| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <20100127170333.19854.40039.stgit@build.ogc.int>
Date: Wed, 27 Jan 2010 11:03:34 -0600
From: Steve Wise <swise@...ngridcomputing.com>
To: rdreier@...co.com
Cc: jeff@...zik.org, linux-rdma@...r.kernel.org,
netdev@...r.kernel.org, divy@...lsio.com
Subject: [PATCH] cxgb3/iw_cxgb3: doorbell overflow avoidance and recovery.
T3 hardware doorbell fifo overflows can cause application stalls due to
lost doorbell ring events. This has been seen when running large NP IMB
alltoall MPI jobs. The T3 hardware supports an xon/xoff type control
mechanism to help avoid overflowing the hw db fifo.
This patch utilizes these interrupts to disable rdma qp db rings when
we near an overflow condition, and then turn them back on (and ring all
the active qp dbs) when when the db fifo emptys out. In addition if an
db ring is dropped by the hardware, the code will now recover.
Design:
cxgb3:
- enable these DB interrupts
- in the intr handler, schedule work tasks to call the ULPs event handlers
with the new events.
- ring all the qset txqs when an overflow is detected.
iw_cxgb3:
- disable db ringing on all active qps when we get the DB_FULL event
- enable db ringing on all active qps and ring all active dbs when we get
the DB_EMPTY event
- On DB_DROP event:
- disable db rings in the event handler
- delay-schedule a work task which rings and enables the dbs on
all active qps.
- in post_send and post_recv logic, don't ring the db if its disabled.
Signed-off-by: Steve Wise <swise@...ngridcomputing.com>
---
drivers/infiniband/hw/cxgb3/cxio_wr.h | 17 +++++++
drivers/infiniband/hw/cxgb3/iwch.c | 79 +++++++++++++++++++++++++++++++--
drivers/infiniband/hw/cxgb3/iwch.h | 2 +
drivers/infiniband/hw/cxgb3/iwch_qp.c | 9 +++-
drivers/net/cxgb3/adapter.h | 5 ++
drivers/net/cxgb3/cxgb3_main.c | 57 +++++++++++++++++++++++-
drivers/net/cxgb3/cxgb3_offload.h | 5 ++
drivers/net/cxgb3/regs.h | 16 +++++++
drivers/net/cxgb3/sge.c | 15 +++++-
drivers/net/cxgb3/t3_hw.c | 5 ++
10 files changed, 196 insertions(+), 14 deletions(-)
diff --git a/drivers/infiniband/hw/cxgb3/cxio_wr.h b/drivers/infiniband/hw/cxgb3/cxio_wr.h
index a197a5b..15073b2 100644
--- a/drivers/infiniband/hw/cxgb3/cxio_wr.h
+++ b/drivers/infiniband/hw/cxgb3/cxio_wr.h
@@ -730,7 +730,22 @@ struct t3_cq {
static inline void cxio_set_wq_in_error(struct t3_wq *wq)
{
- wq->queue->wq_in_err.err = 1;
+ wq->queue->wq_in_err.err |= 1;
+}
+
+static inline void cxio_disable_wq_db(struct t3_wq *wq)
+{
+ wq->queue->wq_in_err.err |= 2;
+}
+
+static inline void cxio_enable_wq_db(struct t3_wq *wq)
+{
+ wq->queue->wq_in_err.err &= ~2;
+}
+
+static inline int cxio_wq_db_enabled(struct t3_wq *wq)
+{
+ return !(wq->queue->wq_in_err.err & 2);
}
static inline struct t3_cqe *cxio_next_hw_cqe(struct t3_cq *cq)
diff --git a/drivers/infiniband/hw/cxgb3/iwch.c b/drivers/infiniband/hw/cxgb3/iwch.c
