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/device.c b/drivers/infiniband/hw/cxgb4/device.c
index 2851bf8..986cfd7 100644
--- a/drivers/infiniband/hw/cxgb4/device.c
+++ b/drivers/infiniband/hw/cxgb4/device.c
@@ -364,7 +364,14 @@
printk(KERN_ERR MOD "error %d initializing rqt pool\n", err);
goto err3;
}
+ err = c4iw_ocqp_pool_create(rdev);
+ if (err) {
+ printk(KERN_ERR MOD "error %d initializing ocqp pool\n", err);
+ goto err4;
+ }
return 0;
+err4:
+ c4iw_rqtpool_destroy(rdev);
err3:
c4iw_pblpool_destroy(rdev);
err2:
@@ -391,6 +398,7 @@
idr_destroy(&dev->cqidr);
idr_destroy(&dev->qpidr);
idr_destroy(&dev->mmidr);
+ iounmap(dev->rdev.oc_mw_kva);
ib_dealloc_device(&dev->ibdev);
}
@@ -406,6 +414,17 @@
}
devp->rdev.lldi = *infop;
+ devp->rdev.oc_mw_pa = pci_resource_start(devp->rdev.lldi.pdev, 2) +
+ (pci_resource_len(devp->rdev.lldi.pdev, 2) -
+ roundup_pow_of_two(devp->rdev.lldi.vr->ocq.size));
+ devp->rdev.oc_mw_kva = ioremap_wc(devp->rdev.oc_mw_pa,
+ devp->rdev.lldi.vr->ocq.size);
+
+ printk(KERN_INFO MOD "ocq memory: "
+ "hw_start 0x%x size %u mw_pa 0x%lx mw_kva %p\n",
+ devp->rdev.lldi.vr->ocq.start, devp->rdev.lldi.vr->ocq.size,
+ devp->rdev.oc_mw_pa, devp->rdev.oc_mw_kva);
+
mutex_lock(&dev_mutex);
ret = c4iw_rdev_open(&devp->rdev);
diff --git a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
index 7780116..1c26922 100644
--- a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
+++ b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
@@ -112,8 +112,11 @@
struct c4iw_dev_ucontext uctx;
struct gen_pool *pbl_pool;
struct gen_pool *rqt_pool;
+ struct gen_pool *ocqp_pool;
u32 flags;
struct cxgb4_lld_info lldi;
+ unsigned long oc_mw_pa;
+ void __iomem *oc_mw_kva;
};
static inline int c4iw_fatal_error(struct c4iw_rdev *rdev)
@@ -675,8 +678,10 @@
int c4iw_init_ctrl_qp(struct c4iw_rdev *rdev);
int c4iw_pblpool_create(struct c4iw_rdev *rdev);
int c4iw_rqtpool_create(struct c4iw_rdev *rdev);
+int c4iw_ocqp_pool_create(struct c4iw_rdev *rdev);
void c4iw_pblpool_destroy(struct c4iw_rdev *rdev);
void c4iw_rqtpool_destroy(struct c4iw_rdev *rdev);
+void c4iw_ocqp_pool_destroy(struct c4iw_rdev *rdev);
void c4iw_destroy_resource(struct c4iw_resource *rscp);
int c4iw_destroy_ctrl_qp(struct c4iw_rdev *rdev);
int c4iw_register_device(struct c4iw_dev *dev);
@@ -742,6 +747,8 @@
void c4iw_rqtpool_free(struct c4iw_rdev *rdev, u32 addr, int size);
u32 c4iw_pblpool_alloc(struct c4iw_rdev *rdev, int size);
void c4iw_pblpool_free(struct c4iw_rdev *rdev, u32 addr, int size);
+u32 c4iw_ocqp_pool_alloc(struct c4iw_rdev *rdev, int size);
+void c4iw_ocqp_pool_free(struct c4iw_rdev *rdev, u32 addr, int size);
int c4iw_ofld_send(struct c4iw_rdev *rdev, struct sk_buff *skb);
void c4iw_flush_hw_cq(struct t4_cq *cq);
void c4iw_count_rcqes(struct t4_cq *cq, struct t4_wq *wq, int *count);
diff --git a/drivers/infiniband/hw/cxgb4/provider.c b/drivers/infiniband/hw/cxgb4/provider.c
index 8f645c8..a49a9c1 100644
--- a/drivers/infiniband/hw/cxgb4/provider.c
+++ b/drivers/infiniband/hw/cxgb4/provider.c
@@ -149,19 +149,28 @@
addr = mm->addr;
kfree(mm);
- if ((addr >= pci_resource_start(rdev->lldi.pdev, 2)) &&
- (addr < (pci_resource_start(rdev->lldi.pdev, 2) +
- pci_resource_len(rdev->lldi.pdev, 2)))) {
+ if ((addr >= pci_resource_start(rdev->lldi.pdev, 0)) &&
+ (addr < (pci_resource_start(rdev->lldi.pdev, 0) +
+ pci_resource_len(rdev->lldi.pdev, 0)))) {
/*
- * Map T4 DB register.
