[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH v2] xen/xenbus: make xs_talkv() interruptible



On 17.12.20 19:25, Andrew Cooper wrote:
On 16/12/2020 08:21, Jürgen Groß wrote:
On 15.12.20 21:59, Andrew Cooper wrote:
On 15/12/2020 11:10, Juergen Gross wrote:
In case a process waits for any Xenstore action in the xenbus driver
it should be interruptible by signals.

Signed-off-by: Juergen Gross <jgross@xxxxxxxx>
---
V2:
- don't special case SIGKILL as libxenstore is handling -EINTR fine
---
   drivers/xen/xenbus/xenbus_xs.c | 9 ++++++++-
   1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/drivers/xen/xenbus/xenbus_xs.c
b/drivers/xen/xenbus/xenbus_xs.c
index 3a06eb699f33..17c8f8a155fd 100644
--- a/drivers/xen/xenbus/xenbus_xs.c
+++ b/drivers/xen/xenbus/xenbus_xs.c
@@ -205,8 +205,15 @@ static bool test_reply(struct xb_req_data *req)
     static void *read_reply(struct xb_req_data *req)
   {
+    int ret;
+
       do {
-        wait_event(req->wq, test_reply(req));
+        ret = wait_event_interruptible(req->wq, test_reply(req));
+
+        if (ret == -ERESTARTSYS && signal_pending(current)) {
+            req->msg.type = XS_ERROR;
+            return ERR_PTR(-EINTR);
+        }

So now I can talk fully about the situations which lead to this, I think
there is a bit more complexity.

It turns out there are a number of issues related to running a Xen
system with no xenstored.

1) If a xenstore-write occurs during startup before init-xenstore-domain
runs, the former blocks on /dev/xen/xenbus waiting for xenstored to
reply, while the latter blocks on /dev/xen/xenbus_backend when trying to
tell the dom0 kernel that xenstored is in dom1.  This effectively
deadlocks the system.

This should be easy to solve: any request to /dev/xen/xenbus should
block upfront in case xenstored isn't up yet (could e.g. wait
interruptible until xenstored_ready is non-zero).

I'm not sure that that would fix the problem.  The problem is that
setting the ring details via /dev/xen/xenbus_backend blocks, which
prevents us launching the xenstored stubdomain, which prevents the
earlier xenbus write being completed.

But _why_ is it blocking? Digging through the code I think it blocks
is xs_suspend() due to the normal xenstore request being pending. If
that request doesn't reach the state to cause blocking in xs_suspend()
all is fine.

So long as /dev/xen/xenbus_backend doesn't block, there's no problem
with other /dev/xen/xenbus activity being pending briefly.


Looking at the current logic, I'm not completely convinced.  Even
finding a filled-in evtchn/gfn doesn't mean that xenstored is actually
ready.

No, but the deadlock is not going to happen anymore (famous last
words).


There are 3 possible cases.

1) PV guest, and details in start_info
2) HVM guest, and details in HVM_PARAMs
3) No details (expected for dom0).  Something in userspace must provide
details at a later point.

So the setup phases go from nothing, to having ring details, to finding
the ring working.

I think it would be prudent to try reading a key between having details
and declaring the xenstored_ready.  Any activity, even XS_ERROR,
indicates that the other end of the ring is listening.

Yes. But I really think the xs_suspend() is the problematic case. And
this will be called _before_ xenstored_ready is being set.



2) If xenstore-watch is running when xenstored dies, it spins at 100%
cpu usage making no system calls at all.  This is caused by bad error
handling from xs_watch(), and attempting to debug found:

Can you expand on "bad error handling from xs_watch()", please?

do_watch() has

     for ( ... ) { // defaults to an infinite loop
         vec = xs_read_watch();
         if (vec == NULL)
             continue;
         ...
     }


My next plan was to experiment with break instead of continue, which
I'll get to at some point.

I'd rather put a sleep() in. Otherwise you might break some use cases.




3) (this issue).  If anyone starts xenstore-watch with no xenstored
running at all, it blocks in D in the kernel.

Should be handled with solution for 1).


The cause is the special handling for watch/unwatch commands which,
instead of just queuing up the data for xenstore, explicitly waits for
an OK for registering the watch.  This causes a write() system call to
block waiting for a non-existent entity to reply.

So while this patch does resolve the major usability issue I found (I
can't even SIGINT and get my terminal back), I think there are issues.

The reason why XS_WATCH/XS_UNWATCH are special cased is because they do
require special handling.  The main kernel thread for processing
incoming data from xenstored does need to know how to associate each
async XS_WATCH_EVENT to the caller who watched the path.

Therefore, depending on when this cancellation hits, we might be in any
of the following states:

1) the watch is queued in the kernel, but not even sent to xenstored yet
2) the watch is queued in the xenstored ring, but not acted upon
3) the watch is queued in the xenstored ring, and the xenstored has seen
it but not replied yet
4) the watch has been processed, but the XS_WATCH reply hasn't been
received yet
5) the watch has been processed, and the XS_WATCH reply received

State 5 (and a little bit) is the normal success path when xenstored has
acted upon the request, and the internal kernel infrastructure is set up
appropriately to handle XS_WATCH_EVENTs.

States 1 and 2 can be very common if there is no xenstored (or at least,
it hasn't started up yet).  In reality, there is either no xenstored, or
it is up and running (and for a period of time during system startup,
these cases occur in sequence).

Yes. this is the reason we can't just reject a user request if xenstored
hasn't been detected yet: it could be just starting.

Right, and I'm not suggesting that we'd want to reject accesses while
xenstored is starting up.



As soon as the XS_WATCH event has been written into the xenstored ring,
it is not safe to cancel.  You've committed to xenstored processing the
request (if it is up).

I'm not sure this is true. Cancelling it might result in a stale watch
in xenstored, but there shouldn't be a problem related to that. In case
that watch fires the event will normally be discarded by the kernel as
no matching watch is found in the kernel's data. In case a new watch
has been setup with the same struct xenbus_watch address (which is used
as the token), then this new watch might fire without the node of the
new watch having changed, but spurious watch events are defined to be
okay (OTOH the path in the event might look strange to the handler).

Watches are a quota'd resource in (at least some) xenstored
configurations.  Losing track of the registration is a resource leak,
even if the kernel can filter and discard the unexpected watch events.

Hmm, true.

The correct way to handle it then would be to mark the request to not
just drop it in case a late answer is arriving, but to do an unwatch.

A similar handling would be required for a transaction start.


If xenstored is actually up and running, its fine and necessary to
block.  The request will be processed in due course (timing subject to
the client and server load).  If xenstored isn't up, blocking isn't ok.

Therefore, I think we need to distinguish "not yet on the ring" from "on
the ring", as our distinction as to whether cancelling is safe, and
ensure we don't queue anything on the ring before we're sure xenstored
has started up.

Does this make sense?

Basically, yes.

Great.  If I get any time, I'll try to look into some fixes along the
above lines.

I won't work on those for the coming 3 weeks, so go ahead. :-)


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: application/pgp-keys

Attachment: OpenPGP_signature
Description: OpenPGP digital signature


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.