Compilers optimize conditional operators at will, but often bpf programmers
want to force compilers to keep the same operator in asm as it's written in C.
Introduce bpf_cmp_likely/unlikely(var1, conditional_op, var2) macros that can be used as:
- if (seen >= 1000)
+ if (bpf_cmp_unlikely(seen, >=, 1000))
The macros take advantage of BPF assembly that is C like.
The macros check the sign of variable 'seen' and emits either
signed or unsigned compare.
For example:
int a;
bpf_cmp_unlikely(a, >, 0) will be translated to 'if rX s> 0 goto' in BPF assembly.
unsigned int a;
bpf_cmp_unlikely(a, >, 0) will be translated to 'if rX > 0 goto' in BPF assembly.
C type conversions coupled with comparison operator are tricky.
int i = -1;
unsigned int j = 1;
if (i < j) // this is false.
long i = -1;
unsigned int j = 1;
if (i < j) // this is true.
Make sure BPF program is compiled with -Wsign-compare then the macros will catch
the mistake.
The macros check LHS (left hand side) only to figure out the sign of compare.
'if 0 < rX goto' is not allowed in the assembly, so the users
have to use a variable on LHS anyway.
The patch updates few tests to demonstrate the use of the macros.
The macro allows to use BPF_JSET in C code, since LLVM doesn't generate it at
present. For example:
if (i & j) compiles into r0 &= r1; if r0 == 0 goto
while
if (bpf_cmp_unlikely(i, &, j)) compiles into if r0 & r1 goto
Note that the macros has to be careful with RHS assembly predicate.
Since:
u64 __rhs = 1ull << 42;
asm goto("if r0 < %[rhs] goto +1" :: [rhs] "ri" (__rhs));
LLVM will silently truncate 64-bit constant into s32 imm.
Note that [lhs] "r"((short)LHS) the type cast is a workaround for LLVM issue.
When LHS is exactly 32-bit LLVM emits redundant <<=32, >>=32 to zero upper 32-bits.
When LHS is 64 or 16 or 8-bit variable there are no shifts.
When LHS is 32-bit the (u64) cast doesn't help. Hence use (short) cast.
It does _not_ truncate the variable before it's assigned to a register.
Traditional likely()/unlikely() macros that use __builtin_expect(!!(x), 1 or 0)
have no effect on these macros, hence macros implement the logic manually.
bpf_cmp_unlikely() macro preserves compare operator as-is while
bpf_cmp_likely() macro flips the compare.
Consider two cases:
A.
for() {
if (foo >= 10) {
bar += foo;
}
other code;
}
B.
for() {
if (foo >= 10)
break;
other code;
}
It's ok to use either bpf_cmp_likely or bpf_cmp_unlikely macros in both cases,
but consider that 'break' is effectively 'goto out_of_the_loop'.
Hence it's better to use bpf_cmp_unlikely in the B case.
While 'bar += foo' is better to keep as 'fallthrough' == likely code path in the A case.
When it's written as:
A.
for() {
if (bpf_cmp_likely(foo, >=, 10)) {
bar += foo;
}
other code;
}
B.
for() {
if (bpf_cmp_unlikely(foo, >=, 10))
break;
other code;
}
The assembly will look like:
A.
for() {
if r1 < 10 goto L1;
bar += foo;
L1:
other code;
}
B.
for() {
if r1 >= 10 goto L2;
other code;
}
L2:
The bpf_cmp_likely vs bpf_cmp_unlikely changes basic block layout, hence it will
greatly influence the verification process. The number of processed instructions
will be different, since the verifier walks the fallthrough first.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/bpf/20231226191148.48536-3-alexei.starovoitov@gmail.com
Convergence for open coded iterators is computed in is_state_visited()
by examining states with branches count > 1 and using states_equal().
states_equal() computes sub-state relation using read and precision marks.
Read and precision marks are propagated from children states,
thus are not guaranteed to be complete inside a loop when branches
count > 1. This could be demonstrated using the following unsafe program:
1. r7 = -16
2. r6 = bpf_get_prandom_u32()
3. while (bpf_iter_num_next(&fp[-8])) {
4. if (r6 != 42) {
5. r7 = -32
6. r6 = bpf_get_prandom_u32()
7. continue
8. }
9. r0 = r10
10. r0 += r7
11. r8 = *(u64 *)(r0 + 0)
12. r6 = bpf_get_prandom_u32()
13. }
Here verifier would first visit path 1-3, create a checkpoint at 3
with r7=-16, continue to 4-7,3 with r7=-32.
Because instructions at 9-12 had not been visitied yet existing
checkpoint at 3 does not have read or precision mark for r7.
Thus states_equal() would return true and verifier would discard
current state, thus unsafe memory access at 11 would not be caught.
This commit fixes this loophole by introducing exact state comparisons
for iterator convergence logic:
- registers are compared using regs_exact() regardless of read or
precision marks;
- stack slots have to have identical type.
Unfortunately, this is too strict even for simple programs like below:
i = 0;
while(iter_next(&it))
i++;
At each iteration step i++ would produce a new distinct state and
eventually instruction processing limit would be reached.
To avoid such behavior speculatively forget (widen) range for
imprecise scalar registers, if those registers were not precise at the
end of the previous iteration and do not match exactly.
This a conservative heuristic that allows to verify wide range of
programs, however it precludes verification of programs that conjure
an imprecise value on the first loop iteration and use it as precise
on the second.
Test case iter_task_vma_for_each() presents one of such cases:
unsigned int seen = 0;
...
bpf_for_each(task_vma, vma, task, 0) {
if (seen >= 1000)
break;
...
seen++;
}
Here clang generates the following code:
<LBB0_4>:
24: r8 = r6 ; stash current value of
... body ... 'seen'
29: r1 = r10
30: r1 += -0x8
31: call bpf_iter_task_vma_next
32: r6 += 0x1 ; seen++;
33: if r0 == 0x0 goto +0x2 <LBB0_6> ; exit on next() == NULL
34: r7 += 0x10
35: if r8 < 0x3e7 goto -0xc <LBB0_4> ; loop on seen < 1000
<LBB0_6>:
... exit ...
Note that counter in r6 is copied to r8 and then incremented,
conditional jump is done using r8. Because of this precision mark for
r6 lags one state behind of precision mark on r8 and widening logic
kicks in.
Adding barrier_var(seen) after conditional is sufficient to force
clang use the same register for both counting and conditional jump.
This issue was discussed in the thread [1] which was started by
Andrew Werner <awerner32@gmail.com> demonstrating a similar bug
in callback functions handling. The callbacks would be addressed
in a followup patch.
[1] https://lore.kernel.org/bpf/97a90da09404c65c8e810cf83c94ac703705dc0e.camel@gmail.com/
Co-developed-by: Andrii Nakryiko <andrii.nakryiko@gmail.com>
Co-developed-by: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Signed-off-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/r/20231024000917.12153-4-eddyz87@gmail.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The open-coded task_vma iter added earlier in this series allows for
natural iteration over a task's vmas using existing open-coded iter
infrastructure, specifically bpf_for_each.
This patch adds a test demonstrating this pattern and validating
correctness. The vma->vm_start and vma->vm_end addresses of the first
1000 vmas are recorded and compared to /proc/PID/maps output. As
expected, both see the same vmas and addresses - with the exception of
the [vsyscall] vma - which is explained in a comment in the prog_tests
program.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20231013204426.1074286-5-davemarchevsky@fb.com
Alexei Starovoitov
Eduard Zingerman
Dave Marchevsky