Hello!
On Wed, Jul 13, 2022 at 08:12:06PM +0400, Roman Arutyunyan wrote:
> Hi,
>
> On Sun, Jul 10, 2022 at 11:35:48AM +0300, Maxim Dounin wrote:
> > Hello!
> >
> > On Fri, Jul 08, 2022 at 07:13:33PM +0000, Lucas Rolff wrote:
> >
> > > I’m having an nginx instance where I utilise the nginx slice
> > > module to slice upstream mp4 files when using proxy_cache.
> > >
> > > However, I have an interesting origin where if sending a range
> > > request (which happens when the slice module is enabled), to a
> > > file that’s less than the slice range, the origin returns a 200
> > > OK, but with the range related headers such as content-range,
> > > but obviously the full file is returned since it’s within the
> > > requested range.
> > >
> > > When playing the MP4s through Google Chrome and Firefox it works
> > > fine when going through the nginx proxy instance, however, it
> > > somehow breaks Safari (both on MacOS, and iOS) - I guess Safari
> > > is more strict.
> > > When playing directly through the origin it works fine in all
> > > browsers.
> > >
> > > The md5 of response from the origin remains the same, so it’s
> > > not that the response itself is an invalid MP4 file, and even if
> > > you compare the cache files on disk with a “working” origin and
> > > the “broken” origin (one sends a 206 Partial Content, another
> > > sends 200 OK) - the content of the cache files remain the same,
> > > except obviously the header section of the cache file.
> > >
> > > The origin returns a 206 status code, only if the file exceeds
> > > the slice size, so if I configure a slice size of 5 megabyte,
> > > only files above 5 megabytes will give 206s. Anything under 5
> > > megabytes will result in a 200 OK with content-range and the
> > > correct content-length,
> > >
> > > Looking in the slice module itself I see:
> > > https://github.com/nginx/nginx/blob/master/src/http/modules/ngx_http_slice_filter_module.c#L116-L126
> > >
> > >
> > > if (r->headers_out.status != NGX_HTTP_PARTIAL_CONTENT) {
> > > if (r == r->main) {
> > > ngx_http_set_ctx(r, NULL, ngx_http_slice_filter_module);
> > > return ngx_http_next_header_filter(r);
> > > }
> > >
> > > ngx_log_error(NGX_LOG_ERR, r->connection->log, 0,
> > > "unexpected status code %ui in slice response",
> > > r->headers_out.status);
> > > return NGX_ERROR;
> > > }
> > >
> > > This seems like the slice module expects a 206 status code to be
> > > returned,
> >
> > For the main request, the code accepts two basic valid variants:
> >
> > - 206, so the slice module will combine multiple responses to
> > range requests as needed;
> >
> > - anything else, so the slice module will give up and simply
> > return the response to the client.
> >
> > If the module sees a non-206 response to a subrequest, this is an
> > error, as the slice module expects underlying resources to be
> > immutable, and does not expect that some ranges can be requested,
> > while some other aren't. This isn't something related to your
> > case though.
> >
> > > however, later in the same function
> > > https://github.com/nginx/nginx/blob/master/src/http/modules/ngx_http_slice_filter_module.c#L200-L211
> > >
> > >
> > > if (r->headers_out.status == NGX_HTTP_PARTIAL_CONTENT) {
> > > if (ctx->start + (off_t) slcf->size <= r->headers_out.content_offset) {
> > > ctx->start = slcf->size
> > > * (r->headers_out.content_offset / slcf->size);
> > > }
> > >
> > > ctx->end = r->headers_out.content_offset
> > > + r->headers_out.content_length_n;
> > >
> > > } else {
> > > ctx->end = cr.complete_length;
> > > }
> > >
> > > There it will do an else statement if the status code isn’t 206.
> > > So would this piece of code ever be reached, since there’s the initial error?
> >
> > Following the initial check, r->headers_out.status is explicitly
> > changed to NGX_HTTP_OK. Later on the
> > ngx_http_next_header_filter() call might again change
> > r->headers_out.status as long as the client used a range request,
> > and this is what checked here.
> >
> > > Additionally I don’t see in RFC7233 that 206 responses are an
> > > absolute requirement, additionally I don’t see content-range
> > > being prohibited/forbidden to be used for 200 OK responses.
> > > Now, if one have a secondary proxy that modifies the response
> > > headers in between the origin returning 200 OK with the
> > > Content-Range header, and then strip out the Content-Range
> > > header, the nginx slice module seems to handle it fine, so
> > > somehow the combination of 200 OK and a Content-Range header
> > > being present seems to break the slice module from functioning.
> > >
> > > I’m just curious why this happens within the slice module, and
> > > if there’s any possible solution for it (like allowing the
> > > combination of 200 OK and Content-Range, since those two would
> > > still indicate that the origin/upstream supports range requests)
> > > - obviously it would be nice to fix the upstream server but
> > > sometimes that’s sadly not possible.
