Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> johan@capvert.se wrote:
>> I, and I suppose a lot of other people, would like know how you look 
>> at the
>> HW support on the Fc side? I don´t know what you mean now, when you talk
>> about HW target support, 
> 
> For qlogic FC, it is the same as scst at this point, except we are based 
> off the mainline qla2xxx driver.

Actually, this is not completely true and could mislead people. Stgt is 
*not* the same as scst at this point and, I think will not be for at 
least considerable amount of time.

They are the same only on a basic subset of functionality, which I would 
call "fast path", and only for block devices that are could be 
considered stateless, i.e. on which the result of a SCSI command's 
execution isn't dependent on the current state of the device. For, eg, 
tapes this isn't true, because on such devices the current state, like 
block size, changes the result of the command a lot, so it shall be 
honored and special handling of Unit Attention conditions shall be done. 
For example, all UAs from a device shall be delivered to all connected 
initiators, not only to one of them who happens to execute the command 
with UA result. Another example is when one initiator changes state of 
the device. After that all other initiators shall receive appropriate 
UA, i.e. the mid-layer shall generate it, because for the device all 
initiators act as one initiator (nexus) and the device is not able to 
distinguish between them to perform all necessary SCSI handling. Thus, 
the mid-layer has to do it. Not doing so is dangerous and could lead to 
data corruption and loss. The same is true for "advanced" commands like 
RESERVE/RELEASE, which also have to be "emulated" by the mid-layer. 
Also, currently stgt doesn't care much about task management.

Scst from the very beginning was targeted for SCSI tapes exported via 
hardware targets and was designed with all that very complicated staff 
in mind. Most of scst complexity is caused by handling of it as well as 
by attempt to make that handling at most performance effective way. For 
example, on the fast path, no task management and UA-related locks are 
taken, although this is done on the way, which could be considered a bit 
  unusual or extravagant, but effective.

Thus, to resume, the following important features are missed in stgt 
comparing to scst:

  - Task management

  - SCSI handling/emulation required for statefull SCSI devices (tapes, 
etc.)

  - Scst has some performance advantages over stgt, at least, on 
hardware targets, because it allows internal handling in SIRQ context as 
well as doesn't need user space program, so it eliminates additional 
context switches (at least 3 per command for WRITEs and 2 per command 
for reads plus switches to user space daemon, probably, 2 per command). 
5 context switches per command looks too much for me, especially 
considering how little work is done on each context. It means ~15000 
CS/sec on regular 200Mb/sec link with 64K block size. Additionally, 
kernel only commands execution allows direct access to page cache, which 
is GREAT performance improvement, because in architecture with user 
space daemon the data should be copied several times between kernel and 
user space. Or, do I miss anything?

  - Access and devices (LUNs) visibility management. It allows for an 
initiator or group of initiators to have different set of LUs, each with 
appropriate access permissions. This feature provides HUGE usability, 
people who tried it will confirm that.

  - Support for most SCSI devices, namely tapes, processors (SCSI type 
3), CDROM's (SCSI type 5), MO disks (SCSI type 7), medium changers (SCSI 
type 8) and RAID controllers (SCSI type 0xC)

  - Stability. Current SCST (0.9.3-pre2) is quite stable, as far as I 
know only task management has some unfixed flaws.

Those are what I've noticed on a brief review. It's possible that I 
missed something. For example, stgt seems doesn't have internal commands 
serialization, so I suspect that if some initiator mixes READs and 
WRITEs on some targets it will be possible that order of the commands 
execution will be broken, because WRITE command have additional phase , 
when the command's data from the initiator is sent to the target. During 
it subsequent READs could be executed out of order. The result could be 
data corruption.

 From other side, stgt has not too much advantages over scst. 
Technically, I, personally, see only few such advantages. One of them is 
support for putting block commands directly to the device request queue. 
I'm going to fix this in the next version (I hope nobody will blame me 
if I borrow this code from stgt :) ). Another one is support for 
different "protocols", although I have not understood which ones, except 
SCSI, are going to be there.

Actually, we would greatly appreciate if Mike or Christoph will tell us 
what is so wrong in scst on their opinion, so they started they own new 
project. (I'm not considering motivation like "I want to make my own in 
any case" seriously). Is scst too complicated? Do you think stgt will be 
simpler with all those features implemented?

Vlad


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_idv37&alloc_id\x16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> Mike Christie wrote:
> 
>> johan@capvert.se wrote:
>>
>>> I, and I suppose a lot of other people, would like know how you look 
>>> at the
>>> HW support on the Fc side? I don´t know what you mean now, when you talk
>>> about HW target support, 
>>
>>
>> For qlogic FC, it is the same as scst at this point, except we are 
>> based off the mainline qla2xxx driver.
> 
> 
> Actually, this is not completely true and could mislead people.  Stgt is
> *not* the same as scst at this point and,

Yes, I agree, I was not talking about functionality. I was only talking 
about how we interface with the initiator driver. My mistake. So we add 
a qla_tgt function callback struct (but we do it for each ha rather than 
global) so the initiator can call back into the target driver, so I am 
basically resusing that code.
-
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

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
>  - Scst has some performance advantages over stgt, at least, on hardware 
> targets, because it allows internal handling in SIRQ context as well as 
> doesn't need user space program, so it eliminates additi
> switches (at least 3 per command for WRITEs and 2 per command for reads 
> plus switches to user space daemon, probably, 2 per command). 5 context 
> switches per command looks too much for me, especially considering how 
> little work is done on each context. It means ~15000 CS/sec on regular 
> 200Mb/sec link with 64K block size. Additionally, kernel only commands 
> execution allows direct access to page cache, which is GREAT performance 
> improvement, because in architecture with user space daemon the data 
> should be copied several times between kernel and user space. Or, do I 
> miss anything?

Userspace only occurs for READ/WRITEs. So for REPORT_LUNS, TUR, etc 
where performance is not a factor.

Also is the page cache comment in reference to us using the page cache 
for our reads and writes or I am not sure why you wrote that if you do 
not do it right now.

> 
>  From other side, stgt has not too much advantages over scst. 

Hey, we just started and have not had too much time.

> Actually, we would greatly appreciate if Mike or Christoph will tell us 
> what is so wrong in scst on their opinion, so they started they own new 
> project. (I'm not considering motivation like "I want to make my own in 
> any case" seriously). Is scst too complicated? Do you think stgt will be 
> simpler with all those features implemented?
> 

Didn't we go over this? To get SCST ready for mainline we would have had 
a large cleanup effort. I started this and sent you a patch to begin the 
cleanup. In the end some of the scsi people liked the idea of throwing 
the non-read/write command to userspace and to do this we just decided 
to start over but I have been cutting and pasting your code and cleaning 
it up as I add more stuff.

Simmer down :) If you had just gotton your code ready when christoph 
asked a year ago we would never have had this problem and I would be 
sending you patches to remove the scsi_request usage as we speak.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[James Bottomley]

On Thu, 2005-12-08 at 21:46 +0300, Vladislav Bolkhovitin wrote:
>   - Scst has some performance advantages over stgt, at least, on 
> hardware targets, because it allows internal handling in SIRQ context as 
> well as doesn't need user space program, so it eliminates additional 
> context switches (at least 3 per command for WRITEs and 2 per command 
> for reads plus switches to user space daemon, probably, 2 per command). 
> 5 context switches per command looks too much for me, especially 
> considering how little work is done on each context. It means ~15000 
> CS/sec on regular 200Mb/sec link with 64K block size. Additionally, 
> kernel only commands execution allows direct access to page cache, which 
> is GREAT performance improvement, because in architecture with user 
> space daemon the data should be copied several times between kernel and 
> user space. Or, do I miimpss anything?

I do have to say that I consider operation in interrupt context (or even
kernel context) to be a disadvantage.  Compared with the response times
that most arrays have to SCSI commands, the kernel context switch time
isn't that significant.

Additionally, it's perfectly possible for all of this to be done zero
copy on the data.  A user space target mmaps the data on its storage
device and then does a SG_IO type scatter gather user virtual region
pass to the underlying target infrastructure.  We already have this
demonstrated in the SG_IO path, someone just needs to come up with the
correct implementation for a target path.

The great advantage of doing SCSI state machines in user space is that
you can prototype anything you want, and user space has much better
state machine implementation (and debugging) tools available than the
kernel does.

James

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[James Bottomley]

On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
> cleanup. In the end some of the scsi people liked the idea of throwing 
> the non-read/write command to userspace and to do this we just decided 
> to start over but I have been cutting and pasting your code and cleaning 
> it up as I add more stuff.

To be honest, I'd like to see all command processing at user level
(including read/write ... for block devices, it shouldn't be that
inefficient, since you're merely going to say remap an area from one
device to another; as long as no data transformation ever occurs, the
user never touches the data and it all remains in the kernel page
cache).

My ideal for the kernel based infrastructure is a simple tap for
transporting commands addressed to devices upwards (and the responses
downwards).  Then everyone can have their own user space processing
implementation that I don't have to care about.

James

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

James Bottomley wrote:
> On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
> 
>>cleanup. In the end some of the scsi people liked the idea of throwing 
>>the non-read/write command to userspace and to do this we just decided 
>>to start over but I have been cutting and pasting your code and cleaning 
>>it up as I add more stuff.
> 
> 
> To be honest, I'd like to see all command processing at user level
> (including read/write ... for block devices, it shouldn't be that
> inefficient, since you're merely going to say remap an area from one
> device to another; as long as no data transformation ever occurs, the
> user never touches the data and it all remains in the kernel page
> cache).

Ok, Tomo and I briefly talked about this when we saw Jeff's post about 
doing block layer drivers in userspace on lkml. I think we were somewhat 
prepared for this given some of your other replies.

So Vlad and other target guys what do you think? Vlad are you going to 
continue to maintain scst as kernel only, or is there some place we can 
work together on this on - if your feelings are not hurt too much that 
is :) ?

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Dave C Boutcher]

On Thu, Dec 08, 2005 at 02:09:32PM -0600, Mike Christie wrote:
> James Bottomley wrote:
> >On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
> >
> >>cleanup. In the end some of the scsi people liked the idea of throwing 
> >>the non-read/write command to userspace and to do this we just decided 
> >>to start over but I have been cutting and pasting your code and cleaning 
> >>it up as I add more stuff.
> >
> >
> >To be honest, I'd like to see all command processing at user level
> >(including read/write ... for block devices, it shouldn't be that
> >inefficient, since you're merely going to say remap an area from one
> >device to another; as long as no data transformation ever occurs, the
> >user never touches the data and it all remains in the kernel page
> >cache).
> 
> Ok, Tomo and I briefly talked about this when we saw Jeff's post about 
> doing block layer drivers in userspace on lkml. I think we were somewhat 
> prepared for this given some of your other replies.
> 
> So Vlad and other target guys what do you think? Vlad are you going to 
> continue to maintain scst as kernel only, or is there some place we can 
> work together on this on - if your feelings are not hurt too much that 
> is :) ?

