h2_bucket_beam.c

/* Copyright 2015 greenbytes GmbH (https://www.greenbytes.de)
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include <apr_lib.h>
#include <apr_atomic.h>
#include <apr_strings.h>
#include <apr_time.h>
#include <apr_buckets.h>
#include <apr_thread_mutex.h>
#include <apr_thread_cond.h>

#include <httpd.h>
#include <http_protocol.h>
#include <http_log.h>

#include "h2_private.h"
#include "h2_util.h"
#include "h2_bucket_beam.h"

static void h2_beam_emitted(h2_bucket_beam *beam, h2_beam_proxy *proxy);

#define H2_BPROXY_NEXT(e)             APR_RING_NEXT((e), link)
#define H2_BPROXY_PREV(e)             APR_RING_PREV((e), link)
#define H2_BPROXY_REMOVE(e)           APR_RING_REMOVE((e), link)

#define H2_BPROXY_LIST_INIT(b)        APR_RING_INIT(&(b)->list, h2_beam_proxy, link);
#define H2_BPROXY_LIST_SENTINEL(b)    APR_RING_SENTINEL(&(b)->list, h2_beam_proxy, link)
#define H2_BPROXY_LIST_EMPTY(b)       APR_RING_EMPTY(&(b)->list, h2_beam_proxy, link)
#define H2_BPROXY_LIST_FIRST(b)       APR_RING_FIRST(&(b)->list)
#define H2_BPROXY_LIST_LAST(b)	      APR_RING_LAST(&(b)->list)
#define H2_PROXY_BLIST_INSERT_HEAD(b, e) do {				\
	h2_beam_proxy *ap__b = (e);                                        \
	APR_RING_INSERT_HEAD(&(b)->list, ap__b, h2_beam_proxy, link);	\
    } while (0)
#define H2_BPROXY_LIST_INSERT_TAIL(b, e) do {				\
	h2_beam_proxy *ap__b = (e);					\
	APR_RING_INSERT_TAIL(&(b)->list, ap__b, h2_beam_proxy, link);	\
    } while (0)
#define H2_BPROXY_LIST_CONCAT(a, b) do {					\
        APR_RING_CONCAT(&(a)->list, &(b)->list, h2_beam_proxy, link);	\
    } while (0)
#define H2_BPROXY_LIST_PREPEND(a, b) do {					\
        APR_RING_PREPEND(&(a)->list, &(b)->list, h2_beam_proxy, link);	\
    } while (0)


/*******************************************************************************
 * beam bucket with reference to beam and bucket it represents
 ******************************************************************************/

const apr_bucket_type_t h2_bucket_type_beam;

#define H2_BUCKET_IS_BEAM(e)     (e->type == &h2_bucket_type_beam)

struct h2_beam_proxy {
    apr_bucket_refcount refcount;
    APR_RING_ENTRY(h2_beam_proxy) link;
    h2_bucket_beam *beam;
    apr_bucket *bsender;
    apr_size_t n;
};

static const char Dummy = '\0';

static apr_status_t beam_bucket_read(apr_bucket *b, const char **str, 
                                     apr_size_t *len, apr_read_type_e block)
{
    h2_beam_proxy *d = b->data;
    if (d->bsender) {
        const char *data;
        apr_status_t status = apr_bucket_read(d->bsender, &data, len, block);
        if (status == APR_SUCCESS) {
            *str = data + b->start;
            *len = b->length;
        }
        return status;
    }
    *str = &Dummy;
    *len = 0;
    return APR_ECONNRESET;
}

static void beam_bucket_destroy(void *data)
{
    h2_beam_proxy *d = data;

    if (apr_bucket_shared_destroy(d)) {
        /* When the beam gets destroyed before this bucket, it will
         * NULLify its reference here. This is not protected by a mutex,
         * so it will not help with race conditions.
         * But it lets us shut down memory pool with circulare beam
         * references. */
        if (d->beam) {
            h2_beam_emitted(d->beam, d);
        }
        apr_bucket_free(d);
    }
}

static apr_bucket * h2_beam_bucket_make(apr_bucket *b, 
                                        h2_bucket_beam *beam,
                                        apr_bucket *bsender, apr_size_t n)
{
    h2_beam_proxy *d;

