check bpf_func_state->callback_depth when pruning states

This patch-set fixes bug in states pruning logic hit in mailing list discussion [0]. The details of the fix are in patch #2.
A change to the test case test_tcp_custom_syncookie.c is necessary, otherwise updated verifier won't be able to process it due to instruction complexity limit. This change is done in patch #1.

The main idea for the fix belongs to Yonghong Song, mine contribution is merely in review and test cases.

There are some changes in verification performance:

File                       Program   Insns   (DIFF)  States (A)  States (B)  States (DIFF)
-------------------------  --------  --------------  ----------  ----------  -------------
pyperf600_bpf_loop.bpf.o   on_event    +53 (+1.09%)         323         330    +7 (+2.17%)
strobemeta_bpf_loop.bpf.o  on_event  +594 (+27.85%)         163         213  +50 (+30.67%)

(when tested using BPF selftests and Cilium object files)

Changelog:
v1 [1] -> v2:
- patch #2 commit message updated to better reflect verifier behavior with regards to checkpoints tree (suggested by Yonghong);
- veristat results added (suggested by Andrii).

[0] https://lore.kernel.org/bpf/[email protected]/
[1] https://lore.kernel.org/bpf/[email protected]/

Eduard Zingerman (3):
selftests/bpf: update tcp_custom_syncookie to use scalar packet offset
bpf: check bpf_func_state->callback_depth when pruning states selftests/bpf: test case for callback_depth states pruning logic

 kernel/bpf/verifier.c                         |  3 +
 .../bpf/progs/test_tcp_custom_syncookie.c     | 83 ++++++++++++-------
 .../bpf/progs/verifier_iterating_callbacks.c  | 70 ++++++++++++++++
 3 files changed, 126 insertions(+), 30 deletions(-)

2.43.0

The next commit in a series fixes bug in bpf_loop() handling. That change makes tcp_custom_syncookie test too complex to verify.

This commit updates tcp_custom_syncookie.c:tcp_parse_option() to use explicit packet offset (ctx->off) for packet access instead of ever moving pointer (ctx->ptr), this reduces verification complexity:
- the tcp_parse_option() is passed as a callback to bpf_loop();
- suppose a checkpoint is created each time at function entry;
- the ctx->ptr is tracked by verifier as PTR_TO_PACKET;
- the ctx->ptr is incremented in tcp_parse_option(),
thus umax_value field tracked for it is incremented as well;
- on each next iteration of tcp_parse_option()
checkpoint from a previous iteration can't be reused for state pruning, because PTR_TO_PACKET registers are considered equivalent only if old->umax_value >= cur->umax_value;
- on the other hand, the ctx->off is a SCALAR,
subject to widen_imprecise_scalars();
- it's exact bounds are eventually forgotten and it is tracked as unknown scalar at entry to tcp_parse_option();
- hence checkpoints created at the start of the function eventually converge.

The change is similar to one applied in [0] to xdp_synproxy_kern.c.

[0] commit 977bc146d4eb ("selftests/bpf: track tcp payload offset as scalar in xdp_synproxy")

---
 .../bpf/progs/test_tcp_custom_syncookie.c     | 83 ++++++++++++-------
 1 file changed, 53 insertions(+), 30 deletions(-)

index a5501b29979a..c8e4553648bf 100644
@@ -10,6 +10,8 @@
 #include "test_siphash.h"
 #include "test_tcp_custom_syncookie.h"
 
+#define MAX_PACKET_OFF 0xffff
+
 /* Hash is calculated for each client and split into ISN and TS.
  *
  *       MSB                                   LSB

@@ -52,16 +54,15 @@ static siphash_key_t test_key_siphash = {
 
 struct tcp_syncookie {
 	struct __sk_buff *skb;
+	void *data;
 	void *data_end;
 	struct ethhdr *eth;
 	struct iphdr *ipv4;
 	struct ipv6hdr *ipv6;
 	struct tcphdr *tcp;
-	union {
-		char *ptr;
-		__be32 *ptr32;
-	};
+	__be32 *ptr32;
 	struct bpf_tcp_req_attrs attrs;
+	u32 off;
 	u32 cookie;
 	u64 first;
 };

@@ -70,6 +71,7 @@ bool handled_syn, handled_ack;
 
 static int tcp_load_headers(struct tcp_syncookie *ctx)
 {
+	ctx->data = (void *)(long)ctx->skb->data;
 	ctx->data_end = (void *)(long)ctx->skb->data_end;
 	ctx->eth = (struct ethhdr *)(long)ctx->skb->data;
 

@@ -134,6 +136,7 @@ static int tcp_reload_headers(struct tcp_syncookie *ctx)
 	if (bpf_skb_change_tail(ctx->skb, data_len + 60 - ctx->tcp->doff * 4, 0))
 		goto err;
 
