lighttpd1.4/src/h2.c

/*
 * h2 - HTTP/2 protocol layer
 *
 * Copyright(c) 2020 Glenn Strauss gstrauss()gluelogic.com  All rights reserved
 * License: BSD 3-clause (same as lighttpd)
 */
#include "first.h"
#include "h2.h"

#include <arpa/inet.h>  /* htonl() */
#include <stdint.h>     /* INT32_MAX INT32_MIN */
#include <stdlib.h>
#include <string.h>

#include "base.h"
#include "buffer.h"
#include "chunk.h"
#include "fdevent.h"    /* FDEVENT_STREAM_REQUEST_BUFMIN */
#include "http_header.h"
#include "log.h"
#include "request.h"
#include "response.h"   /* http_response_omit_header() */


static request_st * h2_init_stream (request_st * const h2r, connection * const con);


static void
h2_send_settings_ack (connection * const con)
{
    static const uint8_t settings_ack[] = {
      /* SETTINGS w/ ACK */
      0x00, 0x00, 0x00        /* frame length */
     ,H2_FTYPE_SETTINGS       /* frame type */
     ,H2_FLAG_ACK             /* frame flags */
     ,0x00, 0x00, 0x00, 0x00  /* stream identifier */
    };

    chunkqueue_append_mem(con->write_queue,
                          (const char *)settings_ack, sizeof(settings_ack));
}


__attribute_cold__
static void
h2_send_rst_stream_id (uint32_t h2id, connection * const con, const request_h2error_t e)
{
    union {
      uint8_t c[16];
      uint32_t u[4];          /*(alignment)*/
    } rst_stream = { {        /*(big-endian numbers)*/
      0x00, 0x00, 0x00        /* padding for alignment; do not send */
      /* RST_STREAM */
     ,0x00, 0x00, 0x04        /* frame length */
     ,H2_FTYPE_RST_STREAM     /* frame type */
     ,0x00                    /* frame flags */
     ,0x00, 0x00, 0x00, 0x00  /* stream identifier (fill in below) */
     ,0x00, 0x00, 0x00, 0x00  /* error code;       (fill in below) */
    } };

    rst_stream.u[2] = htonl(h2id);
    rst_stream.u[3] = htonl(e);
    chunkqueue_append_mem(con->write_queue,  /*(+3 to skip over align padding)*/
                          (const char *)rst_stream.c+3, sizeof(rst_stream)-3);
}


__attribute_cold__
static void
h2_send_rst_stream (request_st * const r, connection * const con, const request_h2error_t e)
{
    r->state = CON_STATE_ERROR;
    r->h2state = H2_STATE_CLOSED;
    h2_send_rst_stream_id(r->h2id, con, e);
}


__attribute_cold__
static void
h2_send_goaway_rst_stream (connection * const con)
{
    h2con * const h2c = con->h2;
    const int sent_goaway = h2c->sent_goaway;
    for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
        request_st * const r = h2c->r[i];
        if (r->h2state == H2_STATE_CLOSED) continue;
        /*(XXX: might consider always sending RST_STREAM)*/
        if (!sent_goaway) {
            r->state = CON_STATE_ERROR;
            r->h2state = H2_STATE_CLOSED;
        }
        else /*(also sets r->h2state = H2_STATE_CLOSED)*/
            h2_send_rst_stream(r, con, H2_E_PROTOCOL_ERROR);
    }
}


void
h2_send_goaway (connection * const con, const request_h2error_t e)
{
    /* future: RFC 7540 Section 6.8 notes that server initiating graceful
     * connection shutdown SHOULD send GOAWAY with stream id 2^31-1 and a
     * NO_ERROR code, and later send another GOAWAY with an updated last
     * stream identifier.  (This is not done here, but doing so would be
     * friendlier to clients that send streaming requests which the client
     * is unable to retry.) */

    if (e != H2_E_NO_ERROR)
        h2_send_goaway_rst_stream(con);
    /*XXX: else should send RST_STREAM w/ CANCEL for any active PUSH_PROMISE */

    h2con * const h2c = con->h2;
    if (h2c->sent_goaway && (h2c->sent_goaway > 0 || e == H2_E_NO_ERROR))
        return;
    h2c->sent_goaway = (e == H2_E_NO_ERROR) ? -1 : (int32_t)e;

    union {
      uint8_t c[20];
      uint32_t u[5];          /*(alignment)*/
    } goaway = { {            /*(big-endian numbers)*/
      0x00, 0x00, 0x00        /* padding for alignment; do not send */
      /* GOAWAY */
     ,0x00, 0x00, 0x08        /* frame length */
     ,H2_FTYPE_GOAWAY         /* frame type */
     ,0x00                    /* frame flags */
     ,0x00, 0x00, 0x00, 0x00  /* stream identifier */
     ,0x00, 0x00, 0x00, 0x00  /* last-stream-id (fill in below) */
     ,0x00, 0x00, 0x00, 0x00  /* error code     (fill in below) */
                              /* additional debug data (*); (optional)
                               * adjust frame length if any additional
                               * debug data is sent */
    } };

    goaway.u[3] = htonl(h2c->h2_cid); /* last-stream-id */
    goaway.u[4] = htonl(e);
    chunkqueue_append_mem(con->write_queue,  /*(+3 to skip over align padding)*/
                          (const char *)goaway.c+3, sizeof(goaway)-3);
}


