[pkg-uWSGI-devel] Bug#1150018: trixie-pu: package uwsgi/2.0.28-9

[Thomas Goirand]

Package: release.debian.org
Severity: normal
Tags: trixie
X-Debbugs-Cc: uwsgi at packages.debian.org
Control: affects -1 + src:uwsgi
User: release.debian.org at packages.debian.org
Usertags: pu

Dear Stable release team,

[ Reason ]
I'd like to update uwsgi in Debian Stable, because there's
a deffect I'd like to fix. When uwsgi serves TLS, and when
a client uses "Transfer-Encoding: chunked", which happens to
be the case with OpenStack Glance, uwsgi just fails. In its
loop, it just sees incoming TLS bytes, and incorrectly decides
this is an error, and quits.

See Debian bug: https://bugs.debian.org/1149707

[ Impact ]
Anyone willing to use uwsgi with TLS and "Transfer-Encoding: chunked"
will see uwsgi close the connection instead of processing it.

[ Tests ]
I manually checked that this fixed the issue, and it did.

[ Risks ]
Not much. Small patch.

[ Checklist ]
  [x] *all* changes are documented in the d/changelog
  [x] I reviewed all changes and I approve them
  [x] attach debdiff against the package in (old)stable
  [ ] the issue is verified as fixed in unstable

[ Changes ]
Just a small patch for the "eat bytes" loop.

[ Other info ]
I haven't uploaded the fix to unstable. I am well aware that
it is the Stable policy that it must happen. Though I'm very
much ok with the stable release team to say something:

"Yeah, we're ok when this patch reaches testing"

I just think it's probably more efficient to just open the bug
now to have your (potential) approval asap.

Cheers,

Thomas Goirand (zigo)
-------------- next part --------------
diff -Nru uwsgi-2.0.28/debian/changelog uwsgi-2.0.28/debian/changelog
--- uwsgi-2.0.28/debian/changelog	2025-03-02 20:27:35.000000000 +0100
+++ uwsgi-2.0.28/debian/changelog	2026-10-02 15:41:35.000000000 +0200
@@ -1,3 +1,11 @@
+uwsgi (2.0.28-9+deb13u1) unstable; urgency=medium
+
+  * With TLS and slow Transfer-Encoding: chunked mode, uwsgi was returning a
+    500 error. Add a patch to fix this (Closes: #1149707):
+    - 2050_Fix_Transfer-Encoding_chunked_brokenness_over_TLS.patch.
+
+ -- Thomas Goirand <zigo at debian.org>  Fri, 02 Oct 2026 15:41:35 +0200
+
 uwsgi (2.0.28-9) unstable; urgency=medium
 
   [ Alexandre Rossi ]
diff -Nru uwsgi-2.0.28/debian/patches/2050_Fix_Transfer-Encoding_chunked_brokenness_over_TLS.patch uwsgi-2.0.28/debian/patches/2050_Fix_Transfer-Encoding_chunked_brokenness_over_TLS.patch
--- uwsgi-2.0.28/debian/patches/2050_Fix_Transfer-Encoding_chunked_brokenness_over_TLS.patch	1970-01-01 01:00:00.000000000 +0100
+++ uwsgi-2.0.28/debian/patches/2050_Fix_Transfer-Encoding_chunked_brokenness_over_TLS.patch	2026-10-02 15:41:35.000000000 +0200
@@ -0,0 +1,51 @@
+Description: Retry the read after wait_read_hook in uwsgi_chunked_input_recv
+ It took me years to understand why OpenStack Glance wouldn't work over
+ uwsgi, and I finally found out. The issue is in uwsgi itself.
+ .
+ After uwsgi.wait_read_hook() reports the socket as readable,
+ uwsgi_chunked_input_recv() performed exactly one proto_read_body() and
+ returned -1 on ANY failure (including EAGAIN), aborting the chunked
+ input with "unable to receive chunked part". This works reasonably
+ well without TLS.
+ .
+ With TLS, POLLIN only means "some TLS bytes arrived"; OpenSSL routinely
+ consumes a partial record (split across TCP segments) and returns
+ SSL_ERROR_WANT_READ right after POLLIN. A single failed post-wait read
+ then aborted the request, so any Transfer-Encoding: chunked request
+ whose body does not arrive in one burst failed fast with an HTTP 500,
+ while the client was still sending the body.
+ .
+ This can be reproduced with a client that streams a large body in chunks
+ with pauses in between, or generates the body on the fly (e.g. piped
+ from another process), and works fine as soon as the whole body is
+ available at once. This description matches 100% what OpenStack Nova
+ and Glance are doing when doing a backup.
+ .
+ To fix this, this patch loops back to the read/wait cycle on EAGAIN
+ instead: EOF (rlen == 0). Hard errors still abort, while a partial
+ read now waits for the rest of the data.
+Author: Thomas Goirand <zigo at debian.org>
+Date: Fri, 2 Oct 2026 15:29:49 +0200
+Forwarded: https://github.com/unbit/uwsgi/pull/2781
+Bug-Debian: https://bugs.debian.org/1149707
+Last-Update: 2026-10-02
+
+diff --git a/core/chunked.c b/core/chunked.c
+index fe693aad4..1b4af07a3 100644
+--- a/core/chunked.c
++++ b/core/chunked.c
+@@ -42,7 +42,13 @@ static ssize_t uwsgi_chunked_input_recv(struct wsgi_request *wsgi_req, int timeo
+                 if (ret > 0) {
+ 			rlen = wsgi_req->socket->proto_read_body(wsgi_req, wsgi_req->chunked_input_buf->buf + wsgi_req->chunked_input_buf->pos, wsgi_req->chunked_input_buf->len - wsgi_req->chunked_input_buf->pos);
+ 			if (rlen > 0) return rlen;
+-			if (rlen <= 0) return -1;
++			if (rlen == 0) return -1;
++			if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINPROGRESS) {
++				// partial read, i.e. split TLS record or TCP segmentation, wait for the rest
++				continue;
++			}
++			uwsgi_error("uwsgi_chunked_input_recv()");
++			return -1;
+ 		}
+                 if (ret < 0) {
+ 			uwsgi_error("uwsgi_chunked_input_recv()");
diff -Nru uwsgi-2.0.28/debian/patches/series uwsgi-2.0.28/debian/patches/series
--- uwsgi-2.0.28/debian/patches/series	2025-03-02 20:14:37.000000000 +0100
+++ uwsgi-2.0.28/debian/patches/series	2026-10-02 15:41:35.000000000 +0200
@@ -13,3 +13,4 @@
 1017_python3-compat.patch
 2001_ensure_verbose_build.patch
 2002_rename_py_plugins.patch
+2050_Fix_Transfer-Encoding_chunked_brokenness_over_TLS.patch