Re: Regression from a9b3c355c2e6 ("asm-generic: pgalloc: provide generic __pgd_{alloc,free}") with CONFIG_DEBUG_VM_PGFLAGS=y and Xen

[Jürgen Groß]

On 16.04.25 14:07, Petr Vaněk wrote:
> Hi all,
>
> I have discovered a regression introduced in commit a9b3c355c2e6
> ("asm-generic: pgalloc: provide generic __pgd_{alloc,free}") [1,2] in
> kernel version 6.14. The problem occurs when the x86 kernel is
> configured with CONFIG_DEBUG_VM_PGFLAGS=y and is run as a PV Dom0 in Xen
> 4.19.1. During the startup, the kernel panics with the error log below.
>
> The commit changed PGD allocation path.  In the new implementation
> _pgd_alloc allocates memory with __pgd_alloc, which indirectly calls
>
>    alloc_pages_noprof(gfp | __GFP_COMP, order);
>
> This is in contrast to the old behavior, where __get_free_pages was
> used, which indirectly called
>
>    alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
>
> The key difference is that the new allocator can return a compound page.
> When xen_pin_page is later called on such a page, it call
> TestSetPagePinned function, which internally uses the PF_NO_COMPOUND
> macro. This macro enforces VM_BUG_ON_PGFLAGS if PageCompound is true,
> triggering the panic when CONFIG_DEBUG_VM_PGFLAGS is enabled.
>
> I am reporting this issue without a patch as I am not sure which part of
> the code should be adapted to resolve the regression.

Thanks for the report AND the analysis.

I believe PMD_ALLOCATION_ORDER needs to be changed: in case the system is
running as a Xen PV domain (or with PTI disabled), PMD_ALLOCATION_ORDER
should be 0.

So I'd suggest to switch PGD_ALLOCATION_ORDER to be defined either as 0
(in case PTI is not configured), or pgd_allocation_order (a new global
variable having the value 0 or 1, depending on PTI active or not).

I'll send a patch.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature.asc
Description: OpenPGP digital signature