On 1/8/19 1:50 PM, Logan Gunthorpe wrote:
> scsi_mq_setup_tags(), which is called by scsi_add_host(), calculates
> the command size to allocate based on the prot_capabilities. In the
> isci driver, scsi_host_set_prot() is called after scsi_add_host()
> so the command size gets calculated to be smaller than it needs to be.
> Eventually, scsi_mq_init_request() locates the 'prot_sdb' after the command
> assuming it was sized correctly and a buffer overrun may occur.
>
> However, seeing blk_mq_alloc_rqs() rounds up to the nearest cache line
> size, the mistake can go unnoticed.
>
> The bug was noticed after the struct request size was reduced by
> commit 9d037ad707ed ("block: remove req->timeout_list")
>
> Which likely reduced the allocated space for the request by an entire
> cache line, enough that the overflow could be hit and it caused a panic,
> on boot, at:
>
> RIP: 0010:t10_pi_complete+0x77/0x1c0
> Call Trace:
> <IRQ>
> sd_done+0xf5/0x340
> scsi_finish_command+0xc3/0x120
> blk_done_softirq+0x83/0xb0
> __do_softirq+0xa1/0x2e6
> irq_exit+0xbc/0xd0
> call_function_single_interrupt+0xf/0x20
> </IRQ>
>
> sd_done() would call scsi_prot_sg_count() which reads the number of
> entities in 'prot_sdb', but seeing 'prot_sdb' is located after the end of
> the allocated space it reads a garbage number and erroneously calls
> t10_pi_complete().
>
> To prevent this, the calls to scsi_host_set_prot() are moved into
> isci_host_alloc() before the call to scsi_add_host(). Out of caution,
> also move the similar call to scsi_host_set_guard().
Nice work!
Reviewed-by: Jens Axboe <axboe@xxxxxxxxx>
--
Jens Axboe
