[PATCH 3/5] zfs: Fix unaligned read of uint64

[mwleeds at mailtundra.com]

On Thursday, April 25th, 2024 at 4:57 AM, Caleb Connolly <caleb.connolly at linaro.org> wrote:

>
> On 25/04/2024 06:02, mwleeds at mailtundra.com wrote:
>
> > Hi Caleb,
> >
> > Thanks for this interesting feedback. I saw these patches were already merged
> > since you sent this but I added a few thoughts below anyway.
>
>
> Ah, a shame.. It would be good to find a scalable solution to this!

While I agree your suggestion of doing two 4 byte reads is more scalable than
the malloc() patch, the code path (zfs_nvlist_lookup_uint64) is only hit when
reading the ZPool label (check_pool_label) which is a once-per-boot operation
operating on some metadata, so I don't think scalability is as much a concern
as it would be if we were hitting this code path when reading every block from
the disk for example.

That said, I may try implementing the patch as you suggested if I get a chance.
And in any case I definitely appreciate your engagement and feedback!

>
> > On Friday, April 12th, 2024 at 11:50 AM, Caleb Connolly caleb.connolly at linaro.org wrote:
> >
> > > Hi Phaedrus,
> > >
> > > On 07/04/2024 03:47, mwleeds at mailtundra.com wrote:
> > >
> > > > Without this patch, when trying to boot zfs using U-Boot on a Jetson TX2
> > > > NX (which is aarch64), I get a CPU reset error like so:
> > > >
> > > > "Synchronous Abort" handler, esr 0x96000021
> > > > elr: 00000000800c9000 lr : 00000000800c8ffc (reloc)
> > > > elr: 00000000fff77000 lr : 00000000fff76ffc
> > > > x0 : 00000000ffb40f04 x1 : 0000000000000000
> > > > x2 : 000000000000000a x3 : 0000000003100000
> > > > x4 : 0000000003100000 x5 : 0000000000000034
> > > > x6 : 00000000fff9cc6e x7 : 000000000000000f
> > > > x8 : 00000000ff7f84a0 x9 : 0000000000000008
> > > > x10: 00000000ffb40f04 x11: 0000000000000006
> > > > x12: 000000000001869f x13: 0000000000000001
> > > > x14: 00000000ff7f84bc x15: 0000000000000010
> > > > x16: 0000000000002080 x17: 00000000001fffff
> > > > x18: 00000000ff7fbdd8 x19: 00000000ffb405f8
> > > > x20: 00000000ffb40dd0 x21: 00000000fffabe5e
> > > > x22: 000000ea77940000 x23: 00000000ffb42090
> > > > x24: 0000000000000000 x25: 0000000000000000
> > > > x26: 0000000000000000 x27: 0000000000000000
> > > > x28: 0000000000bab10c x29: 00000000ff7f85f0
> > > >
> > > > Code: d00001a0 9103a000 94006ac6 f9401ba0 (f9400000)
> > > > Resetting CPU ...
> > > >
> > > > This happens when be64_to_cpu() is called on a value that exists at a
> > > > memory address that's 4 byte aligned but not 8 byte aligned (e.g. an
> > > > address ending in 04). The call stack where that happens is:
> > > > check_pool_label() ->
> > > > zfs_nvlist_lookup_uint64(vdevnvlist, ZPOOL_CONFIG_ASHIFT,...) ->
> > > > be64_to_cpu()
> > >
> > > This is very odd, aarch64 doesn't generally have these restrictions. I
> > > got a bit nerdsniped when I saw this so I did some digging and figured this:
> > >
> > > 1. Your abort exception doesn't include the FAR_ELx register (which
> > > should contain the address that was being accessed when the abort
> > > occured). This means your board is running in EL3.
> > > 2. It turns out there is an "A" flag in the SCTLR_ELx register, when set
> > > this flag causes a fault when trying to load from an address that isn't
> > > aligned to the size of the data element (see "A, bit" on
> > > https://developer.arm.com/documentation/ddi0595/2021-06/AArch64-Registers/SCTLR-EL3--System-Control-Register--EL3-
> > >
> > > I'm not sure who's in the "wrong" here, maybe the driver should avoid
> > > unaligned accesses? But then again, I don't think you should be running
> > > a ZFS driver in EL3.
> > >
> > > I'm not familiar with the Jetson Nano, but maybe there's a documented
> > > way to run U-Boot so that it isn't executing in EL3? Or if not you could
> > > also try unsetting the A flag.
> >
> > I may look into this if I get a chance. However if I write some assembly code
> > to change the execution level or unset the A flag, I worry that the code would
> > work fine on the hardware I have (Jetson TX2 NX) but behave differently on
>
>
> Well, unsetting the A flag might arguably have some niche negative
> outcomes, but I really struggle to see how this is a driver bug to be
> honest...
>
> > another platform. And I don't think I can easily set up testing environments
> > with u-boot + zfs on different platforms to find out.
>
>
> fwiw, if you do intend to investigate this further, I'd be happy to
> validate your assumptions on a Qualcomm board (where we are rather
> boring and running in EL1 or EL2).

