diff --git a/lib/libsys/kqueue.2 b/lib/libsys/kqueue.2 --- a/lib/libsys/kqueue.2 +++ b/lib/libsys/kqueue.2 @@ -97,10 +97,22 @@ a .Fa flags argument, which is a bitwise-inclusive OR of the following flags: -.Bl -tag -width "KQUEUE_CLOEXEC" +.Bl -tag -width "KQUEUE_CPONFORK" .It Dv KQUEUE_CLOEXEC The returned file descriptor is automatically closed on .Xr execve 2 +.It Dv KQUEUE_CPONFORK +With the flag set, the created kqueue is copied to the child process on +.Xr fork 2 +calls. +Copying is deep, that is, each registered event in the original kqueue is +copied (and not shared) into the copy. +A new file descriptor with the same index as in parent, references the +copied kqueue in the child process. +.Pp +By default, in other words, when the flag is not set, kqueues from +the parent are not copied on fork to the child process. +The corresponding file descriptor indexes are invalid in the child. .El .Pp The diff --git a/sys/compat/linux/linux_event.c b/sys/compat/linux/linux_event.c --- a/sys/compat/linux/linux_event.c +++ b/sys/compat/linux/linux_event.c @@ -104,7 +104,7 @@ epoll_create_common(struct thread *td, int flags) { - return (kern_kqueue(td, flags, NULL)); + return (kern_kqueue(td, flags, false, NULL)); } #ifdef LINUX_LEGACY_SYSCALLS diff --git a/sys/dev/netmap/netmap_freebsd.c b/sys/dev/netmap/netmap_freebsd.c --- a/sys/dev/netmap/netmap_freebsd.c +++ b/sys/dev/netmap/netmap_freebsd.c @@ -1406,19 +1406,34 @@ return netmap_knrw(kn, hint, POLLOUT); } +static int +netmap_kncopy(struct knote *kn, struct proc *p1) +{ + struct netmap_priv_d *priv; + struct nm_selinfo *si; + + priv = kn->kn_hook; + si = priv->np_si[kn->kn_filter == EVFILT_WRITE ? NR_TX : NR_RX]; + NMG_LOCK(); + si->kqueue_users++; + NMG_UNLOCK(); + return (0); +} + static const struct filterops netmap_rfiltops = { .f_isfd = 1, .f_detach = netmap_knrdetach, .f_event = netmap_knread, + .f_copy = netmap_kncopy, }; static const struct filterops netmap_wfiltops = { .f_isfd = 1, .f_detach = netmap_knwdetach, .f_event = netmap_knwrite, + .f_copy = netmap_kncopy, }; - /* * This is called when a thread invokes kevent() to record * a change in the configuration of the kqueue(). diff --git a/sys/kern/kern_descrip.c b/sys/kern/kern_descrip.c --- a/sys/kern/kern_descrip.c +++ b/sys/kern/kern_descrip.c @@ -2486,7 +2486,7 @@ if (refcount_load(&p->p_fd->fd_refcnt) == 1) return; - tmp = fdcopy(p->p_fd); + tmp = fdcopy(p->p_fd, p); fdescfree(td); p->p_fd = tmp; } @@ -2515,14 +2515,17 @@ * this is to ease callers, not catch errors. */ struct filedesc * -fdcopy(struct filedesc *fdp) +fdcopy(struct filedesc *fdp, struct proc *p1) { struct filedesc *newfdp; struct filedescent *nfde, *ofde; + struct file *fp; int i, lastfile; + bool fork_pass; MPASS(fdp != NULL); + fork_pass = false; newfdp = fdinit(); FILEDESC_SLOCK(fdp); for (;;) { @@ -2536,6 +2539,8 @@ /* copy all passable descriptors (i.e. not kqueue) */ newfdp->fd_freefile = fdp->fd_freefile; FILEDESC_FOREACH_FDE(fdp, i, ofde) { + if ((ofde->fde_file->f_ops->fo_flags & DFLAG_FORK) != 0) + fork_pass = true; if ((ofde->fde_file->f_ops->fo_flags & DFLAG_PASSABLE) == 0 || (ofde->fde_flags & UF_FOCLOSE) != 0 || !fhold(ofde->fde_file)) { @@ -2549,7 +2554,34 @@ fdused_init(newfdp, i); } MPASS(newfdp->fd_freefile != -1); + if (fork_pass) { + FILEDESC_FOREACH_FDE(fdp, i, ofde) { + if ((ofde->fde_file->f_ops->fo_flags & + DFLAG_FORK) == 0 || + (ofde->fde_flags & UF_FOCLOSE) != 0) + continue; + fp = NULL; + if (ofde->fde_file->f_ops->fo_fork(newfdp, + ofde->fde_file, &fp, p1, curthread) != 0) + continue; + nfde = &newfdp->fd_ofiles[i]; + *nfde = *ofde; + nfde->fde_file = fp; + filecaps_copy(&ofde->fde_caps, &nfde->fde_caps, true); + fdused_init(newfdp, i); + } + } FILEDESC_SUNLOCK(fdp); + if (fork_pass) { + FILEDESC_FOREACH_FDE(newfdp, i, nfde) { + if ((nfde->fde_file->f_ops->fo_flags & DFLAG_FORK) == 0) + continue; + if (nfde->fde_file == NULL) + continue; + nfde->fde_file->f_ops->fo_fork(newfdp, NULL, + &nfde->fde_file, p1, curthread); + } + } return (newfdp); } diff --git a/sys/kern/kern_event.c b/sys/kern/kern_event.c --- a/sys/kern/kern_event.c +++ b/sys/kern/kern_event.c @@ -134,6 +134,7 @@ static fo_stat_t kqueue_stat; static fo_close_t kqueue_close; static fo_fill_kinfo_t kqueue_fill_kinfo; +static fo_fork_t kqueue_fork; static const struct fileops kqueueops = { .fo_read = invfo_rdwr, @@ -148,7 +149,9 @@ .fo_chown = invfo_chown, .fo_sendfile = invfo_sendfile, .fo_cmp = file_kcmp_generic, + .fo_fork = kqueue_fork, .fo_fill_kinfo = kqueue_fill_kinfo, + .fo_flags = DFLAG_FORK, }; static int knote_attach(struct knote *kn, struct kqueue *kq); @@ -176,6 +179,7 @@ static void filt_timerstart(struct knote *kn, sbintime_t to); static void filt_timertouch(struct knote *kn, struct kevent *kev, u_long type); +static int filt_timercopy(struct knote *kn, struct proc *p1); static int filt_timervalidate(struct knote *kn, sbintime_t *to); static int filt_timer(struct knote *kn, long hint); static int filt_userattach(struct knote *kn); @@ -212,6 +216,7 @@ .f_detach = filt_timerdetach, .f_event = filt_timer, .f_touch = filt_timertouch, + .f_copy = filt_timercopy, }; static const struct filterops user_filtops = { .f_attach = filt_userattach, @@ -940,6 +945,30 @@ return (0); } +static int +filt_timercopy(struct knote *kn, struct proc *p) +{ + struct kq_timer_cb_data *kc_src, *kc; + + if (atomic_fetchadd_int(&kq_ncallouts, 1) + 1 > kq_calloutmax) { + atomic_subtract_int(&kq_ncallouts, 1); + return (ENOMEM); + } + + kn->kn_status &= ~KN_DETACHED; + kc_src = kn->kn_ptr.p_v; + kn->kn_ptr.p_v = kc = malloc(sizeof(*kc), M_KQUEUE, M_WAITOK); + kc->kn = kn; + kc->p = p; + kc->flags = kc_src->flags & ~KQ_TIMER_CB_ENQUEUED; + kc->next = kc_src->next; + kc->to = kc_src->to; + kc->cpuid = kc_src->cpuid; + callout_init(&kc->c, 1); + kqtimer_sched_callout(kc); + return (0); +} + static void filt_timerstart(struct knote *kn, sbintime_t to) { @@ -1151,7 +1180,7 @@ sys_kqueue(struct thread *td, struct kqueue_args *uap) { - return (kern_kqueue(td, 0, NULL)); + return (kern_kqueue(td, 0, false, NULL)); } int @@ -1159,55 +1188,76 @@ { int flags; - if ((uap->flags & ~(KQUEUE_CLOEXEC)) != 0) + if ((uap->flags & ~(KQUEUE_CLOEXEC | KQUEUE_CPONFORK)) != 0) return (EINVAL); flags = 0; if ((uap->flags & KQUEUE_CLOEXEC) != 0) flags |= O_CLOEXEC; - return (kern_kqueue(td, flags, NULL)); + return (kern_kqueue(td, flags, (uap->flags & KQUEUE_CPONFORK) != 0, + NULL)); } static void -kqueue_init(struct kqueue *kq) +kqueue_init(struct kqueue *kq, bool cponfork) { mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF | MTX_DUPOK); TAILQ_INIT(&kq->kq_head); knlist_init_mtx(&kq->kq_sel.si_note, &kq->kq_lock); TASK_INIT(&kq->kq_task, 0, kqueue_task, kq); + if (cponfork) + kq->kq_state |= KQ_CPONFORK; } -int -kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps) +static int +kern_kqueue_alloc(struct thread *td, struct filedesc *fdp, int *fdip, + struct file **fpp, int flags, struct filecaps *fcaps, bool cponfork, + struct kqueue **kqp) { - struct filedesc *fdp; - struct kqueue *kq; - struct file *fp; struct ucred *cred; - int fd, error; + struct kqueue *kq; + int error; - fdp = td->td_proc->p_fd; cred = td->td_ucred; if (!chgkqcnt(cred->cr_ruidinfo, 1, lim_cur(td, RLIMIT_KQUEUES))) return (ENOMEM); - error = falloc_caps(td, &fp, &fd, flags, fcaps); + error = fdip != NULL ? falloc_caps(td, fpp, fdip, flags, fcaps) : + _falloc_noinstall(td, fpp, 1); if (error != 0) { chgkqcnt(cred->cr_ruidinfo, -1, 0); return (error); } /* An extra reference on `fp' has been held for us by falloc(). */ - kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO); - kqueue_init(kq); + kq = malloc(sizeof(*kq), M_KQUEUE, M_WAITOK | M_ZERO); + kqueue_init(kq, cponfork); kq->kq_fdp = fdp; kq->kq_cred = crhold(cred); - FILEDESC_XLOCK(fdp); + if (fdip != NULL) + FILEDESC_XLOCK(fdp); TAILQ_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list); - FILEDESC_XUNLOCK(fdp); + if (fdip != NULL) + FILEDESC_XUNLOCK(fdp); + + finit(*fpp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops); + *kqp = kq; + return (0); +} + +int +kern_kqueue(struct thread *td, int flags, bool cponfork, struct filecaps *fcaps) +{ + struct kqueue *kq; + struct file *fp; + int fd, error; + + error = kern_kqueue_alloc(td, td->td_proc->p_fd, &fd, &fp, flags, + fcaps, cponfork, &kq); + if (error != 0) + return (error); - finit(fp, FREAD | FWRITE, DTYPE_KQUEUE, kq, &kqueueops); fdrop(fp, td); td->td_retval[0] = fd; @@ -1488,7 +1538,7 @@ struct kqueue kq = {}; int error; - kqueue_init(&kq); + kqueue_init(&kq, false); kq.kq_refcnt = 1; error = kqueue_kevent(&kq, td, nevents, nevents, k_ops, NULL); kqueue_drain(&kq, td); @@ -1576,7 +1626,7 @@ mtx_lock(&filterops_lock); KASSERT(sysfilt_ops[~filt].for_refcnt > 0, - ("filter object refcount not valid on release")); + ("filter object %d refcount not valid on