Oofff....Architecturally I agree with James...do all command processing
in one place.  On the other hand, the processing involved with a read or
write in the normal case (no aborts/resets/ordering/timeouts/etc) is
almost zero.  Figure out the LBA and length and pass on the I/O.  The
overhead of passing it up and down across the kernel boundary is likely
to be orders of magnitude larger than the actual processing.  I would
personally rather not fix this decision in concrete until we could do
some actual measurements of a SCSI target under heavy load.

-- 
Dave Boutcher

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Dave C Boutcher wrote:
> On Thu, Dec 08, 2005 at 02:09:32PM -0600, Mike Christie wrote:
> 
>>James Bottomley wrote:
>>
>>>On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
>>>
>>>
>>>>cleanup. In the end some of the scsi people liked the idea of throwing 
>>>>the non-read/write command to userspace and to do this we just decided 
>>>>to start over but I have been cutting and pasting your code and cleaning 
>>>>it up as I add more stuff.
>>>
>>>
>>>To be honest, I'd like to see all command processing at user level
>>>(including read/write ... for block devices, it shouldn't be that
>>>inefficient, since you're merely going to say remap an area from one
>>>device to another; as long as no data transformation ever occurs, the
>>>user never touches the data and it all remains in the kernel page
>>>cache).
>>
>>Ok, Tomo and I briefly talked about this when we saw Jeff's post about 
>>doing block layer drivers in userspace on lkml. I think we were somewhat 
>>prepared for this given some of your other replies.
>>
>>So Vlad and other target guys what do you think? Vlad are you going to 
>>continue to maintain scst as kernel only, or is there some place we can 
>>work together on this on - if your feelings are not hurt too much that 
>>is :) ?
> 
> 
> Oofff....Architecturally I agree with James...do all command processing
> in one place.  On the other hand, the processing involved with a read or
> write in the normal case (no aborts/resets/ordering/timeouts/etc) is
> almost zero.  Figure out the LBA and length and pass on the I/O.  The

There is still memory and scatterlist allocations. If we are not going 
to allocate all the memory for a command buffer and request with 
GFP_ATOMIC (and can then run from the the HW interrupt or soft irq) we 
have to pass that on to a thread. I guess there is disagreement whether 
that part is a feature or bad use of GFP_ATOMIC though so... But I just 
mean to say there could be a little more to do.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

James Bottomley wrote:
> 
> Additionally, it's perfectly possible for all of this to be done zero
> copy on the data.  A user space target mmaps the data on its storage
> device and then does a SG_IO type scatter gather user virtual region
> pass to the underlying target infrastructure.  We already have this
> demonstrated in the SG_IO path, someone just needs to come up with the
> correct implementation for a target path.
> 

I guess I am going to try to do some work in userspace so we can 
benchmark things and see how much performance drops.

For your suggestion, are you referring to the sg.c mmap path? I was 
thinking that maybe we could modify dm so it can do mmap like sg.c does. 
The dm device would be the target device and would sit above the real 
device or another MD/DM/Loop/ramdsik device or whatever so when 
userspace decides to execute the command it would inform the dm device 
and that kernel driver would just then send down the bios.

For something like qlogic or mpt would this basically work like the 
following:

1. dm mmap is called and our dm_target (dm drivers like dm-multipath or 
dm-raid are called dm_targets) does like sg.c sg_mmap* and sg_vma* ops.
2. HW interrupt comes in and we allocate a scatterlist with pages from #1.
3. netlink (or whatever is the favorite interface) a message to 
userpsace to tell it we have a command ready.
4. userspace decides if it is a read or write and if it is a read or 
write then userspace tells the dm device to read/write some pages.
5. dm's bi_endio is called when io is finished so we netlink to 
userspace and then userspaces netlinks back to the kernel and tells the 
LLD like qlogic that some data and/or a responce or sense is ready for 
it to transfer.

Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

[Ming Zhang]

On Thu, 2005-12-08 at 21:57 -0600, Mike Christie wrote:
> James Bottomley wrote:
> > 
> > Additionally, it's perfectly possible for all of this to be done zero
> > copy on the data.  A user space target mmaps the data on its storage
> > device and then does a SG_IO type scatter gather user virtual region
> > pass to the underlying target infrastructure.  We already have this
> > demonstrated in the SG_IO path, someone just needs to come up with the
> > correct implementation for a target path.
> > 
> 
> I guess I am going to try to do some work in userspace so we can 
> benchmark things and see how much performance drops.
> 
> For your suggestion, are you referring to the sg.c mmap path? I was 
> thinking that maybe we could modify dm so it can do mmap like sg.c does. 
> The dm device would be the target device and would sit above the real 
> device or another MD/DM/Loop/ramdsik device or whatever so when 
> userspace decides to execute the command it would inform the dm device 
> and that kernel driver would just then send down the bios.
> 
> For something like qlogic or mpt would this basically work like the 
> following:
> 
> 1. dm mmap is called and our dm_target (dm drivers like dm-multipath or 
> dm-raid are called dm_targets) does like sg.c sg_mmap* and sg_vma* ops.
> 2. HW interrupt comes in and we allocate a scatterlist with pages from #1.
> 3. netlink (or whatever is the favorite interface) a message to 
> userpsace to tell it we have a command ready.
> 4. userspace decides if it is a read or write and if it is a read or 
> write then userspace tells the dm device to read/write some pages.
> 5. dm's bi_endio is called when io is finished so we netlink to 
> userspace and then userspaces netlinks back to the kernel and tells the 
> LLD like qlogic that some data and/or a responce or sense is ready for 
> it to transfer.

we can count how many system calls and context switches will occur for
each request here. will this be a source of high response time?

i guess we really need to have both a kernel and user space
implementation here to compare.




-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> Vladislav Bolkhovitin wrote:
> 
>>  - Scst has some performance advantages over stgt, at least, on 
>> hardware targets, because it allows internal handling in SIRQ context 
>> as well as doesn't need user space program, so it eliminates additi
>> switches (at least 3 per command for WRITEs and 2 per command for 
>> reads plus switches to user space daemon, probably, 2 per command). 5 
>> context switches per command looks too much for me, especially 
>> considering how little work is done on each context. It means ~15000 
>> CS/sec on regular 200Mb/sec link with 64K block size. Additionally, 
>> kernel only commands execution allows direct access to page cache, 
>> which is GREAT performance improvement, because in architecture with 
>> user space daemon the data should be copied several times between 
>> kernel and user space. Or, do I miss anything?
> 
> 
> Userspace only occurs for READ/WRITEs. So for REPORT_LUNS, TUR, etc 
> where performance is not a factor.

So, do you mean that READ/WRITE operations are (or will be) performed in 
  kernel space only? I don't see that from the code.

> Also is the page cache comment in reference to us using the page cache 
> for our reads and writes or I am not sure why you wrote that if you do 
> not do it right now.

Hm, at first, the page cache already used somehow in fileio dev handler 
(though, with additional mem copy). At the second, fully utilize the 
page cache is one of two major improvements that are pending in scst, 
because it is required changing the kernel, which until some moment try 
to avoid. Although I prepared what is necessary for that.

The idea basically is the following. When READ operation arrives, pages 
for all requested blocks are at first searched in the page cache 
(probably, in SIRQ context, because it isn't expensive operation) and if 
all pages are found, they are referenced and the result will be sent to 
the initiator. Then the pages will be dereferenced (so, no pages 
allocation will be done at all). Otherwise, the missed pages will be 
allocated and the command will be rescheduled to the thread, which will 
read them. Then, after the response is sent, the pages will remain in 
the page cache for future accesses. For WRITEs the processing is the 
similar, the pages with the data will be put in the page cache.

I doubt it is possible to implement such simple and effective algorithm 
via user space. Pages mapping and unmapping is quite expensive for both 
coding effort and performance.

>>
>>  From other side, stgt has not too much advantages over scst. 
> 
> 
> Hey, we just started and have not had too much time.

No offense, please, you wrote stgt and scst are the same, I wrote, they 
aren't :-)

>> Actually, we would greatly appreciate if Mike or Christoph will tell 
>> us what is so wrong in scst on their opinion, so they started they own 
>> new project. (I'm not considering motivation like "I want to make my 
>> own in any case" seriously). Is scst too complicated? Do you think 
>> stgt will be simpler with all those features implemented?
>>
> 
> Didn't we go over this? To get SCST ready for mainline we would have had 
> a large cleanup effort. I started this and sent you a patch to begin the 
> cleanup. In the end some of the scsi people liked the idea of throwing 
> the non-read/write command to userspace and to do this we just decided 
> to start over but I have been cutting and pasting your code and cleaning 
> it up as I add more stuff.

The patches that I've seen were just pretty mechanic cleanups and 
renames, which could be done in a half of hour time and which I'm going 
to do in any case before preparing the patch. So, that reason doesn't 
look convincing for me to throw away a big chunk of working code. Doing 
so you delayed SCSI targets development for at least to a year-two, 
because there are too much features for you to implement in stgt, which 
are already working and useful in scst.

 From other side, if you look on scst closely you will see that:

- The user space gate could be easily and clearly done using existing 
dev handler's hooks

- There is nothing in current SCST that could be moved in user space 
without sacrificing performance. Neither task management, nor UA 
processing, nor RESERVE/RELEASE emulation. Otherwise, you will have to 
pass *all* the commands to the user space to check allowed this command 
for processing or it shall be returned with an error. SCST core is just 
about 7500 lines of code. Is it too much?

So, in my opinion, there is/was no technical point to drop scst in favor 
of stgt.

> Simmer down :) If you had just gotton your code ready when christoph 
> asked a year ago we would never have had this problem and I would be 
> sending you patches to remove the scsi_request usage as we speak.

Christoph didn't ask me to do anything. Actually, nobody contacted and 
asked anything about that me when you with Christoph conceived stgt.

