* [PATCH 2.6.37 00/11] cxgb4 fixes / enhancements.
@ 2010-09-10 16:14 Steve Wise
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
0 siblings, 1 reply; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:14 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
Here are some bug fixes and enhancements for 2.6.37.
Steve Wise (11):
RDMA/cxgb4: Set the default TCP send window to 128KB.
RDMA/cxgb4: Use a mutex for QP and EP state transitions.
RDMA/cxgb4: Support on-chip SQs.
RDMA/cxgb4: Centralize the wait logic.
RDMA/cxgb4: debugfs files for dumping active stags.
RDMA/cxgb4: log HW lack-of-resource errors.
RDMA/cxgb4: Handle CPL_RDMA_TERMINATE messages.
RDMA/cxgb4: Ignore TERMINATE CQEs
RDMA/cxgb4: Ignore positive return values from cxgb4_*_send() functions.
RDMA/cxgb4: Zero out ISGL padding.
RDMA/cxgb4: Don't use null ep ptr.
drivers/infiniband/hw/cxgb4/cm.c | 172 +++++++++++-------------
drivers/infiniband/hw/cxgb4/cq.c | 24 +--
drivers/infiniband/hw/cxgb4/device.c | 171 +++++++++++++++++++-----
drivers/infiniband/hw/cxgb4/ev.c | 2
drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 68 +++++++---
drivers/infiniband/hw/cxgb4/mem.c | 9 -
drivers/infiniband/hw/cxgb4/provider.c | 28 +++-
drivers/infiniband/hw/cxgb4/qp.c | 227 +++++++++++++++++++-------------
drivers/infiniband/hw/cxgb4/resource.c | 62 +++++++++
drivers/infiniband/hw/cxgb4/t4.h | 40 +++++-
drivers/infiniband/hw/cxgb4/user.h | 7 +
11 files changed, 528 insertions(+), 282 deletions(-)
--
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 [flat|nested] 15+ messages in thread
* [PATCH 2.6.37 01/11] RDMA/cxgb4: Don't use null ep ptr.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
@ 2010-09-10 16:14 ` Steve Wise
2010-09-10 16:14 ` [PATCH 2.6.37 02/11] RDMA/cxgb4: Zero out ISGL padding Steve Wise
` (9 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:14 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
In c4iw_modify_qp() error path, only use qhp->ep if ep is not set. Otherwise
qhp->ep can be NULL and we crash.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/qp.c | 3 ++-
1 files changed, 2 insertions(+), 1 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index 93f6e5b..05aa20b 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -1305,7 +1305,8 @@ err:
/* disassociate the LLP connection */
qhp->attr.llp_stream_handle = NULL;
- ep = qhp->ep;
+ if (!ep)
+ ep = qhp->ep;
qhp->ep = NULL;
qhp->attr.state = C4IW_QP_STATE_ERROR;
free = 1;
--
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] 15+ messages in thread
* [PATCH 2.6.37 02/11] RDMA/cxgb4: Zero out ISGL padding.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
2010-09-10 16:14 ` [PATCH 2.6.37 01/11] RDMA/cxgb4: Don't use null ep ptr Steve Wise
@ 2010-09-10 16:14 ` Steve Wise
2010-09-10 16:14 ` [PATCH 2.6.37 03/11] RDMA/cxgb4: Ignore positive return values from cxgb4_*_send() functions Steve Wise
` (8 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:14 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
The HW design requires zering any pad in SGLs.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/qp.c | 4 ++++
1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index 05aa20b..c92259f 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -263,6 +263,9 @@ static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
rem -= len;
}
}
+ len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
+ if (len)
+ memset(dstp, 0, len);
immdp->op = FW_RI_DATA_IMMD;
immdp->r1 = 0;
immdp->r2 = 0;
@@ -292,6 +295,7 @@ static int build_isgl(__be64 *queue_start, __be64 *queue_end,
if (++flitp == queue_end)
flitp = queue_start;
}
+ *flitp = (__force __be64)0;
isglp->op = FW_RI_DATA_ISGL;
isglp->r1 = 0;
isglp->nsge = cpu_to_be16(num_sge);
--
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] 15+ messages in thread
* [PATCH 2.6.37 03/11] RDMA/cxgb4: Ignore positive return values from cxgb4_*_send() functions.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
2010-09-10 16:14 ` [PATCH 2.6.37 01/11] RDMA/cxgb4: Don't use null ep ptr Steve Wise
2010-09-10 16:14 ` [PATCH 2.6.37 02/11] RDMA/cxgb4: Zero out ISGL padding Steve Wise
@ 2010-09-10 16:14 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 04/11] RDMA/cxgb4: Ignore TERMINATE CQEs Steve Wise
` (7 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:14 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
The cxgb4_*_send() functions return the NET_XMIT_ values, which are
positive integers or negative errno values. So don't treat positive
return values as an error.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/cm.c | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index 32d352a..d548167 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -172,7 +172,7 @@ static int c4iw_l2t_send(struct c4iw_rdev *rdev, struct sk_buff *skb,
error = cxgb4_l2t_send(rdev->lldi.ports[0], skb, l2e);
if (error < 0)
kfree_skb(skb);
- return error;
+ return error < 0 ? error : 0;
}
int c4iw_ofld_send(struct c4iw_rdev *rdev, struct sk_buff *skb)
@@ -187,7 +187,7 @@ int c4iw_ofld_send(struct c4iw_rdev *rdev, struct sk_buff *skb)
error = cxgb4_ofld_send(rdev->lldi.ports[0], skb);
if (error < 0)
kfree_skb(skb);
- return error;
+ return error < 0 ? error : 0;
}
static void release_tid(struct c4iw_rdev *rdev, u32 hwtid, struct sk_buff *skb)
--
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] 15+ messages in thread
* [PATCH 2.6.37 04/11] RDMA/cxgb4: Ignore TERMINATE CQEs
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (2 preceding siblings ...)
2010-09-10 16:14 ` [PATCH 2.6.37 03/11] RDMA/cxgb4: Ignore positive return values from cxgb4_*_send() functions Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 05/11] RDMA/cxgb4: Handle CPL_RDMA_TERMINATE messages Steve Wise
` (6 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
T4 incorrectly inserts TERM CQEs into the CQ. Silently ignore them.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/cq.c | 6 ++++++
1 files changed, 6 insertions(+), 0 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/cq.c b/drivers/infiniband/hw/cxgb4/cq.c
index b3daf39..d902ed7 100644
--- a/drivers/infiniband/hw/cxgb4/cq.c
+++ b/drivers/infiniband/hw/cxgb4/cq.c
@@ -476,6 +476,11 @@ static int poll_cq(struct t4_wq *wq, struct t4_cq *cq, struct t4_cqe *cqe,
goto proc_cqe;
}
+ if (CQE_OPCODE(hw_cqe) == FW_RI_TERMINATE) {
+ ret = -EAGAIN;
+ goto skip_cqe;
+ }
+
/*
* RECV completion.
