Mirror-bot/darling-lol
1
0 Fork 0
mirror of https://github.com/lucannez64/darling-lol.git synced 2026-10-26 19:47:42 +00:00
Code Issues Projects Releases Activity
Files
cdbdd357bb1e4b7ef5b1138016ca4b91920a9751
darling-lol/patches/libkqueue.patch
T

618 lines
23 KiB
Diff
Raw Blame History

diff --git a/CMakeLists.txt b/CMakeLists.txt
index b9cf3f8..62e11d0 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -31,7 +31,12 @@ if (DARLING)
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -nostdinc -DHAVE_SYS_EVENTFD_H -DHAVE_SYS_SIGNALFD_H -DHAVE_SYS_TIMERFD_H -ggdb -O0")
#add_definitions(-DKERNEL_PRIVATE)
- #add_definitions(-Dkqueue=kqueue_impl -Dkevent=kevent_impl -Dkevent64=kevent64_impl)
+ # darling-lol: enable kqueue emulation (was stubbed ENOSYS, breaking
+ # dispatch sources, FilePathWatcher, and socket servers). Maps the
+ # public API to the _impl entry points that override libsystem_kernel's
+ # weak stubs. (Deliberately NOT mapping close=__close_for_kqueue yet;
+ # evaluate fd-lifecycle tracking separately.)
+ add_definitions(-Dkqueue=kqueue_impl -Dkevent=kevent_impl -Dkevent64=kevent64_impl)
#add_definitions(-Dclose=__close_for_kqueue)
endif (DARLING)
diff --git a/src/common/filter.c b/src/common/filter.c
index a9a6512..f324b41 100644
--- a/src/common/filter.c
+++ b/src/common/filter.c
@@ -125,11 +125,14 @@ filter_unregister_all(struct kqueue *kq)
if (kq->kq_filt[i].kf_id == 0)
continue;
- if (kq->kq_filt[i].kf_destroy != NULL)
- kq->kq_filt[i].kf_destroy(&kq->kq_filt[i]);
+ /* Closing the queue owns teardown of every remaining registration,
+ * including the native socket retained by a Mach-port knote. */
+ struct knote *kn;
+ while ((kn = RB_ROOT(&kq->kq_filt[i].kf_knote)) != NULL)
+ knote_delete(&kq->kq_filt[i], kn);
- //XXX-FIXME
- //knote_free_all(&kq->kq_filt[i]);
+ if (kq->kq_filt[i].kf_destroy != NULL)
+ kq->kq_filt[i].kf_destroy(&kq->kq_filt[i]);
if (kqops.filter_free != NULL)
kqops.filter_free(kq, &kq->kq_filt[i]);
diff --git a/src/common/kevent.c b/src/common/kevent.c
index f5d53a2..ca91b33 100644
--- a/src/common/kevent.c
+++ b/src/common/kevent.c
@@ -27,6 +27,8 @@
#include <string.h>
#include "private.h"
+#include <darling/emulation/common/simple.h>
+#define LOLKQ(...) do { } while (0)
static const char *
kevent_filter_dump(const struct kevent64_s *kev)
@@ -127,7 +129,9 @@ kevent_copyin_one(struct kqueue *kq, const struct kevent64_s *src)
dbg_printf("src=%s", kevent_dump(src));
- kn = knote_lookup(filt, src->ident);
+ if((src->filter==EVFILT_READ || src->filter==EVFILT_WRITE) && (src->flags&EV_UDATA_SPECIFIC))
+ kn=knote_lookup_fd_specific(filt,src->ident,src->udata);
+ else kn = knote_lookup(filt, src->ident);
dbg_printf("knote_lookup: ident %d == %p", (int)src->ident, kn);
if (kn == NULL) {
if (src->flags & EV_ADD) {
@@ -170,6 +174,16 @@ kevent_copyin_one(struct kqueue *kq, const struct kevent64_s *src)
}
}
+ /* A delivered dispatch2 event still owns its userdata until acknowledged.