+ * MA_SYNC register...
*/
- if (vma->vm_flags & VM_READ)
- return -EPERM;
-
vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
- vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
- vma->vm_flags &= ~VM_MAYREAD;
+ ret = io_remap_pfn_range(vma, vma->vm_start,
+ addr >> PAGE_SHIFT,
+ len, vma->vm_page_prot);
+ } else if ((addr >= pci_resource_start(rdev->lldi.pdev, 2)) &&
+ (addr < (pci_resource_start(rdev->lldi.pdev, 2) +
+ pci_resource_len(rdev->lldi.pdev, 2)))) {
+
+ /*
+ * Map user DB or OCQP memory...
+ */
+ if (addr >= rdev->oc_mw_pa)
+ vma->vm_page_prot = t4_pgprot_wc(vma->vm_page_prot);
+ else
+ vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
ret = io_remap_pfn_range(vma, vma->vm_start,
addr >> PAGE_SHIFT,
len, vma->vm_page_prot);
@@ -472,6 +481,7 @@
dev->ibdev.post_send = c4iw_post_send;
dev->ibdev.post_recv = c4iw_post_receive;
dev->ibdev.get_protocol_stats = c4iw_get_mib;
+ dev->ibdev.uverbs_abi_ver = C4IW_UVERBS_ABI_VERSION;
dev->ibdev.iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL);
if (!dev->ibdev.iwcm)
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:
diff --git a/drivers/infiniband/hw/cxgb4/resource.c b/drivers/infiniband/hw/cxgb4/resource.c
index 26365f6..4fb50d5 100644
--- a/drivers/infiniband/hw/cxgb4/resource.c
+++ b/drivers/infiniband/hw/cxgb4/resource.c
@@ -422,3 +422,59 @@
{
gen_pool_destroy(rdev->rqt_pool);
}
+
+/*
+ * On-Chip QP Memory.
+ */
+#define MIN_OCQP_SHIFT 12 /* 4KB == min ocqp size */
+
+u32 c4iw_ocqp_pool_alloc(struct c4iw_rdev *rdev, int size)
+{
+ unsigned long addr = gen_pool_alloc(rdev->ocqp_pool, size);
+ PDBG("%s addr 0x%x size %d\n", __func__, (u32)addr, size);
+ return (u32)addr;
+}
+
+void c4iw_ocqp_pool_free(struct c4iw_rdev *rdev, u32 addr, int size)
+{
+ PDBG("%s addr 0x%x size %d\n", __func__, addr, size);
+ gen_pool_free(rdev->ocqp_pool, (unsigned long)addr, size);
+}
+
+int c4iw_ocqp_pool_create(struct c4iw_rdev *rdev)
+{
+ unsigned start, chunk, top;
+
+ rdev->ocqp_pool = gen_pool_create(MIN_OCQP_SHIFT, -1);
+ if (!rdev->ocqp_pool)
+ return -ENOMEM;
+
+ start = rdev->lldi.vr->ocq.start;
+ chunk = rdev->lldi.vr->ocq.size;
+ top = start + chunk;
+
+ while (start < top) {
+ chunk = min(top - start + 1, chunk);
+ if (gen_pool_add(rdev->ocqp_pool, start, chunk, -1)) {
+ PDBG("%s failed to add OCQP chunk (%x/%x)\n",
+ __func__, start, chunk);
+ if (chunk <= 1024 << MIN_OCQP_SHIFT) {
+ printk(KERN_WARNING MOD
+ "Failed to add all OCQP chunks (%x/%x)\n",
+ start, top - start);
+ return 0;
+ }
+ chunk >>= 1;
+ } else {
+ PDBG("%s added OCQP chunk (%x/%x)\n",
+ __func__, start, chunk);
+ start += chunk;
+ }
+ }
+ return 0;
+}
+
+void c4iw_ocqp_pool_destroy(struct c4iw_rdev *rdev)
+{