*/
@@ -696,6 +701,7 @@ static int c4iw_poll_cq_one(struct c4iw_cq *chp, struct ib_wc *wc)
case T4_ERR_MSN_RANGE:
case T4_ERR_IRD_OVERFLOW:
case T4_ERR_OPCODE:
+ case T4_ERR_INTERNAL_ERR:
wc->status = IB_WC_FATAL_ERR;
break;
case T4_ERR_SWFLUSH:
--
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] 15+ messages in thread
* [PATCH 2.6.37 05/11] RDMA/cxgb4: Handle CPL_RDMA_TERMINATE messages.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (3 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 04/11] RDMA/cxgb4: Ignore TERMINATE CQEs Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 06/11] RDMA/cxgb4: log HW lack-of-resource errors Steve Wise
` (5 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
T4 FW sends up CPL_RDMA_TERMINATE to indicate a peer TERM. This triggers
the QP moving to TERMINATE state.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/cm.c | 25 +++++++++++++------------
drivers/infiniband/hw/cxgb4/ev.c | 2 +-
drivers/infiniband/hw/cxgb4/qp.c | 3 ++-
3 files changed, 16 insertions(+), 14 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index d548167..0227c13 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -1725,23 +1725,24 @@ static int close_con_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
static int terminate(struct c4iw_dev *dev, struct sk_buff *skb)
{
- struct c4iw_ep *ep;
- struct cpl_rdma_terminate *term = cplhdr(skb);
+ struct cpl_rdma_terminate *rpl = cplhdr(skb);
struct tid_info *t = dev->rdev.lldi.tids;
- unsigned int tid = GET_TID(term);
+ unsigned int tid = GET_TID(rpl);
+ struct c4iw_ep *ep;
+ struct c4iw_qp_attributes attrs;
ep = lookup_tid(t, tid);
+ BUG_ON(!ep);
- if (state_read(&ep->com) != FPDU_MODE)
- return 0;
+ if (ep->com.qp) {
+ printk(KERN_WARNING MOD "TERM received tid %u qpid %u\n", tid,
+ ep->com.qp->wq.sq.qid);
+ attrs.next_state = C4IW_QP_STATE_TERMINATE;
+ c4iw_modify_qp(ep->com.qp->rhp, ep->com.qp,
+ C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
+ } else
+ printk(KERN_WARNING MOD "TERM received tid %u no qp\n", tid);
- PDBG("%s ep %p tid %u\n", __func__, ep, ep->hwtid);
- skb_pull(skb, sizeof *term);
- PDBG("%s saving %d bytes of term msg\n", __func__, skb->len);
- skb_copy_from_linear_data(skb, ep->com.qp->attr.terminate_buffer,
- skb->len);
- ep->com.qp->attr.terminate_msg_len = skb->len;
- ep->com.qp->attr.is_terminate_local = 0;
return 0;
}
diff --git a/drivers/infiniband/hw/cxgb4/ev.c b/drivers/infiniband/hw/cxgb4/ev.c
index 491e76a..c13041a 100644
--- a/drivers/infiniband/hw/cxgb4/ev.c
+++ b/drivers/infiniband/hw/cxgb4/ev.c
@@ -60,7 +60,7 @@ static void post_qp_event(struct c4iw_dev *dev, struct c4iw_cq *chp,
if (qhp->attr.state == C4IW_QP_STATE_RTS) {
attrs.next_state = C4IW_QP_STATE_TERMINATE;
c4iw_modify_qp(qhp->rhp, qhp, C4IW_QP_ATTR_NEXT_STATE,
- &attrs, 1);
+ &attrs, 0);
}
event.event = ib_event;
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index c92259f..7756d44 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -1238,7 +1238,8 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
t4_set_wq_in_error(&qhp->wq);
ep = qhp->ep;
c4iw_get_ep(&ep->com);
- terminate = 1;
+ if (!internal)
+ terminate = 1;
disconnect = 1;
break;
case C4IW_QP_STATE_ERROR:
--
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] 15+ messages in thread
* [PATCH 2.6.37 06/11] RDMA/cxgb4: log HW lack-of-resource errors.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (4 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 05/11] RDMA/cxgb4: Handle CPL_RDMA_TERMINATE messages Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 07/11] RDMA/cxgb4: debugfs files for dumping active stags Steve Wise
` (4 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
This helps debug cases where HW resources are depleted.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/resource.c | 6 ++++++
1 files changed, 6 insertions(+), 0 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/resource.c b/drivers/infiniband/hw/cxgb4/resource.c
index 83b23df..26365f6 100644
--- a/drivers/infiniband/hw/cxgb4/resource.c
+++ b/drivers/infiniband/hw/cxgb4/resource.c
@@ -311,6 +311,9 @@ u32 c4iw_pblpool_alloc(struct c4iw_rdev *rdev, int size)
{
unsigned long addr = gen_pool_alloc(rdev->pbl_pool, size);
PDBG("%s addr 0x%x size %d\n", __func__, (u32)addr, size);
+ if (!addr && printk_ratelimit())
+ printk(KERN_WARNING MOD "%s: Out of PBL memory\n",
+ pci_name(rdev->lldi.pdev));
return (u32)addr;
}
@@ -370,6 +373,9 @@ u32 c4iw_rqtpool_alloc(struct c4iw_rdev *rdev, int size)
{
unsigned long addr = gen_pool_alloc(rdev->rqt_pool, size << 6);
PDBG("%s addr 0x%x size %d\n", __func__, (u32)addr, size << 6);
+ if (!addr && printk_ratelimit())
+ printk(KERN_WARNING MOD "%s: Out of RQT memory\n",
+ pci_name(rdev->lldi.pdev));
return (u32)addr;
}
--
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] 15+ messages in thread
* [PATCH 2.6.37 07/11] RDMA/cxgb4: debugfs files for dumping active stags.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (5 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 06/11] RDMA/cxgb4: log HW lack-of-resource errors Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 08/11] RDMA/cxgb4: Centralize the wait logic Steve Wise
` (3 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
Add "stags" debugfs file. This is useful for examining the TPTE
and PBL entries in adapter memory. It allows scripts to dump just the
active entries.
Also clean up the "qps" file handlers and shared common code.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/device.c | 152 +++++++++++++++++++++++++---------
1 files changed, 113 insertions(+), 39 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/device.c b/drivers/infiniband/hw/cxgb4/device.c
index 9bbf491..2851bf8 100644
--- a/drivers/infiniband/hw/cxgb4/device.c
+++ b/drivers/infiniband/hw/cxgb4/device.c
@@ -49,29 +49,57 @@ static DEFINE_MUTEX(dev_mutex);
static struct dentry *c4iw_debugfs_root;
-struct debugfs_qp_data {
+struct c4iw_debugfs_data {
struct c4iw_dev *devp;
char *buf;
int bufsize;
int pos;
};
-static int count_qps(int id, void *p, void *data)
+static int count_idrs(int id, void *p, void *data)
{
- struct c4iw_qp *qp = p;
int *countp = data;
- if (id != qp->wq.sq.qid)
- return 0;
-
*countp = *countp + 1;
return 0;
}
-static int dump_qps(int id, void *p, void *data)
+static ssize_t debugfs_read(struct file *file, char __user *buf, size_t count,
+ loff_t *ppos)
+{
+ struct c4iw_debugfs_data *d = file->private_data;
+ loff_t pos = *ppos;
+ loff_t avail = d->pos;
+
+ if (pos < 0)
+ return -EINVAL;
+ if (pos >= avail)
+ return 0;
+ if (count > avail - pos)
+ count = avail - pos;
+
+ while (count) {
+ size_t len = 0;
+
+ len = min((int)count, (int)d->pos - (int)pos);
+ if (copy_to_user(buf, d->buf + pos, len))
+ return -EFAULT;
+ if (len == 0)
+ return -EINVAL;
+
+ buf += len;
+ pos += len;
+ count -= len;
+ }
+ count = pos - *ppos;
+ *ppos = pos;
+ return count;
+}
+
+static int dump_qp(int id, void *p, void *data)
{
struct c4iw_qp *qp = p;
- struct debugfs_qp_data *qpd = data;
+ struct c4iw_debugfs_data *qpd = data;
int space;
int cc;
@@ -101,7 +129,7 @@ static int dump_qps(int id, void *p, void *data)
