Re: SCSI bug

[Helge Deller <deller@gmx.de>]

On 21.02.2016 20:07, James Bottomley wrote:
> On Sun, 2016-02-21 at 13:43 -0500, John David Anglin wrote:
>> I verified that commit 54efd50bfd873e2dbf784e0b21a8027ba4299a3e in
>> linux-block fails to
>> boot and commit 41609892701e26724b8617201f43254cadf2e7ae (blk-cgroup:
>> Drop unlikely
>> before IS_ERR(_OR_NULL)) does boot successfully.  Commit
>> 41609892701e26724b8617201f43254cadf2e7ae
>> is previous commit in tree.
>>
>> I don't believe that the change can be reverted from Linus' tree as
>> this commit allowed other
>> stuff to be removed (see second paragraph of commit description).
> 
> OK, can you just verify you can boot 4.5-rc5 without the sata_sil24
> driver?

I tried it on my c3000, debian kernel 4.4.2, in this case without the pata_ns87415.
I used "modprobe.blacklist=libata,pata_ns87415,ata_generic" as additional 
kernel command line option. 

[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 12288, nbytes = 4096, sum = 16384 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 65536, nbytes = 4096, sum = 69632 > 65536

^^ here.

[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 1, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 12288, nbytes = 4096, sum = 16384 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] NEW SEGMENT sg = 00000000bbe991e8!!!
[   45.980000] __blk_segment_map_sg: length = 4096, nbytes = 4096, sum = 8192 > 65536
[   45.980000] __blk_segment_map_sg: BIOVEC_PHYS_MERGEABLE = 0, BIOVEC_SEG_BOUNDARY = 1
[   45.980000] *** FIXIT *** HELGE: nsegs > rq->nr_phys_segments = 11 > 10
[   45.980000] scsi_init_sgtable: count = 11, nents = 10

^^ here

[   45.980000] timer_interrupt(CPU 0): delayed! cycles 73C0F821 rem 1E1DDF  next/now 1D78B3071C/1D7894E93D
[   48.684000] ------------[ cut here ]------------
[   48.740000] WARNING: at /build/linux-4.4/linux-4.4.2/drivers/scsi/scsi_lib.c:1104
[   48.828000] Modules linked in: sd_mod ohci_pci ehci_pci ohci_hcd ehci_hcd sym53c8xx scsi_transport_spi usbcore scsi_mod usb_common tulip
[   48.976000] CPU: 0 PID: 66 Comm: systemd-udevd Not tainted 4.4.0-1-parisc64-smp #5 Debian 4.4.2-2
[   49.084000] task: 00000000bbe97538 ti: 00000000bbe98000 task.ti: 00000000bbe98000

 
> My theory, based on what Helge produced is that this commit is building
> a large transfer list > 65535 and then splitting it wrongly.  

Yes, this theory sounds right.

> I think
> the problem is that it's not respecting the DMA boundary, so Helge sees
> a transfer size of 69632 which I think slops over on both sides,
> requiring 3 segments to describe.  However, the merge code thinks we
> only need two (because the length is < 2*65536).  The reason we only
> see this with ATA drivers is because virtually no SCSI drivers set the
> DMA boundary.  Most ATA drivers require a dma boundary of 65535.

In my test above there is no ATA driver included, only SCSI discs.

Helge