Muchun Song <[email protected]> writes:
> Hi Ritesh,
>
>> On Oct 27, 2026, at 11:22, Ritesh Harjani (IBM) <[email protected]>
>> wrote:
>>
>> radix__vmemmap_populate() is asked for one 16MB section, which is 16KB of
>> struct pages. With no altmap it rounds the request down to a PMD leaf.
>> That leaf describes 2GB of PFNs with 64KB pages (128MB with 4KB pages),
>> so it also covers sections that are not present yet. It cannot be freed
>> while any of those sections are still in use.
>>
>> A devdax namespace that lands in the same window is marked for vmemmap dedup.
>> vmemmap_populate_compound_pages() treated the leaf as an existing huge
>> mapping
>> and skipped it. The core then initializes only the head vmemmap page and
>> assumes
>> the rest is the shared tail. Those remaining tail slots are real struct pages
>> still holding the page_init_poison() pattern and are not the shared tail
>> pages.
>
> Right.
>
>> __init_single_page() warns on the zone mismatch and the following pageblock
>> init
>> faults in __set_pfnblock_flags_mask().
>
> Yeah, the new adding warning is useful.
>
>>
>> If the section's vmemmap is already a PMD leaf, clear its compound order
>
> I think it's enough to fix this issue.
>
Oh yes, right! since the page init is done later and the optimizable
order can be read from mem_section.
>> and reuse the leaf, so every struct page in the section is initialized.
>> The shared-tail layout needs base-page PTEs, so it cannot be built on top of
>> this leaf and dropping the leaf is not an option while the neighbouring range
>> is still using it. So, fail the compound populate if a leaf is still seen
>> there.
>>
>> <Call Trace>
>> ============
>> [ 3.467060][ T1] WARNING: mm/mm_init.c:600 at
>> __init_single_page+0xd0/0xf4, CPU#0: swapper/0/1
>> <...>
>> [ 3.477419][ T1] NIP [c000000000576b30] __init_single_page+0xd0/0xf4
>> [ 3.477827][ T1] LR [c0000000005778b0]
>> __init_zone_device_page+0x30/0xb8
>> [ 3.478444][ T1] Call Trace:
>> [ 3.478764][ T1] 0x10000 (unreliable)
>> [ 3.479874][ T1] memmap_init_zone_device+0x164/0x430
>> [ 3.480392][ T1] memremap_pages+0x43c/0x93c
>> [ 3.480836][ T1] devm_memremap_pages+0x3c/0xa4
>> [ 3.481328][ T1] dev_dax_probe+0x264/0x3d0
>> [ 3.481817][ T1] dax_bus_probe+0x140/0x1c8
>> [ 3.482237][ T1] really_probe+0x108/0x580
>> [ 3.482681][ T1] __driver_probe_device+0xcc/0x284
>> [ 3.483186][ T1] driver_probe_device+0x5c/0x108
>> [ 3.483656][ T1] __device_attach_driver+0x10c/0x1e8
>> [ 3.484134][ T1] bus_for_each_drv+0xc0/0x13c
>> <...>
>> [ 3.499972][ T1] BUG: Kernel NULL pointer dereference on read at
>> 0x00001400
>> [ 3.500417][ T1] Faulting instruction address: 0xc0000000005f4ecc
>> [ 3.500417][ T1] Faulting instruction address: 0xc0000000005f4ecc
>> cpu 0x0: Vector: 300 (Data Access) at [c000000006756b50]
>> pc: c0000000005f4ecc: __set_pfnblock_flags_mask.constprop.0+0x34/0x160
>> lr: c000000000577f38: memmap_init_zone_device+0x358/0x430
>> pid = 1, comm = swapper/0
>> enter ? for help
>> memmap_init_zone_device+0x358/0x430
>> memmap_init_zone_device+0x220/0x430 (unreliable)
>> memremap_pages+0x43c/0x93c
>> devm_memremap_pages+0x3c/0xa4
>> dev_dax_probe+0x264/0x3d0
>> dax_bus_probe+0x140/0x1c8
>> really_probe+0x108/0x580
>> <...>
>>
>> Fixes: bd0862382adb ("mm/sparse-vmemmap: drop Device DAX-specific population
>> path")
>
> Actually, the right Fixes is:
>
> Fixes: f2b79c0d7968 ("powerpc/book3s64/radix: add support for vmemmap
> optimization for radix")
>
> On older kernels, these errors are silent. They trigger a kernel panic when
> the user
> actually uses Device DAX, rather than during the initialization phase.