static int qp_release(struct inode *inode, struct file *file)
{
- struct debugfs_qp_data *qpd = file->private_data;
+ struct c4iw_debugfs_data *qpd = file->private_data;
if (!qpd) {
printk(KERN_INFO "%s null qpd?\n", __func__);
return 0;
@@ -113,7 +141,7 @@ static int qp_release(struct inode *inode, struct file *file)
static int qp_open(struct inode *inode, struct file *file)
{
- struct debugfs_qp_data *qpd;
+ struct c4iw_debugfs_data *qpd;
int ret = 0;
int count = 1;
@@ -126,7 +154,7 @@ static int qp_open(struct inode *inode, struct file *file)
qpd->pos = 0;
spin_lock_irq(&qpd->devp->lock);
- idr_for_each(&qpd->devp->qpidr, count_qps, &count);
+ idr_for_each(&qpd->devp->qpidr, count_idrs, &count);
spin_unlock_irq(&qpd->devp->lock);
qpd->bufsize = count * 128;
@@ -137,7 +165,7 @@ static int qp_open(struct inode *inode, struct file *file)
}
spin_lock_irq(&qpd->devp->lock);
- idr_for_each(&qpd->devp->qpidr, dump_qps, qpd);
+ idr_for_each(&qpd->devp->qpidr, dump_qp, qpd);
spin_unlock_irq(&qpd->devp->lock);
qpd->buf[qpd->pos++] = 0;
@@ -149,43 +177,84 @@ out:
return ret;
}
-static ssize_t qp_read(struct file *file, char __user *buf, size_t count,
- loff_t *ppos)
+static const struct file_operations qp_debugfs_fops = {
+ .owner = THIS_MODULE,
+ .open = qp_open,
+ .release = qp_release,
+ .read = debugfs_read,
+};
+
+static int dump_stag(int id, void *p, void *data)
{
- struct debugfs_qp_data *qpd = file->private_data;
- loff_t pos = *ppos;
- loff_t avail = qpd->pos;
+ struct c4iw_debugfs_data *stagd = data;
+ int space;
+ int cc;
- if (pos < 0)
- return -EINVAL;
- if (pos >= avail)
+ space = stagd->bufsize - stagd->pos - 1;
+ if (space == 0)
+ return 1;
+
+ cc = snprintf(stagd->buf + stagd->pos, space, "0x%x\n", id<<8);
+ if (cc < space)
+ stagd->pos += cc;
+ return 0;
+}
+
+static int stag_release(struct inode *inode, struct file *file)
+{
+ struct c4iw_debugfs_data *stagd = file->private_data;
+ if (!stagd) {
+ printk(KERN_INFO "%s null stagd?\n", __func__);
return 0;
- if (count > avail - pos)
- count = avail - pos;
+ }
+ kfree(stagd->buf);
+ kfree(stagd);
+ return 0;
+}
- while (count) {
- size_t len = 0;
+static int stag_open(struct inode *inode, struct file *file)
+{
+ struct c4iw_debugfs_data *stagd;
+ int ret = 0;
+ int count = 1;
- len = min((int)count, (int)qpd->pos - (int)pos);
- if (copy_to_user(buf, qpd->buf + pos, len))
- return -EFAULT;
- if (len == 0)
- return -EINVAL;
+ stagd = kmalloc(sizeof *stagd, GFP_KERNEL);
+ if (!stagd) {
+ ret = -ENOMEM;
+ goto out;
+ }
+ stagd->devp = inode->i_private;
+ stagd->pos = 0;
- buf += len;
- pos += len;
- count -= len;
+ spin_lock_irq(&stagd->devp->lock);
+ idr_for_each(&stagd->devp->mmidr, count_idrs, &count);
+ spin_unlock_irq(&stagd->devp->lock);
+
+ stagd->bufsize = count * sizeof("0x12345678\n");
+ stagd->buf = kmalloc(stagd->bufsize, GFP_KERNEL);
+ if (!stagd->buf) {
+ ret = -ENOMEM;
+ goto err1;
}
- count = pos - *ppos;
- *ppos = pos;
- return count;
+
+ spin_lock_irq(&stagd->devp->lock);
+ idr_for_each(&stagd->devp->mmidr, dump_stag, stagd);
+ spin_unlock_irq(&stagd->devp->lock);
+
+ stagd->buf[stagd->pos++] = 0;
+ file->private_data = stagd;
+ goto out;
+err1:
+ kfree(stagd);
+out:
+ return ret;
}
-static const struct file_operations qp_debugfs_fops = {
+static const struct file_operations stag_debugfs_fops = {
.owner = THIS_MODULE,
- .open = qp_open,
- .release = qp_release,
- .read = qp_read,
+ .open = stag_open,
+ .release = stag_release,
+ .read = debugfs_read,
};
static int setup_debugfs(struct c4iw_dev *devp)
@@ -199,6 +268,11 @@ static int setup_debugfs(struct c4iw_dev *devp)
(void *)devp, &qp_debugfs_fops);
if (de && de->d_inode)
de->d_inode->i_size = 4096;
+
+ de = debugfs_create_file("stags", S_IWUSR, devp->debugfs_root,
+ (void *)devp, &stag_debugfs_fops);
+ if (de && de->d_inode)
+ de->d_inode->i_size = 4096;
return 0;
}
--
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] 15+ messages in thread
* [PATCH 2.6.37 08/11] RDMA/cxgb4: Centralize the wait logic.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (6 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 07/11] RDMA/cxgb4: debugfs files for dumping active stags Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs Steve Wise
` (2 subsequent siblings)
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/cm.c | 64 +++++++++++++-------------------
drivers/infiniband/hw/cxgb4/cq.c | 18 +--------
drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 57 ++++++++++++++++++++---------
drivers/infiniband/hw/cxgb4/mem.c | 9 +----
drivers/infiniband/hw/cxgb4/qp.c | 35 +++---------------
5 files changed, 72 insertions(+), 111 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index 0227c13..5547b49 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -252,7 +252,7 @@ static void *alloc_ep(int size, gfp_t gfp)
if (epc) {
kref_init(&epc->kref);
spin_lock_init(&epc->lock);
- init_waitqueue_head(&epc->waitq);
+ c4iw_init_wr_wait(&epc->wr_wait);
}
PDBG("%s alloc ep %p\n", __func__, epc);
return epc;
@@ -1213,9 +1213,9 @@ static int pass_open_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
}
PDBG("%s ep %p status %d error %d\n", __func__, ep,
rpl->status, status2errno(rpl->status));
- ep->com.rpl_err = status2errno(rpl->status);
- ep->com.rpl_done = 1;
- wake_up(&ep->com.waitq);
+ ep->com.wr_wait.ret = status2errno(rpl->status);
+ ep->com.wr_wait.done = 1;
+ wake_up(&ep->com.wr_wait.wait);
return 0;
}
@@ -1249,9 +1249,9 @@ static int close_listsrv_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
struct c4iw_listen_ep *ep = lookup_stid(t, stid);
PDBG("%s ep %p\n", __func__, ep);
- ep->com.rpl_err = status2errno(rpl->status);
- ep->com.rpl_done = 1;
- wake_up(&ep->com.waitq);
+ ep->com.wr_wait.ret = status2errno(rpl->status);
+ ep->com.wr_wait.done = 1;
+ wake_up(&ep->com.wr_wait.wait);
return 0;
}
@@ -1507,17 +1507,17 @@ static int peer_close(struct c4iw_dev *dev, struct sk_buff *skb)
* in rdma connection migration (see c4iw_accept_cr()).
*/
__state_set(&ep->com, CLOSING);
- ep->com.rpl_done = 1;
- ep->com.rpl_err = -ECONNRESET;
+ ep->com.wr_wait.done = 1;
+ ep->com.wr_wait.ret = -ECONNRESET;
PDBG("waking up ep %p tid %u\n", ep, ep->hwtid);
- wake_up(&ep->com.waitq);
+ wake_up(&ep->com.wr_wait.wait);
break;
case MPA_REP_SENT:
__state_set(&ep->com, CLOSING);
- ep->com.rpl_done = 1;
- ep->com.rpl_err = -ECONNRESET;
+ ep->com.wr_wait.done = 1;
+ ep->com.wr_wait.ret = -ECONNRESET;
PDBG("waking up ep %p tid %u\n", ep, ep->hwtid);
- wake_up(&ep->com.waitq);
+ wake_up(&ep->com.wr_wait.wait);
break;
case FPDU_MODE:
start_ep_timer(ep);
@@ -1605,10 +1605,10 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
connect_reply_upcall(ep, -ECONNRESET);
break;
case MPA_REP_SENT:
- ep->com.rpl_done = 1;
- ep->com.rpl_err = -ECONNRESET;
+ ep->com.wr_wait.done = 1;
+ ep->com.wr_wait.ret = -ECONNRESET;
PDBG("waking up ep %p\n", ep);
- wake_up(&ep->com.waitq);
+ wake_up(&ep->com.wr_wait.wait);
break;
case MPA_REQ_RCVD:
@@ -1618,10 +1618,10 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
* rejects the CR. Also wake up anyone waiting
* in rdma connection migration (see c4iw_accept_cr()).
