Mukker, Atul wrote:
Thanks for restating my original question.
1. What interface should be used by the HBA management applications to obtain (non-generic) information from the HBA?
My opinions:
sysfs :
Pro: Good for singular data items and simple status (link state, f/w
rev, etc).
Very good for things that really don't need a tool (simplistic
admin commands):
(show state, reset board, etc).
Con: Doesn't work well for "transactions" that need multiple data elements
Lack of insight to process life cycle, thus multi-step and
concurrent
transactions difficult.
Doesn't work with binary data, buffers, etc.
Difficult to use concurrently by multiple processes.
Can't push async info to user.
No support for complex things.
The list of attributes can get big. Not a big deal, but...
Security based on attribute permissions (not always the best
model)
configfs:
Pro: Basically sysfs but for transactions with multiple data elements
Con: (same as sysfs, just minus multiple data element con).
netlink:
Pro: Very good for "multi-cast" operations - pushing async events to
multiple
receivers.
Handles requests and responses with multiple data elements easily.
Can track per-process life cycles.
Socket based so could even support mgmt from a different machine.
Security checking easy to build in.
Con: Doesn't work well for large payloads.
Payloads can't be referenced via data pointer (they need to be
inline to the pkt).
Direct DMA not supported - has to be staged to driver buffer,
copied in/out
of socket.
Multi-step transactions doable, but difficult. Maintaining
relationships per
pid difficult.
Multiple machines means dealing with endian-ness and data typing.
The netlink sockets do have memory-related issues that must be
watched.
Note: to not burn NETLINK id space, and perhaps collide in different
distro
kernels, please use the mid-layers netlink infrastructure, which
does allow
driver-specific messaging.
bsg:
(Specifically the new midlayer sgio support that was recently added
for ELS passthru)
Pro: Support requests and responses with multiple data elements easily
Supports separate request and response DMA-able payload buffers
Supports big payloads easily
Con: Lack of insight to process lifecycle, thus multi-step and concurrent
transactions difficult.
Async response generation (w/o associated request) very difficult.
It's really a wrappered ioctl, with the midlayer protecting
the kernel from
bad ioctl practice via the way it converts the sgio ioctl
into a midlayer
request. Creates an odd programming interface, as you
really want to
wrapper the ioctl on the user side too.
Thus, when you look across the pros and cons, its easy to see why the
transport
is using different things for different purposes.
2. How should driver notify such applications of asynchronous events happening on the HBA?
This is already there with the midlayer netlink support. Vendor-unique
events
are already supported.
Please keep in mind, all the data transfer between the applications and the HBA is a private protocol.
Private or not, the code for the interface use will have to be in the
driver. Code will
be inspected for proper/safe usage of the interfaces. Coding such that
things in the
messaging are black-boxes will always be a point of contention.
-- james s
Thanks
Atul Mukker
-----Original Message-----
From: James Smart [mailto:James.Smart@xxxxxxxxxx]
Sent: Monday, July 20, 2009 12:58 PM
To: Mukker, Atul
Cc: Brian King; linux-scsi@xxxxxxxxxxxxxxx
Subject: Re: Recommended HBA management interfaces
FYI - netlink (and sysfs, and I believe debugfs) do not exist with
vmware drivers... Additionally, with netlink, many of the distros no
longer include libnl by default in their install images. Even
interfaces that you think exist on vmware, may have very different
semantical behavior (almost all of the transport stuff either doesn't
exist or is only partially implemented).
One big caveat I'd give you: It's not so much the interface being used,
but rather, what are you doing over the interface. One of the goals of
the community is to present a consistent management paradigm for like
things. Thus, if what you are doing is generic, you should do it in a
generic manner so that all drivers for like hardware can utilize it.
This was the motivation for the protocol transports. Interestingly, even
the transports use different interfaces for different things. It all
depends on what it is.
Lastly, some things are considered bad practice from a kernel safety
point of view. Example: driver-specific ioctls passing around user-space
buffer pointers. In these cases, it doesn't matter what interface you
pick, they'll be rejected.
-- james s
Mukker, Atul wrote:
Thanks Brian. Netlink seems to be appropriate for our purpose as well,
almost too good :-)
That make me think, what's the catch? The SCSI drivers are not heavy
usage of this interface for one.
Are the other caveats associated with it?
Best regards,
Atul Mukker
-----Original Message-----
From: Brian King [mailto:brking@xxxxxxxxxxxxxxxxxx]
Sent: Friday, July 17, 2009 11:36 AM
To: Mukker, Atul
Cc: linux-scsi@xxxxxxxxxxxxxxx
Subject: Re: Recommended HBA management interfaces
Mukker, Atul wrote:
Hi All,
We would like expert comments on the following questions regarding
management of HBA from applications.
Traditionally, our drivers create a character device node, whose
file_operations are then used by the management applications to
transfer HBA specific commands. In addition to being quirky, this
interface has a few limitations which we would like to remove, most
important being able to seamlessly handle asynchronous events with
data transfer.
1. What is (are) the other standard/recommended interfaces which
applications can use to transfer HBA specific commands and data.
Depends on what the commands look like. With ipr, the commands that
the management application need to send to the HBA look sufficiently
like SCSI that I was able to report an sg device node for the adapter
and use SG_IO to send these commands.
sysfs, debugfs, and configfs are options as well.
2. How should an LLD implement interfaces to transmit asynchronous
information to the management applications? The requirement is to be
able to transmit data buffer as well as notifications for events.
I've had good success with netlink. In my use I only send a
notification
to userspace and let the application send some commands to figure out
what happened, but netlink does allow to send data as well. It makes it
very
easy to have multiple concurrent readers of the data, which I've found
very
useful.
3. The interface should be able to work even if no SCSI devices are
exported to the kernel.
netlink allows this.
4. Should work seamlessly across vmware and xen kernels.
netlink should work here too.
-Brian
--
Brian King
Linux on Power Virtualization
IBM Linux Technology Center
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html