__attribute_cold__
static void
h2_send_goaway_e (connection * const con, const request_h2error_t e)
{
    h2_send_goaway(con, e);
}


static int
h2_recv_goaway (connection * const con, const uint8_t * const s, uint32_t len)
{
    /*(s must be entire GOAWAY frame and len the frame length field)*/
    /*assert(s[3] == H2_FTYPE_GOAWAY);*/
    UNUSED(len);
    if ((s[5] & ~0x80) | s[6] | s[7] | s[8]) { /*(GOAWAY stream id must be 0)*/
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return 0;
    }
    const uint32_t e = (s[13]<< 24) | (s[14]<< 16) | (s[15]<< 8) | s[16];
  #if 0
    /* XXX: debug: could log error code sent by peer */
  #endif
  #if 0
    /* XXX: debug: could log additional debug info (if any) sent by peer */
    if (len > 8) {
    }
  #endif
  #if 0
    /* XXX: could validate/use Last-Stream-ID sent by peer */
    const uint32_t last_id = (s[9] << 24) | (s[10]<< 16) | (s[11]<< 8) | s[12];
  #endif

    /* send PROTOCOL_ERROR back to peer if peer sent an error code
     * (i.e. not NO_ERROR) in order to terminate connection more quickly */
    h2_send_goaway(con, e==H2_E_NO_ERROR ? H2_E_NO_ERROR : H2_E_PROTOCOL_ERROR);
    h2con * const h2c = con->h2;
    if (0 == h2c->rused) return 0;
    return 1;
}


static void
h2_recv_rst_stream (connection * const con, const uint8_t * const s, const uint32_t len)
{
    /*(s must be entire RST_STREAM frame and len the frame length field)*/
    /*assert(s[3] == H2_FTYPE_RST_STREAM);*/
    if (4 != len) {                  /*(RST_STREAM frame length must be 4)*/
        h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
        return;
    }
    const uint32_t id =
      ((s[5] << 24) | (s[6] << 16) | (s[7] << 8) | s[8]) & ~0x80000000u;
    if (0 == id) {                   /*(RST_STREAM id must not be 0)*/
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return;
    }
    h2con * const h2c = con->h2;
    for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
        request_st * const r = h2c->r[i];
        if (r->h2id != id) continue;
        if (r->h2state == H2_STATE_IDLE) {
            /*(RST_STREAM must not be for stream in "idle" state)*/
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
            return;
        }
        /* XXX: ? add debug trace including error code from RST_STREAM ? */
        r->state = CON_STATE_ERROR;
        r->h2state = H2_STATE_CLOSED;
        return;
    }
    /* unknown/inactive stream id
     * XXX: how should we handle RST_STREAM for unknown/inactive stream id?
     * (stream id may have been closed recently and server forgot about it,
     *  but client (peer) sent RST_STREAM prior to receiving stream end from
     *  server)*/
  #if 0
    if (h2c->sent_goaway && h2c->h2_cid < id) return;
    h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
  #else
    if (h2c->h2_cid < id) {
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return;
    }
  #endif
}


static void
h2_recv_ping (connection * const con, uint8_t * const s, const uint32_t len)
{
  #if 0
    union {
      uint8_t c[20];
      uint32_t u[5];          /*(alignment)*/
    } ping = { {              /*(big-endian numbers)*/
      0x00, 0x00, 0x00        /* padding for alignment; do not send */
      /* PING */
     ,0x00, 0x00, 0x08        /* frame length */
     ,H2_FTYPE_PING           /* frame type */
     ,H2_FLAG_ACK             /* frame flags */
     ,0x00, 0x00, 0x00, 0x00  /* stream identifier */
     ,0x00, 0x00, 0x00, 0x00  /* opaque            (fill in below) */
     ,0x00, 0x00, 0x00, 0x00
    } };
  #endif

    /*(s must be entire PING frame and len the frame length field)*/
    /*assert(s[3] == H2_FTYPE_PING);*/
    if (8 != len) {                  /*(PING frame length must be 8)*/
        h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
        return;
    }
    s[5] &= ~0x80; /* reserved bit must be ignored */
    if (s[5] | s[6] | s[7] | s[8]) { /*(PING stream id must be 0)*/
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return;
    }
    if (s[4] & H2_FLAG_ACK) /*(ignore; unexpected if we did not send PING)*/
        return;
    /* reflect PING back to peer with frame flag ACK */
    /* (9 byte frame header plus 8 byte PING payload = 17 bytes)*/
    s[4] = H2_FLAG_ACK;
    chunkqueue_append_mem(con->write_queue, (const char *)s, 17);
}


