Aarch64 EXT4FS inode checksum failures - seems to be weak memory ordering issues

[Eric Biggers]

On Thu, Jan 07, 2021 at 10:14:46PM +0000, Russell King - ARM Linux admin wrote:
> On Thu, Jan 07, 2021 at 10:48:05PM +0100, Arnd Bergmann wrote:
> > On Thu, Jan 7, 2021 at 5:27 PM Theodore Ts'o <tytso at mit.edu> wrote:
> > >
> > > On Thu, Jan 07, 2021 at 01:37:47PM +0000, Russell King - ARM Linux admin wrote:
> > > > > The gcc bugzilla mentions backports into gcc-linaro, but I do not see
> > > > > them in my git history.
> > > >
> > > > So, do we raise the minimum gcc version for the kernel as a whole to 5.1
> > > > or just for aarch64?
> > >
> > > Russell, Arnd, thanks so much for tracking down the root cause of the
> > > bug!
> > 
> > There is one more thing that I wondered about when looking through
> > the ext4 code: Should it just call the crc32c_le() function directly
> > instead of going through the crypto layer? It seems that with Ard's
> > rework from 2018, that can just call the underlying architecture specific
> > implementation anyway.
> 
> Yes, I've been wondering about that too. To me, it looks like the
> ext4 code performs a layering violation by going "under the covers"
> - there are accessor functions to set the CRC and retrieve it. ext4
> instead just makes the assumption that the CRC value is stored after
> struct shash_desc. Especially as the crypto/crc32c code references
> the value using:
> 
> 	struct chksum_desc_ctx *ctx = shash_desc_ctx(desc);
> 
> Not even crypto drivers are allowed to assume that desc+1 is where
> the CRC is stored.

It violates how the shash API is meant to be used in general, but there is a
test that enforces that the shash_desc_ctx for crc32c must be just the single
u32 crc value.  See alg_test_crc32c() in crypto/testmgr.c.  So it's apparently
intended to work.

> 
> However, struct shash_desc is already 128 bytes in size on aarch64,

Ard Biesheuvel recently sent a patch to reduce the alignment of struct
shash_desc to ARCH_SLAB_MINALIGN
(https://lkml.kernel.org/linux-crypto/20210107124128.19791-1-ardb@kernel.org/),
since apparently most of the bloat is from alignment for DMA, which isn't
necessary.  I think that reduces the size by a lot on arm64.

> and the proper way of doing it via SHASH_DESC_ON_STACK() is overkill,
> being strangely 2 * sizeof(struct shash_desc) + 360 (which looks like
> another bug to me!)

Are you referring to the '2 * sizeof(struct shash_desc)' rather than just
'sizeof(struct shash_desc)'?  As mentioned in the comment above
HASH_MAX_DESCSIZE, there can be a nested shash_desc due to HMAC.
So I believe the value is correct.

> So, I agree with you wrt crc32c_le(), especially as it would be more
> efficient, and as the use of crc32c is already hard coded in the ext4
> code - not only with crypto_alloc_shash("crc32c", 0, 0) but also with
> the fixed-size structure in ext4_chksum().
> 
> However, it's ultimately up to the ext4 maintainers to decide.

As I mentioned in my other response, crc32c_le() isn't a proper library API
(like some of the newer lib/crypto/ stuff) but rather just a wrapper for the
shash API, and it doesn't handle modules being dynamically loaded/unloaded.
So switching to it may cause a performance regression.

What I'd recommend is making crc32c_le() able to call architecture-speccific
implementations directly, similar to blake2s() and chacha20() in lib/crypto/.
Then there would be no concern about when modules get loaded, etc...

- Eric