Call for testing - nginx fix for "no buffer space in script copy" errors

The recent nginx security update for CVE-2026-42533 introduced a regression: on some configurations, nginx randomly answers a small share of requests with HTTP 500 and logs [alert] ... no buffer space in script copy. Users reported it to us after updating, and we were able to reproduce it on every AlmaLinux release that received the update. We’ve found the root cause and built fixed nginx packages that remove the regression while keeping the CVE-2026-42533 fix fully in place. The packages are now in the AlmaLinux testing repository, and we need your help testing them before they go to production.

The problem

After updating to the nginx packages that fix CVE-2026-42533, some users started seeing errors like this in /var/log/nginx/error.log:

[alert] 12345#0: *678 no buffer space in script copy, client: 192.0.2.10, server: example.com, request: "GET /some/path HTTP/1.1"

Each of these alerts means the request failed with HTTP 500. The failures are intermittent: the same URL can work on one request and fail on the next. On one affected site, between a third and a half of the requests to a single location failed.

The problem shows up when a set or rewrite directive uses a variable from a map marked volatile, and the map value itself contains other variables. For example:

map $uri $cache_key {
    volatile;
    default "$scheme://$host$uri$is_args$args";
}

server {
    location /purge/ {
        set $key $cache_key;
        ...
    }
}

Configurations like this worked fine before the CVE-2026-42533 update.

What went wrong

CVE-2026-42533 is a buffer overrun in nginx’s script engine, the code that builds strings from variables for directives like set, rewrite, return and proxy_pass. The upstream fix adds a check: before copying data into the result buffer, nginx makes sure it doesn’t write past the end of that buffer. To do that, upstream stores the buffer end in each script engine instance.

Adding a field to the script engine structure changes nginx’s ABI, which would break third-party dynamic modules built against the current packages. To avoid that, the backported fix shipped in the CentOS Stream and RHEL packages (and therefore in AlmaLinux) keeps the buffer end in a single global variable instead. That works as long as only one script runs at a time.

But scripts can be nested. When set $key $cache_key copies the value of a volatile map variable, nginx evaluates the map right then, and the map runs its own script to build "$scheme://$host$uri...". That inner script overwrites the global buffer end with the end of its own buffer. When control returns to the outer set, its bounds check compares against the wrong buffer. Whether that check fails depends on where the two buffers happen to land in memory, which is why the errors are random.

Without volatile, the map value is computed once and cached before the outer script starts, so there’s no nesting and no error.

What the fix does

The fix is small. Every function that runs a script (ngx_http_complex_value(), ngx_http_script_run(), and their stream module equivalents) now saves the current global buffer end before it starts and restores it when it finishes. A nested script can still use the global for its own buffer, but the outer script gets its correct value back as soon as the inner one returns. This gives the same result as upstream’s per-engine buffer end:

• The CVE-2026-42533 protection stays fully in place. Every bounds check added by the security fix is still there and still checks against the right buffer. We confirmed that a request that triggers the original overflow is still rejected with HTTP 500 and the same alert.
• The ABI doesn’t change, so existing third-party dynamic modules keep working.
• Nested scripts work again, so configurations like the one above stop producing errors.

Here are our results from 2000 requests against the example configuration above:

We’ve also reported the issue and the fix to Red Hat as RHEL-300689, so that it can be fixed in CentOS Stream and RHEL as well.

Packages in testing

If you use a module stream, the testing repository will offer you the update only for the stream you have enabled. You can check which stream that is with dnf module list nginx.

How to test

We especially welcome testing from anyone who has seen the no buffer space in script copy alert, but every report helps, including “everything still works”.

Install the testing repo

sudo dnf install -y almalinux-release-testing

Update nginx and any installed nginx modules

sudo dnf upgrade 'nginx*' --enablerepo=almalinux-testing

Confirm you have the patched version

rpm -q nginx

The result should match the package for your release in the table above.

Check your configuration and restart nginx

The package update doesn’t restart nginx, so restart it to load the new binary:

sudo nginx -t && sudo systemctl restart nginx

Check the error log

After running for a while with your usual traffic, this should not find any new alerts:

sudo grep 'no buffer space in script copy' /var/log/nginx/error.log

Note: We don’t recommend that you keep the testing repo enabled after you’ve updated the packages, unless you’ve done this on a truly non-production environment. If this is a production environment, you can disable the repo with this command:

sudo dnf config-manager --disable almalinux-testing

Share your results!

If you encounter problems or want to share your testing results, please let us know as soon as you can, either in the Development channel on our Mattermost chat, on bugs.almalinux.org, or by emailing packager@almalinux.org. It’s just as important to report if everything works fine as it is to report bugs, so let us know either way!