h2: Rapid reset mitigations (7.3)

bin/varnishd/cache/cache_transport.h

@@ -44,6 +44,7 @@ typedef void vtr_sess_panic_f (struct vsb *, const struct sess *);
 typedef void vtr_req_panic_f (struct vsb *, const struct req *);
 typedef void vtr_req_fail_f (struct req *, stream_close_t);
 typedef void vtr_reembark_f (struct worker *, struct req *);
+typedef int vtr_poll_f (struct req *);
 typedef int vtr_minimal_response_f (struct req *, uint16_t status);
 
 struct transport {
@@ -64,6 +65,7 @@ struct transport {
 	vtr_sess_panic_f		*sess_panic;
 	vtr_req_panic_f			*req_panic;
 	vtr_reembark_f			*reembark;
+	vtr_poll_f			*poll;
 	vtr_minimal_response_f		*minimal_response;
 
 	VTAILQ_ENTRY(transport)		list;

bin/varnishd/cache/cache_vrt_vcl.c

@@ -42,6 +42,7 @@
 #include "vbm.h"
 
 #include "cache_director.h"
+#include "cache_transport.h"
 #include "cache_vcl.h"
 #include "vcc_interface.h"
 
@@ -521,6 +522,40 @@ VRT_VCL_Allow_Discard(struct vclref **refp)
 	FREE_OBJ(ref);
 }
 
+/*--------------------------------------------------------------------
+ */
+
+static int
+req_poll(struct worker *wrk, struct req *req)
+{
+	struct req *top;
+
+	/* NB: Since a fail transition leads to vcl_synth, the request may be
+	 * short-circuited twice.
+	 */
+	if (req->req_reset) {
+		wrk->vpi->handling = VCL_RET_FAIL;
+		return (-1);
+	}
+
+	top = req->top->topreq;
+	CHECK_OBJ_NOTNULL(top, REQ_MAGIC);
+	CHECK_OBJ_NOTNULL(top->transport, TRANSPORT_MAGIC);
+
+	if (!FEATURE(FEATURE_VCL_REQ_RESET))
+		return (0);
+	if (top->transport->poll == NULL)
+		return (0);
+	if (top->transport->poll(top) >= 0)
+		return (0);
+
+	VSLb_ts_req(req, "Reset", W_TIM_real(wrk));
+	wrk->stats->req_reset++;
+	wrk->vpi->handling = VCL_RET_FAIL;
+	req->req_reset = 1;
+	return (-1);
+}
+
 /*--------------------------------------------------------------------
  * Method functions to call into VCL programs.
  *
@@ -552,6 +587,8 @@ vcl_call_method(struct worker *wrk, struct req *req, struct busyobj *bo,
 		CHECK_OBJ_NOTNULL(req->sp, SESS_MAGIC);
 		CHECK_OBJ_NOTNULL(req->vcl, VCL_MAGIC);
 		CHECK_OBJ_NOTNULL(req->top, REQTOP_MAGIC);
+		if (req_poll(wrk, req))
+			return;
 		VCL_Req2Ctx(&ctx, req);
 	}
 	assert(ctx.now != 0);

bin/varnishd/http2/cache_http2.h

@@ -49,6 +49,7 @@ struct h2_error_s {
 
 typedef const struct h2_error_s *h2_error;
 
+#define H2_CUSTOM_ERRORS
 #define H2EC1(U,v,r,d) extern const struct h2_error_s H2CE_##U[1];
 #define H2EC2(U,v,r,d) extern const struct h2_error_s H2SE_##U[1];
 #define H2EC3(U,v,r,d) H2EC1(U,v,r,d) H2EC2(U,v,r,d)
@@ -193,6 +194,15 @@ struct h2_sess {
 	VTAILQ_HEAD(,h2_req)		txqueue;
 
 	h2_error			error;
+
+	// rst rate limit parameters, copied from h2_* parameters
+	vtim_dur			rapid_reset;
+	int64_t				rapid_reset_limit;
+	vtim_dur			rapid_reset_period;
+
+	// rst rate limit state
+	double				rst_budget;
+	vtim_real			last_rst;
 };
 
 #define ASSERT_RXTHR(h2) do {assert(h2->rxthr == pthread_self());} while(0)

bin/varnishd/http2/cache_http2_proto.c

@@ -47,6 +47,7 @@
 #include "vtcp.h"
 #include "vtim.h"
 
+#define H2_CUSTOM_ERRORS
 #define H2EC1(U,v,r,d) const struct h2_error_s H2CE_##U[1] = {{#U,d,v,0,1,r}};
 #define H2EC2(U,v,r,d) const struct h2_error_s H2SE_##U[1] = {{#U,d,v,1,0,r}};
 #define H2EC3(U,v,r,d) H2EC1(U,v,r,d) H2EC2(U,v,r,d)
@@ -317,9 +318,46 @@ h2_rx_push_promise(struct worker *wrk, struct h2_sess *h2, struct h2_req *r2)
 /**********************************************************************
  */
 
