Segfault when interpreting cached X-Accel-Redirect response
Maxim Dounin
mdounin at mdounin.ru
Tue Feb 6 11:08:11 UTC 2024
Hello!
On Tue, Feb 06, 2024 at 11:42:40AM +0100, Jan Prachař wrote:
> Hello Maxim,
>
> On Tue, 2024-02-06 at 00:46 +0300, Maxim Dounin wrote:
> > Hello!
> >
> > On Mon, Feb 05, 2024 at 06:01:54PM +0100, Jan Prachař wrote:
> >
> > > Hello,
> > >
> > > thank you for your responses.
> > >
> > > On Sat, 2024-02-03 at 04:25 +0300, Maxim Dounin wrote:
> > > > Hello!
> > > >
> > > > On Fri, Feb 02, 2024 at 01:47:51PM +0100, Jan Prachař wrote:
> > > >
> > > > > On Fri, 2024-02-02 at 12:48 +0100, Jiří Setnička via nginx-devel wrote:
> > > > > > Hello,
> > > > > >
> > > > > > Also, I believe that the core of the problem is because of the
> > > > > > ngx_http_finalize_request(r, NGX_DONE); call in the
> > > > > > ngx_http_upstream_process_headers function. This call is needed when
> > > > > > doing an internal redirect after the real upstream request (to close the
> > > > > > upstream request), but when serving from the cache, there is no upstream
> > > > > > request to close and this call causes ngx_http_set_lingering_close to be
> > > > > > called from the ngx_http_finalize_connection with no active request on
> > > > > > the connection yielding to the segfault.
> > > > >
> > > > > Hello,
> > > > >
> > > > > I am Jiří's colleague, and so I have taken a closer look at the problem. Another
> > > > > indication of the issue is the alert in the error log for non-keepalive connections,
> > > > > stating "http request count is zero while closing request."
> > > > >
> > > > > Upon reviewing the nginx source code, I discovered that the function
> > > > > ngx_http_upstream_finalize_request(), when called with rc = NGX_DECLINED, does not invoke
> > > > > ngx_http_finalize_request(). However, when there is nothing to clean up (u->cleanup ==
> > > > > NULL), it does. Therefore, I believe the appropriate fix is to follow the patch below.
> > > > >
> > > > > Best, Jan Prachař
> > > > >
> > > > > # User Jan Prachař <jan.prachar at gmail.com>
> > > > > # Date 1706877176 -3600
> > > > > # Fri Feb 02 13:32:56 2024 +0100
> > > > > # Node ID 851c994b48c48c9cd3d32b9aa402f4821aeb8bb2
> > > > > # Parent cf3d537ec6706f8713a757df256f2cfccb8f9b01
> > > > > Upstream: Fix "request count is zero" when procesing X-Accel-Redirect
> > > > >
> > > > > ngx_http_upstream_finalize_request(r, u, NGX_DECLINED) should not call
> > > > > ngx_http_finalize_request().
> > > > >
> > > > > diff -r cf3d537ec670 -r 851c994b48c4 src/http/ngx_http_upstream.c
> > > > > --- a/src/http/ngx_http_upstream.c Thu Nov 26 21:00:25 2020 +0100
> > > > > +++ b/src/http/ngx_http_upstream.c Fri Feb 02 13:32:56 2024 +0100
> > > > > @@ -4340,6 +4340,11 @@
> > > > >
> > > > > if (u->cleanup == NULL) {
> > > > > /* the request was already finalized */
> > > > > +
> > > > > + if (rc == NGX_DECLINED) {
> > > > > + return;
> > > > > + }
> > > > > +
> > > > > ngx_http_finalize_request(r, NGX_DONE);
> > > > > return;
> > > > > }
> > > >
> > > > I somewhat agree: the approach suggested by Jiří certainly looks
> > > > incorrect. The ngx_http_upstream_cache_send() function, which
> > > > calls ngx_http_upstream_process_headers() with r->cached set, can
> > > > be used in two contexts: before the cleanup handler is installed
> > > > (i.e., when sending a cached response during upstream request
> > > > initialization) and after it is installed (i.e., when sending a
> > > > stale cached response on upstream errors). In the latter case
> > > > skipping finalization would mean a socket leak.
