| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <190052a4.36d4.1979ac95438.Coremail.00107082@163.com>
Date: Mon, 23 Jun 2025 11:16:15 +0800 (CST)
From: "David Wang" <00107082@....com>
To: "Suren Baghdasaryan" <surenb@...gle.com>
Cc: oliver.sang@...el.com, urezki@...il.com, ahuang12@...ovo.com,
akpm@...ux-foundation.org, bhe@...hat.com, hch@...radead.org,
linux-kernel@...r.kernel.org, linux-mm@...ck.org, lkp@...el.com,
mjguzik@...il.com, oe-lkp@...ts.linux.dev, harry.yoo@...cle.com,
kent.overstreet@...ux.dev
Subject: Re: CONFIG_TEST_VMALLOC=y conflict/race with alloc_tag_init
At 2025-06-23 10:45:31, "David Wang" <00107082@....com> wrote:
>
>At 2025-06-23 06:50:44, "Suren Baghdasaryan" <surenb@...gle.com> wrote:
>>On Fri, Jun 20, 2025 at 3:03 AM David Wang <00107082@....com> wrote:
>>>
>>> On Wed, Jun 18, 2025 at 02:25:37PM +0800, kernel test robot wrote:
>>> >
>>> > Hello,
>>> >
>>> > for this change, we reported
>>> > "[linux-next:master] [lib/test_vmalloc.c] 7fc85b92db: Mem-Info"
>>> > in
>>> > https://lore.kernel.org/all/202505071555.e757f1e0-lkp@intel.com/
>>> >
>>> > at that time, we made some tests with x86_64 config which runs well.
>>> >
>>> > now we noticed the commit is in mainline now.
>>>
>>> > the config still has expected diff with parent:
>>> >
>>> > --- /pkg/linux/x86_64-randconfig-161-20250614/gcc-12/7a73348e5d4715b5565a53f21c01ea7b54e46cbd/.config 2025-06-17 14:40:29.481052101 +0800
>>> > +++ /pkg/linux/x86_64-randconfig-161-20250614/gcc-12/2d76e79315e403aab595d4c8830b7a46c19f0f3b/.config 2025-06-17 14:41:18.448543738 +0800
>>> > @@ -7551,7 +7551,7 @@ CONFIG_TEST_IDA=m
>>> > CONFIG_TEST_MISC_MINOR=m
>>> > # CONFIG_TEST_LKM is not set
>>> > CONFIG_TEST_BITOPS=m
>>> > -CONFIG_TEST_VMALLOC=m
>>> > +CONFIG_TEST_VMALLOC=y
>>> > # CONFIG_TEST_BPF is not set
>>> > CONFIG_FIND_BIT_BENCHMARK=m
>>> > # CONFIG_TEST_FIRMWARE is not set
>>> >
>>> >
>>> > then we noticed similar random issue with x86_64 randconfig this time.
>>> >
>>> > 7a73348e5d4715b5 2d76e79315e403aab595d4c8830
>>> > ---------------- ---------------------------
>>> > fail:runs %reproduction fail:runs
>>> > | | |
>>> > :199 34% 67:200 dmesg.KASAN:null-ptr-deref_in_range[#-#]
>>> > :199 34% 67:200 dmesg.Kernel_panic-not_syncing:Fatal_exception
>>> > :199 34% 67:200 dmesg.Mem-Info
>>> > :199 34% 67:200 dmesg.Oops:general_protection_fault,probably_for_non-canonical_address#:#[##]SMP_KASAN
>>> > :199 34% 67:200 dmesg.RIP:down_read_trylock
>>> >
>>> > we don't have enough knowledge to understand the relationship between code
>>> > change and the random issues. just report what we obsverved in our tests FYI.
>>> >
>>>
>>> I think this is caused by a race between vmalloc_test_init and alloc_tag_init.
>>>
>>> vmalloc_test actually depends on alloc_tag via alloc_tag_top_users, because when
>>> memory allocation fails show_mem() would invoke alloc_tag_top_users.
>>>
>>> With following configuration:
>>>
>>> CONFIG_TEST_VMALLOC=y
>>> CONFIG_MEM_ALLOC_PROFILING=y
>>> CONFIG_MEM_ALLOC_PROFILING_ENABLED_BY_DEFAULT=y
>>> CONFIG_MEM_ALLOC_PROFILING_DEBUG=y
>>>
>>> If vmalloc_test_init starts before alloc_tag_init, show_mem() would cause
>>> a NULL deference because alloc_tag_cttype was not init yet.
>>>
>>> I add some debug to confirm this theory
>>> diff --git a/lib/alloc_tag.c b/lib/alloc_tag.c
>>> index d48b80f3f007..9b8e7501010f 100644
>>> --- a/lib/alloc_tag.c
>>> +++ b/lib/alloc_tag.c
>>> @@ -133,6 +133,8 @@ size_t alloc_tag_top_users(struct codetag_bytes *tags, size_t count, bool can_sl
>>> struct codetag *ct;
>>> struct codetag_bytes n;
>>> unsigned int i, nr = 0;
>>> + pr_info("memory profiling alloc top %d: %llx\n", mem_profiling_support, (long long)alloc_tag_cttype);
>>> + return 0;
>>>
>>> if (can_sleep)
>>> codetag_lock_module_list(alloc_tag_cttype, true);
>>> @@ -831,6 +833,7 @@ static int __init alloc_tag_init(void)
>>> shutdown_mem_profiling(true);
>>> return PTR_ERR(alloc_tag_cttype);
>>> }
>>> + pr_info("memory profiling ready %d: %llx\n", mem_profiling_support, (long long)alloc_tag_cttype);
>>>
>>> return 0;
>>> }
>>>
>>> When bootup the kernel, the log shows:
>>>
>>> $ sudo dmesg -T | grep profiling
>>> [Fri Jun 20 17:29:35 2025] memory profiling alloc top 1: 0 <--- alloc_tag_cttype == NULL
>>> [Fri Jun 20 17:30:24 2025] memory profiling ready 1: ffff9b1641aa06c0
>>>
>>>
>>> vmalloc_test_init should happened after alloc_tag_init if CONFIG_TEST_VMALLOC=y,
>>> or mem_show() should check whether alloc_tag is done initialized when calling
>>> alloc_tag_top_users
>>
>>Thanks for reporting!
>>So, IIUC https://lore.kernel.org/all/20250620195305.1115151-1-harry.yoo@oracle.com/
>>will address this issue as well. Is that correct?
>
>Yes, the panic can be fix by that patch.
>
>I still feel it better to delay vmalloc_test_init, make it happen after alloc_tag_init.
>Or, maybe we can promote alloc_tag_init to some early init? I remember reporting some allocation
>not registered by memory profiling during boot,
>https://lore.kernel.org/all/213ff7d2.7c6c.1945eb0c2ff.Coremail.00107082@163.com/
>
>I will make some tests, and update later
The memory allocations in sched_init_domains happened quite early, maybe it is core_initcall, while
alloc_tag_init needs rootfs, it needs to be after rootfs_initcall, so no reasonable place to promote.......
But I think this explain why some allocation counter missed during boot: the allocation happened before alloc_tag_init
Thanks
David
>
>
>David
>
>
>>
>>>
>>>
>>>
>>> David
>>>
Powered by blists - more mailing lists