*/
- ep->com.rpl_done = 1;
- ep->com.rpl_err = -ECONNRESET;
+ ep->com.wr_wait.done = 1;
+ ep->com.wr_wait.ret = -ECONNRESET;
PDBG("waking up ep %p tid %u\n", ep, ep->hwtid);
- wake_up(&ep->com.waitq);
+ wake_up(&ep->com.wr_wait.wait);
break;
case MORIBUND:
case CLOSING:
@@ -2043,6 +2043,7 @@ int c4iw_create_listen(struct iw_cm_id *cm_id, int backlog)
}
state_set(&ep->com, LISTEN);
+ c4iw_init_wr_wait(&ep->com.wr_wait);
err = cxgb4_create_server(ep->com.dev->rdev.lldi.ports[0], ep->stid,
ep->com.local_addr.sin_addr.s_addr,
ep->com.local_addr.sin_port,
@@ -2051,15 +2052,8 @@ int c4iw_create_listen(struct iw_cm_id *cm_id, int backlog)
goto fail3;
/* wait for pass_open_rpl */
- wait_event_timeout(ep->com.waitq, ep->com.rpl_done, C4IW_WR_TO);
- if (ep->com.rpl_done)
- err = ep->com.rpl_err;
- else {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(ep->com.dev->rdev.lldi.pdev));
- ep->com.dev->rdev.flags = T4_FATAL_ERROR;
- err = -EIO;
- }
+ err = c4iw_wait_for_reply(&ep->com.dev->rdev, &ep->com.wr_wait, 0, 0,
+ __func__);
if (!err) {
cm_id->provider_data = ep;
goto out;
@@ -2083,20 +2077,12 @@ int c4iw_destroy_listen(struct iw_cm_id *cm_id)
might_sleep();
state_set(&ep->com, DEAD);
- ep->com.rpl_done = 0;
- ep->com.rpl_err = 0;
+ c4iw_init_wr_wait(&ep->com.wr_wait);
err = listen_stop(ep);
if (err)
goto done;
- wait_event_timeout(ep->com.waitq, ep->com.rpl_done, C4IW_WR_TO);
- if (ep->com.rpl_done)
- err = ep->com.rpl_err;
- else {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(ep->com.dev->rdev.lldi.pdev));
- ep->com.dev->rdev.flags = T4_FATAL_ERROR;
- err = -EIO;
- }
+ err = c4iw_wait_for_reply(&ep->com.dev->rdev, &ep->com.wr_wait, 0, 0,
+ __func__);
cxgb4_free_stid(ep->com.dev->rdev.lldi.tids, ep->stid, PF_INET);
done:
cm_id->rem_ref(cm_id);
diff --git a/drivers/infiniband/hw/cxgb4/cq.c b/drivers/infiniband/hw/cxgb4/cq.c
index d902ed7..1da710e 100644
--- a/drivers/infiniband/hw/cxgb4/cq.c
+++ b/drivers/infiniband/hw/cxgb4/cq.c
@@ -64,14 +64,7 @@ static int destroy_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
c4iw_init_wr_wait(&wr_wait);
ret = c4iw_ofld_send(rdev, skb);
if (!ret) {
- wait_event_timeout(wr_wait.wait, wr_wait.done, C4IW_WR_TO);
- if (!wr_wait.done) {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(rdev->lldi.pdev));
- rdev->flags = T4_FATAL_ERROR;
- ret = -EIO;
- } else
- ret = wr_wait.ret;
+ ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
}
kfree(cq->sw_queue);
@@ -157,14 +150,7 @@ static int create_cq(struct c4iw_rdev *rdev, struct t4_cq *cq,
if (ret)
goto err4;
PDBG("%s wait_event wr_wait %p\n", __func__, &wr_wait);
- wait_event_timeout(wr_wait.wait, wr_wait.done, C4IW_WR_TO);
- if (!wr_wait.done) {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(rdev->lldi.pdev));
- rdev->flags = T4_FATAL_ERROR;
- ret = -EIO;
- } else
- ret = wr_wait.ret;
+ ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
if (ret)
goto err4;
diff --git a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
index ed459b8..7780116 100644
--- a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
+++ b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
@@ -79,21 +79,6 @@ static inline void *cplhdr(struct sk_buff *skb)
return skb->data;
}
-#define C4IW_WR_TO (10*HZ)
-
-struct c4iw_wr_wait {
- wait_queue_head_t wait;
- int done;
- int ret;
-};
-
-static inline void c4iw_init_wr_wait(struct c4iw_wr_wait *wr_waitp)
-{
- wr_waitp->ret = 0;
- wr_waitp->done = 0;
- init_waitqueue_head(&wr_waitp->wait);
-}
-
struct c4iw_resource {
struct kfifo tpt_fifo;
spinlock_t tpt_fifo_lock;
@@ -141,6 +126,44 @@ static inline int c4iw_num_stags(struct c4iw_rdev *rdev)
return min((int)T4_MAX_NUM_STAG, (int)(rdev->lldi.vr->stag.size >> 5));
}
+#define C4IW_WR_TO (10*HZ)
+
+struct c4iw_wr_wait {
+ wait_queue_head_t wait;
+ int done;
+ int ret;
+};
+
+static inline void c4iw_init_wr_wait(struct c4iw_wr_wait *wr_waitp)
+{
+ wr_waitp->ret = 0;
+ wr_waitp->done = 0;
+ init_waitqueue_head(&wr_waitp->wait);
+}
+
+static inline int c4iw_wait_for_reply(struct c4iw_rdev *rdev,
+ struct c4iw_wr_wait *wr_waitp,
+ u32 hwtid, u32 qpid,
+ const char *func)
+{
+ unsigned to = C4IW_WR_TO;
+ do {
+
+ wait_event_timeout(wr_waitp->wait, wr_waitp->done, to);
+ if (!wr_waitp->done) {
+ printk(KERN_ERR MOD "%s - Device %s not responding - "
+ "tid %u qpid %u\n", func,
+ pci_name(rdev->lldi.pdev), hwtid, qpid);
+ to = to << 2;
+ }
+ } while (!wr_waitp->done);
+ if (wr_waitp->ret)
+ printk(KERN_WARNING MOD "%s: FW reply %d tid %u qpid %u\n",
+ pci_name(rdev->lldi.pdev), wr_waitp->ret, hwtid, qpid);
+ return wr_waitp->ret;
+}
+
+
struct c4iw_dev {
struct ib_device ibdev;
struct c4iw_rdev rdev;
@@ -582,9 +605,7 @@ struct c4iw_ep_common {
spinlock_t lock;
struct sockaddr_in local_addr;
struct sockaddr_in remote_addr;
- wait_queue_head_t waitq;
- int rpl_done;
- int rpl_err;
+ struct c4iw_wr_wait wr_wait;
unsigned long flags;
};
diff --git a/drivers/infiniband/hw/cxgb4/mem.c b/drivers/infiniband/hw/cxgb4/mem.c
index 269373a..1cbc56f 100644
--- a/drivers/infiniband/hw/cxgb4/mem.c
+++ b/drivers/infiniband/hw/cxgb4/mem.c
@@ -103,14 +103,7 @@ static int write_adapter_mem(struct c4iw_rdev *rdev, u32 addr, u32 len,
len -= C4IW_MAX_INLINE_SIZE;
}
- wait_event_timeout(wr_wait.wait, wr_wait.done, C4IW_WR_TO);
- if (!wr_wait.done) {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(rdev->lldi.pdev));
- rdev->flags = T4_FATAL_ERROR;
- ret = -EIO;
- } else
- ret = wr_wait.ret;
+ ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, 0, __func__);
return ret;
}
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index 7756d44..ee785e2 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -198,14 +198,7 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
ret = c4iw_ofld_send(rdev, skb);
if (ret)
goto err7;
- wait_event_timeout(wr_wait.wait, wr_wait.done, C4IW_WR_TO);
- if (!wr_wait.done) {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(rdev->lldi.pdev));
- rdev->flags = T4_FATAL_ERROR;
- ret = -EIO;
- } else
- ret = wr_wait.ret;
+ ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
if (ret)
goto err7;
@@ -997,20 +990,8 @@ static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
if (ret)
goto out;
- wait_event_timeout(wr_wait.wait, wr_wait.done, C4IW_WR_TO);
- if (!wr_wait.done) {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(rhp->rdev.lldi.pdev));
- rhp->rdev.flags = T4_FATAL_ERROR;
- ret = -EIO;
- } else {
- ret = wr_wait.ret;
- if (ret)
- printk(KERN_WARNING MOD
- "%s: Abnormal close qpid %d ret %u\n",
- pci_name(rhp->rdev.lldi.pdev), qhp->wq.sq.qid,
- ret);
- }
+ ret = c4iw_wait_for_reply(&rhp->rdev, &wr_wait, qhp->ep->hwtid,
+ qhp->wq.sq.qid, __func__);
out:
PDBG("%s ret %d\n", __func__, ret);
return ret;
@@ -1106,14 +1087,8 @@ static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
if (ret)
goto out;
- wait_event_timeout(wr_wait.wait, wr_wait.done, C4IW_WR_TO);
- if (!wr_wait.done) {
- printk(KERN_ERR MOD "Device %s not responding!\n",
- pci_name(rhp->rdev.lldi.pdev));
- rhp->rdev.flags = T4_FATAL_ERROR;
- ret = -EIO;
- } else
- ret = wr_wait.ret;
+ ret = c4iw_wait_for_reply(&rhp->rdev, &wr_wait, qhp->ep->hwtid,
+ qhp->wq.sq.qid, __func__);
out:
PDBG("%s ret %d\n", __func__, ret);
return ret;
--
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] 15+ messages in thread
* [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (7 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 08/11] RDMA/cxgb4: Centralize the wait logic Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
[not found] ` <20100910161530.6829.89294.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
2010-09-10 16:15 ` [PATCH 2.6.37 10/11] RDMA/cxgb4: Use a mutex for QP and EP state transitions Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 11/11] RDMA/cxgb4: Set the default TCP send window to 128KB Steve Wise
10 siblings, 1 reply; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
T4 support on-chip SQs to reduce latency. This patch adds
support for this in iw_cxgb4.
