Re: [PATCH v2 4/5]nbd: make nbd device wait for its users.
From: Pranay Srivastava
Date: Wed Jun 15 2016 - 05:17:35 EST
Hey Markus,
On Wed, Jun 15, 2016 at 12:00 PM, Markus Pargmann <mpa@xxxxxxxxxxxxxx> wrote:
> Hi Pranay,
>
> On Tuesday 14 June 2016 15:03:40 Pranay Srivastava wrote:
>> Hi Markus,
>>
>> On Tue, Jun 14, 2016 at 2:29 PM, Markus Pargmann <mpa@xxxxxxxxxxxxxx> wrote:
>> >
>> > On Thursday 02 June 2016 13:25:00 Pranay Kr. Srivastava wrote:
>> > > When a timeout occurs or a recv fails, then
>> > > instead of abruplty killing nbd block device
>> > > wait for it's users to finish.
>> > >
>> > > This is more required when filesystem(s) like
>> > > ext2 or ext3 don't expect their buffer heads to
>> > > disappear while the filesystem is mounted.
>> > >
>> > > Each open of a nbd device is refcounted, while
>> > > the userland program [nbd-client] doing the
>> > > NBD_DO_IT ioctl would now wait for any other users
>> > > of this device before invalidating the nbd device.
>> > >
>> > > Signed-off-by: Pranay Kr. Srivastava <pranjas@xxxxxxxxx>
>> > > ---
>> > > drivers/block/nbd.c | 58 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>> > > 1 file changed, 58 insertions(+)
>> > >
>> > > diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
>> > > index d1d898d..4da40dc 100644
>> > > --- a/drivers/block/nbd.c
>> > > +++ b/drivers/block/nbd.c
>> > > @@ -70,10 +70,13 @@ struct nbd_device {
>> > > #if IS_ENABLED(CONFIG_DEBUG_FS)
>> > > struct dentry *dbg_dir;
>> > > #endif
>> > > + atomic_t inuse;
>> > > /*
>> > > *This is specifically for calling sock_shutdown, for now.
>> > > */
>> > > struct work_struct ws_shutdown;
>> > > + struct kref users;
>> > > + struct completion user_completion;
>> > > };
>> > >
>> > > #if IS_ENABLED(CONFIG_DEBUG_FS)
>> > > @@ -104,6 +107,7 @@ static DEFINE_SPINLOCK(nbd_lock);
>> > > * Shutdown function for nbd_dev work struct.
>> > > */
>> > > static void nbd_ws_func_shutdown(struct work_struct *);
>> > > +static void nbd_kref_release(struct kref *);
>> > >
>> > > static inline struct device *nbd_to_dev(struct nbd_device *nbd)
>> > > {
>> > > @@ -682,6 +686,8 @@ static void nbd_reset(struct nbd_device *nbd)
>> > > nbd->flags = 0;
>> > > nbd->xmit_timeout = 0;
>> > > INIT_WORK(&nbd->ws_shutdown, nbd_ws_func_shutdown);
>> > > + init_completion(&nbd->user_completion);
>> > > + kref_init(&nbd->users);
>> > > queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
>> > > del_timer_sync(&nbd->timeout_timer);
>> > > }
>> > > @@ -815,6 +821,14 @@ static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
>> > > kthread_stop(thread);
>> > >
>> > > sock_shutdown(nbd);
>> > > + /*
>> > > + * kref_init initializes with ref count as 1,
>> > > + * nbd_client, or the user-land program executing
>> > > + * this ioctl will make the refcount to 2[at least]
>> > > + * so subtracting 2 from refcount.
>> > > + */
>> > > + kref_sub(&nbd->users, 2, nbd_kref_release);
>> >
>> > Why don't you use a kref_put?
>>
>> Ok, so I'll try to explain as I've understood the problem.
>>
>> When the module is loaded the kref is initialized to 1.
>>
>> Suppose now, someone has started nbd-client [nbdC-1] , then this
>> nbd-client will increase the ref count to 2. So far so good...
>>
>> Now let's say this device is being shutdown via nbd-client[nbdC-2].
>>
>> nbdC-1 will subtract the refcount by two, it has to do in NBD_DO_IT
>> since device file will not
>> be closed until after ioctl is over, and it'll wait_for_completion.
>>
>> nbdC-2 now closes it's use of device file, this makes the refcount as
>> zero and completion
>> is triggered with nbdC-1 completed.
>>
>> Now we don't want to trigger kref_put when nbdC-1 closes the device
>> file so kref_put needs
>> to be conditional in this regard so for that in_use is used.
>>
>>
>> >
>> > > + wait_for_completion(&nbd->user_completion);
>> > > mutex_lock(&nbd->tx_lock);
>> > > nbd_clear_que(nbd);
>> > > kill_bdev(bdev);
>> > > @@ -865,13 +879,56 @@ static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
>> > >
>> > > return error;
>> > > }
>> > > +static void nbd_kref_release(struct kref *kref_users)
>> > > +{
>> > > + struct nbd_device *nbd = container_of(kref_users, struct nbd_device,
>> > > + users);
>> >
>> > Not indented to opening bracket.