    d = apr_bucket_alloc(sizeof(*d), b->list);
    H2_BPROXY_LIST_INSERT_TAIL(&beam->proxies, d);
    d->beam = beam;
    d->bsender = bsender;
    d->n = n;
    
    b = apr_bucket_shared_make(b, d, 0, bsender? bsender->length : 0);
    b->type = &h2_bucket_type_beam;

    return b;
}

static apr_bucket *h2_beam_bucket_create(h2_bucket_beam *beam,
                                         apr_bucket *bsender,
                                         apr_bucket_alloc_t *list,
                                         apr_size_t n)
{
    apr_bucket *b = apr_bucket_alloc(sizeof(*b), list);

    APR_BUCKET_INIT(b);
    b->free = apr_bucket_free;
    b->list = list;
    return h2_beam_bucket_make(b, beam, bsender, n);
}

const apr_bucket_type_t h2_bucket_type_beam = {
    "BEAM", 5, APR_BUCKET_DATA,
    beam_bucket_destroy,
    beam_bucket_read,
    apr_bucket_setaside_noop,
    apr_bucket_shared_split,
    apr_bucket_shared_copy
};

/*******************************************************************************
 * h2_blist, a brigade without allocations
 ******************************************************************************/

static apr_array_header_t *beamers;

static apr_status_t cleanup_beamers(void *dummy)
{
    (void)dummy;
    beamers = NULL;
    return APR_SUCCESS;
}

void h2_register_bucket_beamer(h2_bucket_beamer *beamer)
{
    if (!beamers) {
        apr_pool_cleanup_register(apr_hook_global_pool, NULL,
                                  cleanup_beamers, apr_pool_cleanup_null);
        beamers = apr_array_make(apr_hook_global_pool, 10, 
                                 sizeof(h2_bucket_beamer*));
    }
    APR_ARRAY_PUSH(beamers, h2_bucket_beamer*) = beamer;
}

static apr_bucket *h2_beam_bucket(h2_bucket_beam *beam, 
                                  apr_bucket_brigade *dest,
                                  const apr_bucket *src)
{
    apr_bucket *b = NULL;
    int i;
    if (beamers) {
        for (i = 0; i < beamers->nelts && b == NULL; ++i) {
            h2_bucket_beamer *beamer;
            
            beamer = APR_ARRAY_IDX(beamers, i, h2_bucket_beamer*);
            b = beamer(beam, dest, src);
        }
    }
    return b;
}


/*******************************************************************************
 * bucket beam that can transport buckets across threads
 ******************************************************************************/

static void mutex_leave(void *ctx, apr_thread_mutex_t *lock)
{
    apr_thread_mutex_unlock(lock);
}

static apr_status_t mutex_enter(void *ctx, h2_beam_lock *pbl)
{
    h2_bucket_beam *beam = ctx;
    pbl->mutex = beam->lock;
    pbl->leave = mutex_leave;
    return apr_thread_mutex_lock(pbl->mutex);
}

static apr_status_t enter_yellow(h2_bucket_beam *beam, h2_beam_lock *pbl)
{
    return mutex_enter(beam, pbl);
}

static void leave_yellow(h2_bucket_beam *beam, h2_beam_lock *pbl)
{
    if (pbl->leave) {
        pbl->leave(pbl->leave_ctx, pbl->mutex);
    }
}

static apr_off_t bucket_mem_used(apr_bucket *b)
{
    if (APR_BUCKET_IS_FILE(b)) {
        return 0;
    }
    else {
        /* should all have determinate length */
        return b->length;
    }
}

static int report_consumption(h2_bucket_beam *beam, h2_beam_lock *pbl)
{
    int rv = 0;
    apr_off_t len = beam->received_bytes - beam->cons_bytes_reported;
    h2_beam_io_callback *cb = beam->cons_io_cb;
     
    if (len > 0) {
        if (cb) {
            void *ctx = beam->cons_ctx;
            
            if (pbl) leave_yellow(beam, pbl);
            cb(ctx, beam, len);
            if (pbl) enter_yellow(beam, pbl);
            rv = 1;
        }
        beam->cons_bytes_reported += len;
    }
    return rv;
}