Vlad

Re: [Scst-devel] Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

James Bottomley wrote:
> On Thu, 2005-12-08 at 21:46 +0300, Vladislav Bolkhovitin wrote:
> 
>>  - Scst has some performance advantages over stgt, at least, on 
>>hardware targets, because it allows internal handling in SIRQ context as 
>>well as doesn't need user space program, so it eliminates additional 
>>context switches (at least 3 per command for WRITEs and 2 per command 
>>for reads plus switches to user space daemon, probably, 2 per command). 
>>5 context switches per command looks too much for me, especially 
>>considering how little work is done on each context. It means ~15000 
>>CS/sec on regular 200Mb/sec link with 64K block size. Additionally, 
>>kernel only commands execution allows direct access to page cache, which 
>>is GREAT performance improvement, because in architecture with user 
>>space daemon the data should be copied several times between kernel and 
>>user space. Or, do I miimpss anything?
> 
> 
> I do have to say that I consider operation in interrupt context (or even
> kernel context) to be a disadvantage.  Compared with the response times
> that most arrays have to SCSI commands, the kernel context switch time
> isn't that significant.
> 
> Additionally, it's perfectly possible for all of this to be done zero
> copy on the data.  A user space target mmaps the data on its storage
> device and then does a SG_IO type scatter gather user virtual region
> pass to the underlying target infrastructure.  We already have this
> demonstrated in the SG_IO path, someone just needs to come up with the
> correct implementation for a target path.

I'm not completely understand how it will work. Consider, there are 
READ/WRITE commands with random data sizes arrive from an initiator. Are 
you going to do map/unmap for each command individually or alloc data 
buffers for commands from a premapped area and live with possible its 
fragmentation? If map/unmap individually, then I should say that those 
are very expensive operations.

> The great advantage of doing SCSI state machines in user space is that
> you can prototype anything you want, and user space has much better
> state machine implementation (and debugging) tools available than the
> kernel does.
> 
> James
> 
> 
> 
> 
> -------------------------------------------------------
> This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
> for problems?  Stop!  Download the new AJAX search engine that makes
> searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
> http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click
> _______________________________________________
> Scst-devel mailing list
> Scst-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/scst-devel
> 

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> James Bottomley wrote:
> 
>> On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
>>
>>> cleanup. In the end some of the scsi people liked the idea of 
>>> throwing the non-read/write command to userspace and to do this we 
>>> just decided to start over but I have been cutting and pasting your 
>>> code and cleaning it up as I add more stuff.
>>
>>
>>
>> To be honest, I'd like to see all command processing at user level
>> (including read/write ... for block devices, it shouldn't be that
>> inefficient, since you're merely going to say remap an area from one
>> device to another; as long as no data transformation ever occurs, the
>> user never touches the data and it all remains in the kernel page
>> cache).
> 
> 
> Ok, Tomo and I briefly talked about this when we saw Jeff's post about 
> doing block layer drivers in userspace on lkml. I think we were somewhat 
> prepared for this given some of your other replies.

Could you give me a reference on this message, please?

> So Vlad and other target guys what do you think? Vlad are you going to 
> continue to maintain scst as kernel only, or is there some place we can 
> work together on this on - if your feelings are not hurt too much that 
> is :) ?

 From one side, I hate dropping anything in the point where it isn't 
somehow completed. From other side, I don't have much time to spend on 
this my hobby. Nobody pays me for that, it's just for fun.

Vlad

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> Dave C Boutcher wrote:
> 
>> On Thu, Dec 08, 2005 at 02:09:32PM -0600, Mike Christie wrote:
>>
>>> James Bottomley wrote:
>>>
>>>> On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
>>>>
>>>>
>>>>> cleanup. In the end some of the scsi people liked the idea of 
>>>>> throwing the non-read/write command to userspace and to do this we 
>>>>> just decided to start over but I have been cutting and pasting your 
>>>>> code and cleaning it up as I add more stuff.
>>>>
>>>>
>>>>
>>>> To be honest, I'd like to see all command processing at user level
>>>> (including read/write ... for block devices, it shouldn't be that
>>>> inefficient, since you're merely going to say remap an area from one
>>>> device to another; as long as no data transformation ever occurs, the
>>>> user never touches the data and it all remains in the kernel page
>>>> cache).
>>>
>>>
>>> Ok, Tomo and I briefly talked about this when we saw Jeff's post 
>>> about doing block layer drivers in userspace on lkml. I think we were 
>>> somewhat prepared for this given some of your other replies.
>>>
>>> So Vlad and other target guys what do you think? Vlad are you going 
>>> to continue to maintain scst as kernel only, or is there some place 
>>> we can work together on this on - if your feelings are not hurt too 
>>> much that is :) ?
>>
>>
>>
>> Oofff....Architecturally I agree with James...do all command processing
>> in one place.  On the other hand, the processing involved with a read or
>> write in the normal case (no aborts/resets/ordering/timeouts/etc) is
>> almost zero.  Figure out the LBA and length and pass on the I/O.  The
> 
> 
> There is still memory and scatterlist allocations. If we are not going 
> to allocate all the memory for a command buffer and request with 
> GFP_ATOMIC (and can then run from the the HW interrupt or soft irq) we 
> have to pass that on to a thread. I guess there is disagreement whether 
> that part is a feature or bad use of GFP_ATOMIC though so... But I just 
> mean to say there could be a little more to do.

Actually, there is the way to allocate sg vectors with buffers in SIRQ 
and not with GFP_ATOMIC. This is the second major improvement, which is 
pending in scst. I called it sgv_pool. This is a new allocator in the 
kernel similar to mem_pool, but it contains *complete* sg-vectors of 
some size with data buffers (pages). Initiator sends data requests 
usually with some fixed size, like 128K. After a data command completed, 
its sg vector will not be immediately freed, but will be kept in 
sgv_pool until the next request (command) or memory pressure on the 
system. So, all subsequent commands will allocate already built vectors. 
The first allocations will be done in some thread context. This allows 
to allocate huge chunks of memory in SIRQ context as well as save a lot 
of CPU power necessary to always build big sg vectors for each command 
individually.

Vlad

Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> Vladislav Bolkhovitin wrote:
> 
>> Mike Christie wrote:
>>
>>> johan@capvert.se wrote:
>>>
>>>> I, and I suppose a lot of other people, would like know how you look 
>>>> at the
>>>> HW support on the Fc side? I don´t know what you mean now, when you 
>>>> talk
>>>> about HW target support, 
>>>
>>>
>>>
>>> For qlogic FC, it is the same as scst at this point, except we are 
>>> based off the mainline qla2xxx driver.
>>
>>
>>
>> Actually, this is not completely true and could mislead people.  Stgt is
>> *not* the same as scst at this point and,
> 
> 
> Yes, I agree, I was not talking about functionality. I was only talking 
> about how we interface with the initiator driver. My mistake. So we add 
> a qla_tgt function callback struct (but we do it for each ha rather than 
> global) so the initiator can call back into the target driver, so I am 
> basically resusing that code.

OK, if you reuse scst code, I hope you won't forget to insert somewhere 
reference to its original authors :-)




-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_idv37&alloc_id\x16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Dave C Boutcher wrote:
> On Thu, Dec 08, 2005 at 02:09:32PM -0600, Mike Christie wrote:
> 
>>James Bottomley wrote:
>>
>>>On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
>>>
>>>
>>>>cleanup. In the end some of the scsi people liked the idea of throwing 
>>>>the non-read/write command to userspace and to do this we just decided 
>>>>to start over but I have been cutting and pasting your code and cleaning 
>>>>it up as I add more stuff.
>>>
>>>
>>>To be honest, I'd like to see all command processing at user level
>>>(including read/write ... for block devices, it shouldn't be that
>>>inefficient, since you're merely going to say remap an area from one
>>>device to another; as long as no data transformation ever occurs, the
>>>user never touches the data and it all remains in the kernel page
>>>cache).
>>
>>Ok, Tomo and I briefly talked about this when we saw Jeff's post about 
>>doing block layer drivers in userspace on lkml. I think we were somewhat 
>>prepared for this given some of your other replies.
>>
>>So Vlad and other target guys what do you think? Vlad are you going to 
>>continue to maintain scst as kernel only, or is there some place we can 
>>work together on this on - if your feelings are not hurt too much that 
>>is :) ?
> 
> 
> Oofff....Architecturally I agree with James...do all command processing
> in one place.  On the other hand, the processing involved with a read or
> write in the normal case (no aborts/resets/ordering/timeouts/etc) is
> almost zero.  Figure out the LBA and length and pass on the I/O.  The
> overhead of passing it up and down across the kernel boundary is likely
> to be orders of magnitude larger than the actual processing.  I would
> personally rather not fix this decision in concrete until we could do
> some actual measurements of a SCSI target under heavy load.

Totally agree.

Vlad

Re: [Scst-devel] Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[James Bottomley]

On Fri, 2005-12-09 at 18:29 +0300, Vladislav Bolkhovitin wrote:
> > Additionally, it's perfectly possible for all of this to be done zero
> > copy on the data.  A user space target mmaps the data on its storage
> > device and then does a SG_IO type scatter gather user virtual region
> > pass to the underlying target infrastructure.  We already have this
> > demonstrated in the SG_IO path, someone just needs to come up with the
> > correct implementation for a target path.
> 
> I'm not completely understand how it will work. Consider, there are 
> READ/WRITE commands with random data sizes arrive from an initiator. Are 
> you going to do map/unmap for each command individually or alloc data 
> buffers for commands from a premapped area and live with possible its 
> fragmentation? If map/unmap individually, then I should say that those 
> are very expensive operations.

You do it the same way an array does:  the model for SPI is read command
phase, disconnect, process command (i.e. set up areas) reconnect for
data transfer.

map/unmap are really only necessary if you're emulating the data store,
but it's a fairly cheap operation on linux: it just causes the creation
of a vm_area.  If it's a pass through, you can use SG_IO to pull it in
and the SG_IO like output to shoot it out again, effectively using a
piece of user memory as a zero copy buffer.

Fragmentation isn't an issue because the I/O goes via sg lists , all
that's needed is a bunch of pages.

James




-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> Mike Christie wrote:
> 
>> Vladislav Bolkhovitin wrote:
>>
>>>  - Scst has some performance advantages over stgt, at least, on 
>>> hardware targets, because it allows internal handling in SIRQ context 
>>> as well as doesn't need user space program, so it eliminates additi
>>> switches (at least 3 per command for WRITEs and 2 per command for 
>>> reads plus switches to user space daemon, probably, 2 per command). 5 
>>> context switches per command looks too much for me, especially 
>>> considering how little work is done on each context. It means ~15000 
>>> CS/sec on regular 200Mb/sec link with 64K block size. Additionally, 
>>> kernel only commands execution allows direct access to page cache, 
>>> which is GREAT performance improvement, because in architecture with 
>>> user space daemon the data should be copied several times between 
>>> kernel and user space. Or, do I miss anything?
>>
>>
>>
>> Userspace only occurs for READ/WRITEs. So for REPORT_LUNS, TUR, etc 
>> where performance is not a factor.
> 
> 
> So, do you mean that READ/WRITE operations are (or will be) performed in 
>  kernel space only? I don't see that from the code.


They are today. I think they have always been that way. Look again :)


>>>
>>>  From other side, stgt has not too much advantages over scst. 
>>
>>
>>
>> Hey, we just started and have not had too much time.
> 
> 
> No offense, please, you wrote stgt and scst are the same, I wrote, they 
> aren't :-)
> 

Yeah, I know. It was my mistake. If you read other mails, you would see 
I mentioned that we do not have any type of eh and I make Tomo mess up 
the serialization stuff becase I asked him to try to reuse some block 
layer code.

If you read other mails, you would also see I reccomended scst to people 
requesting something stable :)

>>> Actually, we would greatly appreciate if Mike or Christoph will tell 
>>> us what is so wrong in scst on their opinion, so they started they 
>>> own new project. (I'm not considering motivation like "I want to make 
>>> my own in any case" seriously). Is scst too complicated? Do you think 
>>> stgt will be simpler with all those features implemented?
>>>
>>
>> Didn't we go over this? To get SCST ready for mainline we would have 
>> had a large cleanup effort. I started this and sent you a patch to 
>> begin the cleanup. In the end some of the scsi people liked the idea 
>> of throwing the non-read/write command to userspace and to do this we 
>> just decided to start over but I have been cutting and pasting your 
>> code and cleaning it up as I add more stuff.
> 
> 
> The patches that I've seen were just pretty mechanic cleanups and 
> renames, which could be done in a half of hour time and which I'm going 

Yeah it was the beginning of the easy work. I did not mean that as an 
example of evertthing. I thought you would remember when we discussed 
this on linux-scsi before.


> to do in any case before preparing the patch. So, that reason doesn't 
> look convincing for me to throw away a big chunk of working code. Doing 
> so you delayed SCSI targets development for at least to a year-two, 

Hey, looked how long it took iscsi to get in becuase we wasted so much 
time cleaning up iscsi-sfnet :)

> because there are too much features for you to implement in stgt, which 
> are already working and useful in scst.

Well, there was more when you asked on linux-scsi. You have other things 
like refcouting (we only are adding that in today, but we do get 
references to the scsi_host when we access the qla2xxx ha at least). If 
someone ripps out a qla2xxx card you will oops.

We also did not want to hook in as a SCSI ULD interface becuase we did 
not want to worry about what happens when poeple start using 
usb-mass-storage for targets and LUNs. Look how many times we see Alan 
Stern pop up for just the initiator side :) And to be honest DM would do 
a lot of what tgt and scst want as far as giving us a reference to the 
devive we want to use as a LUN and handling all the setup work so we 
probably both messed up there :(

> 
>  From other side, if you look on scst closely you will see that:
> 
> - The user space gate could be easily and clearly done using existing 
> dev handler's hooks

Yeah and the problem is that we just do not believe those are 
implemented in the correct place. We do not like class interface SCSI 
hook in, when we can do the same thing from userspace.

> 
> - There is nothing in current SCST that could be moved in user space 
> without sacrificing performance. Neither task management, nor UA 
> processing, nor RESERVE/RELEASE emulation. Otherwise, you will have to 
> pass *all* the commands to the user space to check allowed this command 
> for processing or it shall be returned with an error.

For non READ/WRITE, we are ok with that performance drops. Even for 
doing READ/WRITEs in the kernel from interrupt context, we were going to 
run from a softirq, but we thought allocating the whole command with 
GFP_ATOMIC would not be liked so we pass it to the thread. And for when 
we do pass through (using elv_add_request/blk_rq_execute_nowait), we can 
do it in just the context swith needed for the memory allocation. But to 
do GFP_ATOMOIC softirq or hw irq would not be a problem, although I do 
not think we want to submit IO from the hw interrupt incase the queue 
gets unplugged at the same time :)

For non READ/WRITEs look how far open-iscsi went. And from James's 
reply, you see that he thinks READs and WRITEs can go to userspace too, 
so you know this is an uphill battle to get everything in the kernel.


  SCST core is just
> about 7500 lines of code. Is it too much?

Ask the people that have to review the code? :) After sfnet, I learned 
that it is sometimes best to get the basics in the kernel first so we do 
not burn out the christoph robot :) I think part of this stems from the 
fact that I touched pretty much every line in that driver to clean it up 
and it took me about a year. And while I was beginning to clean up scst 
I began to remember sfnet.

But there are other cleanups like moving some of the state to per 
target, cleaningup the scattlist allocation code and moving it to 
scsi-ml so the SCSI ULDs can use them and convert them. There is also 
thing like converting to the right APIs for 2.6 (rm kernel_thread, rm 
scsi_request, rm proc, fixup class interface refcouting problems, fixup 
scsi_device lack of refcounting usage, etc).


> 
>> Simmer down :) If you had just gotton your code ready when christoph 
>> asked a year ago we would never have had this problem and I would be 
>> sending you patches to remove the scsi_request usage as we speak.
> 
> 
> Christoph didn't ask me to do anything. Actually, nobody contacted and 

What? Look at the review comments you got on the list when you first 
posted scst. It asked you to take out the 2.4ish code out and clean it 
up didn't it?

> asked anything about that me when you with Christoph conceived stgt.
> 

I do not think Christoph and I conceived stgt. Tomo and I did or 
actually Tomo did the original code to push things to userspace. I just 
bugged him about getting something in mainlin. We only asked Christoph 
what we would should do to get things in mainline. After seeing what 
open-iscsi did and the comments from Arjan about IET, we thought we 
would have to follow a similar route as open-iscsi.


At this point I do not care what we do. I just want to get something in 
mainline for everyone to use and so I can use it at home becuase I am 
lazy and do not want to drive to work in the snow to test FC card 
updates :) If we can come to some conclusions on some of these issues I 
am sure Tomo and I would be happy to work together.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> Mike Christie wrote:
> 
>> Dave C Boutcher wrote:
>>
>>> On Thu, Dec 08, 2005 at 02:09:32PM -0600, Mike Christie wrote:
>>>
>>>> James Bottomley wrote:
>>>>
>>>>> On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
>>>>>
>>>>>
>>>>>> cleanup. In the end some of the scsi people liked the idea of 
>>>>>> throwing the non-read/write command to userspace and to do this we 
>>>>>> just decided to start over but I have been cutting and pasting 
>>>>>> your code and cleaning it up as I add more stuff.
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> To be honest, I'd like to see all command processing at user level
>>>>> (including read/write ... for block devices, it shouldn't be that
>>>>> inefficient, since you're merely going to say remap an area from one
>>>>> device to another; as long as no data transformation ever occurs, the
>>>>> user never touches the data and it all remains in the kernel page
>>>>> cache).
>>>>
>>>>
>>>>
>>>> Ok, Tomo and I briefly talked about this when we saw Jeff's post 
>>>> about doing block layer drivers in userspace on lkml. I think we 
>>>> were somewhat prepared for this given some of your other replies.
>>>>
>>>> So Vlad and other target guys what do you think? Vlad are you going 
>>>> to continue to maintain scst as kernel only, or is there some place 
>>>> we can work together on this on - if your feelings are not hurt too 
>>>> much that is :) ?
>>>
>>>
>>>
>>>
>>> Oofff....Architecturally I agree with James...do all command processing
>>> in one place.  On the other hand, the processing involved with a read or
>>> write in the normal case (no aborts/resets/ordering/timeouts/etc) is
>>> almost zero.  Figure out the LBA and length and pass on the I/O.  The
>>
>>
>>
>> There is still memory and scatterlist allocations. If we are not going 
>> to allocate all the memory for a command buffer and request with 
>> GFP_ATOMIC (and can then run from the the HW interrupt or soft irq) we 
>> have to pass that on to a thread. I guess there is disagreement 
>> whether that part is a feature or bad use of GFP_ATOMIC though so... 
>> But I just mean to say there could be a little more to do.
> 
> 
> Actually, there is the way to allocate sg vectors with buffers in SIRQ 
> and not with GFP_ATOMIC. This is the second major improvement, which is 
> pending in scst. I called it sgv_pool. This is a new allocator in the 
> kernel similar to mem_pool, but it contains *complete* sg-vectors of 
> some size with data buffers (pages). Initiator sends data requests 
> usually with some fixed size, like 128K. After a data command completed, 
> its sg vector will not be immediately freed, but will be kept in 

We considered this, but what did you decide is the upper limit size for 
the pool? Is it dynmaic? We also wanted something that the SCSI ULDs 
could use for their allocations which could go up to 6 MB.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
>>
>> Yes, I agree, I was not talking about functionality. I was only 
>> talking about how we interface with the initiator driver. My mistake. 
>> So we add a qla_tgt function callback struct (but we do it for each ha 
>> rather than global) so the initiator can call back into the target 
>> driver, so I am basically resusing that code.
> 
> 
> OK, if you reuse scst code, I hope you won't forget to insert somewhere 
> reference to its original authors :-)
> 

yes of course, when approporate I kept your copyrights even.

Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

[Mike Christie]