static void
h2_recv_priority (connection * const con, const uint8_t * const s, const uint32_t len)
{
    /*(s must be entire PRIORITY frame and len the frame length field)*/
    /*assert(s[3] == H2_FTYPE_PRIORITY);*/
    if (5 != len) {                  /*(PRIORITY frame length must be 5)*/
        h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
        return;
    }
    const uint32_t id =
      ((s[5] << 24) | (s[6] << 16) | (s[7] << 8) | s[8]) & ~0x80000000u;
    if (0 == id) {                   /*(PRIORITY id must not be 0)*/
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return;
    }
    const uint32_t prio =
      ((s[9] << 24) | (s[10] << 16) | (s[11] << 8) | s[12]) & ~0x80000000u;
  #if 0
    uint32_t exclusive_dependency = (s[9] & 0x80) ? 1 : 0;
    uint32_t weight = s[13];
  #endif
    h2con * const h2c = con->h2;
    for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
        request_st * const r = h2c->r[i];
        if (r->h2id != id) continue;
        /* XXX: TODO: update priority info */
        if (prio == id) {
            h2_send_rst_stream(r, con, H2_E_PROTOCOL_ERROR);
            return;
        }
        return;
    }
    /* XXX: TODO: update priority info for unknown/inactive stream */
    /*if (h2c->sent_goaway && h2c->h2_cid < id) return;*/
    if (prio == id) {
        h2_send_rst_stream_id(id, con, H2_E_PROTOCOL_ERROR);
        return;
    }
}


static void
h2_recv_window_update (connection * const con, const uint8_t * const s, const uint32_t len)
{
    /*(s must be entire WINDOW_UPDATE frame and len the frame length field)*/
    /*assert(s[3] == H2_FTYPE_WINDOW_UPDATE);*/
    if (4 != len) {                  /*(WINDOW_UPDATE frame length must be 4)*/
        h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
        return;
    }
    const uint32_t id =
      ((s[5] << 24) | (s[6] << 16) | (s[7] << 8) | s[8]) & ~0x80000000u;
    const int32_t v =
      (int32_t)(((s[9] << 24)|(s[10] << 16)|(s[11] << 8)|s[12]) & ~0x80000000);
    request_st *r = NULL;
    if (0 == id)
        r = &con->request;
    else {
        h2con * const h2c = con->h2;
        for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
            request_st * const rr = h2c->r[i];
            if (rr->h2id != id) continue;
            r = rr;
            break;
        }
        /* peer should not send WINDOW_UPDATE for an inactive stream,
         * but RFC 7540 does not explicitly call this out.  On the other hand,
         * since there may be a temporary mismatch in stream state between
         * peers, ignore window update if stream id is unknown/inactive.
         * Also, it is not an error if GOAWAY sent and h2c->h2_cid < id */
        if (NULL == r) {
            if (h2c->h2_cid < id && 0 == h2c->sent_goaway)
                h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
          #if 0
            /*(needed for h2spec if testing with response < 16k+1 over TLS
             * or response <= socket send buffer size over cleartext, due to
             * completing response too quickly for the test frame sequence) */
            if (v == 0)        /* h2spec: 6.9-2   (after we retired id 1) */
                h2_send_rst_stream_id(id, con, H2_E_PROTOCOL_ERROR);
            if (v == INT32_MAX)/* h2spec: 6.9.1-3 (after we retired id 1) */
                h2_send_rst_stream_id(id, con, H2_E_FLOW_CONTROL_ERROR);
          #endif
            return;
        }
        /* MUST NOT be treated as error if stream is in closed state; ignore */
        if (r->h2state == H2_STATE_CLOSED
            || r->h2state == H2_STATE_HALF_CLOSED_LOCAL) return;
    }
    if (0 == v || r->h2_swin > INT32_MAX - v) {
        request_h2error_t e = (0 == v)
          ? H2_E_PROTOCOL_ERROR
          : H2_E_FLOW_CONTROL_ERROR;
        if (0 == id)
            h2_send_goaway_e(con, e);
        else
            h2_send_rst_stream(r, con, e);
        return;
    }
    r->h2_swin += v;
}


static void
h2_send_window_update (connection * const con, uint32_t h2id, const uint32_t len)
{
    if (0 == len) return;
    union {
      uint8_t c[16];
      uint32_t u[4];          /*(alignment)*/
    } window_upd = { {        /*(big-endian numbers)*/
      0x00, 0x00, 0x00        /* padding for alignment; do not send */
      /* WINDOW_UPDATE */
     ,0x00, 0x00, 0x04        /* frame length */
     ,H2_FTYPE_WINDOW_UPDATE  /* frame type */
     ,0x00                    /* frame flags */
     ,0x00, 0x00, 0x00, 0x00  /* stream identifier      (fill in below) */
     ,0x00, 0x00, 0x00, 0x00  /* window update increase (fill in below) */
    } };

    window_upd.u[2] = htonl(h2id);
    window_upd.u[3] = htonl(len);
    chunkqueue_append_mem(con->write_queue,  /*(+3 to skip over align padding)*/
                          (const char *)window_upd.c+3, sizeof(window_upd)-3);
}


