Skip to contentSkip to conversation

Autosense requirement

DISCUSSION Luben Tuikov 6 participants 11 messages Jul 29, 2002 – Jul 31, 2002 SCSI

lore
Share & tools
lore
#0 lore raw

How safe am I to assume that SCSI LLDD will return autosense for a failed/incomplete command, as per SCSI core (mid-layer) error handling requrement? (e.g. ide-scsi doesn't return autosense)

In drafts I'm more and more seeing
``targets MUST support and enable autosense.''

If indeed we put a ``MUST'' requirement on SCSI LLDD to support and enable autosense, which will indeed streamline things, then this will save me a an _irq_ spin lock over a simple spin lock...

What is the future of this requrement?

Signature · 1 line
Luben
#1 lore raw
Quoted 9 lines · Luben Tuikov

How safe am I to assume that SCSI LLDD will return autosense for a failed/incomplete command, as per SCSI core (mid-layer) error handling requrement? (e.g. ide-scsi doesn't return autosense)

In drafts I'm more and more seeing
``targets MUST support and enable autosense.''

Quoted 6 lines

If indeed we put a ``MUST'' requirement on SCSI LLDD to support and enable autosense, which will indeed streamline things, then this will save me a an _irq_ spin lock over a simple spin lock...

What is the future of this requrement?

Luben,
In the lk 2.4 series we have both older "non-eh" lower level drivers and a few, newer "eh" drivers. In the lk 2.5 series the mid level interface to the older "non-eh" drivers was dropped.

Hence in the development lk 2.5 series there are many stranded drivers, including ide-scsi, some of which have been converted "on the fly". His is an incomplete list of issues to be addressed in lower level (HBA) driver conversions:
- "eh" interface [mandatory]
- io_request_lock displaced by host_lock [mandatory]
- PCI memory/DMA interface
- driverfs aware
- autosense for high performance drivers (perhaps all)

As you point out autosense support (at the device level) is now mandatory in recent scsi drafts/standards. For simplicity of queueing logic, the sense buffer should not be left for the mid level to fetch (with a REQUEST SENSE).

I have attempted the conversion on several HBA drivers (of which scsi_debug is the only one I maintain). As Ben LaHaise has pointed out, I have skipped addressing the "PCI memory/DMA interface" so far.

Until Doug Ledford pointed out recently what lower level drivers should do with regard to autosense, I was unaware what to do. Since that time I have converted scsi_debug to do autosense properly and submitted a patch for the drivers/scsi/scsi_mid_low_api.txt documentation file.

Now would someone else like to step up to the plate (or "wicket" to use the cricket metaphor) and try and get the ide-scsi driver in lk 2.5 to do autosense?

If kernel people think that a ide-scsi "like" driver is going away then perhaps they should consider the new SAS (SCSI over Serial ATA infrastructure) draft. They have one mode illustrated in that document where the initiator is a SCSI host and it talks to multiple serial ATA disks via an "expander". Could be interesting.

Doug Gilbert

#2 lore raw
Quoted 10 lines · Douglas Gilbert

Hence in the development lk 2.5 series there are many stranded drivers, including ide-scsi, some of which have been converted "on the fly". His is an incomplete list of issues to be addressed in lower level (HBA) driver conversions:
- "eh" interface [mandatory]
- io_request_lock displaced by host_lock [mandatory]
- PCI memory/DMA interface
- driverfs aware
- autosense for high performance drivers (perhaps all)

Quoted 4 lines

As you point out autosense support (at the device level) is now mandatory in recent scsi drafts/standards. For simplicity of queueing logic, the sense buffer should not be left for the mid level to fetch (with a REQUEST SENSE).

It will be mandatory for the LLDD to do autosense. All remnants of code to perform a follow up request sense operation *will* dissappear from the mid layer. OTOH, the sense command and sense buffer will be pre-set for each device so at least the LLDD won't have to build their own command any longer, they can just change the command pointer from the original command to the sense command and the data buffer to the sense buffer and reissue the request.

Quoted 4 lines

I have attempted the conversion on several HBA drivers (of which scsi_debug is the only one I maintain). As Ben LaHaise has pointed out, I have skipped addressing the "PCI memory/DMA interface" so far.

Feel lucky. It's currently pissing me off and I'm thinking of changing it.