Mike Christie wrote:
>>>> Actually, we would greatly appreciate if Mike or Christoph will tell 
>>>> us what is so wrong in scst on their opinion, so they started they 
>>>> own new project. (I'm not considering motivation like "I want to 
>>>> make my own in any case" seriously). Is scst too complicated? Do you 
>>>> think stgt will be simpler with all those features implemented?
>>>>
>>>
>>> Didn't we go over this? To get SCST ready for mainline we would have 
>>> had a large cleanup effort. I started this and sent you a patch to 
>>> begin the cleanup. In the end some of the scsi people liked the idea 
>>> of throwing the non-read/write command to userspace and to do this we 
>>> just decided to start over but I have been cutting and pasting your 
>>> code and cleaning it up as I add more stuff.
>>
>>
>>
>> The patches that I've seen were just pretty mechanic cleanups and 
>> renames, which could be done in a half of hour time and which I'm going 
> 
> 
> Yeah it was the beginning of the easy work. I did not mean that as an 
> example of evertthing. I thought you would remember when we discussed 
> this on linux-scsi before.
> 
> 
>> to do in any case before preparing the patch. So, that reason doesn't 
>> look convincing for me to throw away a big chunk of working code. 
>> Doing so you delayed SCSI targets development for at least to a year-two, 
> 
> 
> Hey, looked how long it took iscsi to get in becuase we wasted so much 
> time cleaning up iscsi-sfnet :)
> 
>> because there are too much features for you to implement in stgt, 
>> which are already working and useful in scst.
> 
> 
> Well, there was more when you asked on linux-scsi. You have other things 
> like refcouting (we only are adding that in today, but we do get 
> references to the scsi_host when we access the qla2xxx ha at least). If 
> someone ripps out a qla2xxx card you will oops.
> 
> We also did not want to hook in as a SCSI ULD interface becuase we did 
> not want to worry about what happens when poeple start using 
> usb-mass-storage for targets and LUNs. Look how many times we see Alan 
> Stern pop up for just the initiator side :) And to be honest DM would do 
> a lot of what tgt and scst want as far as giving us a reference to the 
> devive we want to use as a LUN and handling all the setup work so we 
> probably both messed up there :(
> 
>>
>>  From other side, if you look on scst closely you will see that:
>>
>> - The user space gate could be easily and clearly done using existing 
>> dev handler's hooks
> 
> 
> Yeah and the problem is that we just do not believe those are 
> implemented in the correct place. We do not like class interface SCSI 
> hook in, when we can do the same thing from userspace.
> 
>>
>> - There is nothing in current SCST that could be moved in user space 
>> without sacrificing performance. Neither task management, nor UA 
>> processing, nor RESERVE/RELEASE emulation. Otherwise, you will have to 
>> pass *all* the commands to the user space to check allowed this 
>> command for processing or it shall be returned with an error.
> 
> 
> For non READ/WRITE, we are ok with that performance drops. Even for 
> doing READ/WRITEs in the kernel from interrupt context, we were going to 
> run from a softirq, but we thought allocating the whole command with 
> GFP_ATOMIC would not be liked so we pass it to the thread. And for when 
> we do pass through (using elv_add_request/blk_rq_execute_nowait), we can 
> do it in just the context swith needed for the memory allocation. But to 
> do GFP_ATOMOIC softirq or hw irq would not be a problem, although I do 
> not think we want to submit IO from the hw interrupt incase the queue 
> gets unplugged at the same time :)
> 
> For non READ/WRITEs look how far open-iscsi went. And from James's 
> reply, you see that he thinks READs and WRITEs can go to userspace too, 
> so you know this is an uphill battle to get everything in the kernel.
> 
> 
>  SCST core is just
> 
>> about 7500 lines of code. Is it too much?
> 
> 
> Ask the people that have to review the code? :) After sfnet, I learned 
> that it is sometimes best to get the basics in the kernel first so we do 
> not burn out the christoph robot :) I think part of this stems from the 
> fact that I touched pretty much every line in that driver to clean it up 
> and it took me about a year. And while I was beginning to clean up scst 
> I began to remember sfnet.
> 
> But there are other cleanups like moving some of the state to per 
> target, cleaningup the scattlist allocation code and moving it to 
> scsi-ml so the SCSI ULDs can use them and convert them. There is also 
> thing like converting to the right APIs for 2.6 (rm kernel_thread, rm 
> scsi_request, rm proc, fixup class interface refcouting problems, fixup 
> scsi_device lack of refcounting usage, etc).
> 

Oh yeah I think the other major issue at least I had with scst was that 
it was scsi specific and we wanted try and seperate things so if drivers 
like IET and vscsi are allowed then we could also do other drivers like 
a ATA over ethernet target driver or allow any other target driver that 
wanted to to hook in. I think you noted that we were spererating some 
protocol specific things as a distadvantage or mentioned it for some 
reason but I am not completely sure why and we may not agree on that 
issue too.


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[FUJITA Tomonori]

From: Vladislav Bolkhovitin <vst@vlnb.net>
Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new	version of iscsi target?
Date: Fri, 09 Dec 2005 18:29:57 +0300

> > There is still memory and scatterlist allocations. If we are not going 
> > to allocate all the memory for a command buffer and request with 
> > GFP_ATOMIC (and can then run from the the HW interrupt or soft irq) we 
> > have to pass that on to a thread. I guess there is disagreement whether 
> > that part is a feature or bad use of GFP_ATOMIC though so... But I just 
> > mean to say there could be a little more to do.
> 
> Actually, there is the way to allocate sg vectors with buffers in SIRQ 
> and not with GFP_ATOMIC. This is the second major improvement, which is 
> pending in scst. I called it sgv_pool. This is a new allocator in the 
> kernel similar to mem_pool, but it contains *complete* sg-vectors of 
> some size with data buffers (pages). Initiator sends data requests 
> usually with some fixed size, like 128K. After a data command completed, 
> its sg vector will not be immediately freed, but will be kept in 
> sgv_pool until the next request (command) or memory pressure on the 
> system. So, all subsequent commands will allocate already built vectors. 
> The first allocations will be done in some thread context. This allows 
> to allocate huge chunks of memory in SIRQ context as well as save a lot 
> of CPU power necessary to always build big sg vectors for each command 
> individually.

ibmvscsis code the same thing, though it is mainly for caching
dma-mapped pages, I guess.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[FUJITA Tomonori]

From: Vladislav Bolkhovitin <vst@vlnb.net>
Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?
Date: Fri, 09 Dec 2005 18:28:58 +0300

> > Also is the page cache comment in reference to us using the page cache 
> > for our reads and writes or I am not sure why you wrote that if you do 
> > not do it right now.
> 
> Hm, at first, the page cache already used somehow in fileio dev handler 
> (though, with additional mem copy). At the second, fully utilize the 
> page cache is one of two major improvements that are pending in scst, 
> because it is required changing the kernel, which until some moment try 
> to avoid. Although I prepared what is necessary for that.
> 
> The idea basically is the following. When READ operation arrives, pages 
> for all requested blocks are at first searched in the page cache 
> (probably, in SIRQ context, because it isn't expensive operation) and if 
> all pages are found, they are referenced and the result will be sent to 
> the initiator. Then the pages will be dereferenced (so, no pages 
> allocation will be done at all). Otherwise, the missed pages will be 
> allocated and the command will be rescheduled to the thread, which will 
> read them. Then, after the response is sent, the pages will remain in 
> the page cache for future accesses. For WRITEs the processing is the 
> similar, the pages with the data will be put in the page cache.

The Ardis iSCSI target code does the same thing.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
>> But there are other cleanups like moving some of the state to per 
>> target, cleaningup the scattlist allocation code and moving it to 
>> scsi-ml so the SCSI ULDs can use them and convert them. There is also 
>> thing like converting to the right APIs for 2.6 (rm kernel_thread, rm 
>> scsi_request, rm proc, fixup class interface refcouting problems, 
>> fixup scsi_device lack of refcounting usage, etc).
>>
> 
> Oh yeah I think the other major issue at least I had with scst was that 
> it was scsi specific and we wanted try and seperate things so if drivers 
> like IET and vscsi are allowed then we could also do other drivers like 
> a ATA over ethernet target driver or allow any other target driver that 
> wanted to to hook in. I think you noted that we were spererating some 
> protocol specific things as a distadvantage or mentioned it for some 
> reason but I am not completely sure why and we may not agree on that 
> issue too.

SCSI has a lot of very specific stuff like UA handling and (at least) 
some parts of task management, especially if we consider honoring NACA, 
QErr, TST, UA_INTLCK_CTRL bits, therefore I'm not sure that to have 
common target parts for other protocols worths complicating the 
mid-layer with code and interfaces that will separate SCSI-specifics 
from non-SCSI protocols. So, good luck with it :-)

Vlad

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
>>> There is still memory and scatterlist allocations. If we are not 
>>> going to allocate all the memory for a command buffer and request 
>>> with GFP_ATOMIC (and can then run from the the HW interrupt or soft 
>>> irq) we have to pass that on to a thread. I guess there is 
>>> disagreement whether that part is a feature or bad use of GFP_ATOMIC 
>>> though so... But I just mean to say there could be a little more to do.
>>
>>
>>
>> Actually, there is the way to allocate sg vectors with buffers in SIRQ 
>> and not with GFP_ATOMIC. This is the second major improvement, which 
>> is pending in scst. I called it sgv_pool. This is a new allocator in 
>> the kernel similar to mem_pool, but it contains *complete* sg-vectors 
>> of some size with data buffers (pages). Initiator sends data requests 
>> usually with some fixed size, like 128K. After a data command 
>> completed, its sg vector will not be immediately freed, but will be 
>> kept in 
> 
> 
> We considered this, but what did you decide is the upper limit size for 
> the pool? Is it dynmaic? We also wanted something that the SCSI ULDs 
> could use for their allocations which could go up to 6 MB.

Why do you think it needs any upper limit size? Would you like also any 
upper limits on sizes of the page or slab caches? I don't see any 
difference between them and sgv_pool. Let it allocate and keep unused as 
much memory as possible until it would be notified about memory pressure 
and ask to free some. Similarly as slab does.

BTW, from our experience, initiators tend to send commands without limit 
and in case if actual SCSI device isn't so fast to serve all of them 
with the rate, which they arrive, incoming queued commands very quickly 
eat all system memory with obvious consequences. So, there must be some 
kind of IO-throttling, when after some watermark the initiators start 
receiving TASK QUEUE FULL, which calm them down. I implemented very 
stupid one: don't let more than 32 commands per session. This let scst 
survive under extreme loads. You can think about something smarter  :-)

Vlad

Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

FUJITA Tomonori wrote:
> From: Vladislav Bolkhovitin <vst@vlnb.net>
> Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?
> Date: Fri, 09 Dec 2005 18:28:58 +0300
> 
> 
>>>Also is the page cache comment in reference to us using the page cache 
>>>for our reads and writes or I am not sure why you wrote that if you do 
>>>not do it right now.
>>
>>Hm, at first, the page cache already used somehow in fileio dev handler 
>>(though, with additional mem copy). At the second, fully utilize the 
>>page cache is one of two major improvements that are pending in scst, 
>>because it is required changing the kernel, which until some moment try 
>>to avoid. Although I prepared what is necessary for that.
>>
>>The idea basically is the following. When READ operation arrives, pages 
>>for all requested blocks are at first searched in the page cache 
>>(probably, in SIRQ context, because it isn't expensive operation) and if 
>>all pages are found, they are referenced and the result will be sent to 
>>the initiator. Then the pages will be dereferenced (so, no pages 
>>allocation will be done at all). Otherwise, the missed pages will be 
>>allocated and the command will be rescheduled to the thread, which will 
>>read them. Then, after the response is sent, the pages will remain in 
>>the page cache for future accesses. For WRITEs the processing is the 
>>similar, the pages with the data will be put in the page cache.
> 
> The Ardis iSCSI target code does the same thing.

Perfectly. So, why don't do it on the mid-layer level where all targets 
can benefit from it?

