Message-ID: <c1b69c38-59c8-4d27-9ff5-bcca553da7c2@redhat.com>
Date: Wed, 8 May 2024 19:28:28 +0200
From: David Hildenbrand <david@...hat.com>
To: Prakash Sangappa <prakash.sangappa@...cle.com>
Cc: "linux-mm@...ck.org" <linux-mm@...ck.org>,
 "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
 "muchun.song@...ux.dev" <muchun.song@...ux.dev>,
 "akpm@...ux-foundation.org" <akpm@...ux-foundation.org>,
 "willy@...radead.org" <willy@...radead.org>
Subject: Re: [RFC PATCH 0/1] Address hugetlbfs mmap behavior

On 08.05.24 19:00, Prakash Sangappa wrote:
> 
> 
>> On May 7, 2024, at 5:00 AM, David Hildenbrand <david@...hat.com> wrote:
>>
>> On 03.05.24 03:21, Prakash Sangappa wrote:
>>> This patch proposes to fix hugetlbfs mmap behavior so that the
>>> file size does not get updated in the mmap call.
>>> The current behavior is that hugetlbfs file size will get extended by a
>>> PROT_WRITE mmap(2) call if mmap size is greater then file size. This is
>>> not normal filesystem behavior.
>>> There seem to have been very little discussion about this. There was a
>>> patch discussion[1] a while back, implying hugetlbfs file size needs
>>> extending because of the hugetlb page reservations. Looks like this was
>>> not merged.
>>> It appears there is no correlation between file size and hugetlb page
>>> reservations. Take the case of PROT_READ mmap, where the file size is
>>> not extended even though hugetlb pages are reserved.
>>> On the other hand ftruncate(2) to increase a file size does not reserve
>>> hugetlb pages. Also, mmap with MAP_NORESERVE flag extends the file size
>>> even though hugetlb pages are not reserved.
>>> Hugetlb pages get reserved(if MAP_NORESERVE is not specified) when the
>>> hugeltbfs file is mmapped, and it only covers the file's offset,length
>>> range specified in the mmap call.
>>> Issue:
>>> Some applications would prefer to manage hugetlb page allocations explicity
>>> with use of fallocate(2). The hugetlbfs file would be PROT_WRITE mapped with
>>> MAP_NORESERVE flag, which is accessed only after allocating necessary pages
>>> using fallocate(2) and release the pages by truncating the file size. Any stray
>>> access beyond file size is expected to generate a signal. This does not
>>> work properly due to current behavior which extends file size in mmap call.
>>
>> Would a simple workaround be to mmap(PROT_READ) and then mprotect(PROT_READ|PROT_WRITE)?
> 
> Another workaround could be to ftruncate(2) the file after  mmap(PROT_READ|PROT_WRITE), if MAP_NORESERVE is used. But these will require application changes as a special case for hugetlbfs that can be considered.

I'd assume that most applications that mmap() hugetlb files need to
special-case hugetlb because of the different logical page size
granularity already. But yes, it's all unfortunate.

> 
> However, should this mmap behavior  be addressed? Why mmap(PROT_WRITE) has to extend the file size needs clarification.

The issue is, as you write, that it's existing behavior and changing it
could cause harm to other apps that rely on that. But I do wonder if really
anybody relies on that ...

Let's explore the history:

The current VM_WRITE check was added in:

commit b6174df5eec9cdfd598c03d6d0807e344e109213
Author: Zhang, Yanmin <yanmin.zhang@...el.com>
Date:   Mon Jul 10 04:44:49 2006 -0700

     [PATCH] mmap zero-length hugetlb file with PROT_NONE to protect a hugetlb virtual area
     
     Sometimes, applications need below call to be successful although
     "/mnt/hugepages/file1" doesn't exist.
     
     fd = open("/mnt/hugepages/file1", O_CREAT|O_RDWR, 0755);
     *addr = mmap(NULL, 0x1024*1024*256, PROT_NONE, 0, fd, 0);
     
     As for regular pages (or files), above call does work, but as for huge
     pages, above call would fail because hugetlbfs_file_mmap would fail if
     (!(vma->vm_flags & VM_WRITE) && len > inode->i_size).
     
     This capability on huge page is useful on ia64 when the process wants to
     protect one area on region 4, so other threads couldn't read/write this
     area.  A famous JVM (Java Virtual Machine) implementation on IA64 needs the
     capability.

But it was only moved.

Before that patch:
* mmap(PROT_WRITE) would have failed if the file size would be exceeded
* mmap(PROT_READ/PROT_NONE) would have extended the file

After that patch
* mmap(PROT_WRITE) will extend the file
* mmap(PROT_READ/PROT_NONE) do not extend the file

The code before that predates git times.

Having a mount option to change that really is suboptimal IMHO ... we shouldn't add mount options to work
around all hugetlbfs quirks.

I suggest either

(a) Document it, along with the workaround
(b) Change it an cross fingers.


In QEMU source code is a very interesting comment:

      * ftruncate is not supported by hugetlbfs in older
      * hosts, so don't bother bailing out on errors.
      * If anything goes wrong with it under other filesystems,
      * mmap will fail.

So, was mmap() maybe the way to easily grow a hugetlbfs file before ftruncate() support
was added?

QEMU will only call ftruncate() if the file size is empty, though. So if you'd have a
smaller file QEMU would not try growing it, and mmap() would succeed and grow it. That's
a rare case to happen, though, and likely also undesired here: we want it to behave just
like ordinary files!

-- 
Cheers,

David / dhildenb

Powered by blists - more mailing lists