Re: [PATCH] mm: Remove PFN_MAP, PFN_SPECIAL, PFN_SG_CHAIN and PFN_SG_LAST

[David Hildenbrand <david@redhat.com>]

On 11.06.25 10:03, Marek Szyprowski wrote:
> On 11.06.2025 04:38, Alistair Popple wrote:
>> On Tue, Jun 10, 2025 at 06:18:09PM +0200, Marek Szyprowski wrote:
>>> On 04.06.2025 05:21, Alistair Popple wrote:
>>>> The PFN_MAP flag is no longer used for anything, so remove it.
>>>> The PFN_SG_CHAIN and PFN_SG_LAST flags never appear to have been
>>>> used so also remove them. The last user of PFN_SPECIAL was removed
>>>> by 653d7825c149 ("dcssblk: mark DAX broken, remove FS_DAX_LIMITED
>>>> support").
>>>>
>>>> Signed-off-by: Alistair Popple <apopple@nvidia.com>
>>>> Acked-by: David Hildenbrand <david@redhat.com>
>>>> Reviewed-by: Christoph Hellwig <hch@lst.de>
>>>> Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
>>>> Cc: gerald.schaefer@linux.ibm.com
>>>> Cc: dan.j.williams@intel.com
>>>> Cc: jgg@ziepe.ca
>>>> Cc: willy@infradead.org
>>>> Cc: david@redhat.com
>>>> Cc: linux-kernel@vger.kernel.org
>>>> Cc: nvdimm@lists.linux.dev
>>>> Cc: jhubbard@nvidia.com
>>>> Cc: hch@lst.de
>>>> Cc: zhang.lyra@gmail.com
>>>> Cc: debug@rivosinc.com
>>>> Cc: bjorn@kernel.org
>>>> Cc: balbirs@nvidia.com
>>>> Cc: lorenzo.stoakes@oracle.com
>>>> Cc: John@Groves.net
>>>>
>>>> ---
>>>>
>>>> Splitting this off from the rest of my series[1] as a separate clean-up
>>>> for consideration for the v6.16 merge window as suggested by Christoph.
>>>>
>>>> [1] - https://lore.kernel.org/linux-mm/cover.541c2702181b7461b84f1a6967a3f0e823023fcc.1748500293.git-series.apopple@nvidia.com/
>>>> ---
>>>>     include/linux/pfn_t.h             | 31 +++----------------------------
>>>>     mm/memory.c                       |  2 --
>>>>     tools/testing/nvdimm/test/iomap.c |  4 ----
>>>>     3 files changed, 3 insertions(+), 34 deletions(-)
>>> This patch landed in today's linux-next as commit 28be5676b4a3 ("mm:
>>> remove PFN_MAP, PFN_SPECIAL, PFN_SG_CHAIN and PFN_SG_LAST"). In my tests
>>> I've noticed that it breaks operation of all RISC-V 64bit boards on my
>>> test farm (VisionFive2, BananaPiF3 as well as QEMU's Virt machine). I've
>>> isolated the changes responsible for this issue, see the inline comments
>>> in the patch below. Here is an example of the issues observed in the
>>> logs from those machines:
>> Thanks for the report. I'm really confused by this because this change should
>> just be removal of dead code - nothing sets any of the removed PFN_* flags
>> AFAICT.
>>
>> I don't have access to any RISC-V hardwdare but you say this reproduces under
>> qemu - what do you run on the system to cause the error? Is it just a simple
>> boot and load a module or are you running selftests or something else?
> 
> It fails a simple boot test. Here is a detailed instruction how to
> reproduce this issue with the random Debian rootfs image found on the
> internet (tested on Ubuntu 22.04, with next-20250610
> kernel source):

riscv is one of the archs where pte_mkdevmap() will *not* set the pte as special. (I
raised this recently in the original series, it's all a big mess)

So, before this change here, pfn_t_devmap() would have returned "false" if only
PFN_DEV was set, now it would return "true" if only PFN_DEV is set.

Consequently, in insert_pfn() we would have done a pte_mkspecial(), now we do a
pte_mkdevmap() -- again, which does not imply "special" on riscv.

riscv selects CONFIG_ARCH_HAS_PTE_SPECIAL, so if !pte_special(), it's considered as
normal.

Would the following fix your issue?


diff --git a/mm/memory.c b/mm/memory.c
index 8eba595056fe3..0e972c3493692 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -589,6 +589,10 @@ struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
  {
         unsigned long pfn = pte_pfn(pte);
  
+       /* TODO: remove this crap and set pte_special() instead. */
+       if (pte_devmap(pte))
+               return NULL;
+
         if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
                 if (likely(!pte_special(pte)))
                         goto check_pfn;
@@ -598,16 +602,6 @@ struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
                         return NULL;
                 if (is_zero_pfn(pfn))
                         return NULL;
-               if (pte_devmap(pte))
-               /*
-                * NOTE: New users of ZONE_DEVICE will not set pte_devmap()
-                * and will have refcounts incremented on their struct pages
-                * when they are inserted into PTEs, thus they are safe to
-                * return here. Legacy ZONE_DEVICE pages that set pte_devmap()
-                * do not have refcounts. Example of legacy ZONE_DEVICE is
-                * MEMORY_DEVICE_FS_DAX type in pmem or virtio_fs drivers.
-                */
-                       return NULL;
  
                 print_bad_pte(vma, addr, pte, NULL);
                 return NULL;


But, I would have thought the later patches in Alistairs series would sort that out
(where we remove pte_devmap() ... )

-- 
Cheers,

David / dhildenb