Changes:
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-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/device.c | 19 ++++++
drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 7 ++
drivers/infiniband/hw/cxgb4/provider.c | 28 ++++++---
drivers/infiniband/hw/cxgb4/qp.c | 98 +++++++++++++++++++++++++++-----
drivers/infiniband/hw/cxgb4/resource.c | 56 ++++++++++++++++++
drivers/infiniband/hw/cxgb4/t4.h | 40 +++++++++++--
drivers/infiniband/hw/cxgb4/user.h | 7 ++
7 files changed, 226 insertions(+), 29 deletions(-)
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 @@ static int c4iw_rdev_open(struct c4iw_rdev *rdev)
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 @@ static void c4iw_remove(struct c4iw_dev *dev)
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 @@ static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
}
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_rdev {
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_resource(struct c4iw_rdev *rdev, u32 nr_tpt, u32 nr_pdid);
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 @@ u32 c4iw_rqtpool_alloc(struct c4iw_rdev *rdev, int size);
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 @@ static int c4iw_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
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 @@ int c4iw_register_device(struct c4iw_dev *dev)
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 ee785e2..e0f433f 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 @@ static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
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 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
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 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
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 @@ err7:
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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
__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 @@ void c4iw_rqtpool_destroy(struct c4iw_rdev *rdev)
{
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..51a845f 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 @@ struct t4_swsqe {
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 @@ struct t4_sq {
u16 cidx;
u16 pidx;
u16 wq_pidx;
+ u16 flags;
};
struct t4_swrqe {
@@ -350,6 +378,11 @@ static inline void t4_rq_consume(struct t4_wq *wq)
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 void t4_ring_rq_db(struct t4_wq *wq, u16 inc)
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->sq.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->sq.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 @@ struct c4iw_create_cq_resp {
__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 @@ struct c4iw_create_qp_resp {
__u32 sq_size;
__u32 rq_size;
__u32 qid_mask;
+ __u32 flags;
};
#endif
--
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] 15+ messages in thread
* [PATCH 2.6.37 10/11] RDMA/cxgb4: Use a mutex for QP and EP state transitions.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (8 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 11/11] RDMA/cxgb4: Set the default TCP send window to 128KB Steve Wise
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
Move the connection setup/teardown paths to the workq thread removing
spin lock/irq disable requirements for these paths. This allows calls
down to the LLD for EP and QP state transition actions to be atomic
with respect to processing CPL messages coming up from the HW. Namely,
calls to rdma_init() and rdma_fini() can now be called with the mutex
held avoiding many race conditions with the abort path.
The QP spinlock is still used but only to manipulate the qp state. This
allows the fastpaths, poll, post_send, and pos_recv, to run in the
irq context.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/cm.c | 87 +++++++++++++++----------------
drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 4 +
drivers/infiniband/hw/cxgb4/qp.c | 90 +++++++++++++++-----------------
3 files changed, 88 insertions(+), 93 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index 5547b49..3e6234c 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -219,12 +219,11 @@ static void set_emss(struct c4iw_ep *ep, u16 opt)
static enum c4iw_ep_state state_read(struct c4iw_ep_common *epc)
{
- unsigned long flags;
enum c4iw_ep_state state;
- spin_lock_irqsave(&epc->lock, flags);
+ mutex_lock(&epc->mutex);
state = epc->state;
- spin_unlock_irqrestore(&epc->lock, flags);
+ mutex_unlock(&epc->mutex);
return state;
}
@@ -235,12 +234,10 @@ static void __state_set(struct c4iw_ep_common *epc, enum c4iw_ep_state new)
static void state_set(struct c4iw_ep_common *epc, enum c4iw_ep_state new)
{
- unsigned long flags;
-
- spin_lock_irqsave(&epc->lock, flags);
+ mutex_lock(&epc->mutex);
PDBG("%s - %s -> %s\n", __func__, states[epc->state], states[new]);
__state_set(epc, new);
- spin_unlock_irqrestore(&epc->lock, flags);
+ mutex_unlock(&epc->mutex);
return;
}
@@ -251,7 +248,7 @@ static void *alloc_ep(int size, gfp_t gfp)
epc = kzalloc(size, gfp);
if (epc) {
kref_init(&epc->kref);
- spin_lock_init(&epc->lock);
+ mutex_init(&epc->mutex);
c4iw_init_wr_wait(&epc->wr_wait);
}
PDBG("%s alloc ep %p\n", __func__, epc);
@@ -1131,7 +1128,6 @@ static int abort_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
{
struct c4iw_ep *ep;
struct cpl_abort_rpl_rss *rpl = cplhdr(skb);
- unsigned long flags;
int release = 0;
unsigned int tid = GET_TID(rpl);
struct tid_info *t = dev->rdev.lldi.tids;
@@ -1139,7 +1135,7 @@ static int abort_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
ep = lookup_tid(t, tid);
PDBG("%s ep %p tid %u\n", __func__, ep, ep->hwtid);
BUG_ON(!ep);
- spin_lock_irqsave(&ep->com.lock, flags);
+ mutex_lock(&ep->com.mutex);
switch (ep->com.state) {
case ABORTING:
__state_set(&ep->com, DEAD);
@@ -1150,7 +1146,7 @@ static int abort_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
__func__, ep, ep->com.state);
break;
}
- spin_unlock_irqrestore(&ep->com.lock, flags);
+ mutex_unlock(&ep->com.mutex);
if (release)
release_ep_resources(ep);
@@ -1478,7 +1474,6 @@ static int peer_close(struct c4iw_dev *dev, struct sk_buff *skb)
struct cpl_peer_close *hdr = cplhdr(skb);
struct c4iw_ep *ep;
struct c4iw_qp_attributes attrs;
- unsigned long flags;
int disconnect = 1;
int release = 0;
int closing = 0;
@@ -1489,7 +1484,7 @@ static int peer_close(struct c4iw_dev *dev, struct sk_buff *skb)
PDBG("%s ep %p tid %u\n", __func__, ep, ep->hwtid);
dst_confirm(ep->dst);
- spin_lock_irqsave(&ep->com.lock, flags);
+ mutex_lock(&ep->com.mutex);
switch (ep->com.state) {
case MPA_REQ_WAIT:
__state_set(&ep->com, CLOSING);
@@ -1550,7 +1545,7 @@ static int peer_close(struct c4iw_dev *dev, struct sk_buff *skb)
default:
BUG_ON(1);
}
- spin_unlock_irqrestore(&ep->com.lock, flags);
+ mutex_unlock(&ep->com.mutex);
if (closing) {
attrs.next_state = C4IW_QP_STATE_CLOSING;
c4iw_modify_qp(ep->com.qp->rhp, ep->com.qp,
@@ -1581,7 +1576,6 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
struct c4iw_qp_attributes attrs;
int ret;
int release = 0;
- unsigned long flags;
struct tid_info *t = dev->rdev.lldi.tids;
unsigned int tid = GET_TID(req);
@@ -1591,9 +1585,17 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
ep->hwtid);
return 0;
}
- spin_lock_irqsave(&ep->com.lock, flags);
PDBG("%s ep %p tid %u state %u\n", __func__, ep, ep->hwtid,
ep->com.state);
+
+ /*
+ * Wake up any threads in rdma_init() or rdma_fini().
+ */
+ ep->com.wr_wait.done = 1;
+ ep->com.wr_wait.ret = -ECONNRESET;
+ wake_up(&ep->com.wr_wait.wait);
+
+ mutex_lock(&ep->com.mutex);
switch (ep->com.state) {
case CONNECTING:
break;
@@ -1605,23 +1607,8 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
connect_reply_upcall(ep, -ECONNRESET);
break;
case MPA_REP_SENT:
- ep->com.wr_wait.done = 1;
- ep->com.wr_wait.ret = -ECONNRESET;
- PDBG("waking up ep %p\n", ep);
- wake_up(&ep->com.wr_wait.wait);
break;
case MPA_REQ_RCVD:
-
- /*
- * We're gonna mark this puppy DEAD, but keep
- * the reference on it until the ULP accepts or
- * rejects the CR. Also wake up anyone waiting
- * in rdma connection migration (see c4iw_accept_cr()).
