We need to perform some sanity checks on the udev monitor before every use so that we know nothing changed in the meantime. The reason for moving the code to a separate function is to be able to perform the same check from a worker thread that will replace the udevEventHandleCallback in terms of poking udev for device events. Signed-off-by: Erik Skultety <eskultet(a)redhat.com&gt; --- src/node_device/node_device_udev.c | 57 ++++++++++++++++++++++++++------------ 1 file changed, 39 insertions(+), 18 deletions(-) diff --git a/src/node_device/node_device_udev.c b/src/node_device/node_device_udev.c index f4177455c..465d272ba 100644 --- a/src/node_device/node_device_udev.c +++ b/src/node_device/node_device_udev.c @@ -1615,47 +1615,68 @@ udevHandleOneDevice(struct udev_device *device) } -static void -udevEventHandleCallback(int watch ATTRIBUTE_UNUSED, - int fd, - int events ATTRIBUTE_UNUSED, - void *data ATTRIBUTE_UNUSED) +/* the caller must be holding the driver lock prior to calling this function */ +static bool +udevCheckMonitorFD(struct udev_monitor *udev_monitor, int fd)

{ - struct udev_device *device = NULL; - struct udev_monitor *udev_monitor = DRV_STATE_UDEV_MONITOR(driver); - int udev_fd = -1; + int real_fd = -1; - udev_fd = udev_monitor_get_fd(udev_monitor); - if (fd != udev_fd) { + /* sanity check that the monitor socket hasn't changed */ + real_fd = udev_monitor_get_fd(udev_monitor); + + if (real_fd != fd) { udevPrivate *priv = driver->privateData; virReportError(VIR_ERR_INTERNAL_ERROR, _("File descriptor returned by udev %d does not " "match node device file descriptor %d"), - fd, udev_fd); + real_fd, fd); - /* this is a non-recoverable error, let's remove the handle, so that we - * don't get in here again because of some spurious behaviour and report - * the same error multiple times + /* this is a non-recoverable error, thus the event handle has to be + * removed, so that we don't get into the handler again because of some + * spurious behaviour */ virEventRemoveHandle(priv->watch); priv->watch = -1;

- goto cleanup; + return false; } - device = udev_monitor_receive_device(udev_monitor); - if (device == NULL) { + return true; +} + + +static void +udevEventHandleCallback(int watch ATTRIBUTE_UNUSED, + int fd, + int events ATTRIBUTE_UNUSED, + void *data ATTRIBUTE_UNUSED) +{ + struct udev_device *device = NULL; + struct udev_monitor *udev_monitor = NULL; + + nodeDeviceLock(); + udev_monitor = DRV_STATE_UDEV_MONITOR(driver);

+ + if (!udevCheckMonitorFD(udev_monitor, fd)) + goto unlock;> + + if (!(device = udev_monitor_receive_device(udev_monitor))) { virReportError(VIR_ERR_INTERNAL_ERROR, "%s", _("udev_monitor_receive_device returned NULL"));

- goto cleanup; + goto unlock;

} + nodeDeviceUnlock(); udevHandleOneDevice(device); cleanup: udev_device_unref(device); return; + + unlock: + nodeDeviceUnlock(); + goto cleanup; }

On 08/24/2017 07:23 AM, Erik Skultety wrote: > This patch splits udevEventHandleCallback in two (introduces > udevEventHandleThread) in order to be later able to refactor the latter > to actually become a detached thread which will wait some time for the > kernel to create the whole sysfs tree for a device as we cannot do that > in the event loop directly. > > Signed-off-by: Erik Skultety <eskultet(a)redhat.com&gt; > --- > src/node_device/node_device_udev.c | 54 ++++++++++++++++++++++++++------------ > 1 file changed, 37 insertions(+), 17 deletions(-) > > diff --git a/src/node_device/node_device_udev.c b/src/node_device/node_device_udev.c > index 465d272ba..fe943c78b 100644 > --- a/src/node_device/node_device_udev.c > +++ b/src/node_device/node_device_udev.c > @@ -1647,36 +1647,56 @@ udevCheckMonitorFD(struct udev_monitor *udev_monitor, int fd) > > > static void > +udevEventHandleThread(void *opaque ATTRIBUTE_UNUSED) It's not ATTRIBUTE_UNUSED > +{ > + struct udev_device *device = NULL; > + struct udev_monitor *udev_monitor = NULL; > + int fd = (intptr_t) opaque; ^^^ opaque!

> + > + nodeDeviceLock(); > + udev_monitor = DRV_STATE_UDEV_MONITOR(driver); > + > + if (!udevCheckMonitorFD(udev_monitor, fd)) > + goto unlock; > + > + device = udev_monitor_receive_device(udev_monitor); > + if (device == NULL) { > + virReportError(VIR_ERR_INTERNAL_ERROR, "%s", > + _("udev_monitor_receive_device returned NULL")); > + goto unlock; > + } > + > + nodeDeviceUnlock(); > + udevHandleOneDevice(device); > + > + cleanup: > + udev_device_unref(device); > + return; > + > + unlock: > + nodeDeviceUnlock(); > + goto cleanup; > +} > + > + > +static void > udevEventHandleCallback(int watch ATTRIBUTE_UNUSED, > int fd, > int events ATTRIBUTE_UNUSED, > void *data ATTRIBUTE_UNUSED) > { > - struct udev_device *device = NULL; > struct udev_monitor *udev_monitor = NULL; > > nodeDeviceLock(); > udev_monitor = DRV_STATE_UDEV_MONITOR(driver); > > - if (!udevCheckMonitorFD(udev_monitor, fd)) > - goto unlock; > - > - if (!(device = udev_monitor_receive_device(udev_monitor))) { > - virReportError(VIR_ERR_INTERNAL_ERROR, "%s", > - _("udev_monitor_receive_device returned NULL")); > - goto unlock; > + if (!udevCheckMonitorFD(udev_monitor, fd)) { > + nodeDeviceUnlock(); > + return; > } The above sequence will be done twice - once here and repeated again in HandleThread. I understand why, but as I get to patch 3 - I'm not sure things are going to work as expected.

Adjust udevEventHandleThread to be a proper thread routine running in an infinite loop handling devices. Also introduce udevEventThreadData private structure. Every time there's and incoming event from udev, udevEventHandleCallback only increments the number of events queuing on the monitor and signals the worker thread to handle them. Signed-off-by: Erik Skultety <eskultet(a)redhat.com&gt; --- src/node_device/node_device_udev.c | 132 ++++++++++++++++++++++++++++++------- 1 file changed, 108 insertions(+), 24 deletions(-)

