On 2024/12/06 17:59, lijiang wrote: > Hi, Kazu > Thank you for the fix. > On Thu, Dec 5, 2024 at 5:10 PM <devel-request(a)lists.crash-utility.osci.io <mailto:devel-request@lists.crash-utility.osci.io>> wrote: > > Date: Thu, 5 Dec 2024 01:30:34 +0000 > From: HAGIO KAZUHITO(萩尾　一仁) <k-hagio-ab(a)nec.com <mailto:k-hagio-ab@nec.com>> > Subject: [Crash-utility] [PATCH] Fix infinite loop during module > symbols initialization > To: &quot;devel(a)lists.crash-utility.osci.io <mailto:devel@lists.crash-utility.osci.io>" > <devel(a)lists.crash-utility.osci.io <mailto:devel@lists.crash-utility.osci.io>> > Message-ID: <1733362232-30303-1-git-send-email-k-hagio-ab(a)nec.com <mailto:1733362232-30303-1-git-send-email-k-hagio-ab@nec.com>> > Content-Type: text/plain; charset="iso-2022-jp" > > From: Kazuhito Hagio <k-hagio-ab(a)nec.com <mailto:k-hagio-ab@nec.com>> > > An infinite loop occured in mod_symname_hash_install() during module > symbols initialization on 6.13-rc1 kernel. > > Not sure which kernel patch caused this (probably execmem patches), a > gap region between a module text and another module's one has gone, i.e. > they became consecutive. In this case, as we set a module's pseudo end > symbol to "base + size" value, it's equal to another module's pseudo > start symbol. This and qsort() for st->ext_module_symbtable cause > region overlaps like this: > > sp sp->value sp->name > 823ac68 ffffffffc0800000 _MODULE_TEXT_START_dm_mod > ... > 8242080 ffffffffc0814ad0 __pfx_dm_devt_from_path > 82420a8 ffffffffc0814ae0 dm_devt_from_path > 82420d0 ffffffffc0815000 _MODULE_TEXT_START_libcrc32c << same value > 82420f8 ffffffffc0815000 __pfx_crc32c << and > 8242120 ffffffffc0815000 _MODULE_TEXT_END_dm_mod << overlap > > > Strange, this looks like a kernel module issue? I don't think so, as written above, it should be due to the change of memory placement in kernel and crash's current implementation. > > 8242148 ffffffffc0815010 crc32c > 8242170 ffffffffc0815080 __pfx_libcrc32c_mod_fini > ... > 8242238 ffffffffc0816000 _MODULE_TEXT_END_libcrc32c > > As a result, mod_symtable_hash_install_range() can add a module symbol > two times, which makes a self-loop. This causes an infinite loop in > mod_symname_hash_install(). > > Core was generated by `crash'. > #0 mod_symname_hash_install (spn=0x75e6bc0) at symbols.c:1236 > 1236 if (!sp->name_hash_next || > (gdb) p sp

> $1 = (struct syment *) 0x75e6670 > (gdb) p sp->name > $2 = 0x7c03573 "_MODULE_TEXT_START_libcrc32c" > (gdb) p sp->name_hash_next > $3 = (struct syment *) 0x75e6670 << self-loop > > To avoid this, fix module pseudo end symbols to "base + size - 1". > > So far I haven't thought of a better solution yet. For the patch: Ack. Thank you for your review. Thanks, Kazu > > Thanks > Lianbo > > Signed-off-by: Kazuhito Hagio <k-hagio-ab(a)nec.com <mailto:k-hagio-ab@nec.com>> > --- > symbols.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/symbols.c b/symbols.c > index a26bdc1a3461..0fe5d2951c7e 100644 > --- a/symbols.c > +++ b/symbols.c > @@ -2196,7 +2196,7 @@ store_module_symbols_6_4(ulong total, int mods_installed) > if (!lm->mem[t].size) > continue; > > - st->ext_module_symtable[mcnt].value = lm->mem[t].base + lm->mem[t].size; > + st->ext_module_symtable[mcnt].value = lm->mem[t].base + lm->mem[t].size - 1; > > st->ext_module_symtable[mcnt].type = 'm'; > st->ext_module_symtable[mcnt].flags |= MODULE_SYMBOL; > sprintf(buf2, "%s%s", module_tag[t].end, mod_name); > -- > 2.31.1 > -- Crash-utility mailing list -- devel(a)lists.crash-utility.osci.io To unsubscribe send an email to devel-leave(a)lists.crash-utility.osci.io https://${domain_name}/admin/lists/devel.lists.crash-utility.osci.io/ Contribution Guidelines: https://github.com/crash-utility/crash/wiki

Hi Tao, On 2024/12/10 12:39, Tao Liu wrote: > I can also reproduce this issue on 6.13.0-rc2 machine. The > mod_symname_hash_install() function missed the case if one spn was > installed 2 more times, which as a result, will cause the infinite > loop. > > The reason for installing 2 more times is because the addresses > overlap between 2 kernel modules. See the following: > > crash> sym -l > ffffffffc0800000 MODULE TEXT START: qemu_fw_cfg > ffffffffc0800000 (t) __pfx_fw_cfg_sysfs_attr_show > ffffffffc0800010 (t) fw_cfg_sysfs_attr_show > ... > ffffffffc0801000 MODULE TEXT START: serio_raw > ffffffffc0801000 (t) __pfx_serio_raw_poll > ffffffffc0801000 MODULE TEXT END: qemu_fw_cfg > ... > ffffffffc0802000 MODULE TEXT END: serio_raw > > For module qemu_fw_cfg, 3 symbols of ffffffffc0801000 will be > installed. Then for module serio_raw, 3 symbols of ffffffffc0801000 > will be reinstalled. Due to the lack of repeat symbol check, there > will be symbol reinstall, and infinite loop: > > sp->name_hash_next = spn; <<----- name_hash_next point to itself if sp==spn > > I think Kazu's "base + size - 1" solution is reasonable, this will let > qemu_fw_cfg install range from ffffffffc0800000 to > (ffffffffc0801000-1), so no overlap for 2 modules. In addition, I'd > like to add the following code as well, to skip if a symbol has > already been installed: > > diff --git a/symbols.c b/symbols.c > index a26bdc1..e1a62bb 100644 > --- a/symbols.c > +++ b/symbols.c > @@ -1233,6 +1233,9 @@ mod_symname_hash_install(struct syment *spn) > return; > } > for (; sp; sp = sp->name_hash_next) { > + if (spn == sp) > + return; > + > if (!sp->name_hash_next || > spn->value < sp->name_hash_next->value) { > spn->name_hash_next = sp->name_hash_next; > > What do you think? thank you for checking the issue, it's ok to add it for future safety.

Thanks, Kazu