- */
- ep->com.wr_wait.done = 1;
- ep->com.wr_wait.ret = -ECONNRESET;
- PDBG("waking up ep %p tid %u\n", ep, ep->hwtid);
- wake_up(&ep->com.wr_wait.wait);
break;
case MORIBUND:
case CLOSING:
@@ -1644,7 +1631,7 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
break;
case DEAD:
PDBG("%s PEER_ABORT IN DEAD STATE!!!!\n", __func__);
- spin_unlock_irqrestore(&ep->com.lock, flags);
+ mutex_unlock(&ep->com.mutex);
return 0;
default:
BUG_ON(1);
@@ -1655,7 +1642,7 @@ static int peer_abort(struct c4iw_dev *dev, struct sk_buff *skb)
__state_set(&ep->com, DEAD);
release = 1;
}
- spin_unlock_irqrestore(&ep->com.lock, flags);
+ mutex_unlock(&ep->com.mutex);
rpl_skb = get_skb(skb, sizeof(*rpl), GFP_KERNEL);
if (!rpl_skb) {
@@ -1681,7 +1668,6 @@ static int close_con_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
struct c4iw_ep *ep;
struct c4iw_qp_attributes attrs;
struct cpl_close_con_rpl *rpl = cplhdr(skb);
- unsigned long flags;
int release = 0;
struct tid_info *t = dev->rdev.lldi.tids;
unsigned int tid = GET_TID(rpl);
@@ -1692,7 +1678,7 @@ static int close_con_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
BUG_ON(!ep);
/* The cm_id may be null if we failed to connect */
- spin_lock_irqsave(&ep->com.lock, flags);
+ mutex_lock(&ep->com.mutex);
switch (ep->com.state) {
case CLOSING:
__state_set(&ep->com, MORIBUND);
@@ -1717,7 +1703,7 @@ static int close_con_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
BUG_ON(1);
break;
}
- spin_unlock_irqrestore(&ep->com.lock, flags);
+ mutex_unlock(&ep->com.mutex);
if (release)
release_ep_resources(ep);
return 0;
@@ -2093,12 +2079,11 @@ done:
int c4iw_ep_disconnect(struct c4iw_ep *ep, int abrupt, gfp_t gfp)
{
int ret = 0;
- unsigned long flags;
int close = 0;
int fatal = 0;
struct c4iw_rdev *rdev;
- spin_lock_irqsave(&ep->com.lock, flags);
+ mutex_lock(&ep->com.mutex);
PDBG("%s ep %p state %s, abrupt %d\n", __func__, ep,
states[ep->com.state], abrupt);
@@ -2145,7 +2130,7 @@ int c4iw_ep_disconnect(struct c4iw_ep *ep, int abrupt, gfp_t gfp)
break;
}
- spin_unlock_irqrestore(&ep->com.lock, flags);
+ mutex_unlock(&ep->com.mutex);
if (close) {
if (abrupt)
ret = abort_connection(ep, NULL, gfp);
@@ -2159,6 +2144,13 @@ int c4iw_ep_disconnect(struct c4iw_ep *ep, int abrupt, gfp_t gfp)
return ret;
}
+static int async_event(struct c4iw_dev *dev, struct sk_buff *skb)
+{
+ struct cpl_fw6_msg *rpl = cplhdr(skb);
+ c4iw_ev_dispatch(dev, (struct t4_cqe *)&rpl->data[0]);
+ return 0;
+}
+
/*
* These are the real handlers that are called from a
* work queue.
@@ -2177,7 +2169,8 @@ static c4iw_handler_func work_handlers[NUM_CPL_CMDS] = {
[CPL_ABORT_REQ_RSS] = peer_abort,
[CPL_CLOSE_CON_RPL] = close_con_rpl,
[CPL_RDMA_TERMINATE] = terminate,
- [CPL_FW4_ACK] = fw4_ack
+ [CPL_FW4_ACK] = fw4_ack,
+ [CPL_FW6_MSG] = async_event
};
static void process_timeout(struct c4iw_ep *ep)
@@ -2185,7 +2178,7 @@ static void process_timeout(struct c4iw_ep *ep)
struct c4iw_qp_attributes attrs;
int abort = 1;
- spin_lock_irq(&ep->com.lock);
+ mutex_lock(&ep->com.mutex);
PDBG("%s ep %p tid %u state %d\n", __func__, ep, ep->hwtid,
ep->com.state);
switch (ep->com.state) {
@@ -2212,7 +2205,7 @@ static void process_timeout(struct c4iw_ep *ep)
WARN_ON(1);
abort = 0;
}
- spin_unlock_irq(&ep->com.lock);
+ mutex_unlock(&ep->com.mutex);
if (abort)
abort_connection(ep, NULL, GFP_KERNEL);
c4iw_put_ep(&ep->com);
@@ -2296,6 +2289,7 @@ static int set_tcb_rpl(struct c4iw_dev *dev, struct sk_buff *skb)
printk(KERN_ERR MOD "Unexpected SET_TCB_RPL status %u "
"for tid %u\n", rpl->status, GET_TID(rpl));
}
+ kfree_skb(skb);
return 0;
}
@@ -2313,17 +2307,22 @@ static int fw6_msg(struct c4iw_dev *dev, struct sk_buff *skb)
wr_waitp = (__force struct c4iw_wr_wait *)rpl->data[1];
PDBG("%s wr_waitp %p ret %u\n", __func__, wr_waitp, ret);
if (wr_waitp) {
- wr_waitp->ret = ret;
+ if (ret)
+ wr_waitp->ret = -ret;
+ else
+ wr_waitp->ret = 0;
wr_waitp->done = 1;
wake_up(&wr_waitp->wait);
}
+ kfree_skb(skb);
break;
case 2:
- c4iw_ev_dispatch(dev, (struct t4_cqe *)&rpl->data[0]);
+ sched(dev, skb);
break;
default:
printk(KERN_ERR MOD "%s unexpected fw6 msg type %u\n", __func__,
rpl->type);
+ kfree_skb(skb);
break;
}
return 0;
diff --git a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
index 1c26922..16032cd 100644
--- a/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
+++ b/drivers/infiniband/hw/cxgb4/iw_cxgb4.h
@@ -46,6 +46,7 @@
#include <linux/timer.h>
#include <linux/io.h>
#include <linux/kfifo.h>
+#include <linux/mutex.h>
#include <asm/byteorder.h>
@@ -353,6 +354,7 @@ struct c4iw_qp {
struct c4iw_qp_attributes attr;
struct t4_wq wq;
spinlock_t lock;
+ struct mutex mutex;
atomic_t refcnt;
wait_queue_head_t wait;
struct timer_list timer;
@@ -605,7 +607,7 @@ struct c4iw_ep_common {
struct c4iw_dev *dev;
enum c4iw_ep_state state;
struct kref kref;
- spinlock_t lock;
+ struct mutex mutex;
struct sockaddr_in local_addr;
struct sockaddr_in remote_addr;
struct c4iw_wr_wait wr_wait;
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index e0f433f..44ad386 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -35,6 +35,14 @@ static int ocqp_support;
module_param(ocqp_support, int, 0644);
MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=0)");
+static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
+{
+ unsigned long flag;
+ spin_lock_irqsave(&qhp->lock, flag);
+ qhp->attr.state = state;
+ spin_unlock_irqrestore(&qhp->lock, flag);
+}
+
static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
{
c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
@@ -949,46 +957,38 @@ static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
* Assumes qhp lock is held.
*/
static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
- struct c4iw_cq *schp, unsigned long *flag)
+ struct c4iw_cq *schp)
{
int count;
int flushed;
+ unsigned long flag;
PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
- /* take a ref on the qhp since we must release the lock */
- atomic_inc(&qhp->refcnt);
- spin_unlock_irqrestore(&qhp->lock, *flag);
/* locking hierarchy: cq lock first, then qp lock. */
- spin_lock_irqsave(&rchp->lock, *flag);
+ spin_lock_irqsave(&rchp->lock, flag);
spin_lock(&qhp->lock);
c4iw_flush_hw_cq(&rchp->cq);
c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
spin_unlock(&qhp->lock);
- spin_unlock_irqrestore(&rchp->lock, *flag);
+ spin_unlock_irqrestore(&rchp->lock, flag);
if (flushed)
(*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
/* locking hierarchy: cq lock first, then qp lock. */
- spin_lock_irqsave(&schp->lock, *flag);
+ spin_lock_irqsave(&schp->lock, flag);
spin_lock(&qhp->lock);
c4iw_flush_hw_cq(&schp->cq);
c4iw_count_scqes(&schp->cq, &qhp->wq, &count);
flushed = c4iw_flush_sq(&qhp->wq, &schp->cq, count);
spin_unlock(&qhp->lock);
- spin_unlock_irqrestore(&schp->lock, *flag);
+ spin_unlock_irqrestore(&schp->lock, flag);
if (flushed)
(*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
-
- /* deref */
- if (atomic_dec_and_test(&qhp->refcnt))
- wake_up(&qhp->wait);
-
- spin_lock_irqsave(&qhp->lock, *flag);
}
-static void flush_qp(struct c4iw_qp *qhp, unsigned long *flag)
+static void flush_qp(struct c4iw_qp *qhp)
{
struct c4iw_cq *rchp, *schp;
@@ -1002,7 +1002,7 @@ static void flush_qp(struct c4iw_qp *qhp, unsigned long *flag)
t4_set_cq_in_error(&schp->cq);
return;
}
- __flush_qp(qhp, rchp, schp, flag);
+ __flush_qp(qhp, rchp, schp);
}
static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
@@ -1010,7 +1010,6 @@ static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
{
struct fw_ri_wr *wqe;
int ret;
- struct c4iw_wr_wait wr_wait;
struct sk_buff *skb;
PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
@@ -1029,15 +1028,15 @@ static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
wqe->flowid_len16 = cpu_to_be32(
FW_WR_FLOWID(ep->hwtid) |
FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
- wqe->cookie = (u64)&wr_wait;
+ wqe->cookie = (u64)&ep->com.wr_wait;
wqe->u.fini.type = FW_RI_TYPE_FINI;
- c4iw_init_wr_wait(&wr_wait);
+ c4iw_init_wr_wait(&ep->com.wr_wait);
ret = c4iw_ofld_send(&rhp->rdev, skb);
if (ret)
goto out;
- ret = c4iw_wait_for_reply(&rhp->rdev, &wr_wait, qhp->ep->hwtid,
+ ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
qhp->wq.sq.qid, __func__);
out:
PDBG("%s ret %d\n", __func__, ret);
@@ -1072,7 +1071,6 @@ static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
{
struct fw_ri_wr *wqe;
int ret;
- struct c4iw_wr_wait wr_wait;
struct sk_buff *skb;
PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
@@ -1092,7 +1090,7 @@ static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
FW_WR_FLOWID(qhp->ep->hwtid) |
FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
- wqe->cookie = (u64)&wr_wait;
+ wqe->cookie = (u64)&qhp->ep->com.wr_wait;
wqe->u.init.type = FW_RI_TYPE_INIT;
wqe->u.init.mpareqbit_p2ptype =
@@ -1129,13 +1127,13 @@ static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
if (qhp->attr.mpa_attr.initiator)
build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
- c4iw_init_wr_wait(&wr_wait);
+ c4iw_init_wr_wait(&qhp->ep->com.wr_wait);
ret = c4iw_ofld_send(&rhp->rdev, skb);
if (ret)
goto out;
- ret = c4iw_wait_for_reply(&rhp->rdev, &wr_wait, qhp->ep->hwtid,
- qhp->wq.sq.qid, __func__);
+ ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
+ qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
out:
PDBG("%s ret %d\n", __func__, ret);
return ret;
@@ -1148,7 +1146,6 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
{
int ret = 0;
struct c4iw_qp_attributes newattr = qhp->attr;
- unsigned long flag;
int disconnect = 0;
int terminate = 0;
int abort = 0;
@@ -1159,7 +1156,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
(mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
- spin_lock_irqsave(&qhp->lock, flag);
+ mutex_lock(&qhp->mutex);
/* Process attr changes if in IDLE */
if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
@@ -1210,7 +1207,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
qhp->attr.mpa_attr = attrs->mpa_attr;
qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
qhp->ep = qhp->attr.llp_stream_handle;
- qhp->attr.state = C4IW_QP_STATE_RTS;
+ set_state(qhp, C4IW_QP_STATE_RTS);
/*
* Ref the endpoint here and deref when we
@@ -1219,15 +1216,13 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
* transition.
*/
c4iw_get_ep(&qhp->ep->com);
- spin_unlock_irqrestore(&qhp->lock, flag);
ret = rdma_init(rhp, qhp);
- spin_lock_irqsave(&qhp->lock, flag);
if (ret)
goto err;
break;
case C4IW_QP_STATE_ERROR:
- qhp->attr.state = C4IW_QP_STATE_ERROR;
- flush_qp(qhp, &flag);
+ set_state(qhp, C4IW_QP_STATE_ERROR);
+ flush_qp(qhp);
break;
default:
ret = -EINVAL;
@@ -1238,39 +1233,38 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
switch (attrs->next_state) {
case C4IW_QP_STATE_CLOSING:
BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
- qhp->attr.state = C4IW_QP_STATE_CLOSING;
+ set_state(qhp, C4IW_QP_STATE_CLOSING);
ep = qhp->ep;
if (!internal) {
abort = 0;
disconnect = 1;
- c4iw_get_ep(&ep->com);
+ c4iw_get_ep(&qhp->ep->com);
}
- spin_unlock_irqrestore(&qhp->lock, flag);
ret = rdma_fini(rhp, qhp, ep);
- spin_lock_irqsave(&qhp->lock, flag);
if (ret) {
- c4iw_get_ep(&ep->com);
+ if (internal)
+ c4iw_get_ep(&qhp->ep->com);
disconnect = abort = 1;
goto err;
}
break;
case C4IW_QP_STATE_TERMINATE:
- qhp->attr.state = C4IW_QP_STATE_TERMINATE;
+ set_state(qhp, C4IW_QP_STATE_TERMINATE);
if (qhp->ibqp.uobject)
t4_set_wq_in_error(&qhp->wq);
ep = qhp->ep;
- c4iw_get_ep(&ep->com);
if (!internal)
terminate = 1;
disconnect = 1;
+ c4iw_get_ep(&qhp->ep->com);
break;
case C4IW_QP_STATE_ERROR:
- qhp->attr.state = C4IW_QP_STATE_ERROR;
+ set_state(qhp, C4IW_QP_STATE_ERROR);
if (!internal) {
abort = 1;
disconnect = 1;
ep = qhp->ep;
- c4iw_get_ep(&ep->com);
+ c4iw_get_ep(&qhp->ep->com);
}
goto err;
break;
@@ -1286,8 +1280,8 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
}
switch (attrs->next_state) {
case C4IW_QP_STATE_IDLE:
- flush_qp(qhp, &flag);
- qhp->attr.state = C4IW_QP_STATE_IDLE;
+ flush_qp(qhp);
+ set_state(qhp, C4IW_QP_STATE_IDLE);
qhp->attr.llp_stream_handle = NULL;
c4iw_put_ep(&qhp->ep->com);
qhp->ep = NULL;
@@ -1309,7 +1303,7 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
ret = -EINVAL;
goto out;
}
- qhp->attr.state = C4IW_QP_STATE_IDLE;
+ set_state(qhp, C4IW_QP_STATE_IDLE);
break;
case C4IW_QP_STATE_TERMINATE:
if (!internal) {
@@ -1335,13 +1329,13 @@ err:
if (!ep)
ep = qhp->ep;
qhp->ep = NULL;
- qhp->attr.state = C4IW_QP_STATE_ERROR;
+ set_state(qhp, C4IW_QP_STATE_ERROR);
free = 1;
wake_up(&qhp->wait);
BUG_ON(!ep);
- flush_qp(qhp, &flag);
+ flush_qp(qhp);
out:
- spin_unlock_irqrestore(&qhp->lock, flag);
+ mutex_unlock(&qhp->mutex);
if (terminate)
post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
@@ -1363,7 +1357,6 @@ out:
*/
if (free)
c4iw_put_ep(&ep->com);
-
PDBG("%s exit state %d\n", __func__, qhp->attr.state);
return ret;
}
@@ -1478,6 +1471,7 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
qhp->attr.max_ord = 1;
qhp->attr.max_ird = 1;
spin_lock_init(&qhp->lock);
+ mutex_init(&qhp->mutex);
init_waitqueue_head(&qhp->wait);
atomic_set(&qhp->refcnt, 1);
--
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] 15+ messages in thread
* [PATCH 2.6.37 11/11] RDMA/cxgb4: Set the default TCP send window to 128KB.