static void report_prod_io(h2_bucket_beam *beam, int force, h2_beam_lock *pbl)
{
    apr_off_t len = beam->sent_bytes - beam->prod_bytes_reported;
    if (force || len > 0) {
        h2_beam_io_callback *cb = beam->prod_io_cb; 
        if (cb) {
            void *ctx = beam->prod_ctx;
            
            leave_yellow(beam, pbl);
            cb(ctx, beam, len);
            enter_yellow(beam, pbl);
        }
        beam->prod_bytes_reported += len;
    }
}

static apr_size_t calc_buffered(h2_bucket_beam *beam)
{
    apr_size_t len = 0;
    apr_bucket *b;
    for (b = H2_BLIST_FIRST(&beam->send_list); 
         b != H2_BLIST_SENTINEL(&beam->send_list);
         b = APR_BUCKET_NEXT(b)) {
        if (b->length == ((apr_size_t)-1)) {
            /* do not count */
        }
        else if (APR_BUCKET_IS_FILE(b)) {
            /* if unread, has no real mem footprint. */
        }
        else {
            len += b->length;
        }
    }
    return len;
}

static void r_purge_sent(h2_bucket_beam *beam)
{
    apr_bucket *b;
    /* delete all sender buckets in purge brigade, needs to be called
     * from sender thread only */
    while (!H2_BLIST_EMPTY(&beam->purge_list)) {
        b = H2_BLIST_FIRST(&beam->purge_list);
        apr_bucket_delete(b);
    }
}

static apr_size_t calc_space_left(h2_bucket_beam *beam)
{
    if (beam->max_buf_size > 0) {
        apr_off_t len = calc_buffered(beam);
        return (beam->max_buf_size > len? (beam->max_buf_size - len) : 0);
    }
    return APR_SIZE_MAX;
}

static int buffer_is_empty(h2_bucket_beam *beam)
{
    return ((!beam->recv_buffer || APR_BRIGADE_EMPTY(beam->recv_buffer))
            && H2_BLIST_EMPTY(&beam->send_list));
}

static apr_status_t wait_empty(h2_bucket_beam *beam, apr_read_type_e block,  
                               apr_thread_mutex_t *lock)
{
    apr_status_t rv = APR_SUCCESS;
    
    while (!buffer_is_empty(beam) && APR_SUCCESS == rv) {
        if (APR_BLOCK_READ != block || !lock) {
            rv = APR_EAGAIN;
        }
        else if (beam->timeout > 0) {
            rv = apr_thread_cond_timedwait(beam->change, lock, beam->timeout);
        }
        else {
            rv = apr_thread_cond_wait(beam->change, lock);
        }
    }
    return rv;
}

static apr_status_t wait_not_empty(h2_bucket_beam *beam, apr_read_type_e block,  
                                   apr_thread_mutex_t *lock)
{
    apr_status_t rv = APR_SUCCESS;
    
    while (buffer_is_empty(beam) && APR_SUCCESS == rv) {
        if (beam->aborted) {
            rv = APR_ECONNABORTED;
        }
        else if (beam->closed) {
            rv = APR_EOF;
        }
        else if (APR_BLOCK_READ != block || !lock) {
            rv = APR_EAGAIN;
        }
        else if (beam->timeout > 0) {
            rv = apr_thread_cond_timedwait(beam->change, lock, beam->timeout);
        }
        else {
            rv = apr_thread_cond_wait(beam->change, lock);
        }
    }
    return rv;
}

static apr_status_t wait_not_full(h2_bucket_beam *beam, apr_read_type_e block, 
                                  apr_size_t *pspace_left, h2_beam_lock *bl)
{
    apr_status_t rv = APR_SUCCESS;
    apr_size_t left;
    
    while (0 == (left = calc_space_left(beam)) && APR_SUCCESS == rv) {
        if (beam->aborted) {
            rv = APR_ECONNABORTED;
        }
        else if (block != APR_BLOCK_READ || !bl->mutex) {
            rv = APR_EAGAIN;
        }
        else {
            if (beam->timeout > 0) {
                rv = apr_thread_cond_timedwait(beam->change, bl->mutex, beam->timeout);
            }
            else {
                rv = apr_thread_cond_wait(beam->change, bl->mutex);
            }
        }
    }
    *pspace_left = left;
    return rv;
}

static void h2_beam_emitted(h2_bucket_beam *beam, h2_beam_proxy *proxy)
{
    h2_beam_lock bl;
    apr_bucket *b, *next;

