Alignment Issue with Direct IO to NVMe Drive

Alignment Issue with Direct IO to NVMe Drive

[Laine Walker-Avina]

Hi all,

We are experiencing an issue with doing direct IO to a NVMe device I'm
helping to develop. Every so often, the physical address given by
sg_dma_address() is aligned to 0x800 instead of 0x1000 as specified by
blk_queue_dma_alignement(queue, 4095) when the queue is initialized.
The request is also split over multiple segments to make up for the
missing space (eg: for a 4k IO it's split into two segments 2k in
size, and for an 8k IO it's split into 3 segments--2k,4k,2k). Our
design requires the physical segments given to the device be aligned
to 4k boundaries and be multiples of 4k in size. When not doing direct
IO the physical addresses appear to always be 4k aligned as expected.
One possible issue is the kernel we're primarily testing against is
2.6.32-220 from CentOS, but we have observed similar behavior from a
vanilla 3.3 kernel as well. Any help would be greatly appreciated.

Thanks,
Laine Walker-Avina

Re: Alignment Issue with Direct IO to NVMe Drive

[Jens Axboe]

On 2012-11-27 01:35, Laine Walker-Avina wrote:
> Hi all,
> 
> We are experiencing an issue with doing direct IO to a NVMe device I'm
> helping to develop. Every so often, the physical address given by
> sg_dma_address() is aligned to 0x800 instead of 0x1000 as specified by
> blk_queue_dma_alignement(queue, 4095) when the queue is initialized.
> The request is also split over multiple segments to make up for the
> missing space (eg: for a 4k IO it's split into two segments 2k in
> size, and for an 8k IO it's split into 3 segments--2k,4k,2k). Our
> design requires the physical segments given to the device be aligned
> to 4k boundaries and be multiples of 4k in size. When not doing direct
> IO the physical addresses appear to always be 4k aligned as expected.
> One possible issue is the kernel we're primarily testing against is
> 2.6.32-220 from CentOS, but we have observed similar behavior from a
> vanilla 3.3 kernel as well. Any help would be greatly appreciated.

I'm assuming you set the hardware sector size to 4k as well?

-- 
Jens Axboe

Re: Alignment Issue with Direct IO to NVMe Drive

[Matthew Wilcox]

On Tue, Nov 27, 2012 at 01:09:46PM +0100, Jens Axboe wrote:
> On 2012-11-27 01:35, Laine Walker-Avina wrote:
> > Hi all,
> > 
> > We are experiencing an issue with doing direct IO to a NVMe device I'm
> > helping to develop. Every so often, the physical address given by
> > sg_dma_address() is aligned to 0x800 instead of 0x1000 as specified by
> > blk_queue_dma_alignement(queue, 4095) when the queue is initialized.

FYI, this is a modification to the driver that Laine has made; presumably
for a limitation of the prototype hardware he's working with.  The NVMe
spec requires the device to be able to do I/Os to 4 byte boundaries.

Laine, when this occurs, what is the alignment of 'offset' in the sg
entry you're looking at?  If userspace is passing in an unaligned address,
I don't think there's anything we do to try to align it.

> > The request is also split over multiple segments to make up for the
> > missing space (eg: for a 4k IO it's split into two segments 2k in
> > size, and for an 8k IO it's split into 3 segments--2k,4k,2k). Our
> > design requires the physical segments given to the device be aligned
> > to 4k boundaries and be multiples of 4k in size. When not doing direct
> > IO the physical addresses appear to always be 4k aligned as expected.
> > One possible issue is the kernel we're primarily testing against is
> > 2.6.32-220 from CentOS, but we have observed similar behavior from a
> > vanilla 3.3 kernel as well. Any help would be greatly appreciated.
> 
> I'm assuming you set the hardware sector size to 4k as well?
> 
> -- 
> Jens Axboe

Re: Alignment Issue with Direct IO to NVMe Drive

[Laine Walker-Avina]