> > > >
> > > > Still, checking for NGX_DECLINED explicitly also looks wrong, for
> > > > a number of reasons.
> > > >
> > > > First, the specific code path isn't just for "nothing to clean
> > > > up", it's for the very specific case when the request was already
> > > > finalized due to filter finalization, see 5994:5abf5af257a7. This
> > > > code path is not expected to be triggered when the cleanup handler
> > > > isn't installed yet - before the cleanup handler is installed,
> > > > upstream code is expected to call ngx_http_finalize_request()
> > > > directly instead. And it would be semantically wrong to check for
> > > > NGX_DECLINED: if it's here, it means something already gone wrong.
> > > >
> > > > I think the generic issue here is that
> > > > ngx_http_upstream_process_headers(), which is normally used for
> > > > upstream responses and calls ngx_http_upstream_finalize_request(),
> > > > is also used for cached responses. Still, it assumes it is used
> > > > for an upstream response, and calls
> > > > ngx_http_upstream_finalize_request().
> > > >
> > > > As can be seen from the rest of the
> > > > ngx_http_upstream_process_headers() code, apart from the issue
> > > > with X-Accel-Redirect, it can also call
> > > > ngx_http_upstream_finalize_request(NGX_HTTP_INTERNAL_SERVER_ERROR)
> > > > when hh->copy_handler() or ngx_http_upstream_copy_header_line()
> > > > fails. This will similarly end up in
> > > > ngx_http_finalize_request(NGX_DONE) since there is no u->cleanup,
> > > > leading to a request hang. And it would be certainly wrong to
> > > > check for NGX_HTTP_INTERNAL_SERVER_ERROR similarly to NGX_DECLINED
> > > > in your patch, because it can theoretically happen after filter
> > > > finalization.
> > > >
> > > > Proper solution would probably require re-thinking
> > > > ngx_http_upstream_process_headers() interface.
> > > >
> > > > Some preliminary code below: it disables X-Accel-Redirect
> > > > processing altogether if ngx_http_upstream_process_headers() is
> > > > called when returning a cached response (this essentially means
> > > > that "proxy_ignore_headers X-Accel-Expires" is preserved in the
> > > > cache file, which seems to be the right thing to do as we don't
> > > > save responses with X-Accel-Redirect to cache unless it is
> > > > ignored), and returns NGX_ERROR in other places to trigger
> > > > appropriate error handling instead of calling
> > > > ngx_http_upstream_finalize_request() directly (this no longer
> > > > tries to return 500 Internal Server Error response though, as
> > > > doing so might be unsafe after copying some of the cached headers
> > > > to the response).
> > > >
> > > > Please take a look if it works for you.
> > >
> > > The provided patch works as expected, with no observed issues.
> > >
> > > Considering that proxy_ignore_headers for caching headers is preserved with the
> > > cached file, it seems reasonable to extend the same behavior to
> > > X-Accel-Redirect.
> >
> > Yes, such handling is (mostly) in line with some
> > proxy_ignore_headers handling, that is, X-Accel-Expires, Expires,
> > Cache-Control, Set-Cookie, Vary, and X-Accel-Buffering, as these
> > affect creation of a cache file, but not sending an already cached
> > response to clients.
> >
> > Still, X-Accel-Limit-Rate from a cache file will be applied to the
> > response if not ignored by the current configuration. Similarly,
> > X-Accel-Charset is also applied as long as no longer ignored.
> >
> > As such, I mostly consider this to be a neutral argument.
