changed section .rela.smp_locks not selected for inclusion about kpatch HOT 8 CLOSED

I also met this problem, but the problematic section kpatch_verify_patchability complained is '__bug_table', rather than '.rela.smp_locks'.

[ryan@localhost kpatch]$ ./kpatch-build/kpatch-build -r ./kernel-3.13.6-200.fc20.src.rpm ~/workspace/mlock.patch -d
DEBUG mode enabled
Using cache at /home/ryan/.kpatch/3.13.6-200.fc20.x86_64/src
Testing patch file
patching file mm/mlock.c
Building original kernel
Building patched kernel
Detecting changed objects
Rebuilding changed objects
Extracting new and modified ELF sections
changed function: mlock_vma_page
kpatch_verify_patchability:730: changed section __bug_table not selected for inclusion
/home/ryan/workspace/upstream/kpatch/kpatch-build/create-diff-object: unreconcilable difference
ERROR: kpatch build failed. Check /tmp/kpatch-build-1396494986.log for more details.

from kpatch.

I know of the __bug_table issue and am working to fix it

from kpatch.

I had compared two mlock.o:
[ryan@localhost kpatch]$ readelf -S /tmp/tmp.0TM37HJNqd/orig/mm/mlock.o
...
[10] __bug_table PROGBITS 0000000000000000 00000197
00000000000000c0 0000000000000000 A 0 0 1
...
[16] .rela.smp_locks RELA 0000000000000000 00005730
0000000000000108 0000000000000018 97 15 8
...

[ryan@localhost kpatch]$ readelf -S /tmp/tmp.0TM37HJNqd/patched/mm/mlock.o
...
[10] __bug_table PROGBITS 0000000000000000 00000197
00000000000000b4 0000000000000000 A 0 0 1
...
[16] .rela.smp_locks RELA 0000000000000000 000056e0
0000000000000108 0000000000000018 97 15 8
...

the size of the __bug_table section is changed after patching.
but for .rela.smp_locks section, the size is same.

from kpatch.

Saw this again with the following patch:

Index: src/net/unix/af_unix.c
===================================================================
--- src.orig/net/unix/af_unix.c
+++ src/net/unix/af_unix.c
@@ -1921,6 +1921,8 @@ static int unix_stream_recvmsg(struct ki
    long timeo;
    int skip;

+   printk_once("foo\n");
+
    err = -EINVAL;
    if (sk->sk_state != TCP_ESTABLISHED)
        goto out;

$ kpatch-build unix_stream_recvmsg.patch 
Using cache at /home/jpoimboe/.kpatch/3.13.10-200.fc20.x86_64/src
Testing patch file
patching file net/unix/af_unix.c
Building original kernel
Building patched kernel
Detecting changed objects
Rebuilding changed objects
Extracting new and modified ELF sections
changed function: unix_stream_recvmsg
kpatch_verify_patchability:736: changed section .rela.smp_locks not selected for inclusion
/usr/local/libexec/kpatch/create-diff-object: unreconcilable difference
ERROR: kpatch build failed. Check /tmp/kpatch-build-1398089722.log for more details.

from kpatch.

Each use of an atomic variable or spinlock (any instruction using the x86 "lock" prefix) results in the instruction's address getting added to the .smp_lock section, for potential future code modification (removing or adding the x86 "lock" instruction prefix). Each entry in the section contains only a 4-byte address which is placed there via relocation (.rela.smp.locks).

To support this I think we need to do something similar to what was done with the bug table, except both the .smp_locks and its rela section will need to be regenerated. Only those entries for which the rela refers to an included function should be kept. Regenerating .smp_locks shouldn't be hard since it's just all zeros.

Since this looks like it could be a common pattern emerging, maybe we can replace kpatch_regenerate_bug_table_rela_section with something more generic that covers both cases (and probably more cases we will discover). I think as long as it knows the entry size, that should be enough information for it to examine each of the relas within each entry and copy the needed base and rela entries.

from kpatch.

This is next in my queue. Thanks for the analysis @jpoimboe

from kpatch.

@madper after this fix, i was able to build your patch. Please confirm.

@jpoimboe with your patch, i feel through to a .rela.init.text not included error. I think it is because of the BUILD_BUG_ON in the __init function. however, it does fix the smp_locks failure.

from kpatch.

@Spartacus06 Thanks, when I get a chance I'll take a closer look at the .rela.init.text error. I'm going to go ahead and merge this fix now. @madper , please open a new issue if you still see the same error.

from kpatch.