On 2/6/21 9:08 AM, Pintu Agarwal wrote:
...
# cat meminfo | grep -i cma
CmaTotal: 1048576 kB
CmaFree: 1046872 kB
This CMA info was added by me way back in 2014.
At that time I even thought about adding this cma alloc/fail counter in vmstat.
That time I also had an internal patch about it but later dropped
(never released to mainline).
If required I can re-work again on this part.
And I have few more points to add here.
1) In the past I have faced this cma allocation failure (which could
be common in small embedded devices).
Earlier it was even difficult to figure if cma failure actually happened.
Thus I added this simple patch:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.11-rc6&id=5984af1082f3b115082178ed88c47033d43b924d
2) IMO just reporting CMA alloc/fail may not be enough (at times). The
main point is for which client/dev is this failure happening ?
Sometimes just tuning the size or fixing the alignment can resolve the
failure if we know the client. For global CMA it will be just NULL
(dev).
3) IMO having 2 options SYSFS and DEBUGFS may confuse the
developer/user (personal experience). So are we going to remove the
DEBUGFS or merge it ?
It is confusing to have a whole bunch of different places to find data
about a system. Here, I think the key is to add to the Documentation/
directory. So far, the documentation I see is:
admin-guide/mm/cma_debugfs.rst: only covers debugfs
admin-guide/kernel-parameters.txt:602: for CMA kernel parameters
If we add sysfs items then we will want a new .rst document that covers
that, and lists all the places to look.
So anyway, the underlying guidelines for which fs to user are approximately:
* sysfs: used for monitoring. One value per item, stable ABI, production.
* debugfs: *theoretically* not a stable ABI (we hope no one locks us in
by doing obnoxious things that break userspace if the debugfs APIs change).
The intention is that developers can put *anything* in there.
debugfs has a firm place in debugging, and is probably a keeper here.
I originally thought we should combine CMA's sysfs and debugfs items,
but Minchan listed an example that seems to show a pretty clear need
for monitoring of CMA areas, in production systems, and that's what
sysfs is for.
So it looks like we want both debugfs and sysfs for CMA, plus a new
overall CMA documentation that points to everything.
4) Sometimes CMA (or DMA) allocation failure can happen even very
early in boot time itself. At that time these are anyways not useful.
Some system may not proceed if CMA/DMA allocation is failing (again
personal experience).
==> Anyways this is altogether is different case...
5) The default max CMA areas are defined to be 7 but user can
configure it to any number, may be 20 or 50 (???)
6) Thus I would like to propose another thing here.
Just like we have /proc/vmallocinfo, /proc/slabinfo, etc., we can also
have: /proc/cmainfo
Here in /proc/cmaminfo we can capute more detailed information:
Global CMA Data: Alloc/Free
Client specific data: name, size, success, fail, flags, align, etc
(just a random example).
==> This can dynamically grow as large as possible
==> It can also be enabled/disabled based on CMA config itself (no
additional config required)
Any feedback on point (6) specifically ?
Well, /proc these days is for process-specific items. And CMA areas are
system-wide. So that's an argument against it. However...if there is any
process-specific CMA allocation info to report, then /proc is just the
right place for it.
thanks,
--
John Hubbard
NVIDIA