diff --git a/src/node_device/node_device_udev.c b/src/node_device/node_device_udev.c index fe943c78b..444e5be4d 100644 --- a/src/node_device/node_device_udev.c +++ b/src/node_device/node_device_udev.c @@ -59,6 +59,54 @@ struct _udevPrivate { bool privileged; }; +typedef struct _udevEventThreadData udevEventThreadData; +typedef udevEventThreadData *udevEventThreadDataPtr; + +struct _udevEventThreadData { + virMutex lock; + virCond threadCond; + + bool threadQuit; + int monitor_fd; + unsigned long long nevents; /* number of udev events queuing on monitor */ +}; + + +static udevEventThreadDataPtr +udevEventThreadDataNew(int fd) +{ + udevEventThreadDataPtr ret = NULL; + + if (VIR_ALLOC_QUIET(ret) < 0) + return NULL; + + if (virMutexInit(&ret->lock) < 0) + goto cleanup; + + if (virCondInit(&ret->threadCond) < 0) + goto cleanup_mutex; + + ret->monitor_fd = fd; + + return ret; + + cleanup_mutex: + virMutexDestroy(&ret->lock); + + cleanup: + VIR_FREE(ret); + return NULL; +} + + +static void +udevEventThreadDataFree(udevEventThreadDataPtr data) +{ + virMutexDestroy(&data->lock); + virCondDestroy(&data->threadCond); + VIR_FREE(data); +} + static bool udevHasDeviceProperty(struct udev_device *dev, @@ -1647,35 +1695,53 @@ udevCheckMonitorFD(struct udev_monitor *udev_monitor, int fd) static void -udevEventHandleThread(void *opaque ATTRIBUTE_UNUSED) +udevEventHandleThread(void *opaque) { + udevEventThreadDataPtr privateData = opaque; struct udev_device *device = NULL; struct udev_monitor *udev_monitor = NULL; - int fd = (intptr_t) opaque; - nodeDeviceLock(); - udev_monitor = DRV_STATE_UDEV_MONITOR(driver); + /* continue rather than break from the loop on non-fatal errors */ + while (1) { + virMutexLock(&privateData->lock); + while (privateData->nevents == 0 && !privateData->threadQuit) { + if (virCondWait(&privateData->threadCond, &privateData->lock)) { + virReportSystemError(errno, "%s", + _("handler failed to wait on condition")); + goto cleanup; + } + } - if (!udevCheckMonitorFD(udev_monitor, fd)) - goto unlock; + privateData->nevents--; + virMutexUnlock(&privateData->lock); - device = udev_monitor_receive_device(udev_monitor); - if (device == NULL) { - virReportError(VIR_ERR_INTERNAL_ERROR, "%s", - _("udev_monitor_receive_device returned NULL")); - goto unlock; + nodeDeviceLock(); + udev_monitor = DRV_STATE_UDEV_MONITOR(driver); + + if (!udevCheckMonitorFD(udev_monitor, privateData->monitor_fd)) { + nodeDeviceUnlock(); + goto cleanup; + } + + device = udev_monitor_receive_device(udev_monitor); + nodeDeviceUnlock(); + + if (!device) { + virReportError(VIR_ERR_INTERNAL_ERROR, "%s", + _("udev_monitor_receive_device returned NULL")); + goto next; + } + + udevHandleOneDevice(device); + + next: + udev_device_unref(device); } - nodeDeviceUnlock(); - udevHandleOneDevice(device); - cleanup: udev_device_unref(device); + udevEventThreadDataFree(privateData); return; - - unlock: - nodeDeviceUnlock(); - goto cleanup; } @@ -1683,20 +1749,28 @@ static void udevEventHandleCallback(int watch ATTRIBUTE_UNUSED, int fd, int events ATTRIBUTE_UNUSED, - void *data ATTRIBUTE_UNUSED) + void *opaque) { struct udev_monitor *udev_monitor = NULL; + udevEventThreadDataPtr threadData = opaque; nodeDeviceLock(); udev_monitor = DRV_STATE_UDEV_MONITOR(driver); - if (!udevCheckMonitorFD(udev_monitor, fd)) { + virMutexLock(&threadData->lock); + threadData->threadQuit = true; + virCondSignal(&threadData->threadCond); + virMutexUnlock(&threadData->lock); + nodeDeviceUnlock(); return; } nodeDeviceUnlock(); - udevEventHandleThread((void *)(intptr_t) fd); + virMutexLock(&threadData->lock); + threadData->nevents++; + virCondSignal(&threadData->threadCond); + virMutexUnlock(&threadData->lock); } @@ -1823,6 +1897,9 @@ nodeStateInitialize(bool privileged, { udevPrivate *priv = NULL; struct udev *udev = NULL; + int monitor_fd = -1; + virThread th; + udevEventThreadDataPtr threadData = NULL; if (VIR_ALLOC(priv) < 0) return -1; @@ -1883,6 +1960,14 @@ nodeStateInitialize(bool privileged, 128 * 1024 * 1024); #endif + monitor_fd = udev_monitor_get_fd(priv->udev_monitor); + if (!(threadData = udevEventThreadDataNew(monitor_fd)) || + virThreadCreate(&th, false, udevEventHandleThread, threadData) < 0) { + virReportSystemError(errno, "%s", + _("failed to create udev handling thread")); + goto cleanup; + } + /* We register the monitor with the event callback so we are * notified by udev of device changes before we enumerate existing * devices because libvirt will simply recreate the device if we @@ -1891,9 +1976,8 @@ nodeStateInitialize(bool privileged, * enumeration. The alternative is to register the callback after * we enumerate, in which case we will fail to create any devices * that appear while the enumeration is taking place. */ - priv->watch = virEventAddHandle(udev_monitor_get_fd(priv->udev_monitor), - VIR_EVENT_HANDLE_READABLE, - udevEventHandleCallback, NULL, NULL); + priv->watch = virEventAddHandle(monitor_fd, VIR_EVENT_HANDLE_READABLE, + udevEventHandleCallback, threadData, NULL); if (priv->watch == -1) goto unlock;

On Mon, Aug 28, 2017 at 12:26:08PM -0400, John Ferlan wrote: > > > On 08/24/2017 07:23 AM, Erik Skultety wrote: >> Adjust udevEventHandleThread to be a proper thread routine running in an >> infinite loop handling devices. Also introduce udevEventThreadData >> private structure. >> Every time there's and incoming event from udev, udevEventHandleCallback >> only increments the number of events queuing on the monitor and signals >> the worker thread to handle them. >> >> Signed-off-by: Erik Skultety <eskultet(a)redhat.com&gt; >> --- >> src/node_device/node_device_udev.c | 132 ++++++++++++++++++++++++++++++------- >> 1 file changed, 108 insertions(+), 24 deletions(-) >>

> > Once we get here I'm not sure I understand the udev interaction. I guess > it's not "crystal clear" that the socket relationship is a queue of > device events. I also haven't studied the udev code nor have I been > working through this as much as you have lately - so perhaps something > you've uncovered will help educate me in this manner. Still here's my > thoughts and a small sliver of investigative data... > > So far there's been a 1-to-1 interaction between libvirt and udev event. > With this change there could be an n-to-1 interaction - as in receive @n > devices. > > Up to this point the interaction has been: > > 1. udev event > 2. validate fd/udev_monitor > 3. call udev_monitor_receive_device to get @device > 4. process @device > 5. unref @device > 6. return to udev > > After this point > > 1. udev event > 2. validate fd/udev_monitor > 3. increase event count > 4. signal condition to thread to wakeup > 5. return to udev > > asynchronously in a loop > > 1. wakeup on condition and for each 'event' refcnt > 2. decrease event refcnt > 3. validate fd/udev_monitor > 4. call udev_monitor_receive_device to get @device > 5. process @device > 6. unref @device > > If there's more than one event, does udev_monitor_receive_device > guarantee the order? Are we sure udev buffers/queues in this manner? This should be fairly deterministic as it's not that much of a udev problem as it is a kernel problem (since we're talking about a netlink socket which is internally represented by a 'struct socket' structure in the kernel which in fact implements socket buffers, more specifically - very important - queues). So, if kernel didn't guarantee the order of packets/events on the socket, you'd experience the same problem with calling udev_monitor_receive_device, which basically only does recvmsg(), directly from udevEventHandleCallback. IOW if that was the case, udev itself could get easily confused by the kernel events about a device being deleted that had not been previously added etc. So, yes, the order has to be guaranteed and if it is not, then it's a kernel bug.

> That is since we're not grabbing @device before we return to udev when > the event is signaled, is there any guarantee from udev that *the same* > device will still exist when we do get around to making our call? My > concern being how "long" is the data guaranteed to be there. As I said, unless the socket buffer size limit has been reached, in which case ENOSPC was returned by libudev - I had a BZ on this: https://bugzilla.redhat.com/show_bug.cgi?id=1450960

> > As I read the subsequent patch - I'm thinking perhaps we really want to > keep that 1-to-1 relationship. Could the reason that you're getting > those messages be because we're now calling udev out of the order it > expected? Do we know why we're getting a NULL return? I found: Yeah, I was getting EAGAIN and it wasn't until that point that I realized that the scheduler came into play, IOW the main event loop thread polls the file descriptor list, invoking the corresponding handlers, which means that udevEventHandleCallback just increased the overall number of events waiting to be processed by the handler thread, except that since no one actually extracted the data out of the socket and the thread hadn't been scheduled in the meantime it's still the same data that the main event loop thread finds waiting on the socket the next time it polls the @fd, therefore incrementing the event counter multiple times for the same event. Naturally, when the handler thread finally removes the event from the monitor, not knowing the counter had incremented for this event multiple times, it will fail to remove any more events from the monitor, as none had arrived in the meantime, therefore the EAGAIN.

> > https://sourcecodebrowser.com/udev/141/libudev-monitor_8c.html > > which seems to indicate a number of reasons to return NULL... Maybe udev > is processing another device and while doing so it blocks anyone calling > udev_monitor_receive_device. Conversely while processing an event that > same block wouldn't be present because the expectation is that the > called function will call udev_monitor_receive_device. If this was the case, sooner or later we'd hit the problem anyway. > > I think perhaps the processing thread is fine to have; however, instead > of processing the event it's processing an @device that was managed in > udevEventHandleCallback and put onto some "worker queue". At least that This would work nicely as well, I just went with a simpler solution where I didn't have to create another device queue and only offloaded the whole processing onto the handler thread, so it's just a matter taste :).

Erik > way we know we have a refcnt'd udev_device and we can process in an > expected order. > > John > >> diff --git a/src/node_device/node_device_udev.c b/src/node_device/node_device_udev.c >> index fe943c78b..444e5be4d 100644 >> --- a/src/node_device/node_device_udev.c >> +++ b/src/node_device/node_device_udev.c >> @@ -59,6 +59,54 @@ struct _udevPrivate { >> bool privileged; >> }; >> >> +typedef struct _udevEventThreadData udevEventThreadData; >> +typedef udevEventThreadData *udevEventThreadDataPtr; >> + >> +struct _udevEventThreadData { >> + virMutex lock; >> + virCond threadCond; >> + >> + bool threadQuit; >> + int monitor_fd; >> + unsigned long long nevents; /* number of udev events queuing on monitor */ >> +}; >> + >> + >> +static udevEventThreadDataPtr >> +udevEventThreadDataNew(int fd) >> +{ >> + udevEventThreadDataPtr ret = NULL; >> + >> + if (VIR_ALLOC_QUIET(ret) < 0) >> + return NULL; >> + >> + if (virMutexInit(&ret->lock) < 0) >> + goto cleanup; >> + >> + if (virCondInit(&ret->threadCond) < 0) >> + goto cleanup_mutex; >> + >> + ret->monitor_fd = fd; >> + >> + return ret; >> + >> + cleanup_mutex: >> + virMutexDestroy(&ret->lock); >> + >> + cleanup: >> + VIR_FREE(ret); >> + return NULL; >> +} >> + >> + >> +static void >> +udevEventThreadDataFree(udevEventThreadDataPtr data) >> +{ >> + virMutexDestroy(&data->lock); >> + virCondDestroy(&data->threadCond); >> + VIR_FREE(data); >> +} >> + >> >> static bool >> udevHasDeviceProperty(struct udev_device *dev, >> @@ -1647,35 +1695,53 @@ udevCheckMonitorFD(struct udev_monitor *udev_monitor, int fd) >> >> >> static void >> -udevEventHandleThread(void *opaque ATTRIBUTE_UNUSED) >> +udevEventHandleThread(void *opaque) >> { >> + udevEventThreadDataPtr privateData = opaque; >> struct udev_device *device = NULL; >> struct udev_monitor *udev_monitor = NULL; >> - int fd = (intptr_t) opaque; >> >> - nodeDeviceLock(); >> - udev_monitor = DRV_STATE_UDEV_MONITOR(driver); >> + /* continue rather than break from the loop on non-fatal errors */ >> + while (1) { >> + virMutexLock(&privateData->lock); >> + while (privateData->nevents == 0 && !privateData->threadQuit) { >> + if (virCondWait(&privateData->threadCond, &privateData->lock)) { >> + virReportSystemError(errno, "%s", >> + _("handler failed to wait on condition")); >> + goto cleanup; >> + } >> + } >> >> - if (!udevCheckMonitorFD(udev_monitor, fd)) >> - goto unlock; >> + privateData->nevents--; >> + virMutexUnlock(&privateData->lock); >> >> - device = udev_monitor_receive_device(udev_monitor); >> - if (device == NULL) { >> - virReportError(VIR_ERR_INTERNAL_ERROR, "%s", >> - _("udev_monitor_receive_device returned NULL")); >> - goto unlock; >> + nodeDeviceLock(); >> + udev_monitor = DRV_STATE_UDEV_MONITOR(driver); >> + >> + if (!udevCheckMonitorFD(udev_monitor, privateData->monitor_fd)) { >> + nodeDeviceUnlock(); >> + goto cleanup; >> + } >> + >> + device = udev_monitor_receive_device(udev_monitor); >> + nodeDeviceUnlock(); >> + >> + if (!device) { >> + virReportError(VIR_ERR_INTERNAL_ERROR, "%s", >> + _("udev_monitor_receive_device returned NULL")); >> + goto next; >> + } >> + >> + udevHandleOneDevice(device); >> + >> + next: >> + udev_device_unref(device); >> } >> >> - nodeDeviceUnlock(); >> - udevHandleOneDevice(device); >> - >> cleanup: >> udev_device_unref(device); >> + udevEventThreadDataFree(privateData); >> return; >> - >> - unlock: >> - nodeDeviceUnlock(); >> - goto cleanup; >> } >> >> >> @@ -1683,20 +1749,28 @@ static void >> udevEventHandleCallback(int watch ATTRIBUTE_UNUSED, >> int fd, >> int events ATTRIBUTE_UNUSED, >> - void *data ATTRIBUTE_UNUSED) >> + void *opaque) >> { >> struct udev_monitor *udev_monitor = NULL; >> + udevEventThreadDataPtr threadData = opaque; >> >> nodeDeviceLock(); >> udev_monitor = DRV_STATE_UDEV_MONITOR(driver); >> - >> if (!udevCheckMonitorFD(udev_monitor, fd)) { >> + virMutexLock(&threadData->lock); >> + threadData->threadQuit = true; >> + virCondSignal(&threadData->threadCond); >> + virMutexUnlock(&threadData->lock); >> + >> nodeDeviceUnlock(); >> return; >> } >> nodeDeviceUnlock(); >> >> - udevEventHandleThread((void *)(intptr_t) fd); >> + virMutexLock(&threadData->lock); >> + threadData->nevents++; >> + virCondSignal(&threadData->threadCond); >> + virMutexUnlock(&threadData->lock); >> } >> >> >> @@ -1823,6 +1897,9 @@ nodeStateInitialize(bool privileged, >> { >> udevPrivate *priv = NULL; >> struct udev *udev = NULL; >> + int monitor_fd = -1; >> + virThread th; >> + udevEventThreadDataPtr threadData = NULL; >> >> if (VIR_ALLOC(priv) < 0) >> return -1; >> @@ -1883,6 +1960,14 @@ nodeStateInitialize(bool privileged, >> 128 * 1024 * 1024); >> #endif >> >> + monitor_fd = udev_monitor_get_fd(priv->udev_monitor); >> + if (!(threadData = udevEventThreadDataNew(monitor_fd)) || >> + virThreadCreate(&th, false, udevEventHandleThread, threadData) < 0) { >> + virReportSystemError(errno, "%s", >> + _("failed to create udev handling thread")); >> + goto cleanup; >> + } >> + >> /* We register the monitor with the event callback so we are >> * notified by udev of device changes before we enumerate existing >> * devices because libvirt will simply recreate the device if we >> @@ -1891,9 +1976,8 @@ nodeStateInitialize(bool privileged, >> * enumeration. The alternative is to register the callback after >> * we enumerate, in which case we will fail to create any devices >> * that appear while the enumeration is taking place. */ >> - priv->watch = virEventAddHandle(udev_monitor_get_fd(priv->udev_monitor), >> - VIR_EVENT_HANDLE_READABLE, >> - udevEventHandleCallback, NULL, NULL); >> + priv->watch = virEventAddHandle(monitor_fd, VIR_EVENT_HANDLE_READABLE, >> + udevEventHandleCallback, threadData, NULL); >> if (priv->watch == -1) >> goto unlock; >> >>