+ * Match XNU: a plain delete of a disabled source defers reclamation. */
+ if ((src->flags & EV_DELETE) && !(src->flags & EV_ENABLE) &&
+ (kn->kev.flags & EV_DISPATCH2) == EV_DISPATCH2 &&
+ (kn->kev.flags & EV_DISABLE)) {
+ kn->kn_flags |= KNFL_DEFER_DELETE;
+ errno = EINPROGRESS;
+ return (-1);
+ }
+
if ((src->flags & EV_DELETE) || ((kn->kn_flags & KNFL_DEFER_DELETE) && (src->flags & EV_ENABLE))) {
rv = knote_delete(filt, kn);
dbg_printf("knote_delete returned %d", rv);
@@ -349,6 +363,7 @@ kevent64_impl(int kqfd, const struct kevent64_s *changelist, int nchanges,
kqueue_addref(kq);
pthread_mutex_unlock(&kq_mtx);
+ LOLKQ("kevent64_impl enter kqfd=%d nch=%d nev=%d\n", kqfd, nchanges, nevents);
#ifndef NDEBUG
if (DEBUG_KQUEUE) {
@@ -386,6 +401,11 @@ kevent64_impl(int kqfd, const struct kevent64_s *changelist, int nchanges,
if ((flags & KEVENT_FLAG_ERROR_EVENTS) == 0 && nevents > 0) {
const struct timespec* ts = (flags & KEVENT_FLAG_IMMEDIATE) ? (&timeout_zero) : timeout;
again:
+ /* epoll snapshots contain raw knote pointers. Keep retired notes alive
+ * until every waiter that could have observed them has copied out. */
+ kqueue_lock(kq);
+ kq->kq_waiters++;
+ kqueue_unlock(kq);
rv = kqops.kevent_wait(kq, nevents, ts);
dbg_printf("kqops.kevent_wait returned %d", rv);
@@ -395,6 +415,7 @@ again:
ret += rv;
}
+ kq->kq_waiters--;
kqueue_cleanup(kq);
kqueue_unlock(kq);
diff --git a/src/common/knote.c b/src/common/knote.c
index a328f80..d0141aa 100644
--- a/src/common/knote.c
+++ b/src/common/knote.c
@@ -33,7 +33,17 @@ knote_init(void)
static int
knote_cmp(struct knote *a, struct knote *b)
{
- return memcmp(&a->kev.ident, &b->kev.ident, sizeof(a->kev.ident));
+ int compared=memcmp(&a->kev.ident,&b->kev.ident,sizeof(a->kev.ident));
+ if(compared) return compared;
+ // Darwin allows independent fd registrations distinguished by udata.
+ // Each Linux fd note already owns a separate dupfd/epoll registration.
+ if(a->kev.filter==EVFILT_READ || a->kev.filter==EVFILT_WRITE) {
+ int as=!!(a->kev.flags&EV_UDATA_SPECIFIC);
+ int bs=!!(b->kev.flags&EV_UDATA_SPECIFIC);
+ if(as!=bs) return as-bs;
+ if(as) return a->kev.udata<b->kev.udata?-1:a->kev.udata>b->kev.udata?1:0;
+ }
+ return 0;
}
RB_GENERATE(knt, knote, kn_entries, knote_cmp)
@@ -94,7 +104,7 @@ knote_delete(struct filter *filt, struct knote *kn)
* Verify that the knote wasn't removed by another
* thread before we acquired the knotelist lock.