Signature · 4 lines
  Doug Ledford <[email protected]>     919-754-3700 x44233
         Red Hat, Inc. 
         1801 Varsity Dr.
         Raleigh, NC 27606
#3 lore raw
Quoted 7 lines · Doug Ledford [[email protected]] wrote:

It will be mandatory for the LLDD to do autosense. All remnants of code to perform a follow up request sense operation *will* dissappear from the mid layer. OTOH, the sense command and sense buffer will be pre-set for each device so at least the LLDD won't have to build their own command any longer, they can just change the command pointer from the original command to the sense command and the data buffer to the sense buffer and reissue the request.

Doug,
I assume from your message that you are removing the call from scsi_unjam_host to do request sense (which is good).

Does this imply that your reworking all of the eh handling in scsi_error.c?

I was doing some rework in this area, but it sounds like it might be counter to your work.

-Mike

Signature · 2 lines
Michael Anderson
[email protected]
#4 lore raw
Quoted 6 lines · Mike Anderson

Doug,
I assume from your message that you are removing the call from scsi_unjam_host to do request sense (which is good).

Does this imply that your reworking all of the eh handling in scsi_error.c?

Yeah, I had some stuff I want to do there. However, I'm currently working on other things right now. I was saving this for later :-/

Quoted 2 lines

I was doing some rework in this area, but it sounds like it might be counter to your work.

Go ahead. I'm not into that section yet.

Signature · 4 lines
  Doug Ledford <[email protected]>     919-754-3700 x44233
         Red Hat, Inc. 
         1801 Varsity Dr.
         Raleigh, NC 27606
#5 lore raw
Quoted 8 lines · Doug Ledford
Quoted 6 lines · Mike Anderson

Doug,
I assume from your message that you are removing the call from scsi_unjam_host to do request sense (which is good).

Does this imply that your reworking all of the eh handling in scsi_error.c?

Quoted 7 lines

Yeah, I had some stuff I want to do there. However, I'm currently working on other things right now. I was saving this for later :-/

Quoted 2 lines

I was doing some rework in this area, but it sounds like it might be counter to your work.

Go ahead. I'm not into that section yet.

Doug -

I'm also doing some scsi changes - scsi_scan.c cleanup, and multi-path IO support.

Are you doing work that affects scsi_scan.c?

Do you have a tree or such that we can sync up with?

Any moderate changes to scsi will likely affect the multi-path changes.

-- Patrick Mansfield

#6 lore raw
Quoted 4 lines · Patrick Mansfield

Doug -

I'm also doing some scsi changes - scsi_scan.c cleanup, and multi-path IO support.

I saw the posts you made on this. I was just getting ready to look at/integrate your scsi_scan.c changes then add on the additional stuff I had in mind. I thought I saw you say you were forward porting the changes to 2.5.28 so it will apply cleanly. Once I saw that patch come out I was going to grab it.

Quoted 1 line

Do you have a tree or such that we can sync up with?

Not anything available yet (it's too broken right now).

Quoted 1 line

Any moderate changes to scsi will likely affect the multi-path changes.

Yep, that I can see for sure...

Signature · 4 lines
  Doug Ledford <[email protected]>     919-754-3700 x44233
         Red Hat, Inc. 
         1801 Varsity Dr.
         Raleigh, NC 27606
#7 lore raw
Quoted 7 lines · Doug Ledford

On Mon, Jul 29, 2002 at 03:10:27PM -0700, Patrick Mansfield wrote:

I saw the posts you made on this. I was just getting ready to look at/integrate your scsi_scan.c changes then add on the additional stuff I had in mind. I thought I saw you say you were forward porting the changes to 2.5.28 so it will apply cleanly. Once I saw that patch come out I was going to grab it.

I had some 2.5.28/29 problems (I didn't set the new CONFIG_XXX_SERIAL options, and it took me a bit to realize the system was really booting even though I didn't see any serial line output), I hope to post the patch (against 2.5.29, but it will apply to 2.5.28) by Wednesday.

Quoted 7 lines
Quoted 1 line

Do you have a tree or such that we can sync up with?

Not anything available yet (it's too broken right now).

Quoted 1 line

Any moderate changes to scsi will likely affect the multi-path changes.

Yep, that I can see for sure...

