* [PATCH 1/5] IB/hfi1: Add sysfs interface for affinity setup
[not found] ` <20160925144248.12088.71489.stgit-9QXIwq+3FY+1XWohqUldA0EOCMrvLtNR@public.gmane.org>
@ 2016-09-25 14:44 ` Dennis Dalessandro
2016-09-25 14:44 ` [PATCH 2/5] IB/hfi1: Add a new VL sysfs attribute for sdma engines Dennis Dalessandro
` (4 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Dennis Dalessandro @ 2016-09-25 14:44 UTC (permalink / raw)
To: dledford-H+wXaHxf7aLQT0dZR+AlfA
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA, Dean Luick, Sebastian Sanchez,
Jianxin Xiong, Tadeusz Struk
From: Tadeusz Struk <tadeusz.struk-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Some users want more control over which cpu cores are being used by the
driver. For example, users might want to restrict the driver to some
specified subset of the cores so that they can appropriately partition
processes, irq handlers, and work threads.
To allow the user to fine tune system affinity settings new sysfs
attributes are introduced per sdma engine. This patch adds a new
attribute type for sdma engine and a new cpu_list attribute.
When the user writes a cpu range to the cpu_list attribute the driver
will create an internal cpu->sdma map, which will be used later as a
look-up table to choose an optimal engine for a user requests.
Reviewed-by: Dean Luick <dean.luick-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Sebastian Sanchez <sebastian.sanchez-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Jianxin Xiong <jianxin.xiong-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Tadeusz Struk <tadeusz.struk-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
drivers/infiniband/hw/hfi1/hfi.h | 2
drivers/infiniband/hw/hfi1/sdma.c | 310 ++++++++++++++++++++++++++++++++
drivers/infiniband/hw/hfi1/sdma.h | 8 +
drivers/infiniband/hw/hfi1/sysfs.c | 90 +++++++++
drivers/infiniband/hw/hfi1/user_sdma.c | 9 -
5 files changed, 412 insertions(+), 7 deletions(-)
diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h
index 5711620..59f69fa 100644
--- a/drivers/infiniband/hw/hfi1/hfi.h
+++ b/drivers/infiniband/hw/hfi1/hfi.h
@@ -65,6 +65,7 @@
#include <linux/i2c.h>
#include <linux/i2c-algo-bit.h>
#include <rdma/ib_hdrs.h>
+#include <linux/rhashtable.h>
#include <rdma/rdma_vt.h>
#include "chip_registers.h"
@@ -1174,6 +1175,7 @@ struct hfi1_devdata {
atomic_t aspm_disabled_cnt;
struct hfi1_affinity *affinity;
+ struct rhashtable sdma_rht;
struct kobject kobj;
};
diff --git a/drivers/infiniband/hw/hfi1/sdma.c b/drivers/infiniband/hw/hfi1/sdma.c
index 0990fba..8cfa960 100644
--- a/drivers/infiniband/hw/hfi1/sdma.c
+++ b/drivers/infiniband/hw/hfi1/sdma.c
@@ -726,6 +726,34 @@ u16 sdma_get_descq_cnt(void)
}
/**
+ * sdma_engine_get_vl() - return vl for a given sdma engine
+ * @sde: sdma engine
+ *
+ * This function returns the vl mapped to a given engine, or an error if
+ * the mapping can't be found. The mapping fields are protected by RCU.
+ */
+int sdma_engine_get_vl(struct sdma_engine *sde)
+{
+ struct hfi1_devdata *dd = sde->dd;
+ struct sdma_vl_map *m;
+ u8 vl;
+
+ if (sde->this_idx >= TXE_NUM_SDMA_ENGINES)
+ return -EINVAL;
+
+ rcu_read_lock();
+ m = rcu_dereference(dd->sdma_map);
+ if (unlikely(!m)) {
+ rcu_read_unlock();
+ return -EINVAL;
+ }
+ vl = m->engine_to_vl[sde->this_idx];
+ rcu_read_unlock();
+
+ return vl;
+}
+
+/**
* sdma_select_engine_vl() - select sdma engine
* @dd: devdata
* @selector: a spreading factor
@@ -788,6 +816,283 @@ struct sdma_engine *sdma_select_engine_sc(
return sdma_select_engine_vl(dd, selector, vl);
}
+struct sdma_rht_map_elem {
+ u32 mask;
+ u8 ctr;
+ struct sdma_engine *sde[0];
+};
+
+struct sdma_rht_node {
+ unsigned long cpu_id;
+ struct sdma_rht_map_elem *map[HFI1_MAX_VLS_SUPPORTED];
+ struct rhash_head node;
+};
+
+#define NR_CPUS_HINT 192
+
+static const struct rhashtable_params sdma_rht_params = {
+ .nelem_hint = NR_CPUS_HINT,
+ .head_offset = offsetof(struct sdma_rht_node, node),
+ .key_offset = offsetof(struct sdma_rht_node, cpu_id),
+ .key_len = FIELD_SIZEOF(struct sdma_rht_node, cpu_id),
+ .max_size = NR_CPUS,
+ .min_size = 8,
+ .automatic_shrinking = true,
+};
+
+/*
+ * sdma_select_user_engine() - select sdma engine based on user setup
+ * @dd: devdata
+ * @selector: a spreading factor
+ * @vl: this vl
+ *
+ * This function returns an sdma engine for a user sdma request.
+ * User defined sdma engine affinity setting is honored when applicable,
+ * otherwise system default sdma engine mapping is used. To ensure correct
+ * ordering, the mapping from <selector, vl> to sde must remain unchanged.
+ */
+struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
+ u32 selector, u8 vl)
+{
+ struct sdma_rht_node *rht_node;
+ struct sdma_engine *sde = NULL;
+ const struct cpumask *current_mask = tsk_cpus_allowed(current);
+ unsigned long cpu_id;
+
+ /*
+ * To ensure that always the same sdma engine(s) will be
+ * selected make sure the process is pinned to this CPU only.
+ */
+ if (cpumask_weight(current_mask) != 1)
+ goto out;
+
+ cpu_id = smp_processor_id();
+ rcu_read_lock();
+ rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpu_id,
+ sdma_rht_params);
+
+ if (rht_node && rht_node->map[vl]) {
+ struct sdma_rht_map_elem *map = rht_node->map[vl];
+
+ sde = map->sde[selector & map->mask];
+ }
+ rcu_read_unlock();
+
+ if (sde)
+ return sde;
+
+out:
+ return sdma_select_engine_vl(dd, selector, vl);
+}
+
+static void sdma_populate_sde_map(struct sdma_rht_map_elem *map)
+{
+ int i;
+
+ for (i = 0; i < roundup_pow_of_two(map->ctr ? : 1) - map->ctr; i++)
+ map->sde[map->ctr + i] = map->sde[i];
+}
+
+static void sdma_cleanup_sde_map(struct sdma_rht_map_elem *map,
+ struct sdma_engine *sde)
+{
+ unsigned int i, pow;
+
+ /* only need to check the first ctr entries for a match */
+ for (i = 0; i < map->ctr; i++) {
+ if (map->sde[i] == sde) {
+ memmove(&map->sde[i], &map->sde[i + 1],
+ (map->ctr - i - 1) * sizeof(map->sde[0]));
+ map->ctr--;
+ pow = roundup_pow_of_two(map->ctr ? : 1);
+ map->mask = pow - 1;
+ sdma_populate_sde_map(map);
+ break;
+ }
+ }
+}
+
+/*
+ * Prevents concurrent reads and writes of the sdma engine cpu_mask
+ */
+static DEFINE_MUTEX(process_to_sde_mutex);
+
+ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
+ size_t count)
+{
+ struct hfi1_devdata *dd = sde->dd;
+ cpumask_var_t mask, new_mask;
+ unsigned long cpu;
+ int ret, vl, sz;
+
+ vl = sdma_engine_get_vl(sde);
+ if (unlikely(vl < 0))
+ return -EINVAL;
+
+ ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
+ if (!ret)
+ return -ENOMEM;
+
+ ret = zalloc_cpumask_var(&new_mask, GFP_KERNEL);
+ if (!ret) {
+ free_cpumask_var(mask);
+ return -ENOMEM;
+ }
+ ret = cpulist_parse(buf, mask);
+ if (ret)
+ goto out_free;
+
+ if (!cpumask_subset(mask, cpu_online_mask)) {
+ dd_dev_warn(sde->dd, "Invalid CPU mask\n");
+ ret = -EINVAL;
+ goto out_free;
+ }
+
+ sz = sizeof(struct sdma_rht_map_elem) +
+ (TXE_NUM_SDMA_ENGINES * sizeof(struct sdma_engine *));
+
+ mutex_lock(&process_to_sde_mutex);
+
+ for_each_cpu(cpu, mask) {
+ struct sdma_rht_node *rht_node;
+
+ /* Check if we have this already mapped */
+ if (cpumask_test_cpu(cpu, &sde->cpu_mask)) {
+ cpumask_set_cpu(cpu, new_mask);
+ continue;
+ }
+
+ rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpu,
+ sdma_rht_params);
+ if (!rht_node) {
+ rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL);
+ if (!rht_node) {
+ ret = -ENOMEM;
+ goto out;
+ }
+
+ rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
+ if (!rht_node->map[vl]) {
+ kfree(rht_node);
+ ret = -ENOMEM;
+ goto out;
+ }
+ rht_node->cpu_id = cpu;
+ rht_node->map[vl]->mask = 0;
+ rht_node->map[vl]->ctr = 1;
+ rht_node->map[vl]->sde[0] = sde;
+
+ ret = rhashtable_insert_fast(&dd->sdma_rht,
+ &rht_node->node,
+ sdma_rht_params);
+ if (ret) {
+ kfree(rht_node->map[vl]);
+ kfree(rht_node);
+ dd_dev_err(sde->dd, "Failed to set process to sde affinity for cpu %lu\n",
+ cpu);
+ goto out;
+ }
+
+ } else {
+ int ctr, pow;
+
+ /* Add new user mappings */
+ if (!rht_node->map[vl])
+ rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
+
+ if (!rht_node->map[vl]) {
+ ret = -ENOMEM;
+ goto out;
+ }
+
+ rht_node->map[vl]->ctr++;
+ ctr = rht_node->map[vl]->ctr;
+ rht_node->map[vl]->sde[ctr - 1] = sde;
+ pow = roundup_pow_of_two(ctr);
+ rht_node->map[vl]->mask = pow - 1;
+
+ /* Populate the sde map table */
+ sdma_populate_sde_map(rht_node->map[vl]);
+ }
+ cpumask_set_cpu(cpu, new_mask);
+ }
+
+ /* Clean up old mappings */
+ for_each_cpu(cpu, cpu_online_mask) {
+ struct sdma_rht_node *rht_node;
+
+ /* Don't cleanup sdes that are set in the new mask */
+ if (cpumask_test_cpu(cpu, mask))
+ continue;
+
+ rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpu,
+ sdma_rht_params);
+ if (rht_node) {
+ bool empty = true;
+ int i;
+
+ /* Remove mappings for old sde */
+ for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
+ if (rht_node->map[i])
+ sdma_cleanup_sde_map(rht_node->map[i],
+ sde);
+
+ /* Free empty hash table entries */
+ for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
+ if (!rht_node->map[i])
+ continue;
+
+ if (rht_node->map[i]->ctr) {
+ empty = false;
+ break;
+ }
+ }
+
+ if (empty) {
+ ret = rhashtable_remove_fast(&dd->sdma_rht,
+ &rht_node->node,
+ sdma_rht_params);
+ WARN_ON(ret);
+
+ for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
+ kfree(rht_node->map[i]);
+
+ kfree(rht_node);
+ }
+ }
+ }
+
+ cpumask_copy(&sde->cpu_mask, new_mask);
+out:
+ mutex_unlock(&process_to_sde_mutex);
+out_free:
+ free_cpumask_var(mask);
+ free_cpumask_var(new_mask);
+ return ret ? : strnlen(buf, PAGE_SIZE);
+}
+
+ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
+{
+ mutex_lock(&process_to_sde_mutex);
+ if (cpumask_empty(&sde->cpu_mask))
+ snprintf(buf, PAGE_SIZE, "%s\n", "empty");
+ else
+ cpumap_print_to_pagebuf(true, buf, &sde->cpu_mask);
+ mutex_unlock(&process_to_sde_mutex);
+ return strnlen(buf, PAGE_SIZE);
+}
+
+static void sdma_rht_free(void *ptr, void *arg)
+{
+ struct sdma_rht_node *rht_node = ptr;