> >
> > >From the above explanation it is probably already clear that
> > "disabling slice when an origin returns 200 OK" is what actually
> > happens.
> >
> > The issue does not appear without the slice module in your testing
> > because the Content-Range header seems to be only present in your
> > backend 200 responses when there was a Range header in the
> > request, and this is what happens only with the slice module.
> >
> > I've done some limited testing with Safari and manually added
> > Content-Range header, and there seem to be at least two issues:
> >
> > - Range filter in nginx does not expect the Content-Range header
> > to be already present in 200 responses and simply adds another
> > one. This results in incorrect range responses with multiple
> > Content-Range headers, and this breaks Safari.
> >
> > - Safari also breaks if its test request with "Range: bytes=0-1"
> > results in 200 with the Content-Range header.
> >
> > My initial fix was to simply disable handling of 200 responses
> > with Content-Range headers in the range filter, so such responses
> > wouldn't be touched at all. This is perfectly correct and
> > probably the most secure thing to do, but does not work with
> > Safari due to the second issue outlined above.
> >
> > Another approach would be to clear pre-existing Content-Range
> > headers in the range filter. This seems to work, at least in my
> > testing. See below for the patch.
> >
> > > I know the parts of the slice module haven’t been touched for
> > > years, so obviously it works for most people, just dipping my
> > > toes here to see if there’s a possible solution other than
> > > disabling slice when an origin returns 200 OK for files smaller
> > > than the slice size.
> >
> > Note that that slice module is generally unsafe to use for
> > arbitrary upstream servers: it relies on expectations which are
> > beyond the HTTP standard requirements. In particular:
> >
> > - It requires resources to be immutable, so different range
> > responses can be combined together.
> >
> > - It does not try to handle edge cases, such as 416 returned by
> > the upstream on empty files (which is correct per RFC, but
> > requires complicated additional handling to convert 416 to 200, so
> > it is better to just return 200 OK).
> >
> > In general, the slice module is to be used only in your own
> > infrastructure when you control the backend and can be sure that
> > the slice module expectations are met. As such, disabling it for
> > backends which do something unexpected might actually be a good
> > idea. On the other hand, in this particular case the nginx
> > behaviour can be adjusted to handle things gracefully.
> >
> > Below is a patch to clear pre-existing Content-Range headers
> > in the range filter. Please test if it works for you.
> >
> > # HG changeset patch
> > # User Maxim Dounin <mdounin@mdounin.ru>
> > # Date 1657439390 -10800
> > # Sun Jul 10 10:49:50 2022 +0300
> > # Node ID 219217ea49a8d648f5cadd046f1b1294ef05693c
> > # Parent 9d98d524bd02a562d9cd83f4e369c7e992c5753b
> > Range filter: clearing of pre-existing Content-Range headers.
> >
> > Some servers might emit Conten-Range header on 200 responses, and this
>
> Missing "t" in "Conten-Range".
Fixed, thnx.
> > does not seem to contradict RFC 9110: as per RFC 9110, the Content-Range
> > header have no meaning for status codes other than 206 and 417. Previously
>
> have -> has
> 417 -> 416
Fixed, thnx.
> > this resulted in duplicate Content-Range headers in nginx responses handled
> > by the range filter. Fix is to clear pre-existing headers.
> >
> > diff --git a/src/http/modules/ngx_http_range_filter_module.c b/src/http/modules/ngx_http_range_filter_module.c
> > --- a/src/http/modules/ngx_http_range_filter_module.c
> > +++ b/src/http/modules/ngx_http_range_filter_module.c
> > @@ -425,6 +425,10 @@ ngx_http_range_singlepart_header(ngx_htt
> > return NGX_ERROR;
> > }
> >
> > + if (r->headers_out.content_range) {
> > + r->headers_out.content_range->hash = 0;
> > + }
> > +
> > r->headers_out.content_range = content_range;
> >
> > content_range->hash = 1;
> > @@ -582,6 +586,11 @@ ngx_http_range_multipart_header(ngx_http
> > r->headers_out.content_length = NULL;
> > }
> >
> > + if (r->headers_out.content_range) {
> > + r->headers_out.content_range->hash = 0;
> > + r->headers_out.content_range = NULL;
> > + }
> > +
> > return ngx_http_next_header_filter(r);
> > }
> >
> > @@ -598,6 +607,10 @@ ngx_http_range_not_satisfiable(ngx_http_
> > return NGX_ERROR;
> > }
> >
> > + if (r->headers_out.content_range) {
> > + r->headers_out.content_range->hash = 0;
> > + }
> > +
> > r->headers_out.content_range = content_range;
> >
> > content_range->hash = 1;
>
> The patch looks ok to me
>
> Tested with proxy_force_ranges.
Pushed to http://mdounin.ru/hg/nginx/.
--
Maxim Dounin
http://mdounin.ru/
_______________________________________________
nginx mailing list -- nginx@nginx.org
To unsubscribe send an email to nginx-leave@nginx.org