Re: [RFC]: Maybe revert 'MIPS: add PMD table accounting into MIPS'pmd_alloc_one' from 5.4 LTS?

[Huang Pei <huangpei@loongson.cn>]

On Tue, Jul 20, 2021 at 08:45:45PM -0400, Joshua Kinard wrote:
Hi, 
> (resend using the correct list address)
> 
> Hi all,
> 
> I think there may be an issue with upstream commit
> ed914d48b6a1040d1039d371b56273d422c0081e (MIPS: add PMD table accounting
> into MIPS'pmd_alloc_one), which got backported to 5.4 LTS.
> 
> I am running my SGI Octane on the 5.4 LTS kernel series w/ Gentoo's IP30
> patches, since basic IP30 support didn't get fully mainlined until ~5.7 and
> I am just behind on kernels for the time being.  Generally, the 5.4 series
> is stable on this machine, but as of using 5.4.133 yesterday and 5.4.134
> today, I have had three kernel panics, all with completely different stack
> traces, while building a multilib-enabled glibc-2.33 package.
> 
> Since the stack traces are all completely different, I am suspecting an
> issue with memory allocation, which picking through the 5.4.133 changelog,
> got me to look at commit ed914d48b6a1.  My initial thought is it looks like
> this patch should not have been backported to 5.4, because 5.4 is missing
> the asm-generic version of pmd_free(), which means the MIPS-specific version
> is likely used instead, and that function was not updated alongside
> pmd_alloc_one().
> 
> Now, I am not a memory person, but some quick reading indicates there may be
> some other issues with this patch as well:
> 
> - use of 'GFP_KERNEL | __GFP_ACCOUNT' when there is GFP_KERNEL_ACCOUNT
> available.
Yes, this is better
> 
> - the return value of pgtable_pmd_page_ctor(pg); is not checked to make sure
> it succeeded.  Per Documentation/vm/split_page_table_lock, "NOTE:
> pgtable_page_ctor() and pgtable_pmd_page_ctor() can fail -- it must
> be handled properly."

Yes, I missed this one
> 
> - there is no call to pgtable_pmd_page_dtor() in pmd_free() in MIPS'
> pgalloc.  The commit message indicates that pmd_free in asm-generic will
> handle this, but that function doesn't exist in 5.4, leading to mismatched
> implementations.

Yes, my fix is to fix the mismatch. 

I think all stable branches before 5.10 should *not* apply this my patch
> 
> I think the fix is to either revert ed914d48b6a1 on the 5.4 stable series
> (and any earlier series that are missing asm-generic pmd_free), or the MIPS
> pmd_free() needs to be updated to call pgtable_pmd_page_dtor() in a similar
> manner to x86 and arm64.  I'd opt for reverting.
> 
> Currently testing a build of 5.4.134 with this patch reverted running some
> compile jobs.  If it survives building multilib glibc, then I think my
> suspicions may be valid.
> 
> -- 
> Joshua Kinard
> Gentoo/MIPS
> kumba@gentoo.org
> rsa6144/5C63F4E3F5C6C943 2015-04-27
> 177C 1972 1FB8 F254 BAD0 3E72 5C63 F4E3 F5C6 C943
> 
> "The past tempts us, the present confuses us, the future frightens us.  And
> our lives slip away, moment by moment, lost in that vast, terrible in-between."
> 
> --Emperor Turhan, Centauri Republic