On Wed, Jul 11, 2018 at 02:40:38PM +0200, Dominique Martinet wrote: > Matthew Wilcox wrote on Thu, Jun 28, 2018: > > diff --git a/net/9p/client.c b/net/9p/client.c > > index f8d58b0852fe..bbab82f22c20 100644 > > --- a/net/9p/client.c > > +++ b/net/9p/client.c > > @@ -908,30 +908,27 @@ static struct p9_fid *p9_fid_create(struct p9_client *clnt) > > - ret = p9_idpool_get(clnt->fidpool); > > - if (ret < 0) > > - goto error; > > - fid->fid = ret; > > - ... > > + idr_preload(GFP_KERNEL); > > + spin_lock_irq(&clnt->lock); > > + ret = idr_alloc_u32(&clnt->fids, fid, &fid->fid, UINT_MAX, GFP_NOWAIT); > > There's also a P9_NOFID that we shouldn't use, so the max here should be > P9_NOFID - 1 Happy to fix that. It shouldn't actually happen, of course. I can't imagine us having 4 billion FIDs in use at once. > That aside this introduces a change of behaviour that fid used to be > alloc'd linearily from 0 which no longer holds true, that breaks one > serveur (nfs-ganesha just returns ERANGE) but others seem to handle this > just fine so they'll just need to fix that server. > max aside this looks good. I don't understand your assertion that this is a change of behaviour. The implementation of p9_idpool_get() uses idr_alloc(), not idr_alloc_cyclic(), so I don't believe I've changed which FID would be allocated.
