[PATCH net] bpf: fix liveness marking

while processing Rx = Ry instruction the verifier does
regs[insn->dst_reg] = regs[insn->src_reg]
which often clears write mark (when Ry doesn't have it)
that was just set by check_reg_arg(Rx) prior to the assignment.
That causes mark_reg_read() to keep marking Rx in this block as
REG_LIVE_READ (since the logic incorrectly misses that it's
screened by the write) and in many of its parents (until lucky
write into the same Rx or beginning of the program).
That causes is_state_visited() logic to miss many pruning opportunities.

Furthermore mark_reg_read() logic propagates the read mark
for BPF_REG_FP as well (though it's readonly) which causes
harmless but unnecssary work during is_state_visited().
Note that do_propagate_liveness() skips FP correctly,
so do the same in mark_reg_read() as well.
It saves 0.2 seconds for the test below

program               before  after
bpf_lb-DLB_L3.o       2604    2304
bpf_lb-DLB_L4.o       11159   3723
bpf_lb-DUNKNOWN.o     1116    1110
bpf_lxc-DDROP_ALL.o   34566   28004
bpf_lxc-DUNKNOWN.o    53267   39026
bpf_netdev.o          17843   16943
bpf_overlay.o         8672    7929
time                  ~11 sec  ~4 sec

Fixes: dc503a8ad984 ("bpf/verifier: track liveness for pruning")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
---
 kernel/bpf/verifier.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
index b914fbe1383e..8b8d6ba39e23 100644
--- a/kernel/bpf/verifier.c
+++ b/kernel/bpf/verifier.c

@@ -653,6 +653,10 @@ static void mark_reg_read(const struct bpf_verifier_state *state, u32 regno)
 {
 	struct bpf_verifier_state *parent = state->parent;
 
+	if (regno == BPF_REG_FP)
+		/* We don't need to worry about FP liveness because it's read-only */
+		return;
+
 	while (parent) {
 		/* if read wasn't screened by an earlier write ... */
 		if (state->regs[regno].live & REG_LIVE_WRITTEN)

@@ -2345,6 +2349,7 @@ static int check_alu_op(struct bpf_verifier_env *env, struct bpf_insn *insn)
 				 * copy register state to dest reg
 				 */
 				regs[insn->dst_reg] = regs[insn->src_reg];
+				regs[insn->dst_reg].live |= REG_LIVE_WRITTEN;
 			} else {
 				/* R1 = (u32) R2 */
 				if (is_pointer_value(env, insn->src_reg)) {

-- 
2.9.5

On 06/10/17 00:20, Alexei Starovoitov wrote:

while processing Rx = Ry instruction the verifier does
regs[insn->dst_reg] = regs[insn->src_reg]
which often clears write mark (when Ry doesn't have it)
that was just set by check_reg_arg(Rx) prior to the assignment.
That causes mark_reg_read() to keep marking Rx in this block as
REG_LIVE_READ (since the logic incorrectly misses that it's
screened by the write) and in many of its parents (until lucky
write into the same Rx or beginning of the program).
That causes is_state_visited() logic to miss many pruning opportunities.

Good catch!

Furthermore mark_reg_read() logic propagates the read mark
for BPF_REG_FP as well (though it's readonly) which causes
harmless but unnecssary work during is_state_visited().

Surely it's unnecessary for is_state_visited() to even look at
 BPF_REG_FP anyway, so in addition to your change we could make
 states_equal just do `for (i = 0; i < BPF_REG_FP; i++)`?  That
 might save a bit more time.

Note that do_propagate_liveness() skips FP correctly,
so do the same in mark_reg_read() as well.
It saves 0.2 seconds for the test below

program               before  after
bpf_lb-DLB_L3.o       2604    2304
bpf_lb-DLB_L4.o       11159   3723
bpf_lb-DUNKNOWN.o     1116    1110
bpf_lxc-DDROP_ALL.o   34566   28004
bpf_lxc-DUNKNOWN.o    53267   39026
bpf_netdev.o          17843   16943
bpf_overlay.o         8672    7929
time                  ~11 sec  ~4 sec

Fixes: dc503a8ad984 ("bpf/verifier: track liveness for pruning")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Very nice numbers!
Acked-by: Edward Cree <redacted>

On 10/6/17 9:33 AM, Edward Cree wrote:

On 06/10/17 00:20, Alexei Starovoitov wrote:

quoted

while processing Rx = Ry instruction the verifier does
regs[insn->dst_reg] = regs[insn->src_reg]
which often clears write mark (when Ry doesn't have it)
that was just set by check_reg_arg(Rx) prior to the assignment.
That causes mark_reg_read() to keep marking Rx in this block as
REG_LIVE_READ (since the logic incorrectly misses that it's
screened by the write) and in many of its parents (until lucky
write into the same Rx or beginning of the program).
That causes is_state_visited() logic to miss many pruning opportunities.

