it is effectively infinite in some congested and * misconfigured cases, and this is necessary. Normally * it is about 8 to 100 (pages), which is quite large. */ trigger = vm_cnt.v_page_count * 2 / usevnodes; if (force < 2) trigger = vsmalltrigger; reclaim_nc_src = force >= 3; target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1); target = target / 10 + 1; done = vlrureclaim(reclaim_nc_src, trigger, target); mtx_unlock(&vnode_list_mtx); if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes) uma_reclaim(UMA_RECLAIM_DRAIN); if (done == 0) { if (force == 0 || force == 1) { force = 2; continue; } if (force == 2) { force = 3; continue; } want_reread = true; force = 0; vnlru_nowhere++; tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3); } else { want_reread = true; kern_yield(PRI_USER); } } } static struct kproc_desc vnlru_kp = { "vnlru", vnlru_proc, &vnlruproc }; SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &vnlru_kp); - + /* * Routines having to do with the management of the vnode table. */ /* * Try to recycle a freed vnode. We abort if anyone picks up a reference * before we actually vgone(). This function must be called with the vnode * held to prevent the vnode from being returned to the free list midway * through vgone(). */ static int vtryrecycle(struct vnode *vp) { struct mount *vnmp; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); VNASSERT(vp->v_holdcnt, vp, ("vtryrecycle: Recycling vp %p without a reference.", vp)); /* * This vnode may found and locked via some other list, if so we * can't recycle it yet. */ if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) { CTR2(KTR_VFS, "%s: impossible to recycle, vp %p lock is already held", __func__, vp); return (EWOULDBLOCK); } /* * Don't recycle if its filesystem is being suspended. */ if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) { VOP_UNLOCK(vp); CTR2(KTR_VFS, "%s: impossible to recycle, cannot start the write for %p", __func__, vp); return (EBUSY); } /* * If we got this far, we need to acquire the interlock and see if * anyone picked up this vnode from another list. If not, we will * mark it with DOOMED via vgonel() so that anyone who does find it * will skip over it. */ VI_LOCK(vp); if (vp->v_usecount) { VOP_UNLOCK(vp); VI_UNLOCK(vp); vn_finished_write(vnmp); CTR2(KTR_VFS, "%s: impossible to recycle, %p is already referenced", __func__, vp); return (EBUSY); } if (!VN_IS_DOOMED(vp)) { counter_u64_add(recycles_free_count, 1); vgonel(vp); } VOP_UNLOCK(vp); VI_UNLOCK(vp); vn_finished_write(vnmp); return (0); } /* * Allocate a new vnode. * * The operation never returns an error. Returning an error was disabled * in r145385 (dated 2005) with the following comment: * * XXX Not all VFS_VGET/ffs_vget callers check returns. * * Given the age of this commit (almost 15 years at the time of writing this * comment) restoring the ability to fail requires a significant audit of * all codepaths. * * The routine can try to free a vnode or stall for up to 1 second waiting for * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation. */ static u_long vn_alloc_cyclecount; static struct vnode * __noinline vn_alloc_hard(struct mount *mp) { u_long rnumvnodes, rfreevnodes; mtx_lock(&vnode_list_mtx); rnumvnodes = atomic_load_long(&numvnodes); if (rnumvnodes + 1 < desiredvnodes) { vn_alloc_cyclecount = 0; goto alloc; } rfreevnodes = vnlru_read_freevnodes(); if (vn_alloc_cyclecount++ >= rfreevnodes) { vn_alloc_cyclecount = 0; vstir = 1; } /* * Grow the vnode cache if it will not be above its target max * after growing. Otherwise, if the free list is nonempty, try * to reclaim 1 item from it before growing the cache (possibly * above its target max if the reclamation failed or is delayed). * Otherwise, wait for some space. In all cases, schedule * vnlru_proc() if we are getting short of space. The watermarks * should be chosen so that we never wait or even reclaim from * the free list to below its target minimum. */ if (vnlru_free_locked(1, NULL) > 0) goto alloc; if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) { /* * Wait for space for a new vnode. */ vnlru_kick(); msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz); if (atomic_load_long(&numvnodes) + 1 > desiredvnodes && vnlru_read_freevnodes() > 1) vnlru_free_locked(1, NULL); } alloc: rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; if (vnlru_under(rnumvnodes, vlowat)) vnlru_kick(); mtx_unlock(&vnode_list_mtx); return (uma_zalloc_smr(vnode_zone, M_WAITOK)); } static struct vnode * vn_alloc(struct mount *mp) { u_long rnumvnodes; if (__predict_false(vn_alloc_cyclecount != 0)) return (vn_alloc_hard(mp)); rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1; if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) { atomic_subtract_long(&numvnodes, 1); return (vn_alloc_hard(mp)); } return (uma_zalloc_smr(vnode_zone, M_WAITOK)); } static void vn_free(struct vnode *vp) { atomic_subtract_long(&numvnodes, 1); uma_zfree_smr(vnode_zone, vp); } /* * Return the next vnode from the free list. */ int getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops, struct vnode **vpp) { struct vnode *vp; struct thread *td; struct lock_object *lo; CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag); KASSERT(vops->registered, ("%s: not registered vector op %p\n", __func__, vops)); td = curthread; if (td->td_vp_reserved != NULL) { vp = td->td_vp_reserved; td->td_vp_reserved = NULL; } else { vp = vn_alloc(mp); } counter_u64_add(vnodes_created, 1); /* * Locks are given the generic name "vnode" when created. * Follow the historic practice of using the filesystem * name when they allocated, e.g., "zfs", "ufs", "nfs, etc. * * Locks live in a witness group keyed on their name. Thus, * when a lock is renamed, it must also move from the witness * group of its old name to the witness group of its new name. * * The change only needs to be made when the vnode moves * from one filesystem type to another. We ensure that each * filesystem use a single static name pointer for its tag so * that we can compare pointers rather than doing a strcmp(). */ lo = &vp->v_vnlock->lock_object; #ifdef WITNESS if (lo->lo_name != tag) { #endif lo->lo_name = tag; #ifdef WITNESS WITNESS_DESTROY(lo); WITNESS_INIT(lo, tag); } #endif /* * By default, don't allow shared locks unless filesystems opt-in. */ vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE; /* * Finalize various vnode identity bits. */ KASSERT(vp->v_object == NULL, ("stale v_object %p", vp)); KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp)); KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp)); vp->v_type = VNON; vp->v_op = vops; v_init_counters(vp); vp->v_bufobj.bo_ops = &buf_ops_bio; #ifdef DIAGNOSTIC if (mp == NULL && vops != &dead_vnodeops) printf("NULL mp in getnewvnode(9), tag %s\n", tag); #endif #ifdef MAC mac_vnode_init(vp); if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0) mac_vnode_associate_singlelabel(mp, vp); #endif if (mp != NULL) { vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize; if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0) vp->v_vflag |= VV_NOKNOTE; } /* * For the filesystems which do not use vfs_hash_insert(), * still initialize v_hash to have vfs_hash_index() useful. * E.g., nullfs uses vfs_hash_index() on the lower vnode for * its own hashing. */ vp->v_hash = (uintptr_t)vp >> vnsz2log; *vpp = vp; return (0); } void getnewvnode_reserve(void) { struct thread *td; td = curthread; MPASS(td->td_vp_reserved == NULL); td->td_vp_reserved = vn_alloc(NULL); } void getnewvnode_drop_reserve(void) { struct thread *td; td = curthread; if (td->td_vp_reserved != NULL) { vn_free(td->td_vp_reserved); td->td_vp_reserved = NULL; } } static void freevnode(struct vnode *vp) { struct bufobj *bo; /* * The vnode has been marked for destruction, so free it. * * The vnode will be returned to the zone where it will * normally remain until it is needed for another vnode. We * need to cleanup (or verify that the cleanup has already * been done) any residual data left from its current use * so as not to contaminate the freshly allocated vnode. */ CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp); bo = &vp->v_bufobj; VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't")); VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp); VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count")); VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count")); VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's")); VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0")); VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp, ("clean blk trie not empty")); VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0")); VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp, ("dirty blk trie not empty")); VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst")); VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src")); VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for ..")); VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp, ("Dangling rangelock waiters")); VI_UNLOCK(vp); #ifdef MAC mac_vnode_destroy(vp); #endif if (vp->v_pollinfo != NULL) { destroy_vpollinfo(vp->v_pollinfo); vp->v_pollinfo = NULL; } #ifdef INVARIANTS /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */ vp->v_op = NULL; #endif vp->v_mountedhere = NULL; vp->v_unpcb = NULL; vp->v_rdev = NULL; vp->v_fifoinfo = NULL; vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0; vp->v_irflag = 0; vp->v_iflag = 0; vp->v_vflag = 0; bo->bo_flag = 0; vn_free(vp); } /* * Delete from old mount point vnode list, if on one. */ static void delmntque(struct vnode *vp) { struct mount *mp; VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); mp = vp->v_mount; if (mp == NULL) return; MNT_ILOCK(mp); VI_LOCK(vp); vp->v_mount = NULL; VI_UNLOCK(vp); VNASSERT(mp->mnt_nvnodelistsize > 0, vp, ("bad mount point vnode list size")); TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes); mp->mnt_nvnodelistsize--; MNT_REL(mp); MNT_IUNLOCK(mp); } static void insmntque_stddtr(struct vnode *vp, void *dtr_arg) { vp->v_data = NULL; vp->v_op = &dead_vnodeops; vgone(vp); vput(vp); } /* * Insert into list of vnodes for the new mount point, if available. */ int insmntque1(struct vnode *vp, struct mount *mp, void (*dtr)(struct vnode *, void *), void *dtr_arg) { KASSERT(vp->v_mount == NULL, ("insmntque: vnode already on per mount vnode list")); VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)")); ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp"); /* * We acquire the vnode interlock early to ensure that the * vnode cannot be recycled by another process releasing a * holdcnt on it before we get it on both the vnode list * and the active vnode list. The mount mutex protects only * manipulation of the vnode list and the vnode freelist * mutex protects only manipulation of the active vnode list. * Hence the need to hold the vnode interlock throughout. */ MNT_ILOCK(mp); VI_LOCK(vp); if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 && ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 || mp->mnt_nvnodelistsize == 0)) && (vp->v_vflag & VV_FORCEINSMQ) == 0) { VI_UNLOCK(vp); MNT_IUNLOCK(mp); if (dtr != NULL) dtr(vp, dtr_arg); return (EBUSY); } vp->v_mount = mp; MNT_REF(mp); TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes); VNASSERT(mp->mnt_nvnodelistsize >= 0, vp, ("neg mount point vnode list size")); mp->mnt_nvnodelistsize++; VI_UNLOCK(vp); MNT_IUNLOCK(mp); return (0); } int insmntque(struct vnode *vp, struct mount *mp) { return (insmntque1(vp, mp, insmntque_stddtr, NULL)); } /* * Flush out and invalidate all buffers associated with a bufobj * Called with the underlying object locked. */ int bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo) { int error; BO_LOCK(bo); if (flags & V_SAVE) { error = bufobj_wwait(bo, slpflag, slptimeo); if (error) { BO_UNLOCK(bo); return (error); } if (bo->bo_dirty.bv_cnt > 0) { BO_UNLOCK(bo); if ((error = BO_SYNC(bo, MNT_WAIT)) != 0) return (error); /* * XXX We could save a lock/unlock if this was only * enabled under INVARIANTS */ BO_LOCK(bo); if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) panic("vinvalbuf: dirty bufs"); } } /* * If you alter this loop please notice that interlock is dropped and * reacquired in flushbuflist. Special care is needed to ensure that * no race conditions occur from this. */ do { error = flushbuflist(&bo->bo_clean, flags, bo, slpflag, slptimeo); if (error == 0 && !(flags & V_CLEANONLY)) error = flushbuflist(&bo->bo_dirty, flags, bo, slpflag, slptimeo); if (error != 0 && error != EAGAIN) { BO_UNLOCK(bo); return (error); } } while (error != 0); /* * Wait for I/O to complete. XXX needs cleaning up. The vnode can * have write I/O in-progress but if there is a VM object then the * VM object can also have read-I/O in-progress. */ do { bufobj_wwait(bo, 0, 0); if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) { BO_UNLOCK(bo); vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx"); BO_LOCK(bo); } } while (bo->bo_numoutput > 0); BO_UNLOCK(bo); /* * Destroy the copy in the VM cache, too. */ if (bo->bo_object != NULL && (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) { VM_OBJECT_WLOCK(bo->bo_object); vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ? OBJPR_CLEANONLY : 0); VM_OBJECT_WUNLOCK(bo->bo_object); } #ifdef INVARIANTS BO_LOCK(bo); if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO | V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 || bo->bo_clean.bv_cnt > 0)) panic("vinvalbuf: flush failed"); if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 && bo->bo_dirty.bv_cnt > 0) panic("vinvalbuf: flush dirty failed"); BO_UNLOCK(bo); #endif return (0); } /* * Flush out and invalidate all buffers associated with a vnode. * Called with the underlying object locked. */ int vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo) { CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags); ASSERT_VOP_LOCKED(vp, "vinvalbuf"); if (vp->v_object != NULL && vp->v_object->handle != vp) return (0); return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo)); } /* * Flush out buffers on the specified list. * */ static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag, int slptimeo) { struct buf *bp, *nbp; int retval, error; daddr_t lblkno; b_xflags_t xflags; ASSERT_BO_WLOCKED(bo); retval = 0; TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) { /* * If we are flushing both V_NORMAL and V_ALT buffers then * do not skip any buffers. If we are flushing only V_NORMAL * buffers then skip buffers marked as BX_ALTDATA. If we are * flushing only V_ALT buffers then skip buffers not marked * as BX_ALTDATA. */ if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) && (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) || ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) { continue; } if (nbp != NULL) { lblkno = nbp->b_lblkno; xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN); } retval = EAGAIN; error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), "flushbuf", slpflag, slptimeo); if (error) { BO_LOCK(bo); return (error != ENOLCK ? error : EAGAIN); } KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); /* * XXX Since there are no node locks for NFS, I * believe there is a slight chance that a delayed * write will occur while sleeping just above, so * check for it. */ if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) && (flags & V_SAVE)) { bremfree(bp); bp->b_flags |= B_ASYNC; bwrite(bp); BO_LOCK(bo); return (EAGAIN); /* XXX: why not loop ? */ } bremfree(bp); bp->b_flags |= (B_INVAL | B_RELBUF); bp->b_flags &= ~B_ASYNC; brelse(bp); BO_LOCK(bo); if (nbp == NULL) break; nbp = gbincore(bo, lblkno); if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) != xflags) break; /* nbp invalid */ } return (retval); } int bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn) { struct buf *bp; int error; daddr_t lblkno; ASSERT_BO_LOCKED(bo); for (lblkno = startn;;) { again: bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno); if (bp == NULL || bp->b_lblkno >= endn || bp->b_lblkno < startn) break; error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0); if (error != 0) { BO_RLOCK(bo); if (error == ENOLCK) goto again; return (error); } KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); lblkno = bp->b_lblkno + 1; if ((bp->b_flags & B_MANAGED) == 0) bremfree(bp); bp->b_flags |= B_RELBUF; /* * In the VMIO case, use the B_NOREUSE flag to hint that the * pages backing each buffer in the range are unlikely to be * reused. Dirty buffers will have the hint applied once * they've been written. */ if ((bp->b_flags & B_VMIO) != 0) bp->b_flags |= B_NOREUSE; brelse(bp); BO_RLOCK(bo); } return (0); } /* * Truncate a file's buffer and pages to a specified length. This * is in lieu of the old vinvalbuf mechanism, which performed unneeded * sync activity. */ int vtruncbuf(struct vnode *vp, off_t length, int blksize) { struct buf *bp, *nbp; struct bufobj *bo; daddr_t startlbn; CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__, vp, blksize, (uintmax_t)length); /* * Round up to the *next* lbn. */ startlbn = howmany(length, blksize); ASSERT_VOP_LOCKED(vp, "vtruncbuf"); bo = &vp->v_bufobj; restart_unlocked: BO_LOCK(bo); while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN) ; if (length > 0) { restartsync: TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno > 0) continue; /* * Since we hold the vnode lock this should only * fail if we're racing with the buf daemon. */ if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) == ENOLCK) goto restart_unlocked; VNASSERT((bp->b_flags & B_DELWRI), vp, ("buf(%p) on dirty queue without DELWRI", bp)); bremfree(bp); bawrite(bp); BO_LOCK(bo); goto restartsync; } } bufobj_wwait(bo, 0, 0); BO_UNLOCK(bo); vnode_pager_setsize(vp, length); return (0); } /* * Invalidate the cached pages of a file's buffer within the range of block * numbers [startlbn, endlbn). */ void v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn, int blksize) { struct bufobj *bo; off_t start, end; ASSERT_VOP_LOCKED(vp, "v_inval_buf_range"); start = blksize * startlbn; end = blksize * endlbn; bo = &vp->v_bufobj; BO_LOCK(bo); MPASS(blksize == bo->bo_bsize); while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN) ; BO_UNLOCK(bo); vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1)); } static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo, daddr_t startlbn, daddr_t endlbn) { struct buf *bp, *nbp; bool anyfreed; ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked"); ASSERT_BO_LOCKED(bo); do { anyfreed = false; TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) == ENOLCK) { BO_LOCK(bo); return (EAGAIN); } bremfree(bp); bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = true; BO_LOCK(bo); if (nbp != NULL && (((nbp->b_xflags & BX_VNCLEAN) == 0) || nbp->b_vp != vp || (nbp->b_flags & B_DELWRI) != 0)) return (EAGAIN); } TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) == ENOLCK) { BO_LOCK(bo); return (EAGAIN); } bremfree(bp); bp->b_flags |= B_INVAL | B_RELBUF; bp->b_flags &= ~B_ASYNC; brelse(bp); anyfreed = true; BO_LOCK(bo); if (nbp != NULL && (((nbp->b_xflags & BX_VNDIRTY) == 0) || (nbp->b_vp != vp) || (nbp->b_flags & B_DELWRI) == 0)) return (EAGAIN); } } while (anyfreed); return (0); } static void buf_vlist_remove(struct buf *bp) { struct bufv *bv; KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp)); ASSERT_BO_WLOCKED(bp->b_bufobj); KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) != (BX_VNDIRTY|BX_VNCLEAN), ("buf_vlist_remove: Buf %p is on two lists", bp)); if (bp->b_xflags & BX_VNDIRTY) bv = &bp->b_bufobj->bo_dirty; else bv = &bp->b_bufobj->bo_clean; BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno); TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs); bv->bv_cnt--; bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN); } /* * Add the buffer to the sorted clean or dirty block list. * * NOTE: xflags is passed as a constant, optimizing this inline function! */ static void buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags) { struct bufv *bv; struct buf *n; int error; ASSERT_BO_WLOCKED(bo); KASSERT((bo->bo_flag & BO_NOBUFS) == 0, ("buf_vlist_add: bo %p does not allow bufs", bo)); KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0, ("dead bo %p", bo)); KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags)); bp->b_xflags |= xflags; if (xflags & BX_VNDIRTY) bv = &bo->bo_dirty; else bv = &bo->bo_clean; /* * Keep the list ordered. Optimize empty list insertion. Assume * we tend to grow at the tail so lookup_le should usually be cheaper * than _ge. */ if (bv->bv_cnt == 0 || bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno) TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs); else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL) TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs); else TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs); error = BUF_PCTRIE_INSERT(&bv->bv_root, bp); if (error) panic("buf_vlist_add: Preallocated nodes insufficient."); bv->bv_cnt++; } /* * Look up a buffer using the buffer tries. */ struct buf * gbincore(struct bufobj *bo, daddr_t lblkno) { struct buf *bp; ASSERT_BO_LOCKED(bo); bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno); if (bp != NULL) return (bp); return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno); } /* * Associate a buffer with a vnode. */ void bgetvp(struct vnode *vp, struct buf *bp) { struct bufobj *bo; bo = &vp->v_bufobj; ASSERT_BO_WLOCKED(bo); VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free")); CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags); VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp, ("bgetvp: bp already attached! %p", bp)); vhold(vp); bp->b_vp = vp; bp->b_bufobj = bo; /* * Insert onto list for new vnode. */ buf_vlist_add(bp, bo, BX_VNCLEAN); } /* * Disassociate a buffer from a vnode. */ void brelvp(struct buf *bp) { struct bufobj *bo; struct vnode *vp; CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); KASSERT(bp->b_vp != NULL, ("brelvp: NULL")); /* * Delete from old vnode list, if on one. */ vp = bp->b_vp; /* XXX */ bo = bp->b_bufobj; BO_LOCK(bo); if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) buf_vlist_remove(bp); else panic("brelvp: Buffer %p not on queue.", bp); if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { bo->bo_flag &= ~BO_ONWORKLST; mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; mtx_unlock(&sync_mtx); } bp->b_vp = NULL; bp->b_bufobj = NULL; BO_UNLOCK(bo); vdrop(vp); } /* * Add an item to the syncer work queue. */ static void vn_syncer_add_to_worklist(struct bufobj *bo, int delay) { int slot; ASSERT_BO_WLOCKED(bo); mtx_lock(&sync_mtx); if (bo->bo_flag & BO_ONWORKLST) LIST_REMOVE(bo, bo_synclist); else { bo->bo_flag |= BO_ONWORKLST; syncer_worklist_len++; } if (delay > syncer_maxdelay - 2) delay = syncer_maxdelay - 2; slot = (syncer_delayno + delay) & syncer_mask; LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist); mtx_unlock(&sync_mtx); } static int sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS) { int