Re: [PATCH 2/9] perf/core: Add PERF_SAMPLE_CGROUP feature

[Namhyung Kim]

Hi Tejun,

Sorry for the late reply.


On Fri, Aug 30, 2019 at 09:58:15PM -0700, Tejun Heo wrote:
> Hello,
> 
> On Sat, Aug 31, 2019 at 12:03:26PM +0900, Namhyung Kim wrote:
> > Hmm.. it looks hard to use fhandle as the identifier since perf
> > sampling is done in NMI context.  AFAICS the encode_fh part seems ok
> > but getting dentry/inode from a kernfs_node seems not.
> > 
> > I assume kernfs_node_id's ino and gen are same to its inode's.  Then
> > we might use kernfs_node for encoding but not sure you like it ;-)
> 
> Oh yeah, the whole cgroup id situation is kinda shitty and it's likely
> that it needs to be cleaned up a bit for this to be used widely.  The
> issues are...
> 
> * As identifiers, paths sucks.  It's too big and unwieldy and can be
>   rapidly reused for different instances.
> 
> * ino is compact but can't be easily mapped to path from userland and
>   also not unique.
> 
> * The fhandle identifier - currently ino+gen - is better in that it's
>   finite sized and compact and can be efficiently mapped to path from
>   userspace.  It's also mostly unique.  However, the way gen is
>   currently generated still has some chance of the same ID getting
>   reused and it isn't easily accessible from inside the kernel right
>   now.
> 
> Eventually, where we wanna be at is having a single 64bit identifier
> which can be easily used everywhere.  It should be pretty straight
> forward on 64bit machines - we can just use monotonically increasing
> id and use it for everything - ino, fhandle and internal cgroup id.
> On 32bit, it gets a bit complicated because ino is 32bit, so it'll
> need a custom allocator which bumps gen when the lower 32bit wraps and
> skips in-use inos.  Once we have that, we can use that for cgrp->id
> and fhandle and derive ino from it.
> 
> This is on the to-do list but obviously hasn't happened yet.  If you
> wanna take on it, great, but, otherwise, what can be done now is
> either moving gen+ino generation into cgroup and tell kernfs to use it
> or copy gen+ino into cgroup for easier access.  The former likely is
> the better approach given that it brings us closer to where we wanna
> be eventually.

So is my understanding below correct?

 * currently kernfs ino+gen is different than inode's ino+gen
 * but it'd be better to make them same
 * so move (generic?) inode's ino+gen logic to cgroup
 * and kernfs node use the same logic (and number)
 * so perf sampling code (NMI) just access kernfs node
 * and userspace can use file handle for comparison

Thanks,
Namhyung