+ int i;
+
+ for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
+ kfree(rht_node->map[i]);
+
+ kfree(rht_node);
+}
+
/*
* Free the indicated map struct
*/
@@ -1161,6 +1466,10 @@ int sdma_init(struct hfi1_devdata *dd, u8 port)
dd->num_sdma = num_engines;
if (sdma_map_init(dd, port, ppd->vls_operational, NULL))
goto bail;
+
+ if (rhashtable_init(&dd->sdma_rht, &sdma_rht_params))
+ goto bail;
+
dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma);
return 0;
@@ -1252,6 +1561,7 @@ void sdma_exit(struct hfi1_devdata *dd)
sdma_finalput(&sde->state);
}
sdma_clean(dd, dd->num_sdma);
+ rhashtable_free_and_destroy(&dd->sdma_rht, sdma_rht_free, NULL);
}
/*
diff --git a/drivers/infiniband/hw/hfi1/sdma.h b/drivers/infiniband/hw/hfi1/sdma.h
index b333afa..93025f6 100644
--- a/drivers/infiniband/hw/hfi1/sdma.h
+++ b/drivers/infiniband/hw/hfi1/sdma.h
@@ -413,6 +413,8 @@ struct sdma_engine {
spinlock_t flushlist_lock;
/* private: */
struct list_head flushlist;
+ struct cpumask cpu_mask;
+ struct kobject kobj;
};
int sdma_init(struct hfi1_devdata *dd, u8 port);
@@ -1059,6 +1061,12 @@ struct sdma_engine *sdma_select_engine_vl(
u32 selector,
u8 vl);
+struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
+ u32 selector, u8 vl);
+ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf);
+ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
+ size_t count);
+int sdma_engine_get_vl(struct sdma_engine *sde);
void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *);
#ifdef CONFIG_SDMA_VERBOSITY
diff --git a/drivers/infiniband/hw/hfi1/sysfs.c b/drivers/infiniband/hw/hfi1/sysfs.c
index 74c84c6..836eea5 100644
--- a/drivers/infiniband/hw/hfi1/sysfs.c
+++ b/drivers/infiniband/hw/hfi1/sysfs.c
@@ -766,13 +766,82 @@ bail:
return ret;
}
+struct sde_attribute {
+ struct attribute attr;
+ ssize_t (*show)(struct sdma_engine *sde, char *buf);
+ ssize_t (*store)(struct sdma_engine *sde, const char *buf, size_t cnt);
+};
+
+static ssize_t sde_show(struct kobject *kobj, struct attribute *attr, char *buf)
+{
+ struct sde_attribute *sde_attr =
+ container_of(attr, struct sde_attribute, attr);
+ struct sdma_engine *sde =
+ container_of(kobj, struct sdma_engine, kobj);
+
+ if (!sde_attr->show)
+ return -EINVAL;
+
+ return sde_attr->show(sde, buf);
+}
+
+static ssize_t sde_store(struct kobject *kobj, struct attribute *attr,
+ const char *buf, size_t count)
+{
+ struct sde_attribute *sde_attr =
+ container_of(attr, struct sde_attribute, attr);
+ struct sdma_engine *sde =
+ container_of(kobj, struct sdma_engine, kobj);
+
+ if (!capable(CAP_SYS_ADMIN))
+ return -EPERM;
+
+ if (!sde_attr->store)
+ return -EINVAL;
+
+ return sde_attr->store(sde, buf, count);
+}
+
+static const struct sysfs_ops sde_sysfs_ops = {
+ .show = sde_show,
+ .store = sde_store,
+};
+
+static struct kobj_type sde_ktype = {
+ .sysfs_ops = &sde_sysfs_ops,
+};
+
+#define SDE_ATTR(_name, _mode, _show, _store) \
+ struct sde_attribute sde_attr_##_name = \
+ __ATTR(_name, _mode, _show, _store)
+
+static ssize_t sde_show_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
+{
+ return sdma_get_cpu_to_sde_map(sde, buf);
+}
+
+static ssize_t sde_store_cpu_to_sde_map(struct sdma_engine *sde,
+ const char *buf, size_t count)
+{
+ return sdma_set_cpu_to_sde_map(sde, buf, count);
+}
+
+static SDE_ATTR(cpu_list, S_IWUSR | S_IRUGO,
+ sde_show_cpu_to_sde_map,
+ sde_store_cpu_to_sde_map);
+
+static struct sde_attribute *sde_attribs[] = {
+ &sde_attr_cpu_list
+};
+
/*
* Register and create our files in /sys/class/infiniband.
