Re: [libvirt] hash: failed on concurrent iterating.

On Wed, Apr 04, 2018 at 09:50:29AM +0000, Zhangzijian wrote: > > -----Original Message----- > > From: libvir-list-bounces(a)redhat.com [mailto:libvir-list-bounces@redhat.com] > > On Behalf Of Michal Privoznik > > Sent: Wednesday, April 04, 2018 5:21 PM > > To: zhangzijian (Cloud); libvir-list(a)redhat.com > > Cc: huangyong (Cloud) > > Subject: Re: [libvirt] hash: failed on concurrent iterating. > > > > On 04/04/2018 11:10 AM, Zhangzijian wrote: > > > In function 'virHashForEach/virHashSearch', 'table->iterating' prevent > > concurrent iterating. But most caller evoke it, after hold a read lock. > > > This will lead the second caller failed to iterate the table. So, the caller should > > hold a write lock, then iterate the table. > > > > Looks like there's another thread discussing this issue: > > > > Yes, the issue is still exist in v4.2.0. > Is it applicable to using write lock instead? I hit this a month ago when working on mediated devices and creating a ton of them. You can't replace it with write lock because it would kill the sole purpose of having RW locks in our lists using hash tables in the first place. IMHO the correct way to fix this is to get rid of the table->iterating flag completely. In order to do that, you need to convert hash tables to lockable objects so that even modules that don't define any lockable abstractions on top of hash tables could benefit from this. There are several caveats like dealing with NWFilter and for example the virHashForEach method which allows calling the 'Remove' callback which of course requires a write lock, however, if you issue a standard query, this could have been (and also should) achieved with a read lock to allow concurrent reading of elements in the table. This would also require basically having a *Locked version of each virHash API because the existing APIs are being called both outside and inside of the module, so you need to take into account that you might already be holding a lock, etc.