linux-kernel - fuse readdirplus skip one entry when interrupted by signal

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-ID: <d5c6816c-1af6-5518-e802-cf8fa372ce61@gmail.com>
Date:   Tue, 24 Oct 2017 20:10:49 +0200
From:   Jakob Unterwurzacher <jakobunt@...il.com>
To:     linux-fsdevel@...r.kernel.org,
        Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Miklos Szeredi <miklos@...redi.hu>
Subject: fuse readdirplus skip one entry when interrupted by signal

A user running a Haskell program [1] noticed a problem with fuse's
readdirplus: when it is interrupted by a signal, it skips one
directory entry.

The problem is most apparent with Haskell as it uses
SIGVTALRM to interrupt it's own green threads.

A minimal reproducer in C, "ls-count.c", is available [2]. The problem
has been reproduced against libfuse's "passthrough_fh.c", but also against
gocryptfs, which uses go-fuse instead of libfuse. This suggest
that the bug is in kernel-space, which also the opinion of libfuse
upstream [3].

What "ls-count.c" does is that it loops over readdir while sending itself
SIGVTALRM. When the count of directory entries changes, it exits:

	$ ./ls-count b
	ls-count: counts do not match: 2 vs 1

strace against ls-count shows that we get one entry, when we should get
two ("." and ".."):

	getdents(3, /* 1 entries */, 32768)     = 24
	--- SIGVTALRM ---
	rt_sigreturn({mask=[]})                 = 24
	getdents(3, /* 0 entries */, 32768)     = 0

The debug output from go-fuse [4] shows what seems to be happening:

	Dispatch 548: READDIRPLUS, NodeId: 1. data: {Fh 3 off 0 sz 4096}
	Serialize 548: READDIRPLUS code: OK value:  320 bytes data
	Dispatch 549: READDIRPLUS, NodeId: 1. data: {Fh 3 off 2 sz 4096}
	Serialize 549: READDIRPLUS code: OK value:

The kernel starts reading the directory from "off 0", where it is
interrupted, and only returns one entry to userspace. Then it continues
reading at "off 2". Offset 1 is skipped.

I can reliably reproduce this within 1 second against kernel 4.12.5.

Best regards,
Jakob

[1] https://github.com/hanwen/go-fuse/issues/191
[2]
https://gist.githubusercontent.com/rfjakob/79581292a037ae7cb068067cb6207ef8/raw/f71494a291cfded8a96d02c3f0ee2983457591cc/ls-count.c
[3] https://github.com/libfuse/libfuse/issues/214
[4] gocryptfs -fg -fusedebug -nosyslog