nr_pages calculation in scsi_req_map_sg()

nr_pages calculation in scsi_req_map_sg()

[Jan Beulich]

James,

while reviewing code derived from that function I found this calculation
to be suspicious: I would think that it should get it wrong when both
start and end of the buffer area are misaligned (e.g. consider the case
where sgl->offset equals PAGE_SIZE-1 and bufflen equals 2 - the result
would be 1 when it should have been 2).
Is there something preventing this from happening?

Thanks, Jan

Re: nr_pages calculation in scsi_req_map_sg()

[Boaz Harrosh]

Jan Beulich wrote:
> James,
> 
> while reviewing code derived from that function I found this calculation
> to be suspicious: I would think that it should get it wrong when both
> start and end of the buffer area are misaligned (e.g. consider the case
> where sgl->offset equals PAGE_SIZE-1 and bufflen equals 2 - the result
> would be 1 when it should have been 2).
> Is there something preventing this from happening?
> 
> Thanks, Jan
> 
> --

It has been discussed before for example look here:
http://www.spinics.net/lists/linux-scsi/msg13454.html

But for me the main reason it is not fixed is because
this is only called from scsi_execute_async() which
is a deprecated function. It is still used by old code
which is supposed to be removed soon. Any new code will
not be accepted if it uses scsi_execute_async().

Boaz

Re: nr_pages calculation in scsi_req_map_sg()

[Jan Beulich]

>>> Boaz Harrosh <bharrosh@panasas.com> 22.07.08 15:28 >>>
>Jan Beulich wrote:
>> James,
>> 
>> while reviewing code derived from that function I found this calculation
>> to be suspicious: I would think that it should get it wrong when both
>> start and end of the buffer area are misaligned (e.g. consider the case
>> where sgl->offset equals PAGE_SIZE-1 and bufflen equals 2 - the result
>> would be 1 when it should have been 2).
>> Is there something preventing this from happening?
>> 
>> Thanks, Jan
>> 
>> --
>
>It has been discussed before for example look here:
>http://www.spinics.net/lists/linux-scsi/msg13454.html 
>
>But for me the main reason it is not fixed is because
>this is only called from scsi_execute_async() which
>is a deprecated function. It is still used by old code
>which is supposed to be removed soon. Any new code will
>not be accepted if it uses scsi_execute_async().

No, that's a different issue: Even if the sg elements are all contiguous,
the count can be wrong, as described in the original mail. And as said,
I found this in code cloned from scsi_req_map_sg(), hence would be
interested in confirmation of that fact (or explanation why it's not an
issue) regardless of the function itself sitting in a to-be-removed code
path only.

Thanks, Jan

Re: nr_pages calculation in scsi_req_map_sg()

[Boaz Harrosh]

Jan Beulich wrote:
>>>> Boaz Harrosh <bharrosh@panasas.com> 22.07.08 15:28 >>>
>> Jan Beulich wrote:
>>> James,
>>>
>>> while reviewing code derived from that function I found this calculation
>>> to be suspicious: I would think that it should get it wrong when both
>>> start and end of the buffer area are misaligned (e.g. consider the case
>>> where sgl->offset equals PAGE_SIZE-1 and bufflen equals 2 - the result
>>> would be 1 when it should have been 2).
>>> Is there something preventing this from happening?
>>>
>>> Thanks, Jan
>>>
>>> --
>> It has been discussed before for example look here:
>> http://www.spinics.net/lists/linux-scsi/msg13454.html 
>>
>> But for me the main reason it is not fixed is because
>> this is only called from scsi_execute_async() which
>> is a deprecated function. It is still used by old code
>> which is supposed to be removed soon. Any new code will
>> not be accepted if it uses scsi_execute_async().
> 
> No, that's a different issue: Even if the sg elements are all contiguous,
> the count can be wrong, as described in the original mail. And as said,
> I found this in code cloned from scsi_req_map_sg(), hence would be
> interested in confirmation of that fact (or explanation why it's not an
> issue) regardless of the function itself sitting in a to-be-removed code
> path only.
> 
> Thanks, Jan
> 
lets write it like this:
nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) / PAGE_SIZE;

now you say:
nr_pages = (2 + (PAGE_SIZE-1) + PAGE_SIZE - 1) / PAGE_SIZE;

which is (PAGE_SIZE + PAGE_SIZE) / PAGE_SIZE; No?

What am I missing?

Boaz