Since we have a number of places where we workaround timing issues with devices, attributes (files in general) not being available at the time of processing them by calling usleep in a loop for a fixed number of tries, we could as well have a utility function that would do that. Therefore we won't have to duplicate this ugly workaround even more. Signed-off-by: Erik Skultety <eskultet(a)redhat.com&gt; --- src/libvirt_private.syms | 1 + src/util/virfile.c | 29 +++++++++++++++++++++++++++++ src/util/virfile.h | 2 ++ 3 files changed, 32 insertions(+) diff --git a/src/libvirt_private.syms b/src/libvirt_private.syms index 2149b11b7..b9fb54a79 100644 --- a/src/libvirt_private.syms +++ b/src/libvirt_private.syms @@ -1724,6 +1724,7 @@ virFileStripSuffix; virFileTouch; virFileUnlock; virFileUpdatePerm; +virFileWaitForAccess; virFileWrapperFdClose; virFileWrapperFdFree; virFileWrapperFdNew; diff --git a/src/util/virfile.c b/src/util/virfile.c index 2f28e83f4..15b46bcdd 100644 --- a/src/util/virfile.c +++ b/src/util/virfile.c @@ -4166,3 +4166,32 @@ virFileReadValueString(char **value, const char *format, ...) VIR_FREE(str); return ret; } + + +/** + * virFileWaitForAccess: + * @path: absolute path to a sysfs attribute (can be a symlink) + * @ms: how long to wait (in milliseconds) + * @tries: how many times should we try to wait for @path to become accessible + * + * Checks the existence of @path. In case the file defined by @path + * doesn't exist, we wait for it to appear in @ms milliseconds (for up to + * @tries attempts). + * + * Returns 0 on success, -1 on error, setting errno appropriately. + */ +int +virFileWaitForAccess(const char *path, size_t ms, size_t tries)

+{ + errno = 0; + + /* wait for @path to be accessible in @ms milliseconds, up to @tries */ + while (tries-- > 0 && !virFileExists(path)) { + if (tries == 0 || errno != ENOENT) + return -1; + + usleep(ms * 1000); + } + + return 0; +} diff --git a/src/util/virfile.h b/src/util/virfile.h index 57ceb8072..42630ebb5 100644 --- a/src/util/virfile.h +++ b/src/util/virfile.h @@ -347,6 +347,8 @@ int virFileReadValueScaledInt(unsigned long long *value, const char *format, ... int virFileReadValueString(char **value, const char *format, ...) ATTRIBUTE_FMT_PRINTF(2, 3); +int virFileWaitForAccess(const char *path, size_t ms, size_t tries); + int virFileInData(int fd, int *inData,

If we find ourselves in the situation that the 'add' uevent has been fired earlier than the sysfs tree for a device was created, we should use the best-effort approach and give kernel some predetermined amount of time, thus waiting for the attributes to be ready rather than discarding the device from our device list forever. If those don't appear in the given time frame, we need to move on, since libvirt can't wait indefinitely. Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1463285 Signed-off-by: Erik Skultety <eskultet(a)redhat.com&gt; --- src/node_device/node_device_udev.c | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/src/node_device/node_device_udev.c b/src/node_device/node_device_udev.c index e3a647e3d..96a87f4ab 100644 --- a/src/node_device/node_device_udev.c +++ b/src/node_device/node_device_udev.c @@ -1166,9 +1166,23 @@ udevProcessMediatedDevice(struct udev_device *dev, char *canonicalpath = NULL; virNodeDevCapMdevPtr data = &def->caps->data.mdev; - if (virAsprintf(&linkpath, "%s/mdev_type", udev_device_get_syspath(dev)) < 0) + /* Because of a kernel uevent race, we might get the 'add' event prior to + * the sysfs tree being ready, so any attempt to access any sysfs attribute + * would result in ENOENT and us dropping the device, so let's work around + * it by waiting for the attributes to become available. + */ + + if (virAsprintf(&linkpath, "%s/mdev_type", + udev_device_get_syspath(dev)) < 0) goto cleanup; + if (virFileWaitForAccess(linkpath, 1, 100) < 0) { + virReportSystemError(errno, + _("failed to wait for file '%s' to appear"), + linkpath); + goto cleanup; + } + if (virFileResolveLink(linkpath, &canonicalpath) < 0) { virReportSystemError(errno, _("failed to resolve '%s'"), linkpath); goto cleanup;