Good catch!

quoted

Furthermore mark_reg_read() logic propagates the read mark
for BPF_REG_FP as well (though it's readonly) which causes
harmless but unnecssary work during is_state_visited().

Surely it's unnecessary for is_state_visited() to even look at
 BPF_REG_FP anyway, so in addition to your change we could make
 states_equal just do `for (i = 0; i < BPF_REG_FP; i++)`?  That
 might save a bit more time.

yeah. before this patch it was doing extra
memcmp(rold, rcur, ..) on FP reg. This patch saves this memcpy.
The i < BPF_REG_FP would effectively do the same, but I'm not sure
I want to do it just yet.
For net-next I have a bunch of changes for verifier to support bpf_call
and there two different states may have two different FPs.
One FP from caller and one from callee.
So I might still need to do full
for (i = 0; i < MAX_BPF_REG; i++)
      if (!regsafe(..))

quoted

Note that do_propagate_liveness() skips FP correctly,
so do the same in mark_reg_read() as well.
It saves 0.2 seconds for the test below

program               before  after
bpf_lb-DLB_L3.o       2604    2304
bpf_lb-DLB_L4.o       11159   3723
bpf_lb-DUNKNOWN.o     1116    1110
bpf_lxc-DDROP_ALL.o   34566   28004
bpf_lxc-DUNKNOWN.o    53267   39026
bpf_netdev.o          17843   16943
bpf_overlay.o         8672    7929
time                  ~11 sec  ~4 sec

Fixes: dc503a8ad984 ("bpf/verifier: track liveness for pruning")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Very nice numbers!
Acked-by: Edward Cree <redacted>

Thanks!

On 10/06/2017 01:20 AM, Alexei Starovoitov wrote:

while processing Rx = Ry instruction the verifier does
regs[insn->dst_reg] = regs[insn->src_reg]
which often clears write mark (when Ry doesn't have it)
that was just set by check_reg_arg(Rx) prior to the assignment.
That causes mark_reg_read() to keep marking Rx in this block as
REG_LIVE_READ (since the logic incorrectly misses that it's
screened by the write) and in many of its parents (until lucky
write into the same Rx or beginning of the program).
That causes is_state_visited() logic to miss many pruning opportunities.

Furthermore mark_reg_read() logic propagates the read mark
for BPF_REG_FP as well (though it's readonly) which causes
harmless but unnecssary work during is_state_visited().
Note that do_propagate_liveness() skips FP correctly,
so do the same in mark_reg_read() as well.
It saves 0.2 seconds for the test below

program               before  after
bpf_lb-DLB_L3.o       2604    2304
bpf_lb-DLB_L4.o       11159   3723
bpf_lb-DUNKNOWN.o     1116    1110
bpf_lxc-DDROP_ALL.o   34566   28004
bpf_lxc-DUNKNOWN.o    53267   39026
bpf_netdev.o          17843   16943
bpf_overlay.o         8672    7929
time                  ~11 sec  ~4 sec

Fixes: dc503a8ad984 ("bpf/verifier: track liveness for pruning")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

LGTM, thanks!

Acked-by: Daniel Borkmann <daniel@iogearbox.net>

From: Alexei Starovoitov <redacted>
Date: Thu, 5 Oct 2017 16:20:56 -0700

while processing Rx = Ry instruction the verifier does
regs[insn->dst_reg] = regs[insn->src_reg]
which often clears write mark (when Ry doesn't have it)
that was just set by check_reg_arg(Rx) prior to the assignment.
That causes mark_reg_read() to keep marking Rx in this block as
REG_LIVE_READ (since the logic incorrectly misses that it's
screened by the write) and in many of its parents (until lucky
write into the same Rx or beginning of the program).
That causes is_state_visited() logic to miss many pruning opportunities.

Furthermore mark_reg_read() logic propagates the read mark
for BPF_REG_FP as well (though it's readonly) which causes
harmless but unnecssary work during is_state_visited().
Note that do_propagate_liveness() skips FP correctly,
so do the same in mark_reg_read() as well.
It saves 0.2 seconds for the test below

program               before  after
bpf_lb-DLB_L3.o       2604    2304
bpf_lb-DLB_L4.o       11159   3723
bpf_lb-DUNKNOWN.o     1116    1110
bpf_lxc-DDROP_ALL.o   34566   28004
bpf_lxc-DUNKNOWN.o    53267   39026
bpf_netdev.o          17843   16943
bpf_overlay.o         8672    7929
time                  ~11 sec  ~4 sec

Fixes: dc503a8ad984 ("bpf/verifier: track liveness for pruning")
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

Looks great, applied.