> -----Original Message----- > From: Dan Carpenter [mailto:dan.carpen...@oracle.com] > Sent: Thursday, July 20, 2017 3:49 AM > To: step...@networkplumber.org > Cc: de...@linuxdriverproject.org; KY Srinivasan <k...@microsoft.com>; > Haiyang Zhang <haiya...@microsoft.com> > Subject: [bug report] vmbus: add prefetch to ring buffer iterator > > Hello Stephen Hemminger, > > The patch 15e1674de747: "vmbus: add prefetch to ring buffer iterator" > from Jun 25, 2017, leads to the following static checker warning: > > drivers/hv/ring_buffer.c:366 hv_pkt_iter_first() > warn: 'desc' can't be NULL. > > drivers/hv/ring_buffer.c > 357 struct vmpacket_descriptor *hv_pkt_iter_first(struct vmbus_channel > *channel) > 358 { > 359 struct hv_ring_buffer_info *rbi = &channel->inbound; > 360 struct vmpacket_descriptor *desc; > 361 > 362 if (hv_pkt_iter_avail(rbi) < sizeof(struct > vmpacket_descriptor)) > 363 return NULL; > 364 > 365 desc = hv_get_ring_buffer(rbi) + rbi->priv_read_index; > 366 if (desc) > 367 prefetch((char *)desc + (desc->len8 << 3)); > 368 > 369 return desc; > 370 } > 371 EXPORT_SYMBOL_GPL(hv_pkt_iter_first); > > This static checker warning is pretty bogus, but it's actually correct > this time (like a stopped clock is right twice a day). The first test > ensures that there is enough space in the ring bugger so it can't return > NULL. > > How does that ring buffer work without any sort of locking?? It feels > like you should have to say how much data you want to get in > hv_get_ring_buffer() so that someone else doesn't come and claim it > between the check for if it's avaible and the get.
On the receive side (read side of the ring buffer) all readers for a given ring buffer are serialized as they are bound to the same CPU - channels are affinitized to specific CPUs to spread the interrupt load in the guest. K. Y > > regards, > dan carpenter _______________________________________________ devel mailing list de...@linuxdriverproject.org http://driverdev.linuxdriverproject.org/mailman/listinfo/driverdev-devel