error, len; mtx_lock(&sync_mtx); len = syncer_worklist_len - sync_vnode_count; mtx_unlock(&sync_mtx); error = SYSCTL_OUT(req, &len, sizeof(len)); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0, sysctl_vfs_worklist_len, "I", "Syncer thread worklist length"); static struct proc *updateproc; static void sched_sync(void); static struct kproc_desc up_kp = { "syncer", sched_sync, &updateproc }; SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp); static int sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td) { struct vnode *vp; struct mount *mp; *bo = LIST_FIRST(slp); if (*bo == NULL) return (0); vp = bo2vnode(*bo); if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0) return (1); /* * We use vhold in case the vnode does not * successfully sync. vhold prevents the vnode from * going away when we unlock the sync_mtx so that * we can acquire the vnode interlock. */ vholdl(vp); mtx_unlock(&sync_mtx); VI_UNLOCK(vp); if (vn_start_write(vp, &mp, V_NOWAIT) != 0) { vdrop(vp); mtx_lock(&sync_mtx); return (*bo == LIST_FIRST(slp)); } vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); (void) VOP_FSYNC(vp, MNT_LAZY, td); VOP_UNLOCK(vp); vn_finished_write(mp); BO_LOCK(*bo); if (((*bo)->bo_flag & BO_ONWORKLST) != 0) { /* * Put us back on the worklist. The worklist * routine will remove us from our current * position and then add us back in at a later * position. */ vn_syncer_add_to_worklist(*bo, syncdelay); } BO_UNLOCK(*bo); vdrop(vp); mtx_lock(&sync_mtx); return (0); } static int first_printf = 1; /* * System filesystem synchronizer daemon. */ static void sched_sync(void) { struct synclist *next, *slp; struct bufobj *bo; long starttime; struct thread *td = curthread; int last_work_seen; int net_worklist_len; int syncer_final_iter; int error; last_work_seen = 0; syncer_final_iter = 0; syncer_state = SYNCER_RUNNING; starttime = time_uptime; td->td_pflags |= TDP_NORUNNINGBUF; EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc, SHUTDOWN_PRI_LAST); mtx_lock(&sync_mtx); for (;;) { if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter == 0) { mtx_unlock(&sync_mtx); kproc_suspend_check(td->td_proc); mtx_lock(&sync_mtx); } net_worklist_len = syncer_worklist_len - sync_vnode_count; if (syncer_state != SYNCER_RUNNING && starttime != time_uptime) { if (first_printf) { printf("\nSyncing disks, vnodes remaining... "); first_printf = 0; } printf("%d ", net_worklist_len); } starttime = time_uptime; /* * Push files whose dirty time has expired. Be careful * of interrupt race on slp queue. * * Skip over empty worklist slots when shutting down. */ do { slp = &syncer_workitem_pending[syncer_delayno]; syncer_delayno += 1; if (syncer_delayno == syncer_maxdelay) syncer_delayno = 0; next = &syncer_workitem_pending[syncer_delayno]; /* * If the worklist has wrapped since the * it was emptied of all but syncer vnodes, * switch to the FINAL_DELAY state and run * for one more second. */ if (syncer_state == SYNCER_SHUTTING_DOWN && net_worklist_len == 0 && last_work_seen == syncer_delayno) { syncer_state = SYNCER_FINAL_DELAY; syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP; } } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) && syncer_worklist_len > 0); /* * Keep track of the last time there was anything * on the worklist other than syncer vnodes. * Return to the SHUTTING_DOWN state if any * new work appears. */ if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING) last_work_seen = syncer_delayno; if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY) syncer_state = SYNCER_SHUTTING_DOWN; while (!LIST_EMPTY(slp)) { error = sync_vnode(slp, &bo, td); if (error == 1) { LIST_REMOVE(bo, bo_synclist); LIST_INSERT_HEAD(next, bo, bo_synclist); continue; } if (first_printf == 0) { /* * Drop the sync mutex, because some watchdog * drivers need to sleep while patting */ mtx_unlock(&sync_mtx); wdog_kern_pat(WD_LASTVAL); mtx_lock(&sync_mtx); } } if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0) syncer_final_iter--; /* * The variable rushjob allows the kernel to speed up the * processing of the filesystem syncer process. A rushjob * value of N tells the filesystem syncer to process the next * N seconds worth of work on its queue ASAP. Currently rushjob * is used by the soft update code to speed up the filesystem * syncer process when the incore state is getting so far * ahead of the disk that the kernel memory pool is being * threatened with exhaustion. */ if (rushjob > 0) { rushjob -= 1; continue; } /* * Just sleep for a short period of time between * iterations when shutting down to allow some I/O * to happen. * * If it has taken us less than a second to process the * current work, then wait. Otherwise start right over * again. We can still lose time if any single round * takes more than two seconds, but it does not really * matter as we are just trying to generally pace the * filesystem activity. */ if (syncer_state != SYNCER_RUNNING || time_uptime == starttime) { thread_lock(td); sched_prio(td, PPAUSE); thread_unlock(td); } if (syncer_state != SYNCER_RUNNING) cv_timedwait(&sync_wakeup, &sync_mtx, hz / SYNCER_SHUTDOWN_SPEEDUP); else if (time_uptime == starttime) cv_timedwait(&sync_wakeup, &sync_mtx, hz); } } /* * Request the syncer daemon to speed up its work. * We never push it to speed up more than half of its * normal turn time, otherwise it could take over the cpu. */ int speedup_syncer(void) { int ret = 0; mtx_lock(&sync_mtx); if (rushjob < syncdelay / 2) { rushjob += 1; stat_rush_requests += 1; ret = 1; } mtx_unlock(&sync_mtx); cv_broadcast(&sync_wakeup); return (ret); } /* * Tell the syncer to speed up its work and run though its work * list several times, then tell it to shut down. */ static void syncer_shutdown(void *arg, int howto) { if (howto & RB_NOSYNC) return; mtx_lock(&sync_mtx); syncer_state = SYNCER_SHUTTING_DOWN; rushjob = 0; mtx_unlock(&sync_mtx); cv_broadcast(&sync_wakeup); kproc_shutdown(arg, howto); } void syncer_suspend(void) { syncer_shutdown(updateproc, 0); } void syncer_resume(void) { mtx_lock(&sync_mtx); first_printf = 1; syncer_state = SYNCER_RUNNING; mtx_unlock(&sync_mtx); cv_broadcast(&sync_wakeup); kproc_resume(updateproc); } /* * Reassign a buffer from one vnode to another. * Used to assign file specific control information * (indirect blocks) to the vnode to which they belong. */ void reassignbuf(struct buf *bp) { struct vnode *vp; struct bufobj *bo; int delay; #ifdef INVARIANTS struct bufv *bv; #endif vp = bp->b_vp; bo = bp->b_bufobj; ++reassignbufcalls; CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags); /* * B_PAGING flagged buffers cannot be reassigned because their vp * is not fully linked in. */ if (bp->b_flags & B_PAGING) panic("cannot reassign paging buffer"); /* * Delete from old vnode list, if on one. */ BO_LOCK(bo); if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) buf_vlist_remove(bp); else panic("reassignbuf: Buffer %p not on queue.", bp); /* * If dirty, put on list of dirty buffers; otherwise insert onto list * of clean buffers. */ if (bp->b_flags & B_DELWRI) { if ((bo->bo_flag & BO_ONWORKLST) == 0) { switch (vp->v_type) { case VDIR: delay = dirdelay; break; case VCHR: delay = metadelay; break; default: delay = filedelay; } vn_syncer_add_to_worklist(bo, delay); } buf_vlist_add(bp, bo, BX_VNDIRTY); } else { buf_vlist_add(bp, bo, BX_VNCLEAN); if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) { mtx_lock(&sync_mtx); LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; mtx_unlock(&sync_mtx); bo->bo_flag &= ~BO_ONWORKLST; } } #ifdef INVARIANTS bv = &bo->bo_clean; bp = TAILQ_FIRST(&bv->bv_hd); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bp = TAILQ_LAST(&bv->bv_hd, buflists); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bv = &bo->bo_dirty; bp = TAILQ_FIRST(&bv->bv_hd); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); bp = TAILQ_LAST(&bv->bv_hd, buflists); KASSERT(bp == NULL || bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); #endif BO_UNLOCK(bo); } static void v_init_counters(struct vnode *vp) { VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0, vp, ("%s called for an initialized vnode", __FUNCTION__)); ASSERT_VI_UNLOCKED(vp, __FUNCTION__); refcount_init(&vp->v_holdcnt, 1); refcount_init(&vp->v_usecount, 1); } /* * Increment si_usecount of the associated device, if any. */ static void v_incr_devcount(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __FUNCTION__); if (vp->v_type == VCHR && vp->v_rdev != NULL) { dev_lock(); vp->v_rdev->si_usecount++; dev_unlock(); } } /* * Decrement si_usecount of the associated device, if any. * * The caller is required to hold the interlock when transitioning a VCHR use * count to zero. This prevents a race with devfs_reclaim_vchr() that would * leak a si_usecount reference. The vnode lock will also prevent this race * if it is held while dropping the last ref. * * The race is: * * CPU1 CPU2 * devfs_reclaim_vchr * make v_usecount == 0 * VI_LOCK * sees v_usecount == 0, no updates * vp->v_rdev = NULL; * ... * VI_UNLOCK * VI_LOCK * v_decr_devcount * sees v_rdev == NULL, no updates * * In this scenario si_devcount decrement is not performed. */ static void v_decr_devcount(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, __func__); ASSERT_VI_LOCKED(vp, __FUNCTION__); if (vp->v_type == VCHR && vp->v_rdev != NULL) { dev_lock(); VNPASS(vp->v_rdev->si_usecount > 0, vp); vp->v_rdev->si_usecount--; dev_unlock(); } } /* * Grab a particular vnode from the free list, increment its * reference count and lock it. VIRF_DOOMED is set if the vnode * is being destroyed. Only callers who specify LK_RETRY will * see doomed vnodes. If inactive processing was delayed in * vput try to do it here. * * usecount is manipulated using atomics without holding any locks. * * holdcnt can be manipulated using atomics without holding any locks, * except when transitioning 1<->0, in which case the interlock is held. * * Consumers which don't guarantee liveness of the vnode can use SMR to * try to get a reference. Note this operation can fail since the vnode * may be awaiting getting freed by the time they get to it. */ enum vgetstate vget_prep_smr(struct vnode *vp) { enum vgetstate vs; VFS_SMR_ASSERT_ENTERED(); if (refcount_acquire_if_not_zero(&vp->v_usecount)) { vs = VGET_USECOUNT; } else { if (vhold_smr(vp)) vs = VGET_HOLDCNT; else vs = VGET_NONE; } return (vs); } enum vgetstate vget_prep(struct vnode *vp) { enum vgetstate vs; if (refcount_acquire_if_not_zero(&vp->v_usecount)) { vs = VGET_USECOUNT; } else { vhold(vp); vs = VGET_HOLDCNT; } return (vs); } int vget(struct vnode *vp, int flags, struct thread *td) { enum vgetstate vs; MPASS(td == curthread); vs = vget_prep(vp); return (vget_finish(vp, flags, vs)); } static int __noinline vget_finish_vchr(struct vnode *vp) { VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)")); /* * See the comment in vget_finish before usecount bump. */ if (refcount_acquire_if_not_zero(&vp->v_usecount)) { #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_holdcnt, -1); VNASSERT(old > 0, vp, ("%s: wrong hold count %d", __func__, old)); #else refcount_release(&vp->v_holdcnt); #endif return (0); } VI_LOCK(vp); if (refcount_acquire_if_not_zero(&vp->v_usecount)) { #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_holdcnt, -1); VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); #else refcount_release(&vp->v_holdcnt); #endif VI_UNLOCK(vp); return (0); } v_incr_devcount(vp); refcount_acquire(&vp->v_usecount); VI_UNLOCK(vp); return (0); } int vget_finish(struct vnode *vp, int flags, enum vgetstate vs) { int error, old; if ((flags & LK_INTERLOCK) != 0) ASSERT_VI_LOCKED(vp, __func__); else ASSERT_VI_UNLOCKED(vp, __func__); VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp); VNPASS(vp->v_holdcnt > 0, vp); VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp); error = vn_lock(vp, flags); if (__predict_false(error != 0)) { if (vs == VGET_USECOUNT) vrele(vp); else vdrop(vp); CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__, vp); return (error); } if (vs == VGET_USECOUNT) return (0); if (__predict_false(vp->v_type == VCHR)) return (vget_finish_vchr(vp)); /* * We hold the vnode. If the usecount is 0 it will be utilized to keep * the vnode around. Otherwise someone else lended their hold count and * we have to drop ours. */ old = atomic_fetchadd_int(&vp->v_usecount, 1); VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); if (old != 0) { #ifdef INVARIANTS old = atomic_fetchadd_int(&vp->v_holdcnt, -1); VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); #else refcount_release(&vp->v_holdcnt); #endif } return (0); } /* * Increase the reference (use) and hold count of a vnode. * This will also remove the vnode from the free list if it is presently free. */ static void __noinline vref_vchr(struct vnode *vp, bool interlock) { /* * See the comment in vget_finish before usecount bump. */ if (!interlock) { if (refcount_acquire_if_not_zero(&vp->v_usecount)) { VNODE_REFCOUNT_FENCE_ACQ(); VNASSERT(vp->v_holdcnt > 0, vp, ("%s: active vnode not held", __func__)); return; } VI_LOCK(vp); /* * By the time we get here the vnode might have been doomed, at * which point the 0->1 use count transition is no longer * protected by the interlock. Since it can't bounce back to * VCHR and requires vref semantics, punt it back */ if (__predict_false(vp->v_type == VBAD)) { VI_UNLOCK(vp); vref(vp); return; } } VNASSERT(vp->v_type == VCHR, vp, ("type != VCHR)")); if (refcount_acquire_if_not_zero(&vp->v_usecount)) { VNODE_REFCOUNT_FENCE_ACQ(); VNASSERT(vp->v_holdcnt > 0, vp, ("%s: active vnode not held", __func__)); if (!interlock) VI_UNLOCK(vp); return; } vhold(vp); v_incr_devcount(vp); refcount_acquire(&vp->v_usecount); if (!interlock) VI_UNLOCK(vp); return; } void vref(struct vnode *vp) { int old; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (__predict_false(vp->v_type == VCHR)) { vref_vchr(vp, false); return; } if (refcount_acquire_if_not_zero(&vp->v_usecount)) { VNODE_REFCOUNT_FENCE_ACQ(); VNASSERT(vp->v_holdcnt > 0, vp, ("%s: active vnode not held", __func__)); return; } vhold(vp); /* * See the comment in vget_finish. */ old = atomic_fetchadd_int(&vp->v_usecount, 1); VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old)); if (old != 0) { #ifdef INVARIANTS old = atomic_fetchadd_int(&vp->v_holdcnt, -1); VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old)); #else refcount_release(&vp->v_holdcnt); #endif } } void vrefl(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (__predict_false(vp->v_type == VCHR)) { vref_vchr(vp, true); return; } vref(vp); } void vrefact(struct vnode *vp) { CTR2(KTR_VFS, "%s: vp %p", __func__, vp); #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_usecount, 1); VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); #else refcount_acquire(&vp->v_usecount); #endif } void vrefactn(struct vnode *vp, u_int n) { CTR2(KTR_VFS, "%s: vp %p", __func__, vp); #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_usecount, n); VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old)); #else atomic_add_int(&vp->v_usecount, n); #endif } /* * Return reference count of a vnode. * * The results of this call are only guaranteed when some mechanism is used to * stop other processes from gaining references to the vnode. This may be the * case if the caller holds the only reference. This is also useful when stale * data is acceptable as race conditions may be accounted for by some other * means. */ int vrefcnt(struct vnode *vp) { return (vp->v_usecount); } void vlazy(struct vnode *vp) { struct mount *mp; VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__)); if ((vp->v_mflag & VMP_LAZYLIST) != 0) return; /* * We may get here for inactive routines after the vnode got doomed. */ if (VN_IS_DOOMED(vp)) return; mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); if ((vp->v_mflag & VMP_LAZYLIST) == 0) { vp->v_mflag |= VMP_LAZYLIST; TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist); mp->mnt_lazyvnodelistsize++; } mtx_unlock(&mp->mnt_listmtx); } /* * This routine is only meant to be called from vgonel prior to dooming * the vnode. */ static void vunlazy_gone(struct vnode *vp) { struct mount *mp; ASSERT_VOP_ELOCKED(vp, __func__); ASSERT_VI_LOCKED(vp, __func__); VNPASS(!VN_IS_DOOMED(vp), vp); if (vp->v_mflag & VMP_LAZYLIST) { mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); vp->v_mflag &= ~VMP_LAZYLIST; TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); mp->mnt_lazyvnodelistsize--; mtx_unlock(&mp->mnt_listmtx); } } static void vdefer_inactive(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode without hold count", __func__)); if (VN_IS_DOOMED(vp)) { vdropl(vp); return; } if (vp->v_iflag & VI_DEFINACT) { VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); vdropl(vp); return; } if (vp->v_usecount > 0) { vp->v_iflag &= ~VI_OWEINACT; vdropl(vp); return; } vlazy(vp); vp->v_iflag |= VI_DEFINACT; VI_UNLOCK(vp); counter_u64_add(deferred_inact, 1); } static void vdefer_inactive_unlocked(struct vnode *vp) { VI_LOCK(vp); if ((vp->v_iflag & VI_OWEINACT) == 0) { vdropl(vp); return; } vdefer_inactive(vp); } enum vput_op { VRELE, VPUT, VUNREF }; /* * Handle ->v_usecount transitioning to 0. * * By releasing the last usecount we take ownership of the hold count which * provides liveness of the vnode, meaning we have to vdrop. * * If the vnode is of type VCHR we may need to decrement si_usecount, see * v_decr_devcount for details. * * For all vnodes we may need to perform inactive processing. It requires an * exclusive lock on the vnode, while it is legal to call here with only a * shared lock (or no locks). If locking the vnode in an expected manner fails, * inactive processing gets deferred to the syncer. * * XXX Some filesystems pass in an exclusively locked vnode and strongly depend * on the lock being held all the way until VOP_INACTIVE. This in particular * happens with UFS which adds half-constructed vnodes to the hash, where they * can be found by other code. */ static void vput_final(struct vnode *vp, enum vput_op func) { int error; bool want_unlock; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); VNPASS(vp->v_holdcnt > 0, vp); VI_LOCK(vp); if (__predict_false(vp->v_type == VCHR && func != VRELE)) v_decr_devcount(vp); /* * By the time we got here someone else might have transitioned * the count back to > 0. */ if (vp->v_usecount > 0) goto out; /* * If the vnode is doomed vgone already performed inactive processing * (if needed). */ if (VN_IS_DOOMED(vp)) goto out; if (__predict_true(VOP_NEED_INACTIVE(vp) == 0)) goto out; if (vp->v_iflag & VI_DOINGINACT) goto out; /* * Locking operations here will drop the interlock and possibly the * vnode lock, opening a window where the vnode can get doomed all the * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to * perform inactive. */ vp->v_iflag |= VI_OWEINACT; want_unlock = false; error = 0; switch (func) { case VRELE: switch (VOP_ISLOCKED(vp)) { case LK_EXCLUSIVE: break; case LK_EXCLOTHER: case 0: want_unlock = true; error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK); VI_LOCK(vp); break; default: /* * The lock has at least one sharer, but we have no way * to conclude whether this is us. Play it safe and * defer processing. */ error = EAGAIN; break; } break; case VPUT: want_unlock = true; if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK | LK_NOWAIT); VI_LOCK(vp); } break; case VUNREF: if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) { error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK); VI_LOCK(vp); } break; } if (error == 0) { vinactive(vp); if (want_unlock) VOP_UNLOCK(vp); vdropl(vp); } else { vdefer_inactive(vp); } return; out: if (func == VPUT) VOP_UNLOCK(vp); vdropl(vp); } /* * Decrement ->v_usecount for a vnode. * * Releasing the last use count requires additional processing, see vput_final * above for details. * * Note that releasing use count without the vnode lock requires special casing * for VCHR, see v_decr_devcount for details. * * Comment above each variant denotes lock state on entry and exit. */ static void __noinline vrele_vchr(struct vnode *vp) { if (refcount_release_if_not_last(&vp->v_usecount)) return; VI_LOCK(vp); if (!refcount_release(&vp->v_usecount)) { VI_UNLOCK(vp); return; } v_decr_devcount(vp); VI_UNLOCK(vp); vput_final(vp, VRELE); } /* * in: any * out: same as passed in */ void vrele(struct vnode *vp) { ASSERT_VI_UNLOCKED(vp, __func__); if (__predict_false(vp->v_type == VCHR)) { vrele_vchr(vp); return; } if (!refcount_release(&vp->v_usecount)) return; vput_final(vp, VRELE); } /* * in: locked * out: unlocked */ void vput(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, __func__); ASSERT_VI_UNLOCKED(vp, __func__); if (!refcount_release(&vp->v_usecount)) { VOP_UNLOCK(vp); return; } vput_final(vp, VPUT); } /* * in: locked * out: locked */ void vunref(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, __func__); ASSERT_VI_UNLOCKED(vp, __func__); if (!refcount_release(&vp->v_usecount)) return; vput_final(vp, VUNREF); } void vhold(struct vnode *vp) { struct vdbatch *vd; int old; CTR2(KTR_VFS, "%s: vp %p", __func__, vp); old = atomic_fetchadd_int(&vp->v_holdcnt, 1); VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, ("%s: wrong hold count %d", __func__, old)); if (old != 0) return; critical_enter(); vd = DPCPU_PTR(vd); vd->freevnodes--; critical_exit(); } void vholdl(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); vhold(vp); } void vholdnz(struct vnode *vp) { CTR2(KTR_VFS, "%s: vp %p", __func__, vp); #ifdef INVARIANTS int old = atomic_fetchadd_int(&vp->v_holdcnt, 1); VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp, ("%s: wrong hold count %d", __func__, old)); #else atomic_add_int(&vp->v_holdcnt, 1); #endif } /* * Grab a hold count unless the vnode is freed. * * Only use this routine if vfs smr is the only protection you have against * freeing the vnode. * * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag * is not set. After the flag is set the vnode becomes immutable to anyone but * the thread which managed to set the flag. * * It may be tempting to replace the loop with: * count = atomic_fetchadd_int(&vp->v_holdcnt, 1); * if (count & VHOLD_NO_SMR) { * backpedal and error out; * } * * However, while this is more performant, it hinders debugging by eliminating * the previously mentioned invariant. */ bool vhold_smr(struct vnode *vp) { int count; VFS_SMR_ASSERT_ENTERED(); count = atomic_load_int(&vp->v_holdcnt); for (;;) { if (count & VHOLD_NO_SMR) { VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp, ("non-zero hold count with flags %d\n", count)); return (false); } VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count)); if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) return (true); } } static void __noinline vdbatch_process(struct vdbatch *vd) { struct vnode *vp; int i; mtx_assert(&vd->lock, MA_OWNED); MPASS(curthread->td_pinned > 0); MPASS(vd->index == VDBATCH_SIZE); mtx_lock(&vnode_list_mtx); critical_enter(); freevnodes += vd->freevnodes; for (i = 0; i < VDBATCH_SIZE; i++) { vp = vd->tab[i]; TAILQ_REMOVE(&vnode_list, vp, v_vnodelist); TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist); MPASS(vp->v_dbatchcpu != NOCPU); vp->v_dbatchcpu = NOCPU; } mtx_unlock(&vnode_list_mtx); vd->freevnodes = 0; bzero(vd->tab, sizeof(vd->tab)); vd->index = 0; critical_exit(); } static void vdbatch_enqueue(struct vnode *vp) { struct vdbatch *vd; ASSERT_VI_LOCKED(vp, __func__); VNASSERT(!VN_IS_DOOMED(vp), vp, ("%s: deferring requeue of a doomed vnode", __func__)); critical_enter(); vd = DPCPU_PTR(vd); vd->freevnodes++; if (vp->v_dbatchcpu != NOCPU) { VI_UNLOCK(vp); critical_exit(); return; } sched_pin(); critical_exit(); mtx_lock(&vd->lock); MPASS(vd->index < VDBATCH_SIZE); MPASS(vd->tab[vd->index] == NULL); /* * A hack: we depend on being pinned so that we know what to put in * ->v_dbatchcpu. */ vp->v_dbatchcpu = curcpu; vd->tab[vd->index] = vp; vd->index++; VI_UNLOCK(vp); if (vd->index == VDBATCH_SIZE) vdbatch_process(vd); mtx_unlock(&vd->lock); sched_unpin(); } /* * This routine must only be called for vnodes which are about to be * deallocated. Supporting dequeue for arbitrary vndoes would require * validating that the locked batch matches. */ static void vdbatch_dequeue(struct vnode *vp) { struct vdbatch *vd; int i; short cpu; VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp, ("%s: called for a used vnode\n", __func__)); cpu = vp->v_dbatchcpu; if (cpu == NOCPU) return; vd = DPCPU_ID_PTR(cpu, vd); mtx_lock(&vd->lock); for (i = 0; i < vd->index; i++) { if (vd->tab[i] != vp) continue; vp->v_dbatchcpu = NOCPU; vd->index--; vd->tab[i] = vd->tab[vd->index]; vd->tab[vd->index] = NULL; break; } mtx_unlock(&vd->lock); /* * Either we dequeued the vnode above or the target CPU beat us to it. */ MPASS(vp->v_dbatchcpu == NOCPU); } /* * Drop the hold count of the vnode. If this is the last reference to * the vnode we place it on the free list unless it has been vgone'd * (marked VIRF_DOOMED) in which case we will free it. * * Because the vnode vm object keeps a hold reference on the vnode if * there is at least one resident non-cached page, the vnode cannot * leave the active list without the page cleanup done. */ static void vdrop_deactivate(struct vnode *vp) { struct mount *mp; ASSERT_VI_LOCKED(vp, __func__); /* * Mark a vnode as free: remove it from its active list * and put it up for recycling on the freelist. */ VNASSERT(!VN_IS_DOOMED(vp), vp, ("vdrop: returning doomed vnode")); VNASSERT(vp->v_op != NULL, vp, ("vdrop: vnode already reclaimed.")); VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp, ("vnode with VI_OWEINACT set")); VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("vnode with VI_DEFINACT set")); if (vp->v_mflag & VMP_LAZYLIST) { mp = vp->v_mount; mtx_lock(&mp->mnt_listmtx); VNASSERT(vp->v_mflag & VMP_LAZYLIST, vp, ("lost VMP_LAZYLIST")); /* * Don't remove the vnode from the lazy list if another thread * has increased the hold count. It may have re-enqueued the * vnode to the lazy list and is now responsible for its * removal. */ if (vp->v_holdcnt == 0) { vp->v_mflag &= ~VMP_LAZYLIST; TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist); mp->mnt_lazyvnodelistsize--; } mtx_unlock(&mp->mnt_listmtx); } vdbatch_enqueue(vp); } void vdrop(struct vnode *vp) { ASSERT_VI_UNLOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (refcount_release_if_not_last(&vp->v_holdcnt)) return; VI_LOCK(vp); vdropl(vp); } void vdropl(struct vnode *vp) { ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if (!refcount_release(&vp->v_holdcnt)) { VI_UNLOCK(vp); return; } if (!VN_IS_DOOMED(vp)) { vdrop_deactivate(vp); return; } /* * We may be racing against vhold_smr. * * If they win we can just pretend we never got this far, they will * vdrop later. */ if (!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR)) { /* * We lost the aforementioned race. Note that any subsequent * access is invalid as they might have managed to vdropl on * their own. */ return; } freevnode(vp); } /* * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT * flags. DOINGINACT prevents us from recursing in calls to vinactive. */ static void vinactivef(struct vnode *vp) { struct vm_object *obj; ASSERT_VOP_ELOCKED(vp, "vinactive"); ASSERT_VI_LOCKED(vp, "vinactive"); VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp, ("vinactive: recursed on VI_DOINGINACT")); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); vp->v_iflag |= VI_DOINGINACT; vp->v_iflag &= ~VI_OWEINACT; VI_UNLOCK(vp); /* * Before moving off the active list, we must be sure that any * modified pages are converted into the vnode's dirty * buffers, since these will no longer be checked once the * vnode is on the inactive list. * * The write-out of the dirty pages is asynchronous. At the * point that VOP_INACTIVE() is called, there could still be * pending I/O and dirty pages in the object. */ if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && vm_object_mightbedirty(obj)) { VM_OBJECT_WLOCK(obj); vm_object_page_clean(obj, 0, 0, 0); VM_OBJECT_WUNLOCK(obj); } VOP_INACTIVE(vp, curthread); VI_LOCK(vp); VNASSERT(vp->v_iflag & VI_DOINGINACT, vp, ("vinactive: lost VI_DOINGINACT")); vp->v_iflag &= ~VI_DOINGINACT; } void vinactive(struct vnode *vp) { ASSERT_VOP_ELOCKED(vp, "vinactive"); ASSERT_VI_LOCKED(vp, "vinactive"); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); if ((vp->v_iflag & VI_OWEINACT) == 0) return; if (vp->v_iflag & VI_DOINGINACT) return; if (vp->v_usecount > 0) { vp->v_iflag &= ~VI_OWEINACT; return; } vinactivef(vp); } /* * Remove any vnodes in the vnode table belonging to mount point mp. * * If FORCECLOSE is not specified, there should not be any active ones, * return error if any are found (nb: this is a user error, not a * system error). If FORCECLOSE is specified, detach any active vnodes * that are found. * * If WRITECLOSE is set, only flush out regular file vnodes open for * writing. * * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped. * * `rootrefs' specifies the base reference count for the root vnode * of this filesystem. The root vnode is considered busy if its * v_usecount exceeds this value. On a successful return, vflush(, td) * will call vrele() on the root vnode exactly rootrefs times. * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must * be zero. */ #ifdef DIAGNOSTIC static int busyprt = 0; /* print out busy vnodes */ SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes"); #endif int vflush(struct mount *mp, int rootrefs, int flags, struct thread *td) { struct vnode *vp, *mvp, *rootvp = NULL; struct vattr vattr; int busy = 0, error; CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp, rootrefs, flags); if (rootrefs > 0) { KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0, ("vflush: bad args")); /* * Get the filesystem root vnode. We can vput() it * immediately, since with rootrefs > 0, it won't go away. */ if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) { CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d", __func__, error); return (error); } vput(rootvp); } loop: MNT_VNODE_FOREACH_ALL(vp, mp, mvp) { vholdl(vp); error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE); if (error) { vdrop(vp); MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); goto loop; } /* * Skip over a vnodes marked VV_SYSTEM. */ if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) { VOP_UNLOCK(vp); vdrop(vp); continue; } /* * If WRITECLOSE is set, flush out unlinked but still open * files (even if open only for reading) and regular file * vnodes open for writing. */ if (flags & WRITECLOSE) { if (vp->v_object != NULL) { VM_OBJECT_WLOCK(vp->v_object); vm_object_page_clean(vp->v_object, 0, 0, 0); VM_OBJECT_WUNLOCK(vp->v_object); } error = VOP_FSYNC(vp, MNT_WAIT, td); if (error != 0) { VOP_UNLOCK(vp); vdrop(vp); MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp); return (error); } error = VOP_GETATTR(vp, &vattr, td->td_ucred); VI_LOCK(vp); if ((vp->v_type == VNON || (error == 0 && vattr.va_nlink > 0)) && (vp->v_writecount <= 0 || vp->v_type != VREG)) { VOP_UNLOCK(vp); vdropl(vp); continue; } } else VI_LOCK(vp); /* * With v_usecount == 0, all we need to do is clear out the * vnode data structures and we are done. * * If FORCECLOSE is set, forcibly close the vnode. */ if (vp->v_usecount == 0 || (flags & FORCECLOSE)) { vgonel(vp); } else { busy++; #ifdef DIAGNOSTIC if (busyprt) vn_printf(vp, "vflush: busy vnode "); #endif } VOP_UNLOCK(vp); vdropl(vp); } if (rootrefs > 0 && (flags & FORCECLOSE) == 0) { /* * If just the root vnode is busy, and if its refcount * is equal to `rootrefs', then go ahead and kill it. */ VI_LOCK(rootvp); KASSERT(busy > 0, ("vflush: not busy")); VNASSERT(rootvp->v_usecount >= rootrefs, rootvp, ("vflush: usecount %d < rootrefs %d", rootvp->v_usecount, rootrefs)); if (busy == 1 && rootvp->v_usecount == rootrefs) { VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK); vgone(rootvp); VOP_UNLOCK(rootvp); busy = 0; } else VI_UNLOCK(rootvp); } if (busy) { CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__, busy); return (EBUSY); } for (; rootrefs > 0; rootrefs--) vrele(rootvp); return (0); } /* * Recycle an unused vnode to the front of the free list. */ int vrecycle(struct vnode *vp) { int recycled; VI_LOCK(vp); recycled = vrecyclel(vp); VI_UNLOCK(vp); return (recycled); } /* * vrecycle, with the vp interlock held. */ int vrecyclel(struct vnode *vp) { int recycled; ASSERT_VOP_ELOCKED(vp, __func__); ASSERT_VI_LOCKED(vp, __func__); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); recycled = 0; if (vp->v_usecount == 0) { recycled = 1; vgonel(vp); } return (recycled); } /* * Eliminate all activity associated with a vnode * in preparation for reuse. */ void vgone(struct vnode *vp) { VI_LOCK(vp); vgonel(vp); VI_UNLOCK(vp); } static void notify_lowervp_vfs_dummy(struct mount *mp __unused, struct vnode *lowervp __unused) { } /* * Notify upper mounts about reclaimed or unlinked vnode. */ void vfs_notify_upper(struct vnode *vp, int event) { static struct vfsops vgonel_vfsops = { .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy, .vfs_unlink_lowervp = notify_lowervp_vfs_dummy, }; struct mount *mp, *ump, *mmp; mp = vp->v_mount; if (mp == NULL) return; if (TAILQ_EMPTY(&mp->mnt_uppers)) return; mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO); mmp->mnt_op = &vgonel_vfsops; mmp->mnt_kern_flag |= MNTK_MARKER; MNT_ILOCK(mp); mp->mnt_kern_flag |= MNTK_VGONE_UPPER; for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) { if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) { ump = TAILQ_NEXT(ump, mnt_upper_link); continue; } TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link); MNT_IUNLOCK(mp); switch (event) { case VFS_NOTIFY_UPPER_RECLAIM: VFS_RECLAIM_LOWERVP(ump, vp); break; case VFS_NOTIFY_UPPER_UNLINK: VFS_UNLINK_LOWERVP(ump, vp); break; default: KASSERT(0, ("invalid event %d", event)); break; } MNT_ILOCK(mp); ump = TAILQ_NEXT(mmp, mnt_upper_link); TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link); } free(mmp, M_TEMP); mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER; if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) { mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER; wakeup(&mp->mnt_uppers); } MNT_IUNLOCK(mp); } /* * vgone, with the vp interlock held. */ static void vgonel(struct vnode *vp) { struct thread *td; struct mount *mp; vm_object_t object; bool active, oweinact; ASSERT_VOP_ELOCKED(vp, "vgonel"); ASSERT_VI_LOCKED(vp, "vgonel"); VNASSERT(vp->v_holdcnt, vp, ("vgonel: vp %p has no reference.", vp)); CTR2(KTR_VFS, "%s: vp %p", __func__, vp); td = curthread; /* * Don't vgonel if we're already doomed. */ if (vp->v_irflag & VIRF_DOOMED) return; vunlazy_gone(vp); vp->v_irflag |= VIRF_DOOMED; /* * Check to see if the vnode is in use. If so, we have to call * VOP_CLOSE() and VOP_INACTIVE(). */ active = vp->v_usecount > 0; oweinact = (vp->v_iflag & VI_OWEINACT) != 0; /* * If we need to do inactive VI_OWEINACT will be set. */ if (vp->v_iflag & VI_DEFINACT) { VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count")); vp->v_iflag &= ~VI_DEFINACT; vdropl(vp); } else { VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count")); VI_UNLOCK(vp); } vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM); /* * If purging an active vnode, it must be closed and * deactivated before being reclaimed. */ if (active) VOP_CLOSE(vp, FNONBLOCK, NOCRED, td); if (oweinact || active) { VI_LOCK(vp); vinactivef(vp); VI_UNLOCK(vp); } if (vp->v_type == VSOCK) vfs_unp_reclaim(vp); /* * Clean out any buffers associated with the vnode. * If the flush fails, just toss the buffers. */ mp = NULL; if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd)) (void) vn_start_secondary_write(vp, &mp, V_WAIT); if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) { while (vinvalbuf(vp, 0, 0, 0) != 0) ; } BO_LOCK(&vp->v_bufobj); KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) && vp->v_bufobj.bo_dirty.bv_cnt == 0 && TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) && vp->v_bufobj.bo_clean.bv_cnt == 0, ("vp %p bufobj not invalidated", vp)); /* * For VMIO bufobj, BO_DEAD is set later, or in * vm_object_terminate() after the object's page queue is * flushed. */ object = vp->v_bufobj.bo_object; if (object == NULL) vp->v_bufobj.bo_flag |= BO_DEAD; BO_UNLOCK(&vp->v_bufobj); /* * Handle the VM part. Tmpfs handles v_object on its own (the * OBJT_VNODE check). Nullfs or other bypassing filesystems * should not touch the object borrowed from the lower vnode * (the handle check). */ if (object != NULL && object->type == OBJT_VNODE && object->handle == vp) vnode_destroy_vobject(vp); /* * Reclaim the vnode. */ if (VOP_RECLAIM(vp, td)) panic("vgone: cannot reclaim"); if (mp != NULL) vn_finished_secondary_write(mp); VNASSERT(vp->v_object == NULL, vp, ("vop_reclaim left v_object vp=%p", vp)); /* * Clear the advisory locks and wake up waiting threads. */ (void)VOP_ADVLOCKPURGE(vp); vp->v_lockf = NULL; /* * Delete from old mount point vnode list. */ delmntque(vp); cache_purge(vp); /* * Done with purge, reset to the standard lock and invalidate * the vnode. */ VI_LOCK(vp); vp->v_vnlock = &vp->v_lock; vp->v_op = &dead_vnodeops; vp->v_type = VBAD; } /* * Calculate the total number of references to a special device. */ int vcount(struct vnode *vp) { int count; dev_lock(); count = vp->v_rdev->si_usecount; dev_unlock(); return (count); } /* * Print out a description of a vnode. */ static const char * const typename[] = {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD", "VMARKER"}; _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0, "new hold count flag not added to vn_printf"); void vn_printf(struct vnode *vp, const char *fmt, ...) { va_list ap; char buf[256], buf2[16]; u_long flags; u_int holdcnt; va_start(ap, fmt); vprintf(fmt, ap); va_end(ap); printf("%p: ", (void *)vp); printf("type %s\n", typename[vp->v_type]); holdcnt = atomic_load_int(&vp->v_holdcnt); printf(" usecount %d, writecount %d, refcount %d", vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS); switch (vp->v_type) { case VDIR: printf(" mountedhere %p\n", vp->v_mountedhere); break; case VCHR: printf(" rdev %p\n", vp->v_rdev); break; case VSOCK: printf(" socket %p\n", vp->v_unpcb); break; case VFIFO: printf(" fifoinfo %p\n", vp->v_fifoinfo); break; default: printf("\n"); break; } buf[0] = '\0'; buf[1] = '\0'; if (holdcnt & VHOLD_NO_SMR) strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf)); printf(" hold count flags (%s)\n", buf + 1); buf[0] = '\0'; buf[1] = '\0'; if (vp->v_irflag & VIRF_DOOMED) strlcat(buf, "|VIRF_DOOMED", sizeof(buf)); flags = vp->v_irflag & ~(VIRF_DOOMED); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } if (vp->v_vflag & VV_ROOT) strlcat(buf, "|VV_ROOT", sizeof(buf)); if (vp->v_vflag & VV_ISTTY) strlcat(buf, "|VV_ISTTY", sizeof(buf)); if (vp->v_vflag & VV_NOSYNC) strlcat(buf, "|VV_NOSYNC", sizeof(buf)); if (vp->v_vflag & VV_ETERNALDEV) strlcat(buf, "|VV_ETERNALDEV", sizeof(buf)); if (vp->v_vflag & VV_CACHEDLABEL) strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf)); if (vp->v_vflag & VV_VMSIZEVNLOCK) strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf)); if (vp->v_vflag & VV_COPYONWRITE) strlcat(buf, "|VV_COPYONWRITE", sizeof(buf)); if (vp->v_vflag & VV_SYSTEM) strlcat(buf, "|VV_SYSTEM", sizeof(buf)); if (vp->v_vflag & VV_PROCDEP) strlcat(buf, "|VV_PROCDEP", sizeof(buf)); if (vp->v_vflag & VV_NOKNOTE) strlcat(buf, "|VV_NOKNOTE", sizeof(buf)); if (vp->v_vflag & VV_DELETED) strlcat(buf, "|VV_DELETED", sizeof(buf)); if (vp->v_vflag & VV_MD) strlcat(buf, "|VV_MD", sizeof(buf)); if (vp->v_vflag & VV_FORCEINSMQ) strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf)); if (vp->v_vflag & VV_READLINK) strlcat(buf, "|VV_READLINK", sizeof(buf)); flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV | VV_CACHEDLABEL | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP | VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } if (vp->v_iflag & VI_TEXT_REF) strlcat(buf, "|VI_TEXT_REF", sizeof(buf)); if (vp->v_iflag & VI_MOUNT) strlcat(buf, "|VI_MOUNT", sizeof(buf)); if (vp->v_iflag & VI_DOINGINACT) strlcat(buf, "|VI_DOINGINACT", sizeof(buf)); if (vp->v_iflag & VI_OWEINACT) strlcat(buf, "|VI_OWEINACT", sizeof(buf)); if (vp->v_iflag & VI_DEFINACT) strlcat(buf, "|VI_DEFINACT", sizeof(buf)); flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_DOINGINACT | VI_OWEINACT | VI_DEFINACT); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } if (vp->v_mflag & VMP_LAZYLIST) strlcat(buf, "|VMP_LAZYLIST", sizeof(buf)); flags = vp->v_mflag & ~(VMP_LAZYLIST); if (flags != 0) { snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags); strlcat(buf, buf2, sizeof(buf)); } printf(" flags (%s)\n", buf + 1); if (mtx_owned(VI_MTX(vp))) printf(" VI_LOCKed"); if (vp->v_object != NULL) printf(" v_object %p ref %d pages %d " "cleanbuf %d dirtybuf %d\n", vp->v_object, vp->v_object->ref_count, vp->v_object->resident_page_count, vp->v_bufobj.bo_clean.bv_cnt, vp->v_bufobj.bo_dirty.bv_cnt); printf(" "); lockmgr_printinfo(vp->v_vnlock); if (vp->v_data != NULL) VOP_PRINT(vp); } #ifdef DDB /* * List all of the locked vnodes in the system. * Called when debugging the kernel. */ DB_SHOW_COMMAND(lockedvnods, lockedvnodes) { struct mount *mp; struct vnode *vp; /* * Note: because this is DDB, we can't obey the locking semantics * for these structures, which means we could catch an inconsistent * state and dereference a nasty pointer. Not much to be done * about that. */ db_printf("Locked vnodes\n"); TAILQ_FOREACH(mp, &mountlist, mnt_list) { TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && VOP_ISLOCKED(vp)) vn_printf(vp, "vnode "); } } } /* * Show details about the given vnode. */ DB_SHOW_COMMAND(vnode, db_show_vnode) { struct vnode *vp; if (!have_addr) return; vp = (struct vnode *)addr; vn_printf(vp, "vnode "); } /* * Show details about the given mount point. */ DB_SHOW_COMMAND(mount, db_show_mount) { struct mount *mp; struct vfsopt *opt; struct statfs *sp; struct vnode *vp; char buf[512]; uint64_t mflags; u_int flags; if (!have_addr) { /* No address given, print short info about all mount points. */ TAILQ_FOREACH(mp, &mountlist, mnt_list) { db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); if (db_pager_quit) break; } db_printf("\nMore info: show mount \n"); return; } mp = (struct mount *)addr; db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname, mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename); buf[0] = '\0'; mflags = mp->mnt_flag; #define MNT_FLAG(flag) do { \ if (mflags & (flag)) { \ if (buf[0] != '\0') \ strlcat(buf, ", ", sizeof(buf)); \ strlcat(buf, (#flag) + 4, sizeof(buf)); \ mflags &= ~(flag); \ } \ } while (0) MNT_FLAG(MNT_RDONLY); MNT_FLAG(MNT_SYNCHRONOUS); MNT_FLAG(MNT_NOEXEC); MNT_FLAG(MNT_NOSUID); MNT_FLAG(MNT_NFS4ACLS); MNT_FLAG(MNT_UNION); MNT_FLAG(MNT_ASYNC); MNT_FLAG(MNT_SUIDDIR); MNT_FLAG(MNT_SOFTDEP); MNT_FLAG(MNT_NOSYMFOLLOW); MNT_FLAG(MNT_GJOURNAL); MNT_FLAG(MNT_MULTILABEL); MNT_FLAG(MNT_ACLS); MNT_FLAG(MNT_NOATIME); MNT_FLAG(MNT_NOCLUSTERR); MNT_FLAG(MNT_NOCLUSTERW); MNT_FLAG(MNT_SUJ); MNT_FLAG(MNT_EXRDONLY); MNT_FLAG(MNT_EXPORTED); MNT_FLAG(MNT_DEFEXPORTED); MNT_FLAG(MNT_EXPORTANON); MNT_FLAG(MNT_EXKERB); MNT_FLAG(MNT_EXPUBLIC); MNT_FLAG(MNT_LOCAL); MNT_FLAG(MNT_QUOTA); MNT_FLAG(MNT_ROOTFS); MNT_FLAG(MNT_USER); MNT_FLAG(MNT_IGNORE); MNT_FLAG(MNT_UPDATE); MNT_FLAG(MNT_DELEXPORT); MNT_FLAG(MNT_RELOAD); MNT_FLAG(MNT_FORCE); MNT_FLAG(MNT_SNAPSHOT); MNT_FLAG(MNT_BYFSID); #undef MNT_FLAG if (mflags != 0) { if (buf[0] != '\0') strlcat(buf, ", ", sizeof(buf)); snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "0x%016jx", mflags); } db_printf(" mnt_flag = %s\n", buf); buf[0] = '\0'; flags = mp->mnt_kern_flag; #define MNT_KERN_FLAG(flag) do { \ if (flags & (flag)) { \ if (buf[0] != '\0') \ strlcat(buf, ", ", sizeof(buf)); \ strlcat(buf, (#flag) + 5, sizeof(buf)); \ flags &= ~(flag); \ } \ } while (0) MNT_KERN_FLAG(MNTK_UNMOUNTF); MNT_KERN_FLAG(MNTK_ASYNC); MNT_KERN_FLAG(MNTK_SOFTDEP); MNT_KERN_FLAG(MNTK_DRAINING); MNT_KERN_FLAG(MNTK_REFEXPIRE); MNT_KERN_FLAG(MNTK_EXTENDED_SHARED); MNT_KERN_FLAG(MNTK_SHARED_WRITES); MNT_KERN_FLAG(MNTK_NO_IOPF); MNT_KERN_FLAG(MNTK_VGONE_UPPER); MNT_KERN_FLAG(MNTK_VGONE_WAITER); MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT); MNT_KERN_FLAG(MNTK_MARKER); MNT_KERN_FLAG(MNTK_USES_BCACHE); MNT_KERN_FLAG(MNTK_NOASYNC); MNT_KERN_FLAG(MNTK_UNMOUNT); MNT_KERN_FLAG(MNTK_MWAIT); MNT_KERN_FLAG(MNTK_SUSPEND); MNT_KERN_FLAG(MNTK_SUSPEND2); MNT_KERN_FLAG(MNTK_SUSPENDED); MNT_KERN_FLAG(MNTK_LOOKUP_SHARED); MNT_KERN_FLAG(MNTK_NOKNOTE); #undef MNT_KERN_FLAG if (flags != 0) { if (buf[0] != '\0') strlcat(buf, ", ", sizeof(buf)); snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "0x%08x", flags); } db_printf(" mnt_kern_flag = %s\n", buf); db_printf(" mnt_opt = "); opt = TAILQ_FIRST(mp->mnt_opt); if (opt != NULL) { db_printf("%s", opt->name); opt = TAILQ_NEXT(opt, link); while (opt != NULL) { db_printf(", %s", opt->name); opt = TAILQ_NEXT(opt, link); } } db_printf("\n"); sp = &mp->mnt_stat; db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx " "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju " "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju " "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n", (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags, (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize, (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree, (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files, (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites, (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads, (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax, (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]); db_printf(" mnt_cred = { uid=%u ruid=%u", (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid); if (jailed(mp->mnt_cred)) db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id); db_printf(" }\n"); db_printf(" mnt_ref = %d (with %d in the struct)\n", vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref); db_printf(" mnt_gen = %d\n", mp->mnt_gen); db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize); db_printf(" mnt_lazyvnodelistsize = %d\n", mp->mnt_lazyvnodelistsize); db_printf(" mnt_writeopcount = %d (with %d in the struct)\n", vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount); db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen); db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max); db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed); db_printf(" mnt_lockref = %d (with %d in the struct)\n", vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref); db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes); db_printf(" mnt_secondary_accwrites = %d\n", mp->mnt_secondary_accwrites); db_printf(" mnt_gjprovider = %s\n", mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL"); db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops); db_printf("\n\nList of active vnodes\n"); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && vp->v_holdcnt > 0) { vn_printf(vp, "vnode "); if (db_pager_quit) break; } } db_printf("\n\nList of inactive vnodes\n"); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (vp->v_type != VMARKER && vp->v_holdcnt == 0) { vn_printf(vp, "vnode "); if (db_pager_quit) break; } } } #endif /* DDB */ /* * Fill in a struct xvfsconf based on a struct vfsconf. */ static int vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp) { struct xvfsconf xvfsp; bzero(&xvfsp, sizeof(xvfsp)); strcpy(xvfsp.vfc_name, vfsp->vfc_name); xvfsp.vfc_typenum = vfsp->vfc_typenum; xvfsp.vfc_refcount = vfsp->vfc_refcount; xvfsp.vfc_flags = vfsp->vfc_flags; /* * These are unused in userland, we keep them * to not break binary compatibility. */ xvfsp.vfc_vfsops = NULL; xvfsp.vfc_next = NULL; return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); } #ifdef COMPAT_FREEBSD32 struct xvfsconf32 { uint32_t vfc_vfsops; char vfc_name[MFSNAMELEN]; int32_t vfc_typenum; int32_t vfc_refcount; int32_t vfc_flags; uint32_t vfc_next; }; static int vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp) { struct xvfsconf32 xvfsp; bzero(&xvfsp, sizeof(xvfsp)); strcpy(xvfsp.vfc_name, vfsp->vfc_name); xvfsp.vfc_typenum = vfsp->vfc_typenum; xvfsp.vfc_refcount = vfsp->vfc_refcount; xvfsp.vfc_flags = vfsp->vfc_flags; return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp))); } #endif /* * Top level filesystem related information gathering. */ static int sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS) { struct vfsconf *vfsp; int error; error = 0; vfsconf_slock(); TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { #ifdef COMPAT_FREEBSD32 if (req->flags & SCTL_MASK32) error = vfsconf2x32(req, vfsp); else #endif error = vfsconf2x(req, vfsp); if (error) break; } vfsconf_sunlock(); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist, "S,xvfsconf", "List of all configured filesystems"); #ifndef BURN_BRIDGES static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS); static int vfs_sysctl(SYSCTL_HANDLER_ARGS) { int *name = (int *)arg1 - 1; /* XXX */ u_int namelen = arg2 + 1; /* XXX */ struct vfsconf *vfsp; log(LOG_WARNING, "userland calling deprecated sysctl, " "please rebuild world\n"); #if 1 || defined(COMPAT_PRELITE2) /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */ if (namelen == 1) return (sysctl_ovfs_conf(oidp, arg1, arg2, req)); #endif switch (name[1]) { case VFS_MAXTYPENUM: if (namelen != 2) return (ENOTDIR); return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int))); case VFS_CONF: if (namelen != 3) return (ENOTDIR); /* overloaded */ vfsconf_slock(); TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { if (vfsp->vfc_typenum == name[2]) break; } vfsconf_sunlock(); if (vfsp == NULL) return (EOPNOTSUPP); #ifdef COMPAT_FREEBSD32 if (req->flags & SCTL_MASK32) return (vfsconf2x32(req, vfsp)); else #endif return (vfsconf2x(req, vfsp)); } return (EOPNOTSUPP); } static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP | CTLFLAG_MPSAFE, vfs_sysctl, "Generic filesystem"); #if 1 || defined(COMPAT_PRELITE2) static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS) { int error; struct vfsconf *vfsp; struct ovfsconf ovfs; vfsconf_slock(); TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) { bzero(&ovfs, sizeof(ovfs)); ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */ strcpy(ovfs.vfc_name, vfsp->vfc_name); ovfs.vfc_index = vfsp->vfc_typenum; ovfs.vfc_refcount = vfsp->vfc_refcount; ovfs.vfc_flags = vfsp->vfc_flags; error = SYSCTL_OUT(req, &ovfs, sizeof ovfs); if (error != 0) { vfsconf_sunlock(); return (error); } } vfsconf_sunlock(); return (0); } #endif /* 1 || COMPAT_PRELITE2 */ #endif /* !BURN_BRIDGES */ #define KINFO_VNODESLOP 10 #ifdef notyet /* * Dump vnode list (via sysctl). */ /* ARGSUSED */ static int sysctl_vnode(SYSCTL_HANDLER_ARGS) { struct xvnode *xvn; struct mount *mp; struct vnode *vp; int error, len, n; /* * Stale numvnodes access is not fatal here. */ req->lock = 0; len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn; if (!req->oldptr) /* Make an estimate */ return (SYSCTL_OUT(req, 0, len)); error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK); n = 0; mtx_lock(&mountlist_mtx); TAILQ_FOREACH(mp, &mountlist, mnt_list) { if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) continue; MNT_ILOCK(mp); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { if (n == len) break; vref(vp); xvn[n].xv_size = sizeof *xvn; xvn[n].xv_vnode = vp; xvn[n].xv_id = 0; /* XXX compat */ #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field XV_COPY(usecount); XV_COPY(writecount); XV_COPY(holdcnt); XV_COPY(mount); XV_COPY(numoutput); XV_COPY(type); #undef XV_COPY xvn[n].xv_flag = vp->v_vflag; switch (vp->v_type) { case VREG: case VDIR: case VLNK: break; case VBLK: case VCHR: if (vp->v_rdev == NULL) { vrele(vp); continue; } xvn[n].xv_dev = dev2udev(vp->v_rdev); break; case VSOCK: xvn[n].xv_socket = vp->v_socket; break; case VFIFO: xvn[n].xv_fifo = vp->v_fifoinfo; break; case VNON: case VBAD: default: /* shouldn't happen? */ vrele(vp); continue; } vrele(vp); ++n; } MNT_IUNLOCK(mp); mtx_lock(&mountlist_mtx); vfs_unbusy(mp); if (n == len) break; } mtx_unlock(&mountlist_mtx); error = SYSCTL_OUT(req, xvn, n * sizeof *xvn); free(xvn, M_TEMP); return (error); } SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode", ""); #endif static void unmount_or_warn(struct mount *mp) { int error; error = dounmount(mp, MNT_FORCE, curthread); if (error != 0) { printf("unmount of %s failed (", mp->mnt_stat.f_mntonname); if (error == EBUSY) printf("BUSY)\n"); else printf("%d)\n", error); } } /* * Unmount all filesystems. The list is traversed in reverse order * of mounting to avoid dependencies. */ void vfs_unmountall(void) { struct mount *mp, *tmp; CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__); /* * Since this only runs when rebooting, it is not interlocked. */ TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) { vfs_ref(mp); /* * Forcibly unmounting "/dev" before "/" would prevent clean * unmount of the latter. */ if (mp == rootdevmp) continue; unmount_or_warn(mp); } if (rootdevmp != NULL) unmount_or_warn(rootdevmp); } static void vfs_deferred_inactive(struct vnode *vp, int lkflags) { ASSERT_VI_LOCKED(vp, __func__); VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set")); if ((vp->v_iflag & VI_OWEINACT) == 0) { vdropl(vp); return; } if (vn_lock(vp, lkflags) == 0) { VI_LOCK(vp); vinactive(vp); VOP_UNLOCK(vp); vdropl(vp); return; } vdefer_inactive_unlocked(vp); } static int vfs_periodic_inactive_filter(struct vnode *vp, void *arg) { return (vp->v_iflag & VI_DEFINACT); } static void __noinline vfs_periodic_inactive(struct mount *mp, int flags) { struct vnode *vp, *mvp; int lkflags; lkflags = LK_EXCLUSIVE | LK_INTERLOCK; if (flags != MNT_WAIT) lkflags |= LK_NOWAIT; MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) { if ((vp->v_iflag & VI_DEFINACT) == 0) { VI_UNLOCK(vp); continue; } vp->v_iflag &= ~VI_DEFINACT; vfs_deferred_inactive(vp, lkflags); } } static inline bool vfs_want_msync(struct vnode *vp) { struct vm_object *obj; /* * This test may be performed without any locks held. * We rely on vm_object's type stability. */ if (vp->v_vflag & VV_NOSYNC) return (false); obj = vp->v_object; return (obj != NULL && vm_object_mightbedirty(obj)); } static int vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused) { if (vp->v_vflag & VV_NOSYNC) return (false); if (vp->v_iflag & VI_DEFINACT) return (true); return (vfs_want_msync(vp)); } static void __noinline vfs_periodic_msync_inactive(struct mount *mp, int flags) { struct vnode *vp, *mvp; struct vm_object *obj; struct thread *td; int lkflags, objflags; bool seen_defer; td = curthread; lkflags = LK_EXCLUSIVE | LK_INTERLOCK; if (flags != MNT_WAIT) { lkflags |= LK_NOWAIT; objflags = OBJPC_NOSYNC; } else { objflags = OBJPC_SYNC; } MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) { seen_defer = false; if (vp->v_iflag & VI_DEFINACT) { vp->v_iflag &= ~VI_DEFINACT; seen_defer = true; } if (!vfs_want_msync(vp)) { if (seen_defer) vfs_deferred_inactive(vp, lkflags); else VI_UNLOCK(vp); continue; } if (vget(vp, lkflags, td) == 0) { obj = vp->v_object; if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) { VM_OBJECT_WLOCK(obj); vm_object_page_clean(obj, 0, 0, objflags); VM_OBJECT_WUNLOCK(obj); } vput(vp); if (seen_defer) vdrop(vp); } else { if (seen_defer) vdefer_inactive_unlocked(vp); } } } void vfs_periodic(struct mount *mp, int flags) { CTR2(KTR_VFS, "%s: mp %p", __func__, mp); if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0) vfs_periodic_inactive(mp, flags); else vfs_periodic_msync_inactive(mp, flags); } static void destroy_vpollinfo_free(struct vpollinfo *vi) { knlist_destroy(&vi->vpi_selinfo.si_note); mtx_destroy(&vi->vpi_lock); uma_zfree(vnodepoll_zone, vi); } static void destroy_vpollinfo(struct vpollinfo *vi) { knlist_clear(&vi->vpi_selinfo.si_note, 1); seldrain(&vi->vpi_selinfo); destroy_vpollinfo_free(vi); } /* * Initialize per-vnode helper structure to hold poll-related state. */ void v_addpollinfo(struct vnode *vp) { struct vpollinfo *vi; if (vp->v_pollinfo != NULL) return; vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO); mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF); knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock, vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked); VI_LOCK(vp); if (vp->v_pollinfo != NULL) { VI_UNLOCK(vp); destroy_vpollinfo_free(vi); return; } vp->v_pollinfo = vi; VI_UNLOCK(vp); } /* * Record a process's interest in events which might happen to * a vnode. Because poll uses the historic select-style interface * internally, this routine serves as both the ``check for any * pending events'' and the ``record my interest in future events'' * functions. (These are done together, while the lock is held, * to avoid race conditions.) */ int vn_pollrecord(struct vnode *vp, struct thread *td, int events) { v_addpollinfo(vp); mtx_lock(&vp->v_pollinfo->vpi_lock); if (vp->v_pollinfo->vpi_revents & events) { /* * This leaves events we are not interested * in available for the other process which * which presumably had requested them * (otherwise they would never have been * recorded). */ events &= vp->v_pollinfo->vpi_revents; vp->v_pollinfo->vpi_revents &= ~events; mtx_unlock(&vp->v_pollinfo->vpi_lock); return (events); } vp->v_pollinfo->vpi_events |= events; selrecord(td, &vp->v_pollinfo->vpi_selinfo); mtx_unlock(&vp->v_pollinfo->vpi_lock); return (0); } /* * Routine to create and manage a filesystem syncer vnode. */ #define sync_close ((int (*)(struct vop_close_args *))nullop) static int sync_fsync(struct vop_fsync_args *); static int sync_inactive(struct vop_inactive_args *); static int sync_reclaim(struct vop_reclaim_args *); static struct vop_vector sync_vnodeops = { .vop_bypass = VOP_EOPNOTSUPP, .vop_close = sync_close, /* close */ .vop_fsync = sync_fsync, /* fsync */ .vop_inactive = sync_inactive, /* inactive */ .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */ .vop_reclaim = sync_reclaim, /* reclaim */ .vop_lock1 = vop_stdlock, /* lock */ .vop_unlock = vop_stdunlock, /* unlock */ .vop_islocked = vop_stdislocked, /* islocked */ }; VFS_VOP_VECTOR_REGISTER(sync_vnodeops); /* * Create a new filesystem syncer vnode for the specified mount point. */ void vfs_allocate_syncvnode(struct mount *mp) { struct vnode *vp; struct bufobj *bo; static long start, incr, next; int error; /* Allocate a new vnode */ error = getnewvnode("syncer", mp, &sync_vnodeops, &vp); if (error != 0) panic("vfs_allocate_syncvnode: getnewvnode() failed"); vp->v_type = VNON; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); vp->v_vflag |= VV_FORCEINSMQ; error = insmntque(vp, mp); if (error != 0) panic("vfs_allocate_syncvnode: insmntque() failed"); vp->v_vflag &= ~VV_FORCEINSMQ; VOP_UNLOCK(vp); /* * Place the vnode onto the syncer worklist. We attempt to * scatter them about on the list so that they will go off * at evenly distributed times even if all the filesystems * are mounted at once. */ next += incr; if (next == 0 || next > syncer_maxdelay) { start /= 2; incr /= 2; if (start == 0) { start = syncer_maxdelay / 2; incr = syncer_maxdelay; } next = start; } bo = &vp->v_bufobj; BO_LOCK(bo); vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0); /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */ mtx_lock(&sync_mtx); sync_vnode_count++; if (mp->mnt_syncer == NULL) { mp->mnt_syncer = vp; vp = NULL; } mtx_unlock(&sync_mtx); BO_UNLOCK(bo); if (vp != NULL) { vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); vgone(vp); vput(vp); } } void vfs_deallocate_syncvnode(struct mount *mp) { struct vnode *vp; mtx_lock(&sync_mtx); vp = mp->mnt_syncer; if (vp != NULL) mp->mnt_syncer = NULL; mtx_unlock(&sync_mtx); if (vp != NULL) vrele(vp); } /* * Do a lazy sync of the filesystem. */ static int sync_fsync(struct vop_fsync_args *ap) { struct vnode *syncvp = ap->a_vp; struct mount *mp = syncvp->v_mount; int error, save; struct bufobj *bo; /* * We only need to do something if this is a lazy evaluation. */ if (ap->a_waitfor != MNT_LAZY) return (0); /* * Move ourselves to the back of the sync list. */ bo = &syncvp->v_bufobj; BO_LOCK(bo); vn_syncer_add_to_worklist(bo, syncdelay); BO_UNLOCK(bo); /* * Walk the list of vnodes pushing all that are dirty and * not already on the sync list. */ if (vfs_busy(mp, MBF_NOWAIT) != 0) return (0); if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) { vfs_unbusy(mp); return (0); } save = curthread_pflags_set(TDP_SYNCIO); /* * The filesystem at hand may be idle with free vnodes stored in the * batch. Return them instead of letting them stay there indefinitely. */ vfs_periodic(mp, MNT_NOWAIT); error = VFS_SYNC(mp, MNT_LAZY); curthread_pflags_restore(save); vn_finished_write(mp); vfs_unbusy(mp); return (error); } /* * The syncer vnode is no referenced. */ static int sync_inactive(struct vop_inactive_args *ap) { vgone(ap->a_vp); return (0); } /* * The syncer vnode is no longer needed and is being decommissioned. * * Modifications to the worklist must be protected by sync_mtx. */ static int sync_reclaim(struct vop_reclaim_args *ap) { struct vnode *vp = ap->a_vp; struct bufobj *bo; bo = &vp->v_bufobj; BO_LOCK(bo); mtx_lock(&sync_mtx); if (vp->v_mount->mnt_syncer == vp) vp->v_mount->mnt_syncer = NULL; if (bo->bo_flag & BO_ONWORKLST) { LIST_REMOVE(bo, bo_synclist); syncer_worklist_len--; sync_vnode_count--; bo->bo_flag &= ~BO_ONWORKLST; } mtx_unlock(&sync_mtx); BO_UNLOCK(bo); return (0); } int vn_need_pageq_flush(struct vnode *vp) { struct vm_object *obj; int need; MPASS(mtx_owned(VI_MTX(vp))); need = 0; if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 && vm_object_mightbedirty(obj)) need = 1; return (need); } /* * Check if vnode represents a disk device */ int vn_isdisk(struct vnode *vp, int *errp) { int error; if (vp->v_type != VCHR) { error = ENOTBLK; goto out; } error = 0; dev_lock(); if (vp->v_rdev == NULL) error = ENXIO; else if (vp->v_rdev->si_devsw == NULL) error = ENXIO; else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK)) error = ENOTBLK; dev_unlock(); out: if (errp != NULL) *errp = error; return (error == 0); } /* * Common filesystem object access control check routine. Accepts a * vnode's type, "mode", uid and gid, requested access mode, credentials, * and optional call-by-reference privused argument allowing vaccess() * to indicate to the caller whether privilege was used to satisfy the * request (obsoleted). Returns 0 on success, or an errno on failure. */ int vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid, accmode_t accmode, struct ucred *cred, int *privused) { accmode_t dac_granted; accmode_t priv_granted; KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0, ("invalid bit in accmode")); KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE), ("VAPPEND without VWRITE")); /* * Look for a normal, non-privileged way to access the file/directory * as requested. If it exists, go with that. */ if (privused != NULL) *privused = 0; dac_granted = 0; /* Check the owner. */ if (cred->cr_uid == file_uid) { dac_granted |= VADMIN; if (file_mode & S_IXUSR) dac_granted |= VEXEC; if (file_mode & S_IRUSR) dac_granted |= VREAD; if (file_mode & S_IWUSR) dac_granted |= (VWRITE | VAPPEND); if ((accmode & dac_granted) == accmode) return (0); goto privcheck; } /* Otherwise, check the groups (first match) */ if (groupmember(file_gid, cred)) { if (file_mode & S_IXGRP) dac_granted |= VEXEC; if (file_mode & S_IRGRP) dac_granted |= VREAD; if (file_mode & S_IWGRP) dac_granted |= (VWRITE | VAPPEND); if ((accmode & dac_granted) == accmode) return (0); goto privcheck; } /* Otherwise, check everyone else. */ if (file_mode & S_IXOTH) dac_granted |= VEXEC; if (file_mode & S_IROTH) dac_granted |= VREAD; if (file_mode & S_IWOTH) dac_granted |= (VWRITE | VAPPEND); if ((accmode & dac_granted) == accmode) return (0); privcheck: /* * Build a privilege mask to determine if the set of privileges * satisfies the requirements when combined with the granted mask * from above. For each privilege, if the privilege is required, * bitwise or the request type onto the priv_granted mask. */ priv_granted = 0; if (type == VDIR) { /* * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC * requests, instead of PRIV_VFS_EXEC. */ if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && !priv_check_cred(cred, PRIV_VFS_LOOKUP)) priv_granted |= VEXEC; } else { /* * Ensure that at least one execute bit is on. Otherwise, * a privileged user will always succeed, and we don't want * this to happen unless the file really is executable. */ if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) && (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 && !priv_check_cred(cred, PRIV_VFS_EXEC)) priv_granted |= VEXEC; } if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) && !priv_check_cred(cred, PRIV_VFS_READ)) priv_granted |= VREAD; if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) && !priv_check_cred(cred, PRIV_VFS_WRITE)) priv_granted |= (VWRITE | VAPPEND); if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) && !priv_check_cred(cred, PRIV_VFS_ADMIN)) priv_granted |= VADMIN; if ((accmode & (priv_granted | dac_granted)) == accmode) { /* XXX audit: privilege used */ if (privused != NULL) *privused = 1; return (0); } return ((accmode & VADMIN) ? EPERM : EACCES); } /* * Credential check based on process requesting service, and per-attribute * permissions. */ int extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred, struct thread *td, accmode_t accmode) { /* * Kernel-invoked always succeeds. */ if (cred == NOCRED) return (0); /* * Do not allow privileged processes in jail to directly manipulate * system attributes. */ switch (attrnamespace) { case EXTATTR_NAMESPACE_SYSTEM: /* Potentially should be: return (EPERM); */ return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM)); case EXTATTR_NAMESPACE_USER: return (VOP_ACCESS(vp, accmode, cred, td)); default: return (EPERM); } } #ifdef DEBUG_VFS_LOCKS /* * This only exists to suppress warnings from unlocked specfs accesses. It is * no longer ok to have an unlocked VFS. */ #define IGNORE_LOCK(vp) (KERNEL_PANICKED() || (vp) == NULL || \ (vp)->v_type == VCHR || (vp)->v_type == VBAD) int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0, "Drop into debugger on lock violation"); int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex, 0, "Check for interlock across VOPs"); int vfs_badlock_print = 1; /* Print lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print, 0, "Print lock violations"); int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode, 0, "Print vnode details on lock violations"); #ifdef KDB int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */ SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW, &vfs_badlock_backtrace, 0, "Print backtrace at lock violations"); #endif static void vfs_badlock(const char *msg, const char *str, struct vnode *vp) { #ifdef KDB if (vfs_badlock_backtrace) kdb_backtrace(); #endif if (vfs_badlock_vnode) vn_printf(vp, "vnode "); if (vfs_badlock_print) printf("%s: %p %s\n", str, (void *)vp, msg); if (vfs_badlock_ddb) kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); } void assert_vi_locked(struct vnode *vp, const char *str) { if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp))) vfs_badlock("interlock is not locked but should be", str, vp); } void assert_vi_unlocked(struct vnode *vp, const char *str) { if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp))) vfs_badlock("interlock is locked but should not be", str, vp); } void assert_vop_locked(struct vnode *vp, const char *str) { int locked; if (!IGNORE_LOCK(vp)) { locked = VOP_ISLOCKED(vp); if (locked == 0 || locked == LK_EXCLOTHER) vfs_badlock("is not locked but should be", str, vp); } } void assert_vop_unlocked(struct vnode *vp, const char *str) { if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE) vfs_badlock("is locked but should not be", str, vp); } void assert_vop_elocked(struct vnode *vp, const char *str) { if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) vfs_badlock("is not exclusive locked but should be", str, vp); } #endif /* DEBUG_VFS_LOCKS */ void vop_rename_fail(struct vop_rename_args *ap) { if (ap->a_tvp != NULL) vput(ap->a_tvp); if (ap->a_tdvp == ap->a_tvp) vrele(ap->a_tdvp); else vput(ap->a_tdvp); vrele(ap->a_fdvp); vrele(ap->a_fvp); } void vop_rename_pre(void *ap) { struct vop_rename_args *a = ap; #ifdef DEBUG_VFS_LOCKS if (a->a_tvp) ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME"); ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME"); /* Check the source (from). */ if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock && (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock)) ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked"); if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock) ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked"); /* Check the target. */ if (a->a_tvp) ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked"); ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked"); #endif if (a->a_tdvp != a->a_fdvp) vhold(a->a_fdvp); if (a->a_tvp != a->a_fvp) vhold(a->a_fvp); vhold(a->a_tdvp); if (a->a_tvp) vhold(a->a_tvp); } #ifdef DEBUG_VFS_LOCKS void vop_strategy_pre(void *ap) { struct vop_strategy_args *a; struct buf *bp; a = ap; bp = a->a_bp; /* * Cluster ops lock their component buffers but not the IO container. */ if ((bp->b_flags & B_CLUSTER) != 0) return; if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) { if (vfs_badlock_print) printf( "VOP_STRATEGY: bp is not locked but should be\n"); if (vfs_badlock_ddb) kdb_enter(KDB_WHY_VFSLOCK, "lock violation"); } } void vop_lock_pre(void *ap) { struct vop_lock1_args *a = ap; if ((a->a_flags & LK_INTERLOCK) == 0) ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); else ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK"); } void vop_lock_post(void *ap, int rc) { struct vop_lock1_args *a = ap; ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK"); if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0) ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK"); } void vop_unlock_pre(void *ap) { struct vop_unlock_args *a = ap; ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK"); } void vop_need_inactive_pre(void *ap) { struct vop_need_inactive_args *a = ap; ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); } void vop_need_inactive_post(void *ap, int rc) { struct vop_need_inactive_args *a = ap; ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE"); } #endif void vop_create_post(void *ap, int rc) { struct vop_create_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); } void vop_deleteextattr_post(void *ap, int rc) { struct vop_deleteextattr_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); } void vop_link_post(void *ap, int rc) { struct vop_link_args *a = ap; if (!rc) { VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK); VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE); } } void vop_mkdir_post(void *ap, int rc) { struct vop_mkdir_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); } void vop_mknod_post(void *ap, int rc) { struct vop_mknod_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); } void vop_reclaim_post(void *ap, int rc) { struct vop_reclaim_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE); } void vop_remove_post(void *ap, int rc) { struct vop_remove_args *a = ap; if (!rc) { VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); } } void vop_rename_post(void *ap, int rc) { struct vop_rename_args *a = ap; long hint; if (!rc) { hint = NOTE_WRITE; if (a->a_fdvp == a->a_tdvp) { if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR) hint |= NOTE_LINK; VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); } else { hint |= NOTE_EXTEND; if (a->a_fvp->v_type == VDIR) hint |= NOTE_LINK; VFS_KNOTE_UNLOCKED(a->a_fdvp, hint); if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL && a->a_tvp->v_type == VDIR) hint &= ~NOTE_LINK; VFS_KNOTE_UNLOCKED(a->a_tdvp, hint); } VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME); if (a->a_tvp) VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE); } if (a->a_tdvp != a->a_fdvp) vdrop(a->a_fdvp); if (a->a_tvp != a->a_fvp) vdrop(a->a_fvp); vdrop(a->a_tdvp); if (a->a_tvp) vdrop(a->a_tvp); } void vop_rmdir_post(void *ap, int rc) { struct vop_rmdir_args *a = ap; if (!rc) { VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK); VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE); } } void vop_setattr_post(void *ap, int rc) { struct vop_setattr_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); } void vop_setextattr_post(void *ap, int rc) { struct vop_setextattr_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB); } void vop_symlink_post(void *ap, int rc) { struct vop_symlink_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE); } void vop_open_post(void *ap, int rc) { struct vop_open_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN); } void vop_close_post(void *ap, int rc) { struct vop_close_args *a = ap; if (!rc && (a->a_cred != NOCRED || /* filter out revokes */ !VN_IS_DOOMED(a->a_vp))) { VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ? NOTE_CLOSE_WRITE : NOTE_CLOSE); } } void vop_read_post(void *ap, int rc) { struct vop_read_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); } void vop_readdir_post(void *ap, int rc) { struct vop_readdir_args *a = ap; if (!rc) VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ); } static struct knlist fs_knlist; static void vfs_event_init(void *arg) { knlist_init_mtx(&fs_knlist, NULL); } /* XXX - correct order? */ SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL); void vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused) { KNOTE_UNLOCKED(&fs_knlist, event); } static int filt_fsattach(struct knote *kn); static void filt_fsdetach(struct knote *kn); static int filt_fsevent(struct knote *kn, long hint); struct filterops fs_filtops = { .f_isfd = 0, .f_attach = filt_fsattach, .f_detach = filt_fsdetach, .f_event = filt_fsevent }; static int filt_fsattach(struct knote *kn) { kn->kn_flags |= EV_CLEAR; knlist_add(&fs_knlist, kn, 0); return (0); } static void filt_fsdetach(struct knote *kn) { knlist_remove(&fs_knlist, kn, 0); } static int filt_fsevent(struct knote *kn, long hint) { kn->kn_fflags |= hint; return (kn->kn_fflags != 0); } static int sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS) { struct vfsidctl vc; int error; struct mount *mp; error = SYSCTL_IN(req, &vc, sizeof(vc)); if (error) return (error); if (vc.vc_vers != VFS_CTL_VERS1) return (EINVAL); mp = vfs_getvfs(&vc.vc_fsid); if (mp == NULL) return (ENOENT); /* ensure that a specific sysctl goes to the right filesystem. */ if (strcmp(vc.vc_fstypename, "*") != 0 && strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) { vfs_rel(mp); return (EINVAL); } VCTLTOREQ(&vc, req); error = VFS_SYSCTL(mp, vc.vc_op, req); vfs_rel(mp); return (error); } SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0, sysctl_vfs_ctl, "", "Sysctl by fsid"); /* * Function to initialize a va_filerev field sensibly. * XXX: Wouldn't a random number make a lot more sense ?? */ u_quad_t init_va_filerev(void) { struct bintime bt; getbinuptime(&bt); return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL)); } static int filt_vfsread(struct knote *kn, long hint); static int filt_vfswrite(struct knote *kn, long hint); static int filt_vfsvnode(struct knote *kn, long hint); static void filt_vfsdetach(struct knote *kn); static struct filterops vfsread_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfsread }; static struct filterops vfswrite_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfswrite }; static struct filterops vfsvnode_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfsvnode }; static void vfs_knllock(void *arg) { struct vnode *vp = arg; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } static void vfs_knlunlock(void *arg) { struct vnode *vp = arg; VOP_UNLOCK(vp); } static void vfs_knl_assert_locked(void *arg) { #ifdef DEBUG_VFS_LOCKS struct vnode *vp = arg; ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked"); #endif } static void vfs_knl_assert_unlocked(void *arg) { #ifdef DEBUG_VFS_LOCKS struct vnode *vp = arg; ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked"); #endif } int vfs_kqfilter(struct vop_kqfilter_args *ap) { struct vnode *vp = ap->a_vp; struct knote *kn = ap->a_kn; struct knlist *knl; switch (kn->kn_filter) { case EVFILT_READ: kn->kn_fop = &vfsread_filtops; break; case EVFILT_WRITE: kn->kn_fop = &vfswrite_filtops; break; case EVFILT_VNODE: kn->kn_fop = &vfsvnode_filtops; break; default: return (EINVAL); } kn->kn_hook = (caddr_t)vp; v_addpollinfo(vp); if (vp->v_pollinfo == NULL) return (ENOMEM); knl = &vp->v_pollinfo->vpi_selinfo.si_note; vhold(vp); knlist_add(knl, kn, 0); return (0); } /* * Detach knote from vnode */ static void filt_vfsdetach(struct knote *kn) { struct vnode *vp = (struct vnode *)kn->kn_hook; KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo")); knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0); vdrop(vp); } /*ARGSUSED*/ static int filt_vfsread(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; struct vattr va; int res; /* * filesystem is gone, so set the EOF flag and schedule * the knote for deletion. */ if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { VI_LOCK(vp); kn->kn_flags |= (EV_EOF | EV_ONESHOT); VI_UNLOCK(vp); return (1); } if (VOP_GETATTR(vp, &va, curthread->td_ucred)) return (0); VI_LOCK(vp); kn->kn_data = va.va_size - kn->kn_fp->f_offset; res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0; VI_UNLOCK(vp); return (res); } /*ARGSUSED*/ static int filt_vfswrite(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; VI_LOCK(vp); /* * filesystem is gone, so set the EOF flag and schedule * the knote for deletion. */ if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) kn->kn_flags |= (EV_EOF | EV_ONESHOT); kn->kn_data = 0; VI_UNLOCK(vp); return (1); } static int filt_vfsvnode(struct knote *kn, long hint) { struct vnode *vp = (struct vnode *)kn->kn_hook; int res; VI_LOCK(vp); if (kn->kn_sfflags & hint) kn->kn_fflags |= hint; if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) { kn->kn_flags |= EV_EOF; VI_UNLOCK(vp); return (1); } res = (kn->kn_fflags != 0); VI_UNLOCK(vp); return (res); } /* * Returns whether the directory is empty or not. * If it is empty, the return value is 0; otherwise * the return value is an error value (which may * be ENOTEMPTY). */ int vfs_emptydir(struct vnode *vp) { struct uio uio; struct iovec iov; struct dirent *dirent, *dp, *endp; int error, eof; error = 0; eof = 0; ASSERT_VOP_LOCKED(vp, "vfs_emptydir"); dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK); iov.iov_base = dirent; iov.iov_len = sizeof(struct dirent); uio.uio_iov = &iov; uio.uio_iovcnt = 1; uio.uio_offset = 0; uio.uio_resid = sizeof(struct dirent); uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = curthread; while (eof == 0 && error == 0) { error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof, NULL, NULL); if (error != 0) break; endp = (void *)((uint8_t *)dirent + sizeof(struct dirent) - uio.uio_resid); for (dp = dirent; dp < endp; dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) { if (dp->d_type == DT_WHT) continue; if (dp->d_namlen == 0) continue; if (dp->d_type != DT_DIR && dp->d_type != DT_UNKNOWN) { error = ENOTEMPTY; break; } if (dp->d_namlen > 2) { error = ENOTEMPTY; break; } if (dp->d_namlen == 1 && dp->d_name[0] != '.') { error = ENOTEMPTY; break; } if (dp->d_namlen == 2 && dp->d_name[1] != '.') { error = ENOTEMPTY; break; } uio.uio_resid = sizeof(struct dirent); } } free(dirent, M_TEMP); return (error); } int vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) { int error; if (dp->d_reclen > ap->a_uio->uio_resid) return (ENAMETOOLONG); error = uiomove(dp, dp->d_reclen, ap->a_uio); if (error) { if (ap->a_ncookies != NULL) { if (ap->a_cookies != NULL) free(ap->a_cookies, M_TEMP); ap->a_cookies = NULL; *ap->a_ncookies = 0; } return (error); } if (ap->a_ncookies == NULL) return (0); KASSERT(ap->a_cookies, ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!")); *ap->a_cookies = realloc(*ap->a_cookies, (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO); (*ap->a_cookies)[*ap->a_ncookies] = off; *ap->a_ncookies += 1; return (0); } /* * The purpose of this routine is to remove granularity from accmode_t, * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE, * VADMIN and VAPPEND. * * If it returns 0, the caller is supposed to continue with the usual * access checks using 'accmode' as modified by this routine. If it * returns nonzero value, the caller is supposed to return that value * as errno. * * Note that after this routine runs, accmode may be zero. */ int vfs_unixify_accmode(accmode_t *accmode) { /* * There is no way to specify explicit "deny" rule using * file mode or POSIX.1e ACLs. */ if (*accmode & VEXPLICIT_DENY) { *accmode = 0; return (0); } /* * None of these can be translated into usual access bits. * Also, the common case for NFSv4 ACLs is to not contain * either of these bits. Caller should check for VWRITE * on the containing directory instead. */ if (*accmode & (VDELETE_CHILD | VDELETE)) return (EPERM); if (*accmode & VADMIN_PERMS) { *accmode &= ~VADMIN_PERMS; *accmode |= VADMIN; } /* * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL * or VSYNCHRONIZE using file mode or POSIX.1e ACL. */ *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE); return (0); } /* * Clear out a doomed vnode (if any) and replace it with a new one as long * as the fs is not being unmounted. Return the root vnode to the caller. */ static int __noinline vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp) { struct vnode *vp; int error; restart: if (mp->mnt_rootvnode != NULL) { MNT_ILOCK(mp); vp = mp->mnt_rootvnode; if (vp != NULL) { if (!VN_IS_DOOMED(vp)) { vrefact(vp); MNT_IUNLOCK(mp); error = vn_lock(vp, flags); if (error == 0) { *vpp = vp; return (0); } vrele(vp); goto restart; } /* * Clear the old one. */ mp->mnt_rootvnode = NULL; } MNT_IUNLOCK(mp); if (vp != NULL) { vfs_op_barrier_wait(mp); vrele(vp); } } error = VFS_CACHEDROOT(mp, flags, vpp); if (error != 0) return (error); if (mp->mnt_vfs_ops == 0) { MNT_ILOCK(mp); if (mp->mnt_vfs_ops != 0) { MNT_IUNLOCK(mp); return (0); } if (mp->mnt_rootvnode == NULL) { vrefact(*vpp); mp->mnt_rootvnode = *vpp; } else { if (mp->mnt_rootvnode != *vpp) { if (!VN_IS_DOOMED(mp->mnt_rootvnode)) { panic("%s: mismatch between vnode returned " " by VFS_CACHEDROOT and the one cached " " (%p != %p)", __func__, *vpp, mp->mnt_rootvnode); } } } MNT_IUNLOCK(mp); } return (0); } int vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp) { struct vnode *vp; int error; if (!vfs_op_thread_enter(mp)) return (vfs_cache_root_fallback(mp, flags, vpp)); vp = atomic_load_ptr(&mp->mnt_rootvnode); if (vp == NULL || VN_IS_DOOMED(vp)) { vfs_op_thread_exit(mp); return (vfs_cache_root_fallback(mp, flags, vpp)); } vrefact(vp); vfs_op_thread_exit(mp); error = vn_lock(vp, flags); if (error != 0) { vrele(vp); return (vfs_cache_root_fallback(mp, flags, vpp)); } *vpp = vp; return (0); } struct vnode * vfs_cache_root_clear(struct mount *mp) { struct vnode *vp; /* * ops > 0 guarantees there is nobody who can see this vnode */ MPASS(mp->mnt_vfs_ops > 0); vp = mp->mnt_rootvnode; mp->mnt_rootvnode = NULL; return (vp); } void vfs_cache_root_set(struct mount *mp, struct vnode *vp) { MPASS(mp->mnt_vfs_ops > 0); vrefact(vp); mp->mnt_rootvnode = vp; } /* * These are helper functions for filesystems to traverse all * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h. * * This interface replaces MNT_VNODE_FOREACH. */ struct vnode * __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp) { struct vnode *vp; if (should_yield()) kern_yield(PRI_USER); MNT_ILOCK(mp); KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL; vp = TAILQ_NEXT(vp, v_nmntvnodes)) { /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) continue; VI_LOCK(vp); if (VN_IS_DOOMED(vp)) { VI_UNLOCK(vp); continue; } break; } if (vp == NULL) { __mnt_vnode_markerfree_all(mvp, mp); /* MNT_IUNLOCK(mp); -- done in above function */ mtx_assert(MNT_MTX(mp), MA_NOTOWNED); return (NULL); } TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); MNT_IUNLOCK(mp); return (vp); } struct vnode * __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp) { struct vnode *vp; *mvp = vn_alloc_marker(mp); MNT_ILOCK(mp); MNT_REF(mp); TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) { /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */ if (vp->v_type == VMARKER || VN_IS_DOOMED(vp)) continue; VI_LOCK(vp); if (VN_IS_DOOMED(vp)) { VI_UNLOCK(vp); continue; } break; } if (vp == NULL) { MNT_REL(mp); MNT_IUNLOCK(mp); vn_free_marker(*mvp); *mvp = NULL; return (NULL); } TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes); MNT_IUNLOCK(mp); return (vp); } void __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp) { if (*mvp == NULL) { MNT_IUNLOCK(mp); return; } mtx_assert(MNT_MTX(mp), MA_OWNED); KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes); MNT_REL(mp); MNT_IUNLOCK(mp); vn_free_marker(*mvp); *mvp = NULL; } /* * These are helper functions for filesystems to traverse their * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h */ static void mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) { KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); MNT_ILOCK(mp); MNT_REL(mp); MNT_IUNLOCK(mp); vn_free_marker(*mvp); *mvp = NULL; } /* * Relock the mp mount vnode list lock with the vp vnode interlock in the * conventional lock order during mnt_vnode_next_lazy iteration. * * On entry, the mount vnode list lock is held and the vnode interlock is not. * The list lock is dropped and reacquired. On success, both locks are held. * On failure, the mount vnode list lock is held but the vnode interlock is * not, and the procedure may have yielded. */ static bool mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp, struct vnode *vp) { VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER && TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp, ("%s: bad marker", __func__)); VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp, ("%s: inappropriate vnode", __func__)); ASSERT_VI_UNLOCKED(vp, __func__); mtx_assert(&mp->mnt_listmtx, MA_OWNED); TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist); TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist); /* * Note we may be racing against vdrop which transitioned the hold * count to 0 and now waits for the ->mnt_listmtx lock. This is fine, * if we are the only user after we get the interlock we will just * vdrop. */ vhold(vp); mtx_unlock(&mp->mnt_listmtx); VI_LOCK(vp); if (VN_IS_DOOMED(vp)) { VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp); goto out_lost; } VNPASS(vp->v_mflag & VMP_LAZYLIST, vp); /* * There is nothing to do if we are the last user. */ if (!refcount_release_if_not_last(&vp->v_holdcnt)) goto out_lost; mtx_lock(&mp->mnt_listmtx); return (true); out_lost: vdropl(vp); maybe_yield(); mtx_lock(&mp->mnt_listmtx); return (false); } static struct vnode * mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg) { struct vnode *vp; mtx_assert(&mp->mnt_listmtx, MA_OWNED); KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch")); restart: vp = TAILQ_NEXT(*mvp, v_lazylist); while (vp != NULL) { if (vp->v_type == VMARKER) { vp = TAILQ_NEXT(vp, v_lazylist); continue; } /* * See if we want to process the vnode. Note we may encounter a * long string of vnodes we don't care about and hog the list * as a result. Check for it and requeue the marker. */ VNPASS(!VN_IS_DOOMED(vp), vp); if (!cb(vp, cbarg)) { if (!should_yield()) { vp = TAILQ_NEXT(vp, v_lazylist); continue; } TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); mtx_unlock(&mp->mnt_listmtx); kern_yield(PRI_USER); mtx_lock(&mp->mnt_listmtx); goto restart; } /* * Try-lock because this is the wrong lock order. */ if (!VI_TRYLOCK(vp) && !mnt_vnode_next_lazy_relock(*mvp, mp, vp)) goto restart; KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp)); KASSERT(vp->v_mount == mp || vp->v_mount == NULL, ("alien vnode on the lazy list %p %p", vp, mp)); VNPASS(vp->v_mount == mp, vp); VNPASS(!VN_IS_DOOMED(vp), vp); break; } TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); /* Check if we are done */ if (vp == NULL) { mtx_unlock(&mp->mnt_listmtx); mnt_vnode_markerfree_lazy(mvp, mp); return (NULL); } TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist); mtx_unlock(&mp->mnt_listmtx); ASSERT_VI_LOCKED(vp, "lazy iter"); return (vp); } struct vnode * __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg) { if (should_yield()) kern_yield(PRI_USER); mtx_lock(&mp->mnt_listmtx); return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); } struct vnode * __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb, void *cbarg) { struct vnode *vp; if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist)) return (NULL); *mvp = vn_alloc_marker(mp); MNT_ILOCK(mp); MNT_REF(mp); MNT_IUNLOCK(mp); mtx_lock(&mp->mnt_listmtx); vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist); if (vp == NULL) { mtx_unlock(&mp->mnt_listmtx); mnt_vnode_markerfree_lazy(mvp, mp); return (NULL); } TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist); return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg)); } void __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp) { if (*mvp == NULL) return; mtx_lock(&mp->mnt_listmtx); TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist); mtx_unlock(&mp->mnt_listmtx); mnt_vnode_markerfree_lazy(mvp, mp); } int vn_dir_check_exec(struct vnode *vp, struct componentname *cnp) { if ((cnp->cn_flags & NOEXECCHECK) != 0) { cnp->cn_flags &= ~NOEXECCHECK; return (0); } return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, cnp->cn_thread)); } Index: head/sys/kern/vfs_vnops.c =================================================================== --- head/sys/kern/vfs_vnops.c (revision 363070) +++ head/sys/kern/vfs_vnops.c (revision 363071) @@ -1,3271 +1,3271 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Copyright (c) 2012 Konstantin Belousov * Copyright (c) 2013, 2014 The FreeBSD Foundation * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif static fo_rdwr_t vn_read; static fo_rdwr_t vn_write; static fo_rdwr_t vn_io_fault; static fo_truncate_t vn_truncate; static fo_ioctl_t vn_ioctl; static fo_poll_t vn_poll; static fo_kqfilter_t vn_kqfilter; static fo_stat_t vn_statfile; static fo_close_t vn_closefile; static fo_mmap_t vn_mmap; static fo_fallocate_t vn_fallocate; struct fileops vnops = { .fo_read = vn_io_fault, .fo_write = vn_io_fault, .fo_truncate = vn_truncate, .fo_ioctl = vn_ioctl, .fo_poll = vn_poll, .fo_kqfilter = vn_kqfilter, .fo_stat = vn_statfile, .fo_close = vn_closefile, .fo_chmod = vn_chmod, .fo_chown = vn_chown, .fo_sendfile = vn_sendfile, .fo_seek = vn_seek, .fo_fill_kinfo = vn_fill_kinfo, .fo_mmap = vn_mmap, .fo_fallocate = vn_fallocate, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; static const int io_hold_cnt = 16; static int vn_io_fault_enable = 1; SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW, &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); static int vn_io_fault_prefault = 0; SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW, &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); static u_long vn_io_faults_cnt; SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); static int vfs_allow_read_dir = 0; SYSCTL_INT(_security_bsd, OID_AUTO, allow_read_dir, CTLFLAG_RW, &vfs_allow_read_dir, 0, "Enable read(2) of directory by root for filesystems that support it"); /* * Returns true if vn_io_fault mode of handling the i/o request should * be used. */ static bool do_vn_io_fault(struct vnode *vp, struct uio *uio) { struct mount *mp; return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && (mp = vp->v_mount) != NULL && (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); } /* * Structure used to pass arguments to vn_io_fault1(), to do either * file- or vnode-based I/O calls. */ struct vn_io_fault_args { enum { VN_IO_FAULT_FOP, VN_IO_FAULT_VOP } kind; struct ucred *cred; int flags; union { struct fop_args_tag { struct file *fp; fo_rdwr_t *doio; } fop_args; struct vop_args_tag { struct vnode *vp; } vop_args; } args; }; static int vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, struct thread *td); int vn_open(struct nameidata *ndp, int *flagp, int cmode, struct file *fp) { struct thread *td = ndp->ni_cnd.cn_thread; return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); } /* * Common code for vnode open operations via a name lookup. * Lookup the vnode and invoke VOP_CREATE if needed. * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. - * + * * Note that this does NOT free nameidata for the successful case, * due to the NDINIT being done elsewhere. */ int vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, struct ucred *cred, struct file *fp) { struct vnode *vp; struct mount *mp; struct thread *td = ndp->ni_cnd.cn_thread; struct vattr vat; struct vattr *vap = &vat; int fmode, error; restart: fmode = *flagp; if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | O_EXCL | O_DIRECTORY)) return (EINVAL); else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { ndp->ni_cnd.cn_nameiop = CREATE; /* * Set NOCACHE to avoid flushing the cache when * rolling in many files at once. */ ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE; if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) ndp->ni_cnd.cn_flags |= FOLLOW; if ((fmode & O_BENEATH) != 0) ndp->ni_cnd.cn_flags |= BENEATH; if (!