index b0ea010..6a3214a 100644
--- a/drivers/infiniband/hw/cxgb3/iwch.c
+++ b/drivers/infiniband/hw/cxgb3/iwch.c
@@ -65,6 +65,46 @@ struct cxgb3_client t3c_client = {
static LIST_HEAD(dev_list);
static DEFINE_MUTEX(dev_mutex);
+static int disable_qp_db(int id, void *p, void *data)
+{
+ struct iwch_qp *qhp = p;
+
+ cxio_disable_wq_db(&qhp->wq);
+ return 0;
+}
+
+static int enable_qp_db(int id, void *p, void *data)
+{
+ struct iwch_qp *qhp = p;
+
+ if (data)
+ ring_doorbell(qhp->rhp->rdev.ctrl_qp.doorbell, qhp->wq.qpid);
+ cxio_enable_wq_db(&qhp->wq);
+ return 0;
+}
+
+static void disable_dbs(struct iwch_dev *rnicp)
+{
+ spin_lock_irq(&rnicp->lock);
+ idr_for_each(&rnicp->qpidr, disable_qp_db, NULL);
+ spin_unlock_irq(&rnicp->lock);
+}
+
+static void enable_dbs(struct iwch_dev *rnicp, int ring_db)
+{
+ spin_lock_irq(&rnicp->lock);
+ idr_for_each(&rnicp->qpidr, enable_qp_db,
+ (void *)(unsigned long)ring_db);
+ spin_unlock_irq(&rnicp->lock);
+}
+
+static void iwch_db_drop_task(struct work_struct *work)
+{
+ struct iwch_dev *rnicp = container_of(work, struct iwch_dev,
+ db_drop_task.work);
+ enable_dbs(rnicp, 1);
+}
+
static void rnic_init(struct iwch_dev *rnicp)
{
PDBG("%s iwch_dev %p\n", __func__, rnicp);
@@ -72,6 +112,7 @@ static void rnic_init(struct iwch_dev *rnicp)
idr_init(&rnicp->qpidr);
idr_init(&rnicp->mmidr);
spin_lock_init(&rnicp->lock);
+ INIT_DELAYED_WORK(&rnicp->db_drop_task, iwch_db_drop_task);
rnicp->attr.max_qps = T3_MAX_NUM_QP - 32;
rnicp->attr.max_wrs = T3_MAX_QP_DEPTH;
@@ -147,6 +188,7 @@ static void close_rnic_dev(struct t3cdev *tdev)
mutex_lock(&dev_mutex);
list_for_each_entry_safe(dev, tmp, &dev_list, entry) {
if (dev->rdev.t3cdev_p == tdev) {
+ cancel_delayed_work_sync(&dev->db_drop_task);
list_del(&dev->entry);
iwch_unregister_device(dev);
cxio_rdev_close(&dev->rdev);
@@ -165,7 +207,8 @@ static void iwch_event_handler(struct t3cdev *tdev, u32 evt, u32 port_id)
struct cxio_rdev *rdev = tdev->ulp;
struct iwch_dev *rnicp;
struct ib_event event;
- u32 portnum = port_id + 1;
+ u32 portnum = port_id + 1;
+ int dispatch = 0;
if (!rdev)
return;
@@ -174,21 +217,49 @@ static void iwch_event_handler(struct t3cdev *tdev, u32 evt, u32 port_id)
case OFFLOAD_STATUS_DOWN: {
rdev->flags = CXIO_ERROR_FATAL;
event.event = IB_EVENT_DEVICE_FATAL;
+ dispatch = 1;
break;
}
case OFFLOAD_PORT_DOWN: {
event.event = IB_EVENT_PORT_ERR;
+ dispatch = 1;
break;
}
case OFFLOAD_PORT_UP: {
event.event = IB_EVENT_PORT_ACTIVE;
+ dispatch = 1;
+ break;
+ }
+ case OFFLOAD_DB_FULL: {
+ disable_dbs(rnicp);
+ break;
+ }
+ case OFFLOAD_DB_EMPTY: {
+ enable_dbs(rnicp, 1);
+ break;
+ }
+ case OFFLOAD_DB_DROP: {
+ unsigned long delay = 1000;
+ unsigned short r;
+
+ disable_dbs(rnicp);
+ get_random_bytes(&r, 2);
+ delay += r & 1023;
+
+ /*
+ * delay is between 1000-2023 usecs.
