From: CQ Tang <cq.t...@intel.com>
Improve performance by changing the behavour of the driver when all SDMA
descriptors are in use, and the processes adding new descriptors are
single- or multi-rail.
For single-rail processes, the driver will block the call and finish
posting all SDMA descriptors onto the hardware queue before returning
back to PSM. Repeated kernel calls are slower than blocking.
For multi-rail processes, the driver will return to PSM as quick as
possible so PSM can feed packets to other rail. If all hardware queues
are full, PSM will buffer the remaining SDMA descriptors until notified
by interrupt that space is available.
This patch builds a red-black tree to track the number rails opened by
a particular PID. If the number is more than one, it is a multi-rail
PSM process, otherwise, it is a single-rail process.
Reviewed-by: Dean Luick <dean.lu...@intel.com>
Reviewed-by: John A Gregor <john.a.gre...@intel.com>
Reviewed-by: Mitko Haralanov <mitko.harala...@intel.com>
Signed-off-by: CQ Tang <cq.t...@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marcinis...@intel.com>
---
drivers/infiniband/hw/qib/qib_user_sdma.c | 136 ++++++++++++++++++++++++++---
1 file changed, 123 insertions(+), 13 deletions(-)
diff --git a/drivers/infiniband/hw/qib/qib_user_sdma.c
b/drivers/infiniband/hw/qib/qib_user_sdma.c
index 165aee2..b225645 100644
--- a/drivers/infiniband/hw/qib/qib_user_sdma.c
+++ b/drivers/infiniband/hw/qib/qib_user_sdma.c
@@ -52,6 +52,17 @@
/* attempt to drain the queue for 5secs */
#define QIB_USER_SDMA_DRAIN_TIMEOUT 500
+/*
+ * track how many times a process open this driver.
+ */
+struct rb_root qib_user_sdma_rb_root = RB_ROOT;
+
+struct qib_user_sdma_rb_node {
+ struct rb_node node;
+ int refcount;
+ pid_t pid;
+};
+
struct qib_user_sdma_pkt {
struct list_head list; /* list element */
@@ -120,15 +131,60 @@ struct qib_user_sdma_queue {
/* dma page table */
struct rb_root dma_pages_root;
+ struct qib_user_sdma_rb_node *sdma_rb_node;
+
/* protect everything above... */
struct mutex lock;
};
+static struct qib_user_sdma_rb_node *
+qib_user_sdma_rb_search(struct rb_root *root, pid_t pid)
+{
+ struct qib_user_sdma_rb_node *sdma_rb_node;
+ struct rb_node *node = root->rb_node;
+
+ while (node) {
+ sdma_rb_node = container_of(node,
+ struct qib_user_sdma_rb_node, node);
+ if (pid < sdma_rb_node->pid)
+ node = node->rb_left;
+ else if (pid > sdma_rb_node->pid)
+ node = node->rb_right;
+ else
+ return sdma_rb_node;
+ }
+ return NULL;
+}
+
+static int
+qib_user_sdma_rb_insert(struct rb_root *root, struct qib_user_sdma_rb_node
*new)
+{
+ struct rb_node **node = &(root->rb_node);
+ struct rb_node *parent = NULL;
+ struct qib_user_sdma_rb_node *got;
+
+ while (*node) {
+ got = container_of(*node, struct qib_user_sdma_rb_node, node);
+ parent = *node;
+ if (new->pid < got->pid)
+ node = &((*node)->rb_left);
+ else if (new->pid > got->pid)
+ node = &((*node)->rb_right);
+ else
+ return 0;
+ }
+
+ rb_link_node(&new->node, parent, node);
+ rb_insert_color(&new->node, root);
+ return 1;
+}
+
struct qib_user_sdma_queue *
qib_user_sdma_queue_create(struct device *dev, int unit, int ctxt, int sctxt)
{
struct qib_user_sdma_queue *pq =
kmalloc(sizeof(struct qib_user_sdma_queue), GFP_KERNEL);
+ struct qib_user_sdma_rb_node *sdma_rb_node;
if (!pq)
goto done;
@@ -138,6 +194,7 @@ qib_user_sdma_queue_create(struct device *dev, int unit,
int ctxt, int sctxt)
pq->num_pending = 0;
pq->num_sending = 0;
pq->added = 0;
+ pq->sdma_rb_node = NULL;
INIT_LIST_HEAD(&pq->sent);
spin_lock_init(&pq->sent_lock);
@@ -163,8 +220,30 @@ qib_user_sdma_queue_create(struct device *dev, int unit,
int ctxt, int sctxt)
pq->dma_pages_root = RB_ROOT;
+ sdma_rb_node = qib_user_sdma_rb_search(&qib_user_sdma_rb_root,
+ current->pid);
+ if (sdma_rb_node) {
+ sdma_rb_node->refcount++;
+ } else {
+ int ret;
+ sdma_rb_node = kmalloc(sizeof(
+ struct qib_user_sdma_rb_node), GFP_KERNEL);
+ if (!sdma_rb_node)
+ goto err_rb;
+
+ sdma_rb_node->refcount = 1;
+ sdma_rb_node->pid = current->pid;
+
+ ret = qib_user_sdma_rb_insert(&qib_user_sdma_rb_root,
+ sdma_rb_node);
+ BUG_ON(ret == 0);
+ }
+ pq->sdma_rb_node = sdma_rb_node;
+
goto done;
+err_rb:
+ dma_pool_destroy(pq->header_cache);
err_slab:
kmem_cache_destroy(pq->pkt_slab);
err_kfree:
@@ -1020,8 +1099,13 @@ void qib_user_sdma_queue_destroy(struct
qib_user_sdma_queue *pq)
if (!pq)
return;
- kmem_cache_destroy(pq->pkt_slab);
+ pq->sdma_rb_node->refcount--;
+ if (pq->sdma_rb_node->refcount == 0) {
+ rb_erase(&pq->sdma_rb_node->node, &qib_user_sdma_rb_root);
+ kfree(pq->sdma_rb_node);
+ }
dma_pool_destroy(pq->header_cache);
+ kmem_cache_destroy(pq->pkt_slab);
kfree(pq);
}
@@ -1241,26 +1325,52 @@ static int qib_user_sdma_push_pkts(struct qib_pportdata
*ppd,
struct qib_user_sdma_queue *pq,
struct list_head *pktlist, int count)
{
- int ret = 0;
unsigned long flags;
if (unlikely(!(ppd->lflags & QIBL_LINKACTIVE)))
return -ECOMM;
- spin_lock_irqsave(&ppd->sdma_lock, flags);
-
- if (unlikely(!__qib_sdma_running(ppd))) {
- ret = -ECOMM;
- goto unlock;
+ /* non-blocking mode */
+ if (pq->sdma_rb_node->refcount > 1) {
+ spin_lock_irqsave(&ppd->sdma_lock, flags);
+ if (unlikely(!__qib_sdma_running(ppd))) {
+ spin_unlock_irqrestore(&ppd->sdma_lock, flags);
+ return -ECOMM;
+ }
+ pq->num_pending += count;
+ list_splice_tail_init(pktlist, &ppd->sdma_userpending);
+ qib_user_sdma_send_desc(ppd, &ppd->sdma_userpending);
+ spin_unlock_irqrestore(&ppd->sdma_lock, flags);
+ return 0;
}
+ /* In this case, descriptors from this process are not
+ * linked to ppd pending queue, interrupt handler
+ * won't update this process, it is OK to directly
+ * modify without sdma lock.
+ */
+
+
pq->num_pending += count;
- list_splice_tail_init(pktlist, &ppd->sdma_userpending);
- qib_user_sdma_send_desc(ppd, &ppd->sdma_userpending);
+ /*
+ * Blocking mode for single rail process, we must
+ * release/regain sdma_lock to give other process
+ * chance to make progress. This is important for
+ * performance.
+ */
+ do {
+ spin_lock_irqsave(&ppd->sdma_lock, flags);
+ if (unlikely(!__qib_sdma_running(ppd))) {
+ spin_unlock_irqrestore(&ppd->sdma_lock, flags);
+ return -ECOMM;
+ }
+ qib_user_sdma_send_desc(ppd, pktlist);
+ if (!list_empty(pktlist))
+ qib_sdma_make_progress(ppd);
+ spin_unlock_irqrestore(&ppd->sdma_lock, flags);
+ } while (!list_empty(pktlist));
-unlock:
- spin_unlock_irqrestore(&ppd->sdma_lock, flags);
- return ret;
+ return 0;
}
int qib_user_sdma_writev(struct qib_ctxtdata *rcd,
@@ -1290,7 +1400,7 @@ int qib_user_sdma_writev(struct qib_ctxtdata *rcd,
qib_user_sdma_queue_clean(ppd, pq);
while (dim) {
- int mxp = 8;
+ int mxp = 1;
int ndesc = 0;
ret = qib_user_sdma_queue_pkts(dd, ppd, pq,
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