    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        /* even when beam buckets are split, only the one where
         * refcount drops to 0 will call us */
        H2_BPROXY_REMOVE(proxy);
        /* invoked from receiver thread, the last beam bucket for the send
         * bucket is about to be destroyed.
         * remove it from the hold, where it should be now */
        if (proxy->bsender) {
            for (b = H2_BLIST_FIRST(&beam->hold_list); 
                 b != H2_BLIST_SENTINEL(&beam->hold_list);
                 b = APR_BUCKET_NEXT(b)) {
                 if (b == proxy->bsender) {
                    break;
                 }
            }
            if (b != H2_BLIST_SENTINEL(&beam->hold_list)) {
                /* bucket is in hold as it should be, mark this one
                 * and all before it for purging. We might have placed meta
                 * buckets without a receiver proxy into the hold before it 
                 * and schedule them for purging now */
                for (b = H2_BLIST_FIRST(&beam->hold_list); 
                     b != H2_BLIST_SENTINEL(&beam->hold_list);
                     b = next) {
                    next = APR_BUCKET_NEXT(b);
                    if (b == proxy->bsender) {
                        APR_BUCKET_REMOVE(b);
                        H2_BLIST_INSERT_TAIL(&beam->purge_list, b);
                        break;
                    }
                    else if (APR_BUCKET_IS_METADATA(b)) {
                        APR_BUCKET_REMOVE(b);
                        H2_BLIST_INSERT_TAIL(&beam->purge_list, b);
                    }
                    else {
                        /* another data bucket before this one in hold. this
                         * is normal since DATA buckets need not be destroyed
                         * in order */
                    }
                }
                
                proxy->bsender = NULL;
            }
            else {
                /* it should be there unless we screwed up */
                ap_log_perror(APLOG_MARK, APLOG_WARNING, 0, beam->send_pool, 
                              APLOGNO(03384) "h2_beam(%d-%s): emitted bucket not "
                              "in hold, n=%d", beam->id, beam->tag, 
                              (int)proxy->n);
                ap_assert(!proxy->bsender);
            }
        }
        /* notify anyone waiting on space to become available */
        if (!bl.mutex) {
            r_purge_sent(beam);
        }
        else {
            apr_thread_cond_broadcast(beam->change);
        }
        leave_yellow(beam, &bl);
    }
}

static void h2_blist_cleanup(h2_blist *bl)
{
    apr_bucket *e;

    while (!H2_BLIST_EMPTY(bl)) {
        e = H2_BLIST_FIRST(bl);
        apr_bucket_delete(e);
    }
}

static apr_status_t beam_close(h2_bucket_beam *beam)
{
    if (!beam->closed) {
        beam->closed = 1;
        apr_thread_cond_broadcast(beam->change);
    }
    return APR_SUCCESS;
}

int h2_beam_is_closed(h2_bucket_beam *beam)
{
    return beam->closed;
}

static int pool_register(h2_bucket_beam *beam, apr_pool_t *pool, 
                         apr_status_t (*cleanup)(void *))
{
    if (pool && pool != beam->pool) {
        apr_pool_pre_cleanup_register(pool, beam, cleanup);
        return 1;
    }
    return 0;
}

static int pool_kill(h2_bucket_beam *beam, apr_pool_t *pool,
                     apr_status_t (*cleanup)(void *)) {
    if (pool && pool != beam->pool) {
        apr_pool_cleanup_kill(pool, beam, cleanup);
        return 1;
    }
    return 0;
}

static apr_status_t beam_recv_cleanup(void *data)
{
    h2_bucket_beam *beam = data;
    /* receiver pool has gone away, clear references */
    beam->recv_buffer = NULL;
    beam->recv_pool = NULL;
    return APR_SUCCESS;
}

static apr_status_t beam_send_cleanup(void *data)
{
    h2_bucket_beam *beam = data;
    /* sender is going away, clear up all references to its memory */
    r_purge_sent(beam);
    h2_blist_cleanup(&beam->send_list);
    report_consumption(beam, NULL);
    while (!H2_BPROXY_LIST_EMPTY(&beam->proxies)) {
        h2_beam_proxy *proxy = H2_BPROXY_LIST_FIRST(&beam->proxies);
        H2_BPROXY_REMOVE(proxy);
        proxy->beam = NULL;
        proxy->bsender = NULL;
    }
    h2_blist_cleanup(&beam->purge_list);
    h2_blist_cleanup(&beam->hold_list);
    beam->send_pool = NULL;
    return APR_SUCCESS;
}

static void beam_set_send_pool(h2_bucket_beam *beam, apr_pool_t *pool) 
{
    if (beam->send_pool != pool) {
        if (beam->send_pool && beam->send_pool != beam->pool) {
            pool_kill(beam, beam->send_pool, beam_send_cleanup);
            beam_send_cleanup(beam);
        }
        beam->send_pool = pool;
        pool_register(beam, beam->send_pool, beam_send_cleanup);
    }
}

static void recv_buffer_cleanup(h2_bucket_beam *beam, h2_beam_lock *bl)
{
    if (beam->recv_buffer && !APR_BRIGADE_EMPTY(beam->recv_buffer)) {
        apr_bucket_brigade *bb = beam->recv_buffer;
        apr_off_t bblen = 0;
        
        beam->recv_buffer = NULL;
        apr_brigade_length(bb, 0, &bblen);
        beam->received_bytes += bblen;
        
        /* need to do this unlocked since bucket destroy might 
         * call this beam again. */
        if (bl) leave_yellow(beam, bl);
        apr_brigade_destroy(bb);
        if (bl) enter_yellow(beam, bl);
        
        if (beam->cons_ev_cb) { 
            beam->cons_ev_cb(beam->cons_ctx, beam);
        }
    }
}

static apr_status_t beam_cleanup(void *data)
{
    h2_bucket_beam *beam = data;
    apr_status_t status = APR_SUCCESS;
    int safe_send = (beam->owner == H2_BEAM_OWNER_SEND);
    int safe_recv = (beam->owner == H2_BEAM_OWNER_RECV);
    
    /* 
     * Owner of the beam is going away, depending on which side it owns,
     * cleanup strategies will differ.
     *
     * In general, receiver holds references to memory from sender. 
     * Clean up receiver first, if safe, then cleanup sender, if safe.
     */
     
    /* When modify send is not safe, this means we still have multi-thread
     * protection and the owner is receiving the buckets. If the sending
     * side has not gone away, this means we could have dangling buckets
     * in our lists that never get destroyed. This should not happen. */
    ap_assert(safe_send || !beam->send_pool);
    if (!H2_BLIST_EMPTY(&beam->send_list)) {
        ap_assert(beam->send_pool);
    }
    
    if (safe_recv) {
        if (beam->recv_pool) {
            pool_kill(beam, beam->recv_pool, beam_recv_cleanup);
            beam->recv_pool = NULL;
        }
        recv_buffer_cleanup(beam, NULL);
    }
    else {
        beam->recv_buffer = NULL;
        beam->recv_pool = NULL;
    }
    
    if (safe_send && beam->send_pool) {
        pool_kill(beam, beam->send_pool, beam_send_cleanup);
        status = beam_send_cleanup(beam);
    }
    
    if (safe_recv) {
        ap_assert(H2_BPROXY_LIST_EMPTY(&beam->proxies));
        ap_assert(H2_BLIST_EMPTY(&beam->send_list));
        ap_assert(H2_BLIST_EMPTY(&beam->hold_list));
        ap_assert(H2_BLIST_EMPTY(&beam->purge_list));
    }
    return status;
}

apr_status_t h2_beam_destroy(h2_bucket_beam *beam)
{
    apr_pool_cleanup_kill(beam->pool, beam, beam_cleanup);
    return beam_cleanup(beam);
}

apr_status_t h2_beam_create(h2_bucket_beam **pbeam, apr_pool_t *pool, 
                            int id, const char *tag, 
                            h2_beam_owner_t owner,
                            apr_size_t max_buf_size,
                            apr_interval_time_t timeout)
{
    h2_bucket_beam *beam;
    apr_status_t rv = APR_SUCCESS;
    