+ */
+ schedule_delayed_work(&rnicp->db_drop_task,
+ usecs_to_jiffies(delay));
break;
}
}
- event.device = &rnicp->ibdev;
- event.element.port_num = portnum;
- ib_dispatch_event(&event);
+ if (dispatch) {
+ event.device = &rnicp->ibdev;
+ event.element.port_num = portnum;
+ ib_dispatch_event(&event);
+ }
return;
}
diff --git a/drivers/infiniband/hw/cxgb3/iwch.h b/drivers/infiniband/hw/cxgb3/iwch.h
index 8473550..a1c4457 100644
--- a/drivers/infiniband/hw/cxgb3/iwch.h
+++ b/drivers/infiniband/hw/cxgb3/iwch.h
@@ -36,6 +36,7 @@
#include <linux/list.h>
#include <linux/spinlock.h>
#include <linux/idr.h>
+#include <linux/workqueue.h>
#include <rdma/ib_verbs.h>
@@ -110,6 +111,7 @@ struct iwch_dev {
struct idr mmidr;
spinlock_t lock;
struct list_head entry;
+ struct delayed_work db_drop_task;
};
static inline struct iwch_dev *to_iwch_dev(struct ib_device *ibdev)
diff --git a/drivers/infiniband/hw/cxgb3/iwch_qp.c b/drivers/infiniband/hw/cxgb3/iwch_qp.c
index 3eb8cec..b4d893d 100644
--- a/drivers/infiniband/hw/cxgb3/iwch_qp.c
+++ b/drivers/infiniband/hw/cxgb3/iwch_qp.c
@@ -452,7 +452,8 @@ int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
++(qhp->wq.sq_wptr);
}
spin_unlock_irqrestore(&qhp->lock, flag);
- ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
+ if (cxio_wq_db_enabled(&qhp->wq))
+ ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
out:
if (err)
@@ -514,7 +515,8 @@ int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
num_wrs--;
}
spin_unlock_irqrestore(&qhp->lock, flag);
- ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
+ if (cxio_wq_db_enabled(&qhp->wq))
+ ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
out:
if (err)
@@ -597,7 +599,8 @@ int iwch_bind_mw(struct ib_qp *qp,
++(qhp->wq.sq_wptr);
spin_unlock_irqrestore(&qhp->lock, flag);
- ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
+ if (cxio_wq_db_enabled(&qhp->wq))
+ ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
return err;
}
diff --git a/drivers/net/cxgb3/adapter.h b/drivers/net/cxgb3/adapter.h
index 3e8618b..11d148c 100644
--- a/drivers/net/cxgb3/adapter.h
+++ b/drivers/net/cxgb3/adapter.h
@@ -263,6 +263,10 @@ struct adapter {
struct work_struct ext_intr_handler_task;
struct work_struct fatal_error_handler_task;
struct work_struct link_fault_handler_task;
+
+ struct work_struct db_full_task;
+ struct work_struct db_empty_task;
+ struct work_struct db_drop_task;
struct dentry *debugfs_root;
@@ -335,6 +339,7 @@ int t3_sge_alloc_qset(struct adapter *adapter, unsigned int id, int nports,
int t3_get_desc(const struct sge_qset *qs, unsigned int qnum, unsigned int idx,
unsigned char *data);
irqreturn_t t3_sge_intr_msix(int irq, void *cookie);
+extern struct workqueue_struct *cxgb3_wq;
int t3_get_edc_fw(struct cphy *phy, int edc_idx, int size);
diff --git a/drivers/net/cxgb3/cxgb3_main.c b/drivers/net/cxgb3/cxgb3_main.c
index 2aaa747..f47dc76 100644
--- a/drivers/net/cxgb3/cxgb3_main.c
+++ b/drivers/net/cxgb3/cxgb3_main.c
@@ -45,6 +45,7 @@
#include <linux/firmware.h>
#include <linux/log2.h>
#include <linux/stringify.h>
+#include <linux/sched.h>
#include <asm/uaccess.h>
#include "common.h"
@@ -140,7 +141,7 @@ MODULE_PARM_DESC(ofld_disable, "whether to enable offload at init time or not");
* will block keventd as it needs the rtnl lock, and we'll deadlock waiting
* for our work to complete. Get our own work queue to solve this.
*/
-static struct workqueue_struct *cxgb3_wq;
+struct workqueue_struct *cxgb3_wq;
/**
* link_report - show link status and link speed/duplex
@@ -590,6 +591,19 @@ static void setup_rss(struct adapter *adap)
V_RRCPLCPUSIZE(6) | F_HASHTOEPLITZ, cpus, rspq_map);
}
+static void ring_dbs(struct adapter *adap)
+{
+ int i, j;
+
+ for (i = 0; i < SGE_QSETS; i++) {
+ struct sge_qset *qs = &adap->sge.qs[i];
+
+ if (qs->adap)
+ for (j=0; j < SGE_TXQ_PER_SET; j++)
+ t3_write_reg(adap, A_SG_KDOORBELL, F_SELEGRCNTX | V_EGRCNTX(qs->txq[j].cntxt_id));
+ }
+}
+
static void init_napi(struct adapter *adap)
{
int i;
@@ -2778,6 +2792,42 @@ static void t3_adap_check_task(struct work_struct *work)
spin_unlock_irq(&adapter->work_lock);
}
+static void db_full_task(struct work_struct *work)
+{
+ struct adapter *adapter = container_of(work, struct adapter,
+ db_full_task);
+
+ cxgb3_event_notify(&adapter->tdev, OFFLOAD_DB_FULL, 0);
+}
+
+static void db_empty_task(struct work_struct *work)
+{
+ struct adapter *adapter = container_of(work, struct adapter,
+ db_empty_task);
+
+ cxgb3_event_notify(&adapter->tdev, OFFLOAD_DB_EMPTY, 0);
+}
+
+static void db_drop_task(struct work_struct *work)
+{
+ struct adapter *adapter = container_of(work, struct adapter,
+ db_drop_task);
+ unsigned long delay = 1000;
+ unsigned short r;
+
+ cxgb3_event_notify(&adapter->tdev, OFFLOAD_DB_DROP, 0);
+
+ /*
+ * Sleep a while before ringing the driver qset dbs.
