EDAC a new sysfs 'subsystem' under /sys/devices
EDAC is for detecting and logging memory ECC, PCI Parity and (future) other hardware errors.
In my process of getting edac ready for kernel submission:
I have been exploring mechanisms for placing EDAC controls and attribute files into sysfs for a few days now.
In the process, I "discovered" how sysfs subystems are currently used. I have determined that EDAC itself should be a subsystem, since more than one type of EDAC device can be created.
I made a copy of drivers/base/sys.c to a new file, drivers/base/edac.c (and corresponding .h file) and refactored it to become a 'edacdev' subsystem.
I have mounted edac to: /sys/devices/edac because the /sys/device/system subsystem is not exported (at the moment. I suppose that could be changed).
I now have:
/sys/devices/edac/mc/mc0/csrow0 kobjects working and am adding attribute files to their respective kobjects.
My RFC is:
Should I leave edac under /sys/devices OR refactor drivers/base/sys.c and export the "system" subsystem and mount edac under /sys/devices/system?
Another RFC is:
Should EDAC really have a "new" subsystem in the kernel? (I believe it fits bets, but am looking for alternatives)
It requires a new drivers/base/edac.c (and .h file + plumbing). I would like to see that, as it really works nicely. The EDAC module currently in the -mm tree will depend on it, in order to implement the requested sysfs features. (Other options are possible I suppose)
The only problem I ran into was that I needed further subdirectories under 'mc/mc0', and I had to resort to using kobject_register() since the subsystem didn't have methods for such creation. Yet that code now works nicely.
I suppose I am looking for verification of the direction I am progressing in.
Thanks for any input
doug t
"If you think Education is expensive, just try Ignorance"
"Don't tell people HOW to do things, tell them WHAT you want and they will surprise you with their ingenuity."
Gen George Patton
Quoted 15 lines · Doug Thompson
EDAC is for detecting and logging memory ECC, PCI Parity and (future) other hardware errors.
In my process of getting edac ready for kernel submission:
I have been exploring mechanisms for placing EDAC controls and attribute files into sysfs for a few days now.
In the process, I "discovered" how sysfs subystems are currently used. I have determined that EDAC itself should be a subsystem, since more than one type of EDAC device can be created.
Quoted 7 lines
I made a copy of
drivers/base/sys.cto a new file,drivers/base/edac.c(and corresponding .h file) and refactored it to become a 'edacdev' subsystem.I have mounted edac to:
/sys/devices/edacbecause the/sys/device/systemsubsystem is not exported (at the moment. I suppose that could be changed).
Yes, that would be trivial to do :)
Quoted 11 lines
I now have:
/sys/devices/edac/mc/mc0/csrow0kobjects working and am adding attribute files to their respective kobjects.My RFC is:
Should I leave edac under
/sys/devicesOR refactordrivers/base/sys.cand export the "system" subsystem and mount edac under/sys/devices/system?
Please put it under /sys/devices/system.
Also, see the patch on lkml from Pat Mochel which might make your deep sysfs trees much easier to create and use.
Quoted 5 lines
Another RFC is:
Should EDAC really have a "new" subsystem in the kernel? (I believe it fits bets, but am looking for alternatives)
Why not just a system device?
Quoted 6 lines
It requires a new
drivers/base/edac.c(and .h file + plumbing). I would like to see that, as it really works nicely. The EDAC module currently in the -mm tree will depend on it, in order to implement the requested sysfs features. (Other options are possible I suppose)
No, you should not be adding stuff to drivers/base.
Quoted 5 lines
The only problem I ran into was that I needed further subdirectories under 'mc/mc0', and I had to resort to using
kobject_register()since the subsystem didn't have methods for such creation. Yet that code now works nicely.
See the patch from Pat for how to do this easier. But even without that patch, creating new kobjects is one way to do it.
thanks,
greg k-h
Quoted 8 lines · --- Greg KH
Quoted 1 line · Doug Thompson
I have mounted edac to:
/sys/devices/edacbecausethe
Quoted 1 line
/sys/device/systemsubsystem is not exportedYes, that would be trivial to do :)
I have changed to using /sys/devices/system for the edac_subsystem. I had to remove the 'static' for the declaration of system_subsystem in drivers/base/sys.c, and linked to it instead of the 'devices_subsys' variable.
Quoted 4 lines
Quoted 4 lines
I now have:
/sys/devices/system/edac/mc/mc0/csrow0kobjects
working
Quoted 2 lines
Please put it under
/sys/devices/system.
With the mod:
/sys/devices/system/edac
subsystem is working and I can add:
mc/mc0/csrows0/...
to it.
Quoted 4 lines
Also, see the patch on lkml from Pat Mochel which
might make your deep
sysfs trees much easier to create and use.
Quoted 8 lines
Quoted 4 lines
Another RFC is:
Should EDAC really have a "new" subsystem in the
kernel? (I believe it fits bets, but am lookingfor
Quoted 1 line
alternatives)
Why not just a system device?
One reason is the naming convention used:
creating an 'edac' class followed by adding a edac device creates the following name:
/sys/devices/system/edac/edac0/.../sys/devices/system/edac/edac1/...
whereas I desire:
/system/devices/system/edac/mc/mc0/...
I would just register a class 'edac' and then all else would be custom (or utilize the patch you have mentioned)
Another reason is that EDAC is not necessarily just for memory ECC checking only. The EDAC module currently has 1) ECC checking and 2) PCI parity checking (not driver notification) on bus devices. For the future, other EDAC types of devices can be added.
If I add an EDAC kobject 'subsystem' in the base, then future EDAC type devices can be added and be decoupled from the memory ECC EDAC module.
(This poses the question: if the current EDAC module's name is too generic just for the memory reporter maybe it should should be edac_mem instead or some such)
If the EDAC module "creates" the EDAC device under /sys/devices/system, then that module will "own" that name and future devices will need to be added to that module. There currently is no "core" edac sysfs subystem to provide the abstraction.
I guess the issue becomes: Is the EDAC module JUST for memory ECC and PCI parity reporting, in utilizing sysfs?
Should other future EDAC devices then also be added to the EDAC module in order to utilize the sysfs /sys/devices/system/edac device?
Or should they be modular in their own right and plug into a sysfs EDAC subsystem?
I might be wrong, or need redirection, but I see this as a clean model by using an abstraction of EDAC type devices in sysfs.
Quoted 6 lines
Quoted 2 lines
It requires a new
drivers/base/edac.c(and .h file+
Quoted 1 line
plumbing).
No, you should not be adding stuff to
drivers/base.
I understand. I did look at other options, but I saw more coupling in the others. If there were a de-coupled EDAC sysfs 'core or base class' (ie EDAC subsystem), then the memory ECC reporter could plug in cleanly. The PCI Parity report could plug in. Other future, EDAC reporters could plug in as well.
Quoted 5 lines
Quoted 2 lines
The only problem I ran into was that I needed
further
Quoted 1 line
subdirectories under 'mc/mc0', and I had to resort
to
Quoted 8 lines
Quoted 1 line
using
kobject_register()since the subsystemdidn't
Quoted 2 lines
have methods for such creation. Yet that code now
works nicely.See the patch from Pat for how to do this easier.
But even without that
patch, creating new kobjects is one way to do it.
Ok, I will take a look at the patch to see its options are. I look forward to feedback.
Quoted 5 lines
thanks,
greg k-h
my thanks
doug t
"If you think Education is expensive, just try Ignorance"
"Don't tell people HOW to do things, tell them WHAT you want and they will surprise you with their ingenuity."
Gen George Patton