Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Eric Van Hensbergen <ericvh@kernel.org>
Subject: [PATCH v2 0/4] fs/9p: fix mmap regression
Date:   Tue, 18 Jul 2023 20:50:14 +0000
Message-Id: <20230716-fixes-overly-restrictive-mmap-v2-0-147d6b93f699@kernel.org>
MIME-Version: 1.0
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
To:     Latchesar Ionkov <lucho@ionkov.net>,
        Dominique Martinet <asmadeus@codewreck.org>,
        Christian Schoenebeck <linux_oss@crudebyte.com>
Cc:     v9fs@lists.linux.dev, linux-kernel@vger.kernel.org,
        kernel@pengutronix.de, Robert Schwebel <r.schwebel@pengutronix.de>,
        Eric Van Hensbergen <ericvh@kernel.org>
Precedence: bulk

This series attempts to fix a reported exception with mmap
on newer kernels. 

original regression report follows:

TL;DR: mmap() seems to be broken on 9pfs on Linux 6.4. setting
"rootflags=ignoreqv" fixes it as well, but it feels like a regression.

I'm tracking down an issue which recently turned up in DistroKit [1] (an
embedded Linux distro based on the ptxdist build system). The issue was a bit
uggly, as my CI didn't find it (systems boot up normally after a while, and I
only use 9p for virtual qemu machines, while most of the test farm is real
hardware).

The qemu machine in question is qemu-system-arm, emulating an ARM v7a machine.

When starting the systems interactively, I get a lot of error output from
ldconfig, like this:

[   17.412964] systemd-rc-once[127]: ldconfig: Cannot mmap file /lib/libgcc_s.so.1.
[   17.418851] systemd-rc-once[127]: ldconfig: Cannot mmap file /lib/libstdc++.so.
[   17.425009] systemd-rc-once[127]: ldconfig: Cannot mmap file /lib/libstdc++.so.6.0.30.
[   17.436671] systemd-rc-once[127]: ldconfig: Cannot mmap file /lib/libstdc++.so.6.
[   17.448451] systemd-rc-once[127]: ldconfig: Cannot mmap file /lib/libatomic.so.
[   17.456418] systemd-rc-once[127]: ldconfig: Cannot mmap file /lib/libatomic.so.1.2.0.
...

Running ldconfig with strace shows this, for all libraries::

| statx(AT_FDCWD, "/lib/libnm.so.0", AT_STATX_SYNC_AS_STAT|AT_SYMLINK_NOFOLLOW|AT_NO_AUTOMOUNT, STATX_BASIC_STATS, {stx_mask=STATX_BASIC_STATS|STATX_MNT_ID, stx_attributes=0, stx_mode=S_IFLNK|0777, stx_size=14, ...}) = 0
| statx(AT_FDCWD, "/lib/libnm.so.0", AT_STATX_SYNC_AS_STAT|AT_NO_AUTOMOUNT, STATX_BASIC_STATS, {stx_mask=STATX_BASIC_STATS|STATX_MNT_ID, stx_attributes=0, stx_mode=S_IFREG|064 4, stx_size=862228, ...}) = 0
| openat(AT_FDCWD, "/lib/libnm.so.0", O_RDONLY|O_LARGEFILE) = 4
| statx(4, "", AT_STATX_SYNC_AS_STAT|AT_NO_AUTOMOUNT|AT_EMPTY_PATH, STATX_BASIC_STATS, {stx_mask=STATX_BASIC_STATS|STATX_MNT_ID, stx_attributes=0, stx_mode=S_IFREG|0644, stx_s ize=862228, ...}) = 0
| mmap2(NULL, 862228, PROT_READ, MAP_SHARED, 4, 0) = -1 ENODEV (No such device)
| write(2, "ldconfig: ", 10ldconfig: )              = 10
| write(2, "Cannot mmap file /lib/libnm.so.0"..., 34Cannot mmap file /lib/libnm.so.0.) = 34
| write(2, "\n", 1)                       = 1
| close(4)                                = 0

I could track down the breakage to

  1543b4c5071c54d76aad7a7a26a6e43082269b0c

My test setup has, in addition to the patch above, the following patches also
reverted on top of a vanilla 6.4 kernel:

  4eb3117888a923f6b9b1ad2dd093641c49a63ae5
  21e26d5e54ab7cfe6b488fd27d4d70956d07e03b

as 1543b cannot be reverted without those; however, the effect only goes away
when I also revert 1543b. The kernel has no other patches applied, only these
three reverts.

Reported-by: Robert Schwebel <r.schwebel@pengutronix.de>
Signed-off-by: Eric Van Hensbergen <ericvh@kernel.org>
---
Changes in v2:
- fix requested changes in commit messages
- add patch to remove unnecessary invalidate_inode_pages in mmap readonly path
- Link to v1: https://lore.kernel.org/r/20230716-fixes-overly-restrictive-mmap-v1-0-0683b283b932@kernel.org

---
Eric Van Hensbergen (4):
      fs/9p: remove unnecessary and overrestrictive check
      fs/9p: fix typo in comparison logic for cache mode
      fs/9p: fix type mismatch in file cache mode helper
      fs/9p: remove unnecessary invalidate_inode_pages2

 fs/9p/fid.h      | 6 +++---
 fs/9p/vfs_file.c | 3 ---
 2 files changed, 3 insertions(+), 6 deletions(-)
---
base-commit: 95f41d87810083d8b3dedcce46a4e356cf4a9673
change-id: 20230716-fixes-overly-restrictive-mmap-30a23501e787

Best regards,
-- 
Eric Van Hensbergen <ericvh@kernel.org>