Vlad


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

James Bottomley wrote:
> On Fri, 2005-12-09 at 18:29 +0300, Vladislav Bolkhovitin wrote:
> 
>>>Additionally, it's perfectly possible for all of this to be done zero
>>>copy on the data.  A user space target mmaps the data on its storage
>>>device and then does a SG_IO type scatter gather user virtual region
>>>pass to the underlying target infrastructure.  We already have this
>>>demonstrated in the SG_IO path, someone just needs to come up with the
>>>correct implementation for a target path.
>>
>>I'm not completely understand how it will work. Consider, there are 
>>READ/WRITE commands with random data sizes arrive from an initiator. Are 
>>you going to do map/unmap for each command individually or alloc data 
>>buffers for commands from a premapped area and live with possible its 
>>fragmentation? If map/unmap individually, then I should say that those 
>>are very expensive operations.
> 
> You do it the same way an array does:  the model for SPI is read command
> phase, disconnect, process command (i.e. set up areas) reconnect for
> data transfer.
> 
> map/unmap are really only necessary if you're emulating the data store,
> but it's a fairly cheap operation on linux: it just causes the creation
> of a vm_area.  If it's a pass through, you can use SG_IO to pull it in
> and the SG_IO like output to shoot it out again, effectively using a
> piece of user memory as a zero copy buffer.
> 
> Fragmentation isn't an issue because the I/O goes via sg lists , all
> that's needed is a bunch of pages.

OK, I see what you meant, thanks.

> I do have to say that I consider operation in interrupt context (or even
> kernel context) to be a disadvantage.  Compared with the response times
> that most arrays have to SCSI commands, the kernel context switch time
> isn't that significant.

Are you sure that there are no now or will be available in the nearest 
feature such (eg iSCSI) SCSI arrays with response time/latency so small 
that having 5 (five) context switches or more per command, some of which 
include map/unmap operations, will not increase the latency too much? I 
mean, eg NFS server, which originally was user space daemon and many 
people didn't want it in the kernel. Eventually, it's in. I don't see 
any fundamental difference between NFS server and SCSI target server, 
especially considering that SCSI is synchroneous protocol, where the 
latency matters a lot, and usually there is some FS over exported SCSI 
device, which not always does bulk data operations.

Vlad


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[FUJITA Tomonori]

From: Vladislav Bolkhovitin <vst@vlnb.net>
Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?
Date: Sat, 10 Dec 2005 18:32:46 +0300

> >>>Also is the page cache comment in reference to us using the page cache 
> >>>for our reads and writes or I am not sure why you wrote that if you do 
> >>>not do it right now.
> >>
> >>Hm, at first, the page cache already used somehow in fileio dev handler 
> >>(though, with additional mem copy). At the second, fully utilize the 
> >>page cache is one of two major improvements that are pending in scst, 
> >>because it is required changing the kernel, which until some moment try 
> >>to avoid. Although I prepared what is necessary for that.
> >>
> >>The idea basically is the following. When READ operation arrives, pages 
> >>for all requested blocks are at first searched in the page cache 
> >>(probably, in SIRQ context, because it isn't expensive operation) and if 
> >>all pages are found, they are referenced and the result will be sent to 
> >>the initiator. Then the pages will be dereferenced (so, no pages 
> >>allocation will be done at all). Otherwise, the missed pages will be 
> >>allocated and the command will be rescheduled to the thread, which will 
> >>read them. Then, after the response is sent, the pages will remain in 
> >>the page cache for future accesses. For WRITEs the processing is the 
> >>similar, the pages with the data will be put in the page cache.
> > 
> > The Ardis iSCSI target code does the same thing.
> 
> Perfectly. So, why don't do it on the mid-layer level where all targets 
> can benefit from it?

Because I think that it not of much performance benefit.

Re: stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> James Bottomley wrote:
> 
>> On Fri, 2005-12-09 at 18:29 +0300, Vladislav Bolkhovitin wrote:
>>
>>>> Additionally, it's perfectly possible for all of this to be done zero
>>>> copy on the data.  A user space target mmaps the data on its storage
>>>> device and then does a SG_IO type scatter gather user virtual region
>>>> pass to the underlying target infrastructure.  We already have this
>>>> demonstrated in the SG_IO path, someone just needs to come up with the
>>>> correct implementation for a target path.
>>>
>>>
>>> I'm not completely understand how it will work. Consider, there are 
>>> READ/WRITE commands with random data sizes arrive from an initiator. 
>>> Are you going to do map/unmap for each command individually or alloc 
>>> data buffers for commands from a premapped area and live with 
>>> possible its fragmentation? If map/unmap individually, then I should 
>>> say that those are very expensive operations.
>>
>>
>> You do it the same way an array does:  the model for SPI is read command
>> phase, disconnect, process command (i.e. set up areas) reconnect for
>> data transfer.
>>
>> map/unmap are really only necessary if you're emulating the data store,
>> but it's a fairly cheap operation on linux: it just causes the creation
>> of a vm_area.  If it's a pass through, you can use SG_IO to pull it in
>> and the SG_IO like output to shoot it out again, effectively using a
>> piece of user memory as a zero copy buffer.
>>
>> Fragmentation isn't an issue because the I/O goes via sg lists , all
>> that's needed is a bunch of pages.
> 
> 
> OK, I see what you meant, thanks.
> 
>> I do have to say that I consider operation in interrupt context (or even
>> kernel context) to be a disadvantage.  Compared with the response times
>> that most arrays have to SCSI commands, the kernel context switch time
>> isn't that significant.
> 
> 
> Are you sure that there are no now or will be available in the nearest 
> feature such (eg iSCSI) SCSI arrays with response time/latency so small 
> that having 5 (five) context switches or more per command, some of which 
> include map/unmap operations, will not increase the latency too much? I 
> mean, eg NFS server, which originally was user space daemon and many 
> people didn't want it in the kernel. Eventually, it's in. I don't see 
> any fundamental difference between NFS server and SCSI target server, 

Isn't the reason a NFS server is still in the kernel is becuase some of 
the locking difficulties?

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> FUJITA Tomonori wrote:
> 
>> From: Vladislav Bolkhovitin <vst@vlnb.net>
>> Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new 
>> version of iscsi target?
>> Date: Fri, 09 Dec 2005 18:28:58 +0300
>>
>>
>>>> Also is the page cache comment in reference to us using the page 
>>>> cache for our reads and writes or I am not sure why you wrote that 
>>>> if you do not do it right now.
>>>
>>>
>>> Hm, at first, the page cache already used somehow in fileio dev 
>>> handler (though, with additional mem copy). At the second, fully 
>>> utilize the page cache is one of two major improvements that are 
>>> pending in scst, because it is required changing the kernel, which 
>>> until some moment try to avoid. Although I prepared what is necessary 
>>> for that.
>>>
>>> The idea basically is the following. When READ operation arrives, 
>>> pages for all requested blocks are at first searched in the page 
>>> cache (probably, in SIRQ context, because it isn't expensive 
>>> operation) and if all pages are found, they are referenced and the 
>>> result will be sent to the initiator. Then the pages will be 
>>> dereferenced (so, no pages allocation will be done at all). 
>>> Otherwise, the missed pages will be allocated and the command will be 
>>> rescheduled to the thread, which will read them. Then, after the 
>>> response is sent, the pages will remain in the page cache for future 
>>> accesses. For WRITEs the processing is the similar, the pages with 
>>> the data will be put in the page cache.
>>
>>
>> The Ardis iSCSI target code does the same thing.
> 
> 
> Perfectly. So, why don't do it on the mid-layer level where all targets 
> can benefit from it?
> 

Any target can hook into stgt too? What is your point since neither of 
us are in mainline or even close given the scsi guy's veiwpoint on where 
to do reads and writes?

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> Mike Christie wrote:
> 
>>>> There is still memory and scatterlist allocations. If we are not 
>>>> going to allocate all the memory for a command buffer and request 
>>>> with GFP_ATOMIC (and can then run from the the HW interrupt or soft 
>>>> irq) we have to pass that on to a thread. I guess there is 
>>>> disagreement whether that part is a feature or bad use of GFP_ATOMIC 
>>>> though so... But I just mean to say there could be a little more to do.
>>>
>>>
>>>
>>>
>>> Actually, there is the way to allocate sg vectors with buffers in 
>>> SIRQ and not with GFP_ATOMIC. This is the second major improvement, 
>>> which is pending in scst. I called it sgv_pool. This is a new 
>>> allocator in the kernel similar to mem_pool, but it contains 
>>> *complete* sg-vectors of some size with data buffers (pages). 
>>> Initiator sends data requests usually with some fixed size, like 
>>> 128K. After a data command completed, its sg vector will not be 
>>> immediately freed, but will be kept in 
>>
>>
>>
>> We considered this, but what did you decide is the upper limit size 
>> for the pool? Is it dynmaic? We also wanted something that the SCSI 
>> ULDs could use for their allocations which could go up to 6 MB.
> 
> 
> Why do you think it needs any upper limit size? Would you like also any 
> upper limits on sizes of the page or slab caches? I don't see any 

Considering I have not seen the code and am probably misunderstanding 
what you wrote above, I will shutup and wait until it is out to see it.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> Mike Christie wrote:
> 
>>> But there are other cleanups like moving some of the state to per 
>>> target, cleaningup the scattlist allocation code and moving it to 
>>> scsi-ml so the SCSI ULDs can use them and convert them. There is also 
>>> thing like converting to the right APIs for 2.6 (rm kernel_thread, rm 
>>> scsi_request, rm proc, fixup class interface refcouting problems, 
>>> fixup scsi_device lack of refcounting usage, etc).
>>>
>>
>> Oh yeah I think the other major issue at least I had with scst was 
>> that it was scsi specific and we wanted try and seperate things so if 
>> drivers like IET and vscsi are allowed then we could also do other 
>> drivers like a ATA over ethernet target driver or allow any other 
>> target driver that wanted to to hook in. I think you noted that we 
>> were spererating some protocol specific things as a distadvantage or 
>> mentioned it for some reason but I am not completely sure why and we 
>> may not agree on that issue too.
> 
> 
> SCSI has a lot of very specific stuff like UA handling and (at least) 
> some parts of task management, especially if we consider honoring NACA, 
> QErr, TST, UA_INTLCK_CTRL bits, therefore I'm not sure that to have 
> common target parts for other protocols worths complicating the 
> mid-layer with code and interfaces that will separate SCSI-specifics 
> from non-SCSI protocols. So, good luck with it :-)
> 

I am talking about the memory allocation bits for starters. For example 
the way you allocate the scatterlists and memory is useful to other 
parts of the kernel not just scst. The userspace interface, if we added 
the netlink parts to scst could be useful too.

Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
>>
>> Are you sure that there are no now or will be available in the nearest 
>> feature such (eg iSCSI) SCSI arrays with response time/latency so 
>> small that having 5 (five) context switches or more per command, some 
>> of which include map/unmap operations, will not increase the latency 
>> too much? I mean, eg NFS server, which originally was user space 
>> daemon and many people didn't want it in the kernel. Eventually, it's 
>> in. I don't see any fundamental difference between NFS server and SCSI 
>> target server, 
> 
> 
> Isn't the reason a NFS server is still in the kernel is becuase some of 
> the locking difficulties?