I have some patches from the multi-path that I could break out, especially some list iterators - multi-path needs the Scsi_Device separated from Scsi_Host. Iterators were added so the underlying structures are not visible - that is, so host_queue is not directly referenced, and I can put any type of data structure in its place. This is one of the broadest changes (potentially affects all adapter drivers) in the multi-path patch.

Having Scsi_Device on its own list could also help in hot-replacement of single pathed devices (Scsi_Host can be removed without having to remove a Scsi_Device off of its host_queue).

If you make changes in the host/device list(s) please consider that I need to have the Scsi_Devices on their own list separate from the Scsi_host list.

Thanks.

-- Patrick Mansfield

#8 lore raw
Quoted 7 lines · [email protected] said:

I have some patches from the multi-path that I could break out, especially some list iterators - multi-path needs the Scsi_Device separated from Scsi_Host. Iterators were added so the underlying structures are not visible - that is, so host_queue is not directly referenced, and I can put any type of data structure in its place. This is one of the broadest changes (potentially affects all adapter drivers) in the multi-path patch.

I was planning to introduce the Scsi_[Host,Device,Cmnd] structures to the concept of struct list_head as part of the error handler re-write. The error handler needs lists for the concept of reset locality (i.e. dump all commands for all LUNS of this target, dump all commands for all devices on this bus), although its a non time critical section, so it wouldn't necessarily matter having to have it loop over every device to find only the ones it's interested in.

Quoted 3 lines

Having Scsi_Device on its own list could also help in hot-replacement of single pathed devices (Scsi_Host can be removed without having to remove a Scsi_Device off of its host_queue).

Isn't this going to cause serious problems for the driverfs stuff? Effectively you're changing the parent pointer under the device which alters the layout of the device tree.

Quoted 3 lines

If you make changes in the host/device list(s) please consider that I need to have the Scsi_Devices on their own list separate from the Scsi_host list.

Could you post some of the details of this (separated from the multi-path patch if possible)? Thanks,

James

#9 lore raw

James -

Quoted 8 lines · James Bottomley
Quoted 3 lines · [email protected] said:

Having Scsi_Device on its own list could also help in hot-replacement of single pathed devices (Scsi_Host can be removed without having to remove a Scsi_Device off of its host_queue).

Isn't this going to cause serious problems for the driverfs stuff? Effectively you're changing the parent pointer under the device which alters the layout of the device tree.

Yes, for any multi-path implementation, driverfs needs changes so that a child can have multiple parents, or something along those lines.

For the next version of the patch, I'm planning on merging with the current identifier code - the value put into the driverfs "name", and then I plan on looking into driverfs changes.

Perhaps driverfs could seperate device registration (or creation) from link (or path) registration.

Quoted 8 lines
Quoted 3 lines

If you make changes in the host/device list(s) please consider that I need to have the Scsi_Devices on their own list separate from the Scsi_host list.

Could you post some of the details of this (separated from the multi-path patch if possible)? Thanks,

James

There is a document here:

http://www-124.ibm.com/storageio/multipath/scsi-multipath/docs/index.html

It describes the current host_queue list (I think you had similiar slides at OLS), and the new list introduced by the multi-path patch, mainly to show the path structures, not the Scsi_Device list.

Generally, if we want one Scsi_Device per actual LUN, and a LUN is connected to multiple adapters, the Scsi_Device can't easily be linked off of Scsi_Host - there is no longer one Scsi_Host for a given Scsi_Device.

So, I added a list of Scsi_Devices; on each Scsi_Device there is a list of paths, where each path includes a pointer back to Scsi_Host.

On removal of a single Scsi_Host (if there was such code), all paths pointing to the host must first be removed, and Scsi_Device can remain (even with no left paths).

If I linked off of Scsi_Host, there would have to be a separate list structure so that a Scsi_Device could be on multiple Scsi_Host lists. This could still be done even with a separate Scsi_Device list, but is only usefull if we need a fast way to lookup all Scsi_Devices on a host.

-- Patrick Mansfield

#10 lore raw
In reply to Doug Ledford
Quoted 8 lines · Doug Ledford

It will be mandatory for the LLDD to do autosense. All remnants of code to perform a follow up request sense operation *will* dissappear from the mid layer. OTOH, the sense command and sense buffer will be pre-set for each device so at least the LLDD won't have to build their own command any longer, they can just change the command pointer from the original command to the sense command and the data buffer to the sense buffer and reissue the request.

Oh, this is soooo nice to hear.

Signature · 1 line
Luben