    beam = apr_pcalloc(pool, sizeof(*beam));
    if (!beam) {
        return APR_ENOMEM;
    }

    beam->id = id;
    beam->tag = tag;
    beam->pool = pool;
    beam->owner = owner;
    H2_BLIST_INIT(&beam->send_list);
    H2_BLIST_INIT(&beam->hold_list);
    H2_BLIST_INIT(&beam->purge_list);
    H2_BPROXY_LIST_INIT(&beam->proxies);
    beam->tx_mem_limits = 1;
    beam->max_buf_size = max_buf_size;
    beam->timeout = timeout;

    rv = apr_thread_mutex_create(&beam->lock, APR_THREAD_MUTEX_DEFAULT, pool);
    if (APR_SUCCESS == rv) {
        rv = apr_thread_cond_create(&beam->change, pool);
        if (APR_SUCCESS == rv) {
            apr_pool_pre_cleanup_register(pool, beam, beam_cleanup);
            *pbeam = beam;
        }
    }
    return rv;
}

void h2_beam_buffer_size_set(h2_bucket_beam *beam, apr_size_t buffer_size)
{
    h2_beam_lock bl;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        beam->max_buf_size = buffer_size;
        leave_yellow(beam, &bl);
    }
}

apr_size_t h2_beam_buffer_size_get(h2_bucket_beam *beam)
{
    h2_beam_lock bl;
    apr_size_t buffer_size = 0;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        buffer_size = beam->max_buf_size;
        leave_yellow(beam, &bl);
    }
    return buffer_size;
}

void h2_beam_timeout_set(h2_bucket_beam *beam, apr_interval_time_t timeout)
{
    h2_beam_lock bl;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        beam->timeout = timeout;
        leave_yellow(beam, &bl);
    }
}

apr_interval_time_t h2_beam_timeout_get(h2_bucket_beam *beam)
{
    h2_beam_lock bl;
    apr_interval_time_t timeout = 0;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        timeout = beam->timeout;
        leave_yellow(beam, &bl);
    }
    return timeout;
}

void h2_beam_abort(h2_bucket_beam *beam)
{
    h2_beam_lock bl;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        if (!beam->aborted) {
            beam->aborted = 1;
            r_purge_sent(beam);
            h2_blist_cleanup(&beam->send_list);
            report_consumption(beam, &bl);
        }
        apr_thread_cond_broadcast(beam->change);
        leave_yellow(beam, &bl);
    }
}

apr_status_t h2_beam_close(h2_bucket_beam *beam)
{
    h2_beam_lock bl;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        r_purge_sent(beam);
        beam_close(beam);
        report_consumption(beam, &bl);
        leave_yellow(beam, &bl);
    }
    return beam->aborted? APR_ECONNABORTED : APR_SUCCESS;
}

apr_status_t h2_beam_leave(h2_bucket_beam *beam)
{
    h2_beam_lock bl;
    
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        recv_buffer_cleanup(beam, &bl);
        beam->aborted = 1;
        beam_close(beam);
        leave_yellow(beam, &bl);
    }
    return APR_SUCCESS;
}

apr_status_t h2_beam_wait_empty(h2_bucket_beam *beam, apr_read_type_e block)
{
    apr_status_t status;
    h2_beam_lock bl;
    
    if ((status = enter_yellow(beam, &bl)) == APR_SUCCESS) {
        status = wait_empty(beam, block, bl.mutex);
        leave_yellow(beam, &bl);
    }
    return status;
}

static void move_to_hold(h2_bucket_beam *beam, 
                         apr_bucket_brigade *sender_bb)
{
    apr_bucket *b;
    while (sender_bb && !APR_BRIGADE_EMPTY(sender_bb)) {
        b = APR_BRIGADE_FIRST(sender_bb);
        APR_BUCKET_REMOVE(b);
        H2_BLIST_INSERT_TAIL(&beam->send_list, b);
    }
}

static apr_status_t append_bucket(h2_bucket_beam *beam, 
                                  apr_bucket *b,
                                  apr_read_type_e block,
                                  apr_size_t *pspace_left,
                                  h2_beam_lock *pbl)
{
    const char *data;
    apr_size_t len;
    apr_status_t status;
    int can_beam, check_len;
    