Might be. But from what I remember, the major reason was the 
performance. After googling a bit I found many acknowledgments of that.

Vlad


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: Ang: Re: [Stgt-devel] Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> Vladislav Bolkhovitin wrote:
> 
>> FUJITA Tomonori wrote:
>>
>>> From: Vladislav Bolkhovitin <vst@vlnb.net>
>>> Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new 
>>> version of iscsi target?
>>> Date: Fri, 09 Dec 2005 18:28:58 +0300
>>>
>>>
>>>>> Also is the page cache comment in reference to us using the page 
>>>>> cache for our reads and writes or I am not sure why you wrote that 
>>>>> if you do not do it right now.
>>>>
>>>>
>>>>
>>>> Hm, at first, the page cache already used somehow in fileio dev 
>>>> handler (though, with additional mem copy). At the second, fully 
>>>> utilize the page cache is one of two major improvements that are 
>>>> pending in scst, because it is required changing the kernel, which 
>>>> until some moment try to avoid. Although I prepared what is 
>>>> necessary for that.
>>>>
>>>> The idea basically is the following. When READ operation arrives, 
>>>> pages for all requested blocks are at first searched in the page 
>>>> cache (probably, in SIRQ context, because it isn't expensive 
>>>> operation) and if all pages are found, they are referenced and the 
>>>> result will be sent to the initiator. Then the pages will be 
>>>> dereferenced (so, no pages allocation will be done at all). 
>>>> Otherwise, the missed pages will be allocated and the command will 
>>>> be rescheduled to the thread, which will read them. Then, after the 
>>>> response is sent, the pages will remain in the page cache for future 
>>>> accesses. For WRITEs the processing is the similar, the pages with 
>>>> the data will be put in the page cache.
>>>
>>>
>>>
>>> The Ardis iSCSI target code does the same thing.
>>
>>
>>
>> Perfectly. So, why don't do it on the mid-layer level where all 
>> targets can benefit from it?
>>
> 
> Any target can hook into stgt too? What is your point since neither of
> us are in mainline or even close given the scsi guy's veiwpoint on where 
> to do reads and writes?

My point is to show one of the benefits of the kernel side 
implementation. Any target driver (of scst or stgt, doesn't matter) 
should benefit from it. Obviously, there is no point to overcomplicate 
any target driver with the functions of the the mid-level.

Vlad


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: [Scst-devel] Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

FUJITA Tomonori wrote:
> From: Vladislav Bolkhovitin <vst@vlnb.net>
> Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?
> Date: Sat, 10 Dec 2005 18:32:46 +0300
> 
> 
>>>>>Also is the page cache comment in reference to us using the page cache 
>>>>>for our reads and writes or I am not sure why you wrote that if you do 
>>>>>not do it right now.
>>>>
>>>>Hm, at first, the page cache already used somehow in fileio dev handler 
>>>>(though, with additional mem copy). At the second, fully utilize the 
>>>>page cache is one of two major improvements that are pending in scst, 
>>>>because it is required changing the kernel, which until some moment try 
>>>>to avoid. Although I prepared what is necessary for that.
>>>>
>>>>The idea basically is the following. When READ operation arrives, pages 
>>>>for all requested blocks are at first searched in the page cache 
>>>>(probably, in SIRQ context, because it isn't expensive operation) and if 
>>>>all pages are found, they are referenced and the result will be sent to 
>>>>the initiator. Then the pages will be dereferenced (so, no pages 
>>>>allocation will be done at all). Otherwise, the missed pages will be 
>>>>allocated and the command will be rescheduled to the thread, which will 
>>>>read them. Then, after the response is sent, the pages will remain in 
>>>>the page cache for future accesses. For WRITEs the processing is the 
>>>>similar, the pages with the data will be put in the page cache.
>>>
>>>The Ardis iSCSI target code does the same thing.
>>
>>Perfectly. So, why don't do it on the mid-layer level where all targets 
>>can benefit from it?
> 
> 
> Because I think that it not of much performance benefit.

Are you sure? Why do you think using the cache together with its nice 
features like read-ahead doesn't bring major performance benefits?

Vlad


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: stgt a new version of iscsi target?

[Mike Christie]

Vladislav Bolkhovitin wrote:
> Mike Christie wrote:
> 
>>>
>>> Are you sure that there are no now or will be available in the 
>>> nearest feature such (eg iSCSI) SCSI arrays with response 
>>> time/latency so small that having 5 (five) context switches or more 
>>> per command, some of which include map/unmap operations, will not 
>>> increase the latency too much? I mean, eg NFS server, which 
>>> originally was user space daemon and many people didn't want it in 
>>> the kernel. Eventually, it's in. I don't see any fundamental 
>>> difference between NFS server and SCSI target server, 
>>
>>
>>
>> Isn't the reason a NFS server is still in the kernel is becuase some 
>> of the locking difficulties?
> 
> 
> Might be. But from what I remember, the major reason was the 
> performance. After googling a bit I found many acknowledgments of that.
> 

I do not think we are going to get anywhere with this type of thread :(

We should try to compare at least one of the userspace *nbd 
implementations with the unh target in scst. I see some that just do 
some basic socket ops (no sendfile type hook in even) for the network 
part then just async or normal read/writes. I do not want to comapre FC 
to nbd, but maybe comparing software iscsi to userspace nbd is a little 
more fair. I think ata over ethernet has a userspace target too. Is the 
unh target defaults set ok for performance testing, or could you send 
some off list, so we can at least test those.


-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

Mike Christie wrote:
> Vladislav Bolkhovitin wrote:
> 
>> Mike Christie wrote:
>>
>>>>
>>>> Are you sure that there are no now or will be available in the 
>>>> nearest feature such (eg iSCSI) SCSI arrays with response 
>>>> time/latency so small that having 5 (five) context switches or more 
>>>> per command, some of which include map/unmap operations, will not 
>>>> increase the latency too much? I mean, eg NFS server, which 
>>>> originally was user space daemon and many people didn't want it in 
>>>> the kernel. Eventually, it's in. I don't see any fundamental 
>>>> difference between NFS server and SCSI target server, 
>>>
>>>
>>>
>>>
>>> Isn't the reason a NFS server is still in the kernel is becuase some 
>>> of the locking difficulties?
>>
>>
>>
>> Might be. But from what I remember, the major reason was the 
>> performance. After googling a bit I found many acknowledgments of that.
>>
> 
> I do not think we are going to get anywhere with this type of thread :(
> 
> We should try to compare at least one of the userspace *nbd 
> implementations with the unh target in scst. I see some that just do 
> some basic socket ops (no sendfile type hook in even) for the network 
> part then just async or normal read/writes. I do not want to comapre FC 
> to nbd, but maybe comparing software iscsi to userspace nbd is a little 
> more fair. I think ata over ethernet has a userspace target too. Is the 
> unh target defaults set ok for performance testing, or could you send 
> some off list, so we can at least test those.

Agree that we need to have some numbers. But currently it is impossible 
  to measure them correctly without very considerable effort. For 
instance, the comparision of nbd with iscsi includes in the measurements 
not only user space/kernel space differences, but also many additional 
parts, like different implementation architectures. For the correct 
comparison we need some target driver (for scst or sgt), which commands 
would be processed in both user and kernel space. Additionally, because 
we discuss not only user vs kernel implementations, but also SIRQ vs 
thread implementations, the target needs to be the hardware one.

Right now without big effort we can only compare SIRQ vs thread 
implementations over FC, because the QLA target driver and scst support 
both modes of SCSI commands execution. See DEBUG_WORK_IN_THREAD symbol. 
We did some comparisons some time ago and, if I recall correctly, on 
small blocks (especially 16K and smaller) the performance drop was quite 
visible, because ~40000+ cs/sec are not very good for the system health 
:). You can easily repeat those experiments using scst, the qlogic 
driver and disk_perf or tape_perf dev handler.

But, since FC has quite a big latencies, this comparision will not fully 
suit our needs. We need some some low latency link. Probably, some of 
hardware iSCSI cards, like Qlogic 4100. But this is not the nearest future.

Vlad




-------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc. Do you grep through log files
for problems?  Stop!  Download the new AJAX search engine that makes
searching your log files as easy as surfing the  web.  DOWNLOAD SPLUNK!
http://ads.osdn.com/?ad_id=7637&alloc_id=16865&op=click

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[FUJITA Tomonori]

From: James Bottomley <James.Bottomley@SteelEye.com>
Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new	version of iscsi target?
Date: Thu, 08 Dec 2005 14:48:10 -0500

> On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
> > cleanup. In the end some of the scsi people liked the idea of throwing 
> > the non-read/write command to userspace and to do this we just decided 
> > to start over but I have been cutting and pasting your code and cleaning 
> > it up as I add more stuff.
> 
> To be honest, I'd like to see all command processing at user level
> (including read/write ... for block devices, it shouldn't be that
> inefficient, since you're merely going to say remap an area from one
> device to another; as long as no data transformation ever occurs, the
> user never touches the data and it all remains in the kernel page
> cache).

The current version of tgt that performs READ/WRITE commands in kernel
space using the vfs interface and other commands in user space. I'd
implemented a prototype version of tgt with the mmap scheme.

With the mmap tgt, the kernel module asks a single user-space daemon
(tgtd) to map a file (logical unit) through netlink, then it call
get_user_pages().

I did some initial performance tests with both tgt versions and IET
(as you know, another iSCSI software implementation runs in kernel
space). All implementations run with write back policy so probably,
there should be little real disk I/O effect. I started disktest
benchmark software with cold cache state. The machine has 4 GB memory,
1.5K SCSI disk, 4 CPUs (x86_64).

(disktest -PT -T10 -h1 -K8 -B8192 -ID /dev/sdc -w)

o IET
| 2005/12/21-18:05:15 | STAT  | 7259 | v1.2.8 | /dev/sdc | Total write throughput: 48195993.6B/s (45.96MB/s), IOPS 5883.3/s.

o tgt (I/O in kernel space)
| 2005/12/21-18:03:23 | STAT  | 7013 | v1.2.8 | /dev/sdc | Total write throughput: 45829324.8B/s (43.71MB/s), IOPS 5594.4/s.

o mmap tgt
| 2005/12/21-18:22:28 | STAT  | 7990 | v1.2.8 | /dev/sdc | Total write throughput: 25373900.8B/s (24.20MB/s), IOPS 3097.4/s.


