BUG_ON at nvme_setup_prp getting hit.

[Keith Busch]

On Tue, 7 Jan 2014, Hrishikesh Gokhale wrote:
> Hi All,
>
> Requesting for help in resolving the following:
>
> Setup:
> Running FIO tester tool on NVMe device driver (/block/nvme.c) with I/O
> engine “libaio”. Using Odd block size like 17K, 129K, 257K,  etc.
> Distribution ‘Fedora 18’ (Kernel v3.6.10).
>
> Observation:
> I’ve observed intermittent hit on BUG_ON which is pointing to
> nvme_setup_prps() function in nvme.c driver.
>
>>>> CODE
> static int nvme_setup_prps(...)
>
> {
>       int dma_len = sg_dma_len(sg);
>
>       u64 dma_addr = sg_dma_address(sg);
>
>       int offset = offset_in_page(dma_addr);
>       ...
>       dma_len -= PAGE_SIZE;
>       ...
>       if (dma_len > 0)
>           continue;
>       BUG_ON(dma_len < 0);
> }
> <<< CODE
>
> After adding prints , I found that this problem is reproduced on use
> of PRP list.
>
> Offset value in DMA page is changed (in nvme_setup_prps() function)
> from the original offset value which was present while sg_set_page()
> function was executed. The offset value of virtual address is found
> different than the offset value of physical/DMA address.

This BUG_ON shouldn't happen unless we gave nvme_setup_prps an sgl that
can't be mapped to a PRP list, so the request should have been split
instead. A virtual address with different alignment than the dma address
might cause that and sounds similar to this issue from way back:

merlin.infradead.org/pipermail/linux-nvme/2012-November/000121.html

I didn't see that anything came of it, though.

Maybe our BIOVEC_NOT_VIRT_MERGEABLE check should use the physical
addresses instead of virtual to see if the bio_vec is usable. ?

> Query:
> Is it possible to avoid the ‘offset value change’ in DMA-able memory
> (in nvme_setup_prps() function) to avoid BUG_ON()check? ”

> Thanks & Regards,
> Hrishikesh