static void
h2_parse_frame_settings (connection * const con, const uint8_t *s, uint32_t len)
{
    /*(s and len must be SETTINGS frame payload)*/
    /*(caller must validate frame len, frame type == 0x04, frame id == 0)*/
    h2con * const h2c = con->h2;
    for (; len >= 6; len -= 6, s += 6) {
        uint32_t v = (s[2] << 24) | (s[3] << 16) | (s[4] << 8) | s[5];
        switch (((s[0] << 8) | s[1])) {
          case H2_SETTINGS_HEADER_TABLE_SIZE:
            /* encoder may use any table size <= value sent by peer */
            /* For simple compliance with RFC and constrained memory use,
             * choose to not increase table size beyond the default 4096,
             * but allow smaller sizes to be set and then reset up to 4096,
             * e.g. set to 0 to evict all dynamic table entries,
             * and then set to 4096 to restore dynamic table use */
            if (v > 4096) v = 4096;
            if (v == h2c->s_header_table_size) break;
            h2c->s_header_table_size = v;
            lshpack_enc_set_max_capacity(&h2c->encoder, v);
            break;
          case H2_SETTINGS_ENABLE_PUSH:
            if ((v|1) != 1) { /*(v == 0 || v == 1)*/
                h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
                return;
            }
            h2c->s_enable_push = v;
            break;
          case H2_SETTINGS_MAX_CONCURRENT_STREAMS:
            h2c->s_max_concurrent_streams = v;
            break;
          case H2_SETTINGS_INITIAL_WINDOW_SIZE:
            if (v > INT32_MAX) { /*(2^31 - 1)*/
                h2_send_goaway_e(con, H2_E_FLOW_CONTROL_ERROR);
                return;
            }
            else if (h2c->rused) { /*(update existing streams)*/
                /*(underflow is ok; unsigned integer math)*/
                /*(h2c->s_initial_window_size is >= 0)*/
                int32_t diff =
                  (int32_t)((uint32_t)v - (uint32_t)h2c->s_initial_window_size);
                for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
                    request_st * const r = h2c->r[i];
                    const int32_t swin = r->h2_swin;
                    if (r->h2state == H2_STATE_HALF_CLOSED_LOCAL
                        || r->h2state == H2_STATE_CLOSED) continue;
                    if (diff >= 0
                        ? swin > INT32_MAX - diff
                        : swin < INT32_MIN - diff) {
                        h2_send_rst_stream(r, con, H2_E_FLOW_CONTROL_ERROR);
                        continue;
                    }
                    r->h2_swin += diff;
                }
            }
            h2c->s_initial_window_size = (int32_t)v;
            break;
          case H2_SETTINGS_MAX_FRAME_SIZE:
            if (v < 16384 || v > 16777215) { /*[(2^14),(2^24-1)]*/
                h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
                return;
            }
            h2c->s_max_frame_size = v;
            break;
          case H2_SETTINGS_MAX_HEADER_LIST_SIZE:
            h2c->s_max_header_list_size = v;
            break;
          default:
            break;
        }
    }

    if (len) {
        h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
        return;
    }

    /* caller must send SETTINGS frame with ACK flag,
     * if appropriate, and if h2c->sent_goaway is not set
     * (Do not send ACK for Upgrade: h2c and HTTP2-Settings header) */
}


static void
h2_recv_settings (connection * const con, const uint8_t * const s, const uint32_t len)
{
    /*(s must be entire SETTINGS frame, len must be the frame length field)*/
    /*assert(s[3] == H2_FTYPE_SETTINGS);*/
    if ((s[5] & ~0x80) | s[6] | s[7] | s[8]) {/*(SETTINGS stream id must be 0)*/
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return;
    }

    h2con * const h2c = con->h2;
    if (!(s[4] & H2_FLAG_ACK)) {
        h2_parse_frame_settings(con, s+9, len);
        if (h2c->sent_goaway <= 0)
            h2_send_settings_ack(con);
    }
    else {
        /* lighttpd currently sends SETTINGS in server preface, and not again,
         * so this does not have to handle another SETTINGS frame being sent
         * before receiving an ACK from prior SETTINGS frame.  (If it does,
         * then we will need some sort of counter.) */
        if (0 != len)
            h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
        else if (h2c->sent_settings)
            h2c->sent_settings = 0;
        else /* SETTINGS with ACK for SETTINGS frame we did not send */
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
    }
}


static int
h2_recv_end_data (request_st * const r, connection * const con, const uint32_t alen)
{
    chunkqueue * const reqbody_queue = r->reqbody_queue;
    r->h2state = (r->h2state == H2_STATE_OPEN)
      ? H2_STATE_HALF_CLOSED_REMOTE
      : H2_STATE_CLOSED;
    if (r->reqbody_length == -1)
        r->reqbody_length = reqbody_queue->bytes_in + (off_t)alen;
    else if (r->reqbody_length != reqbody_queue->bytes_in + (off_t)alen) {
        if (0 == reqbody_queue->bytes_out) {
            h2_send_rst_stream(r, con, H2_E_PROTOCOL_ERROR);
            return 0;
        } /* else let reqbody streaming consumer handle truncated reqbody */
    }

    return 1;
}


static int
h2_recv_data (connection * const con, const uint8_t * const s, const uint32_t len)
{
    /*(s must be entire DATA frame, len must be the frame length field)*/
    /*assert(s[3] == H2_FTYPE_DATA);*/

    /* future: consider string refs rather than copying DATA from chunkqueue
     * or try to consume entire chunk, or to split chunks with less copying */

    h2con * const h2c = con->h2;
    const uint32_t id =
      ((s[5] << 24) | (s[6] << 16) | (s[7] << 8) | s[8]) & ~0x80000000u;
    if (0 == id || h2c->h2_cid < id) { /*(RST_STREAM id must not be 0)*/
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return 0;
    }

    uint32_t alen = len; /* actual data len, minus padding */
    uint32_t pad = 0;
    if (s[4] & H2_FLAG_PADDED) {
        pad = s[9];
        if (pad >= len) {
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
            return 0;
        }
        alen -= (1 + pad);
    }

    request_st * const h2r = &con->request;
    if (h2r->h2_rwin <= 0 && 0 != alen) { /*(always proceed if 0 == alen)*/
        /*(connection_state_machine_h2() must ensure con is rescheduled,
         * when backends consume data if con->read_queue is not empty,
         * whether or not con->fd has data to read from the network)*/
        /*(leave frame in cq to be re-read later)*/
        return 0;
    }
    /*(allow h2r->h2_rwin to dip below 0 so that entire frame is processed)*/
    /*(not worried about underflow while
     * SETTINGS_MAX_FRAME_SIZE is small (e.g. 16k or 32k) and
     * SETTINGS_MAX_CONCURRENT_STREAMS is small (h2c->r[8]))*/
    /*h2r->h2_rwin -= (int32_t)len;*//* update connection recv window (below) */

    request_st *r = NULL;
    for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
        request_st * const rr = h2c->r[i];
        if (rr->h2id != id) continue;
        r = rr;
        break;
    }
    chunkqueue * const cq = con->read_queue;
    if (NULL == r) {
        /* XXX: TODO: might need to keep a list of recently retired streams
         * for a few seconds so that if we send RST_STREAM, then we ignore
         * further DATA and do not send connection error, though recv windows
         * still must be updated. */
        if (h2c->h2_cid < id || (!h2c->sent_goaway && 0 != alen))
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        chunkqueue_mark_written(cq, 9+len);
        return 0;
    }

    if (r->h2state == H2_STATE_CLOSED
        || r->h2state == H2_STATE_HALF_CLOSED_REMOTE) {
        h2_send_rst_stream_id(id, con, H2_E_STREAM_CLOSED);
        chunkqueue_mark_written(cq, 9+len);
        h2_send_window_update(con, 0, len); /*(h2r->h2_rwin)*/
        return 1;
    }

    if (r->h2_rwin <= 0 && 0 != alen) {/*(always proceed if 0==alen)*/
        if (r->conf.stream_request_body & FDEVENT_STREAM_REQUEST_BUFMIN) {
            /*(connection_state_machine_h2() must ensure con is rescheduled,
             * when backends consume data if con->read_queue is not empty,
             * whether or not con->fd has data to read from the network)*/
            /*(leave frame in cq to be re-read later)*/
            return 0;
        }
    }
    /*(allow h2r->h2_rwin to dip below 0 so that entire frame is processed)*/
    /*(undeflow will not occur (with reasonable SETTINGS_MAX_FRAME_SIZE used)
     * since windows updated elsewhere and data is streamed to temp files if
     * not FDEVENT_STREAM_REQUEST_BUFMIN)*/
    /*r->h2_rwin -= (int32_t)len;*/
    h2_send_window_update(con, r->h2id, len); /*(r->h2_rwin)*/
    h2_send_window_update(con, 0, len);       /*(h2r->h2_rwin)*/

    chunkqueue * const dst = r->reqbody_queue;

    if (r->reqbody_length >= 0 && r->reqbody_length < dst->bytes_in + alen) {
        /* data exceeds Content-Length specified (client mistake) */
      #if 0 /* truncate */
        alen = r->reqbody_length - dst->bytes_in;
        /*(END_STREAM may follow in 0-length DATA frame or HEADERS (trailers))*/
      #else /* reject */
        h2_send_rst_stream(r, con, H2_E_PROTOCOL_ERROR);
        chunkqueue_mark_written(cq, 9+len);
        return 1;
      #endif
    }

    /*(accounting for mod_accesslog and mod_rrdtool)*/
    chunkqueue * const rq = r->read_queue;
    rq->bytes_in  += (off_t)alen;
    rq->bytes_out += (off_t)alen;

    /* r->conf.max_request_size is in kBytes */
    const off_t max_request_size = (off_t)r->conf.max_request_size << 10;
    if (0 != max_request_size
        && dst->bytes_in + (off_t)alen > max_request_size) {
        if (0 == r->http_status) {
            r->http_status = 413; /* Payload Too Large */
            log_error(r->conf.errh, __FILE__, __LINE__,
              "request-size too long: %lld -> 413",
              (long long) (dst->bytes_in + (off_t)alen));
        }
        chunkqueue_mark_written(cq, 9+len);
        return 1;
    }

    if ((s[4] & H2_FLAG_END_STREAM) && !h2_recv_end_data(r, con, alen)) {
        chunkqueue_mark_written(cq, 9+len);
        return 1;
    }

    chunkqueue_mark_written(cq, 9 + ((s[4] & H2_FLAG_PADDED) ? 1 : 0));

