You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
cloudfall_nginx_site always enables HTTP/2; make it a switch, off by default #551
The managed vhost template (engine/ansible/roles/cloudfall_nginx_site/templates/domain-site.conf.j2) always enables HTTP/2 on the TLS server block: listen 443 ssl http2; before nginx 1.25.1, http2 on; from it.
NGINX2Root (CVE-2026-123007, published 2026-11-18) is an RCE in nginx's HTTP/2 flow control, affecting nginx 1.9.5 to 1.31.8 and fixed in 1.31.9. Distribution packages with the fix are not out yet, and the mitigation is to stop serving HTTP/2. An operator who removes http2 by hand from a Cloudfall-managed vhost loses the edit on the next domains run, since the role rewrites the file.
Proposed: a role variable cloudfall_nginx_site_http2, default false, that renders HTTP/2 (in the version-appropriate form) only when true. Turning HTTP/2 back on after the fixed nginx is installed is then --extra-vars cloudfall_nginx_site_http2=true (or the inventory), not a code change.
Verified on main (92481fc): engine/ansible/roles/cloudfall_nginx_site/templates/domain-site.conf.j2 always renders HTTP/2 in the TLS server block (listen 443 ssl http2; below nginx 1.25.1, http2 on; from it), with no variable to turn it off, so a hand edit is lost on the next domains run.
The vulnerability this issue cites (CVE-2026-123007, "NGINX2Root") could not be found in any public CVE source on 2026-11-19. The nginx HTTP/2 CVEs that were found this year (CVE-2026-42926, CVE-2026-42055) need proxy_http_version 2 or grpc_pass toward the upstream; this template proxies over HTTP/1.1. So the switch goes in, but its default keeps today's behaviour (see the decided comment below).
The plan:
A role variable cloudfall_nginx_site_http2 (bool), in defaults/main.yml and meta/argument_specs.yml
The TLS server block renders HTTP/2 in its version-appropriate form only when it is true; when false, listen 443 ssl; and listen [::]:443 ssl; with no http2 anywhere, on both sides of 1.25.1
Template tests for both values on 1.22.1 and 1.26.3, golden files unchanged for the default
A CHANGELOG entry and a line in the domains docs on how to turn it off (--extra-vars cloudfall_nginx_site_http2=false or the inventory)
Design. New role variable cloudfall_nginx_site_http2: true (bool). Every existing vhost renders byte for byte as before; an operator who wants HTTP/2 off sets it to false and the next domains run keeps it off. Tested by rendering the template for both values on each nginx version.
@romamo Decided without waiting: the new cloudfall_nginx_site_http2 switch defaults to true, so every existing vhost keeps serving HTTP/2 and turning it off is an operator's choice.
Why: the issue's reason for turning HTTP/2 off by default, CVE-2026-123007, could not be found in any public CVE source, and the nginx HTTP/2 CVEs that do exist this year need HTTP/2 toward the upstream, which this template doesn't use; flipping a default for every user on an unverified advisory is a contract change with no confirmed reason.
Not taken: default false as the issue proposes; every managed site would drop to HTTP/1.1 on its next domains run until the operator sets the variable back.
Reply with a comment starting with veto before 2026-11-20 11:45 UTC to reverse it, adding what you want instead if you like (for example "veto: default false"). A later veto still counts and gets a follow-up.
The fix is in #552: the role takes cloudfall_nginx_site_http2 (bool, default true); set to false in host or group vars, or with --extra-vars cloudfall_nginx_site_http2=false on ansible-playbook, the HTTPS server renders listen 443 ssl; with no http2 on any nginx version. The default vhost is byte-identical (golden files unchanged). The template reads the value through | bool, since a key=value --extra-vars hands the role the string "false". Not yet checked with nginx -t on a real host for the off case; only the rendered text is tested. This issue closes when #552 merges.
#553, a second PR for this issue with default false, is closed unmerged: it contradicts the decision above, which has no veto. #552 (default true) stays the fix and closes this issue when it merges.
cloudfall_nginx_site_http2 defaults to false. The advisory is verified: nginx's own release notes for 1.31.9 (2026-11-18) list "memory corruption in HTTP/2 flow control (CVE-2026-123007)", with the details in https://nginx.org/en/security_advisories.html. It is reached from the client side on any HTTP/2 listener, not through HTTP/2 toward the upstream, so the template's TLS block is exposed on every nginx before 1.31.9, which includes Debian 12's 1.22.1 and Debian 13's 1.26.3 (no fixed Debian package yet).
So the safe default is off, and an operator turns HTTP/2 back on once the fixed nginx is installed. #553 had this shape; as the refusal says, flip the default on #552's branch.