I guess that one of the reasons for the mmap tgt poor performance is
that it uses single user-space daemon so all commands are
serialized. I will implement multi-threaded user-space daemon version
if you like to see its performance.

One potential disadvantage of the mmap scheme is that it reads
unnecessarily from disk about WRITE commands. That is, when it updates
the whole page frame, the vfs interface can avoid reading cleverly,
however, the mmap scheme cannot (If I understand correctly).

Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Vladislav Bolkhovitin]

FUJITA,

It would be helpful if you also gather CPU utilization statistics during 
the tests (user/kernel/idle/iowait time + context switches/sec rate).

Vlad

FUJITA Tomonori wrote:
> I did some initial performance tests with both tgt versions and IET
> (as you know, another iSCSI software implementation runs in kernel
> space). All implementations run with write back policy so probably,
> there should be little real disk I/O effect. I started disktest
> benchmark software with cold cache state. The machine has 4 GB memory,
> 1.5K SCSI disk, 4 CPUs (x86_64).
> 
> (disktest -PT -T10 -h1 -K8 -B8192 -ID /dev/sdc -w)
> 
> o IET
> | 2005/12/21-18:05:15 | STAT  | 7259 | v1.2.8 | /dev/sdc | Total write throughput: 48195993.6B/s (45.96MB/s), IOPS 5883.3/s.
> 
> o tgt (I/O in kernel space)
> | 2005/12/21-18:03:23 | STAT  | 7013 | v1.2.8 | /dev/sdc | Total write throughput: 45829324.8B/s (43.71MB/s), IOPS 5594.4/s.
> 
> o mmap tgt
> | 2005/12/21-18:22:28 | STAT  | 7990 | v1.2.8 | /dev/sdc | Total write throughput: 25373900.8B/s (24.20MB/s), IOPS 3097.4/s.

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[FUJITA Tomonori]

From: James Bottomley <James.Bottomley@SteelEye.com>
Subject: Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new	version of iscsi target?
Date: Thu, 08 Dec 2005 14:48:10 -0500

> On Thu, 2005-12-08 at 13:10 -0600, Mike Christie wrote:
> > cleanup. In the end some of the scsi people liked the idea of throwing 
> > the non-read/write command to userspace and to do this we just decided 
> > to start over but I have been cutting and pasting your code and cleaning 
> > it up as I add more stuff.
> 
> To be honest, I'd like to see all command processing at user level
> (including read/write ... for block devices, it shouldn't be that
> inefficient, since you're merely going to say remap an area from one
> device to another; as long as no data transformation ever occurs, the
> user never touches the data and it all remains in the kernel page
> cache).
> 
> My ideal for the kernel based infrastructure is a simple tap for
> transporting commands addressed to devices upwards (and the responses
> downwards).  Then everyone can have their own user space processing
> implementation that I don't have to care about.

Mike and I have worked on the tgt mmap version.

o It does read/write commands like sg by using mmap in user space and
get_user_pages in kernel space.

o It does non-read/write commands like direct I/O by allocating
aligned buffers in user space and using get_user_pages in kernel space.

It works like the simple tap that you suggested. It does not allocate
buffers in kernel space at all and does zero copy on all sorts of
commands.

Here are some performance results with open-iscsi (which are better
than the previous results that I got with sfnet).

o IET

| 2005/12/27-07:50:59 | STAT  | 6827 | v1.2.8 | /dev/sdc | Total write throughput: 53790310.4B/s (51.30MB/s), IOPS 6566.2/s.

o current tgt (I/O in kernel space)

| 2005/12/27-08:07:50 | STAT  | 7294 | v1.2.8 | /dev/sdc | Total write throughput: 49666457.6B/s (47.37MB/s), IOPS 6062.8/s.

o tgt mmap

| 2005/12/27-08:42:51 | STAT  | 5286 | v1.2.8 | /dev/sdc | Total write
throughput: 44701286.4B/s (42.63MB/s), IOPS 5456.7/s.

We can get something like this if we avoid calling mmap/munmap per
command (by using some sorts of caching).

o tgt mmap (mmap caching)

| 2005/12/27-07:53:19 | STAT  | 6996 | v1.2.8 | /dev/sdc | Total write throughput: 48253337.6B/s (46.02MB/s), IOPS 5890.3/s.


James, can we get your approval of the this mmap design?

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[James Bottomley]

On Tue, 2005-12-27 at 08:53 +0900, FUJITA Tomonori wrote:
> Mike and I have worked on the tgt mmap version.
> 
> o It does read/write commands like sg by using mmap in user space and
> get_user_pages in kernel space.
> 
> o It does non-read/write commands like direct I/O by allocating
> aligned buffers in user space and using get_user_pages in kernel space.
> 
> It works like the simple tap that you suggested. It does not allocate
> buffers in kernel space at all and does zero copy on all sorts of
> commands.
> 
> Here are some performance results with open-iscsi (which are better
> than the previous results that I got with sfnet).
> 
> o IET
> 
> | 2005/12/27-07:50:59 | STAT  | 6827 | v1.2.8 | /dev/sdc | Total write throughput: 53790310.4B/s (51.30MB/s), IOPS 6566.2/s.
> 
> o current tgt (I/O in kernel space)
> 
> | 2005/12/27-08:07:50 | STAT  | 7294 | v1.2.8 | /dev/sdc | Total write throughput: 49666457.6B/s (47.37MB/s), IOPS 6062.8/s.
> 
> o tgt mmap
> 
> | 2005/12/27-08:42:51 | STAT  | 5286 | v1.2.8 | /dev/sdc | Total write
> throughput: 44701286.4B/s (42.63MB/s), IOPS 5456.7/s.
> 
> We can get something like this if we avoid calling mmap/munmap per
> command (by using some sorts of caching).
> 
> o tgt mmap (mmap caching)
> 
> | 2005/12/27-07:53:19 | STAT  | 6996 | v1.2.8 | /dev/sdc | Total write throughput: 48253337.6B/s (46.02MB/s), IOPS 5890.3/s.
> 
> 
> James, can we get your approval of the this mmap design?

Yes, that looks fine ... it runs in user space, which was really all I
was looking for.

There is another half to this, which is that I'd like the tap to come
via a SCSI API.  This isn't strictly necessary for iSCSI but it would
allow us to integrate a generic target approach that could work for all
SCSI HBA's as well as just iSCSI.

Thanks,

James

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[Mike Christie]

James Bottomley wrote:
> On Tue, 2005-12-27 at 08:53 +0900, FUJITA Tomonori wrote:
> 
>>Mike and I have worked on the tgt mmap version.
>>
>>o It does read/write commands like sg by using mmap in user space and
>>get_user_pages in kernel space.
>>
>>o It does non-read/write commands like direct I/O by allocating
>>aligned buffers in user space and using get_user_pages in kernel space.
>>
>>It works like the simple tap that you suggested. It does not allocate
>>buffers in kernel space at all and does zero copy on all sorts of
>>commands.
>>
>>Here are some performance results with open-iscsi (which are better
>>than the previous results that I got with sfnet).
>>
>>o IET
>>
>>| 2005/12/27-07:50:59 | STAT  | 6827 | v1.2.8 | /dev/sdc | Total write throughput: 53790310.4B/s (51.30MB/s), IOPS 6566.2/s.
>>
>>o current tgt (I/O in kernel space)
>>
>>| 2005/12/27-08:07:50 | STAT  | 7294 | v1.2.8 | /dev/sdc | Total write throughput: 49666457.6B/s (47.37MB/s), IOPS 6062.8/s.
>>
>>o tgt mmap
>>
>>| 2005/12/27-08:42:51 | STAT  | 5286 | v1.2.8 | /dev/sdc | Total write
>>throughput: 44701286.4B/s (42.63MB/s), IOPS 5456.7/s.
>>
>>We can get something like this if we avoid calling mmap/munmap per
>>command (by using some sorts of caching).
>>
>>o tgt mmap (mmap caching)
>>
>>| 2005/12/27-07:53:19 | STAT  | 6996 | v1.2.8 | /dev/sdc | Total write throughput: 48253337.6B/s (46.02MB/s), IOPS 5890.3/s.
>>
>>
>>James, can we get your approval of the this mmap design?
> 
> 
> Yes, that looks fine ... it runs in user space, which was really all I
> was looking for.
> 
> There is another half to this, which is that I'd like the tap to come
> via a SCSI API.  This isn't strictly necessary for iSCSI but it would
> allow us to integrate a generic target approach that could work for all
> SCSI HBA's as well as just iSCSI.
> 

The code we currently have is designed to work with software iscsi 
targets or software AoE and HW cards like qlogic or emulex's FC cards. 
There are a lot of places we could use scsi-ml or block layer structs 
like the request or scsi_cmnd.

To support HW like qlogic or emulex's FC target mode, are you thinking 
you might want us to add on to the scsi-ml's scsi_host_template or add a 
scsi_target_template? If we add on to the scsi_host_template and if that 
one PCI device would be in initiator and target mode at the same time 
would we have one scsi_host for that resource and just add our target 
related fields to the scsi_host? Is this what you mean?

Re: Ang: Re: [Stgt-devel] Re: [Iscsitarget-devel] stgt a new version of iscsi target?

[James Bottomley]

On Fri, 2005-12-30 at 21:27 -0600, Mike Christie wrote:
> > Yes, that looks fine ... it runs in user space, which was really all I
> > was looking for.
> > 
> > There is another half to this, which is that I'd like the tap to come
> > via a SCSI API.  This isn't strictly necessary for iSCSI but it would
> > allow us to integrate a generic target approach that could work for all
> > SCSI HBA's as well as just iSCSI.
> > 
> 
> The code we currently have is designed to work with software iscsi 
> targets or software AoE and HW cards like qlogic or emulex's FC cards. 
> There are a lot of places we could use scsi-ml or block layer structs 
> like the request or scsi_cmnd.
> 
> To support HW like qlogic or emulex's FC target mode, are you thinking 
> you might want us to add on to the scsi-ml's scsi_host_template or add a 
> scsi_target_template? If we add on to the scsi_host_template and if that 
> one PCI device would be in initiator and target mode at the same time 
> would we have one scsi_host for that resource and just add our target 
> related fields to the scsi_host? Is this what you mean?

I'm thinking one device would do both intiator and target (although not
at the same time, but probably via some sort of internal role change
mechanism---Although that would be up to the driver writer; it could
certainly be set up to be initiator or target only) we probably need one
or two additional callbacks for sending incoming commands upwards and a
control channel for specifying what we do next (since for write
commands, we need command first, then userspace processing and setup
then body into allocated buffer).  The idea is that at the end of the
project we have a well defined target infrastructure for any SCSI device
(with an iSCSI reference implementation).

James