[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] [PATCH 00/28] blk_end_request: full I/O completion handler (take 3)

Hello Jens,

The following is the updated patch-set for blk_end_request().
Changes since the last version are only minor updates to catch up
with the base kernel changes.
Do you agree the implementation of blk_end_request()?
If there's no problem, could you merge it to your tree?
Or does it have to be merged to -mm tree first?

Could you review the newly added PATCH 27 which converts the bidi part,
and give me your comments?
It uses blk_end_request_callback() in PATCH 25, which was only for
the tricky ide-cd driver.
If bidi added a 'resid' member to struct request instead of reusing
'data_len' for the other purpose, it could use the standard
blk_end_request() instead.

------------------ Changes from the previous post ---------------------
Changes between take2 and take3:
  o Rebased on top of 2.6.24-rc3-mm2
  o Added a bidi patch, which changes bidi to use blk_end_request()
    (PATCH 27)
  o Dropped blk_rq_size() which was to get size of entire request
    because rq_byte_size() has been added to ll_rw_blk.c (PATCH 02)
  o Removed 'dequeue' argument, which was added in 2.6.23-rc7-mm1,
    from __end_request() (PATCH 03)
  o Removed lguest patch because lguest has been changed not to use
    end_that_request_{chunk/last} directly.

Changes between take1 and take2:
  o Rebased on top of 2.6.23-rc4-mm1
  o Don't pass the lock held information (PATCH 01)
  o Removed sect2byte() macro (PATCH 02)
  o fixed blk_rq_size() and blk_rq_cur_size() for blk_pc_requests (PATCH 02)
  o Separated the patch for changes of end_that_request_* user (PATCH 03-26)
  o Removed the patch which changes the role of rq->end_io()
    from this patch-set because some more discussions are needed
    about it.

Summary of each patch are below:
  01/28: add new request completion interface, blk_end_request()
  02/28: add some functions to get the size of request in bytes
  03/28: convert to use blk_end_request() (core parts of block layer)
  04/28: convert to use blk_end_request() (arm)
  05/28: convert to use blk_end_request() (um)
  06/28: convert to use blk_end_request() (DAC960)
  07/28: convert to use blk_end_request() (floppy)
  08/28: convert to use blk_end_request() (nbd)
  09/28: convert to use blk_end_request() (ps3disk)
  10/28: convert to use blk_end_request() (sunvdc)
  11/28: convert to use blk_end_request() (sx8)
  12/28: convert to use blk_end_request() (ub)
  13/28: convert to use blk_end_request() (viodasd)
  14/28: convert to use blk_end_request() (xen-blkfront)
  15/28: convert to use blk_end_request() (viocd)
  16/28: convert to use blk_end_request() (i2o_block)
  17/28: convert to use blk_end_request() (mmc)
  18/28: convert to use blk_end_request() (s390)
  19/28: convert to use blk_end_request() (scsi mid-layer)
  20/28: convert to use blk_end_request() (ide-scsi)
  21/28: convert to use blk_end_request() (xsysace)
  22/28: convert to use blk_end_request() (cciss)
  23/28: convert to use blk_end_request() (cpqarray)
  24/28: convert to use blk_end_request() (normal parts of ide)
  25/28: add a valiant of blk_end_request() having callback feature
  26/28: convert to use blk_end_request() (ide-cd, cdrom_newpc_intr())
  27/28: convert to use blk_end_request() (scsi bidi)
  28/28: remove/unexport no longer needed end_that_request_*

I have tested this patch-set on two machines,
IA64+QLA1280+QLA2200 box and x86_64+SATA+IDE-CDROM box.

Below is the explanation about needs and details of this patch-set.

This set of patches changes request completion interface
between device drivers and block layer to 1 step procedure
from current 2 step procedures using end_that_request_{first/chunk}
and end_that_request_last().

This patch-set prepares for realizing another patch-set which changes
the role of rq->end_io().  It allows request-based multipath to hook
in before completing each chunk of request, check errors for it and
retry it using another path if error is detected.

The patch-set which changes the role of rq->end_io() is necessary
to allow device stacking at request level, that is request-based
device-mapper multipath.
Currently, device-mapper is implemented as a stacking block device
at BIO level.  OTOH, request-based dm will stack at request level
to allow better multipathing decision.
To allow device stacking at request level, the completion procedure
need to provide a hook for it.
For example, dm-multipath has to check errors and retry with other
paths if necessary before returning the I/O result to upper layer.
struct request has 'end_io' hook currently.  But it's called at
the very late stage of completion handling where the I/O result
is already returned to the upper layer.
So we need something here.

The first approach to hook in completion of each chunk of request
was adding a new rq->end_io_first() hook and calling it on the top
of __end_that_request_first().
  - http://marc.theaimsgroup.com/?l=linux-scsi&m=115520444515914&w=2
  - http://marc.theaimsgroup.com/?l=linux-kernel&m=116656637425880&w=2
However, Jens pointed out that redesigning rq->end_io() as a full
completion handler would be better:

On Thu, 21 Dec 2006 08:49:47 +0100, Jens Axboe <jens axboe oracle com> wrote:
> Ok, I see what you are getting at. The current ->end_io() is called when
> the request has fully completed, you want notification for each chunk
> potentially completed.
> I think a better design here would be to use ->end_io() as the full
> completion handler, similar to how bio->bi_end_io() works. A request
> originating from __make_request() would set something ala:
> instead of calling the functions manually. That would allow you to get
> notification right at the beginning and do what you need, without adding
> a special hook for this.

I thought his comment was reasonable.
So I modified the patches based on his suggestion.

The change is basically illustlated by the following pseudo code:

  if (end_that_request_{first/chunk} succeeds) { <-- completes bios
     <do something driver specific>
     end_that_request_last() <-- calls end_io()
     <the request is free from the driver>
  } else {
     <the request was incomplete, retry for leftover or ignoring>

  if (blk_end_request() succeeds) { <-- calls end_io(), completes bios
     <the request is free from the driver>
  } else {
     <the request was incomplete, retry for leftover or ignoring>

In detail, request completion procedures are changed like below.

  o 2 steps completion using end_that_request_{first/chunk}
    and end_that_request_last().
  o Device drivers have ownership of a request until they
    call end_that_request_last().
  o rq->end_io() is called at the last stage of
    end_that_request_last() for some block layer codes need
    specific request handling when completing it.

  o 1 step completion using blk_end_request().
    (end_that_request_* are no longer used from device drivers.)
  o Device drivers give over ownership of a request
    when calling blk_end_request().
    If it returns 0, the request is completed.
    If it returns 1, the request isn't completed and
    the ownership is returned to the device driver again.
  o rq->end_io() is called at the top of blk_end_request() to
    allow to hook all parts of request completion.
    Existing users of rq->end_io() must be changed to do
    all parts of request completion.

Kiyoshi Ueda

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]