> >
> > Further, we might reconsider X-Accel-Redirect handling if caching
> > of X-Accel-Redirect responses will be introduced (see
> > https://trac.nginx.org/nginx/ticket/407 for a feature request).
> >
> > > From my perspective, the updated code in ngx_http_upstream_process_headers() is
> > > a bit confusing. The function can return NGX_DONE, but this return code is only
> > > handled in one place where ngx_http_upstream_process_headers() is called.
> >
> > I've removed NGX_DONE handling from the other call since NGX_DONE
> > return code isn't possible there due to r->cached being set just
> > before the call.
> >
> > We can instead assume it can be returned and handle appropriately:
> > this will also make handling X-Accel-Redirect from cached files
> > easier if we'll decide to (instead of checking r->cached, we'll
> > have to call ngx_http_upstream_finalize_request(NGX_DECLINED)
> > conditionally, only if u->cleanup is set).
> >
> > > If I may suggest, splitting the function might be helpful – redirect processing
> > > would only occur for direct upstream responses, while the rest of the header
> > > processing would be called always (i.e., also for cached responses).
> >
> > I can't say I like this idea. Processing of X-Accel-Redirect is a
> > part of headers processing, and quite naturally handled in
> > ngx_http_upstream_process_headers(). Moving it to a separate function
> > will needlessly complicate things.
> >
> > > Additionally, I believe the special handling of NGX_DECLINED in
> > > ngx_http_upstream_finalize_request() can be removed. The updated patch is
> > > provided below.
> >
> > Not really. The ngx_http_upstream_finalize_request(NGX_DECLINED)
> > call ensures that the upstream handling is properly finalized,
> > notably the upstream connection is closed. For short responses
> > after X-Accel-Redirect, this might not be important, because the
> > upstream connection will be closed anyway during request
> > finalization. But if the redirected request processing takes a
> > while, the upstream connection will still be open, and might
> > receive further events - leading to unexpected behaviour (not to
> > mention that various upstream timing variables, such as
> > $upstream_response_time, will be wrong).
>
> In my previous patch I replaced
>
> ngx_http_upstream_finalize_request(NGX_DECLINED);
>
> by
>
> r->count++;
> ngx_http_upstream_finalize_request(NGX_DONE);
>
> The upstream connection is still finalized and closed, allowing
> for the removal of the special handling of NGX_DECLINED from
> ngx_http_upstream_finalize_request().
Ah, sorry, missed this.
Yes, r->count++ followed by a real request finalization is a
possible alternative to special handling of NGX_DECLINED without
calling ngx_http_finalize_request(). Still, without special
handling in ngx_http_upstream_finalize_request() this won't be
entirely correct: as can be seen from the code, c->log->action
will be incorrectly set to "sending to client".
> >
> > Below is a patch which preserves proper NGX_DONE processing, and
> > handles X-Accel-Redirect from cached files by checking r->cleanup
> > when calling ngx_http_upstream_finalize_request(NGX_DECLINED). I
> > tend to think this might be the best solution after all, providing
> > better compatibility for further improvements.
> >
> > # HG changeset patch
> > # User Maxim Dounin <mdounin at mdounin.ru>
> > # Date 1707167064 -10800
> > # Tue Feb 06 00:04:24 2024 +0300
> > # Node ID 6e7f0d6d857473517048b8838923253d5230ace0
> > # Parent 631ee3c6d38cfdf97dec67c3d2c457af5d91db01
> > Upstream: fixed X-Accel-Redirect handling from cache files.
> >
> > The X-Accel-Redirect header might appear in cache files if its handling
> > is ignored with the "proxy_ignore_headers" directive. If the cache file
> > is later served with different settings, ngx_http_upstream_process_headers()
> > used to call ngx_http_upstream_finalize_request(NGX_DECLINED), which
> > is not expected to happen before the cleanup handler is installed and
> > resulted in ngx_http_finalize_request(NGX_DONE), leading to unexpected
> > request counter decrement, "request count is zero" alerts, and segmentation
> > faults.