>
> So, for older kernels, we should guide them to backport the patch.
>
>> Reported-by: Pavithra <[email protected]>
>> Assisted-by: LLM
>> Signed-off-by: Ritesh Harjani (IBM) <[email protected]>
>> ---
>> Fixes tag here is added since the fix is dependent upon the change mentioned
>> in
>> "Fixes" tag. The Fixes commit-id is from mm-new branch. Although this could
>> be
>> an older pre-existing issue, but got exposed after this patch series.
>
> I knew this issue, so I have a local commit to fix it as well. I'd like to
> share
> it with you, because it looks a little simple, does it work for you?
>
> powerpc/mm: fix uninitialized ZONE_DEVICE struct pages with PMDs
>
> PowerPC radix vmemmap population can reuse an existing PMD mapping while
> populating a device DAX range whose section is marked as HVO-enabled. The
> reused mapping is an ordinary vmemmap mapping: the redundant tail struct
> pages are not remapped to the shared tail page.
>
> Leaving the section marked as HVO-enabled makes later ZONE_DEVICE
> initialization treat the range as optimized and initialize only the struct
> pages that would be backed by a real HVO mapping. The remaining tail struct
> pages can then be left uninitialized even though they are still used as
> metadata for the device DAX compound page.
>
Here, feel free to use the call trace from my commit msg to show what goes
wrong w/o this fix.
> The Fixes target predates the section-based vmemmap optimization
> infrastructure, so older kernels do not have section state that can be
> marked non-HVO. Backports to such kernels need to handle the PMD-reuse
> case in powerpc radix code directly, for example by initializing the tail
> struct pages that core ZONE_DEVICE initialization would otherwise skip.
>
> Mark the section as non-HVO when reusing an existing PMD mapping, so the
> section's HVO state matches the actual vmemmap layout and all required
> struct pages are initialized.
>
> Fixes: f2b79c0d7968 ("powerpc/book3s64/radix: add support for vmemmap
> optimization for radix")
> Signed-off-by: Muchun Song <[email protected]>
> ---
> arch/powerpc/mm/book3s64/radix_pgtable.c | 11 ++++++++++-
> 1 file changed, 10 insertions(+), 1 deletion(-)
yes, love this approach. Thanks for sharing the fix.
>
> diff --git a/arch/powerpc/mm/book3s64/radix_pgtable.c
> b/arch/powerpc/mm/book3s64/radix_pgtable.c
> index 72793ace1fb6..fc3a2d4b422d 100644
> --- a/arch/powerpc/mm/book3s64/radix_pgtable.c
> +++ b/arch/powerpc/mm/book3s64/radix_pgtable.c
> @@ -1200,7 +1200,16 @@ static int __meminit
> vmemmap_populate_compound_pages(unsigned long start_pfn,
> return -ENOMEM;
>
> if (pmd_leaf(READ_ONCE(*pmd))) {
> - /* existing huge mapping. Skip the range */
> + unsigned long addr_pfn = page_to_pfn((struct page
> *)addr);
> + struct mem_section *ms = __pfn_to_section(addr_pfn);
> +
> + /*
> + * Existing PMD mappings can cover the edges of a valid
> + * HVO populate range. They are ordinary vmemmap
> mappings,
> + * so the current section must not be treated as
> HVO-backed
> + * when its struct pages are initialized later.
> + */
> + section_set_order(ms, 0);
I think you meant "section_set_compound_order(ms, 0);"
yes, otherwise the patch looks good to me. Can you please post this
patch separately then? Feel free to add:
Reviewed-by: Ritesh Harjani (IBM) <[email protected]>
I have validated the fix with my custom reproducer and I see the issue
is fixed with this. Meanwhile, Pavithra (cc'd) is testing this and few
other patches of your series. Once her testing is finished, she can help
provide her Tested-by/feedback if any.
> next = pmd_addr_end(addr, end);
> continue;
> }
-ritesh