*/
int hfi1_verbs_register_sysfs(struct hfi1_devdata *dd)
{
struct ib_device *dev = &dd->verbs_dev.rdi.ibdev;
- int i, ret;
+ struct device *class_dev = &dev->dev;
+ int i, j, ret;
for (i = 0; i < ARRAY_SIZE(hfi1_attributes); ++i) {
ret = device_create_file(&dev->dev, hfi1_attributes[i]);
@@ -780,10 +849,29 @@ int hfi1_verbs_register_sysfs(struct hfi1_devdata *dd)
goto bail;
}
+ for (i = 0; i < dd->num_sdma; i++) {
+ ret = kobject_init_and_add(&dd->per_sdma[i].kobj,
+ &sde_ktype, &class_dev->kobj,
+ "sdma%d", i);
+ if (ret)
+ goto bail;
+
+ for (j = 0; j < ARRAY_SIZE(sde_attribs); j++) {
+ ret = sysfs_create_file(&dd->per_sdma[i].kobj,
+ &sde_attribs[j]->attr);
+ if (ret)
+ goto bail;
+ }
+ }
+
return 0;
bail:
for (i = 0; i < ARRAY_SIZE(hfi1_attributes); ++i)
device_remove_file(&dev->dev, hfi1_attributes[i]);
+
+ for (i = 0; i < dd->num_sdma; i++)
+ kobject_del(&dd->per_sdma[i].kobj);
+
return ret;
}
diff --git a/drivers/infiniband/hw/hfi1/user_sdma.c b/drivers/infiniband/hw/hfi1/user_sdma.c
index bc7e5c1..a761f80 100644
--- a/drivers/infiniband/hw/hfi1/user_sdma.c
+++ b/drivers/infiniband/hw/hfi1/user_sdma.c
@@ -548,7 +548,7 @@ int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec,
u8 opcode, sc, vl;
int req_queued = 0;
u16 dlid;
- u8 selector;
+ u32 selector;
if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
hfi1_cdbg(
@@ -753,12 +753,9 @@ int hfi1_user_sdma_process_request(struct file *fp, struct iovec *iovec,
dlid = be16_to_cpu(req->hdr.lrh[1]);
selector = dlid_to_selector(dlid);
+ selector += uctxt->ctxt + fd->subctxt;
+ req->sde = sdma_select_user_engine(dd, selector, vl);
- /* Have to select the engine */
- req->sde = sdma_select_engine_vl(dd,
- (u32)(uctxt->ctxt + fd->subctxt +
- selector),
- vl);
if (!req->sde || !sdma_running(req->sde)) {
ret = -ECOMM;
goto free_req;
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 7+ messages in thread* [PATCH 3/5] IB/hfi1: Add irq affinity notification handler
[not found] ` <20160925144248.12088.71489.stgit-9QXIwq+3FY+1XWohqUldA0EOCMrvLtNR@public.gmane.org>
2016-09-25 14:44 ` [PATCH 1/5] IB/hfi1: Add sysfs interface for affinity setup Dennis Dalessandro
2016-09-25 14:44 ` [PATCH 2/5] IB/hfi1: Add a new VL sysfs attribute for sdma engines Dennis Dalessandro
@ 2016-09-25 14:44 ` Dennis Dalessandro
2016-09-25 14:44 ` [PATCH 4/5] IB/hfi1: Add new debugfs sdma_cpu_list file Dennis Dalessandro
` (2 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Dennis Dalessandro @ 2016-09-25 14:44 UTC (permalink / raw)
To: dledford-H+wXaHxf7aLQT0dZR+AlfA
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA, Sebastian Sanchez,
Jianxin Xiong, Tadeusz Struk
From: Tadeusz Struk <tadeusz.struk-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
This patch adds an irq affinity notification handler.