  #if 0
    if (pad) {
        /* XXX: future optimization: if data is at end of chunk, then adjust
         * size of chunk by reducing c->mem->used to avoid copying chunk
         * when it is split (below) since the split would be due to padding
         * (also adjust cq->bytes_out)*/
        /*(might quickly check 9+len == cqlen if cqlen passed in as param)*/
        /*(then check if cq->last contains all of padding, or leave alone)*/
        /*(if handled here, then set pad = 0 here)*/
    }
  #endif

    /*(similar decision logic to that in http_chunk_uses_tempfile())*/
    const chunk * const c = dst->last;
    if ((c && c->type == FILE_CHUNK && c->file.is_temp)
        || dst->bytes_in - dst->bytes_out + alen > 65536) {
        log_error_st * const errh = r->conf.errh;
        if (0 != chunkqueue_steal_with_tempfiles(dst, cq, (off_t)alen, errh)) {
            h2_send_rst_stream(r, con, H2_E_INTERNAL_ERROR);
            return 0;
        }
    }
    else
        chunkqueue_steal(dst, cq, (off_t)alen);

    if (pad)
        chunkqueue_mark_written(cq, pad);
    return 1;
}


__attribute_cold__
static uint32_t
h2_frame_cq_compact (chunkqueue * const cq, uint32_t len)
{
    /*(marked cold since most frames not expect to cross chunk boundary)*/

    /*(must be guaranteed by caller)*/
        /*assert(chunkqueue_length(cq) >= len);*/
        /*assert(cq->first != cq->last);*//*(multiple chunks)*/
    /* caller must guarantee that chunks in chunkqueue are all MEM_CHUNK */

    /* move data to beginning of buffer if offset is large or data is short */
    chunk *c = cq->first;
    uint32_t mlen = buffer_string_length(c->mem);
    if (mlen < c->offset) {
        memmove(c->mem->ptr, c->mem->ptr + c->offset, mlen);
        buffer_string_set_length(c->mem, mlen);
        c->offset = 0;
    }

    /* combine first mem chunk with next non-empty mem chunks up to len
     * (loop if next chunk is empty) */
    /* (modified from connection_handle_read_post_cq_compact()) */
    uint32_t clen = mlen;
    do {
        buffer * const mem = c->next->mem;
        const off_t offset = c->next->offset;
        mlen = buffer_string_length(mem) - (uint32_t)offset;
        force_assert(c->type == MEM_CHUNK);
        force_assert(c->next->type == MEM_CHUNK);
        if (mlen > clen - len) {
            mlen = clen - len;
            buffer_append_string_len(c->mem, mem->ptr+offset, mlen);
            c->next->offset += mlen;
            return len;
        }

        buffer_append_string_len(c->mem, mem->ptr+offset, mlen);
        clen += mlen;
        /*(swap first and second chunk, then remove first chunk)*/
        c->next->offset = c->offset;
        c->next->mem = c->mem;
        c->mem = mem;
        c->offset = offset + (off_t)mlen;
        chunkqueue_remove_finished_chunks(cq);
    } while ((c = cq->first)); /*(need to re-read cq->first)*/
    return clen;
}


__attribute_cold__
static uint32_t
h2_recv_continuation (uint32_t n, uint32_t clen, const off_t cqlen, chunkqueue * const cq, connection * const con)
{
    chunk *c = cq->first;
    uint8_t *s = (uint8_t *)(c->mem->ptr + c->offset);
    uint32_t m = n;
    uint32_t flags;
    h2con * const h2c = con->h2;
    const uint32_t fsize = h2c->s_max_frame_size;
    const uint32_t id =
      ((s[5] << 24) | (s[6] << 16) | (s[7] << 8) | s[8]) & ~0x80000000u;
    do {
        if (cqlen < n+9) return n+9; /* incomplete frame; go on */
        if (clen < n+9) {
            clen = h2_frame_cq_compact(cq, n+9);
            c = cq->first; /*(reload after h2_frame_cq_compact())*/
            s = (uint8_t *)(c->mem->ptr + c->offset);
        }
        if (s[n+3] != H2_FTYPE_CONTINUATION) {
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
            return 0;
        }
        flags = s[n+4];
        const uint32_t flen = (s[n+0]<<16)|(s[n+1]<<8)|s[n+2];
        if (id != (uint32_t)((s[n+5]<<24)|(s[n+6]<<16)|(s[n+7]<<8)|s[n+8])) {
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
            return 0;
        }
        if (flen > fsize) {
            h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
            return 0;
        }
        n += 9+flen;
        if (n >= 65536) { /*(very oversized for hpack)*/
            h2_send_goaway_e(con, H2_E_FRAME_SIZE_ERROR);
            return 0;
        }
        if (clen < n) {
            clen = h2_frame_cq_compact(cq, n);
            c = cq->first; /*(reload after h2_frame_cq_compact())*/
            s = (uint8_t *)(c->mem->ptr + c->offset);
        }
    } while (!(flags & H2_FLAG_END_HEADERS));

