Re: [PATCH] microblaze: restore the page alignment of swapper_pg_dir

[Guenter Roeck]

On 7/27/26 22:28, Michal Simek wrote:
> On 7/27/26 23:58, Ramin Moussavi wrote:
> > microblaze handles TLB misses in software, and the handler builds the
> > address of the L1 entry by ORing the index into the page directory base
> > instead of adding it (hw_exception_handler.S):
> >
> >         bsrli    r5, r3, PGDIR_SHIFT - 2
> >         andi    r5, r5, PAGE_SIZE - 4
> >     /* Assume pgdir aligned on 4K boundary, no need for
> >        "andi r4,r4,0xfffff003" */
> >         or    r4, r4, r5
> >         lwi    r4, r4, 0        /* Get L1 entry */
> >
> > The index is masked to the low 12 bits, so the OR only works if those bits
> > of the base are zero -- which is exactly the assumption the comment states
> > and the reason the masking of the base can be skipped.
> >
> > swapper_pg_dir had no alignment directive of its own.  It was aligned
> > because it followed empty_zero_page in head.S, and that one carried the
> > .align 12:
> >
> >     .section .data
> >     .global empty_zero_page
> >     .align 12
> >     empty_zero_page:
> >         .space    PAGE_SIZE
> >     .global swapper_pg_dir
> >     swapper_pg_dir:
> >         .space    PAGE_SIZE
> >
> > Commit 6215d9f4470f ("arch, mm: consolidate empty_zero_page") removed
> > empty_zero_page from head.S, and with it the .align 12 that -- despite
> > sitting next to empty_zero_page -- was what page aligned swapper_pg_dir.
> > Since then swapper_pg_dir lands wherever .data happens to put it, its low
> > bits are no longer zero, and every kernel TLB miss ORs the index into a
> > base with a nonzero offset.  The resulting L1 lookups read the wrong
> > words, no valid translation is ever installed, and the kernel spins in
> > exceptions long before it can print anything.  On qemu-system-microblazeel
> > (petalogix-s3adsp1800) the console stays completely silent at 100% CPU;
> > there is no oops and no guest error reported by qemu, which makes this
> > awkward to diagnose.
> >
> > Give swapper_pg_dir the alignment it requires, rather than relying on a
> > neighbour to provide it.
> >
> > microblaze has no noMMU variant left in mainline -- CONFIG_MMU is
> > def_bool y and mmu_defconfig is the only defconfig -- so this is not a
> > corner case: every mainline microblaze kernel since v7.1-rc1 fails to
> > boot, including the v7.1 release.
> >
> >     v7.0:                    swapper_pg_dir = 0xc05fd000  (aligned)
> >     v7.1-rc1:                swapper_pg_dir = 0xc0603140  (offset 320)
> >     v7.1-rc1 + this fix:     swapper_pg_dir = 0xc0604000  (aligned)
> >     next-20260726:           swapper_pg_dir = 0xc0615140  (offset 320)
> >     next-20260726 + this fix: swapper_pg_dir = 0xc0616000  (aligned)
>
> hm. This is the first time I get information about this regression.
>
> Guenter: I know you had Microblaze in your testing. Is it still running?

No, sorry. I gave up on it late last year. Too difficult to support - little
if any interest, big endian support was dropped from qemu after v10.0, and
too many "who cares" responses when bringing up issues. With marginal support
in both qemu and the Linux kernel I ended up spending way too much time trying
to figure out if a problem is a kernel problem or a qemu problem.

Guenter