When a user changes interrupt affinity settings for an sdma engine,
the driver needs to make changes to its internal sde structures and
also update the affinity_hint.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Sebastian Sanchez <sebastian.sanchez-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Jianxin Xiong <jianxin.xiong-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Tadeusz Struk <tadeusz.struk-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
drivers/infiniband/hw/hfi1/affinity.c | 112 ++++++++++++++++++++++++++++++---
drivers/infiniband/hw/hfi1/hfi.h | 1
2 files changed, 103 insertions(+), 10 deletions(-)
diff --git a/drivers/infiniband/hw/hfi1/affinity.c b/drivers/infiniband/hw/hfi1/affinity.c
index 17c805a..a26a9a0 100644
--- a/drivers/infiniband/hw/hfi1/affinity.c
+++ b/drivers/infiniband/hw/hfi1/affinity.c
@@ -47,6 +47,7 @@
#include <linux/topology.h>
#include <linux/cpumask.h>
#include <linux/module.h>
+#include <linux/interrupt.h>
#include "hfi.h"
#include "affinity.h"
@@ -309,6 +310,101 @@ int hfi1_dev_affinity_init(struct hfi1_devdata *dd)
}
/*
+ * Function updates the irq affinity hint for msix after it has been changed
+ * by the user using the /proc/irq interface. This function only accepts
+ * one cpu in the mask.
+ */
+static void hfi1_update_sdma_affinity(struct hfi1_msix_entry *msix, int cpu)
+{
+ struct sdma_engine *sde = msix->arg;
+ struct hfi1_devdata *dd = sde->dd;
+ struct hfi1_affinity_node *entry;
+ struct cpu_mask_set *set;
+ int i, old_cpu;
+
+ if (cpu > num_online_cpus() || cpu == sde->cpu)
+ return;
+
+ mutex_lock(&node_affinity.lock);
+ entry = node_affinity_lookup(dd->node);
+ if (!entry)
+ goto unlock;
+
+ old_cpu = sde->cpu;
+ sde->cpu = cpu;
+ cpumask_clear(&msix->mask);
+ cpumask_set_cpu(cpu, &msix->mask);
+ dd_dev_dbg(dd, "IRQ vector: %u, type %s engine %u -> cpu: %d\n",
+ msix->msix.vector, irq_type_names[msix->type],
+ sde->this_idx, cpu);
+ irq_set_affinity_hint(msix->msix.vector, &msix->mask);
+
+ /*
+ * Set the new cpu in the hfi1_affinity_node and clean
+ * the old cpu if it is not used by any other IRQ
+ */
+ set = &entry->def_intr;
+ cpumask_set_cpu(cpu, &set->mask);
+ cpumask_set_cpu(cpu, &set->used);
+ for (i = 0; i < dd->num_msix_entries; i++) {
+ struct hfi1_msix_entry *other_msix;
+
+ other_msix = &dd->msix_entries[i];
+ if (other_msix->type != IRQ_SDMA || other_msix == msix)
+ continue;
+
+ if (cpumask_test_cpu(old_cpu, &other_msix->mask))
+ goto unlock;
+ }
+ cpumask_clear_cpu(old_cpu, &set->mask);
+ cpumask_clear_cpu(old_cpu, &set->used);
+unlock:
+ mutex_unlock(&node_affinity.lock);
+}
+
+static void hfi1_irq_notifier_notify(struct irq_affinity_notify *notify,
+ const cpumask_t *mask)
+{
+ int cpu = cpumask_first(mask);
+ struct hfi1_msix_entry *msix = container_of(notify,
+ struct hfi1_msix_entry,
+ notify);
+
+ /* Only one CPU configuration supported currently */
+ hfi1_update_sdma_affinity(msix, cpu);
+}
+
+static void hfi1_irq_notifier_release(struct kref *ref)
+{
+ /*
+ * This is required by affinity notifier. We don't have anything to
+ * free here.
+ */
+}
+
+static void hfi1_setup_sdma_notifier(struct hfi1_msix_entry *msix)
+{
+ struct irq_affinity_notify *notify = &msix->notify;
+
+ notify->irq = msix->msix.vector;
+ notify->notify = hfi1_irq_notifier_notify;
+ notify->release = hfi1_irq_notifier_release;
+
+ if (irq_set_affinity_notifier(notify->irq, notify))
+ pr_err("Failed to register sdma irq affinity notifier for irq %d\n",
+ notify->irq);
+}
+
+static void hfi1_cleanup_sdma_notifier(struct hfi1_msix_entry *msix)
+{
+ struct irq_affinity_notify *notify = &msix->notify;
+
+ if (irq_set_affinity_notifier(notify->irq, NULL))
+ pr_err("Failed to cleanup sdma irq affinity notifier for irq %d\n",
+ notify->irq);
+}
+
+/*
* Function sets the irq affinity for msix.
* It *must* be called with node_affinity.lock held.
*/
@@ -374,22 +470,17 @@ static int get_irq_affinity(struct hfi1_devdata *dd,
cpumask_set_cpu(cpu, &set->used);
}
- switch (msix->type) {
- case IRQ_SDMA:
- sde->cpu = cpu;
- break;
- case IRQ_GENERAL:
- case IRQ_RCVCTXT:
- case IRQ_OTHER:
- break;
- }
-
cpumask_set_cpu(cpu, &msix->mask);
dd_dev_info(dd, "IRQ vector: %u, type %s %s -> cpu: %d\n",
msix->msix.vector, irq_type_names[msix->type],
extra, cpu);
irq_set_affinity_hint(msix->msix.vector, &msix->mask);
+ if (msix->type == IRQ_SDMA) {
+ sde->cpu = cpu;
+ hfi1_setup_sdma_notifier(msix);
+ }
+
free_cpumask_var(diff);
return 0;
}
@@ -417,6 +508,7 @@ void hfi1_put_irq_affinity(struct hfi1_devdata *dd,
switch (msix->type) {
case IRQ_SDMA:
set = &entry->def_intr;
+ hfi1_cleanup_sdma_notifier(msix);
break;
case IRQ_GENERAL:
/* Don't do accounting for general contexts */
diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h
index 59f69fa..7eef11b 100644
--- a/drivers/infiniband/hw/hfi1/hfi.h
+++ b/drivers/infiniband/hw/hfi1/hfi.h
@@ -531,6 +531,7 @@ struct hfi1_msix_entry {
void *arg;
char name[MAX_NAME_SIZE];
cpumask_t mask;
+ struct irq_affinity_notify notify;
};
/* per-SL CCA information */
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 7+ messages in thread* [PATCH 4/5] IB/hfi1: Add new debugfs sdma_cpu_list file
[not found] ` <20160925144248.12088.71489.stgit-9QXIwq+3FY+1XWohqUldA0EOCMrvLtNR@public.gmane.org>
` (2 preceding siblings ...)
2016-09-25 14:44 ` [PATCH 3/5] IB/hfi1: Add irq affinity notification handler Dennis Dalessandro
@ 2016-09-25 14:44 ` Dennis Dalessandro
2016-09-25 14:44 ` [PATCH 5/5] IB/hfi1: Document new sysfs entries for hfi1 driver Dennis Dalessandro
2016-10-03 14:31 ` [PATCH 0/5] IB/hfi1: sdma engine affinity improvements Doug Ledford
5 siblings, 0 replies; 7+ messages in thread
From: Dennis Dalessandro @ 2016-09-25 14:44 UTC (permalink / raw)
To: dledford-H+wXaHxf7aLQT0dZR+AlfA
Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA, Sebastian Sanchez,
Jianxin Xiong, Tadeusz Struk
From: Tadeusz Struk <tadeusz.struk-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Add a debugfs sdma_cpu_list file that can be used to examine the CPU to
sdma engine assignments for the whole device.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Sebastian Sanchez <sebastian.sanchez-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Reviewed-by: Jianxin Xiong <jianxin.xiong-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Tadeusz Struk <tadeusz.struk-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
drivers/infiniband/hw/hfi1/debugfs.c | 38 ++++++++++++++++++++++++++++++
drivers/infiniband/hw/hfi1/sdma.c | 43 ++++++++++++++++++++++++++++++++++
drivers/infiniband/hw/hfi1/sdma.h | 2 ++
3 files changed, 83 insertions(+), 0 deletions(-)
diff --git a/drivers/infiniband/hw/hfi1/debugfs.c b/drivers/infiniband/hw/hfi1/debugfs.c
index 5e9be16..632ba21 100644
--- a/drivers/infiniband/hw/hfi1/debugfs.c
+++ b/drivers/infiniband/hw/hfi1/debugfs.c
@@ -933,6 +933,43 @@ static const struct counter_info port_cntr_ops[] = {
DEBUGFS_OPS("asic_flags", asic_flags_read, asic_flags_write),
};
+static void *_sdma_cpu_list_seq_start(struct seq_file *s, loff_t *pos)
