> On 3. Oct 2018, at 11:17, Stanislaw Gruszka <sgruszka@xxxxxxxxxx> wrote: > > Make mt76 frag cache similar to netdev frag cache. This should > make frag allocation safe regarding concurrent access and also > be more efficient since we will use pages that most likely are > hot on particular cpu. > > And we don't need to clean up the cache up during device removal. The problem with feeding multiple queues with buffers from the same fragment cache is the fact that this re-introduces the problem of large compound pages staying pinned in memory (often with only one or two fragments actually being used) far too long, leading to excessive RAM usage. That is much more critical than the spinlock on alloc issue. Please keep allocation the way it is now. - Felix
