RDMA/cxgb4: Support on-chip SQs
T4 support on-chip SQs to reduce latency. This patch adds support for
this in iw_cxgb4:
- Manage ocqp memory like other adapter mem resources.
- Allocate user mode SQs from ocqp mem if available.
- Map ocqp mem to user process using write combining.
- Map PCIE_MA_SYNC reg to user process.
Bump uverbs ABI.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index 40187e2..7e45f73 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -31,6 +31,55 @@
*/
#include "iw_cxgb4.h"
+static int ocqp_support;
+module_param(ocqp_support, int, 0644);
+MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=0)");
+
+static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
+{
+ c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
+}
+
+static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
+{
+ dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
+ pci_unmap_addr(sq, mapping));
+}
+
+static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
+{
+ if (t4_sq_onchip(sq))
+ dealloc_oc_sq(rdev, sq);
+ else
+ dealloc_host_sq(rdev, sq);
+}
+
+static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
+{
+ if (!ocqp_support || !t4_ocqp_supported())
+ return -ENOSYS;
+ sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
+ if (!sq->dma_addr)
+ return -ENOMEM;
+ sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
+ rdev->lldi.vr->ocq.start;
+ sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
+ rdev->lldi.vr->ocq.start);
+ sq->flags |= T4_SQ_ONCHIP;
+ return 0;
+}
+
+static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
+{
+ sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
+ &(sq->dma_addr), GFP_KERNEL);
+ if (!sq->queue)
+ return -ENOMEM;
+ sq->phys_addr = virt_to_phys(sq->queue);
+ pci_unmap_addr_set(sq, mapping, sq->dma_addr);
+ return 0;
+}
+
static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
struct c4iw_dev_ucontext *uctx)
{
@@ -41,9 +90,7 @@
dma_free_coherent(&(rdev->lldi.pdev->dev),
wq->rq.memsize, wq->rq.queue,
dma_unmap_addr(&wq->rq, mapping));
- dma_free_coherent(&(rdev->lldi.pdev->dev),
- wq->sq.memsize, wq->sq.queue,
- dma_unmap_addr(&wq->sq, mapping));
+ dealloc_sq(rdev, &wq->sq);
c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
kfree(wq->rq.sw_rq);
kfree(wq->sq.sw_sq);
@@ -93,11 +140,12 @@
if (!wq->rq.rqt_hwaddr)
goto err4;
- wq->sq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
- wq->sq.memsize, &(wq->sq.dma_addr),
- GFP_KERNEL);
- if (!wq->sq.queue)
- goto err5;
+ if (user) {
+ if (alloc_oc_sq(rdev, &wq->sq) && alloc_host_sq(rdev, &wq->sq))
+ goto err5;
+ } else
+ if (alloc_host_sq(rdev, &wq->sq))
+ goto err5;
memset(wq->sq.queue, 0, wq->sq.memsize);
dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
@@ -158,6 +206,7 @@
V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */
V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */
V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */
+ t4_sq_onchip(&wq->sq) ? F_FW_RI_RES_WR_ONCHIP : 0 |
V_FW_RI_RES_WR_IQID(scq->cqid));
res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
V_FW_RI_RES_WR_DCAEN(0) |
@@ -212,9 +261,7 @@
wq->rq.memsize, wq->rq.queue,
dma_unmap_addr(&wq->rq, mapping));
err6:
- dma_free_coherent(&(rdev->lldi.pdev->dev),
- wq->sq.memsize, wq->sq.queue,
- dma_unmap_addr(&wq->sq, mapping));
+ dealloc_sq(rdev, &wq->sq);
err5:
c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
err4:
@@ -1361,7 +1408,7 @@
int sqsize, rqsize;
struct c4iw_ucontext *ucontext;
int ret;
- struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4;
+ struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
PDBG("%s ib_pd %p\n", __func__, pd);
@@ -1459,7 +1506,15 @@
ret = -ENOMEM;
goto err6;
}
-
+ if (t4_sq_onchip(&qhp->wq.sq)) {
+ mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
+ if (!mm5) {
+ ret = -ENOMEM;
+ goto err7;
+ }
+ uresp.flags = C4IW_QPF_ONCHIP;
+ } else
+ uresp.flags = 0;
uresp.qid_mask = rhp->rdev.qpmask;
uresp.sqid = qhp->wq.sq.qid;
uresp.sq_size = qhp->wq.sq.size;
@@ -1468,6 +1523,10 @@
uresp.rq_size = qhp->wq.rq.size;
uresp.rq_memsize = qhp->wq.rq.memsize;
spin_lock(&ucontext->mmap_lock);
+ if (mm5) {
+ uresp.ma_sync_key = ucontext->key;
+ ucontext->key += PAGE_SIZE;
+ }
uresp.sq_key = ucontext->key;
ucontext->key += PAGE_SIZE;
uresp.rq_key = ucontext->key;
@@ -1479,9 +1538,9 @@
spin_unlock(&ucontext->mmap_lock);
ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
if (ret)
- goto err7;
+ goto err8;
mm1->key = uresp.sq_key;
- mm1->addr = virt_to_phys(qhp->wq.sq.queue);
+ mm1->addr = qhp->wq.sq.phys_addr;
mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
insert_mmap(ucontext, mm1);
mm2->key = uresp.rq_key;
@@ -1496,6 +1555,13 @@
mm4->addr = qhp->wq.rq.udb;
mm4->len = PAGE_SIZE;
insert_mmap(ucontext, mm4);
+ if (mm5) {
+ mm5->key = uresp.ma_sync_key;
+ mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
+ + A_PCIE_MA_SYNC) & PAGE_MASK;
+ mm5->len = PAGE_SIZE;
+ insert_mmap(ucontext, mm5);
+ }
}
qhp->ibqp.qp_num = qhp->wq.sq.qid;
init_timer(&(qhp->timer));
@@ -1503,6 +1569,8 @@
__func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
qhp->wq.sq.qid);
return &qhp->ibqp;
+err8:
+ kfree(mm5);
err7:
kfree(mm4);
err6: