> > @@ -1721,6 +1721,10 @@ static ssize_t target_cpu_store(struct vmbus_channel *channel,
> > * in on a CPU that is different from the channel target_cpu value.
> > */
> >
> > + if (channel->change_target_cpu_callback)
> > + (*channel->change_target_cpu_callback)(channel,
> > + channel->target_cpu, target_cpu);
> > +
> > channel->target_cpu = target_cpu;
> > channel->target_vp = hv_cpu_number_to_vp_number(target_cpu);
> > channel->numa_node = cpu_to_node(target_cpu);
>
> I think there's an ordering problem here. The change_target_cpu_callback
> will allow storvsc to flush the cache that it is keeping, but there's a window
> after the storvsc callback releases the spin lock and before this function
> changes channel->target_cpu to the new value. In that window, the cache
> could get refilled based on the old value of channel->target_cpu, which is
> exactly what we don't want. Generally with caches, you have to set the new
> value first, then flush the cache, and I think that works in this case. The
> callback function doesn't depend on the value of channel->target_cpu,
> and any cache filling that might happen after channel->target_cpu is set
> to the new value but before the callback function runs is OK. But please
> double-check my thinking. :-)
Sorry, I don't see the problem. AFAICT, the "cache" gets refilled based
on the values of alloced_cpus and on the current state of the cache but
not based on the value of channel->target_cpu. The callback invocation
uses the value of the "old" target_cpu; I think I ended up placing the
callback call where it is for not having to introduce a local variable
"old_cpu". ;-)
> > @@ -621,6 +621,63 @@ static inline struct storvsc_device *get_in_stor_device(
> >
> > }
> >
> > +void storvsc_change_target_cpu(struct vmbus_channel *channel, u32 old, u32 new)
> > +{
> > + struct storvsc_device *stor_device;
> > + struct vmbus_channel *cur_chn;
> > + bool old_is_alloced = false;
> > + struct hv_device *device;
> > + unsigned long flags;
> > + int cpu;
> > +
> > + device = channel->primary_channel ?
> > + channel->primary_channel->device_obj
> > + : channel->device_obj;
> > + stor_device = get_out_stor_device(device);
> > + if (!stor_device)
> > + return;
> > +
> > + /* See storvsc_do_io() -> get_og_chn(). */
> > + spin_lock_irqsave(&device->channel->lock, flags);
> > +
> > + /*
> > + * Determines if the storvsc device has other channels assigned to
> > + * the "old" CPU to update the alloced_cpus mask and the stor_chns
> > + * array.
> > + */
> > + if (device->channel != channel && device->channel->target_cpu == old) {
> > + cur_chn = device->channel;
> > + old_is_alloced = true;
> > + goto old_is_alloced;
> > + }
> > + list_for_each_entry(cur_chn, &device->channel->sc_list, sc_list) {
> > + if (cur_chn == channel)
> > + continue;
> > + if (cur_chn->target_cpu == old) {
> > + old_is_alloced = true;
> > + goto old_is_alloced;
> > + }
> > + }
> > +
> > +old_is_alloced:
> > + if (old_is_alloced)
> > + WRITE_ONCE(stor_device->stor_chns[old], cur_chn);
> > + else
> > + cpumask_clear_cpu(old, &stor_device->alloced_cpus);
>
> I think target_cpu_store() can get called in parallel on multiple CPUs for different
> channels on the same storvsc device, but multiple changes to a single channel are
> serialized by higher levels of sysfs. So this function could run after multiple
> channels have been changed, in which case there's not just a single "old" value,
> and the above algorithm might not work, especially if channel->target_cpu is
> updated before calling this function per my earlier comment. I can see a
> couple of possible ways to deal with this. One is to put the update of
> channel->target_cpu in this function, within the spin lock boundaries so
> that the cache flush and target_cpu update are atomic. Another idea is to
> process multiple changes in this function, by building a temp copy of
> alloced_cpus by walking the channel list, use XOR to create a cpumask
> with changes, and then process all the changes in a loop instead of
> just handling a single change as with the current code at the old_is_alloced
> label. But I haven't completely thought through this idea.
Same here: the invocations of target_cpu_store() are serialized on the
per-connection channel_mutex...
> > @@ -1268,8 +1330,10 @@ static struct vmbus_channel *get_og_chn(struct storvsc_device
> > *stor_device,
> > if (cpumask_test_cpu(tgt_cpu, node_mask))
> > num_channels++;
> > }
> > - if (num_channels == 0)
> > + if (num_channels == 0) {
> > + stor_device->stor_chns[q_num] = stor_device->device->channel;
>
> Is the above added line just fixing a bug in the existing code? I'm not seeing how
> it would derive from the other changes in this patch.
It was rather intended as an optimization: Each time I/O for a device
is initiated on a CPU that have "num_channels == 0" channel, the current
code ends up calling get_og_chn() (in the attempt to fill the cache) and
returns the device's primary channel. In the current code, the cost of
this operations is basically the cost of parsing alloced_cpus, but with
the changes introduced here this also involves acquiring (and releasing)
the primary channel's lock. I should probably put my hands forward and
say that I haven't observed any measurable effects due this addition in
my experiments; OTOH, caching the returned/"found" value made sense...
> > @@ -1324,7 +1390,10 @@ static int storvsc_do_io(struct hv_device *device,
> > continue;
> > if (tgt_cpu == q_num)
> > continue;
> > - channel = stor_device->stor_chns[tgt_cpu];
> > + channel = READ_ONCE(
> > + stor_device->stor_chns[tgt_cpu]);
> > + if (channel == NULL)
> > + continue;
>
> The channel == NULL case is new because a cache flush could be happening
> in parallel on another CPU. I'm wondering about the tradeoffs of
> continuing in the loop (as you have coded in this patch) vs. a "goto" back to
> the top level "if" statement. With the "continue" you might finish the
> loop without finding any matches, and fall through to the next approach.
> But it's only a single I/O operation, and if it comes up with a less than
> optimal channel choice, it's no big deal. So I guess it's really a wash.
Yes, I considered both approaches; they both "worked" here. I was a
bit concerned about the number of "possible" gotos (again, mainly a
theoretical issue, since I can imagine that the cash flushes will be
relatively "rare" events in most cases and, in any case, they happen
to be serialized); the "continue" looked like a suitable and simpler
approach/compromise, at least for the time being.
>
> > if (hv_get_avail_to_write_percent(
> > &channel->outbound)
> > > ring_avail_percent_lowater) {
> > @@ -1350,7 +1419,10 @@ static int storvsc_do_io(struct hv_device *device,
> > for_each_cpu(tgt_cpu, &stor_device->alloced_cpus) {
> > if (cpumask_test_cpu(tgt_cpu, node_mask))
> > continue;
> > - channel = stor_device->stor_chns[tgt_cpu];
> > + channel = READ_ONCE(
> > + stor_device->stor_chns[tgt_cpu]);
> > + if (channel == NULL)
> > + continue;
>
> Same comment here.
Similarly here.
Thoughts?
Thanks,
Andrea