*/
- query.kev.ident = kn->kev.ident;
+ query.kev = kn->kev;
pthread_rwlock_wrlock(&filt->kf_knote_mtx);
tmp = RB_FIND(knt, &filt->kf_knote, &query);
if (tmp == kn) {
@@ -114,10 +124,11 @@ knote_delete(struct filter *filt, struct knote *kn)
struct knote *
knote_lookup(struct filter *filt, uintptr_t ident)
{
- struct knote query;
+ struct knote query={0};
struct knote *ent = NULL;
query.kev.ident = ident;
+ query.kev.filter=filt->kf_id;
pthread_rwlock_rdlock(&filt->kf_knote_mtx);
ent = RB_FIND(knt, &filt->kf_knote, &query);
@@ -160,3 +171,14 @@ knote_disable(struct filter *filt, struct knote *kn)
//TODO: knote_enable()
+
+struct knote *
+knote_lookup_fd_specific(struct filter *filt,uintptr_t ident,uint64_t udata) {
+ struct knote query={0};
+ query.kev.ident=ident;query.kev.filter=filt->kf_id;
+ query.kev.flags=EV_UDATA_SPECIFIC;query.kev.udata=udata;
+ pthread_rwlock_rdlock(&filt->kf_knote_mtx);
+ struct knote* result=RB_FIND(knt,&filt->kf_knote,&query);
+ pthread_rwlock_unlock(&filt->kf_knote_mtx);
+ return result;
+}
diff --git a/src/common/kqueue.c b/src/common/kqueue.c
index 61cc9dc..f05576d 100644
--- a/src/common/kqueue.c
+++ b/src/common/kqueue.c
@@ -23,6 +23,8 @@
#include <string.h>
#include "private.h"
+#include <darling/emulation/common/simple.h>
+#define LOLKQ(...) do { } while (0)
#ifdef DARLING
#include "../darling/listenregistry.h"
@@ -113,6 +115,7 @@ kqueue_cmp(struct kqueue *a, struct kqueue *b)
void kqueue_cleanup(struct kqueue* kq)
{
struct knote *next, *tmp;
+ if (kq->kq_waiters) return;
for (next = kq->kq_tofree.lh_first; next != NULL; next = tmp)
{
tmp = next->kn_entries2free.le_next;
@@ -135,6 +138,7 @@ kqueue_free(struct kqueue *kq)
// XXX: we should be doing this more efficiently (i.e. without walking through all open kqueues)
map_foreach(kqmap, kqueue_free_cb, kq);
filter_unregister_all(kq);
+ kqueue_cleanup(kq);
kqops.kqueue_free(kq);
free(kq);
}
@@ -329,6 +333,7 @@ void kqueue_delref(struct kqueue *kq)
int VISIBLE
kqueue_impl(void)
{
+ LOLKQ("kqueue_impl enter\n");
struct kqueue *kq;
struct kqueue *tmp;
diff --git a/src/common/map.c b/src/common/map.c
index cc254dd..ed0f21c 100644
--- a/src/common/map.c
+++ b/src/common/map.c
@@ -18,6 +18,7 @@
struct map {
size_t len;
+ int high_water;
void **data;
};
@@ -51,12 +52,23 @@ map_new(size_t len)
return (dst);
}
+/* Iteration is used on every descriptor close. Do not scan the entire
+ * RLIMIT_NOFILE-sized allocation for a map with only a few populated slots. */
+static void map_note_index(struct map *m, int idx)
+{
+ int end = idx + 1;
+ int old = __atomic_load_n(&m->high_water, __ATOMIC_RELAXED);
+ while (old < end && !__atomic_compare_exchange_n(&m->high_water, &old, end,
+ 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED)) {}
+}
+
int
map_insert(struct map *m, int idx, void *ptr)
{
- if (slowpath(idx < 0 || idx > (int)m->len))
+ if (slowpath(idx < 0 || idx >= (int)m->len))
return (-1);
+ map_note_index(m, idx);
if (atomic_ptr_cas(&(m->data[idx]), 0, ptr) == NULL) {
dbg_printf("inserted %p in location %d", ptr, idx);
return (0);
@@ -71,7 +83,7 @@ map_insert(struct map *m, int idx, void *ptr)
int
map_remove(struct map *m, int idx, void *ptr)
{
- if (slowpath(idx < 0 || idx > (int)m->len))
+ if (slowpath(idx < 0 || idx >= (int)m->len))
return (-1);
if (atomic_ptr_cas(&(m->data[idx]), ptr, 0) == NULL) {
@@ -88,9 +100,10 @@ map_replace(struct map *m, int idx, void *oldp, void *newp)
{
void *tmp;
- if (slowpath(idx < 0 || idx > (int)m->len))
+ if (slowpath(idx < 0 || idx >= (int)m->len))
return (-1);
+ map_note_index(m, idx);
tmp = atomic_ptr_cas(&(m->data[idx]), oldp, newp);
if (tmp == oldp) {
dbg_printf("replaced value %p in location %d with value %p",
@@ -106,7 +119,7 @@ map_replace(struct map *m, int idx, void *oldp, void *newp)
void *
map_lookup(struct map *m, int idx)
{
- if (slowpath(idx < 0 || idx > (int)m->len))
+ if (slowpath(idx < 0 || idx >= (int)m->len))
return (NULL);
return m->data[idx];
@@ -116,7 +129,8 @@ void
map_foreach(struct map *m, void(*cb)(int, void*, void*), void* private)
{
int i;
- for (i = 0; i < (int)m->len; i++) {
+ int end = __atomic_load_n(&m->high_water, __ATOMIC_ACQUIRE);
+ for (i = 0; i < end; i++) {
if (m->data[i])
cb(i, m->data[i], private);
}
@@ -128,7 +142,7 @@ map_delete(struct map *m, int idx)
void *oval;
void *nval;
- if (slowpath(idx < 0 || idx > (int)m->len))
+ if (slowpath(idx < 0 || idx >= (int)m->len))
return ((void *)-1);
/* Hopefully we aren't racing with another thread, but you never know.. */
diff --git a/src/common/private.h b/src/common/private.h
index 5c0718e..e7a0841 100644
--- a/src/common/private.h
+++ b/src/common/private.h
@@ -161,6 +161,7 @@ struct kqueue {
#endif
RB_ENTRY(kqueue) entries;
LIST_HEAD(knt2free, knote) kq_tofree;
+ unsigned int kq_waiters; /* protected by kq_mtx; keep existing offsets */
};
struct kqueue_vtable {
@@ -198,6 +199,7 @@ int kevent_copyin_one(struct kqueue *kq, const struct kevent64_s *src);
* knote internal API
*/
struct knote * knote_lookup(struct filter *, uintptr_t);
+struct knote * knote_lookup_fd_specific(struct filter *, uintptr_t,uint64_t);
//DEADWOOD: struct knote * knote_get_by_data(struct filter *filt, intptr_t);
struct knote * knote_new(void);
#define knote_retain(kn) atomic_inc(&kn->kn_ref)
diff --git a/src/linux/machport.c b/src/linux/machport.c
index 074958e..1e99023 100644
--- a/src/linux/machport.c
+++ b/src/linux/machport.c
@@ -34,6 +34,9 @@
#include "private.h"
+/* Keep the Mach-port path free of per-event diagnostic I/O. */
+#define M58MP(...) do { } while (0)
+
int
evfilt_machport_copyout(struct kevent64_s *dst, struct knote *src, void *ptr)
{
@@ -45,13 +48,16 @@ evfilt_machport_copyout(struct kevent64_s *dst, struct knote *src, void *ptr)
epoll_event_dump(ev);
kevent_int_to_64(&src->kev, dst);
+ M58MP("copyout fd=%d\n", src->kdata.kn_dupfd);
// first, read the notification
rv = recv(src->kdata.kn_dupfd, &notification, sizeof(notification), 0);
if (rv < 0) {
+ M58MP("copyout notif-recv FAILED fd=%d errno=%d\n", src->kdata.kn_dupfd, errno);
dbg_printf("evfilt_machport_copyout() reading notification failed: %d (%s)", errno, strerror(errno));
return -1;
}
+ M58MP("copyout notif fd=%d rv=%d num=0x%x\n", src->kdata.kn_dupfd, rv, notification.header.number);
if (notification.header.number != dserver_kqchan_msgnum_notification) {
dbg_puts("evfilt_machport_copyout() read invalid notification");
@@ -124,18 +130,24 @@ evfilt_machport_knote_create(struct filter *filt, struct knote *kn)
int status = _dserver_rpc_kqchan_mach_port_open_4libkqueue(port, (void*)kn->kev.ext[0], kn->kev.ext[1], kn->kev.fflags, &kn->kdata.kn_dupfd);
if (status < 0) {
+ M58MP("create open FAILED port=%d status=%d\n", port, status);
dbg_printf("evfilt_machport_open: %s", strerror(-status));
return (-1);
}
+ M58MP("create open port=%d fd=%d events=0x%x\n", port, kn->kdata.kn_dupfd, ev.events);
dbg_printf("evfilt_machport_open: listening to FD %d for events %d", kn->kdata.kn_dupfd, ev.events);
fcntl(kn->kdata.kn_dupfd, F_SETFD, FD_CLOEXEC);
if (epoll_ctl(kn->kn_epollfd, EPOLL_CTL_ADD, kn->kdata.kn_dupfd, &ev) < 0) {
+ M58MP("create epoll_add FAILED epfd=%d fd=%d errno=%d\n", kn->kn_epollfd, kn->kdata.kn_dupfd, errno);
dbg_printf("epoll_ctl(2): %s", strerror(errno));
+ (void) __close_for_kqueue(kn->kdata.kn_dupfd);
+ kn->kdata.kn_dupfd = -1;
return (-1);
}
+ M58MP("create epoll_add OK epfd=%d fd=%d\n", kn->kn_epollfd, kn->kdata.kn_dupfd);
return 0;
}
@@ -186,7 +198,8 @@ evfilt_machport_knote_delete(struct filter *filt, struct knote *kn)
if ((kn->kev.flags & EV_DISABLE) == 0) {
if (epoll_ctl(kn->kn_epollfd, EPOLL_CTL_DEL, kn->kdata.kn_dupfd, NULL) < 0) {
dbg_perror("epoll_ctl(2)");
- return (-1);
+ /* Teardown must still close the channel if its epoll entry was
+ * already removed. knote_delete frees the registration anyway. */
}
}
diff --git a/src/linux/platform.c b/src/linux/platform.c
index 482ffb2..dd750a4 100644
--- a/src/linux/platform.c
+++ b/src/linux/platform.c
@@ -16,6 +16,8 @@
# define _GNU_SOURCE
# include <poll.h>
+#include <darling/emulation/common/simple.h>
+#define LOLKQ(...) do { } while (0)
#include <pthread.h>
#include "../common/private.h"
@@ -179,6 +181,7 @@ linux_kevent_wait(
dbg_puts("waiting for events");
restart:
nret = epoll_wait(kqueue_epfd(kq), &epevt[0], nevents, timeout);
+ LOLKQ("ew ret=%d epfd=%d\n", nret, kqueue_epfd(kq));
if (nret < 0) {
if (errno == EINTR) // only happens when ptrace() is called
goto restart;
@@ -204,6 +207,13 @@ linux_kevent_copyout(struct kqueue *kq, int nready,
ev = &epevt_get()[i];
kn = (struct knote *) ev->data.ptr;
event = eventlist;
+ /* Another waiter or changelist may have consumed/removed this note
+ * after epoll_wait took its snapshot. Never deliver its userdata. */
+ if ((kn->kn_flags & KNFL_KNOTE_DELETED) ||
+ (kn->kev.flags & EV_DISABLE)) {
+ nret--;
+ continue;
+ }
if (kn->kev.filter == 0) {
dbg_puts("kevent copyout for zero filter, discarding!");
nret--;
@@ -352,8 +362,11 @@ linux_get_descriptor_type(struct knote *kn)
*/
if (fstat(kn->kev.ident, &sb) < 0) {
dbg_perror("fstat(2)");
+ LOLKQ("desctype: fstat(%d) FAILED errno=%d\n", (int)kn->kev.ident, (int)errno);
return (-1);
}
+ LOLKQ("desctype: fd=%d mode=0%o sock=%d reg=%d\n", (int)kn->kev.ident,
+ (unsigned)sb.st_mode, !!S_ISSOCK(sb.st_mode), !!S_ISREG(sb.st_mode));
if (S_ISREG(sb.st_mode)) {
kn->kn_flags |= KNFL_REGULAR_FILE;
dbg_printf("fd %d is a regular file\n", (int)kn->kev.ident);
@@ -382,6 +395,8 @@ linux_get_descriptor_type(struct knote *kn)
#endif
default:
dbg_perror("getsockopt(3)");
+ LOLKQ("desctype: getsockopt(%d,SO_ACCEPTCONN) FAILED errno=%d\n",
+ (int)kn->kev.ident, (int)errno);
return (-1);
}
}
@@ -436,12 +451,35 @@ epoll_update(int op, struct filter *filt, struct knote *kn, struct epoll_event *
int
linux_fd_to_path(char *buf, size_t bufsz, int fd)
{
+ /* NOTE: must use /proc/self, not /proc/<getpid()>: under Darling
+ * getpid() returns the *guest* pid, which does not exist in the host
+ * /proc. Using the guest pid broke every EVFILT_VNODE watch
+ * (linux_fd_to_path failed -> kevent returned EBADF). */
char path[1024]; //TODO: Maxpathlen, etc.
+#ifdef DARLING
+ /* darling-lol: readlink(/proc/self/fd) yields a HOST path, but our
+ * consumers (the ext inotify/epoll wrappers) treat paths as GUEST
+ * paths and re-apply guest->host translation (double translation ->
+ * ENOENT). Re-enter the guest namespace via /Volumes/SystemRoot so
+ * the round-trip resolves back to the same host file. */
+ char hpath[1024];
+ int n;
+#endif
- if (snprintf(&path[0], sizeof(path), "/proc/%d/fd/%d", getpid(), fd) < 0)
+ if (snprintf(&path[0], sizeof(path), "/proc/self/fd/%d", fd) < 0)
return (-1);
memset(buf, 0, bufsz);
+#ifdef DARLING
+ memset(hpath, 0, sizeof(hpath));
+ n = readlink(path, hpath, sizeof(hpath) - sizeof("/Volumes/SystemRoot") - 1);
+ if (n < 0)
+ return (-1);
+ hpath[n] = '\0';
+ if (snprintf(buf, bufsz, "/Volumes/SystemRoot%s", hpath) < 0)
+ return (-1);
+ return (n + (int)sizeof("/Volumes/SystemRoot") - 1);
+#else
return (readlink(path, buf, bufsz));
+#endif
}
-
diff --git a/src/linux/proc.c b/src/linux/proc.c
index 433a9b3..9a70d77 100644
--- a/src/linux/proc.c
+++ b/src/linux/proc.c
@@ -256,7 +256,7 @@ evfilt_proc_knote_delete(struct filter *filt, struct knote *kn)
if ((kn->kev.flags & EV_DISABLE) == 0) {
if (epoll_ctl(kn->kn_epollfd, EPOLL_CTL_DEL, kn->kdata.kn_dupfd, NULL) < 0) {
dbg_perror("epoll_ctl(2)");
- return (-1);
+ /* Always release the channel during registration teardown. */
}
}
@@ -302,4 +302,3 @@ const struct filter evfilt_proc = {
evfilt_proc_knote_enable,
evfilt_proc_knote_disable,
};
-
diff --git a/src/linux/read.c b/src/linux/read.c
index b6d2b73..8c7e97d 100644
--- a/src/linux/read.c
+++ b/src/linux/read.c
@@ -15,6 +15,11 @@
*/
#include <errno.h>
+#ifdef DARLING
+#include <darling/emulation/common/simple.h>
+#define LOLKQ(...) do { } while (0)
+#else
+#endif
#include <fcntl.h>
#include <sys/ioctl.h>
#include <pthread.h>
@@ -58,6 +63,7 @@ get_eof_offset(int fd)
int
evfilt_read_copyout(struct kevent64_s *dst, struct knote *src, void *ptr)
{
+ LOLKQ("read-copyout enter\n");
struct epoll_event * const ev = (struct epoll_event *) ptr;
/* Special case: for regular files, return the offset from current position to end of file */
@@ -144,8 +150,10 @@ evfilt_read_knote_create(struct filter *filt, struct knote *kn)
{
struct epoll_event ev;
- if (linux_get_descriptor_type(kn) < 0)
+ if (linux_get_descriptor_type(kn) < 0) {
+ LOLKQ("knote-create: desctype FAILED for ident=%d\n", (int)kn->kev.ident);
return (-1);
+ }
/* Convert the kevent into an epoll_event */
#if defined(HAVE_EPOLLRDHUP)
@@ -189,10 +197,13 @@ evfilt_read_knote_create(struct filter *filt, struct knote *kn)
else {
kn->kdata.kn_dupfd = _dup_4libkqueue(kn->kev.ident);
fcntl(kn->kdata.kn_dupfd, F_SETFD, FD_CLOEXEC);
+ LOLKQ("knote-create ident=%d dupfd=%d events=0x%x\n", (int)kn->kev.ident, (int)kn->kdata.kn_dupfd, ev.events);
if (epoll_ctl(kn->kn_epollfd, EPOLL_CTL_ADD, kn->kdata.kn_dupfd, &ev) < 0) {
dbg_printf("epoll_ctl(2): %s", strerror(errno));
+ LOLKQ("knote-create FAILED dupfd=%d errno=%d\n", (int)kn->kdata.kn_dupfd, (int)errno);
return (-1);
}
+ LOLKQ("knote-create OK ident=%d dupfd=%d\n", (int)kn->kev.ident, (int)kn->kdata.kn_dupfd);
}
return 0;
diff --git a/src/linux/user.c b/src/linux/user.c
index f109ce4..7285f2d 100644
--- a/src/linux/user.c
+++ b/src/linux/user.c
@@ -102,10 +102,14 @@ linux_evfilt_user_copyout(struct kevent64_s *dst, struct knote *src, void *ptr U
}
if (src->kev.flags & EV_CLEAR)
src->kev.fflags &= ~NOTE_TRIGGER;
- if (src->kev.flags & (EV_DISPATCH | EV_CLEAR | EV_ONESHOT)) {
- if (eventfd_lower(src->kdata.kn_eventfd) < 0)
- return (-1);
- }
+ /* darling-lol: Chromium registers its EVFILT_USER wakeup with EV_ADD only.
