IB/hfi1: Change default number of user contexts

Change the default number of user contexts to the number of real
(non-HT) cpu cores in order to reduce the division of hfi1 hardware
contexts in the case of high core counts with hyper-threading enabled.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

diff --git a/drivers/staging/rdma/hfi1/affinity.c b/drivers/staging/rdma/hfi1/affinity.c
index 2cb8ca77f876e7eeec1c553445f25587e080224c..6e7050ab9e16049039406b8cbb207cdfced18109 100644 (file)
--- a/drivers/staging/rdma/hfi1/affinity.c
+++ b/drivers/staging/rdma/hfi1/affinity.c
@@ -53,20 +53,6 @@
 #include "sdma.h"
 #include "trace.h"
 
-struct cpu_mask_set {
-       struct cpumask mask;
-       struct cpumask used;
-       uint gen;
-};
-
-struct hfi1_affinity {
-       struct cpu_mask_set def_intr;
-       struct cpu_mask_set rcv_intr;
-       struct cpu_mask_set proc;
-       /* spin lock to protect affinity struct */
-       spinlock_t lock;
-};
-
 /* Name of IRQ types, indexed by enum irq_type */
 static const char * const irq_type_names[] = {
        "SDMA",
@@ -82,6 +68,48 @@ static inline void init_cpu_mask_set(struct cpu_mask_set *set)
        set->gen = 0;
 }
 
+/* Initialize non-HT cpu cores mask */
+int init_real_cpu_mask(struct hfi1_devdata *dd)
+{
+       struct hfi1_affinity *info;
+       int possible, curr_cpu, i, ht;
+
+       info = kzalloc(sizeof(*info), GFP_KERNEL);
+       if (!info)
+               return -ENOMEM;
+
+       cpumask_clear(&info->real_cpu_mask);
+
+       /* Start with cpu online mask as the real cpu mask */
+       cpumask_copy(&info->real_cpu_mask, cpu_online_mask);
+
+       /*
+        * Remove HT cores from the real cpu mask.  Do this in two steps below.
+        */
+       possible = cpumask_weight(&info->real_cpu_mask);
+       ht = cpumask_weight(topology_sibling_cpumask(
+                                       cpumask_first(&info->real_cpu_mask)));
+       /*
+        * Step 1.  Skip over the first N HT siblings and use them as the
+        * "real" cores.  Assumes that HT cores are not enumerated in
+        * succession (except in the single core case).
+        */
+       curr_cpu = cpumask_first(&info->real_cpu_mask);
+       for (i = 0; i < possible / ht; i++)
+               curr_cpu = cpumask_next(curr_cpu, &info->real_cpu_mask);
+       /*
+        * Step 2.  Remove the remaining HT siblings.  Use cpumask_next() to
+        * skip any gaps.
+        */
+       for (; i < possible; i++) {
+               cpumask_clear_cpu(curr_cpu, &info->real_cpu_mask);
+               curr_cpu = cpumask_next(curr_cpu, &info->real_cpu_mask);
+       }
+
+       dd->affinity = info;
+       return 0;
+}
+
 /*
  * Interrupt affinity.
  *
@@ -93,20 +121,17 @@ static inline void init_cpu_mask_set(struct cpu_mask_set *set)
  * to the node relative 1 as necessary.
  *
  */
-int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
+void hfi1_dev_affinity_init(struct hfi1_devdata *dd)
 {
        int node = pcibus_to_node(dd->pcidev->bus);
-       struct hfi1_affinity *info;
+       struct hfi1_affinity *info = dd->affinity;
        const struct cpumask *local_mask;
-       int curr_cpu, possible, i, ht;
+       int curr_cpu, possible, i;
 
        if (node < 0)
                node = numa_node_id();
        dd->node = node;
 
-       info = kzalloc(sizeof(*info), GFP_KERNEL);
-       if (!info)
-               return -ENOMEM;
        spin_lock_init(&info->lock);
 
        init_cpu_mask_set(&info->def_intr);
@@ -116,30 +141,8 @@ int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
        local_mask = cpumask_of_node(dd->node);
        if (cpumask_first(local_mask) >= nr_cpu_ids)
                local_mask = topology_core_cpumask(0);
-       /* use local mask as default */
-       cpumask_copy(&info->def_intr.mask, local_mask);
-       /*
-        * Remove HT cores from the default mask.  Do this in two steps below.
-        */
-       possible = cpumask_weight(&info->def_intr.mask);
-       ht = cpumask_weight(topology_sibling_cpumask(
-                                       cpumask_first(&info->def_intr.mask)));
-       /*
-        * Step 1.  Skip over the first N HT siblings and use them as the
-        * "real" cores.  Assumes that HT cores are not enumerated in
-        * succession (except in the single core case).
-        */
-       curr_cpu = cpumask_first(&info->def_intr.mask);
-       for (i = 0; i < possible / ht; i++)
-               curr_cpu = cpumask_next(curr_cpu, &info->def_intr.mask);
-       /*
-        * Step 2.  Remove the remaining HT siblings.  Use cpumask_next() to
-        * skip any gaps.
-        */
-       for (; i < possible; i++) {
-               cpumask_clear_cpu(curr_cpu, &info->def_intr.mask);
-               curr_cpu = cpumask_next(curr_cpu, &info->def_intr.mask);
-       }
+       /* Use the "real" cpu mask of this node as the default */
+       cpumask_and(&info->def_intr.mask, &info->real_cpu_mask, local_mask);
 