(vn_open_flags & VN_OPEN_NOAUDIT)) ndp->ni_cnd.cn_flags |= AUDITVNODE1; if (vn_open_flags & VN_OPEN_NOCAPCHECK) ndp->ni_cnd.cn_flags |= NOCAPCHECK; if ((vn_open_flags & VN_OPEN_INVFS) == 0) bwillwrite(); if ((error = namei(ndp)) != 0) return (error); if (ndp->ni_vp == NULL) { VATTR_NULL(vap); vap->va_type = VREG; vap->va_mode = cmode; if (fmode & O_EXCL) vap->va_vaflags |= VA_EXCLUSIVE; if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { NDFREE(ndp, NDF_ONLY_PNBUF); vput(ndp->ni_dvp); if ((error = vn_start_write(NULL, &mp, V_XSLEEP | PCATCH)) != 0) return (error); goto restart; } if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) ndp->ni_cnd.cn_flags |= MAKEENTRY; #ifdef MAC error = mac_vnode_check_create(cred, ndp->ni_dvp, &ndp->ni_cnd, vap); if (error == 0) #endif error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, &ndp->ni_cnd, vap); vput(ndp->ni_dvp); vn_finished_write(mp); if (error) { NDFREE(ndp, NDF_ONLY_PNBUF); return (error); } fmode &= ~O_TRUNC; vp = ndp->ni_vp; } else { if (ndp->ni_dvp == ndp->ni_vp) vrele(ndp->ni_dvp); else vput(ndp->ni_dvp); ndp->ni_dvp = NULL; vp = ndp->ni_vp; if (fmode & O_EXCL) { error = EEXIST; goto bad; } if (vp->v_type == VDIR) { error = EISDIR; goto bad; } fmode &= ~O_CREAT; } } else { ndp->ni_cnd.cn_nameiop = LOOKUP; ndp->ni_cnd.cn_flags = ISOPEN | ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF; if (!(fmode & FWRITE)) ndp->ni_cnd.cn_flags |= LOCKSHARED; if ((fmode & O_BENEATH) != 0) ndp->ni_cnd.cn_flags |= BENEATH; if (!(vn_open_flags & VN_OPEN_NOAUDIT)) ndp->ni_cnd.cn_flags |= AUDITVNODE1; if (vn_open_flags & VN_OPEN_NOCAPCHECK) ndp->ni_cnd.cn_flags |= NOCAPCHECK; if ((error = namei(ndp)) != 0) return (error); vp = ndp->ni_vp; } error = vn_open_vnode(vp, fmode, cred, td, fp); if (error) goto bad; *flagp = fmode; return (0); bad: NDFREE(ndp, NDF_ONLY_PNBUF); vput(vp); *flagp = fmode; ndp->ni_vp = NULL; return (error); } static int vn_open_vnode_advlock(struct vnode *vp, int fmode, struct file *fp) { struct flock lf; int error, lock_flags, type; ASSERT_VOP_LOCKED(vp, "vn_open_vnode_advlock"); if ((fmode & (O_EXLOCK | O_SHLOCK)) == 0) return (0); KASSERT(fp != NULL, ("open with flock requires fp")); if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) return (EOPNOTSUPP); lock_flags = VOP_ISLOCKED(vp); VOP_UNLOCK(vp); lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; lf.l_type = (fmode & O_EXLOCK) != 0 ? F_WRLCK : F_RDLCK; type = F_FLOCK; if ((fmode & FNONBLOCK) == 0) type |= F_WAIT; error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); if (error == 0) fp->f_flag |= FHASLOCK; vn_lock(vp, lock_flags | LK_RETRY); if (error == 0 && VN_IS_DOOMED(vp)) error = ENOENT; return (error); } /* * Common code for vnode open operations once a vnode is located. * Check permissions, and call the VOP_OPEN routine. */ int vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, struct thread *td, struct file *fp) { accmode_t accmode; int error; if (vp->v_type == VLNK) return (EMLINK); if (vp->v_type == VSOCK) return (EOPNOTSUPP); if (vp->v_type != VDIR && fmode & O_DIRECTORY) return (ENOTDIR); accmode = 0; if (fmode & (FWRITE | O_TRUNC)) { if (vp->v_type == VDIR) return (EISDIR); accmode |= VWRITE; } if (fmode & FREAD) accmode |= VREAD; if (fmode & FEXEC) accmode |= VEXEC; if ((fmode & O_APPEND) && (fmode & FWRITE)) accmode |= VAPPEND; #ifdef MAC if (fmode & O_CREAT) accmode |= VCREAT; if (fmode & O_VERIFY) accmode |= VVERIFY; error = mac_vnode_check_open(cred, vp, accmode); if (error) return (error); accmode &= ~(VCREAT | VVERIFY); #endif if ((fmode & O_CREAT) == 0 && accmode != 0) { error = VOP_ACCESS(vp, accmode, cred, td); if (error != 0) return (error); } if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) vn_lock(vp, LK_UPGRADE | LK_RETRY); error = VOP_OPEN(vp, fmode, cred, td, fp); if (error != 0) return (error); error = vn_open_vnode_advlock(vp, fmode, fp); if (error == 0 && (fmode & FWRITE) != 0) { error = VOP_ADD_WRITECOUNT(vp, 1); if (error == 0) { CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", __func__, vp, vp->v_writecount); } } /* * Error from advlock or VOP_ADD_WRITECOUNT() still requires * calling VOP_CLOSE() to pair with earlier VOP_OPEN(). * Arrange for that by having fdrop() to use vn_closefile(). */ if (error != 0) { fp->f_flag |= FOPENFAILED; fp->f_vnode = vp; if (fp->f_ops == &badfileops) { fp->f_type = DTYPE_VNODE; fp->f_ops = &vnops; } vref(vp); } ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); return (error); } /* * Check for write permissions on the specified vnode. * Prototype text segments cannot be written. * It is racy. */ int vn_writechk(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, "vn_writechk"); /* * If there's shared text associated with * the vnode, try to free it up once. If * we fail, we can't allow writing. */ if (VOP_IS_TEXT(vp)) return (ETXTBSY); return (0); } /* * Vnode close call */ static int vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td, bool keep_ref) { struct mount *mp; int error, lock_flags; if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && MNT_EXTENDED_SHARED(vp->v_mount)) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, lock_flags | LK_RETRY); AUDIT_ARG_VNODE1(vp); if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { VOP_ADD_WRITECOUNT_CHECKED(vp, -1); CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", __func__, vp, vp->v_writecount); } error = VOP_CLOSE(vp, flags, file_cred, td); if (keep_ref) VOP_UNLOCK(vp); else vput(vp); vn_finished_write(mp); return (error); } int vn_close(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td) { return (vn_close1(vp, flags, file_cred, td, false)); } /* * Heuristic to detect sequential operation. */ static int sequential_heuristic(struct uio *uio, struct file *fp) { enum uio_rw rw; ASSERT_VOP_LOCKED(fp->f_vnode, __func__); rw = uio->uio_rw; if (fp->f_flag & FRDAHEAD) return (fp->f_seqcount[rw] << IO_SEQSHIFT); /* * Offset 0 is handled specially. open() sets f_seqcount to 1 so * that the first I/O is normally considered to be slightly * sequential. Seeking to offset 0 doesn't change sequentiality * unless previous seeks have reduced f_seqcount to 0, in which * case offset 0 is not special. */ if ((uio->uio_offset == 0 && fp->f_seqcount[rw] > 0) || uio->uio_offset == fp->f_nextoff[rw]) { /* * f_seqcount is in units of fixed-size blocks so that it * depends mainly on the amount of sequential I/O and not * much on the number of sequential I/O's. The fixed size * of 16384 is hard-coded here since it is (not quite) just * a magic size that works well here. This size is more * closely related to the best I/O size for real disks than * to any block size used by software. */ if (uio->uio_resid >= IO_SEQMAX * 16384) fp->f_seqcount[rw] = IO_SEQMAX; else { fp->f_seqcount[rw] += howmany(uio->uio_resid, 16384); if (fp->f_seqcount[rw] > IO_SEQMAX) fp->f_seqcount[rw] = IO_SEQMAX; } return (fp->f_seqcount[rw] << IO_SEQSHIFT); } /* Not sequential. Quickly draw-down sequentiality. */ if (fp->f_seqcount[rw] > 1) fp->f_seqcount[rw] = 1; else fp->f_seqcount[rw] = 0; return (0); } /* * Package up an I/O request on a vnode into a uio and do it. */ int vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, struct ucred *file_cred, ssize_t *aresid, struct thread *td) { struct uio auio; struct iovec aiov; struct mount *mp; struct ucred *cred; void *rl_cookie; struct vn_io_fault_args args; int error, lock_flags; if (offset < 0 && vp->v_type != VCHR) return (EINVAL); auio.uio_iov = &aiov; auio.uio_iovcnt = 1; aiov.iov_base = base; aiov.iov_len = len; auio.uio_resid = len; auio.uio_offset = offset; auio.uio_segflg = segflg; auio.uio_rw = rw; auio.uio_td = td; error = 0; if ((ioflg & IO_NODELOCKED) == 0) { if ((ioflg & IO_RANGELOCKED) == 0) { if (rw == UIO_READ) { rl_cookie = vn_rangelock_rlock(vp, offset, offset + len); } else { rl_cookie = vn_rangelock_wlock(vp, offset, offset + len); } } else rl_cookie = NULL; mp = NULL; if (rw == UIO_WRITE) { if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto out; if (MNT_SHARED_WRITES(mp) || ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; } else lock_flags = LK_SHARED; vn_lock(vp, lock_flags | LK_RETRY); } else rl_cookie = NULL; ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); #ifdef MAC if ((ioflg & IO_NOMACCHECK) == 0) { if (rw == UIO_READ) error = mac_vnode_check_read(active_cred, file_cred, vp); else error = mac_vnode_check_write(active_cred, file_cred, vp); } #endif if (error == 0) { if (file_cred != NULL) cred = file_cred; else cred = active_cred; if (do_vn_io_fault(vp, &auio)) { args.kind = VN_IO_FAULT_VOP; args.cred = cred; args.flags = ioflg; args.args.vop_args.vp = vp; error = vn_io_fault1(vp, &auio, &args, td); } else if (rw == UIO_READ) { error = VOP_READ(vp, &auio, ioflg, cred); } else /* if (rw == UIO_WRITE) */ { error = VOP_WRITE(vp, &auio, ioflg, cred); } } if (aresid) *aresid = auio.uio_resid; else if (auio.uio_resid && error == 0) error = EIO; if ((ioflg & IO_NODELOCKED) == 0) { VOP_UNLOCK(vp); if (mp != NULL) vn_finished_write(mp); } out: if (rl_cookie != NULL) vn_rangelock_unlock(vp, rl_cookie); return (error); } /* * Package up an I/O request on a vnode into a uio and do it. The I/O * request is split up into smaller chunks and we try to avoid saturating * the buffer cache while potentially holding a vnode locked, so we * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() * to give other processes a chance to lock the vnode (either other processes * core'ing the same binary, or unrelated processes scanning the directory). */ int vn_rdwr_inchunks(enum uio_rw rw, struct vnode *vp, void *base, size_t len, off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, struct ucred *file_cred, size_t *aresid, struct thread *td) { int error = 0; ssize_t iaresid; do { int chunk; /* * Force `offset' to a multiple of MAXBSIZE except possibly * for the first chunk, so that filesystems only need to * write full blocks except possibly for the first and last * chunks. */ chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; if (chunk > len) chunk = len; if (rw != UIO_READ && vp->v_type == VREG) bwillwrite(); iaresid = 0; error = vn_rdwr(rw, vp, base, chunk, offset, segflg, ioflg, active_cred, file_cred, &iaresid, td); len -= chunk; /* aresid calc already includes length */ if (error) break; offset += chunk; base = (char *)base + chunk; kern_yield(PRI_USER); } while (len); if (aresid) *aresid = len + iaresid; return (error); } #if OFF_MAX <= LONG_MAX off_t foffset_lock(struct file *fp, int flags) { volatile short *flagsp; off_t res; short state; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); if ((flags & FOF_NOLOCK) != 0) return (atomic_load_long(&fp->f_offset)); /* * According to McKusick the vn lock was protecting f_offset here. * It is now protected by the FOFFSET_LOCKED flag. */ flagsp = &fp->f_vnread_flags; if (atomic_cmpset_acq_16(flagsp, 0, FOFFSET_LOCKED)) return (atomic_load_long(&fp->f_offset)); sleepq_lock(&fp->f_vnread_flags); state = atomic_load_16(flagsp); for (;;) { if ((state & FOFFSET_LOCKED) == 0) { if (!atomic_fcmpset_acq_16(flagsp, &state, FOFFSET_LOCKED)) continue; break; } if ((state & FOFFSET_LOCK_WAITING) == 0) { if (!atomic_fcmpset_acq_16(flagsp, &state, state | FOFFSET_LOCK_WAITING)) continue; } DROP_GIANT(); sleepq_add(&fp->f_vnread_flags, NULL, "vofflock", 0, 0); sleepq_wait(&fp->f_vnread_flags, PUSER -1); PICKUP_GIANT(); sleepq_lock(&fp->f_vnread_flags); state = atomic_load_16(flagsp); } res = atomic_load_long(&fp->f_offset); sleepq_release(&fp->f_vnread_flags); return (res); } void foffset_unlock(struct file *fp, off_t val, int flags) { volatile short *flagsp; short state; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); if ((flags & FOF_NOUPDATE) == 0) atomic_store_long(&fp->f_offset, val); if ((flags & FOF_NEXTOFF_R) != 0) fp->f_nextoff[UIO_READ] = val; if ((flags & FOF_NEXTOFF_W) != 0) fp->f_nextoff[UIO_WRITE] = val; if ((flags & FOF_NOLOCK) != 0) return; flagsp = &fp->f_vnread_flags; state = atomic_load_16(flagsp); if ((state & FOFFSET_LOCK_WAITING) == 0 && atomic_cmpset_rel_16(flagsp, state, 0)) return; sleepq_lock(&fp->f_vnread_flags); MPASS((fp->f_vnread_flags & FOFFSET_LOCKED) != 0); MPASS((fp->f_vnread_flags & FOFFSET_LOCK_WAITING) != 0); fp->f_vnread_flags = 0; sleepq_broadcast(&fp->f_vnread_flags, SLEEPQ_SLEEP, 0, 0); sleepq_release(&fp->f_vnread_flags); } #else off_t foffset_lock(struct file *fp, int flags) { struct mtx *mtxp; off_t res; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if ((flags & FOF_NOLOCK) == 0) { while (fp->f_vnread_flags & FOFFSET_LOCKED) { fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; msleep(&fp->f_vnread_flags, mtxp, PUSER -1, "vofflock", 0); } fp->f_vnread_flags |= FOFFSET_LOCKED; } res = fp->f_offset; mtx_unlock(mtxp); return (res); } void foffset_unlock(struct file *fp, off_t val, int flags) { struct mtx *mtxp; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if ((flags & FOF_NOUPDATE) == 0) fp->f_offset = val; if ((flags & FOF_NEXTOFF_R) != 0) fp->f_nextoff[UIO_READ] = val; if ((flags & FOF_NEXTOFF_W) != 0) fp->f_nextoff[UIO_WRITE] = val; if ((flags & FOF_NOLOCK) == 0) { KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, ("Lost FOFFSET_LOCKED")); if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) wakeup(&fp->f_vnread_flags); fp->f_vnread_flags = 0; } mtx_unlock(mtxp); } #endif void foffset_lock_uio(struct file *fp, struct uio *uio, int flags) { if ((flags & FOF_OFFSET) == 0) uio->uio_offset = foffset_lock(fp, flags); } void foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) { if ((flags & FOF_OFFSET) == 0) foffset_unlock(fp, uio->uio_offset, flags); } static int get_advice(struct file *fp, struct uio *uio) { struct mtx *mtxp; int ret; ret = POSIX_FADV_NORMAL; if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) return (ret); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if (fp->f_advice != NULL && uio->uio_offset >= fp->f_advice->fa_start && uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) ret = fp->f_advice->fa_advice; mtx_unlock(mtxp); return (ret); } /* * File table vnode read routine. */ static int vn_read(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct vnode *vp; off_t orig_offset; int error, ioflag; int advice; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); vp = fp->f_vnode; ioflag = 0; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; advice = get_advice(fp, uio); vn_lock(vp, LK_SHARED | LK_RETRY); switch (advice) { case POSIX_FADV_NORMAL: case POSIX_FADV_SEQUENTIAL: case POSIX_FADV_NOREUSE: ioflag |= sequential_heuristic(uio, fp); break; case POSIX_FADV_RANDOM: /* Disable read-ahead for random I/O. */ break; } orig_offset = uio->uio_offset; #ifdef MAC error = mac_vnode_check_read(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_READ(vp, uio, ioflag, fp->f_cred); fp->f_nextoff[UIO_READ] = uio->uio_offset; VOP_UNLOCK(vp); if (error == 0 && advice == POSIX_FADV_NOREUSE && orig_offset != uio->uio_offset) /* * Use POSIX_FADV_DONTNEED to flush pages and buffers * for the backing file after a POSIX_FADV_NOREUSE * read(2). */ error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, POSIX_FADV_DONTNEED); return (error); } /* * File table vnode write routine. */ static int vn_write(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { struct vnode *vp; struct mount *mp; off_t orig_offset; int error, ioflag, lock_flags; int advice; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); vp = fp->f_vnode; if (vp->v_type == VREG) bwillwrite(); ioflag = IO_UNIT; if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) ioflag |= IO_APPEND; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; if ((fp->f_flag & O_FSYNC) || (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) ioflag |= IO_SYNC; mp = NULL; if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto unlock; advice = get_advice(fp, uio); if (MNT_SHARED_WRITES(mp) || (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { lock_flags = LK_SHARED; } else { lock_flags = LK_EXCLUSIVE; } vn_lock(vp, lock_flags | LK_RETRY); switch (advice) { case POSIX_FADV_NORMAL: case POSIX_FADV_SEQUENTIAL: case POSIX_FADV_NOREUSE: ioflag |= sequential_heuristic(uio, fp); break; case POSIX_FADV_RANDOM: /* XXX: Is this correct? */ break; } orig_offset = uio->uio_offset; #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); fp->f_nextoff[UIO_WRITE] = uio->uio_offset; VOP_UNLOCK(vp); if (vp->v_type != VCHR) vn_finished_write(mp); if (error == 0 && advice == POSIX_FADV_NOREUSE && orig_offset != uio->uio_offset) /* * Use POSIX_FADV_DONTNEED to flush pages and buffers * for the backing file after a POSIX_FADV_NOREUSE * write(2). */ error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, POSIX_FADV_DONTNEED); unlock: return (error); } /* * The vn_io_fault() is a wrapper around vn_read() and vn_write() to * prevent the following deadlock: * * Assume that the thread A reads from the vnode vp1 into userspace * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is * currently not resident, then system ends up with the call chain * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) * which establishes lock order vp1->vn_lock, then vp2->vn_lock. * If, at the same time, thread B reads from vnode vp2 into buffer buf2 * backed by the pages of vnode vp1, and some page in buf2 is not * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. * * To prevent the lock order reversal and deadlock, vn_io_fault() does * not allow page faults to happen during VOP_READ() or VOP_WRITE(). * Instead, it first tries to do the whole range i/o with pagefaults * disabled. If all pages in the i/o buffer are resident and mapped, * VOP will succeed (ignoring the genuine filesystem errors). * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do * i/o in chunks, with all pages in the chunk prefaulted and held * using vm_fault_quick_hold_pages(). * * Filesystems using this deadlock avoidance scheme should use the * array of the held pages from uio, saved in the curthread->td_ma, * instead of doing uiomove(). A helper function * vn_io_fault_uiomove() converts uiomove request into * uiomove_fromphys() over td_ma array. * * Since vnode locks do not cover the whole i/o anymore, rangelocks * make the current i/o request atomic with respect to other i/os and * truncations. */ /* * Decode vn_io_fault_args and perform the corresponding i/o. */ static int vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, struct thread *td) { int error, save; error = 0; save = vm_fault_disable_pagefaults(); switch (args->kind) { case VN_IO_FAULT_FOP: error = (args->args.fop_args.doio)(args->args.fop_args.fp, uio, args->cred, args->flags, td); break; case VN_IO_FAULT_VOP: if (uio->uio_rw == UIO_READ) { error = VOP_READ(args->args.vop_args.vp, uio, args->flags, args->cred); } else if (uio->uio_rw == UIO_WRITE) { error = VOP_WRITE(args->args.vop_args.vp, uio, args->flags, args->cred); } break; default: panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind, uio->uio_rw); } vm_fault_enable_pagefaults(save); return (error); } static int vn_io_fault_touch(char *base, const struct uio *uio) { int r; r = fubyte(base); if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) return (EFAULT); return (0); } static int vn_io_fault_prefault_user(const struct uio *uio) { char *base; const struct iovec *iov; size_t len; ssize_t resid; int error, i; KASSERT(uio->uio_segflg == UIO_USERSPACE, ("vn_io_fault_prefault userspace")); error = i = 0; iov = uio->uio_iov; resid = uio->uio_resid; base = iov->iov_base; len = iov->iov_len; while (resid > 0) { error = vn_io_fault_touch(base, uio); if (error != 0) break; if (len < PAGE_SIZE) { if (len != 0) { error = vn_io_fault_touch(base + len - 1, uio); if (error != 0) break; resid -= len; } if (++i >= uio->uio_iovcnt) break; iov = uio->uio_iov + i; base = iov->iov_base; len = iov->iov_len; } else { len -= PAGE_SIZE; base += PAGE_SIZE; resid -= PAGE_SIZE; } } return (error); } /* * Common code for vn_io_fault(), agnostic to the kind of i/o request. * Uses vn_io_fault_doio() to make the call to an actual i/o function. * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request * into args and call vn_io_fault1() to handle faults during the user * mode buffer accesses. */ static int vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, struct thread *td) { vm_page_t ma[io_hold_cnt + 2]; struct uio *uio_clone, short_uio; struct iovec short_iovec[1]; vm_page_t *prev_td_ma; vm_prot_t prot; vm_offset_t addr, end; size_t len, resid; ssize_t adv; int error, cnt, saveheld, prev_td_ma_cnt; if (vn_io_fault_prefault) { error = vn_io_fault_prefault_user(uio); if (error != 0) return (error); /* Or ignore ? */ } prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; /* * The UFS follows IO_UNIT directive and replays back both * uio_offset and uio_resid if an error is encountered during the * operation. But, since the iovec may be already advanced, * uio is still in an inconsistent state. * * Cache a copy of the original uio, which is advanced to the redo * point using UIO_NOCOPY below. */ uio_clone = cloneuio(uio); resid = uio->uio_resid; short_uio.uio_segflg = UIO_USERSPACE; short_uio.uio_rw = uio->uio_rw; short_uio.uio_td = uio->uio_td; error = vn_io_fault_doio(args, uio, td); if (error != EFAULT) goto out; atomic_add_long(&vn_io_faults_cnt, 1); uio_clone->uio_segflg = UIO_NOCOPY; uiomove(NULL, resid - uio->uio_resid, uio_clone); uio_clone->uio_segflg = uio->uio_segflg; saveheld = curthread_pflags_set(TDP_UIOHELD); prev_td_ma = td->td_ma; prev_td_ma_cnt = td->td_ma_cnt; while (uio_clone->uio_resid != 0) { len = uio_clone->uio_iov->iov_len; if (len == 0) { KASSERT(uio_clone->uio_iovcnt >= 1, ("iovcnt underflow")); uio_clone->uio_iov++; uio_clone->uio_iovcnt--; continue; } if (len > io_hold_cnt * PAGE_SIZE) len = io_hold_cnt * PAGE_SIZE; addr = (uintptr_t)uio_clone->uio_iov->iov_base; end = round_page(addr + len); if (end < addr) { error = EFAULT; break; } cnt = atop(end - trunc_page(addr)); /* * A perfectly misaligned address and length could cause * both the start and the end of the chunk to use partial * page. +2 accounts for such a situation. */ cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, addr, len, prot, ma, io_hold_cnt + 2); if (cnt == -1) { error = EFAULT; break; } short_uio.uio_iov = &short_iovec[0]; short_iovec[0].iov_base = (void *)addr; short_uio.uio_iovcnt = 1; short_uio.uio_resid = short_iovec[0].iov_len = len; short_uio.uio_offset = uio_clone->uio_offset; td->td_ma = ma; td->td_ma_cnt = cnt; error = vn_io_fault_doio(args, &short_uio, td); vm_page_unhold_pages(ma, cnt); adv = len - short_uio.uio_resid; uio_clone->uio_iov->iov_base = (char *)uio_clone->uio_iov->iov_base + adv; uio_clone->uio_iov->iov_len -= adv; uio_clone->uio_resid -= adv; uio_clone->uio_offset += adv; uio->uio_resid -= adv; uio->uio_offset += adv; if (error != 0 || adv == 0) break; } td->td_ma = prev_td_ma; td->td_ma_cnt = prev_td_ma_cnt; curthread_pflags_restore(saveheld); out: free(uio_clone, M_IOV); return (error); } static int vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { fo_rdwr_t *doio; struct vnode *vp; void *rl_cookie; struct vn_io_fault_args args; int error; doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; vp = fp->f_vnode; /* * The ability to read(2) on a directory has historically been * allowed for all users, but this can and has been the source of * at least one security issue in the past. As such, it is now hidden * away behind a sysctl for those that actually need it to use it, and * restricted to root when it's turned on to make it relatively safe to * leave on for longer sessions of need. */ if (vp->v_type == VDIR) { KASSERT(uio->uio_rw == UIO_READ, ("illegal write attempted on a directory")); if (!vfs_allow_read_dir) return (EISDIR); if ((error = priv_check(td, PRIV_VFS_READ_DIR)) != 0) return (EISDIR); } foffset_lock_uio(fp, uio, flags); if (do_vn_io_fault(vp, uio)) { args.kind = VN_IO_FAULT_FOP; args.args.fop_args.fp = fp; args.args.fop_args.doio = doio; args.cred = active_cred; args.flags = flags | FOF_OFFSET; if (uio->uio_rw == UIO_READ) { rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, uio->uio_offset + uio->uio_resid); } else if ((fp->f_flag & O_APPEND) != 0 || (flags & FOF_OFFSET) == 0) { /* For appenders, punt and lock the whole range. */ rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); } else { rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, uio->uio_offset + uio->uio_resid); } error = vn_io_fault1(vp, uio, &args, td); vn_rangelock_unlock(vp, rl_cookie); } else { error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); } foffset_unlock_uio(fp, uio, flags); return (error); } /* * Helper function to perform the requested uiomove operation using * the held pages for io->uio_iov[0].iov_base buffer instead of * copyin/copyout. Access to the pages with uiomove_fromphys() * instead of iov_base prevents page faults that could occur due to * pmap_collect() invalidating the mapping created by * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or * object cleanup revoking the write access from page mappings. * * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() * instead of plain uiomove(). */ int vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) { struct uio transp_uio; struct iovec transp_iov[1]; struct thread *td; size_t adv; int error, pgadv; td = curthread; if ((td->td_pflags & TDP_UIOHELD) == 0 || uio->uio_segflg != UIO_USERSPACE) return (uiomove(data, xfersize, uio)); KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); transp_iov[0].iov_base = data; transp_uio.uio_iov = &transp_iov[0]; transp_uio.uio_iovcnt = 1; if (xfersize > uio->uio_resid) xfersize = uio->uio_resid; transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; transp_uio.uio_offset = 0; transp_uio.uio_segflg = UIO_SYSSPACE; /* * Since transp_iov points to data, and td_ma page array * corresponds to original uio->uio_iov, we need to invert the * direction of the i/o operation as passed to * uiomove_fromphys(). */ switch (uio->uio_rw) { case UIO_WRITE: transp_uio.uio_rw = UIO_READ; break; case UIO_READ: transp_uio.uio_rw = UIO_WRITE; break; } transp_uio.uio_td = uio->uio_td; error = uiomove_fromphys(td->td_ma, ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, xfersize, &transp_uio); adv = xfersize - transp_uio.uio_resid; pgadv = (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); td->td_ma += pgadv; KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, pgadv)); td->td_ma_cnt -= pgadv; uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; uio->uio_iov->iov_len -= adv; uio->uio_resid -= adv; uio->uio_offset += adv; return (error); } int vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, struct uio *uio) { struct thread *td; vm_offset_t iov_base; int cnt, pgadv; td = curthread; if ((td->td_pflags & TDP_UIOHELD) == 0 || uio->uio_segflg != UIO_USERSPACE) return (uiomove_fromphys(ma, offset, xfersize, uio)); KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; iov_base = (vm_offset_t)uio->uio_iov->iov_base; switch (uio->uio_rw) { case UIO_WRITE: pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, offset, cnt); break; case UIO_READ: pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, cnt); break; } pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); td->td_ma += pgadv; KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, pgadv)); td->td_ma_cnt -= pgadv; uio->uio_iov->iov_base = (char *)(iov_base + cnt); uio->uio_iov->iov_len -= cnt; uio->uio_resid -= cnt; uio->uio_offset += cnt; return (0); } /* * File table truncate routine. */ static int vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct mount *mp; struct vnode *vp; void *rl_cookie; int error; vp = fp->f_vnode; /* * Lock the whole range for truncation. Otherwise split i/o * might happen partly before and partly after the truncation. */ rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error) goto out1; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); if (vp->v_type == VDIR) { error = EISDIR; goto out; } #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error) goto out; #endif error = vn_truncate_locked(vp, length, (fp->f_flag & O_FSYNC) != 0, fp->f_cred); out: VOP_UNLOCK(vp); vn_finished_write(mp); out1: vn_rangelock_unlock(vp, rl_cookie); return (error); } /* * Truncate a file that is already locked. */ int vn_truncate_locked(struct vnode *vp, off_t length, bool sync, struct ucred *cred) { struct vattr vattr; int error; error = VOP_ADD_WRITECOUNT(vp, 1); if (error == 0) { VATTR_NULL(&vattr); vattr.va_size = length; if (sync) vattr.va_vaflags |= VA_SYNC; error = VOP_SETATTR(vp, &vattr, cred); VOP_ADD_WRITECOUNT_CHECKED(vp, -1); } return (error); } /* * File table vnode stat routine. */ static int vn_statfile(struct file *fp, struct stat *sb, struct ucred *active_cred, struct thread *td) { struct vnode *vp = fp->f_vnode; int error; vn_lock(vp, LK_SHARED | LK_RETRY); error = vn_stat(vp, sb, active_cred, fp->f_cred, td); VOP_UNLOCK(vp); return (error); } /* * Stat a vnode; implementation for the stat syscall */ int vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred, struct ucred *file_cred, struct thread *td) { struct vattr vattr; struct vattr *vap; int error; u_short mode; AUDIT_ARG_VNODE1(vp); #ifdef MAC error = mac_vnode_check_stat(active_cred, file_cred, vp); if (error) return (error); #endif vap = &vattr; /* * Initialize defaults for new and unusual fields, so that file * systems which don't support these fields don't need to know * about them. */ vap->va_birthtime.tv_sec = -1; vap->va_birthtime.tv_nsec = 0; vap->va_fsid = VNOVAL; vap->va_rdev = NODEV; error = VOP_GETATTR(vp, vap, active_cred); if (error) return (error); /* * Zero the spare stat fields */ bzero(sb, sizeof *sb); /* * Copy from vattr table */ if (vap->va_fsid != VNOVAL) sb->st_dev = vap->va_fsid; else sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0]; sb->st_ino = vap->va_fileid; mode = vap->va_mode; switch (vap->va_type) { case VREG: mode |= S_IFREG; break; case VDIR: mode |= S_IFDIR; break; case VBLK: mode |= S_IFBLK; break; case VCHR: mode |= S_IFCHR; break; case VLNK: mode |= S_IFLNK; break; case VSOCK: mode |= S_IFSOCK; break; case VFIFO: mode |= S_IFIFO; break; default: return (EBADF); } sb->st_mode = mode; sb->st_nlink = vap->va_nlink; sb->st_uid = vap->va_uid; sb->st_gid = vap->va_gid; sb->st_rdev = vap->va_rdev; if (vap->va_size > OFF_MAX) return (EOVERFLOW); sb->st_size = vap->va_size; sb->st_atim.tv_sec = vap->va_atime.tv_sec; sb->st_atim.tv_nsec = vap->va_atime.tv_nsec; sb->st_mtim.tv_sec = vap->va_mtime.tv_sec; sb->st_mtim.tv_nsec = vap->va_mtime.tv_nsec; sb->st_ctim.tv_sec = vap->va_ctime.tv_sec; sb->st_ctim.tv_nsec = vap->va_ctime.tv_nsec; sb->st_birthtim.tv_sec = vap->va_birthtime.tv_sec; sb->st_birthtim.tv_nsec = vap->va_birthtime.tv_nsec; - /* + /* * According to www.opengroup.org, the meaning of st_blksize is * "a filesystem-specific preferred I/O block size for this * object. In some filesystem types, this may vary from file * to file" * Use minimum/default of PAGE_SIZE (e.g. for VCHR). */ sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize); - + sb->st_flags = vap->va_flags; if (priv_check_cred_vfs_generation(td->td_ucred)) sb->st_gen = 0; else sb->st_gen = vap->va_gen; sb->st_blocks = vap->va_bytes / S_BLKSIZE; return (0); } /* * File table vnode ioctl routine. */ static int vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { struct vattr vattr; struct vnode *vp; struct fiobmap2_arg *bmarg; int error; vp = fp->f_vnode; switch (vp->v_type) { case VDIR: case VREG: switch (com) { case FIONREAD: vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, active_cred); VOP_UNLOCK(vp); if (error == 0) *(int *)data = vattr.va_size - fp->f_offset; return (error); case FIOBMAP2: bmarg = (struct fiobmap2_arg *)data; vn_lock(vp, LK_SHARED | LK_RETRY); #ifdef MAC error = mac_vnode_check_read(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_BMAP(vp, bmarg->bn, NULL, &bmarg->bn, &bmarg->runp, &bmarg->runb); VOP_UNLOCK(vp); return (error); case FIONBIO: case FIOASYNC: return (0); default: return (VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td)); } break; case VCHR: return (VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td)); default: return (ENOTTY); } } /* * File table vnode poll routine. */ static int vn_poll(struct file *fp, int events, struct ucred *active_cred, struct thread *td) { struct vnode *vp; int error; vp = fp->f_vnode; #ifdef MAC vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); VOP_UNLOCK(vp); if (!error) #endif error = VOP_POLL(vp, events, fp->f_cred, td); return (error); } /* * Acquire the requested lock and then check for validity. LK_RETRY * permits vn_lock to return doomed vnodes. */ static int __noinline _vn_lock_fallback(struct vnode *vp, int flags, const char *file, int line, int error) { KASSERT((flags & LK_RETRY) == 0 || error == 0, ("vn_lock: error %d incompatible with flags %#x", error, flags)); if (error == 0) VNASSERT(VN_IS_DOOMED(vp), vp, ("vnode not doomed")); if ((flags & LK_RETRY) == 0) { if (error == 0) { VOP_UNLOCK(vp); error = ENOENT; } return (error); } /* * LK_RETRY case. * * Nothing to do if we got the lock. */ if (error == 0) return (0); /* * Interlock was dropped by the call in _vn_lock. */ flags &= ~LK_INTERLOCK; do { error = VOP_LOCK1(vp, flags, file, line); } while (error != 0); return (0); } int _vn_lock(struct vnode *vp, int flags, const char *file, int line) { int error; VNASSERT((flags & LK_TYPE_MASK) != 0, vp, ("vn_lock: no locktype (%d passed)", flags)); VNPASS(vp->v_holdcnt > 0, vp); error = VOP_LOCK1(vp, flags, file, line); if (__predict_false(error != 0 || VN_IS_DOOMED(vp))) return (_vn_lock_fallback(vp, flags, file, line, error)); return (0); } /* * File table vnode close routine. */ static int vn_closefile(struct file *fp, struct thread *td) { struct vnode *vp; struct flock lf; int error; bool ref; vp = fp->f_vnode; fp->f_ops = &badfileops; ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); if (__predict_false(ref)) { lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; lf.l_type = F_UNLCK; (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); vrele(vp); } return (error); } /* * Preparing to start a filesystem write operation. If the operation is * permitted, then we bump the count of operations in progress and * proceed. If a suspend request is in progress, we wait until the * suspension is over, and then proceed. */ static int vn_start_write_refed(struct mount *mp, int flags, bool mplocked) { int error, mflags; if (__predict_true(!mplocked) && (flags & V_XSLEEP) == 0 && vfs_op_thread_enter(mp)) { MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); vfs_mp_count_add_pcpu(mp, writeopcount, 1); vfs_op_thread_exit(mp); return (0); } if (mplocked) mtx_assert(MNT_MTX(mp), MA_OWNED); else MNT_ILOCK(mp); error = 0; /* * Check on status of suspension. */ if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || mp->mnt_susp_owner != curthread) { mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0) | (PUSER - 1); while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { if (flags & V_NOWAIT) { error = EWOULDBLOCK; goto unlock; } error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0); if (error) goto unlock; } } if (flags & V_XSLEEP) goto unlock; mp->mnt_writeopcount++; unlock: if (error != 0 || (flags & V_XSLEEP) != 0) MNT_REL(mp); MNT_IUNLOCK(mp); return (error); } int vn_start_write(struct vnode *vp, struct mount **mpp, int flags) { struct mount *mp; int error; KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), ("V_MNTREF requires mp")); error = 0; /* * If a vnode is provided, get and return the mount point that * to which it will write. */ if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } if ((mp = *mpp) == NULL) return (0); /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ if (vp == NULL && (flags & V_MNTREF) == 0) vfs_ref(mp); return (vn_start_write_refed(mp, flags, false)); } /* * Secondary suspension. Used by operations such as vop_inactive * routines that are needed by the higher level functions. These * are allowed to proceed until all the higher level functions have * completed (indicated by mnt_writeopcount dropping to zero). At that * time, these operations are halted until the suspension is over. */ int vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) { struct mount *mp; int error; KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), ("V_MNTREF requires mp")); retry: if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } /* * If we are not suspended or have not yet reached suspended * mode, then let the operation proceed. */ if ((mp = *mpp) == NULL) return (0); /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ MNT_ILOCK(mp); if (vp == NULL && (flags & V_MNTREF) == 0) MNT_REF(mp); if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { mp->mnt_secondary_writes++; mp->mnt_secondary_accwrites++; MNT_IUNLOCK(mp); return (0); } if (flags & V_NOWAIT) { MNT_REL(mp); MNT_IUNLOCK(mp); return (EWOULDBLOCK); } /* * Wait for the suspension to finish. */ error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), "suspfs", 0); vfs_rel(mp); if (error == 0) goto retry; return (error); } /* * Filesystem write operation has completed. If we are suspending and this * operation is the last one, notify the suspender that the suspension is * now in effect. */ void vn_finished_write(struct mount *mp) { int c; if (mp == NULL) return; if (vfs_op_thread_enter(mp)) { vfs_mp_count_sub_pcpu(mp, writeopcount, 1); vfs_mp_count_sub_pcpu(mp, ref, 1); vfs_op_thread_exit(mp); return; } MNT_ILOCK(mp); vfs_assert_mount_counters(mp); MNT_REL(mp); c = --mp->mnt_writeopcount; if (mp->mnt_vfs_ops == 0) { MPASS((mp->mnt_kern_flag & MNTK_SUSPEND) == 0); MNT_IUNLOCK(mp); return; } if (c < 0) vfs_dump_mount_counters(mp); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && c == 0) wakeup(&mp->mnt_writeopcount); MNT_IUNLOCK(mp); } /* * Filesystem secondary write operation has completed. If we are * suspending and this operation is the last one, notify the suspender * that the suspension is now in effect. */ void vn_finished_secondary_write(struct mount *mp) { if (mp == NULL) return; MNT_ILOCK(mp); MNT_REL(mp); mp->mnt_secondary_writes--; if (mp->mnt_secondary_writes < 0) panic("vn_finished_secondary_write: neg cnt"); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && mp->mnt_secondary_writes <= 0) wakeup(&mp->mnt_secondary_writes); MNT_IUNLOCK(mp); } /* * Request a filesystem to suspend write operations. */ int vfs_write_suspend(struct mount *mp, int flags) { int error; vfs_op_enter(mp); MNT_ILOCK(mp); vfs_assert_mount_counters(mp); if (mp->mnt_susp_owner == curthread) { vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); return (EALREADY); } while (mp->mnt_kern_flag & MNTK_SUSPEND) msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); /* * Unmount holds a write reference on the mount point. If we * own busy reference and drain for writers, we deadlock with * the reference draining in the unmount path. Callers of * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if * vfs_busy() reference is owned and caller is not in the * unmount context. */ if ((flags & VS_SKIP_UNMOUNT) != 0 && (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { vfs_op_exit_locked(mp); MNT_IUNLOCK(mp); return (EBUSY); } mp->mnt_kern_flag |= MNTK_SUSPEND; mp->mnt_susp_owner = curthread; if (mp->mnt_writeopcount > 0) (void) msleep(&mp->mnt_writeopcount, MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); else MNT_IUNLOCK(mp); if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) { vfs_write_resume(mp, 0); /* vfs_write_resume does vfs_op_exit() for us */ } return (error); } /* * Request a filesystem to resume write operations. */ void vfs_write_resume(struct mount *mp, int flags) { MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | MNTK_SUSPENDED); mp->mnt_susp_owner = NULL; wakeup(&mp->mnt_writeopcount); wakeup(&mp->mnt_flag); curthread->td_pflags &= ~TDP_IGNSUSP; if ((flags & VR_START_WRITE) != 0) { MNT_REF(mp); mp->mnt_writeopcount++; } MNT_IUNLOCK(mp); if ((flags & VR_NO_SUSPCLR) == 0) VFS_SUSP_CLEAN(mp); vfs_op_exit(mp); } else if ((flags & VR_START_WRITE) != 0) { MNT_REF(mp); vn_start_write_refed(mp, 0, true); } else { MNT_IUNLOCK(mp); } } /* * Helper loop around vfs_write_suspend() for filesystem unmount VFS * methods. */ int vfs_write_suspend_umnt(struct mount *mp) { int error; KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, ("vfs_write_suspend_umnt: recursed")); /* dounmount() already called vn_start_write(). */ for (;;) { vn_finished_write(mp); error = vfs_write_suspend(mp, 0); if (error != 0) { vn_start_write(NULL, &mp, V_WAIT); return (error); } MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) break; MNT_IUNLOCK(mp); vn_start_write(NULL, &mp, V_WAIT); } mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); wakeup(&mp->mnt_flag); MNT_IUNLOCK(mp); curthread->td_pflags |= TDP_IGNSUSP; return (0); } /* * Implement kqueues for files by translating it to vnode operation. */ static int vn_kqfilter(struct file *fp, struct knote *kn) { return (VOP_KQFILTER(fp->f_vnode, kn)); } /* * Simplified in-kernel wrapper calls for extended attribute access. * Both calls pass in a NULL credential, authorizing as "kernel" access. * Set IO_NODELOCKED in ioflg if the vnode is already locked. */ int vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int *buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; int error; iov.iov_len = *buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = *buflen; if ((ioflg & IO_NODELOCKED) == 0) vn_lock(vp, LK_SHARED | LK_RETRY); ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute retrieval as kernel */ error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) VOP_UNLOCK(vp); if (error == 0) { *buflen = *buflen - auio.uio_resid; } return (error); } /* * XXX failure mode if partially written? */ int vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; struct mount *mp; int error; iov.iov_len = buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_WRITE; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = buflen; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute setting as kernel */ error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp); } return (error); } int vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, struct thread *td) { struct mount *mp; int error; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute removal as kernel */ error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); if (error == EOPNOTSUPP) error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp); } return (error); } static int vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, struct vnode **rvp) { return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); } int vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) { return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, lkflags, rvp)); } int vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, int lkflags, struct vnode **rvp) { struct mount *mp; int ltype, error; ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); mp = vp->v_mount; ltype = VOP_ISLOCKED(vp); KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, ("vn_vget_ino: vp not locked")); error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { vfs_ref(mp); VOP_UNLOCK(vp); error = vfs_busy(mp, 0); vn_lock(vp, ltype | LK_RETRY); vfs_rel(mp); if (error != 0) return (ENOENT); if (VN_IS_DOOMED(vp)) { vfs_unbusy(mp); return (ENOENT); } } VOP_UNLOCK(vp); error = alloc(mp, alloc_arg, lkflags, rvp); vfs_unbusy(mp); if (error != 0 || *rvp != vp) vn_lock(vp, ltype | LK_RETRY); if (VN_IS_DOOMED(vp)) { if (error == 0) { if (*rvp == vp) vunref(vp); else vput(*rvp); } error = ENOENT; } return (error); } int vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, struct thread *td) { if (vp->v_type != VREG || td == NULL) return (0); if ((uoff_t)uio->uio_offset + uio->uio_resid > lim_cur(td, RLIMIT_FSIZE)) { PROC_LOCK(td->td_proc); kern_psignal(td->td_proc, SIGXFSZ); PROC_UNLOCK(td->td_proc); return (EFBIG); } return (0); } int vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct vnode *vp; vp = fp->f_vnode; #ifdef AUDIT vn_lock(vp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(vp); VOP_UNLOCK(vp); #endif return (setfmode(td, active_cred, vp, mode)); } int vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct vnode *vp; vp = fp->f_vnode; #ifdef AUDIT vn_lock(vp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(vp); VOP_UNLOCK(vp); #endif return (setfown(td, active_cred, vp, uid, gid)); } void vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) { vm_object_t