    /* If some CONTINUATION frames were concatenated to earlier frames while
     * processing above, but END_HEADERS were not received, then the next time
     * data was read, initial frame size might exceed SETTINGS_MAX_FRAME_SIZE.
     * (This describes the current lighttpd implementation in h2_parse_frames())
     * While a flag could be set and checked to avoid this, such situations of
     * large HEADERS (and CONTINUATION) across multiple network reads is
     * expected to be rare.  Reparse and concatenate below.
     *
     * Aside: why would the authors of RFC 7540 go through the trouble of
     * creating a CONTINUATION frame that must be special-cased when use of
     * CONTINUATION is so restricted e.g. no other intervening frames and
     * that HEADERS and PUSH_PROMISE HPACK must be parsed as a single block?
     * IMHO, it would have been simpler to avoid CONTINUATION entirely, and have
     * a special-case for HEADERS and PUSH_PROMISE to be allowed to exceed
     * SETTINGS_MAX_FRAME_SIZE with implementations providing a different limit.
     * While intermediates would not know such a limit of origin servers,
     * there could have been a reasonable default set with a different SETTINGS
     * parameter aimed just at HEADERS and PUSH_PROMISE.  The parameter
     * SETTINGS_MAX_HEADER_LIST_SIZE could even have been (re)used, had it been
     * given a reasonable initial value instead of "unlimited", since HPACK
     * encoded headers are smaller than the HPACK decoded headers to which the
     * limit SETTINGS_MAX_HEADER_LIST_SIZE applies. */

    n = m; /* reset n to beginning of first CONTINUATION frame */

    /* Eliminate padding from first frame (HEADERS or PUSH_PROMISE) if PADDED */
    if (s[4] & H2_FLAG_PADDED) {
        const uint32_t plen = s[9];
        /* validate padding */
        const uint32_t flen = (s[0]<<16)|(s[1]<<8)|s[2];
        if (flen < 1 + plen + ((s[n+4] & H2_FLAG_PRIORITY) ? 5 : 0)) {
            /* Padding that exceeds the size remaining for the header block
             * fragment MUST be treated as a PROTOCOL_ERROR. */
            h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
            return 0;
        }
        /* set padding to 0 since we will overwrite padding in merge below */
        /* (alternatively, could memmove() 9 bytes of frame header over the
         *  pad length octet, remove PADDED flag, add 1 to c->offset,
         *  add 1 to s, subtract 1 from clen and substract 1 from cqlen,
         *  substract 1 from n) */
        s[9] = 0;
        /* set offset to beginning of padding at end of first frame */
        m -= plen;
    }

    do {
        const uint32_t flen = (s[n+0]<<16)|(s[n+1]<<8)|s[n+2];
        flags = s[n+4];
        memmove(s+m, s+n+9, flen);
        m += flen;
        n += 9+flen;
    } while (!(flags & H2_FLAG_END_HEADERS));
    /* overwrite frame size */
    m -= 9; /*(temporarily remove frame header from len)*/
    s[0] = (m >> 16) & 0xFF;
    s[1] = (m >>  8) & 0xFF;
    s[2] = (m      ) & 0xFF;
    m += 9;
    /* adjust chunk c->mem */
    if (n < clen) { /*(additional frames after CONTINUATION)*/
        memmove(s+m, s+n, clen-n);
        n = m + (clen-n);
    }
    else
        n = m;
    buffer_string_set_length(c->mem, n + (uint32_t)c->offset);

    return m;
}


__attribute_cold__
static request_st *
h2_recv_trailers_r (connection * const con, h2con * const h2c, const uint32_t id, const uint32_t flags)
{
    /* rant: RFC 7230 HTTP/1.1 trailer-part would have been much simpler
     * to support in RFC 7540 HTTP/2 as a distinct frame type rather than
     * HEADERS.  As trailers are not known at the time the request is made,
     * reuse of such trailers is limited and so a theoretical TRAILERS frame
     * could have been implemented without HPACK encoding, and would have
     * been more straightforward to implement than overloading and having to
     * handle multiple cases for HEADERS.  TRAILERS support could then also
     * be optional, like in HTTP/1.1 */
    request_st *r = NULL;
    for (uint32_t i = 0, rused = h2c->rused; i < rused; ++i) {
        request_st * const rr = h2c->r[i];
        if (rr->h2id != id) continue;
        r = rr;
        break;
    }
    if (NULL == r) {
        h2_send_goaway_e(con, H2_E_PROTOCOL_ERROR);
        return NULL;
    }
    if (r->h2state != H2_STATE_OPEN
        && r->h2state != H2_STATE_HALF_CLOSED_LOCAL) {
        h2_send_rst_stream(r, con, H2_E_STREAM_CLOSED);
        return NULL;
    }
    /* RFC 7540 is not explicit in restricting HEADERS (trailers) following
     * (optional) DATA frames, but in following HTTP/1.1, we limit to single
     * (optional) HEADERS (+ CONTINUATIONs) after (optional) DATA frame(s)
     * and require that the HEADERS frame set END_STREAM flag. */
    if (!(flags & H2_FLAG_END_STREAM)) {
        h2_send_rst_stream(r, con, H2_E_PROTOCOL_ERROR);
        return NULL;
    }