release", filt)); sysfilt_ops[~filt].for_refcnt--; mtx_unlock(&filterops_lock); } @@ -1855,17 +1905,8 @@ } static int -kqueue_acquire(struct file *fp, struct kqueue **kqp) +kqueue_acquire_ref(struct kqueue *kq) { - int error; - struct kqueue *kq; - - error = 0; - - kq = fp->f_data; - if (fp->f_type != DTYPE_KQUEUE || kq == NULL) - return (EINVAL); - *kqp = kq; KQ_LOCK(kq); if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) { KQ_UNLOCK(kq); @@ -1873,8 +1914,22 @@ } kq->kq_refcnt++; KQ_UNLOCK(kq); + return (0); +} - return error; +static int +kqueue_acquire(struct file *fp, struct kqueue **kqp) +{ + struct kqueue *kq; + int error; + + kq = fp->f_data; + if (fp->f_type != DTYPE_KQUEUE || kq == NULL) + return (EINVAL); + error = kqueue_acquire_ref(kq); + if (error == 0) + *kqp = kq; + return (error); } static void @@ -2937,6 +2992,153 @@ return (error); } +static int +kqueue_fork_alloc(struct filedesc *fdp, struct file *fp, struct file **fp1, + struct thread *td) +{ + struct kqueue *kq, *kq1; + int error; + + MPASS(fp->f_type == DTYPE_KQUEUE); + kq = fp->f_data; + if ((kq->kq_state & KQ_CPONFORK) == 0) + return (EOPNOTSUPP); + error = kqueue_acquire_ref(kq); + if (error != 0) + return (error); + error = kern_kqueue_alloc(td, fdp, NULL, fp1, 0, NULL, true, &kq1); + if (error == 0) { +KQ_LOCK(kq); +kq->copies++; +KQ_UNLOCK(kq); + kq1->kq_forksrc = kq; + (*fp1)->f_flag = fp->f_flag & (FREAD | FWRITE | FEXEC | + O_CLOEXEC | O_CLOFORK); + } else { + kqueue_release(kq, 0); + } + return (error); +} + +static void +kqueue_fork_copy_knote(struct kqueue *kq1, struct knote *kn, struct proc *p1, + struct filedesc *fdp) +{ + struct knote *kn1; + const struct filterops *fop; + int error; + + fop = kn->kn_fop; + if (fop->f_isfd && + fdp->fd_files->fdt_ofiles[kn->kn_kevent.ident].fde_file == NULL) + return; + error = kqueue_expand(kq1, fop, kn->kn_kevent.ident, M_WAITOK); + if (error != 0) + return; + + kn1 = knote_alloc(M_WAITOK); + *kn1 = *kn; + kn1->kn_status |= KN_DETACHED; + kn1->kn_status &= ~KN_QUEUED; + kn1->kn_kq = kq1; + if (fop->f_copy != NULL) + error = fop->f_copy(kn1, p1); + else + error = 0; + if (error != 0) { + knote_free(kn1); + return; + } + (void)kqueue_fo_find(kn->kn_kevent.filter); + if (fop->f_isfd && !fhold(kn1->kn_fp)) { + fop->f_detach(kn1); + kqueue_fo_release(kn->kn_kevent.filter); + knote_free(kn1); + return; + } + if (kn->kn_knlist != NULL) + knlist_add(kn->kn_knlist, kn1, 0); + KQ_LOCK(kq1); + knote_attach(kn1, kq1); + kn1->kn_influx = 0; + if ((kn->kn_status & KN_QUEUED) != 0) + knote_enqueue(kn1); + KQ_UNLOCK(kq1); +} + +static void +kqueue_fork_copy_list(struct klist *knlist, struct knote *marker, + struct kqueue *kq, struct kqueue *kq1, struct proc *p1, + struct filedesc *fdp) +{ + struct knote *kn; + + KQ_OWNED(kq); + kn = SLIST_FIRST(knlist); + while (kn != NULL) { + if ((kn->kn_status & KN_DETACHED) != 0 || + (kn_in_flux(kn) && (kn->kn_status & KN_SCAN) == 0)) { + kn = SLIST_NEXT(kn, kn_link); + continue; + } + kn_enter_flux(kn); + SLIST_INSERT_AFTER(kn, marker, kn_link); + KQ_UNLOCK(kq); + kqueue_fork_copy_knote(kq1, kn, p1, fdp); + KQ_LOCK(kq); + kn_leave_flux(kn); + kn = SLIST_NEXT(marker, kn_link); + /* XXXKIB switch kn_link to LIST? */ + SLIST_REMOVE(knlist, marker, knote, kn_link); + } +} + +static int +kqueue_fork_copy(struct filedesc *fdp, struct file *fp, struct file *fp1, + struct proc *p1, struct thread *td) +{ + struct kqueue *kq, *kq1; + struct knote *marker; + int error, i; + + error = 0; + MPASS(fp == NULL); + MPASS(fp1->f_type == DTYPE_KQUEUE); + + kq1 = fp1->f_data; + kq = kq1->kq_forksrc; + marker = knote_alloc(M_WAITOK); + marker->kn_status = KN_MARKER; + + KQ_LOCK(kq); + for (i = 0; i < kq->kq_knlistsize; i++) { + kqueue_fork_copy_list(&kq->kq_knlist[i], marker, kq, kq1, + p1, fdp); + } + if (kq->kq_knhashmask != 0) { + for (i = 0; i <= kq->kq_knhashmask; i++) { + kqueue_fork_copy_list(&kq->kq_knhash[i], marker, kq, + kq1, p1, fdp); + } + } +kq->copies--; + kqueue_release(kq, 1); + kq1->kq_forksrc = NULL; + KQ_UNLOCK(kq); + + knote_free(marker); + return (error); +} + +static int +kqueue_fork(struct filedesc *fdp, struct file *fp, struct file **fp1, + struct proc *p1, struct thread *td) +{ + if (*fp1 == NULL) + return (kqueue_fork_alloc(fdp, fp, fp1, td)); + return (kqueue_fork_copy(fdp, fp, *fp1, p1, td)); +} + struct knote_status_export_bit { int kn_status_bit; int knt_status_bit; diff --git a/sys/kern/kern_fork.c b/sys/kern/kern_fork.c --- a/sys/kern/kern_fork.c +++ b/sys/kern/kern_fork.c @@ -423,7 +423,7 @@ pd = pdshare(p1->p_pd); else pd = pdcopy(p1->p_pd); - fd = fdcopy(p1->p_fd); + fd = fdcopy(p1->p_fd, p2); fdtol = NULL; } else { if (fr->fr_flags2 & FR2_SHARE_PATHS) diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c --- a/sys/kern/vfs_subr.c +++ b/sys/kern/vfs_subr.c @@ -6624,24 +6624,28 @@ static void filt_vfsdetach(struct knote *kn); static int filt_vfsdump(struct proc *p, struct knote *kn, struct kinfo_knote *kin); +static int filt_vfscopy(struct knote *kn, struct proc *p1); static const struct filterops vfsread_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfsread, .f_userdump = filt_vfsdump, + .f_copy = filt_vfscopy, }; static const struct filterops vfswrite_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfswrite, .f_userdump = filt_vfsdump, + .f_copy = filt_vfscopy, }; static const struct filterops vfsvnode_filtops = { .f_isfd = 1, .f_detach = filt_vfsdetach, .f_event = filt_vfsvnode, .f_userdump = filt_vfsdump, + .f_copy = filt_vfscopy, }; static void @@ -6825,6 +6829,16 @@ return (0); } +static int +filt_vfscopy(struct knote *kn, struct proc *p1) +{ + struct vnode *vp; + + vp = (struct vnode *)kn->kn_hook; + vhold(vp); + return (0); +} + int vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off) { diff --git a/sys/sys/event.h b/sys/sys/event.h --- a/sys/sys/event.h +++ b/sys/sys/event.h @@ -228,6 +228,7 @@ /* Flags for kqueuex(2) */ #define