POLLERR and POLLHUP handling in rpc_service() could not deal with
session failures or auto reconnect.
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
It makes no sense to have socket.c keep invoking this callback over and over.
Just change it to become one-shot.
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
The linux kernel does not check the UDP checksum until the application tries
to read if from the socket.
This means that the socket might be readable, but when we try to read
the data, or inspect how much data is available, the packets will be discarded
by the kernel.
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
If we are trying to read (part of?) the RM, we can not assume that as long
as recv() returned non-error that we have the full RM.
We must check before we proceed to try to read the actual PDU data.
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
We can not have a static rpc->inbuf buffer since that will no longer guarantee
that the received buffer is valid for the duration of callbacks.
One of the problems is that if we issue new (sync) RPCs from within a
callback, that will overwrite and invalidate the receive buffer that
we passed to the callback.
Revert "init: do not leak rpc->inbuf"
This reverts commit f7bc4c8bb1.
Revert "socket: we have to use memmove in rpc_read_from_socket"
This reverts commit 24429e95b8.
Revert "socket: make rpc->inbuf static and simplify receive logic"
This reverts commit 7000a0aa04.
This funciton is called from rpc_service when it has detected that
a socket has errored out during reading/writing.
However, since this fucntion returns 0 (==success) for the case where
autoreconnect is not enabled, this means that for an errored socket we
will return 0 (==success) from rpc_service() back to the application.
Change rpc_reconnect_requeue to return -1 when invoked and autoreconnect
is disabled so that applications will receive an error back from rpc_service.
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
There is no guarantee that we get the same fd again when
reestablishing a session. But if the fd changes during a
reconnect we might up with a client application busy polling
on the old fd.
Qemu registers a read handler on the current fd, but is
not realizing fd changes. So we busy poll on the old fd for good.
Things are working (except for the busy polling) until
a drain all is issued. At this point Qemu deadlocks.
Signed-off-by: Peter Lieven <pl@kamp.de>
otherwise we end up eating up all socket errors in rpc_service and then
believe we are connected, but the next call to rpc_read_from_socket
fails because the socket is closed. we then reconnect anyway.
Signed-off-by: Peter Lieven <pl@kamp.de>
the requeueing code is broken because we access pdu->next
after we mangled it in rpc_return_to_queue.
This leads to losing of waitqueue elements and more severe
a deadlock as soon as more than one waitpdu queue has elements.
Reason for that is that the first elements of the first
two queues are linked to each other.
Example:
waitpdu[0]->head = pduA ; pduA->next = pduB; pduB->next = NULL;
waitpdu[1]->head = pduC ; pduC->next = NULL;
outqueue->head = NULL;
After the for loop for waitpdu[0] queue the outqueue looks like
outqueue->head = pduA; pduA->next = NULL;
At this point pduB is lost!
In the for loop for waitpdu[1] queue the outqueue looks like this
after the first iteration:
outqueue->head = pduC; pduC->next = pduA; pduA->next = NULL;
We now fetch pdu->next of pduC which is pduA.
In the next iteration we put pduA in front of pduC. pduA->next
is then pduC and pduC->next is pduA. => Deadlock.
Signed-off-by: Peter Lieven <pl@kamp.de>
An EOF is signalled through a POLLIN event and subsequen recvs return
always 0. Handle this condition and reconnect. Otherwise we might
deadlock here.
Signed-off-by: Peter Lieven <pl@kamp.de>
only logging to stderr is supported at the moment. Per default
there is no output. Its possible to set the log level via
debug url parameter.
Example:
nfs-ls nfs://127.0.0.1/export?debug=2
Signed-off-by: Peter Lieven <pl@kamp.de>
the write limit of libnfs has been 1M since a long time.
Restrict rtmax and wrmax to 1M and error out otherwise.
Limit the PDU size when reading from socket to rule out
malicious servers forcing us to allocate a lot of memory.
Signed-off-by: Peter Lieven <pl@kamp.de>
under Linux poll might return POLLIN even if there are no bytes available for read.
See select(2) manpage for surious readiness under BUGS.
As a consequence we start dropping TCP connections which are still alive.
Signed-off-by: Peter Lieven <pl@kamp.de>
Conflicts:
lib/socket.c
Update the configure to add some sanity -W arguments.
A good start is probably :
-Wall -Werror -Wshadow -Wno-write-strings -Wstrict-prototypes
-Wpointer-arith -Wcast-align -Wno-strict-aliasing
Fixup the paces in the code that triggers.
(one of which is readahead code which is perhaps broken?)
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
Include the proper headers to fix warnings like:
libnfs-sync.c:1529:3: warning: implicit declaration of function 'gettimeofday' [-Wimplicit-function-declaration]
libnfs-zdr.c:506:2: warning: implicit declaration of function 'getuid' [-Wimplicit-function-declaration]
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
sys/time.h needs to be protected with an ifdef
Signed-off-by: Ronnie Sahlberg <ronniesahlberg@gmail.com>
commit 1c1e09a completely broke connects for non broadcast
traffic since it forgot to copy the server address into
the socket_storage struct.
Signed-off-by: Peter Lieven <pl@kamp.de>
NFS servers can respond to requests in any order, and they do. In our
tests there is also some clustering to the responses; it could be
because eg. requests are served synchronously if the data is in the cache.
Introduce a hash table so that we are able to find the pdu quickly in
all cases, assuming random distribution of the responses.
When making many concurrent requests (as is likely in any performance
criticial application), the use of SLIST_REMOVE and SLIST_ADD_END are
a severe bottleneck because of their linear search.
I considered using a double-linked list but it was unnecessary to
allocate the additional memory for each list entry.
Instead, continue to use a single-linked list but retain:
* a pointer to the end of the list; and
* a pointer to the previous entry during a linear search.
The former would makes append operations O(1) time, and the latter
does the same for removal. We can do this because removal only happens
within the linear search, and there is no random access to the queue.
This adds basic IPv6 support to libnfs.
Since libnfs currently only support PORTMAPPER protocol up to version 2
the IPv6 support only works if the server runs Both MOUNT and NFS protocols
on the same ports for IPv6 as for IPv4.
To get full IPv6 support we need to add support for PORTMAPPER version 3
and use it for discovery when using IPv6
These were uncovered by the previously added __attribute__((format(printf))).
Emacs also removed trailing whitespace while at it.
Signed-off-by: Arne Redlich <arne.redlich@googlemail.com>
This is mainly needed when having to track and control the file descriptors that are used by libnfs, for example when trying to emulate dup2() ontop
of libnfs.
This allows indirect support for a configurable connect timeout.
Linux uses a exponential backoff for SYN retries starting
with 1 second.
This means for a value n for TCP_SYNCNT, the connect will
effectively timeout after 2^(n+1)-1 seconds.
Example:
examples/nfs-ls nfs://10.0.0.1/export?tcp-syncnt=1
Signed-off-by: Peter Lieven <pl@kamp.de>