>> >
>> > > + pr_debug("Releasing kref [%s]\n", __func__);
>> > > + atomic_set(&nbd->inuse, 0);
>> > > + complete(&nbd->user_completion);
>> > > +
>> > > +}
>> > > +
>> > > +static int nbd_open(struct block_device *bdev, fmode_t mode)
>> > > +{
>> > > + struct nbd_device *nbd_dev = bdev->bd_disk->private_data;
>> > > +
>> > > + if (kref_get_unless_zero(&nbd_dev->users))
>> > > + atomic_set(&nbd_dev->inuse, 1);
>> > > +
>> > > + pr_debug("Opening nbd_dev %s. Active users = %u\n",
>> > > + bdev->bd_disk->disk_name,
>> > > + atomic_read(&nbd_dev->users.refcount) - 1);
>> >
>> > Indent to opening bracket.
>> >
>> > > + return 0;
>> > > +}
>> > > +
>> > > +static void nbd_release(struct gendisk *disk, fmode_t mode)
>> > > +{
>> > > + struct nbd_device *nbd_dev = disk->private_data;
>> > > + /*
>> > > + *kref_init initializes ref count to 1, so we
>> > > + *we check for refcount to be 2 for a final put.
>> > > + *
>> > > + *kref needs to be re-initialized just here as the
>> > > + *other process holding it must see the ref count as 2.
>> > > + */
>> > > + if (atomic_read(&nbd_dev->inuse))
>> > > + kref_put(&nbd_dev->users, nbd_kref_release);
>> >
>>
>> > What is this inuse atomic for? Everyone that releases the nbd device
>> > will need to execute a kref_put().
>>
>> To do away with inuse, perhaps we can do
>>
>> kref_get just before leaving the NBD_DO_IT? so that when device file
>> is closed everyone
>> would do a kref_put? However there's a small race window while the
>> kref is being initialized,
>> and another process [not just nbd-client] is trying to open the device.
>>
>> Do you think it's better to do this by introducing a spin_lock instead
>> of atomic?
>>
>> Let me know if my understanding is correct.
>
> Thanks for the explanations. I think my understanding was off by one ;).
> I didn't realize that the DO_IT thread from the userspace has the block
> device open as well.
>
> I thought a bit about this, does it make sense to delay the essential
> cleanup steps until really all open file handles were closed? So that
> even if the DO_IT thread exits, the block device is still there. Only if
> the file is closed everything is cleaned up. Maybe this makes the code
> simpler and we can directly use krefs without any strange constructs.
> What do you think?
>
I chose open/close as that is the common interface to all processes that need
to use nbd device and not just nbd-client.
Let's take example of a mount, so some user has just done a mount of this device
and right now it's not doing anything. So someone issues an NBD_DISCONNECT
and wham... The idea is to propagate errors to user space correctly.
So the solution I've proposed says, if there's someone
apart_from_nbd-client (which
is currently waiting for NBD_DO_IT to complete) is
using this device then nbd-client should honor that and shouldn't
allow this device
to be reused until after all such processes have left this device.
> This would also allow the client to setup a new socket as long as it
> does not close the nbd file handle.
I think that is still possible right, which is why there's a kref_sub
of 2, so the wait
is only for the "other processes" using this device and _not_ this
nbd-client whose
NBD_DO_IT just got over.
Now I'm just concerned over the processes which are trying to use this
nbd device
but the re-initialzation code at the end of NBD_DO_IT hasn't been done. So it's
possible some device may skip getting a kref, due to kref_get_unless_zero.
Actually I wanted to put this kref under bd_mutex to avoid such races and it'll
always be a call to kref_get in open and kref_put in close. Still
kref_sub(2) would
be required :-)
>
> Could this behavior be potentially problematic for any client
> implementation? Does it solve our other issue with setting up a new
> sockets for an existing nbd blockdevice?
I don't think that would be a problem. sock_shutdown is before the wait
right? So in error condition the worker thread would close it and set to null
while for a normal case, sock_shutdown in NBD_DO_IT would set that.
So it should be OK.
>
> Cc Wouter
>
> Best Regards,
>
> Markus
>
>>
>>
>> >
>> > Best Regards,
>> >
>> > Markus
>> >
>> > > +
>> > > + pr_debug("Closing nbd_dev %s. Active users = %u\n",
>> > > + disk->disk_name,
>> > > + atomic_read(&nbd_dev->users.refcount) - 1);
>> > > +}
>> > >
>> > > static const struct block_device_operations nbd_fops = {
>> > > .owner = THIS_MODULE,
>> > > .ioctl = nbd_ioctl,
>> > > .compat_ioctl = nbd_ioctl,
>> > > + .open = nbd_open,
>> > > + .release = nbd_release
>> > > };
>> > >
>> > > +
>> > > static void nbd_ws_func_shutdown(struct work_struct *ws_nbd)
>> > > {
>> > > struct nbd_device *nbd_dev = container_of(ws_nbd, struct nbd_device,
>> > > @@ -1107,6 +1164,7 @@ static int __init nbd_init(void)
>> > > disk->fops = &nbd_fops;
>> > > disk->private_data = &nbd_dev[i];
>> > > sprintf(disk->disk_name, "nbd%d", i);
>> > > + atomic_set(&nbd_dev[i].inuse, 0);
>> > > nbd_reset(&nbd_dev[i]);
>> > > add_disk(disk);
>> > > }
>> > >
>> >
>> > --
>> > Pengutronix e.K. | |
>> > Industrial Linux Solutions | http://www.pengutronix.de/ |
>> > Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
>> > Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
>>
>>
>>
>>
>>
>
> --
> Pengutronix e.K. | |
> Industrial Linux Solutions | http://www.pengutronix.de/ |
> Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 |
> Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
--
---P.K.S