+	ctx->data = (void *)(long)ctx->skb->data;
 	ctx->data_end = (void *)(long)ctx->skb->data_end;
 	ctx->eth = (struct ethhdr *)(long)ctx->skb->data;
 	if (ctx->ipv4) {

@@ -195,47 +198,68 @@ static int tcp_validate_header(struct tcp_syncookie *ctx)
 	return -1;
 }
 
-static int tcp_parse_option(__u32 index, struct tcp_syncookie *ctx)
+static __always_inline void *next(struct tcp_syncookie *ctx, __u32 sz)
 {
-	char opcode, opsize;
+	__u64 off = ctx->off;
+	__u8 *data;
 
-	if (ctx->ptr + 1 > ctx->data_end)
-		goto stop;
+	/* Verifier forbids access to packet when offset exceeds MAX_PACKET_OFF */
+	if (off > MAX_PACKET_OFF - sz)
+		return NULL;
+
+	data = ctx->data + off;
+	barrier_var(data);
+	if (data + sz >= ctx->data_end)
+		return NULL;
 
-	opcode = *ctx->ptr++;
+	ctx->off += sz;
+	return data;
+}
 
-	if (opcode == TCPOPT_EOL)
+static int tcp_parse_option(__u32 index, struct tcp_syncookie *ctx)
+{
+	__u8 *opcode, *opsize, *wscale;
+	__u32 *tsval, *tsecr;
+	__u16 *mss;
+	__u32 off;
+
+	off = ctx->off;
+	opcode = next(ctx, 1);
+	if (!opcode)
 		goto stop;
 
-	if (opcode == TCPOPT_NOP)
+	if (*opcode == TCPOPT_EOL)
+		goto stop;
+
+	if (*opcode == TCPOPT_NOP)
 		goto next;
 
-	if (ctx->ptr + 1 > ctx->data_end)
+	opsize = next(ctx, 1);
+	if (!opsize)
 		goto stop;
 
-	opsize = *ctx->ptr++;
-
-	if (opsize < 2)
+	if (*opsize < 2)
 		goto stop;
 
-	switch (opcode) {
+	switch (*opcode) {
 	case TCPOPT_MSS:
-		if (opsize == TCPOLEN_MSS && ctx->tcp->syn &&
-		    ctx->ptr + (TCPOLEN_MSS - 2) < ctx->data_end)
-			ctx->attrs.mss = get_unaligned_be16(ctx->ptr);
+		mss = next(ctx, 2);
+		if (*opsize == TCPOLEN_MSS && ctx->tcp->syn && mss)
+			ctx->attrs.mss = get_unaligned_be16(mss);
 		break;
 	case TCPOPT_WINDOW:
-		if (opsize == TCPOLEN_WINDOW && ctx->tcp->syn &&
-		    ctx->ptr + (TCPOLEN_WINDOW - 2) < ctx->data_end) {
+		wscale = next(ctx, 1);
+		if (*opsize == TCPOLEN_WINDOW && ctx->tcp->syn && wscale) {
 			ctx->attrs.wscale_ok = 1;
-			ctx->attrs.snd_wscale = *ctx->ptr;
+			ctx->attrs.snd_wscale = *wscale;
 		}
 		break;
 	case TCPOPT_TIMESTAMP:
-		if (opsize == TCPOLEN_TIMESTAMP &&
-		    ctx->ptr + (TCPOLEN_TIMESTAMP - 2) < ctx->data_end) {
-			ctx->attrs.rcv_tsval = get_unaligned_be32(ctx->ptr);
-			ctx->attrs.rcv_tsecr = get_unaligned_be32(ctx->ptr + 4);
+		tsval = next(ctx, 4);
+		tsecr = next(ctx, 4);
+		if (*opsize == TCPOLEN_TIMESTAMP && tsval && tsecr) {
+			ctx->attrs.rcv_tsval = get_unaligned_be32(tsval);
+			ctx->attrs.rcv_tsecr = get_unaligned_be32(tsecr);
 
 			if (ctx->tcp->syn && ctx->attrs.rcv_tsecr)
 				ctx->attrs.tstamp_ok = 0;

@@ -244,13 +268,12 @@ static int tcp_parse_option(__u32 index, struct tcp_syncookie *ctx)
 		}
 		break;
 	case TCPOPT_SACK_PERM:
-		if (opsize == TCPOLEN_SACK_PERM && ctx->tcp->syn &&
-		    ctx->ptr + (TCPOLEN_SACK_PERM - 2) < ctx->data_end)
+		if (*opsize == TCPOLEN_SACK_PERM && ctx->tcp->syn)
 			ctx->attrs.sack_ok = 1;
 		break;
 	}
 
-	ctx->ptr += opsize - 2;
+	ctx->off = off + *opsize;
 next:
 	return 0;
 stop:

@@ -259,7 +282,7 @@ static int tcp_parse_option(__u32 index, struct tcp_syncookie *ctx)
 
 static void tcp_parse_options(struct tcp_syncookie *ctx)
 {
-	ctx->ptr = (char *)(ctx->tcp + 1);
+	ctx->off = (__u8 *)(ctx->tcp + 1) - (__u8 *)ctx->data,
 
 	bpf_loop(40, tcp_parse_option, ctx, 0);
 }

2.43.0

When comparing current and cached states verifier should consider bpf_func_state->callback_depth. Current state cannot be pruned against cached state, when current states has more iterations left compared to cached state. Current state has more iterations left when it's callback_depth is smaller.

