libbpf-tools: add cpu filter for hardirqs/softirqs #5107
Conversation
ZhangYet
requested review from
drzaeus77,
goldshtn,
yonghong-song,
4ast,
brendangregg and
davemarchevsky
as code owners
ZhangYet
force-pushed
the
dante-dev
branch
10 times, most recently
from
cee19ff
to
6bb7124
Compare
libbpf-tools/softirqs.bpf.c
Outdated
| u64 delta, *tsp; | ||
| u32 key = 0; | ||
|
|
||
| if (vec_nr >= NR_SOFTIRQS) | ||
| return 0; | ||
| vec_nr &= 0xf; // to avoid compiler opti make verifier error |
There was a problem hiding this comment.
which compiler do you use and what kind of error you hit on which kernel?
There was a problem hiding this comment.
Please give details about the above verifier failure (kernel version, compiler version, error message, etc.)
There was a problem hiding this comment.
I have replied your #5107 (comment) yesterday as below (but it shows pending I think that's why you didn't see it):
clang: 18.1.8
kernel: 6.10.10-arch1-1
Here is the error log:
if (vec_nr >= NR_SOFTIRQS)
return 0;https://gist.github.com/ZhangYet/5d293455680ef9953f7aed13a56463ac
There was a problem hiding this comment.
Sorry, I missed the original comments.
From the verifier log, the following is the reason:
18: (bf) r1 = r6 ; frame1: R1_w=scalar(id=1) R6=scalar(id=1)
19: (67) r1 <<= 32 ; frame1: R1_w=scalar(smax=0x7fffffff00000000,umax=0xffffffff00000000,smin32=0,smax32=umax32=0,var_off=(0x0; 0xffffffff00000000))
20: (77) r1 >>= 32 ; frame1: R1_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
21: (25) if r1 > 0x9 goto pc+72 ; frame1: R1_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=9,var_off=(0x0; 0xf))
22: (bf) r2 = r10 ; frame1: R2_w=fp0 R10=fp0
...
39: (67) r6 <<= 32 ; frame1: R6_w=scalar(smax=0x7fffffff00000000,umax=0xffffffff00000000,smin32=0,smax32=umax32=0,var_off=(0x0; 0xffffffff00000000))
40: (77) r6 >>= 32 ; frame1: R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
...
In the above, at insn 22, we know r1 <= 0x9, and not that insn 18, we have r1 = r6, so lower 32bit of r6 should be <= 0x9 as well.
After insn 40, we know r6 top 32bits are 0, so in ideal case verifier should conclude r6 <= 0x9, but unfortunately verifier is not sophisticated enough to derive the precise bound.
I suspect -mcpu=v3 might work as we could remove some left/right shifts.
There was a problem hiding this comment.
Great. -mcpu=v3 works well.
I added
BPFCFLAGS_softirqs := $(BPFCFLAGS) -mcpu=v3
$(OUTPUT)/softirqs.bpf.o: BPFCFLAGS = $(BPFCFLAGS_softirqs)
I wonder if there is a better way to add -mcpu=v3 for softirqs. If so, I will modify and rebase the commit.
ZhangYet
force-pushed
the
dante-dev
branch
3 times, most recently
from
e44323a
to
9812c9b
Compare
libbpf-tools/softirqs.bpf.c
Outdated
| u64 delta, *tsp; | ||
| u32 key = 0; | ||
|
|
||
| if (vec_nr >= NR_SOFTIRQS) | ||
| return 0; | ||
| vec_nr &= 0xf; // to avoid compiler opti make verifier error |
There was a problem hiding this comment.
Please give details about the above verifier failure (kernel version, compiler version, error message, etc.)
ZhangYet
force-pushed
the
dante-dev
branch
2 times, most recently
from
cfe636f
to
4159296
Compare
tools/softirqs.py supports `-c/--cpu` otption to filter cpu, It's reasonable to support a same otption for libbpf-tools/softirqs. Add `-mcpu=v3` in Makefile for softirqs or the shift op will lead to verifier errors. Reported-by: Tang Yizhou <[email protected]> Signed-off-by: Dantezy <[email protected]>
Signed-off-by: Dantezy <[email protected]>
Sorry for the format issues. I kept switching development env between office and home, sometimes the fix was forget to push(like the space between Maybe we should provide a tool like |
Since tools/hardirqs and tools/softirqs supports a
--cpuoption, it's reasonable to support a same option in libbpf version.