- Fix CVE-2026-80225, Possible degradation of service from continuous

queries on the same TCP/DoT connection. Thanks to Qifan Zhang from
  Palo Alto Networks for the report.
This commit is contained in:
W.C.A. Wijngaards
2026-09-16 09:16:29 +02:00
parent eb86d9bc50
commit e619ead2db
6 changed files with 104 additions and 3 deletions
+11 -1
View File
@@ -2133,7 +2133,7 @@ void listen_start_accept(struct listen_dnsport* listen)
}
struct tcp_req_info*
tcp_req_info_create(struct sldns_buffer* spoolbuf)
tcp_req_info_create(struct comm_base* base, struct sldns_buffer* spoolbuf)
{
struct tcp_req_info* req = (struct tcp_req_info*)malloc(sizeof(*req));
if(!req) {
@@ -2141,6 +2141,12 @@ tcp_req_info_create(struct sldns_buffer* spoolbuf)
return NULL;
}
memset(req, 0, sizeof(*req));
req->read_again_timer = comm_timer_create(base, tcp_read_again_cb, req);
if(!req->read_again_timer) {
log_err("malloc failure");
free(req);
return NULL;
}
req->spool_buffer = spoolbuf;
return req;
}
@@ -2150,6 +2156,7 @@ tcp_req_info_delete(struct tcp_req_info* req)
{
if(!req) return;
tcp_req_info_clear(req);
comm_timer_delete(req->read_again_timer);
/* cp is pointer back to commpoint that owns this struct and
* called delete on us */
/* spool_buffer is shared udp buffer, not deleted here */
@@ -2189,6 +2196,9 @@ void tcp_req_info_clear(struct tcp_req_info* req)
req->done_req_list = NULL;
req->num_done_req = 0;
req->read_is_closed = 0;
if(comm_timer_is_set(req->read_again_timer))
comm_timer_disable(req->read_again_timer);
}
void
+7 -1
View File
@@ -347,6 +347,10 @@ struct tcp_req_info {
int num_done_req;
/** list of pending writable result packets, malloced one at a time */
struct tcp_req_done_item* done_req_list;
/** the read again timer, when the number of pipelined TCP queries
* is large, it waits, zero time, for a new event loop to service
* the remainder of the TCP traffic on the fd. */
struct comm_timer* read_again_timer;
};
/**
@@ -377,10 +381,12 @@ struct tcp_req_done_item {
* Create tcp request info structure that keeps track of open
* requests on the TCP channel that are resolved at the same time,
* and the pending results that have to get written back to that client.
* @param base: comm base for read again timer.
* @param spoolbuf: shared buffer
* @return new structure or NULL on alloc failure.
*/
struct tcp_req_info* tcp_req_info_create(struct sldns_buffer* spoolbuf);
struct tcp_req_info* tcp_req_info_create(struct comm_base* base,
struct sldns_buffer* spoolbuf);
/**
* Delete tcp request structure. Called by owning commpoint.
+10
View File
@@ -786,3 +786,13 @@ size_t doq_table_quic_size_get(struct doq_table* ATTR_UNUSED(table))
return 0;
}
#endif
void tcp_read_again_cb(void* ATTR_UNUSED(arg))
{
/* nothing */
}
void tcp_more_read_again_cb(void* ATTR_UNUSED(arg))
{
/* nothing */
}
+2
View File
@@ -141,6 +141,8 @@ fptr_whitelist_comm_timer(void (*fptr)(void*))
#ifdef UB_ON_WINDOWS
else if(fptr == &wsvc_cron_cb) return 1;
#endif
else if(fptr == &tcp_read_again_cb) return 1;
else if(fptr == &tcp_more_read_again_cb) return 1;
else if(fptr == &auth_xfer_timer) return 1;
else if(fptr == &auth_xfer_probe_timer_callback) return 1;
else if(fptr == &auth_xfer_transfer_timer_callback) return 1;
+65 -1
View File
@@ -122,6 +122,10 @@
#define NUM_UDP_PER_SELECT 1
#endif
/** The number of TCP queries over a TCP connection, per read indication
* from select. */
#define NUM_TCP_PER_SELECT 100
/** timeout in millisec to wait for write to unblock, packets dropped after.*/
#define SEND_BLOCKED_WAIT_TIMEOUT 200
/** max number of times to wait for write to unblock, packets dropped after.*/
@@ -4621,6 +4625,10 @@ comm_point_tcp_handle_write(int fd, struct comm_point* c)
static int
tcp_req_info_read_again(int fd, struct comm_point* c)
{
/* One event-loop visit drains at most this many pipelined queries;
* the rest is re-queued, so that other file descriptors get
* serviced in between. */
int budget = NUM_TCP_PER_SELECT;
while(c->tcp_req_info->read_again) {
int r;
c->tcp_req_info->read_again = 0;
@@ -4637,6 +4645,16 @@ tcp_req_info_read_again(int fd, struct comm_point* c)
}
return 0;
}
if(--budget <= 0 && c->tcp_req_info->read_again) {
/* Defer the rest of the drain to the next loop turn.