    return h2_recv_end_data(r, con, 0) ? r : NULL;
}


static void
h2_parse_headers_frame (request_st * const restrict r, const unsigned char *psrc, const uint32_t plen, const int trailers)
{
    h2con * const h2c = r->con->h2;
    struct lshpack_dec * const restrict decoder = &h2c->decoder;
    const unsigned char * const endp = psrc + plen;
    http_header_parse_ctx hpctx;
    hpctx.hlen     = 0;
    hpctx.pseudo   = 1;
    hpctx.scheme   = 0;
    hpctx.trailers = trailers;
    hpctx.max_request_field_size = r->conf.max_request_field_size;
    hpctx.http_parseopts = r->conf.http_parseopts;
    const int log_request_header = r->conf.log_request_header;
    int rc = LSHPACK_OK;
    /*buffer_clear(&r->target);*//*(initial state)*/

    /*(h2_init_con() resized h2r->tmp_buf to 64k; shared with r->tmp_buf)*/
    buffer * const tb = r->tmp_buf;
    force_assert(tb->size >= 65536);/*(sanity check; remove in future)*/
    const lsxpack_strlen_t tbsz = (tb->size <= LSXPACK_MAX_STRLEN)
      ? tb->size
      : LSXPACK_MAX_STRLEN;

    /* note: #define LSHPACK_DEC_HTTP1X_OUTPUT 1 (default) configures
     * decoder to produce output in format: "field-name: value\r\n"
     * future: modify build system to define value to 0 in lshpack.h
     * against which lighttpd builds (or define value in build systems)
     * Then adjust code below to not use the HTTP/1.x compatibility,
     * as it is less efficient to copy into HTTP/1.1 request and reparse
     * than it is to directly parse each decoded header line. */
    lsxpack_header_t lsx;
    while (psrc < endp) {
        memset(&lsx, 0, sizeof(lsxpack_header_t));
        lsx.buf = tb->ptr;
        lsx.val_len = tbsz;
        rc = lshpack_dec_decode(decoder, &psrc, endp, &lsx);
        if (0 == lsx.name_len)
            rc = LSHPACK_ERR_BAD_DATA;
        if (rc == LSHPACK_OK) {
            hpctx.k = lsx.buf+lsx.name_offset;
            hpctx.v = lsx.buf+lsx.val_offset;
            hpctx.klen = lsx.name_len;
            hpctx.vlen = lsx.val_len;

            if (log_request_header)
                log_error(r->conf.errh, __FILE__, __LINE__,
                  "fd:%d id:%u rqst: %.*s: %.*s", r->con->fd, r->h2id,
                  (int)hpctx.klen, hpctx.k, (int)hpctx.vlen, hpctx.v);

            r->http_status = http_request_parse_header(r, &hpctx);
            if (0 != r->http_status)
                break;
        }
      #if 0 /*(see catch-all below)*/
        /* Send GOAWAY (further below) (decoder state not maintained on error)
         * (see comments above why decoder state must be maintained) */
        /* XXX: future: could try to send :status 431 here
         * and reset other active streams in H2_STATE_OPEN */
        else if (rc == LSHPACK_ERR_MORE_BUF) {
            /* XXX: TODO if (r->conf.log_request_header_on_error) */
            r->http_status = 431; /* Request Header Fields Too Large */
            /*(try to avoid reading/buffering more data for this request)*/
            r->h2_rwin = 0; /*(out-of-sync with peer, but is error case)*/
            /*r->h2state = H2_STATE_HALF_CLOSED_REMOTE*/
            /* psrc was not advanced if LSHPACK_ERR_MORE_BUF;
             * processing must stop (since not retrying w/ larger buf)*/
            break;
        }
      #endif
        else { /* LSHPACK_ERR_BAD_DATA */
            /* GOAWAY with H2_E_PROTOCOL_ERROR is not specific enough
             * to tell peer to not retry request, so send RST_STREAM
             * (slightly more specific, but not by much) before GOAWAY*/
            /* LSHPACK_ERR_MORE_BUF is treated as an attack, send GOAWAY
             * (h2r->tmp_buf was resized to 64k in h2_init_con()) */
            request_h2error_t err = H2_E_COMPRESSION_ERROR;
            if (rc != LSHPACK_ERR_BAD_DATA) {
                /* LSHPACK_ERR_TOO_LARGE, LSHPACK_ERR_MORE_BUF */
                err = H2_E_PROTOCOL_ERROR;
                h2_send_rst_stream(r, r->con, err);
            }
            if (!h2c->sent_goaway && !trailers)
                h2c->h2_cid = r->h2id;
            h2_send_goaway_e(r->con, err);
            if (!trailers) {
                h2_retire_stream(r, r->con);
                return;
            }
            else {
                r->state = CON_STATE_ERROR;
                r->h2state = H2_STATE_CLOSED;
            }
            break;
        }
    }