Re: [PATCH] scsi/hpsa: fix to check dma mapping error

[scameron@beardog.cce.hp.com]

On Wed, Feb 13, 2013 at 03:55:55PM -0700, Shuah Khan wrote:
> Add missing dma mapping error check to fix the following warning from
> dma-debug.
> 
> [   16.614739] hpsa 0000:03:00.0: DMA-API: device driver failed to check map error[device address=0x00000000001c2ec0] [size=64 bytes] [mapped as single]
> [   16.614833] Modules linked in: pata_atiixp tg3(+) ptp pps_core hpsa(+)
> [   16.615381] Pid: 324, comm: modprobe Not tainted 3.8.0-rc7+ #6
> [   16.615466] Call Trace:
> [   16.615552]  [<ffffffff810581ef>] warn_slowpath_common+0x7f/0xc0
> [   16.615639]  [<ffffffff810582e6>] warn_slowpath_fmt+0x46/0x50
> [   16.615747]  [<ffffffff8134c129>] check_unmap+0x459/0x8a0
> [   16.615832]  [<ffffffff81698aad>] ? schedule_timeout+0x1ed/0x250
> [   16.615918]  [<ffffffff8134c70c>] debug_dma_unmap_page+0x5c/0x60
> [   16.616008]  [<ffffffffa0002776>] hpsa_pci_unmap+0x86/0xd0 [hpsa]
> [   16.616094]  [<ffffffffa00028c0>] hpsa_scsi_do_simple_cmd_with_retry+0x100/0x2a0 [hpsa]
> [   16.616195]  [<ffffffffa0004863>] hpsa_scsi_do_inquiry+0x83/0x100 [hpsa]
> [   16.616282]  [<ffffffffa0006ce9>] hpsa_init_one+0x1839/0x1ce0 [hpsa]
> [   16.616369]  [<ffffffff811fc2d8>] ? sysfs_new_dirent+0x58/0x140
> [   16.616456]  [<ffffffff8135db9b>] local_pci_probe+0x4b/0x80
> [   16.616542]  [<ffffffff8135f471>] pci_device_probe+0x101/0x120
> [   16.616629]  [<ffffffff814328db>] driver_probe_device+0x7b/0x240
> [   16.616715]  [<ffffffff81432b4b>] __driver_attach+0xab/0xb0
> [   16.616801]  [<ffffffff81432aa0>] ? driver_probe_device+0x240/0x240
> [   16.616888]  [<ffffffff81430cf6>] bus_for_each_dev+0x56/0x90
> [   16.617001]  [<ffffffff8143240e>] driver_attach+0x1e/0x20
> [   16.617087]  [<ffffffff81431f80>] bus_add_driver+0x190/0x290
> [   16.617173]  [<ffffffffa0011000>] ? 0xffffffffa0010fff
> [   16.617259]  [<ffffffff814330aa>] driver_register+0x7a/0x160
> [   16.617345]  [<ffffffffa0011000>] ? 0xffffffffa0010fff
> [   16.617431]  [<ffffffff8135e42c>] __pci_register_driver+0x4c/0x50
> [   16.617519]  [<ffffffffa001101e>] hpsa_init+0x1e/0x1000 [hpsa]
> [   16.617605]  [<ffffffff8100206f>] do_one_initcall+0x3f/0x170
> [   16.617692]  [<ffffffff810bef7e>] load_module+0x16ae/0x1c40
> [   16.617777]  [<ffffffff810bbb70>] ? show_initstate+0x50/0x50
> [   16.617863]  [<ffffffff8169f71e>] ? do_page_fault+0xe/0x10
> [   16.617949]  [<ffffffff810bf5de>] sys_init_module+0xce/0x100
> [   16.618035]  [<ffffffff816a3d99>] system_call_fastpath+0x16/0x1b
> [   16.618157] ---[ end trace 260311c4be71d0dc ]---
> [   16.618228] Mapped at:
> [   16.618311]  [<ffffffff8134b189>] debug_dma_map_page+0xb9/0x160
> [   16.618451]  [<ffffffffa0004132>] fill_cmd.isra.31+0x252/0x400 [hpsa]
> [   16.618613]  [<ffffffffa0004853>] hpsa_scsi_do_inquiry+0x73/0x100 [hpsa]
> [   16.618766]  [<ffffffffa0006ce9>] hpsa_init_one+0x1839/0x1ce0 [hpsa]
> [   16.618919]  [<ffffffff8135db9b>] local_pci_probe+0x4b/0x80
> [   16.619286] scsi3 : hpsa
> 
> 
> Signed-off-by: Shuah Khan <shuah.khan@hp.com>
> CC: stable@vger.kernel.org
> ---
>  drivers/scsi/hpsa.c |    5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/drivers/scsi/hpsa.c b/drivers/scsi/hpsa.c
> index 4f33806..3b4d195 100644
> --- a/drivers/scsi/hpsa.c
> +++ b/drivers/scsi/hpsa.c
> @@ -1405,6 +1405,11 @@ static void hpsa_map_one(struct pci_dev *pdev,
>  	}
>  
>  	addr64 = (u64) pci_map_single(pdev, buf, buflen, data_direction);
> +	if (dma_mapping_error(&pdev->dev, addr64)) {
> +		cp->Header.SGList = 0;
> +		cp->Header.SGTotal = 0;
> +		return;
> +	}
>  	cp->SG[0].Addr.lower =
>  	  (u32) (addr64 & (u64) 0x00000000FFFFFFFF);
>  	cp->SG[0].Addr.upper =

Probably should bubble that error up through hpsa_map_one return value
(currently void, but I guess that needs to change) and that's called from
one place, fill_cmd(), also currently returning void (needs to change).
I think all the places that call fill_cmd() have failure paths that
can be used.

There are a couple other places in hpsa that we call pci_map_single and do
not subsequently call dma_mapping_error().

I don't thinking just setting SGList = 0 and SGTotal = 0 is going to 
do something good (I expect it will end up causing the cnotroller to 
reject the command as invalid -- and pretty much all the commands that
go through this path are driver generated commands, so, if the driver
is generating invalid commands, that would be a bug.

So, I don't think this patch is quite correct as is.

But, thanks for bringing this failure to call dma_mapping_error to my
attention.

So what cmdline options or config options did you enable to get this check?

was it something like:

dma_debug=1 dma_debug_entries=???

-- steve