filter: fix silent truncation at 131072 bytes for inputs >ZSTD_CStreamInSize

[t0mtaylor]

Closes #25.

When ZSTD_compressStream returns rc>0 (a hint that ~rc bytes are still pending in libzstd's internal buffer, the natural value being ZSTD_CStreamInSize() == 131072), the body filter flips action to FLUSH and re-enters. After ZSTD_flushStream returns rc=0 (drained), the existing block

if (rc == 0 && (ctx->flush || ctx->last)) {
    b->last_buf = ctx->last;
    ctx->done   = ctx->last;
}

unconditionally marks the chain buffer as last and the context as done — even though ctx->buffer_in.pos < ctx->buffer_in.size and more chain links may still be queued. The next filter sees last_buf=1, stops reading, and ZSTD_endStream is never invoked on the remaining input.

Symptom: any single-buf body filter call carrying a chunk larger than ZSTD_CStreamInSize and last_buf=1 silently terminates the zstd frame at exactly 131072 decompressed bytes. Multi-buf streams (where the first chunk has last_buf=0) dodge the bug because the FLUSH-rc=0 branch is gated on ctx->last, so the action is restored to COMPRESS and the next chain link is consumed.

Reproduced live with nginx 1.30.0 + libzstd 1.4.4 against a 141186-byte CSS file via Accept-Encoding: zstd: wire 84969 bytes, decompressed to exactly 131072 bytes. After patch: wire 92265 bytes, decompressed to the full 141186. Verified across input sizes 5 KB / 141 KB / 208 KB / 290 KB / 1.5 MB; all decompress to the full original size.

Two-part fix:

1. Gate the END transition on the current ZSTD_inBuffer being fully drained AND no more chain links queued, so we don't finalize the frame while libzstd still has unconsumed input:

   } else if (ctx->last && ctx->action != END
              && ctx->buffer_in.pos >= ctx->buffer_in.size
              && ctx->in == NULL) {
       ctx->action = END;
       return NGX_AGAIN;
   } else if (ctx->action != END) {
       ctx->action = COMPRESS; /* restore */
   }
   

   The new else if (... != END) clause is required because the original unconditional action = COMPRESS restore would reset the END action immediately after ZSTD_endStream returned rc=0, defeating the done-flag gate below and causing infinite output (observed: 422 MB output for 141 KB input during patch development).

2. Gate the EOF marker on action == END so b->last_buf=1 and ctx->done=1 only fire after ZSTD_endStream has emitted the frame epilogue:

   if (rc == 0
       && (ctx->flush
           || (ctx->last && ctx->action == END))) {
       ...
   }
   

   The mid-stream flush case (ctx->flush) is unaffected — that is the client-flush path, not the EOF path.

[eilandert]

somehow this works for me in proxy mode but not in php-fpm fcgi mode, any idea?

[hsw]

#52 also fixes #25, via a different route. Instead of patching the legacy API, it migrates the filter to ZSTD_compressStream2 (ngx_brotli per-call-op), which drops the input-drain bookkeeping and also closes #23 / #41 in the same change.

Verified the same 131072 case you did. Thanks for the repro write-up — flagging the overlap so the two don't get merged on top of each other.