        /*  fill in the receive list */
        possible = cpumask_weight(&info->def_intr.mask);
@@ -167,8 +170,6 @@ int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
        }
 
        cpumask_copy(&info->proc.mask, cpu_online_mask);
-       dd->affinity = info;
-       return 0;
 }
 
 void hfi1_dev_affinity_free(struct hfi1_devdata *dd)

diff --git a/drivers/staging/rdma/hfi1/affinity.h b/drivers/staging/rdma/hfi1/affinity.h
index b287e4963024d4760a3db6fac88860d54f3482de..20f52fe7409161cb772b90aae96f8e4c7a012e05 100644 (file)
--- a/drivers/staging/rdma/hfi1/affinity.h
+++ b/drivers/staging/rdma/hfi1/affinity.h
@@ -64,10 +64,27 @@ enum affinity_flags {
        AFF_IRQ_LOCAL
 };
 
+struct cpu_mask_set {
+       struct cpumask mask;
+       struct cpumask used;
+       uint gen;
+};
+
+struct hfi1_affinity {
+       struct cpu_mask_set def_intr;
+       struct cpu_mask_set rcv_intr;
+       struct cpu_mask_set proc;
+       struct cpumask real_cpu_mask;
+       /* spin lock to protect affinity struct */
+       spinlock_t lock;
+};
+
 struct hfi1_msix_entry;
 
+/* Initialize non-HT cpu cores mask */
+int init_real_cpu_mask(struct hfi1_devdata *);
 /* Initialize driver affinity data */
-int hfi1_dev_affinity_init(struct hfi1_devdata *);
+void hfi1_dev_affinity_init(struct hfi1_devdata *);
 /* Free driver affinity data */
 void hfi1_dev_affinity_free(struct hfi1_devdata *);
 /*

diff --git a/drivers/staging/rdma/hfi1/chip.c b/drivers/staging/rdma/hfi1/chip.c
index 38c0e9726cee0f5037b4e44ca82bc71134fc150b..1984454ff1166d77eab3c6ce3ba83552d2812261 100644 (file)
--- a/drivers/staging/rdma/hfi1/chip.c
+++ b/drivers/staging/rdma/hfi1/chip.c
@@ -12656,12 +12656,13 @@ static int set_up_context_variables(struct hfi1_devdata *dd)
                num_kernel_contexts = dd->chip_send_contexts - num_vls - 1;
        }
        /*
-        * User contexts: (to be fixed later)
-        *      - default to 1 user context per CPU if num_user_contexts is
-        *        negative
+        * User contexts:
+        *      - default to 1 user context per real (non-HT) CPU core if
+        *        num_user_contexts is negative
         */
        if (num_user_contexts < 0)
-               num_user_contexts = num_online_cpus();
+               num_user_contexts =
+                       cpumask_weight(&dd->affinity->real_cpu_mask);
 
        total_contexts = num_kernel_contexts + num_user_contexts;
 
@@ -14089,6 +14090,19 @@ struct hfi1_devdata *hfi1_init_dd(struct pci_dev *pdev,
                 (dd->revision >> CCE_REVISION_SW_SHIFT)
                    & CCE_REVISION_SW_MASK);
 
+       /*
+        * The real cpu mask is part of the affinity struct but has to be
+        * initialized earlier than the rest of the affinity struct because it
+        * is needed to calculate the number of user contexts in
+        * set_up_context_variables(). However, hfi1_dev_affinity_init(),
+        * which initializes the rest of the affinity struct members,
+        * depends on set_up_context_variables() for the number of kernel
+        * contexts, so it cannot be called before set_up_context_variables().
+        */
+       ret = init_real_cpu_mask(dd);
+       if (ret)
+               goto bail_cleanup;
+
        ret = set_up_context_variables(dd);
        if (ret)
                goto bail_cleanup;
@@ -14102,9 +14116,7 @@ struct hfi1_devdata *hfi1_init_dd(struct pci_dev *pdev,
        /* set up KDETH QP prefix in both RX and TX CSRs */
        init_kdeth_qp(dd);
 
-       ret = hfi1_dev_affinity_init(dd);
-       if (ret)
-               goto bail_cleanup;
+       hfi1_dev_affinity_init(dd);
 
        /* send contexts must be set up before receive contexts */
        ret = init_send_contexts(dd);