I think this behavior is not part of the API, it is a bug. I agree that detach() frees the direct mbuf when refcnt becomes 0, Konstantin suggests. It is a right behavior of reference counting.
Regards, Hiroyuki 2016-05-18 0:45 GMT+09:00 Ananyev, Konstantin <konstantin.ananyev at intel.com>: > > >> -----Original Message----- >> From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com] >> Sent: Tuesday, May 17, 2016 3:19 PM >> To: Ananyev, Konstantin >> Cc: dev at dpdk.org; Hiroyuki Mikita; olivier.matz at 6wind.com >> Subject: Re: [dpdk-dev] [PATCH v2] mbuf: decrease refcnt when detaching >> >> 2016-05-17 13:44, Ananyev, Konstantin: >> > From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com] >> > > 2016-05-17 12:59, Ananyev, Konstantin: >> > > > > > The rte_pktmbuf_detach() function should decrease refcnt on a >> > > > > > direct >> > > > > > buffer. >> > > > > >> > > > > As you have noticed, "whenever the indirect buffer is detached, >> > > > > the reference counter on the direct buffer is decremented." >> > > > > So the current behaviour of rte_pktmbuf_detach() is buggy. >> > > > > Why not fix it without renaming? >> > > > > If you consider this behavioral bug is part of the API, we >> > > > > can fix it in a new function unattach and deprecate detach. >> > > > > But Konstantin, why do you want to keep a restore function? >> > > > > What is the need? >> > > > >> > > > I think it might be a useful functionality in some situations: >> > > > some users can attach/detach to external memory buffers (no mbufs) >> > > > and similar functionality is required. >> > > >> > > Attach to external memory buffer (raw buffer) is not currently supported. >> > > >> > > > Let say right now examples/vhost/main.c has its own >> > > > pktmbuf_detach_zcp() >> > > >> > > You should look at the commit http://dpdk.org/commit/68363d85 >> > > "examples/vhost: remove the non-working zero copy code" >> > > >> > > > which is doing pretty much the same - restore original values, after >> > > > detaching >> > > > mbuf from external (virtio) memory buffer. >> > > > Would be good if we'll use a standard API function here. >> > > >> > > You are welcome to implement mbuf attach to raw buffer. >> > > But it is not a requirement for this fix. >> > >> > Hmm, still not sure why we can't keep an existing function? >> >> Because it does not do what its name (and doc) suggest. >> >> > Obviously it wouldn't cost anything and I still think might be useful. >> >> It costs to overcomplicate API for only a half support. > > I still think it is better to have it then not, but wouldn't insist here. > Konstantin > >> If you need the feature "attach to raw", please implement it completely. >