On Tue, Nov 27, 2012 at 4:09 AM, Jens Axboe <axboe@kernel.dk> wrote:
> On 2012-11-27 01:35, Laine Walker-Avina wrote:
>> Hi all,
>>
>> We are experiencing an issue with doing direct IO to a NVMe device I'm
>> helping to develop. Every so often, the physical address given by
>> sg_dma_address() is aligned to 0x800 instead of 0x1000 as specified by
>> blk_queue_dma_alignement(queue, 4095) when the queue is initialized.
>> The request is also split over multiple segments to make up for the
>> missing space (eg: for a 4k IO it's split into two segments 2k in
>> size, and for an 8k IO it's split into 3 segments--2k,4k,2k). Our
>> design requires the physical segments given to the device be aligned
>> to 4k boundaries and be multiples of 4k in size. When not doing direct
>> IO the physical addresses appear to always be 4k aligned as expected.
>> One possible issue is the kernel we're primarily testing against is
>> 2.6.32-220 from CentOS, but we have observed similar behavior from a
>> vanilla 3.3 kernel as well. Any help would be greatly appreciated.
>
> I'm assuming you set the hardware sector size to 4k as well?
>
> --
> Jens Axboe
>

Yes, as well as the logical block size and io_min.

Laine Walker-Avina

Re: Alignment Issue with Direct IO to NVMe Drive

[Laine Walker-Avina]

On Tue, Nov 27, 2012 at 9:05 AM, Matthew Wilcox <willy@linux.intel.com> wrote:
> Laine, when this occurs, what is the alignment of 'offset' in the sg
> entry you're looking at?  If userspace is passing in an unaligned address,
> I don't think there's anything we do to try to align it.

I thought this was the case as well, but it appears to be aligned
looking at the request in the user space program (fio using libaio in
this case) before the IO is submitted.

Laine

Re: Alignment Issue with Direct IO to NVMe Drive

[Matthew Wilcox]

On Tue, Nov 27, 2012 at 09:47:48AM -0800, Laine Walker-Avina wrote:
> On Tue, Nov 27, 2012 at 9:05 AM, Matthew Wilcox <willy@linux.intel.com> wrote:
> > Laine, when this occurs, what is the alignment of 'offset' in the sg
> > entry you're looking at?  If userspace is passing in an unaligned address,
> > I don't think there's anything we do to try to align it.
> 
> I thought this was the case as well, but it appears to be aligned
> looking at the request in the user space program (fio using libaio in
> this case) before the IO is submitted.

OK, so we have an aligned virtual address being translated into an
unaligned DMA address.  I have a suspicion ... are you using the swiotlb,
or are you using a real IOMMU?  Grep dmesg for 'IOMMU' if you're not sure.

Re: Alignment Issue with Direct IO to NVMe Drive

[Laine Walker-Avina]

On Tue, Nov 27, 2012 at 1:39 PM, Matthew Wilcox <willy@linux.intel.com> wrote:
> On Tue, Nov 27, 2012 at 09:47:48AM -0800, Laine Walker-Avina wrote:
>> On Tue, Nov 27, 2012 at 9:05 AM, Matthew Wilcox <willy@linux.intel.com> wrote:
>> > Laine, when this occurs, what is the alignment of 'offset' in the sg
>> > entry you're looking at?  If userspace is passing in an unaligned address,
>> > I don't think there's anything we do to try to align it.
>>
>> I thought this was the case as well, but it appears to be aligned
>> looking at the request in the user space program (fio using libaio in
>> this case) before the IO is submitted.
>
> OK, so we have an aligned virtual address being translated into an
> unaligned DMA address.  I have a suspicion ... are you using the swiotlb,
> or are you using a real IOMMU?  Grep dmesg for 'IOMMU' if you're not sure.

There is no mention of IOMMU in dmesg on a fresh boot. The processor
in the test system I'm using is an i5-2400 (Sandy Bridge).

-Laine