# haproxy-0004 — http_capture_headers O(H×C) per-request header capture scan

## Ecosystem
haproxy (C)

## Severity
MEDIUM — hot path: called for every HTTP request and response with `capture request header` or `capture response header` directives configured

## Location
`src/http_ana.c`
- Function: `http_capture_headers` (~line 5096)
- Inner loop: `for (h = cap_hdr; h; h = h->next)` at line 5113

## Description

`http_capture_headers` is called on every HTTP request and response when
header capture is configured. It iterates over all H headers in the HTX
message, and for each header performs a linear O(C) walk of the `cap_hdr`
linked list to check for a name match:

```c
static void http_capture_headers(struct htx *htx, char **cap, struct cap_hdr *cap_hdr)
{
    for (pos = htx_get_first(htx); pos != -1; pos = htx_get_next(htx, pos)) {
        // ... get header name n ...
        for (h = cap_hdr; h; h = h->next) {         // O(C) per header
            if (h->namelen && (h->namelen == n.len) &&
                (strncasecmp(n.ptr, h->name, h->namelen) == 0)) {
                // capture value
            }
        }
    }
}
```

Total cost per request: O(H × C) where:
- H = number of headers in the request/response (typically 10-50, up to 100+)
- C = number of configured capture headers (can be dozens with complex configs)

Since this runs on every HTTP request+response, it compounds:
- Large H (100 headers) × large C (50 capture directives) = 5000 strncasecmp calls per request

## Fix

Build a hash map from `cap_hdr` name → `cap_hdr *` at config parse time. On each
request, look up each header name in O(1):

```c
--- a/include/haproxy/proxy-t.h
+++ b/include/haproxy/proxy-t.h
@@ -437,6 +437,8 @@
    struct cap_hdr *req_cap;    /* chained list of request headers to be captured */
    struct cap_hdr *rsp_cap;    /* chained list of response headers to be captured */
+   struct eb_root req_cap_tree;  /* name→cap_hdr for O(log C) lookup */
+   struct eb_root rsp_cap_tree;  /* name→cap_hdr for O(log C) lookup */

--- a/src/http_ana.c
+++ b/src/http_ana.c
@@ -5096,15 +5096,14 @@
 static void http_capture_headers(struct htx *htx, char **cap, struct cap_hdr *cap_hdr)
 {
-   struct cap_hdr *h;
    int32_t pos;

    for (pos = htx_get_first(htx); pos != -1; pos = htx_get_next(htx, pos)) {
        struct htx_blk *blk = htx_get_blk(htx, pos);
        if (htx_get_blk_type(blk) == HTX_BLK_EOH) break;
        if (htx_get_blk_type(blk) != HTX_BLK_HDR) continue;

        struct ist n = htx_get_blk_name(htx, blk);

-       for (h = cap_hdr; h; h = h->next) {          // O(C) walk eliminated
+       struct cap_hdr *h = cap_hdr_lookup(cap_hdr, n); // O(1) hash lookup
+       if (h) {
            if (h->namelen && /* ... match ... */) {
                // capture value
            }
        }
    }
}
```

## Complexity

| Variant | Cost |
|---------|------|
| Before  | O(H × C) per request+response pair |
| After   | O(H) per request+response pair — O(1) hash lookup per header |
| Speedup | C× (number of capture headers) |

## Notes
- With C=1 (one capture directive) this is trivially already O(H); the defect
  matters when C grows with complex proxy configurations
- The cap_hdr list is immutable after config parse, making it safe to index
- HAProxy already uses various tree structures (ceb-trees) for similar purposes