    if (beam->aborted) {
        return APR_ECONNABORTED;
    }
    
    if (APR_BUCKET_IS_METADATA(b)) {
        if (APR_BUCKET_IS_EOS(b)) {
            beam->closed = 1;
        }
        APR_BUCKET_REMOVE(b);
        H2_BLIST_INSERT_TAIL(&beam->send_list, b);
        return APR_SUCCESS;
    }
    else if (APR_BUCKET_IS_FILE(b)) {
        /* For file buckets the problem is their internal readpool that
         * is used on the first read to allocate buffer/mmap.
         * Since setting aside a file bucket will de-register the
         * file cleanup function from the previous pool, we need to
         * call that only from the sender thread.
         *
         * Currently, we do not handle file bucket with refcount > 1 as
         * the beam is then not in complete control of the file's lifetime.
         * Which results in the bug that a file get closed by the receiver
         * while the sender or the beam still have buckets using it. 
         * 
         * Additionally, we allow callbacks to prevent beaming file
         * handles across. The use case for this is to limit the number 
         * of open file handles and rather use a less efficient beam
         * transport. */
        apr_bucket_file *bf = b->data;
        apr_file_t *fd = bf->fd;
        can_beam = (bf->refcount.refcount == 1);
        if (can_beam && beam->can_beam_fn) {
            can_beam = beam->can_beam_fn(beam->can_beam_ctx, beam, fd);
        }
        check_len = !can_beam;
    }
    else {
        if (b->length == ((apr_size_t)-1)) {
            const char *data;
            status = apr_bucket_read(b, &data, &len, APR_BLOCK_READ);
            if (status != APR_SUCCESS) {
                return status;
            }
        }
        check_len = 1;
    }
    
    if (check_len) {
        if (b->length > *pspace_left) {
            apr_bucket_split(b, *pspace_left);
        }
        *pspace_left -= b->length;
    }

    /* The fundamental problem is that reading a sender bucket from
     * a receiver thread is a total NO GO, because the bucket might use
     * its pool/bucket_alloc from a foreign thread and that will
     * corrupt. */
    status = APR_ENOTIMPL;
    if (APR_BUCKET_IS_TRANSIENT(b)) {
        /* this takes care of transient buckets and converts them
         * into heap ones. Other bucket types might or might not be
         * affected by this. */
        status = apr_bucket_setaside(b, beam->send_pool);
    }
    else if (APR_BUCKET_IS_HEAP(b)) {
        /* For heap buckets read from a receiver thread is fine. The
         * data will be there and live until the bucket itself is
         * destroyed. */
        status = APR_SUCCESS;
    }
    else if (APR_BUCKET_IS_POOL(b)) {
        /* pool buckets are bastards that register at pool cleanup
         * to morph themselves into heap buckets. That may happen anytime,
         * even after the bucket data pointer has been read. So at
         * any time inside the receiver thread, the pool bucket memory
         * may disappear. yikes. */
        status = apr_bucket_read(b, &data, &len, APR_BLOCK_READ);
        if (status == APR_SUCCESS) {
            apr_bucket_heap_make(b, data, len, NULL);
        }
    }
    else if (APR_BUCKET_IS_FILE(b) && can_beam) {
        status = apr_bucket_setaside(b, beam->send_pool);
    }
    
    if (status == APR_ENOTIMPL) {
        /* we have no knowledge about the internals of this bucket,
         * but hope that after read, its data stays immutable for the
         * lifetime of the bucket. (see pool bucket handling above for
         * a counter example).
         * We do the read while in the sender thread, so that the bucket may
         * use pools/allocators safely. */
        status = apr_bucket_read(b, &data, &len, APR_BLOCK_READ);
        if (status == APR_SUCCESS) {
            status = apr_bucket_setaside(b, beam->send_pool);
        }
    }
    
    if (status != APR_SUCCESS && status != APR_ENOTIMPL) {
        return status;
    }
    
