About SG Elements of request_buffer

About SG Elements of request_buffer

[Aboo Valappil]

Hi All,

Thanks for all your support for my previous posts.

I have two questions.

1. Is it safe to assume that all the SG elements of request_buffer are 
page aligned? (Except the first one and last one),  I mean, it is always 
multiple of pages?

2. What are situations reading a scsi disk (using dd ) cause a write to 
the disk?  When I read my virtual scsi disk using dd command, block 
layer is generating lot of SCSI write, as well as reads. Any thoughts? 
Besides this observation, everything is working fine.

Thanks in advance.

Aboo

Re: About SG Elements of request_buffer

[Douglas Gilbert]

Aboo Valappil wrote:
> Hi All,
> 
> Thanks for all your support for my previous posts.
> 
> I have two questions.
> 
> 1. Is it safe to assume that all the SG elements of request_buffer are
> page aligned? (Except the first one and last one),  I mean, it is always
> multiple of pages?
> 
> 2. What are situations reading a scsi disk (using dd ) cause a write to
> the disk?  When I read my virtual scsi disk using dd command, block
> layer is generating lot of SCSI write, as well as reads. Any thoughts?
> Besides this observation, everything is working fine.

Aboo,
A comment on your second point: to bypass the vagaries of
the block layer caching with dd (or sg_dd) use 'iflag=direct'
and/or 'oflag=direct' as required. There may be alignment
constraints.

Doug Gilbert

Re: About SG Elements of request_buffer

[James Bottomley]

On Sun, 2006-10-15 at 22:03 +1000, Aboo Valappil wrote:
> 1. Is it safe to assume that all the SG elements of request_buffer are 
> page aligned? (Except the first one and last one),  I mean, it is always 
> multiple of pages?

No.

It's an artifact of block sector size >= PAGE_SIZE and the access method
only using pages.  It isn't even true of ext3 which does its metadata
through the buffer cache.

James

Re: About SG Elements of request_buffer

[Aboo Valappil]

[-- Attachment #1: Type: text/plain, Size: 2895 bytes --]

Hi All,

I have posted this a while ago on strange behavior I see when "dd" opens 
and reads the disk presented by my virtual HBA. I am attaching the code 
in this email.

Basically, I have this virtual HBA which is being served by a user space 
SCSI engine. It basically transports SCSI commands out side kernel and 
get serviced by a user space scsi engine.

I have successfully implemented this using read and write system call to 
a character device. So the SCSI request buffer is copied back and forth 
user and kernel space for every request. It works alright.

Now i wanted a method of sharing SCSI request_buffer between user space 
and kernel without doing copy_to_user and copy_from_user of the 
request_buffer. I have implemented memmap and facing some challenges. I 
am memory mapping the scsi request_buffer to the user space. But i am 
facing a problem with page cache.

The SCSI device presented by my HBA is /dev/sdb (for eg:). If I open 
this device and read a block ( 512 bytes). I can see a SCSI read coming 
down my HBA. But I also a see a SCSI write later on when I close 
/dev/sdb. Up on investigation I found that it is the effect of page 
cache. When /dev/sdb opened and initiated a read, a page cache is 
already allocated. When I do memmap, I am actually mem mapping that page 
buffer to my user space SCSI engine. When the SCSI engine service the 
SCSI read, it writes the SCSI READ output to the buffer making it dirty. 
When the application  closed /dev/sdb, the kernel causes all the dirty 
pages to be written and causing the SCSI write which is what I am 
seeing. This means that for every SCSI read, i will see a SCSI write.It 
is a very strange implementation from my side. Even Xen source is doing 
the same memory mapping technique and I do not know how it avoid this 
page cache interference.

Can some one throw some lights on this for me to prevent this page cache 
interference? I have done bit of research on marking the pages/buffer 
clean after writing to the memory mapped area. But I could not find 
anything conclusive.

Thanks in advance,

Aboo


Aboo Valappil wrote:
> Hi All,
>
> Thanks for all your support for my previous posts.
>
> I have two questions.
>
> 1. Is it safe to assume that all the SG elements of request_buffer are 
> page aligned? (Except the first one and last one),  I mean, it is 
> always multiple of pages?
>
> 2. What are situations reading a scsi disk (using dd ) cause a write 
> to the disk?  When I read my virtual scsi disk using dd command, block 
> layer is generating lot of SCSI write, as well as reads. Any thoughts? 
> Besides this observation, everything is working fine.
>
> Thanks in advance.
>
> Aboo
>
>
> -
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html


[-- Attachment #2: scsitap.tgz --]
[-- Type: application/octet-stream, Size: 4755 bytes --]