KQUEUE_CLOEXEC 0x00000001 /* close on exec */ +#define KQUEUE_CPONFORK 0x00000002 /* copy on fork */ struct knote; SLIST_HEAD(klist, knote); @@ -283,6 +284,7 @@ void (*f_touch)(struct knote *kn, struct kevent *kev, u_long type); int (*f_userdump)(struct proc *p, struct knote *kn, struct kinfo_knote *kin); + int (*f_copy)(struct knote *kn, struct proc *p1); }; /* diff --git a/sys/sys/eventvar.h b/sys/sys/eventvar.h --- a/sys/sys/eventvar.h +++ b/sys/sys/eventvar.h @@ -55,12 +55,15 @@ #define KQ_CLOSING 0x10 #define KQ_TASKSCHED 0x20 /* task scheduled */ #define KQ_TASKDRAIN 0x40 /* waiting for task to drain */ +#define KQ_CPONFORK 0x80 int kq_knlistsize; /* size of knlist */ struct klist *kq_knlist; /* list of knotes */ u_long kq_knhashmask; /* size of knhash */ struct klist *kq_knhash; /* hash table for knotes */ struct task kq_task; struct ucred *kq_cred; + struct kqueue *kq_forksrc; +u_int copies; }; #endif /* !_SYS_EVENTVAR_H_ */ diff --git a/sys/sys/file.h b/sys/sys/file.h --- a/sys/sys/file.h +++ b/sys/sys/file.h @@ -139,6 +139,8 @@ off_t *offset, off_t *length, int flags, struct ucred *active_cred, struct thread *td); typedef int fo_cmp_t(struct file *fp, struct file *fp1, struct thread *td); +typedef int fo_fork_t(struct filedesc *fdp, struct file *fp, struct file **fp1, + struct proc *p1, struct thread *td); typedef int fo_spare_t(struct file *fp); typedef int fo_flags_t; @@ -163,12 +165,14 @@ fo_fallocate_t *fo_fallocate; fo_fspacectl_t *fo_fspacectl; fo_cmp_t *fo_cmp; + fo_fork_t *fo_fork; fo_spare_t *fo_spares[8]; /* Spare slots */ fo_flags_t fo_flags; /* DFLAG_* below */ }; #define DFLAG_PASSABLE 0x01 /* may be passed via unix sockets. */ #define DFLAG_SEEKABLE 0x02 /* seekable / nonsequential */ +#define DFLAG_FORK 0x04 /* copy on fork */ #endif /* _KERNEL */ #if defined(_KERNEL) || defined(_WANT_FILE) diff --git a/sys/sys/filedesc.h b/sys/sys/filedesc.h --- a/sys/sys/filedesc.h +++ b/sys/sys/filedesc.h @@ -265,7 +265,7 @@ void fdclose(struct thread *td, struct file *fp, int idx); void fdcloseexec(struct thread *td); void fdsetugidsafety(struct thread *td); -struct filedesc *fdcopy(struct filedesc *fdp); +struct filedesc *fdcopy(struct filedesc *fdp, struct proc *p1); void fdunshare(struct thread *td); void fdescfree(struct thread *td); int fdlastfile(struct filedesc *fdp); diff --git a/sys/sys/syscallsubr.h b/sys/sys/syscallsubr.h --- a/sys/sys/syscallsubr.h +++ b/sys/sys/syscallsubr.h @@ -211,7 +211,8 @@ int nevents, struct kevent_copyops *k_ops, const struct timespec *timeout); int kern_kill(struct thread *td, pid_t pid, int signum); -int kern_kqueue(struct thread *td, int flags, struct filecaps *fcaps); +int kern_kqueue(struct thread *td, int flags, bool cponfork, + struct filecaps *fcaps); int kern_kldload(struct thread *td, const char *file, int *fileid); int kern_kldstat(struct thread *td, int fileid, struct kld_file_stat *stat); int kern_kldunload(struct thread *td, int fileid, int flags);