+static h2_error
+h2_rapid_reset(struct worker *wrk, struct h2_sess *h2, struct h2_req *r2)
+{
+	vtim_real now;
+	vtim_dur d;
+
+	CHECK_OBJ_NOTNULL(wrk, WORKER_MAGIC);
+	ASSERT_RXTHR(h2);
+	CHECK_OBJ_NOTNULL(r2, H2_REQ_MAGIC);
+
+	if (h2->rapid_reset_limit == 0)
+		return (0);
+
+	now = VTIM_real();
+	CHECK_OBJ_NOTNULL(r2->req, REQ_MAGIC);
+	AN(r2->req->t_first);
+	if (now - r2->req->t_first > h2->rapid_reset)
+		return (0);
+
+	d = now - h2->last_rst;
+	h2->rst_budget += h2->rapid_reset_limit * d /
+	    h2->rapid_reset_period;
+	h2->rst_budget = vmin_t(double, h2->rst_budget,
+	    h2->rapid_reset_limit);
+	h2->last_rst = now;
+
+	if (h2->rst_budget < 1.0) {
+		Lck_Lock(&h2->sess->mtx);
+		VSLb(h2->vsl, SLT_Error, "H2: Hit RST limit. Closing session.");
+		Lck_Unlock(&h2->sess->mtx);
+		return (H2CE_RAPID_RESET);
+	}
+	h2->rst_budget -= 1.0;
+	return (0);
+}
+
 static h2_error v_matchproto_(h2_rxframe_f)
 h2_rx_rst_stream(struct worker *wrk, struct h2_sess *h2, struct h2_req *r2)
 {
+	h2_error h2e;
 
 	CHECK_OBJ_NOTNULL(wrk, WORKER_MAGIC);
 	ASSERT_RXTHR(h2);
@@ -329,8 +367,9 @@ h2_rx_rst_stream(struct worker *wrk, struct h2_sess *h2, struct h2_req *r2)
 		return (H2CE_FRAME_SIZE_ERROR);
 	if (r2 == NULL)
 		return (0);
+	h2e = h2_rapid_reset(wrk, h2, r2);
 	h2_kill_req(wrk, h2, r2, h2_streamerror(vbe32dec(h2->rxf_data)));
-	return (0);
+	return (h2e);
 }
 
 /**********************************************************************

bin/varnishd/http2/cache_http2_session.c

@@ -128,6 +128,14 @@ h2_init_sess(struct sess *sp,
 	h2->remote_settings = H2_proto_settings;
 	h2->decode = decode;
 
+	h2->rapid_reset = cache_param->h2_rapid_reset;
+	h2->rapid_reset_limit = cache_param->h2_rapid_reset_limit;
+	h2->rapid_reset_period = cache_param->h2_rapid_reset_period;
+
+	h2->rst_budget = h2->rapid_reset_limit;
+	h2->last_rst = sp->t_open;
+	AZ(isnan(h2->last_rst));
+
 	AZ(VHT_Init(h2->dectbl, h2->local_settings.header_table_size));
 
 	*up = (uintptr_t)h2;
@@ -438,6 +446,16 @@ h2_new_session(struct worker *wrk, void *arg)
 	wrk->vsl = NULL;
 }
 
+static int v_matchproto_(vtr_poll_f)
+h2_poll(struct req *req)
+{
+	struct h2_req *r2;
+
+	CHECK_OBJ_NOTNULL(req, REQ_MAGIC);
+	CAST_OBJ_NOTNULL(r2, req->transport_priv, H2_REQ_MAGIC);
+	return (r2->error ? -1 : 1);
+}
+
 struct transport HTTP2_transport = {
 	.name =			"HTTP/2",
 	.magic =		TRANSPORT_MAGIC,
@@ -447,4 +465,5 @@ struct transport HTTP2_transport = {
 	.req_body =		h2_req_body,
 	.req_fail =		h2_req_fail,
 	.sess_panic =		h2_sess_panic,
+	.poll =			h2_poll,
 };

bin/varnishtest/tests/r03996.vtc

@@ -0,0 +1,90 @@
+varnishtest "h2 rapid reset"
+
+barrier b1 sock 2 -cyclic
+barrier b2 sock 5 -cyclic
+
+server s1 {
+	rxreq
+	txresp
+} -start
+
+varnish v1 -cliok "param.set feature +http2"
+varnish v1 -cliok "param.set debug +syncvsl"
+varnish v1 -cliok "param.set h2_rapid_reset_limit 3"
+varnish v1 -cliok "param.set h2_rapid_reset 5"
+
+varnish v1 -vcl+backend {
+	import vtc;
+
+	sub vcl_recv {
+		if (req.http.barrier) {
+			vtc.barrier_sync(req.http.barrier);
+		}
+		vtc.barrier_sync("${b2_sock}");
+	}
+
+} -start
+
+client c1 {
+	stream 0 {
+		rxgoaway
+		expect goaway.err == ENHANCE_YOUR_CALM
+	} -start
+
+	stream 1 {
+		txreq -hdr barrier ${b1_sock}
+		barrier b1 sync
+		txrst
+	} -run
+	stream 3 {
+		txreq -hdr barrier ${b1_sock}
+		barrier b1 sync
+		txrst
+	} -run
+	stream 5 {
+		txreq -hdr barrier ${b1_sock}
+		barrier b1 sync
+		txrst
+	} -run
+	stream 7 {
+		txreq -hdr barrier ${b1_sock}
+		barrier b1 sync
+		txrst
+	} -run
+
+	barrier b2 sync
+	stream 0 -wait
+} -run
+
+varnish v1 -expect sc_rapid_reset == 1
+
+varnish v1 -cliok "param.set feature -vcl_req_reset"
+
+client c2 {
+	stream 0 {
+		rxgoaway
+		expect goaway.err == ENHANCE_YOUR_CALM
+	} -start
+
+	stream 1 {
+		txreq
+		txrst
+	} -run
+	stream 3 {
+		txreq
+		txrst
+	} -run
+	stream 5 {
+		txreq
+		txrst
+	} -run
+	stream 7 {
+		txreq
+		txrst
+	} -run
+
+	barrier b2 sync
+	stream 0 -wait
+} -run
+
+varnish v1 -expect sc_rapid_reset == 2