lkml.org 
[lkml]   [2013]   [Mar]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [patch] [SCSI] scsi_transport_sas: check for allocation failure
From
Date
On Mon, 2013-03-11 at 10:48 -0400, Douglas Gilbert wrote:
> On 13-03-11 09:10 AM, Dan Carpenter wrote:
> > On Fri, Mar 08, 2013 at 10:50:19PM +0000, James Bottomley wrote:
> >> On Fri, 2013-03-08 at 12:57 -0500, Douglas Gilbert wrote:
> >>> On 13-03-08 07:02 AM, Dan Carpenter wrote:
> >>>> Static checkers complain that this allocation isn't checked. We
> >>>> should return zero if the allocation fails.
> >>>>
> >>>> Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
> >>>>
> >>>> diff --git a/drivers/scsi/scsi_transport_sas.c b/drivers/scsi/scsi_transport_sas.c
> >>>> index 1b68142..a022997 100644
> >>>> --- a/drivers/scsi/scsi_transport_sas.c
> >>>> +++ b/drivers/scsi/scsi_transport_sas.c
> >>>> @@ -379,9 +379,12 @@ sas_tlr_supported(struct scsi_device *sdev)
> >>>> {
> >>>> const int vpd_len = 32;
> >>>> struct sas_end_device *rdev = sas_sdev_to_rdev(sdev);
> >>>> - char *buffer = kzalloc(vpd_len, GFP_KERNEL);
> >>>> + char *buffer;
> >>>> int ret = 0;
> >>>>
> >>>> + buffer = kzalloc(vpd_len, GFP_KERNEL);
> >>>> + if (!buffer)
> >>>> + goto out;
> >>>> if (scsi_get_vpd_page(sdev, 0x90, buffer, vpd_len))
> >>>> goto out;
> >>>>
> >>>
> >>> For 32 bytes, why not use the stack?
> >>
> >> Because the buffer is a DMA target. You can't DMA to stack because of
> >> padding and cacheline issues.
> >>
> >
> > I think stack data works here. scsi_execute() calls
> > blk_rq_map_kern() which handles stack memory and alignment issues.
>
> That being the case, several other callers of
> scsi_get_vpd_page() 9and friends) could be
> simplified and sped up.

Um, I'm not sure you'll speed stuff up by doing this: have you seen what
bio_copy_kern() does? it will allocate a page per iov entry (in this
case probably only a single entry) and copy the data into and out of
that page, so not only do we get the copy overhead. We also get page
allocation overheads instead of the small number of bytes we might need,
which could cause quite a stall in a low memory situation.

I'm not saying don't do it, I'm just saying think of the consequences
over the fiddle of doing a small kmalloc.

> Also since VPD pages don't change and they can carry
> a lot of disparate information (e.g. the Extended
> Inquiry and Block Limits pages) perhaps they could
> be cached by the appropriate level (e.g. Extended
> Inquiry cached by mid-level; Block Limits cached
> by sd driver).

We cache the ones we care about, but we could always add more if there's
enough call, I suppose.

James



\
 
 \ /
  Last update: 2013-03-11 17:01    [W:0.078 / U:0.080 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site