This adds a new notify channel to the transfuse protocol. It is not
optional yet but could be made to be optional. A notify protocol and
notify channel are required because writing FUSE response messages to
the FUSE device has different semantics from writing asynchronous
notifications. In particular, response message writes only error on
malformed messages, do not take locks, and do not block. In contrast,
asynchronous notifications can error under normal conditions
(e.g. invalidating a cache entry that doesn't exist), can take locks
during the write call, and can block.
If responses and notifications occur in the same thread, the file system
can become deadlocked when syscalls lock resources waiting for a
response and a notification is written that blocks attempting to acquire
those same locks. The response that would unlock the contended lock
could be queued behind the notification write but the notification write
can't unblock until the response is written in the future. This patch
enables file systems to avoid that fate by offering a secondary channel
on which to send notifications.
Signed-off-by: David Sheets <dsheets@docker.com>
We distinguish export suitability requests from bind mount suitability
requests in the transfuse control protocol. This distinction allows us to
permit both bind mounts of empty directories and export mounts onto empty
directories. Addresses docker/pinata#4213.
Signed-off-by: David Sheets <dsheets@docker.com>
Linux xargs calls the command with no arguments if it gets no inputs, which
`docker rmi` complains about. It provides -r / --no-run-if-empty to prevent
this but unfortunately this isn't supported on OSX.
Ignore errors from `docker rmi` so that `make clean` will keep going and clean
up later stuff.
Signed-off-by: Ian Campbell <ian.campbell@docker.com>
Add a mount suitability predicate, a pong reply message type, and a log
notice message type. Also, fixes the multi-line mount point printing issue.
Signed-off-by: David Sheets <dsheets@docker.com>
Mount points are now made dynamically if either the leaf doesn't exist
or the leaf has no children. Any proper prefix of the target mount point may
or may not exist prior to mounting.
Signed-off-by: David Sheets <dsheets@docker.com>
During start-up, logs are sent to syslog and stderr and the console if
syslog is not available. -l logfile will also send the logs to logfile but
only prior to any log trigger events. -m mount_trigger lets the user
specify the mount point to watch for to begin sending logs to
-t triggered_log file which may be on the newly mounted file system. Log
concatenation and rollover are not implemented, yet. Logging functionality
has been separated into transfused_log.c. Locks have been placed around
shared fds. The writer thread is now symmetric to the reader thread instead
of inheriting the parent's thread in order to trigger log switchover after
the file system proxy has started. The coredump size rlimit has been lifted.
Signed-off-by: David Sheets <david.sheets@docker.com>
The event string in the 9p socket file system now contains a 1 byte
channel type immediately preceding the connection ID. This channel type
determines which protocol will be used on the channel -- m for FUSE
protocol, e for events. The event messages are host-initiated and have
the following structure:
2 bytes for total length
2 bytes for path length + NUL (x)
x bytes for path
1 byte for syscall
stderr logging was also changed to syslog-based logging in this patch.
Signed-off-by: David Sheets <david.sheets@docker.com>
