Jeremy Allison
2021-Dec-03 17:47 UTC
[Samba] Why samba only use single thread to process request？ Why vfs_cephfs
On Fri, Dec 03, 2021 at 04:50:50PM +0800, lushasha08--- via samba wrote:>Hi, Lists > I recently begin to study samba source code. I am focus on samba FileServer code. I have some confusion. Can you give me some help? > > 1. Why samba only use single thread to process request from one connection? Why not use threadpools? I think using threadpool can improve performance extremly.Samba is 30 years old. Portable, working threads were merely a fantasy at that time :-). Samba is moving towards moving threads in places where it makes sense - check out the use of lib/pthreadpool/pthreadpool.c in Samba, but in many cases (not all of course) the same benefits can be gained by using an async event model - look at lib/tevent in the Samba source code.> 2. Why async ceph write faked up by calling the sync API? Can async ceph write just use conn's threadpool like glusterfs?Ceph needs async primitives, as gluster already has. I'm not a ceph expert so can't say if it has them but they aren't suiitable yet.
John Mulligan
2021-Dec-04 15:42 UTC
[Samba] Why samba only use single thread to process request？ Why vfs_cephfs
On Friday, December 3, 2021 12:47:08 PM EST Jeremy Allison via samba wrote:> On Fri, Dec 03, 2021 at 04:50:50PM +0800, lushasha08--- via samba wrote: > >Hi, Lists > > > > I recently begin to study samba source code. I am focus on samba > > FileServer code. I have some confusion. Can you give me some > > help?> > > 1. Why samba only use single thread to process request from one > > connection? Why not use threadpools? I think using threadpool can > > improve performance extremly. > Samba is 30 years old. Portable, working threads were > merely a fantasy at that time :-). > > Samba is moving towards moving threads in places > where it makes sense - check out the use of > lib/pthreadpool/pthreadpool.c in Samba, but > in many cases (not all of course) the same > benefits can be gained by using an async > event model - look at lib/tevent in the > Samba source code. > > > 2. Why async ceph write faked up by calling the sync API? Can > > async ceph write just use conn's threadpool like glusterfs? > Ceph needs async primitives, as gluster > already has. I'm not a ceph expert so > can't say if it has them but they aren't > suiitable yet.This piqued my curiosity as I have a passing familiarity with the cephfs api and have done some work with both gluster and ceph. I skimmed the vfs module sources and it appears to me that the vfs module for gluster does not directly use async calls from libfgapi. Rather, it uses a threadpool to mimic async behavior, which is something lushasha08 implied [1]. So, please correct me if I'm wrong, but I think the vfs ceph module could use a similar technique if someone was to implement the changes [2]. That said, Jeremy's suggestion that the api for cephfs could updated to include async functions would probably be even better. The cephfs API has two sets of function call, a "high level" api that mimics many of the traditional unix/posix calls and a "low level" api that operates on specialized structs rather than virtual file handles. However, I don't see anything that looks like an async call even in the low level set of calls today. I am aware that the api is continuing to evolve as ceph "pacific" now makes an "openat" style call available [3]. So perhaps the cephfs maintainers would be open to having async apis too. [1] - https://git.samba.org/?p=samba.git;a=blob;f=source3/modules/ vfs_glusterfs.c;h=25b82bdd4719c7b02a4513d94ca0230dec353540;hb=HEAD#l909 [2] - I'm not volunteering anyone. :-) [2] - https://github.com/ceph/ceph/blob/pacific/src/include/cephfs/ libcephfs.h#L1180