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
` (9 preceding siblings ...)
2010-09-10 16:15 ` [PATCH 2.6.37 10/11] RDMA/cxgb4: Use a mutex for QP and EP state transitions Steve Wise
@ 2010-09-10 16:15 ` Steve Wise
10 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-10 16:15 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
This helps with large IO throughput.
Signed-off-by: Steve Wise <swise-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
---
drivers/infiniband/hw/cxgb4/cm.c | 4 ++--
1 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb4/cm.c b/drivers/infiniband/hw/cxgb4/cm.c
index 3e6234c..89c6762 100644
--- a/drivers/infiniband/hw/cxgb4/cm.c
+++ b/drivers/infiniband/hw/cxgb4/cm.c
@@ -117,9 +117,9 @@ static int rcv_win = 256 * 1024;
module_param(rcv_win, int, 0644);
MODULE_PARM_DESC(rcv_win, "TCP receive window in bytes (default=256KB)");
-static int snd_win = 32 * 1024;
+static int snd_win = 128 * 1024;
module_param(snd_win, int, 0644);
-MODULE_PARM_DESC(snd_win, "TCP send window in bytes (default=32KB)");
+MODULE_PARM_DESC(snd_win, "TCP send window in bytes (default=128KB)");
static struct workqueue_struct *workq;
--
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] 15+ messages in thread
* Re: [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs.
[not found] ` <20100910161530.6829.89294.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
@ 2010-09-10 19:57 ` Steve Wise
[not found] ` <4C8A8DC1.4040302-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
0 siblings, 1 reply; 15+ messages in thread
From: Steve Wise @ 2010-09-10 19:57 UTC (permalink / raw)
To: linux-rdma-u79uwXL29TY76Z2rM5mHXA
On 09/10/2010 11:15 AM, Steve Wise wrote:
> T4 support on-chip SQs to reduce latency. This patch adds
> support for this in iw_cxgb4.
>
> Changes:
>
> 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-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
> ---
>
> drivers/infiniband/hw/cxgb4/device.c | 19 ++++++
> drivers/infiniband/hw/cxgb4/iw_cxgb4.h | 7 ++
> drivers/infiniband/hw/cxgb4/provider.c | 28 ++++++---
> drivers/infiniband/hw/cxgb4/qp.c | 98 +++++++++++++++++++++++++++-----
> drivers/infiniband/hw/cxgb4/resource.c | 56 ++++++++++++++++++
> drivers/infiniband/hw/cxgb4/t4.h | 40 +++++++++++--
> drivers/infiniband/hw/cxgb4/user.h | 7 ++
> 7 files changed, 226 insertions(+), 29 deletions(-)
>
> 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 @@ static int c4iw_rdev_open(struct c4iw_rdev *rdev)
> 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 @@ static void c4iw_remove(struct c4iw_dev *dev)
> 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 @@ static struct c4iw_dev *c4iw_alloc(const struct cxgb4_lld_info *infop)
> }
> 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_rdev {
> 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_resource(struct c4iw_rdev *rdev, u32 nr_tpt, u32 nr_pdid);
> 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 @@ u32 c4iw_rqtpool_alloc(struct c4iw_rdev *rdev, int size);
> 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 @@ static int c4iw_mmap(struct ib_ucontext *context, struct vm_area_struct *vma)
> 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 @@ int c4iw_register_device(struct c4iw_dev *dev)
> 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 ee785e2..e0f433f 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 @@ static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
> 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 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
> 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 @@ static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
> 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 @@ err7:
> 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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
> 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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
> 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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
> 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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
> 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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
> 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 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
> __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 @@ void c4iw_rqtpool_destroy(struct c4iw_rdev *rdev)
> {
> 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..51a845f 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 @@ struct t4_swsqe {
> 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 @@ struct t4_sq {
> u16 cidx;
> u16 pidx;
> u16 wq_pidx;
> + u16 flags;
> };
>
> struct t4_swrqe {
> @@ -350,6 +378,11 @@ static inline void t4_rq_consume(struct t4_wq *wq)
> 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 void t4_ring_rq_db(struct t4_wq *wq, u16 inc)
>
> 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->sq.size].status.qp_err;
>
Oops, cought this during regression testing: The above line should be
indexing by wq->rq.size, not wq->sq.size. This error caused
intermittent post failures and I missed it on my first round of testing.
> }
>
> 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->sq.size].status.db_off;
>
Same issue here.
> }
>
> 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 @@ struct c4iw_create_cq_resp {
> __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 @@ struct c4iw_create_qp_resp {
> __u32 sq_size;
> __u32 rq_size;
> __u32 qid_mask;
> + __u32 flags;
> };
> #endif
>
> --
> 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
>
--
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 [flat|nested] 15+ messages in thread
* Re: [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs.
[not found] ` <4C8A8DC1.4040302-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
@ 2010-09-13 16:13 ` Roland Dreier
[not found] ` <adaiq29wrjm.fsf-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org>
0 siblings, 1 reply; 15+ messages in thread
From: Roland Dreier @ 2010-09-13 16:13 UTC (permalink / raw)
To: Steve Wise; +Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA
> Oops, cought this during regression testing: The above line should be
> indexing by wq->rq.size, not wq->sq.size. This error caused
> intermittent post failures and I missed it on my first round of
> testing.
Steve, can you send a fixed version of this patch?
Thanks.
--
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 [flat|nested] 15+ messages in thread
* Re: [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs.
[not found] ` <adaiq29wrjm.fsf-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org>
@ 2010-09-13 16:25 ` Steve Wise
0 siblings, 0 replies; 15+ messages in thread
From: Steve Wise @ 2010-09-13 16:25 UTC (permalink / raw)
To: Roland Dreier; +Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA
I just emailed it out...
Steve.
On 9/13/2010 11:13 AM, Roland Dreier wrote:
> > Oops, cought this during regression testing: The above line should be
> > indexing by wq->rq.size, not wq->sq.size. This error caused
> > intermittent post failures and I missed it on my first round of
> > testing.
>
> Steve, can you send a fixed version of this patch?
>
> Thanks.
> --
> 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
--
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 [flat|nested] 15+ messages in thread
end of thread, other threads:[~2010-09-13 16:25 UTC | newest]
Thread overview: 15+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2010-09-10 16:14 [PATCH 2.6.37 00/11] cxgb4 fixes / enhancements Steve Wise
[not found] ` <20100910161442.6829.91594.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
2010-09-10 16:14 ` [PATCH 2.6.37 01/11] RDMA/cxgb4: Don't use null ep ptr Steve Wise
2010-09-10 16:14 ` [PATCH 2.6.37 02/11] RDMA/cxgb4: Zero out ISGL padding Steve Wise
2010-09-10 16:14 ` [PATCH 2.6.37 03/11] RDMA/cxgb4: Ignore positive return values from cxgb4_*_send() functions Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 04/11] RDMA/cxgb4: Ignore TERMINATE CQEs Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 05/11] RDMA/cxgb4: Handle CPL_RDMA_TERMINATE messages Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 06/11] RDMA/cxgb4: log HW lack-of-resource errors Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 07/11] RDMA/cxgb4: debugfs files for dumping active stags Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 08/11] RDMA/cxgb4: Centralize the wait logic Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 09/11] RDMA/cxgb4: Support on-chip SQs Steve Wise
[not found] ` <20100910161530.6829.89294.stgit-T4OLL4TyM9aNDNWfRnPdfg@public.gmane.org>
2010-09-10 19:57 ` Steve Wise
[not found] ` <4C8A8DC1.4040302-7bPotxP6k4+P2YhJcF5u+vpXobYPEAuW@public.gmane.org>
2010-09-13 16:13 ` Roland Dreier
[not found] ` <adaiq29wrjm.fsf-FYB4Gu1CFyUAvxtiuMwx3w@public.gmane.org>
2010-09-13 16:25 ` Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 10/11] RDMA/cxgb4: Use a mutex for QP and EP state transitions Steve Wise
2010-09-10 16:15 ` [PATCH 2.6.37 11/11] RDMA/cxgb4: Set the default TCP send window to 128KB Steve Wise
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.