Re: [PATCH 2/2] perf/tools: put new buildid locks to use

[Milos Vyletel <milos@redhat.com>]

On Wed, Jun 03, 2015 at 01:21:41PM +0200, Ingo Molnar wrote:

> 
> * Milos Vyletel <milos@redhat.com> wrote:
> 
> > On Thu, May 14, 2015 at 07:38:08PM +0200, Ingo Molnar wrote:
> > 
> > > 
> > > * Milos Vyletel <milos@redhat.com> wrote:
> > > 
> > > > On Thu, May 14, 2015 at 01:38:21PM +0200, Jiri Olsa wrote:
> > > > 
> > > > > On Thu, May 14, 2015 at 12:40:59PM +0200, Ingo Molnar wrote:
> > > > > > 
> > > > > > * Milos Vyletel <milos@redhat.com> wrote:
> > > > > > 
> > > > > > > Use new read/write locks when accesing buildid directory on places where
> > > > > > > we may race if multiple instances are run simultaneously.
> > > > > > 
> > > > > > Dunno, this will create locking interaction between multiple instances 
> > > > > > of perf - hanging each other, etc.
> > > > > > 
> > > > > > And it seems unnecessary: the buildid hierarchy is already spread out. 
> > > > > > What kind of races might there be?
> > > > > 
> > > > > there was just recently one fixed by commit:
> > > > >   0635b0f71424 perf tools: Fix race in build_id_cache__add_s()
> > > > > 
> > > > > havent checked the final patch yet, but the idea is to
> > > > > protect us from similar bugs
> > > > 
> > > > right. on top of race with EEXIST couple more are possible (EMLINK, 
> > > > ENOSPC, EDQUOT, ENOMEM... the only way to prevent them all is to 
> > > > lock this kind of operations and make sure we run one at a time.
> > > 
> > > Yeah, so the race pointed out in 0635b0f71424 can be (and should be) 
> > > fixed without locking:
> > > 
> > >  - first create the file under a process-private name under 
> > >    ~/.debug/tmp/ if the target does not exist yet
> > > 
> > >  - then fully fill it in with content
> > > 
> > >  - then link(2) it to the public target name, which VFS operation is
> > >    atomic and may fail safely: at which point it got already created
> > >    by someone else.
> > > 
> > >  - finally unlink() the private instance name and the target will now
> > >    be the only instance left: either created by us, or by some other 
> > >    perf instance in the rare racy case.
> > > 
> > > Since all of ~/.debug is on the same filesystem this should work fine.
> > > 
> > > Beyond avoiding locking this approach has another advantage: it's 
> > > transaction safe, so a crashed/interrupted perf instance won't corrupt 
> > > the debug database, it will only put fully constructed files into the 
> > > public build-id namespace. It at most leaves a stale private file 
> > > around in ~/.debug/tmp/.
> > > 
> > 
> > Ingo,
> > 
> > I finally found some time to make this change. While going over the code I've 
> > noticed one thing that would make concurrent creation even easier to solve. 
> > Instead of copying the file to temp file what about simply opening file with 
> > O_CREAT|O_EXCL? creat itself
> > 
> > "creat() is equivalent to open() with flags equal to O_CREAT|O_WRONLY|O_TRUNC."
> > 
> > addition of O_EXCL would
> > 
> > "Ensure that this call creates the file: if this flag is specified in 
> > conjunction with O_CREAT, and pathname already exists, then open() will fail."
> > 
> > This we would prevent truncation of already linked file in case link() races as 
> > in 0635b0f71424. What do you think?
> 
> But it would not prevent the problem of creating a not yet fully constructed file 
> - which some other tool invocation could attempt to parse in an incomplete 
> fashion.
> 
> Using create+link+unlink avoids that race, the files in the publicly visible 
> namespace will always be fully constructed by the time they are made visible 
> (atomically).
> 

Got it. Will use the approach proposed by you.

Milos