    APR_BUCKET_REMOVE(b);
    H2_BLIST_INSERT_TAIL(&beam->send_list, b);
    beam->sent_bytes += b->length;

    return APR_SUCCESS;
}

void h2_beam_send_from(h2_bucket_beam *beam, apr_pool_t *p)
{
    h2_beam_lock bl;
    /* Called from the sender thread to add buckets to the beam */
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        r_purge_sent(beam);
        beam_set_send_pool(beam, p);
        leave_yellow(beam, &bl);
    }
}

apr_status_t h2_beam_send(h2_bucket_beam *beam, 
                          apr_bucket_brigade *sender_bb, 
                          apr_read_type_e block)
{
    apr_bucket *b;
    apr_status_t rv = APR_SUCCESS;
    apr_size_t space_left = 0;
    h2_beam_lock bl;

    /* Called from the sender thread to add buckets to the beam */
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        ap_assert(beam->send_pool);
        r_purge_sent(beam);
        
        if (beam->aborted) {
            move_to_hold(beam, sender_bb);
            rv = APR_ECONNABORTED;
        }
        else if (sender_bb) {
            int force_report = !APR_BRIGADE_EMPTY(sender_bb);
            
            space_left = calc_space_left(beam);
            while (!APR_BRIGADE_EMPTY(sender_bb) && APR_SUCCESS == rv) {
                if (space_left <= 0) {
                    report_prod_io(beam, force_report, &bl);
                    rv = wait_not_full(beam, block, &space_left, &bl);
                    if (APR_SUCCESS != rv) {
                        break;
                    }
                }
                b = APR_BRIGADE_FIRST(sender_bb);
                rv = append_bucket(beam, b, block, &space_left, &bl);
            }
            
            report_prod_io(beam, force_report, &bl);
            apr_thread_cond_broadcast(beam->change);
        }
        report_consumption(beam, &bl);
        leave_yellow(beam, &bl);
    }
    return rv;
}

apr_status_t h2_beam_receive(h2_bucket_beam *beam, 
                             apr_bucket_brigade *bb, 
                             apr_read_type_e block,
                             apr_off_t readbytes)
{
    h2_beam_lock bl;
    apr_bucket *bsender, *brecv, *ng;
    int transferred = 0;
    apr_status_t status = APR_SUCCESS;
    apr_off_t remain;
    int transferred_buckets = 0;
    
    /* Called from the receiver thread to take buckets from the beam */
    if (enter_yellow(beam, &bl) == APR_SUCCESS) {
        if (readbytes <= 0) {
            readbytes = APR_SIZE_MAX;
        }
        remain = readbytes;
        
transfer:
        if (beam->aborted) {
            recv_buffer_cleanup(beam, &bl);
            status = APR_ECONNABORTED;
            goto leave;
        }

        /* transfer enough buckets from our receiver brigade, if we have one */
        while (remain >= 0 
               && beam->recv_buffer 
               && !APR_BRIGADE_EMPTY(beam->recv_buffer)) {
               
            brecv = APR_BRIGADE_FIRST(beam->recv_buffer);
            if (brecv->length > 0 && remain <= 0) {
                break;
            }            
            APR_BUCKET_REMOVE(brecv);
            APR_BRIGADE_INSERT_TAIL(bb, brecv);
            remain -= brecv->length;
            ++transferred;
        }

        /* transfer from our sender brigade, transforming sender buckets to
         * receiver ones until we have enough */
        while (remain >= 0 && !H2_BLIST_EMPTY(&beam->send_list)) {
               
            brecv = NULL;
            bsender = H2_BLIST_FIRST(&beam->send_list);            
            if (bsender->length > 0 && remain <= 0) {
                break;
            }
                        
            if (APR_BUCKET_IS_METADATA(bsender)) {
                if (APR_BUCKET_IS_EOS(bsender)) {
                    brecv = apr_bucket_eos_create(bb->bucket_alloc);
                    beam->close_sent = 1;
                }
                else if (APR_BUCKET_IS_FLUSH(bsender)) {
                    brecv = apr_bucket_flush_create(bb->bucket_alloc);
                }
                else if (AP_BUCKET_IS_ERROR(bsender)) {
                    ap_bucket_error *eb = (ap_bucket_error *)bsender;
                    brecv = ap_bucket_error_create(eb->status, eb->data,
                                                    bb->p, bb->bucket_alloc);
                }
            }
            else if (bsender->length == 0) {