Re: [fuse-devel] [PATCH] cuse: do not register multiple devices with the same name

[David H. <dh....@go...>]

Hi Tejun

On Mon, Nov 12, 2012 at 7:33 PM, Tejun Heo <tj...@ke...> wrote:
> Hello, David.
>
> On Mon, Nov 12, 2012 at 05:15:48PM +0100, David Herrmann wrote:
>> We do not check whether we already registered a CUSE device with a given
>> name so we might end up with two devices with the same name. Sysfs will
>> then complain as it cannot create suitable directories.
>>
>> This patch makes the init-command fail if there is already a device with
>> the given name. To avoid race-conditions during initialization, we
>> actually need to add the device to the list while still holding the lock
>> for the name-check.
>> The new "ready" field guarantees that the device is still not opened until
>> it is fully initialized.
>>
>> Following the sysfs warnings when registering two devices with the same
>> name:
>>
>>  ------------[ cut here ]------------
>>  WARNING: at fs/sysfs/dir.c:529 sysfs_add_one+0xc8/0xf0()
>>  Hardware name: N150P/N210P/N220P
>>  sysfs: cannot create duplicate filename '/devices/virtual/cuse/ttyFseat0'
>>  Modules linked in: btusb bluetooth
>>  Pid: 14089, comm: lt-kmscon Tainted: G        W    3.5.3+ #60
>>  Call Trace:
>>   [<ffffffff81136400>] ? sysfs_add_one+0x60/0xf0
>>   [<ffffffff8102f99d>] warn_slowpath_common+0x7d/0xc0
>>   [<ffffffff8102fa83>] warn_slowpath_fmt+0x43/0x50
>>   [<ffffffff81136468>] sysfs_add_one+0xc8/0xf0
>>   [<ffffffff81136686>] create_dir+0x76/0xd0
>>   [<ffffffff81136a14>] sysfs_create_dir+0x84/0xe0
>>   [<ffffffff811fe67b>] kobject_add_internal+0x9b/0x200
>>   [<ffffffff811feb38>] kobject_add+0x68/0xc0
>>   [<ffffffff81310c73>] device_add+0xe3/0x680
>>   [<ffffffff8130f5ae>] ? dev_set_name+0x3e/0x40
>>   [<ffffffff811c6834>] cuse_process_init_reply+0x204/0x410
>>   [<ffffffff811c6630>] ? cuse_open+0xe0/0xe0
>>   [<ffffffff811bb23c>] request_end+0xfc/0x1a0
>>   [<ffffffff811bc6e2>] fuse_dev_do_write+0xa32/0xd10
>>   [<ffffffff811ba435>] ? fuse_copy_one+0x45/0x60
>>   [<ffffffff8109cf06>] ? find_get_page+0x66/0xb0
>>   [<ffffffff811bccc0>] ? fuse_dev_splice_write+0x300/0x300
>>   [<ffffffff811bcd29>] fuse_dev_write+0x69/0x80
>>   [<ffffffff810d569c>] do_sync_readv_writev+0xdc/0x120
>>   [<ffffffff810d57db>] ? rw_copy_check_uvector+0x6b/0x130
>>   [<ffffffff810b9c5e>] ? handle_mm_fault+0x12e/0x1f0
>>   [<ffffffff810d5973>] do_readv_writev+0xd3/0x1e0
>>   [<ffffffff810d5ab0>] vfs_writev+0x30/0x60
>>   [<ffffffff810d5c38>] sys_writev+0x48/0xb0
>>   [<ffffffff815846a2>] system_call_fastpath+0x16/0x1b
>>  ---[ end trace 368eb04507b14c94 ]---
>>
>> Signed-off-by: David Herrmann <dh....@go...>
>> ---
>> Hi
>>
>> I am not sure whether this qualifies for the stable-tree, so please CC
>> st...@vg... if you think so.
>>
>> The patch is against linux-next from today.
>>
>> Regards
>> David
>>
>>  fs/fuse/cuse.c | 35 +++++++++++++++++++++++++++++------
>>  1 file changed, 29 insertions(+), 6 deletions(-)
>>
>> diff --git a/fs/fuse/cuse.c b/fs/fuse/cuse.c
>> index 1326051..11fbc52 100644
>> --- a/fs/fuse/cuse.c
>> +++ b/fs/fuse/cuse.c
>> @@ -33,6 +33,7 @@
>>   * closed.
>>   */
>>
>> +#include <linux/atomic.h>
>>  #include <linux/fuse.h>
>>  #include <linux/cdev.h>
>>  #include <linux/device.h>
>> @@ -45,6 +46,7 @@
>>  #include <linux/miscdevice.h>
>>  #include <linux/mutex.h>
>>  #include <linux/slab.h>
>> +#include <linux/smp.h>
>>  #include <linux/spinlock.h>
>>  #include <linux/stat.h>
>>  #include <linux/module.h>
>> @@ -54,6 +56,7 @@
>>  #define CUSE_CONNTBL_LEN     64
>>
>>  struct cuse_conn {
>> +     atomic_t                ready;  /* device is ready for open() */
>
> Hmmm... how about converting cuse_lock to a mutex and wrapping the
> whole registration inside it instead of splitting the synchronization
> into two places?

That's probably the easier way. I will resend the patch later.

Thanks
David