Jens Axboe <axboe@kernel.dk> writes:

> On 2010-08-06 12:42, Dmitry Monakhov wrote:
>> Hi Jens,
>> Seems that my first mail was missed somewhere.  
>> I've found couple of trivial issues in blkdev_issue_zeroout()
>> implementation. Unfortunately I've miss during initial testing phase
>> because always called it with BARRIER|WAIT flags.
>
> BTW, this:
>
> @@ -218,15 +222,18 @@ submit:
>  		/* One of bios in the batch was completed with error.*/
>  		ret = -EIO;
>  
> -	if (ret)
> +	if (ret && ret != -ENOMEM)
>  		goto out;
>  
>  	if (test_bit(BIO_EOPNOTSUPP, &bb.flags)) {
>  		ret = -EOPNOTSUPP;
>  		goto out;
>  	}
> -	if (nr_sects != 0)
> +	if (nr_sects != 0) {
> +		if (ret == -ENOMEM)
> +			io_schedule();
>  		goto submit;
> +	}
>  out:
>  	return ret;
>  }
>
> is broken. Either the caller sets __GFP_WAIT and then bio_alloc() will
> not fail, or GFP_ATOMIC is used knowing that the call can fail and
> return ENOMEM. Don't code in retry logic like this.
Ok, my fault and in fact i've done in explicitly. I just thought
that blk-layer is some times an exception from general GFP_ATOMIC rule
because in some places in blk-layer we stick to GFP_NOFAIL semantics
regardless to actual gfp flags.

New version attached.