> >
> > Similarly, errors in ngx_http_upstream_process_headers() resulted in
> > ngx_http_upstream_finalize_request(NGX_HTTP_INTERNAL_SERVER_ERROR) being
> > called. This is also not expected to happen before the cleanup handler is
> > installed, and resulted in ngx_http_finalize_request(NGX_DONE) without
> > proper request finalization.
> >
> > Fix is to avoid calling ngx_http_upstream_finalize_request() from
> > ngx_http_upstream_process_headers(), notably when the cleanup handler
> > is not yet installed. Errors are now simply return NGX_ERROR, so the
> > caller is responsible for proper finalization by calling either
> > ngx_http_finalize_request() or ngx_http_upstream_finalize_request().
> > And X-Accel-Redirect handling now does not call
> > ngx_http_upstream_finalize_request(NGX_DECLINED) if no cleanup handler
> > is installed.
> >
> > Reported by Jiří Setnička
> > (https://mailman.nginx.org/pipermail/nginx-devel/2024-February/HWLYHOO3DDB3XTFT6X3GRMXIEJ3SJRUA.html).
> >
> > diff --git a/src/http/ngx_http_upstream.c b/src/http/ngx_http_upstream.c
> > --- a/src/http/ngx_http_upstream.c
> > +++ b/src/http/ngx_http_upstream.c
> > @@ -1087,8 +1087,10 @@ ngx_http_upstream_cache_send(ngx_http_re
> >
> > if (rc == NGX_OK) {
> >
> > - if (ngx_http_upstream_process_headers(r, u) != NGX_OK) {
> > - return NGX_DONE;
> > + rc = ngx_http_upstream_process_headers(r, u);
> > +
> > + if (rc != NGX_OK) {
> > + return rc;
> > }
> >
> > return ngx_http_cache_send(r);
> > @@ -2516,7 +2518,14 @@ ngx_http_upstream_process_header(ngx_htt
> > }
> > }
> >
> > - if (ngx_http_upstream_process_headers(r, u) != NGX_OK) {
> > + rc = ngx_http_upstream_process_headers(r, u);
> > +
> > + if (rc == NGX_DONE) {
> > + return;
> > + }
> > +
> > + if (rc == NGX_ERROR) {
> > + ngx_http_upstream_finalize_request(r, u, NGX_ERROR);
> > return;
> > }
> >
> > @@ -2829,7 +2838,9 @@ ngx_http_upstream_process_headers(ngx_ht
> > if (u->headers_in.x_accel_redirect
> > && !(u->conf->ignore_headers & NGX_HTTP_UPSTREAM_IGN_XA_REDIRECT))
> > {
> > - ngx_http_upstream_finalize_request(r, u, NGX_DECLINED);
> > + if (u->cleanup) {
> > + ngx_http_upstream_finalize_request(r, u, NGX_DECLINED);
> > + }
> >
> > part = &u->headers_in.headers.part;
> > h = part->elts;
>
> Just a note. If you move ngx_http_upstream_finalize_request() bellow
> the for loop that copies upstream headers, then this change is also possible:
>
> @@ -2855,13 +2851,15 @@ ngx_http_upstream_process_headers(ngx_http_request_t *r,
> ngx_http_upstream_t *u)
>
> if (hh && hh->redirect) {
> if (hh->copy_handler(r, &h[i], hh->conf) != NGX_OK) {
> - ngx_http_finalize_request(r,
> - NGX_HTTP_INTERNAL_SERVER_ERROR);
> - return NGX_DONE;
> + return NGX_ERROR;
> }
> }
>
>
I don't think it worth the effort, especially given
ngx_http_finalize_request(NGX_HTTP_NOT_FOUND) below.
[...]
--
Maxim Dounin
http://mdounin.ru/
More information about the nginx-devel
mailing list