Re: [PATCH v3 10/15] block: Add fops atomic write support

From: Nilay Shroff <nilay@linux.ibm.com>
To: John Garry <john.g.garry@oracle.com>
Cc: axboe@kernel.dk, brauner@kernel.org, bvanassche@acm.org,
	dchinner@redhat.com, djwong@kernel.org, hch@lst.de, jack@suse.cz,
	jbongio@google.com, jejb@linux.ibm.com, kbusch@kernel.org,
	linux-block@vger.kernel.org, linux-fsdevel@vger.kernel.org,
	linux-kernel@vger.kernel.org, linux-nvme@lists.infradead.org,
	linux-scsi@vger.kernel.org, linux-xfs@vger.kernel.org,
	martin.petersen@oracle.com, ming.lei@redhat.com,
	ojaswin@linux.ibm.com, sagi@grimberg.me, tytso@mit.edu,
	viro@zeniv.linux.org.uk
Subject: Re: [PATCH v3 10/15] block: Add fops atomic write support
Date: Tue, 13 Feb 2024 16:38:31 +0530	[thread overview]
Message-ID: <e99cf4ef-40ec-4e66-956f-c9e2aebb4621@linux.ibm.com> (raw)
In-Reply-To: <9ffc3102-2936-4f83-b69d-bbf64793b9ca@oracle.com>

On 2/13/24 15:28, John Garry wrote:
> On 13/02/2024 09:36, Nilay Shroff wrote:
>>> +static bool blkdev_atomic_write_valid(struct block_device *bdev, loff_t pos,
>>
>>> +                      struct iov_iter *iter)
>>
>>> +{
>>
>>> +    struct request_queue *q = bdev_get_queue(bdev);
>>
>>> +    unsigned int min_bytes = queue_atomic_write_unit_min_bytes(q);
>>
>>> +    unsigned int max_bytes = queue_atomic_write_unit_max_bytes(q);
>>
>>> +
>>
>>> +    if (!iter_is_ubuf(iter))
>>
>>> +        return false;
>>
>>> +    if (iov_iter_count(iter) & (min_bytes - 1))
>>
>>> +        return false;
>>
>>> +    if (!is_power_of_2(iov_iter_count(iter)))
>>
>>> +        return false;
>>
>>> +    if (pos & (iov_iter_count(iter) - 1))
>>
>>> +        return false;
>>
>>> +    if (iov_iter_count(iter) > max_bytes)
>>
>>> +        return false;
>>
>>> +    return true;
>>
>>> +}
>>
>>
>>
>> Here do we need to also validate whether the IO doesn't straddle
>>
>> the atmic bondary limit (if it's non-zero)? We do check that IO
>>
>> doesn't straddle the atomic boundary limit but that happens very
>>
>> late in the IO code path either during blk-merge or in NVMe driver
>>
>> code.
> 
> It's relied that atomic_write_unit_max is <= atomic_write_boundary and both are a power-of-2. Please see the NVMe patch, which this is checked. Indeed, it would not make sense if atomic_write_unit_max > atomic_write_boundary (when non-zero).
> 
> So if the write is naturally aligned and its size is <= atomic_write_unit_max, then it cannot be straddling a boundary.

Ok fine but in case the device doesn't support namespace atomic boundary size (i.e. NABSPF is zero) then still do we need 
to restrict IO which crosses the atomic boundary? 

I am quoting this from NVMe spec (Command Set Specification, revision 1.0a, Section 2.1.4.3) : 
"To ensure backwards compatibility, the values reported for AWUN, AWUPF, and ACWU shall be set such that 
they  are  supported  even  if  a  write  crosses  an  atomic  boundary.  If  a  controller  does  not  
guarantee atomicity across atomic boundaries, the controller shall set AWUN, AWUPF, and ACWU to 0h (1 LBA)." 

Thanks,
--Nilay