# Background
I was investigating spurious non-deterministic EINTR returns from
various 9p file system operations in a Linux guest served from the
qemu 9p server.
## EINTR, ERESTARTSYS and the linux kernel
When a signal arrives that the Linux kernel needs to deliver to user-space
while a given thread is blocked (in the 9p case waiting for a reply to its
request in 9p_client_rpc -> wait_event_interruptible), it asks whatever
driver is currently running to abort its current operation (in the 9p case
causing the submission of a TFLUSH message) and return to user space.
In these situations, the error message reported is generally ERESTARTSYS.
If the userspace processes specified SA_RESTART, this means that the
system call will get restarted upon completion of the signal handler
delivery (assuming the signal handler doesn't modify the process state
in complicated ways not relevant here). If SA_RESTART is not specified,
ERESTARTSYS gets translated to EINTR and user space is expected to handle
the restart itself.
## The 9p TFLUSH command
The 9p TFLUSH commands requests that the server abort an ongoing operation.
The man page [1] specifies:
```
If it recognizes oldtag as the tag of a pending transaction, it should
abort any pending response and discard that tag.
[...]
When the client sends a Tflush, it must wait to receive the corresponding
Rflush before reusing oldtag for subsequent messages. If a response to the
flushed request is received before the Rflush, the client must honor the
response as if it had not been flushed, since the completed request may
signify a state change in the server
```
In particular, this means that the server must not send a reply with the
orignal tag in response to the cancellation request, because the client is
obligated to interpret such a reply as a coincidental reply to the original
request.
# The bug
When qemu receives a TFlush request, it sets the `cancelled` flag on the
relevant pdu. This flag is periodically checked, e.g. in
`v9fs_co_name_to_path`, and if set, the operation is aborted and the error
is set to EINTR. However, the server then violates the spec, by returning
to the client an Rerror response, rather than discarding the message
entirely. As a result, the client is required to assume that said Rerror
response is a result of the original request, not a result of the
cancellation and thus passes the EINTR error back to user space.
This is not the worst thing it could do, however as discussed above, the
correct error code would have been ERESTARTSYS, such that user space
programs with SA_RESTART set get correctly restarted upon completion of
the signal handler.
Instead, such programs get spurious EINTR results that they were not
expecting to handle.
It should be noted that there are plenty of user space programs that do not
set SA_RESTART and do not correctly handle EINTR either. However, that is
then a userspace bug. It should also be noted that this bug has been
mitigated by a recent commit to the Linux kernel [2], which essentially
prevents the kernel from sending Tflush requests unless the process is about
to die (in which case the process likely doesn't care about the response).
Nevertheless, for older kernels and to comply with the spec, I believe this
change is beneficial.
# Implementation
The fix is fairly simple, just skipping notification of a reply if
the pdu was previously cancelled. We do however, also notify the transport
layer that we're doing this, so it can clean up any resources it may be
holding. I also added a new trace event to distinguish
operations that caused an error reply from those that were cancelled.
One complication is that we only omit sending the message on EINTR errors in
order to avoid confusing the rest of the code (which may assume that a
client knows about a fid if it sucessfully passed it off to pud_complete
without checking for cancellation status). This does mean that if the server
acts upon the cancellation flag, it always needs to set err to EINTR. I
believe this is true of the current code.
[1] https://9fans.github.io/plan9port/man/man9/flush.html
[2] https://github.com/torvalds/linux/commit/9523feac272ccad2ad8186ba4fcc891
Signed-off-by: Keno Fischer <keno@juliacomputing.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
[groug, send a zero-sized reply instead of detaching the buffer]
Signed-off-by: Greg Kurz <groug@kaod.org>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
With the move of some docs/ to docs/devel/ on ac06724a71,
no references were updated.
Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
We shouldn't allow guests to create extended attribute with arbitrary sizes.
On linux hosts, the limit is XATTR_SIZE_MAX. Let's use it.
Signed-off-by: Greg Kurz <groug@kaod.org>
Move all trace-events for files in the hw/9pfs/ directory to
their own file.
Signed-off-by: Daniel P. Berrange <berrange@redhat.com>
Message-id: 1466066426-16657-26-git-send-email-berrange@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Keno Fischer