Taken up @romamo's answer on #552 (option 1): #552 now defaults cloudfall_nginx_site_http2 to false (93fc258). Every managed vhost renders without HTTP/2 on its next domains run; true in the inventory or --extra-vars cloudfall_nginx_site_http2=true renders it as before. This supersedes the decided comment above. #552 still waits on #535 (the merge gate reads main red on a stale release run) and closes this issue when it merges; #555 tracks turning HTTP/2 back on by default.
Activity
shipmill-romamo commentedon Nov 19, 2026
Triage: implement.
Verified on main (92481fc):
engine/ansible/roles/cloudfall_nginx_site/templates/domain-site.conf.j2always renders HTTP/2 in the TLS server block (listen 443 ssl http2;below nginx 1.25.1,http2 on;from it), with no variable to turn it off, so a hand edit is lost on the nextdomainsrun.The vulnerability this issue cites (CVE-2026-123007, "NGINX2Root") could not be found in any public CVE source on 2026-11-19. The nginx HTTP/2 CVEs that were found this year (CVE-2026-42926, CVE-2026-42055) need
proxy_http_version 2orgrpc_passtoward the upstream; this template proxies over HTTP/1.1. So the switch goes in, but its default keeps today's behaviour (see the decided comment below).The plan:
cloudfall_nginx_site_http2(bool), indefaults/main.ymlandmeta/argument_specs.ymllisten 443 ssl;andlisten [::]:443 ssl;with nohttp2anywhere, on both sides of 1.25.1--extra-vars cloudfall_nginx_site_http2=falseor the inventory)Design. New role variable
cloudfall_nginx_site_http2: true(bool). Every existing vhost renders byte for byte as before; an operator who wants HTTP/2 off sets it to false and the nextdomainsrun keeps it off. Tested by rendering the template for both values on each nginx version.I'll link the PR here.
shipmill-romamo commentedon Nov 19, 2026
@romamo Decided without waiting: the new
cloudfall_nginx_site_http2switch defaults totrue, so every existing vhost keeps serving HTTP/2 and turning it off is an operator's choice.Why: the issue's reason for turning HTTP/2 off by default, CVE-2026-123007, could not be found in any public CVE source, and the nginx HTTP/2 CVEs that do exist this year need HTTP/2 toward the upstream, which this template doesn't use; flipping a default for every user on an unverified advisory is a contract change with no confirmed reason.
Not taken: default
falseas the issue proposes; every managed site would drop to HTTP/1.1 on its nextdomainsrun until the operator sets the variable back.Reply with a comment starting with
vetobefore 2026-11-20 11:45 UTC to reverse it, adding what you want instead if you like (for example "veto: default false"). A later veto still counts and gets a follow-up.shipmill-romamo commentedon Nov 19, 2026
The fix is in #552: the role takes
cloudfall_nginx_site_http2(bool, defaulttrue); set tofalsein host or group vars, or with--extra-vars cloudfall_nginx_site_http2=falseonansible-playbook, the HTTPS server renderslisten 443 ssl;with nohttp2on any nginx version. The default vhost is byte-identical (golden files unchanged). The template reads the value through| bool, since a key=value--extra-varshands the role the string "false". Not yet checked withnginx -ton a real host for the off case; only the rendered text is tested. This issue closes when #552 merges.shipmill-romamo commentedon Nov 19, 2026
#553, a second PR for this issue with default
false, is closed unmerged: it contradicts the decision above, which has no veto. #552 (defaulttrue) stays the fix and closes this issue when it merges.shipmill-romamo commentedon Nov 19, 2026
veto: default false
Decision by @romamo, relayed by Claude Code.
cloudfall_nginx_site_http2defaults tofalse. The advisory is verified: nginx's own release notes for 1.31.9 (2026-11-18) list "memory corruption in HTTP/2 flow control (CVE-2026-123007)", with the details in https://nginx.org/en/security_advisories.html. It is reached from the client side on any HTTP/2 listener, not through HTTP/2 toward the upstream, so the template's TLS block is exposed on every nginx before 1.31.9, which includes Debian 12's 1.22.1 and Debian 13's 1.26.3 (no fixed Debian package yet).So the safe default is off, and an operator turns HTTP/2 back on once the fixed nginx is installed. #553 had this shape; as the refusal says, flip the default on #552's branch.
Default cloudfall_nginx_site_http2 to false
shipmill-romamo commentedon Nov 19, 2026
Taken up
@romamo's answer on #552 (option 1): #552 now defaultscloudfall_nginx_site_http2tofalse(93fc258). Every managed vhost renders without HTTP/2 on its nextdomainsrun;truein the inventory or--extra-vars cloudfall_nginx_site_http2=truerenders it as before. This supersedes the decided comment above. #552 still waits on #535 (the merge gate reads main red on a stale release run) and closes this issue when it merges; #555 tracks turning HTTP/2 back on by default.