+ gen_pool_destroy(rdev->ocqp_pool);
+}
diff --git a/drivers/infiniband/hw/cxgb4/t4.h b/drivers/infiniband/hw/cxgb4/t4.h
index 24f3690..17ea5fc 100644
--- a/drivers/infiniband/hw/cxgb4/t4.h
+++ b/drivers/infiniband/hw/cxgb4/t4.h
@@ -52,6 +52,7 @@
#define T4_STAG_UNSET 0xffffffff
#define T4_FW_MAJ 0
#define T4_EQ_STATUS_ENTRIES (L1_CACHE_BYTES > 64 ? 2 : 1)
+#define A_PCIE_MA_SYNC 0x30b4
struct t4_status_page {
__be32 rsvd1; /* flit 0 - hw owns */
@@ -266,10 +267,36 @@
u16 idx;
};
+static inline pgprot_t t4_pgprot_wc(pgprot_t prot)
+{
+#if defined(__i386__) || defined(__x86_64__)
+ return pgprot_writecombine(prot);
+#elif defined(CONFIG_PPC64)
+ return __pgprot((pgprot_val(prot) | _PAGE_NO_CACHE) &
+ ~(pgprot_t)_PAGE_GUARDED);
+#else
+ return pgprot_noncached(prot);
+#endif
+}
+
+static inline int t4_ocqp_supported(void)
+{
+#if defined(__i386__) || defined(__x86_64__) || defined(CONFIG_PPC64)
+ return 1;
+#else
+ return 0;
+#endif
+}
+
+enum {
+ T4_SQ_ONCHIP = (1<<0),
+};
+
struct t4_sq {
union t4_wr *queue;
dma_addr_t dma_addr;
DEFINE_DMA_UNMAP_ADDR(mapping);
+ unsigned long phys_addr;
struct t4_swsqe *sw_sq;
struct t4_swsqe *oldest_read;
u64 udb;
@@ -280,6 +307,7 @@
u16 cidx;
u16 pidx;
u16 wq_pidx;
+ u16 flags;
};
struct t4_swrqe {
@@ -350,6 +378,11 @@
wq->rq.cidx = 0;
}
+static inline int t4_sq_onchip(struct t4_sq *sq)
+{
+ return sq->flags & T4_SQ_ONCHIP;
+}
+
static inline int t4_sq_empty(struct t4_wq *wq)
{
return wq->sq.in_use == 0;
@@ -396,30 +429,27 @@
static inline int t4_wq_in_error(struct t4_wq *wq)
{
- return wq->sq.queue[wq->sq.size].status.qp_err;
+ return wq->rq.queue[wq->rq.size].status.qp_err;
}
static inline void t4_set_wq_in_error(struct t4_wq *wq)
{
- wq->sq.queue[wq->sq.size].status.qp_err = 1;
wq->rq.queue[wq->rq.size].status.qp_err = 1;
}
static inline void t4_disable_wq_db(struct t4_wq *wq)
{
- wq->sq.queue[wq->sq.size].status.db_off = 1;
wq->rq.queue[wq->rq.size].status.db_off = 1;
}
static inline void t4_enable_wq_db(struct t4_wq *wq)
{
- wq->sq.queue[wq->sq.size].status.db_off = 0;
wq->rq.queue[wq->rq.size].status.db_off = 0;
}
static inline int t4_wq_db_enabled(struct t4_wq *wq)
{
- return !wq->sq.queue[wq->sq.size].status.db_off;
+ return !wq->rq.queue[wq->rq.size].status.db_off;
}
struct t4_cq {
diff --git a/drivers/infiniband/hw/cxgb4/user.h b/drivers/infiniband/hw/cxgb4/user.h
index ed6414a..e6669d5 100644
--- a/drivers/infiniband/hw/cxgb4/user.h
+++ b/drivers/infiniband/hw/cxgb4/user.h
@@ -50,7 +50,13 @@
__u32 qid_mask;
};
+
+enum {
+ C4IW_QPF_ONCHIP = (1<<0)
+};
+
struct c4iw_create_qp_resp {
+ __u64 ma_sync_key;
__u64 sq_key;
__u64 rq_key;
__u64 sq_db_gts_key;
@@ -62,5 +68,6 @@
__u32 sq_size;
__u32 rq_size;
__u32 qid_mask;
+ __u32 flags;
};
#endif