The random_choose selection policy is meant to implement
power-of-d-choices: sample d available upstreams uniformly, then pick
the least-loaded of that sample. The sampling loop, however, was not a
correct reservoir sample (Algorithm R): it never filled the first k
reservoir slots unconditionally, instead writing every candidate to a
random slot j = rand(i+1), which can evict an earlier candidate while
leaving another slot nil. It also derived j from the upstream's index
in the pool rather than from the number of available upstreams seen,
skewing the sample whenever unavailable upstreams precede available
ones.
As a result the reservoir frequently held fewer than min(k, available)
upstreams, so the least-load comparison often never happened. Most
notably, with two upstreams and 'random_choose 2' (the canonical
power-of-two-choices setup), half of all requests were routed to the
more-loaded upstream even when it was saturated and the other idle --
identical behavior to plain 'random'. The nil slots this leaves behind
were also the cause of the panic reported in #3810, which was patched
by skipping nils in leastRequests rather than by fixing the sampling.
Replace the loop with a standard Algorithm R reservoir sample over the
available upstreams: fill the first k slots, then replace a random slot
with probability k/n. Every available upstream is now sampled uniformly
and the reservoir always holds min(k, available) candidates.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* reverseproxy: validate on weighted_round_robin policy
Validate that weighted_round_robin has a non-zero total weight.
This prevents configurations such as:
weighted_round_robin 0 0
from being accepted and causing a divide-by-zero panic during request handling.
* test: validation test on zero weight upstreams.
* test: provision called instead of totalweight setting
* reverseproxy: validate on negative upstream weights
* test: regression test on weighted_round_robin selection policy
* reverseproxy: replace placeholders specified for sni while using http3
* add test for placeholder
* reverseproxy: replace placeholders specified for sni while using http3
* add test for placeholder
* reverseproxy: test HTTP/3 SNI host placeholder
---------
Co-authored-by: Zen Dodd <mail@steadytao.com>
* feat: drop headers with underscore in their names
* feat: Caddyfile binding and tests for underscore-in-header drop
Add the `allow_underscore_in_headers` global server option, refine the
doc comment, and cover the filter end-to-end: server-level unit tests
(drop, opt-out, debug log, RFC-7230 space rejection), a fastcgi unit
test for the trimmed header name replacer, and forward_auth integration
tests for both the default-drop and opt-out paths.
* remove allow_underscore_in_headers option for now
Both fallbacks in splitPos relied on golang.org/x/text/search with
search.IgnoreCase, which performs Unicode equivalence matching far beyond
ASCII case folding. Combined with the validated-ASCII guarantee on every
SplitPath entry, that fallback turned non-PHP filenames into PHP scripts:
- when the inner loop hit a non-ASCII byte and the IndexString fallback
returned -1, the loop broke without resetting match=false, so a stale
match=true caused a non-existent .php to be reported (PoC:
"/name.<U+00A1>.txt").
- search.IgnoreCase folded fullwidth, mathematical and circled letters
onto ASCII, so "/shell.<math sans-serif php>",
"/shell.<fullwidth p>hp", "/shell.<circled php>" were all detected as
".php" files.
Replace the fallback with strict byte-level ASCII case-insensitive
matching: any byte >= utf8.RuneSelf in the path can never be part of a
match, since SplitPath entries are validated ASCII-only and lower-cased
in Provision(). This keeps the hot path branch-light and removes the
x/text/search dependency from the main module.
Reported against FrankenPHP as GHSA-3g8v-8r37-cgjm and
GHSA-v4h7-cj44-8fc8. The vulnerable function in this module was adapted
from the same FrankenPHP code.
* reverseproxy: Add ability to clear dynamic upstreams cache during retries
This is an optional interface for dynamic upstream modules to implement if they cache results.
TODO: More documentation; this is an experiment.
* Add some godoc
* Export interface; update godoc
* admin: Redact sensitive request headers in API logs
* Fix govulncheck and typed atomic lint failures
* Sync Go module metadata after dependency downgrade
* add 'root' key to Helper.State for access in frankenphp's `php_server` directive
* clone state before passing it to child directives, but keep sharing it among sibling directives
* propagate named route state from children to parent
* use BlockState to set "root" instead
* gofmt -w .
* go fmt ./...
* here we go
When using copy_headers in a forward_auth block, client-supplied headers with
the same names were not being removed before being forwarded to the backend.
This happens because PR #6608 added a MatchNot guard that skips the Set
operation when the auth service does not return a given header. That guard
prevents setting headers to empty strings, which is the correct behavior,
but it also means a client can send X-User-Id: admin in their request and
if the auth service validates the token without returning X-User-Id, Caddy
skips the Set and the client value passes through unchanged to the backend.
The fix adds an unconditional delete route for each copy_headers entry,
placed just before the existing conditional set route. The delete always runs
regardless of what the auth service returns. The conditional set still only
runs when the auth service provides that header.
The end result is:
- Client-supplied headers are always removed
- When the auth service returns the header, the backend gets that value
- When the auth service does not return the header, the backend sees nothing
Existing behavior is unchanged for any deployment where the auth service
returns all of the configured copy_headers entries.
Fixes GHSA-7r4p-vjf4-gxv4
This refactors the initial approach in PR #7281, replacing the UsagePool
with a dedicated package-level sync.Map and atomic.Int64 to track
in-flight requests without global lock contention.
It also introduces a lookup map in the admin API to fix a potential
O(n^2) iteration over upstreams, ensuring that draining upstreams
are correctly exposed across config reloads without leaking memory.
Co-authored-by: Y.Horie <u5.horie@gmail.com>
reverseproxy: optimize in-flight tracking and admin API
- Replaced sync.RWMutex with sync.Map and atomic.Int64 to avoid lock contention under high RPS.
- Introduced a lookup map in the admin API to fix a potential O(n^2) iteration over upstreams.
When a request arrives via a Unix domain socket (RemoteAddr == "@"),
net.SplitHostPort fails, causing addForwardedHeaders to strip all
X-Forwarded-* headers even when the connection is trusted via
trusted_proxies_unix.
Handle Unix socket connections before parsing RemoteAddr: if untrusted,
strip headers for security; if trusted, let clientIP remain empty (no
peer IP for a Unix socket hop) and fall through to the shared header
logic, preserving the existing XFF chain without appending a spurious
entry.
Amp-Thread-ID: https://ampcode.com/threads/T-019c4225-a0ad-7283-ac56-e2c01eae1103
Co-authored-by: Amp <amp@ampcode.com>
* capture the buffered body once, then reset clonedReq.Body before each retry
* no copy
* keep receiver name
* set the buf to nil after extraction and only return it to pool if not nil
---------
Co-authored-by: WeidiDeng <weidi_deng@icloud.com>
