[V4,00/16] use sg helper to operate scatterlist

Message

Ming Lei June 17, 2019, 3:03 a.m. UTC

Hi,

Scsi MQ makes a large static allocation for the first scatter gather
list chunk for the driver to use.  This is a performance headache we'd
like to fix by reducing the size of the allocation to a 2 element
array.  Doing this will break the current guarantee that any driver
using SG_ALL doesn't need to use the scatterlist iterators and can get
away with directly dereferencing the array.  Thus we need to update all
drivers to use the scatterlist iterators and remove direct indexing of
the scatterlist array before reducing the initial scatterlist
allocation size in SCSI.

So convert drivers to use scatterlist helper.

There are two types of direct access on scatterlist in SCSI drivers:

1) operate on the scatterlist via scsi_sglist(scmd) directly, then one
local variable of 'struct scatterlist *' is involved.

2) scsi_sglist(scmd) is stored to cmd->SCp.buffer and the scatterlist is
used via cmd->SCp.buffer.

The following coccinelle semantic patch is developed for finding the
above two types of direct scatterlist uses:

	@@ struct scatterlist *p; @@
	(
	- ++p
	+ p = sg_next(p)
	|
	- p++
	+ p = sg_next(p)
	|
	- p = p + 1
	+ p = sg_next(p)
	|
	- p += 1
	+ p = sg_next(p)
	|
	- --p
	+ p = sg_non_exist_prev(p)
	|
	- p--
	+ p = sg_non_exist_prev(p)
	|
	- p = p - 1
	+ p = sg_non_exist_prev(p)
	|
	- p -= 1
	+ p = sg_non_exist_prev(p)
	)
	
	@@
	struct scatterlist *p;
	expression data != 0;
	@@
	- p[data]
	+ '!!!!!!use sg iterator helper!!!!!!'
	
	@@
	struct scatterlist[] p;
	expression data != 0;
	@@
	- p[data]
	+ '!!!!!!use sg iterator helper!!!!!!'
	
	
	@@ struct scsi_cmnd *scmd; @@
	(
	-	scmd->SCp.buffer++
	+	scmd->SCp.buffer = sg_next(scmd->SCp.buffer)
	|
	-	++scmd->SCp.buffer
	+	scmd->SCp.buffer = sg_next(scmd->SCp.buffer)
	|
	-	scmd->SCp.buffer += 1
	+	scmd->SCp.buffer = sg_next(scmd->SCp.buffer)
	|
	-	scmd->SCp.buffer = scmd->SCp.buffer + 1
	+	scmd->SCp.buffer = sg_next(scmd->SCp.buffer)
	|
	-	scmd->SCp.buffer--
	+	scmd->SCp.buffer = sg_no_exit_prev(scmd->SCp.buffer)
	|
	-	--scmd->SCp.buffer
	+	scmd->SCp.buffer = sg_no_exit_prev(scmd->SCp.buffer)
	|
	-	scmd->SCp.buffer -= 1
	+	scmd->SCp.buffer = sg_no_exit_prev(scmd->SCp.buffer)
	|
	-	scmd->SCp.buffer = scmd->SCp.buffer - 1
	+	scmd->SCp.buffer = sg_no_exit_prev(scmd->SCp.buffer)
	)
	
	@@
	struct scsi_cmnd *scmd;
	expression data != 0;
	@@
	- scmd->SCp.buffer[data]
	+ '!!!!!!use sg iterator helper!!!!!!'


The 1st 10 patches are for handling type #1, and the other 6 patches
for handling type #2, and all the 16 are found by the above coccinelle
semantic patch.

V4:
	- fix building failure on pmcraid's conversion 
	- improve the coccinelle semantic patch to cover both two types of
	scatterlist direct use
	- driver 'staging: rtsx' is covered

V3:
	- update commit log and cover letter, most of words are from
	James Bottomley	

V2:
	- use coccinelle semantic patch for finding direct sgl uses from
	scsi command(9 drivers found)
	- run 'git grep -E "SCp.buffer"' to find direct sgl uses
	from SCp.buffer(6 drivers are found)


Finn Thain (1):
  NCR5380: Support chained sg lists

Ming Lei (15):
  scsi: vmw_pscsi: use sg helper to operate scatterlist
  scsi: advansys: use sg helper to operate scatterlist
  scsi: lpfc: use sg helper to operate scatterlist
  scsi: mvumi: use sg helper to operate scatterlist
  scsi: ipr: use sg helper to operate scatterlist
  scsi: pmcraid: use sg helper to operate scatterlist
  usb: image: microtek: use sg helper to operate scatterlist
  staging: unisys: visorhba: use sg helper to operate scatterlist
  staging: rtsx: use sg helper to operate scatterlist
  s390: zfcp_fc: use sg helper to operate scatterlist
  scsi: aha152x: use sg helper to operate scatterlist
  scsi: imm: use sg helper to operate scatterlist
  scsi: pcmcia: nsp_cs: use sg helper to operate scatterlist
  scsi: ppa: use sg helper to operate scatterlist
  scsi: wd33c93: use sg helper to operate scatterlist

 drivers/s390/scsi/zfcp_fc.c                   |  4 +-
 drivers/scsi/NCR5380.c                        | 41 ++++++++----------
 drivers/scsi/advansys.c                       |  2 +-
 drivers/scsi/aha152x.c                        | 42 +++++++++----------
 drivers/scsi/imm.c                            |  2 +-
 drivers/scsi/ipr.c                            | 28 +++++++------
 drivers/scsi/lpfc/lpfc_nvmet.c                |  3 +-
 drivers/scsi/mvumi.c                          |  9 ++--
 drivers/scsi/pcmcia/nsp_cs.c                  |  4 +-
 drivers/scsi/pmcraid.c                        | 14 +++----
 drivers/scsi/ppa.c                            |  2 +-
 drivers/scsi/vmw_pvscsi.c                     |  2 +-
 drivers/scsi/wd33c93.c                        |  2 +-
 drivers/staging/rts5208/rtsx_transport.c      |  4 +-
 .../staging/unisys/visorhba/visorhba_main.c   |  9 ++--
 drivers/usb/image/microtek.c                  | 20 ++++-----
 drivers/usb/image/microtek.h                  |  2 +-
 17 files changed, 90 insertions(+), 100 deletions(-)

Comments

Bart Van Assche June 17, 2019, 11:52 p.m. UTC | #1

On 6/16/19 8:03 PM, Ming Lei wrote:
> Scsi MQ makes a large static allocation for the first scatter gather
> list chunk for the driver to use.  This is a performance headache we'd
> like to fix by reducing the size of the allocation to a 2 element
> array.  Doing this will break the current guarantee that any driver
> using SG_ALL doesn't need to use the scatterlist iterators and can get
> away with directly dereferencing the array.  Thus we need to update all
> drivers to use the scatterlist iterators and remove direct indexing of
> the scatterlist array before reducing the initial scatterlist
> allocation size in SCSI.

For all patches in this series except 9/16 and 11/16, please add the 
following:

Reviewed-by: Bart Van Assche <[e-mail]>