Re: CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP is broken, was Re: [RFC PATCH 0/6] Deep talk about folio vmap
From: Huan Yang
Date: Mon Apr 07 2025 - 03:10:41 EST
在 2025/4/7 14:43, Muchun Song 写道:
On Apr 7, 2025, at 11:37, Muchun Song <muchun.song@xxxxxxxxx> wrote:
On Apr 7, 2025, at 11:21, Huan Yang <link@xxxxxxxx> wrote:
在 2025/4/7 10:57, Muchun Song 写道:
On Apr 7, 2025, at 09:59, Huan Yang <link@xxxxxxxx> wrote:
在 2025/4/4 18:07, Muchun Song 写道:
On Apr 4, 2025, at 17:38, Muchun Song <muchun.song@xxxxxxxxx> wrote:
On Apr 4, 2025, at 17:01, Christoph Hellwig <hch@xxxxxx> wrote:
After the btrfs compressed bio discussion I think the hugetlb changes that
skip the tail pages are fundamentally unsafe in the current kernel.
That is because the bio_vec representation assumes tail pages do exist, so
as soon as you are doing direct I/O that generates a bvec starting beyond
the present head page things will blow up. Other users of bio_vecs might
do the same, but the way the block bio_vecs are generated are very suspect
to that. So we'll first need to sort that out and a few other things
before we can even think of enabling such a feature.
I would like to express my gratitude to Christoph for including me in the
thread. I have carefully read the cover letter in [1], which indicates
that an issue has arisen due to the improper use of `vmap_pfn()`. I'm
wondering if we could consider using `vmap()` instead. In the HVO scenario,
the tail struct pages do **exist**, but they are read-only. I've examined
the code of `vmap()`, and it appears that it only reads the struct page.
Therefore, it seems feasible for us to use `vmap()` (I am not a expert in
udmabuf.). Right?
I believe my stance is correct. I've also reviewed another thread in [2].
Allow me to clarify and correct the viewpoints you presented. You stated:
"
So by HVO, it also not backed by pages, only contains folio head, each
tail pfn's page struct go away.
"
This statement is entirely inaccurate. The tail pages do not cease to exist;
rather, they are read-only. For your specific use-case, please use `vmap()`
to resolve the issue at hand. If you wish to gain a comprehensive understanding
I see the document give a simple graph to point:
+-----------+ ---virt_to_page---> +-----------+ mapping to +-----------+
| | | 0 | -------------> | 0 |
| | +-----------+ +-----------+
| | | 1 | -------------> | 1 |
| | +-----------+ +-----------+
| | | 2 | ----------------^ ^ ^ ^ ^ ^
| | +-----------+ | | | | |
| | | 3 | ------------------+ | | | |
| | +-----------+ | | | |
| | | 4 | --------------------+ | | |
| PMD | +-----------+ | | |
| level | | 5 | ----------------------+ | |
| mapping | +-----------+ | |
| | | 6 | ------------------------+ |
| | +-----------+ |
| | | 7 | --------------------------+
| | +-----------+
| |
| |
| |
+-----------+
If I understand correct, each 2-7 tail's page struct is freed, so if I just need map page 2-7, can we use vmap do
something correctly?
The answer is you can. It is essential to distinguish between virtual
Thanks for your reply, but I still can't understand it. For example, I need vmap a hugetlb HVO folio's
2-7 page:
struct page **pages = kvmalloc(sizeof(*pages), 6, GFP_KENREL);
for (i = 2; i < 8; ++i)
pages[i] = folio_page(folio, i); //set 2-7 range page into pages,
void *vaddr = vmap(pages, 6, 0, PAGE_KERNEL);
For no HVO pages, this can work. If HVO enabled, do "pages[i] = folio_page(folio, i);" just
got the head page? and how vmap can correctly map each page?
Why do you think folio_page(folio, i) (i ≠ 0) returns the head page?
Is it speculation or tested? Please base it on the actual situation
instead of indulging in wild thoughts.
By the way, in case you truly struggle to comprehend the fundamental
aspects of HVO, I would like to summarize for you the user-visible
behaviors in comparison to the situation where HVO is disabled.
HVO Status Tail Page Structures Head Page Structures
Enabled Read-Only (RO) Read-Write (RW)
Disabled Read-Write (RW) Read-Write (RW)
The sole distinction between the two scenarios lies in whether the
tail page structures are allowed to be written or not. Please refrain
from getting bogged down in the details of the implementation of HVO.
Thanks, I do a test, an figure out that I'm totally misunderstand it.
Even if HVO enabled, tail page struct freed and point to head, linear mapping still exist, so that any page_to_pfn,
page_to_virt(also folio's version), if start from head page can compute each need page like folio_page, can still work:
hvo head 0xfffff9de849d0000, pfn=0x127400, wish offset_pfn 0x1275f1, idx 497 is 0xfffff9de849d7c40, pfn=0x1275f1.
When vmap, we no need to touch actually page's content, just turn to pfn, so, work well.
BTW, even if we need to touch actually input page struct, it point to head page, I guess will effect nothing.
If anything still misunderstand, please corrent me. :)
Muchun, thank you for your patience,
Huan Yang
Thanks,
Muchun.
Thanks,
Muchun.
Please correct me. :)
Thanks,
Huan Yang
address (VA) and physical address (PA). The VAs of tail struct pages
aren't freed but remapped to the physical page mapped by the VA of the
head struct page (since contents of those tail physical pages are the
same). Thus, the freed pages are the physical pages mapped by original
tail struct pages, not their virtual addresses. Moreover, while it
is possible to read the virtual addresses of these tail struct pages,
any write operations are prohibited since it is within the realm of
acceptability that the kernel is expected to perform write operations
solely on the head struct page of a compound head and conduct read
operations only on the tail struct pages. BTW, folio infrastructure
is also based on this assumption.
Thanks,
Muchun.
Or something I still misunderstand, please correct me.
Thanks,
Huan Yang
of the fundamentals of HVO, I kindly suggest a thorough review of the document
in [3].
[2] https://lore.kernel.org/lkml/5229b24f-1984-4225-ae03-8b952de56e3b@xxxxxxxx/#t
[3] Documentation/mm/vmemmap_dedup.rst
[1] https://lore.kernel.org/linux-mm/20250327092922.536-1-link@xxxxxxxx/T/#m055b34978cf882fd44d2d08d929b50292d8502b4
Thanks,
Muchun.