Below is an example illustrating this bug, minimized from mailing list discussion [0].
The example is not a safe program: if loop_cb point (1) is followed by loop_cb point (2), then division by zero is possible at point (4).

    struct ctx {
    	__u64 a;
    	__u64 b;
    	__u64 c;
    };

    static void loop_cb(int i, struct ctx *ctx)
    {
    	/* assume that generated code is "fallthrough-first":
    	 * if ... == 1 goto
    	 * if ... == 2 goto
    	 * <default>
    	 */
    	switch (bpf_get_prandom_u32()) {
    	case 1:  /* 1 */ ctx->a = 42; return 0; break;
    	case 2:  /* 2 */ ctx->b = 42; return 0; break;
    	default: /* 3 */ ctx->c = 42; return 0; break;
    	}
    }

    SEC("tc")
    __failure
    __flag(BPF_F_TEST_STATE_FREQ)
    int test(struct __sk_buff *skb)
    {
    	struct ctx ctx = { 7, 7, 7 };

    	bpf_loop(2, loop_cb, &ctx, 0);              /* 0 */
    	/* assume generated checks are in-order: .a first */
    	if (ctx.a == 42 && ctx.b == 42 && ctx.c == 7)
    		asm volatile("r0 /= 0;":::"r0");    /* 4 */
    	return 0;
    }

Prior to this commit verifier built the following checkpoint tree for this example:

.------------------------------------- Checkpoint / State name

 |    .-------------------------------- Code point number
 |    |   .---------------------------- Stack state {ctx.a,ctx.b,ctx.c}
 |    |   |        .------------------- Callback depth in frame #0
 v    v   v        v

- (0) {7P,7P,7},depth=0
- (3) {7P,7P,7},depth=1
- (0) {7P,7P,42},depth=1
- (3) {7P,7,42},depth=2
- (0) {7P,7,42},depth=2 loop terminates because of depth limit
- (4) {7P,7,42},depth=0 predicted false, ctx.a marked precise
- (6) exit
- (2) {7P,7,42},depth=2
(a) - (0) {7P,42,42},depth=2 loop terminates because of depth limit
- (4) {7P,42,42},depth=0 predicted false, ctx.a marked precise
- (6) exit
(b) - (1) {7P,7P,42},depth=2
- (0) {42P,7P,42},depth=2 loop terminates because of depth limit
- (4) {42P,7P,42},depth=0 predicted false, ctx.{a,b} marked precise
- (6) exit
- (2) {7P,7,7},depth=1
- (0) {7P,42,7},depth=1 considered safe, pruned using checkpoint (a)
(c) - (1) {7P,7P,7},depth=1 considered safe, pruned using checkpoint (b)

Here checkpoint (b) has callback_depth of 2, meaning that it would never reach state {42,42,7}.
While checkpoint (c) has callback_depth of 1, and thus could yet explore the state {42,42,7} if not pruned prematurely. This commit makes forbids such premature pruning, allowing verifier to explore states sub-tree starting at (c):

(c) - (1) {7,7,7P},depth=1
- (0) {42P,7,7P},depth=1
...
- (2) {42,7,7},depth=2
- (0) {42,42,7},depth=2 loop terminates because of depth limit
- (4) {42,42,7},depth=0 predicted true, ctx.{a,b,c} marked precise
- (5) division by zero

[0] https://lore.kernel.org/bpf/[email protected]/

---
 kernel/bpf/verifier.c | 3 +++
 1 file changed, 3 insertions(+)

index 011d54a1dc53..c1fa1de590dc 100644
@@ -16705,6 +16705,9 @@ static bool func_states_equal(struct bpf_verifier_env *env, struct bpf_func_stat
 {
 	int i;
 
+	if (old->callback_depth > cur->callback_depth)
+		return false;
+
 	for (i = 0; i < MAX_BPF_REG; i++)
 		if (!regsafe(env, &old->regs[i], &cur->regs[i],
 			     &env->idmap_scratch, exact))

2.43.0

Missing Fixes: tag? Also, shouldn't this go into bpf tree instead of bpf-next?

Otherwise everything looks good.

