dmaengine: make dma_channel_rebalance() NUMA aware

dma_channel_rebalance() currently distributes channels by processor ID.
These IDs often change with the BIOS, and the order isn't related to
the DMA channel list (related to PCI bus ids).
* On my SuperMicro dual E5 machine, first socket has processor IDs [0-7]
  (and [16-23] for hyperthreads), second socket has [8-15]+[24-31]
  => channels are properly allocated to local CPUs.
* On Dells R720 with same processors, first socket has even processor IDs,
  second socket has odd numbers
  => half the processors get channels on the remote socket, causing
     cross-NUMA traffic and lower DMA performance.

Change nth_chan() to return the channel with min table_count and in the
NUMA node of the given CPU, if any. If none, the (non-local) channel with
min table_count is returned. nth_chan() is therefore renamed into min_chan()
since we don't iterate until the nth channel anymore. In practice, the
behavior is the same because first channels are taken first and are then
ignored because they got an additional reference.

The new code has a slightly higher complexity since we always scan the
entire list of channels for finding the minimal table_count (instead
of stopping after N chans), and because we check whether the CPU is in the
DMA device locality mask. Overall we still have time complexity =
number of chans x number of processors. This rebalance is rarely used,
so this won't hurt.

On the above SuperMicro machine, channels are still allocated the same.
On the Dells, there are no locality issue anymore (MEMCPY channel X goes
to processor X and to its hyperthread sibling).

Signed-off-by: Brice Goglin <Brice.Goglin@inria.fr>
Signed-off-by: Dan Williams <djbw@fb.com>

diff --git a/drivers/dma/dmaengine.c b/drivers/dma/dmaengine.c
index 9e56745f87bf164aa3b4728b52a113c9a5b68cc6..e428cf2a458bff515b68b1ebe0760f0b7ce7286d 100644 (file)
--- a/drivers/dma/dmaengine.c
+++ b/drivers/dma/dmaengine.c
@@ -376,20 +376,30 @@ void dma_issue_pending_all(void)
 EXPORT_SYMBOL(dma_issue_pending_all);
 
 /**
- * nth_chan - returns the nth channel of the given capability
+ * dma_chan_is_local - returns true if the channel is in the same numa-node as the cpu
+ */
+static bool dma_chan_is_local(struct dma_chan *chan, int cpu)
+{
+       int node = dev_to_node(chan->device->dev);
+       return node == -1 || cpumask_test_cpu(cpu, cpumask_of_node(node));
+}
+
+/**
+ * min_chan - returns the channel with min count and in the same numa-node as the cpu
  * @cap: capability to match
- * @n: nth channel desired
+ * @cpu: cpu index which the channel should be close to
  *
- * Defaults to returning the channel with the desired capability and the
- * lowest reference count when 'n' cannot be satisfied.  Must be called
- * under dma_list_mutex.
+ * If some channels are close to the given cpu, the one with the lowest
+ * reference count is returned. Otherwise, cpu is ignored and only the
+ * reference count is taken into account.
+ * Must be called under dma_list_mutex.
  */
-static struct dma_chan *nth_chan(enum dma_transaction_type cap, int n)
+static struct dma_chan *min_chan(enum dma_transaction_type cap, int cpu)
 {
        struct dma_device *device;
        struct dma_chan *chan;
-       struct dma_chan *ret = NULL;
        struct dma_chan *min = NULL;
+       struct dma_chan *localmin = NULL;
 
        list_for_each_entry(device, &dma_device_list, global_node) {
                if (!dma_has_cap(cap, device->cap_mask) ||
@@ -398,27 +408,22 @@ static struct dma_chan *nth_chan(enum dma_transaction_type cap, int n)
                list_for_each_entry(chan, &device->channels, device_node) {
                        if (!chan->client_count)
                                continue;
-                       if (!min)
-                               min = chan;
-                       else if (chan->table_count < min->table_count)
+                       if (!min || chan->table_count < min->table_count)
                                min = chan;
 
-                       if (n-- == 0) {
-                               ret = chan;
-                               break; /* done */
-                       }
+                       if (dma_chan_is_local(chan, cpu))
+                               if (!localmin ||
+                                   chan->table_count < localmin->table_count)
+                                       localmin = chan;
                }
-               if (ret)
-                       break; /* done */
        }
 
-       if (!ret)
-               ret = min;
+       chan = localmin ? localmin : min;
 
-       if (ret)
-               ret->table_count++;
+       if (chan)
+               chan->table_count++;
 
-       return ret;
+       return chan;
 }
 
 /**
@@ -435,7 +440,6 @@ static void dma_channel_rebalance(void)
        struct dma_device *device;
        int cpu;
        int cap;
-       int n;
 
        /* undo the last distribution */
        for_each_dma_cap_mask(cap, dma_cap_mask_all)
@@ -454,14 +458,9 @@ static void dma_channel_rebalance(void)
                return;
 
        /* redistribute available channels */
-       n = 0;
        for_each_dma_cap_mask(cap, dma_cap_mask_all)
                for_each_online_cpu(cpu) {
-                       if (num_possible_cpus() > 1)
-                               chan = nth_chan(cap, n++);
-                       else
-                               chan = nth_chan(cap, -1);
-
+                       chan = min_chan(cap, cpu);
                        per_cpu_ptr(channel_table[cap], cpu)->chan = chan;
                }
 }