* This uses a zero delay timer. For TLS the undrained
* remainder sits in OpenSSL's user-space buffer. */
struct timeval tv;
memset(&tv, 0, sizeof(tv));
verbose(VERB_ALGO, "Defer tcp_req_info read again");
comm_timer_set(c->tcp_req_info->read_again_timer, &tv);
return 1;
}
}
return 1;
}
@@ -4650,6 +4668,7 @@ tcp_more_read_again(int fd, struct comm_point* c)
/* this continues until the read routines get EAGAIN or so,
* and thus does not call the callback, and the bool is 0 */
int* moreread = c->tcp_more_read_again;
int budget = NUM_TCP_PER_SELECT;
while(moreread && *moreread) {
*moreread = 0;
if(!comm_point_tcp_handle_read(fd, c, 0)) {
@@ -4662,6 +4681,30 @@ tcp_more_read_again(int fd, struct comm_point* c)
}
return;
}
if(--budget <= 0 && *moreread) {
/* Defer the rest of the drain to the next loop turn.
* This uses a zero delay timer. For TLS the undrained
* remainder sits in OpenSSL's user-space buffer. */
struct timeval tv;
memset(&tv, 0, sizeof(tv));
if(!c->tcp_more_read_again_timer) {
c->tcp_more_read_again_timer = comm_timer_create(c->ev->base, tcp_more_read_again_cb, c);
if(!c->tcp_more_read_again_timer) {
log_err("out of memory for tcp more read again timer");
reclaim_tcp_handler(c);
if(!c->tcp_do_close) {
fptr_ok(fptr_whitelist_comm_point(
c->callback));
(void)(*c->callback)(c, c->cb_arg,
NETEVENT_CLOSED, NULL);
}
return;
}
}
verbose(VERB_ALGO, "Defer more read again");
comm_timer_set(c->tcp_more_read_again_timer, &tv);
return;
}
}
}
@@ -4689,6 +4732,23 @@ tcp_more_write_again(int fd, struct comm_point* c)
}
}
void
tcp_read_again_cb(void* arg)
{
struct tcp_req_info* req = (struct tcp_req_info*)arg;
verbose(VERB_ALGO, "tcp_read_again_cb");
if(!tcp_req_info_read_again(req->cp->fd, req->cp))
return;
}
void
tcp_more_read_again_cb(void* arg)
{
struct comm_point* c = (struct comm_point*)arg;
verbose(VERB_ALGO, "tcp_more_read_again_cb");
tcp_more_read_again(c->fd, c);
}
void
comm_point_tcp_handle_callback(int fd, short event, void* arg)
{
@@ -6131,7 +6191,7 @@ comm_point_create_tcp_handler(struct comm_base *base,
c->pp2_enabled = parent->pp2_enabled;
c->pp2_header_state = pp2_header_none;
if(spoolbuf) {
c->tcp_req_info = tcp_req_info_create(spoolbuf);
c->tcp_req_info = tcp_req_info_create(base, spoolbuf);
if(!c->tcp_req_info) {
log_err("could not create tcp commpoint");
sldns_buffer_free(c->buffer);
@@ -6693,6 +6753,9 @@ comm_point_close(struct comm_point* c)
*c->tcp_more_read_again = 0;
if(c->tcp_more_write_again && *c->tcp_more_write_again)
*c->tcp_more_write_again = 0;
if(c->tcp_more_read_again_timer &&
comm_timer_is_set(c->tcp_more_read_again_timer))
comm_timer_disable(c->tcp_more_read_again_timer);
/* close fd after removing from event lists, or epoll.. is messed up */
if(c->fd != -1 && !c->do_not_close) {
@@ -6732,6 +6795,7 @@ comm_point_delete(struct comm_point* c)
free(c->tcp_handlers);
}
free(c->timeout);
comm_timer_delete(c->tcp_more_read_again_timer);
if(c->type == comm_tcp || c->type == comm_local || c->type == comm_http) {
sldns_buffer_free(c->buffer);
#ifdef USE_DNSCRYPT
+9
View File
@@ -384,6 +384,9 @@ struct comm_point {
* Or leave NULL if it is not used at all. */
int* tcp_more_write_again;
/** resume timer for tcp_more_read_again */
struct comm_timer* tcp_more_read_again_timer;
/** if set, read/write completes:
read/write state of tcp is toggled.
buffer reset/bytecount reset.
@@ -1133,6 +1136,12 @@ void doq_send_pkt(struct comm_point* c, struct doq_pkt_addr* paddr,
/** doq timer callback function. */
void doq_timer_cb(void* arg);
/** tcp read again callback function. For tcp req info listen. */
void tcp_read_again_cb(void* arg);
/** tcp more read again callback function. For outside network. */
void tcp_more_read_again_cb(void* arg);
/**
* This routine is published for checks and tests, and is only used internally.
* handle libevent callback for timer comm.