On Fri, 2024-02-16 at 10:16 -0800, Andrii Nakryiko wrote:
[...]

Right, sorry aboutt that

Will re-send v3 with fixes tag to 'bpf'

On Sat, 2024-02-17 at 20:19 +0200, Eduard Zingerman wrote:
[...]

Sending via 'bpf' tree would require dropping patch #1. The test_tcp_custom_syncookie is not yet in 'bpf'. Note that patch #2 breaks syncookie test w/o patch #1. Should I split this in two parts?
- patch #1 - send via bpf-next
- patch #2,3 - send via bpf

Sounds good to me. Patch 1 will likely to be merged in bpf-next before patch 2/3 merged in bpf tree and circulared back to bpf-next. Please add related commit message in Patch 1 to explain actual fix will go to bpf tree and will be back to bpf-next later.

The test case was minimized from mailing list discussion [0]. It is equivalent to the following C program:

    struct iter_limit_bug_ctx { __u64 a; __u64 b; __u64 c; };

    static __naked void iter_limit_bug_cb(void)
    {
    	switch (bpf_get_prandom_u32()) {
    	case 1:  ctx->a = 42; break;
    	case 2:  ctx->b = 42; break;
    	default: ctx->c = 42; break;
    	}
    }

    int iter_limit_bug(struct __sk_buff *skb)
    {
    	struct iter_limit_bug_ctx ctx = { 7, 7, 7 };

    	bpf_loop(2, iter_limit_bug_cb, &ctx, 0);
    	if (ctx.a == 42 && ctx.b == 42 && ctx.c == 7)
    	  asm volatile("r1 /= 0;":::"r1");
    	return 0;
    }

The main idea is that each loop iteration changes one of the state variables in a non-deterministic manner. Hence it is premature to prune the states that have two iterations left comparing them to states with one iteration left.
E.g. {{7,7,7}, callback_depth=0} can reach state {42,42,7}, while {{7,7,7}, callback_depth=1} can't.

[0] https://lore.kernel.org/bpf/[email protected]/

---
 .../bpf/progs/verifier_iterating_callbacks.c  | 70 +++++++++++++++++++
 1 file changed, 70 insertions(+)

index 5905e036e0ea..a955a6358206 100644
@@ -239,4 +239,74 @@ int bpf_loop_iter_limit_nested(void *unused)
 	return 1000 * a + b + c;
 }
 
+struct iter_limit_bug_ctx {
+	__u64 a;
+	__u64 b;
+	__u64 c;
+};
+
+static __naked void iter_limit_bug_cb(void)
+{
+	/* This is the same as C code below, but written
+	 * in assembly to control which branches are fall-through.
+	 *
+	 *   switch (bpf_get_prandom_u32()) {
+	 *   case 1:  ctx->a = 42; break;
+	 *   case 2:  ctx->b = 42; break;
+	 *   default: ctx->c = 42; break;
+	 *   }
+	 */
+	asm volatile (
+	"r9 = r2;"
+	"call %[bpf_get_prandom_u32];"
+	"r1 = r0;"
+	"r2 = 42;"
+	"r0 = 0;"
+	"if r1 == 0x1 goto 1f;"
+	"if r1 == 0x2 goto 2f;"
+	"*(u64 *)(r9 + 16) = r2;"
+	"exit;"
+	"1: *(u64 *)(r9 + 0) = r2;"
+	"exit;"
+	"2: *(u64 *)(r9 + 8) = r2;"
+	"exit;"
+	:
+	: __imm(bpf_get_prandom_u32)
+	: __clobber_all
+	);
+}
+
+SEC("tc")
+__failure
+__flag(BPF_F_TEST_STATE_FREQ)
+int iter_limit_bug(struct __sk_buff *skb)
+{
+	struct iter_limit_bug_ctx ctx = { 7, 7, 7 };
+
+	bpf_loop(2, iter_limit_bug_cb, &ctx, 0);
+
+	/* This is the same as C code below,
+	 * written in assembly to guarantee checks order.
+	 *
+	 *   if (ctx.a == 42 && ctx.b == 42 && ctx.c == 7)
+	 *     asm volatile("r1 /= 0;":::"r1");
+	 */
+	asm volatile (
+	"r1 = *(u64 *)%[ctx_a];"
+	"if r1 != 42 goto 1f;"
+	"r1 = *(u64 *)%[ctx_b];"
+	"if r1 != 42 goto 1f;"
+	"r1 = *(u64 *)%[ctx_c];"
+	"if r1 != 7 goto 1f;"
+	"r1 /= 0;"
+	"1:"
+	:
+	: [ctx_a]"m"(ctx.a),
+	  [ctx_b]"m"(ctx.b),
+	  [ctx_c]"m"(ctx.c)
+	: "r1"
+	);
+	return 0;
+}
+
 char _license[] SEC("license") = "GPL";

2.43.0