Re: pahole 1.25 SIGSEGV when building kernel (-j when cores > 81)

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Wed, Jul 05, 2023 at 05:00:40AM +0300, Vitaly Chikunov escreveu:
> Hi,
> 
> After updating Dwarves to v1.25 we are getting SIGSEGV while building
> kernel on 32-bit x86 architecture. It looks like this:

128 cores, 32-bit system? Interesting :-)
 
>     BTF     .btf.vmlinux.bin.o
>   scripts/link-vmlinux.sh: line 111: 395728 Segmentation fault      LLVM_OBJCOPY="${OBJCOPY}" ${PAHOLE} -J ${PAHOLE_FLAGS} ${1}
>     LD      .tmp_vmlinux.kallsyms1
>     NM      .tmp_vmlinux.kallsyms1.syms
>     KSYMS   .tmp_vmlinux.kallsyms1.S
>     AS      .tmp_vmlinux.kallsyms1.S
>     LD      .tmp_vmlinux.kallsyms2
>     NM      .tmp_vmlinux.kallsyms2.syms
>     KSYMS   .tmp_vmlinux.kallsyms2.S
>     AS      .tmp_vmlinux.kallsyms2.S
>     LD      vmlinux
>     BTFIDS  vmlinux
>   libbpf: failed to find '.BTF' ELF section in vmlinux
>   FAILED: load BTF from vmlinux: No data available
> 
> What crashes is this command:
> 
>   pahole -J --btf_gen_floats -j --lang_exclude=rust .tmp_vmlinux.btf
> 
> I found that cause of the crash is that build box having 128 cores. By
> experiment I found that with -j81 pahole works OK, but with -j82 or
> greater it crashes.
> 
>   $ gdb -q --args pahole -J --btf_gen_floats -j111 --lang_exclude=rust .tmp_vmlinux.btf
>   Thread 15 "pahole" received signal SIGSEGV, Segmentation fault.
>   [Switching to Thread 0xda5ffb40 (LWP 3102466)]
>   0xf7f3c944 in btf_encoder__btf (encoder=0x0) at /usr/src/debug/dwarves-1.25/btf_encoder.c:1890
>   1890            return encoder->btf;
>   (gdb) bt
>   #0  0xf7f3c944 in btf_encoder__btf (encoder=0x0) at /usr/src/debug/dwarves-1.25/btf_encoder.c:1890
>   #1  0x5655c25d in pahole_stealer (cu=0xd9a01f80, conf_load=0x565640c0 <conf_load>, thr_data=0x56567c18) at /usr/src/debug/dwarves-1.25/pahole.c:3100
>   #2  0xf7f452d7 in cu__finalize (cu=cu@entry=0xd9a01f80, conf=0x565640c0 <conf_load>, thr_data=thr_data@entry=0x56567c18)
>       at /usr/src/debug/dwarves-1.25/dwarf_loader.c:3001
>   #3  0xf7f4541d in cus__finalize (thr_data=0x56567c18, conf=<optimized out>, cu=0xd9a01f80, cus=0x565651c0) at /usr/src/debug/dwarves-1.25/dwarf_loader.c:3008
>   #4  dwarf_cus__create_and_process_cu (dcus=dcus@entry=0xffffd19c, cu_die=cu_die@entry=0xda5ff38c, pointer_size=<optimized out>, thr_data=0x56567c18)
>       at /usr/src/debug/dwarves-1.25/dwarf_loader.c:3207
>   #5  0xf7f461af in dwarf_cus__process_cu_thread (arg=0xffffcbf8) at /usr/src/debug/dwarves-1.25/dwarf_loader.c:3250
>   #6  0xf7db4258 in start_thread (arg=<optimized out>) at pthread_create.c:444
>   #7  0xf7e3a878 in clone3 () from /lib/libc.so.6
>   (gdb) p encoder
>   $1 = (struct btf_encoder *) 0x0
>   (gdb) f 1
>   #1  0x5655c25d in pahole_stealer (cu=0xd9a01f80, conf_load=0x565640c0 <conf_load>, thr_data=0x56567c18) at /usr/src/debug/dwarves-1.25/pahole.c:3100
>   3100                                    thread->btf = btf_encoder__btf(thread->encoder);
>   (gdb) list -2
>   3093                                    thread->encoder =
>   3094                                            btf_encoder__new(cu, detached_btf_filename,
>   3095                                                             NULL,
>   3096                                                             skip_encoding_btf_vars,
>   3097                                                             btf_encode_force,
>   3098                                                             btf_gen_floats,
>   3099                                                             global_verbose);
>   3100                                    thread->btf = btf_encoder__btf(thread->encoder);
>   3101                            }
>   3102                            encoder = thread->encoder;
> 
> I think that return value of btf_encoder__new is not checked. But did
> not investigate further why is this happening. It would be great to have
> this fixed.

Can you try with this minimal patch? Maybe we can with some more work
auto-limit the number of threads or make the threads without a
btf_encoder be processed at the end, reusing the main btf_encoder, but
this first patch would at least help diagnosing the problem more
quickly.

- Arnaldo

diff --git a/pahole.c b/pahole.c
index e843999fde2a8a37..887b3403dcff18c0 100644
--- a/pahole.c
+++ b/pahole.c
@@ -3097,6 +3097,15 @@ static enum load_steal_kind pahole_stealer(struct cu *cu,
 							 btf_encode_force,
 							 btf_gen_floats,
 							 global_verbose);
+				if (thread->encoder == NULL) {
+					fprintf(stderr, "Not enough memory to instantiate new BTF encoder.\n");
+
+					if (conf_load->nr_jobs > 1)
+						fprintf(stderr, "HINT: Maybe reduce the -j/--jobs command line value?.\n");
+
+					ret = LSK__STOP_LOADING;
+					goto out_btf;
+				}
 				thread->btf = btf_encoder__btf(thread->encoder);
 			}
 			encoder = thread->encoder;