+ * A returned user knote means its eventfd was readable, regardless
+ * of whether NOTE_TRIGGER is still visible in kev.fflags. A concurrent
+ * trigger can raise the eventfd just before copyout clears the shared bit,
+ * otherwise leaving a readable fd with no NOTE_TRIGGER forever. */
+ if (eventfd_lower(src->kdata.kn_eventfd) < 0)
+ return (-1);
+ src->kev.fflags &= ~NOTE_TRIGGER;
if (src->kev.flags & EV_DISPATCH)
src->kev.fflags &= ~NOTE_TRIGGER;
diff --git a/src/linux/vnode.c b/src/linux/vnode.c
index 58b1544..8337ea0 100644
--- a/src/linux/vnode.c
+++ b/src/linux/vnode.c
@@ -95,7 +95,10 @@ add_watch(struct filter *filt, struct knote *kn)
return (-1);
/* Convert the fflags to the inotify mask */
- mask = IN_CLOSE;
+ /* Closing another descriptor for this inode is not a vnode event.
+ * Watching IN_CLOSE used to poison kn_flags with EV_ONESHOT (the same
+ * bit as KNFL_KNOTE_DELETED), leaving an undrained epoll registration. */
+ mask = IN_ATTRIB;
if (kn->kev.fflags & NOTE_DELETE)
mask |= IN_ATTRIB | IN_DELETE_SELF;
if (kn->kev.fflags & NOTE_WRITE)
@@ -178,18 +181,9 @@ evfilt_vnode_copyout(struct kevent64_s *dst, struct knote *src, void *ptr UNUSED
return (0);
}
- /* Check if the watched file has been closed, and
- XXX-this may not exactly match the kevent() behavior if multiple file de
-scriptors reference the same file.
- */
- if (evt->mask & IN_CLOSE_WRITE || evt->mask & IN_CLOSE_NOWRITE) {
- src->kn_flags |= EV_ONESHOT; /* KLUDGE: causes the knote to be deleted */
- dst->filter = EVFILT_DROP_POSTPROCESS; /* KLUDGE: causes the event to be discarded */
- return (0);
- }
-
kevent_int_to_64(&src->kev, dst);
dst->data = 0;
+ dst->fflags = 0;
/* No error checking because fstat(2) should rarely fail */
//FIXME: EINTR
@@ -203,7 +197,7 @@ scriptors reference the same file.
if (sb.st_nsize == 0 && src->kev.fflags & NOTE_TRUNCATE)
dst->fflags |= NOTE_TRUNCATE;
#endif
- if (sb.st_size > src->data.vnode.size && src->kev.fflags & NOTE_WRITE)
+ if (sb.st_size > src->data.vnode.size && src->kev.fflags & NOTE_EXTEND)
dst->fflags |= NOTE_EXTEND;
src->data.vnode.nlink = sb.st_nlink;
src->data.vnode.size = sb.st_size;
@@ -227,6 +221,7 @@ scriptors reference the same file.
if (evt->mask & IN_DELETE_SELF && src->kev.fflags & NOTE_DELETE)
dst->fflags |= NOTE_DELETE;
+ if (!dst->fflags) dst->filter = EVFILT_DROP;
return (0);
}
Reference in new issue View Git Blame Copy Permalink