object; if ((object = vp->v_object) == NULL) return; VM_OBJECT_WLOCK(object); vm_object_page_remove(object, start, end, 0); VM_OBJECT_WUNLOCK(object); } int vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) { struct vattr va; daddr_t bn, bnp; uint64_t bsize; off_t noff; int error; KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, ("Wrong command %lu", cmd)); if (vn_lock(vp, LK_SHARED) != 0) return (EBADF); if (vp->v_type != VREG) { error = ENOTTY; goto unlock; } error = VOP_GETATTR(vp, &va, cred); if (error != 0) goto unlock; noff = *off; if (noff >= va.va_size) { error = ENXIO; goto unlock; } bsize = vp->v_mount->mnt_stat.f_iosize; for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize - noff % bsize) { error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); if (error == EOPNOTSUPP) { error = ENOTTY; goto unlock; } if ((bnp == -1 && cmd == FIOSEEKHOLE) || (bnp != -1 && cmd == FIOSEEKDATA)) { noff = bn * bsize; if (noff < *off) noff = *off; goto unlock; } } if (noff > va.va_size) noff = va.va_size; /* noff == va.va_size. There is an implicit hole at the end of file. */ if (cmd == FIOSEEKDATA) error = ENXIO; unlock: VOP_UNLOCK(vp); if (error == 0) *off = noff; return (error); } int vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct ucred *cred; struct vnode *vp; struct vattr vattr; off_t foffset, size; int error, noneg; cred = td->td_ucred; vp = fp->f_vnode; foffset = foffset_lock(fp, 0); noneg = (vp->v_type != VCHR); error = 0; switch (whence) { case L_INCR: if (noneg && (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset))) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, cred); VOP_UNLOCK(vp); if (error) break; /* * If the file references a disk device, then fetch * the media size and use that to determine the ending * offset. */ if (vattr.va_size == 0 && vp->v_type == VCHR && fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) vattr.va_size = size; if (noneg && (vattr.va_size > OFF_MAX || (offset > 0 && vattr.va_size > OFF_MAX - offset))) { error = EOVERFLOW; break; } offset += vattr.va_size; break; case L_SET: break; case SEEK_DATA: error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); if (error == ENOTTY) error = EINVAL; break; case SEEK_HOLE: error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); if (error == ENOTTY) error = EINVAL; break; default: error = EINVAL; } if (error == 0 && noneg && offset < 0) error = EINVAL; if (error != 0) goto drop; VFS_KNOTE_UNLOCKED(vp, 0); td->td_uretoff.tdu_off = offset; drop: foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } int vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *td) { int error; /* * Grant permission if the caller is the owner of the file, or * the super-user, or has ACL_WRITE_ATTRIBUTES permission on * on the file. If the time pointer is null, then write * permission on the file is also sufficient. * * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES * will be allowed to set the times [..] to the current * server time. */ error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) error = VOP_ACCESS(vp, VWRITE, cred, td); return (error); } int vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) { struct vnode *vp; int error; if (fp->f_type == DTYPE_FIFO) kif->kf_type = KF_TYPE_FIFO; else kif->kf_type = KF_TYPE_VNODE; vp = fp->f_vnode; vref(vp); FILEDESC_SUNLOCK(fdp); error = vn_fill_kinfo_vnode(vp, kif); vrele(vp); FILEDESC_SLOCK(fdp); return (error); } static inline void vn_fill_junk(struct kinfo_file *kif) { size_t len, olen; /* * Simulate vn_fullpath returning changing values for a given * vp during e.g. coredump. */ len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; olen = strlen(kif->kf_path); if (len < olen) strcpy(&kif->kf_path[len - 1], "$"); else for (; olen < len; olen++) strcpy(&kif->kf_path[olen], "A"); } int vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) { struct vattr va; char *fullpath, *freepath; int error; kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); freepath = NULL; fullpath = "-"; error = vn_fullpath(curthread, vp, &fullpath, &freepath); if (error == 0) { strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); } if (freepath != NULL) free(freepath, M_TEMP); KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, vn_fill_junk(kif); ); /* * Retrieve vnode attributes. */ va.va_fsid = VNOVAL; va.va_rdev = NODEV; vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &va, curthread->td_ucred); VOP_UNLOCK(vp); if (error != 0) return (error); if (va.va_fsid != VNOVAL) kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; else kif->kf_un.kf_file.kf_file_fsid = vp->v_mount->mnt_stat.f_fsid.val[0]; kif->kf_un.kf_file.kf_file_fsid_freebsd11 = kif->kf_un.kf_file.kf_file_fsid; /* truncate */ kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); kif->kf_un.kf_file.kf_file_size = va.va_size; kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; kif->kf_un.kf_file.kf_file_rdev_freebsd11 = kif->kf_un.kf_file.kf_file_rdev; /* truncate */ return (0); } int vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { #ifdef HWPMC_HOOKS struct pmckern_map_in pkm; #endif struct mount *mp; struct vnode *vp; vm_object_t object; vm_prot_t maxprot; boolean_t writecounted; int error; #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) /* * POSIX shared-memory objects are defined to have * kernel persistence, and are not defined to support * read(2)/write(2) -- or even open(2). Thus, we can * use MAP_ASYNC to trade on-disk coherence for speed. * The shm_open(3) library routine turns on the FPOSIXSHM * flag to request this behavior. */ if ((fp->f_flag & FPOSIXSHM) != 0) flags |= MAP_NOSYNC; #endif vp = fp->f_vnode; /* * Ensure that file and memory protections are * compatible. Note that we only worry about * writability if mapping is shared; in this case, * current and max prot are dictated by the open file. * XXX use the vnode instead? Problem is: what * credentials do we use for determination? What if * proc does a setuid? */ mp = vp->v_mount; if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { maxprot = VM_PROT_NONE; if ((prot & VM_PROT_EXECUTE) != 0) return (EACCES); } else maxprot = VM_PROT_EXECUTE; if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_READ; else if ((prot & VM_PROT_READ) != 0) return (EACCES); /* * If we are sharing potential changes via MAP_SHARED and we * are trying to get write permission although we opened it * without asking for it, bail out. */ if ((flags & MAP_SHARED) != 0) { if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; else if ((prot & VM_PROT_WRITE) != 0) return (EACCES); } else { maxprot |= VM_PROT_WRITE; cap_maxprot |= VM_PROT_WRITE; } maxprot &= cap_maxprot; /* * For regular files and shared memory, POSIX requires that * the value of foff be a legitimate offset within the data * object. In particular, negative offsets are invalid. * Blocking negative offsets and overflows here avoids * possible wraparound or user-level access into reserved * ranges of the data object later. In contrast, POSIX does * not dictate how offsets are used by device drivers, so in * the case of a device mapping a negative offset is passed * on. */ if ( #ifdef _LP64 size > OFF_MAX || #endif foff < 0 || foff > OFF_MAX - size) return (EINVAL); writecounted = FALSE; error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, &foff, &object, &writecounted); if (error != 0) return (error); error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, foff, writecounted, td); if (error != 0) { /* * If this mapping was accounted for in the vnode's * writecount, then undo that now. */ if (writecounted) vm_pager_release_writecount(object, 0, size); vm_object_deallocate(object); } #ifdef HWPMC_HOOKS /* Inform hwpmc(4) if an executable is being mapped. */ if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { pkm.pm_file = vp; pkm.pm_address = (uintptr_t) *addr; PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); } } #endif return (error); } void vn_fsid(struct vnode *vp, struct vattr *va) { fsid_t *f; f = &vp->v_mount->mnt_stat.f_fsid; va->va_fsid = (uint32_t)f->val[1]; va->va_fsid <<= sizeof(f->val[1]) * NBBY; va->va_fsid += (uint32_t)f->val[0]; } int vn_fsync_buf(struct vnode *vp, int waitfor) { struct buf *bp, *nbp; struct bufobj *bo; struct mount *mp; int error, maxretry; error = 0; maxretry = 10000; /* large, arbitrarily chosen */ mp = NULL; if (vp->v_type == VCHR) { VI_LOCK(vp); mp = vp->v_rdev->si_mountpt; VI_UNLOCK(vp); } bo = &vp->v_bufobj; BO_LOCK(bo); loop1: /* * MARK/SCAN initialization to avoid infinite loops. */ TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { bp->b_vflags &= ~BV_SCANNED; bp->b_error = 0; } /* * Flush all dirty buffers associated with a vnode. */ loop2: TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { if ((bp->b_vflags & BV_SCANNED) != 0) continue; bp->b_vflags |= BV_SCANNED; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL)) { if (waitfor != MNT_WAIT) continue; if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_INTERLOCK | LK_SLEEPFAIL, BO_LOCKPTR(bo)) != 0) { BO_LOCK(bo); goto loop1; } BO_LOCK(bo); } BO_UNLOCK(bo); KASSERT(bp->b_bufobj == bo, ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo)); if ((bp->b_flags & B_DELWRI) == 0) panic("fsync: not dirty"); if ((vp->v_object != NULL) && (bp->b_flags & B_CLUSTEROK)) { vfs_bio_awrite(bp); } else { bremfree(bp); bawrite(bp); } if (maxretry < 1000) pause("dirty", hz < 1000 ? 1 : hz / 1000); BO_LOCK(bo); goto loop2; } /* * If synchronous the caller expects us to completely resolve all * dirty buffers in the system. Wait for in-progress I/O to * complete (which could include background bitmap writes), then * retry if dirty blocks still exist. */ if (waitfor == MNT_WAIT) { bufobj_wwait(bo, 0, 0); if (bo->bo_dirty.bv_cnt > 0) { /* * If we are unable to write any of these buffers * then we fail now rather than trying endlessly * to write them out. */ TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) if ((error = bp->b_error) != 0) break; if ((mp != NULL && mp->mnt_secondary_writes > 0) || (error == 0 && --maxretry >= 0)) goto loop1; if (error == 0) error = EAGAIN; } } BO_UNLOCK(bo); if (error != 0) vn_printf(vp, "fsync: giving up on dirty (error = %d) ", error); return (error); } /* * Copies a byte range from invp to outvp. Calls VOP_COPY_FILE_RANGE() * or vn_generic_copy_file_range() after rangelocking the byte ranges, * to do the actual copy. * vn_generic_copy_file_range() is factored out, so it can be called * from a VOP_COPY_FILE_RANGE() call as well, but handles vnodes from * different file systems. */ int vn_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) { int error; size_t len; uint64_t uvalin, uvalout; len = *lenp; *lenp = 0; /* For error returns. */ error = 0; /* Do some sanity checks on the arguments. */ uvalin = *inoffp; uvalin += len; uvalout = *outoffp; uvalout += len; if (invp->v_type == VDIR || outvp->v_type == VDIR) error = EISDIR; else if (*inoffp < 0 || uvalin > INT64_MAX || uvalin < (uint64_t)*inoffp || *outoffp < 0 || uvalout > INT64_MAX || uvalout < (uint64_t)*outoffp || invp->v_type != VREG || outvp->v_type != VREG) error = EINVAL; if (error != 0) goto out; /* * If the two vnode are for the same file system, call * VOP_COPY_FILE_RANGE(), otherwise call vn_generic_copy_file_range() * which can handle copies across multiple file systems. */ *lenp = len; if (invp->v_mount == outvp->v_mount) error = VOP_COPY_FILE_RANGE(invp, inoffp, outvp, outoffp, lenp, flags, incred, outcred, fsize_td); else error = vn_generic_copy_file_range(invp, inoffp, outvp, outoffp, lenp, flags, incred, outcred, fsize_td); out: return (error); } /* * Test len bytes of data starting at dat for all bytes == 0. * Return true if all bytes are zero, false otherwise. * Expects dat to be well aligned. */ static bool mem_iszero(void *dat, int len) { int i; const u_int *p; const char *cp; for (p = dat; len > 0; len -= sizeof(*p), p++) { if (len >= sizeof(*p)) { if (*p != 0) return (false); } else { cp = (const char *)p; for (i = 0; i < len; i++, cp++) if (*cp != '\0') return (false); } } return (true); } /* * Look for a hole in the output file and, if found, adjust *outoffp * and *xferp to skip past the hole. * *xferp is the entire hole length to be written and xfer2 is how many bytes * to be written as 0's upon return. */ static off_t vn_skip_hole(struct vnode *outvp, off_t xfer2, off_t *outoffp, off_t *xferp, off_t *dataoffp, off_t *holeoffp, struct ucred *cred) { int error; off_t delta; if (*holeoffp == 0 || *holeoffp <= *outoffp) { *dataoffp = *outoffp; error = VOP_IOCTL(outvp, FIOSEEKDATA, dataoffp, 0, cred, curthread); if (error == 0) { *holeoffp = *dataoffp; error = VOP_IOCTL(outvp, FIOSEEKHOLE, holeoffp, 0, cred, curthread); } if (error != 0 || *holeoffp == *dataoffp) { /* * Since outvp is unlocked, it may be possible for * another thread to do a truncate(), lseek(), write() * creating a hole at startoff between the above * VOP_IOCTL() calls, if the other thread does not do * rangelocking. * If that happens, *holeoffp == *dataoffp and finding * the hole has failed, so disable vn_skip_hole(). */ *holeoffp = -1; /* Disable use of vn_skip_hole(). */ return (xfer2); } KASSERT(*dataoffp >= *outoffp, ("vn_skip_hole: dataoff=%jd < outoff=%jd", (intmax_t)*dataoffp, (intmax_t)*outoffp)); KASSERT(*holeoffp > *dataoffp, ("vn_skip_hole: holeoff=%jd <= dataoff=%jd", (intmax_t)*holeoffp, (intmax_t)*dataoffp)); } /* * If there is a hole before the data starts, advance *outoffp and * *xferp past the hole. */ if (*dataoffp > *outoffp) { delta = *dataoffp - *outoffp; if (delta >= *xferp) { /* Entire *xferp is a hole. */ *outoffp += *xferp; *xferp = 0; return (0); } *xferp -= delta; *outoffp += delta; xfer2 = MIN(xfer2, *xferp); } /* * If a hole starts before the end of this xfer2, reduce this xfer2 so * that the write ends at the start of the hole. * *holeoffp should always be greater than *outoffp, but for the * non-INVARIANTS case, check this to make sure xfer2 remains a sane * value. */ if (*holeoffp > *outoffp && *holeoffp < *outoffp + xfer2) xfer2 = *holeoffp - *outoffp; return (xfer2); } /* * Write an xfer sized chunk to outvp in blksize blocks from dat. * dat is a maximum of blksize in length and can be written repeatedly in * the chunk. * If growfile == true, just grow the file via vn_truncate_locked() instead * of doing actual writes. * If checkhole == true, a hole is being punched, so skip over any hole * already in the output file. */ static int vn_write_outvp(struct vnode *outvp, char *dat, off_t outoff, off_t xfer, u_long blksize, bool growfile, bool checkhole, struct ucred *cred) { struct mount *mp; off_t dataoff, holeoff, xfer2; int error, lckf; /* * Loop around doing writes of blksize until write has been completed. * Lock/unlock on each loop iteration so that a bwillwrite() can be * done for each iteration, since the xfer argument can be very * large if there is a large hole to punch in the output file. */ error = 0; holeoff = 0; do { xfer2 = MIN(xfer, blksize); if (checkhole) { /* * Punching a hole. Skip writing if there is * already a hole in the output file. */ xfer2 = vn_skip_hole(outvp, xfer2, &outoff, &xfer, &dataoff, &holeoff, cred); if (xfer == 0) break; if (holeoff < 0) checkhole = false; KASSERT(xfer2 > 0, ("vn_write_outvp: xfer2=%jd", (intmax_t)xfer2)); } bwillwrite(); mp = NULL; error = vn_start_write(outvp, &mp, V_WAIT); if (error == 0) { if (MNT_SHARED_WRITES(mp)) lckf = LK_SHARED; else lckf = LK_EXCLUSIVE; error = vn_lock(outvp, lckf); } if (error == 0) { if (growfile) error = vn_truncate_locked(outvp, outoff + xfer, false, cred); else { error = vn_rdwr(UIO_WRITE, outvp, dat, xfer2, outoff, UIO_SYSSPACE, IO_NODELOCKED, curthread->td_ucred, cred, NULL, curthread); outoff += xfer2; xfer -= xfer2; } VOP_UNLOCK(outvp); } if (mp != NULL) vn_finished_write(mp); } while (!growfile && xfer > 0 && error == 0); return (error); } /* * Copy a byte range of one file to another. This function can handle the * case where invp and outvp are on different file systems. * It can also be called by a VOP_COPY_FILE_RANGE() to do the work, if there * is no better file system specific way to do it. */ int vn_generic_copy_file_range(struct vnode *invp, off_t *inoffp, struct vnode *outvp, off_t *outoffp, size_t *lenp, unsigned int flags, struct ucred *incred, struct ucred *outcred, struct thread *fsize_td) { struct vattr va; struct mount *mp; struct uio io; off_t startoff, endoff, xfer, xfer2; u_long blksize; int error; bool cantseek, readzeros, eof, lastblock; ssize_t aresid; size_t copylen, len, savlen; char *dat; long holein, holeout; holein = holeout = 0; savlen = len = *lenp; error = 0; dat = NULL; error = vn_lock(invp, LK_SHARED); if (error != 0) goto out; if (VOP_PATHCONF(invp, _PC_MIN_HOLE_SIZE, &holein) != 0) holein = 0; VOP_UNLOCK(invp); mp = NULL; error = vn_start_write(outvp, &mp, V_WAIT); if (error == 0) error = vn_lock(outvp, LK_EXCLUSIVE); if (error == 0) { /* * If fsize_td != NULL, do a vn_rlimit_fsize() call, * now that outvp is locked. */ if (fsize_td != NULL) { io.uio_offset = *outoffp; io.uio_resid = len; error = vn_rlimit_fsize(outvp, &io, fsize_td); if (error != 0) error = EFBIG; } if (VOP_PATHCONF(outvp, _PC_MIN_HOLE_SIZE, &holeout) != 0) holeout = 0; /* * Holes that are past EOF do not need to be written as a block * of zero bytes. So, truncate the output file as far as * possible and then use va.va_size to decide if writing 0 * bytes is necessary in the loop below. */ if (error == 0) error = VOP_GETATTR(outvp, &va, outcred); if (error == 0 && va.va_size > *outoffp && va.va_size <= *outoffp + len) { #ifdef MAC error = mac_vnode_check_write(curthread->td_ucred, outcred, outvp); if (error == 0) #endif error = vn_truncate_locked(outvp, *outoffp, false, outcred); if (error == 0) va.va_size = *outoffp; } VOP_UNLOCK(outvp); } if (mp != NULL) vn_finished_write(mp); if (error != 0) goto out; /* * Set the blksize to the larger of the hole sizes for invp and outvp. * If hole sizes aren't available, set the blksize to the larger * f_iosize of invp and outvp. * This code expects the hole sizes and f_iosizes to be powers of 2. * This value is clipped at 4Kbytes and 1Mbyte. */ blksize = MAX(holein, holeout); if (blksize == 0) blksize = MAX(invp->v_mount->mnt_stat.f_iosize, outvp->v_mount->mnt_stat.f_iosize); if (blksize < 4096) blksize = 4096; else if (blksize > 1024 * 1024) blksize = 1024 * 1024; dat = malloc(blksize, M_TEMP, M_WAITOK); /* * If VOP_IOCTL(FIOSEEKHOLE) works for invp, use it and FIOSEEKDATA * to find holes. Otherwise, just scan the read block for all 0s * in the inner loop where the data copying is done. * Note that some file systems such as NFSv3, NFSv4.0 and NFSv4.1 may * support holes on the server, but do not support FIOSEEKHOLE. */ eof = false; while (len > 0 && error == 0 && !eof) { endoff = 0; /* To shut up compilers. */ cantseek = true; startoff = *inoffp; copylen = len; /* * Find the next data area. If there is just a hole to EOF, * FIOSEEKDATA should fail and then we drop down into the * inner loop and create the hole on the outvp file. * (I do not know if any file system will report a hole to * EOF via FIOSEEKHOLE, but I am pretty sure FIOSEEKDATA * will fail for those file systems.) * * For input files that don't support FIOSEEKDATA/FIOSEEKHOLE, * the code just falls through to the inner copy loop. */ error = EINVAL; if (holein > 0) error = VOP_IOCTL(invp, FIOSEEKDATA, &startoff, 0, incred, curthread); if (error == 0) { endoff = startoff; error = VOP_IOCTL(invp, FIOSEEKHOLE, &endoff, 0, incred, curthread); /* * Since invp is unlocked, it may be possible for * another thread to do a truncate(), lseek(), write() * creating a hole at startoff between the above * VOP_IOCTL() calls, if the other thread does not do * rangelocking. * If that happens, startoff == endoff and finding * the hole has failed, so set an error. */ if (error == 0 && startoff == endoff) error = EINVAL; /* Any error. Reset to 0. */ } if (error == 0) { if (startoff > *inoffp) { /* Found hole before data block. */ xfer = MIN(startoff - *inoffp, len); if (*outoffp < va.va_size) { /* Must write 0s to punch hole. */ xfer2 = MIN(va.va_size - *outoffp, xfer); memset(dat, 0, MIN(xfer2, blksize)); error = vn_write_outvp(outvp, dat, *outoffp, xfer2, blksize, false, holeout > 0, outcred); } if (error == 0 && *outoffp + xfer > va.va_size && xfer == len) /* Grow last block. */ error = vn_write_outvp(outvp, dat, *outoffp, xfer, blksize, true, false, outcred); if (error == 0) { *inoffp += xfer; *outoffp += xfer; len -= xfer; } } copylen = MIN(len, endoff - startoff); cantseek = false; } else { cantseek = true; startoff = *inoffp; copylen = len; error = 0; } xfer = blksize; if (cantseek) { /* * Set first xfer to end at a block boundary, so that * holes are more likely detected in the loop below via * the for all bytes 0 method. */ xfer -= (*inoffp % blksize); } /* Loop copying the data block. */ while (copylen > 0 && error == 0 && !eof) { if (copylen < xfer) xfer = copylen; error = vn_lock(invp, LK_SHARED); if (error != 0) goto out; error = vn_rdwr(UIO_READ, invp, dat, xfer, startoff, UIO_SYSSPACE, IO_NODELOCKED, curthread->td_ucred, incred, &aresid, curthread); VOP_UNLOCK(invp); lastblock = false; if (error == 0 && aresid > 0) { /* Stop the copy at EOF on the input file. */ xfer -= aresid; eof = true; lastblock = true; } if (error == 0) { /* * Skip the write for holes past the initial EOF * of the output file, unless this is the last * write of the output file at EOF. */ readzeros = cantseek ? mem_iszero(dat, xfer) : false; if (xfer == len) lastblock = true; if (!cantseek || *outoffp < va.va_size || lastblock || !readzeros) error = vn_write_outvp(outvp, dat, *outoffp, xfer, blksize, readzeros && lastblock && *outoffp >= va.va_size, false, outcred); if (error == 0) { *inoffp += xfer; startoff += xfer; *outoffp += xfer; copylen -= xfer; len -= xfer; } } xfer = blksize; } } out: *lenp = savlen - len; free(dat, M_TEMP); return (error); } static int vn_fallocate(struct file *fp, off_t offset, off_t len, struct thread *td) { struct mount *mp; struct vnode *vp; off_t olen, ooffset; int error; #ifdef AUDIT int audited_vnode1 = 0; #endif vp = fp->f_vnode; if (vp->v_type != VREG) return (ENODEV); /* Allocating blocks may take a long time, so iterate. */ for (;;) { olen = len; ooffset = offset; bwillwrite(); mp = NULL; error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error != 0) break; error = vn_lock(vp, LK_EXCLUSIVE); if (error != 0) { vn_finished_write(mp); break; } #ifdef AUDIT if (!audited_vnode1) { AUDIT_ARG_VNODE1(vp); audited_vnode1 = 1; } #endif #ifdef MAC error = mac_vnode_check_write(td->td_ucred, fp->f_cred, vp); if (error == 0) #endif error = VOP_ALLOCATE(vp, &offset, &len); VOP_UNLOCK(vp); vn_finished_write(mp); if (olen + ooffset != offset + len) { panic("offset + len changed from %jx/%jx to %jx/%jx", ooffset, olen, offset, len); } if (error != 0 || len == 0) break; KASSERT(olen > len, ("Iteration did not make progress?")); maybe_yield(); } return (error); }