+{
+ if (*pos >= num_online_cpus())
+ return NULL;
+
+ return pos;
+}
+
+static void *_sdma_cpu_list_seq_next(struct seq_file *s, void *v, loff_t *pos)
+{
+ ++*pos;
+ if (*pos >= num_online_cpus())
+ return NULL;
+
+ return pos;
+}
+
+static void _sdma_cpu_list_seq_stop(struct seq_file *s, void *v)
+{
+ /* nothing allocated */
+}
+
+static int _sdma_cpu_list_seq_show(struct seq_file *s, void *v)
+{
+ struct hfi1_ibdev *ibd = (struct hfi1_ibdev *)s->private;
+ struct hfi1_devdata *dd = dd_from_dev(ibd);
+ loff_t *spos = v;
+ loff_t i = *spos;
+
+ sdma_seqfile_dump_cpu_list(s, dd, (unsigned long)i);
+ return 0;
+}
+
+DEBUGFS_SEQ_FILE_OPS(sdma_cpu_list);
+DEBUGFS_SEQ_FILE_OPEN(sdma_cpu_list)
+DEBUGFS_FILE_OPS(sdma_cpu_list);
+
void hfi1_dbg_ibdev_init(struct hfi1_ibdev *ibd)
{
char name[sizeof("port0counters") + 1];
@@ -961,6 +998,7 @@ void hfi1_dbg_ibdev_init(struct hfi1_ibdev *ibd)
DEBUGFS_SEQ_FILE_CREATE(ctx_stats, ibd->hfi1_ibdev_dbg, ibd);
DEBUGFS_SEQ_FILE_CREATE(qp_stats, ibd->hfi1_ibdev_dbg, ibd);
DEBUGFS_SEQ_FILE_CREATE(sdes, ibd->hfi1_ibdev_dbg, ibd);
+ DEBUGFS_SEQ_FILE_CREATE(sdma_cpu_list, ibd->hfi1_ibdev_dbg, ibd);
/* dev counter files */
for (i = 0; i < ARRAY_SIZE(cntr_ops); i++)
DEBUGFS_FILE_CREATE(cntr_ops[i].name,
diff --git a/drivers/infiniband/hw/hfi1/sdma.c b/drivers/infiniband/hw/hfi1/sdma.c
index 8cfa960..fd39bca 100644
--- a/drivers/infiniband/hw/hfi1/sdma.c
+++ b/drivers/infiniband/hw/hfi1/sdma.c
@@ -1093,6 +1093,49 @@ static void sdma_rht_free(void *ptr, void *arg)
kfree(rht_node);
}
+/**
+ * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings
+ * @s: seq file
+ * @dd: hfi1_devdata
+ * @cpuid: cpu id
+ *
+ * This routine dumps the process to sde mappings per cpu
+ */
+void sdma_seqfile_dump_cpu_list(struct seq_file *s,
+ struct hfi1_devdata *dd,
+ unsigned long cpuid)
+{
+ struct sdma_rht_node *rht_node;
+ int i, j;
+
+ rht_node = rhashtable_lookup_fast(&dd->sdma_rht, &cpuid,
+ sdma_rht_params);
+ if (!rht_node)
+ return;
+
+ seq_printf(s, "cpu%3lu: ", cpuid);
+ for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
+ if (!rht_node->map[i] || !rht_node->map[i]->ctr)
+ continue;
+
+ seq_printf(s, " vl%d: [", i);
+
+ for (j = 0; j < rht_node->map[i]->ctr; j++) {
+ if (!rht_node->map[i]->sde[j])
+ continue;
+
+ if (j > 0)
+ seq_puts(s, ",");
+
+ seq_printf(s, " sdma%2d",
+ rht_node->map[i]->sde[j]->this_idx);
+ }
+ seq_puts(s, " ]");
+ }
+
+ seq_puts(s, "\n");
+}
+
/*
* Free the indicated map struct
*/
diff --git a/drivers/infiniband/hw/hfi1/sdma.h b/drivers/infiniband/hw/hfi1/sdma.h
index 93025f6..56257ea 100644
--- a/drivers/infiniband/hw/hfi1/sdma.h
+++ b/drivers/infiniband/hw/hfi1/sdma.h
@@ -1068,6 +1068,8 @@ ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
size_t count);
int sdma_engine_get_vl(struct sdma_engine *sde);
void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *);
+void sdma_seqfile_dump_cpu_list(struct seq_file *s, struct hfi1_devdata *dd,
+ unsigned long cpuid);
#ifdef CONFIG_SDMA_VERBOSITY
void sdma_dumpstate(struct sdma_engine *);
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply related [flat|nested] 7+ messages in thread