Re: libgfapi changes to add lk_owner and lease ID

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Dec 03, 2015 at 06:34:56PM -0500, Ira Cooper wrote:
> Kaleb KEITHLEY <kkeithle@xxxxxxxxxx> writes:
> 
> > On 12/03/2015 03:10 PM, Ira Cooper wrote:
> >> Poornima Gurusiddaiah <pgurusid@xxxxxxxxxx> writes:
> >> 
> >>> Hi, 
> >>>
> >>> Brief Background: 
> >>> ============ 
> >>> For the below two features, we need ligfapi to take 2 other parameters from the applications for most number of fops. 
> >>> http://review.gluster.org/#/c/12014/ 
> >>> http://review.gluster.org/#/c/11980/ 
> >>>
> >>> For leases to work as explained in the above doc, every file data read/write fop needs to be associated with a lease ID. This is specially required for Samba and NFS-Ganesha as they inturn serve other clients who request for leases. For Gluster to identify the end client (which is not Samba/NFS Ganesha) we need lease ID to be filled by Samba/NFS Ganesha. 
> >>>
> >>> For mandatory locks feature to work as explained, every file data read/write fop needs to be associated with a lk_owner. In linux Kernel VFS takes care of filling the lk_ownere for the file system. In libgfapi case, the applications calling into libgfapi should be providing lk_owner with every fop. This is again required mainly for Samba and NFS Ganesha, as they serve multiple clients. 
> >>>
> >>> Possible solutions: 
> >>> ============= 
> >>> 1. Modify all the required APIs to take 2 other parameter, lease ID and lk_owner. But that would mean backward compatibility issues and is a programming overhead for applications not interested in Leases and mandatory lock feature. 
> >>> 2. Add an API called glfs_set_fop_attrs (lease ID, lk_owner) which works similar to glfs_set_uid(uid). The API sets a thread local storage (pthread_key) with the values provided, the further fops on that thread will pick the lease ID and lk_owner from the thread local storage (pthread_key). There are few minor details that needs to be worked out: 
> >>> - In case of async API will end up using lease ID and lk_owner from wrong thread. 
> >>> - unset lease ID and lk_owner after every fop to ensure there is no stale lease ID or lk_owner set? 
> >>> - For fd based fops we can store the lease ID and lk_owner in glfd, so that the application writed need not set it for every fop. But for handle based fops lease ID and lk_owner needs to be set explicitly every-time. 
> >>>
> >>> Solution 2 is more preferable except for that it adds overhead of calling another API, for the libgfapi users who intends to use these features. 
> >>> A prototype of solution 2 can be found at http://review.gluster.org/#/c/12876/ 
> >>>
> >
> > why not use storage in the client_t instead of thread local?
> 
> It comes down to the use case.  For Samba, the right spot is almost
> certainly the fd, because lease keys are a per-handle (which we map to
> fd) property.
> 
> client_t is a horror show, due to race conditions between threads, IMHO.

If there are known races, should we not address that? Got a bug that
explains it in more detail?

Thanks,
Niels

Attachment: signature.asc
Description: PGP signature

_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel

[Index of Archives]     [Gluster Users]     [Ceph Users]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux