On Mon, Dec 19, 2011 at 02:56:35PM -0800, Tejun Heo wrote: > Hello, Vivek. > > On Mon, Dec 19, 2011 at 01:27:17PM -0500, Vivek Goyal wrote: > > Ok, that's good to know. If per cpu allocator can support this use case, > > it will be good for 3.3 onwards. This seems to be right way to go to fix > > the problem. > > Ummm... if we're gonna make percpu usable w/ GFP_NOIO, the right > interim solution would be making a simplistic mempool so that later > when percpu can do it, it can be swapped easily. I really can't see > much benefit of adding refcnting on top of everything just for this. Ok. So are you suggesting that I should write a simple mempool kind of implementation of my own for group and per cpu data allocation. Keep certain number of elements in the cache and trigger a worker thread to allocate more elements once minimum number of elements go below threshold. If pool has run out of pre-allocated elements then allocation will fail and IO will be accounted to root group? I am looking at the mempool implementation (mempool_create()) and looks like that is not suitable for my use case. mempool_alloc() will call into alloc function provided by me and pass the flags. I can't implement an alloc function and honor that flag as per cpu alloc does not take any flags. So IIUC, existing mempool implementation is not directly usable for my requirement and I need to write some code of my own for the caching layer which always allocates objects from reserve and fills in the pool asynchronously with the help of a worker thread. Thanks Vivek -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