Thanks, I'll keep that in mind

>
> > > If this really is something to fix in the driver, I don't think
> > > hotpatching every unaligned access with a malloc() is the right solution.
> >
> > I'm certainly open to other ideas. The difficulty is the data structure we're
> > parsing in this file is read from disk and it's only 4-byte aligned.
>
>
> According to the ARM docs, the issue is that you're loading an 8-byte
> value from a 4-byte aligned address. If you split it into two 4-byte
> reads (for the upper and lower words) then it should work fine. This
> would avoid the malloc at any rate.

Makes sense

>
> Thanks and regards,
>
> > > > Signed-off-by: Phaedrus Leeds mwleeds at mailtundra.com
> > > > Tested-by: Phaedrus Leeds mwleeds at mailtundra.com
> > >
> > > regarding your question about re-sending to remove these tags, I'd say
> > > probably yes, and especially if you're going to send a new revision anyway.
> > >
> > > fwiw you seem to have gotten pretty much everything else about the patch
> > > submission process spot on :)
> > >
> > > Kind regards,
> > >
> > > > ---
> > > > fs/zfs/zfs.c | 18 ++++++++++++++++++
> > > > 1 file changed, 18 insertions(+)
> > > >
> > > > diff --git a/fs/zfs/zfs.c b/fs/zfs/zfs.c
> > > > index 61d58fce68..9a50deac18 100644
> > > > --- a/fs/zfs/zfs.c
> > > > +++ b/fs/zfs/zfs.c
> > > > @@ -1552,35 +1552,53 @@ nvlist_find_value(char *nvlist, char *name, int valtype, char **val,
> > > > if (nelm_out)
> > > > *nelm_out = nelm;
> > > > return 1;
> > > > }
> > > >
> > > > nvlist += encode_size; /* goto the next nvpair */
> > > > }
> > > > return 0;
> > > > }
> > > >
> > > > +int is_word_aligned_ptr(void *ptr) {
> > > > + return ((uintptr_t)ptr & (sizeof(void *) - 1)) == 0;
> > > > +}
> > > > +
> > > > int
> > > > zfs_nvlist_lookup_uint64(char *nvlist, char *name, uint64_t *out)
> > > > {
> > > > char *nvpair;
> > > > size_t size;
> > > > int found;
> > > >
> > > > found = nvlist_find_value(nvlist, name, DATA_TYPE_UINT64, &nvpair, &size, 0);
> > > > if (!found)
> > > > return 0;
> > > > if (size < sizeof(uint64_t)) {
> > > > printf("invalid uint64\n");
> > > > return ZFS_ERR_BAD_FS;
> > > > }
> > > >
> > > > + /* On arm64, calling be64_to_cpu() on a value stored at a memory address
> > > > + * that's not 8-byte aligned causes the CPU to reset. Avoid that by copying the
> > > > + * value somewhere else if needed.
> > > > + */
> > > > + if (!is_word_aligned_ptr((void *)nvpair)) {
> > > > + uint64_t *alignedptr = malloc(sizeof(uint64_t));
> > > > + if (!alignedptr)
> > > > + return 0;
> > > > + memcpy(alignedptr, nvpair, sizeof(uint64_t));
> > > > + *out = be64_to_cpu(*alignedptr);
> > > > + free(alignedptr);
> > > > + return 1;
> > > > + }
> > > > +
> > > > out = be64_to_cpu((uint64_t *) nvpair);
> > > > return 1;
> > > > }
> > > >
> > > > char *
> > > > zfs_nvlist_lookup_string(char *nvlist, char *name)
> > > > {
> > > > char *nvpair;
> > > > char *ret;
> > > > size_t slen;
> > >
> > > --
> > > // Caleb (they/them)
>
>
> --
> // Caleb (they/them)