+ * The delay is between 1000-2023 usecs.
+ */
+ get_random_bytes(&r, 2);
+ delay += r & 1023;
+ set_current_state(TASK_UNINTERRUPTIBLE);
+ schedule_timeout(usecs_to_jiffies(delay));
+ ring_dbs(adapter);
+}
+
/*
* Processes external (PHY) interrupts in process context.
*/
@@ -3246,6 +3296,11 @@ static int __devinit init_one(struct pci_dev *pdev,
INIT_LIST_HEAD(&adapter->adapter_list);
INIT_WORK(&adapter->ext_intr_handler_task, ext_intr_task);
INIT_WORK(&adapter->fatal_error_handler_task, fatal_error_task);
+
+ INIT_WORK(&adapter->db_full_task, db_full_task);
+ INIT_WORK(&adapter->db_empty_task, db_empty_task);
+ INIT_WORK(&adapter->db_drop_task, db_drop_task);
+
INIT_DELAYED_WORK(&adapter->adap_check_task, t3_adap_check_task);
for (i = 0; i < ai->nports0 + ai->nports1; ++i) {
diff --git a/drivers/net/cxgb3/cxgb3_offload.h b/drivers/net/cxgb3/cxgb3_offload.h
index 670aa62..929c298 100644
--- a/drivers/net/cxgb3/cxgb3_offload.h
+++ b/drivers/net/cxgb3/cxgb3_offload.h
@@ -73,7 +73,10 @@ enum {
OFFLOAD_STATUS_UP,
OFFLOAD_STATUS_DOWN,
OFFLOAD_PORT_DOWN,
- OFFLOAD_PORT_UP
+ OFFLOAD_PORT_UP,
+ OFFLOAD_DB_FULL,
+ OFFLOAD_DB_EMPTY,
+ OFFLOAD_DB_DROP
};
struct cxgb3_client {
diff --git a/drivers/net/cxgb3/regs.h b/drivers/net/cxgb3/regs.h
index 1b5327b..cb42353 100644
--- a/drivers/net/cxgb3/regs.h
+++ b/drivers/net/cxgb3/regs.h
@@ -254,6 +254,22 @@
#define V_LOPIODRBDROPERR(x) ((x) << S_LOPIODRBDROPERR)
#define F_LOPIODRBDROPERR V_LOPIODRBDROPERR(1U)
+#define S_HIPRIORITYDBFULL 7
+#define V_HIPRIORITYDBFULL(x) ((x) << S_HIPRIORITYDBFULL)
+#define F_HIPRIORITYDBFULL V_HIPRIORITYDBFULL(1U)
+
+#define S_HIPRIORITYDBEMPTY 6
+#define V_HIPRIORITYDBEMPTY(x) ((x) << S_HIPRIORITYDBEMPTY)
+#define F_HIPRIORITYDBEMPTY V_HIPRIORITYDBEMPTY(1U)
+
+#define S_LOPRIORITYDBFULL 5
+#define V_LOPRIORITYDBFULL(x) ((x) << S_LOPRIORITYDBFULL)
+#define F_LOPRIORITYDBFULL V_LOPRIORITYDBFULL(1U)
+
+#define S_LOPRIORITYDBEMPTY 4
+#define V_LOPRIORITYDBEMPTY(x) ((x) << S_LOPRIORITYDBEMPTY)
+#define F_LOPRIORITYDBEMPTY V_LOPRIORITYDBEMPTY(1U)
+
#define S_RSPQDISABLED 3
#define V_RSPQDISABLED(x) ((x) << S_RSPQDISABLED)
#define F_RSPQDISABLED V_RSPQDISABLED(1U)
diff --git a/drivers/net/cxgb3/sge.c b/drivers/net/cxgb3/sge.c
index bdbd147..08978ad 100644
--- a/drivers/net/cxgb3/sge.c
+++ b/drivers/net/cxgb3/sge.c
@@ -42,6 +42,7 @@
#include "sge_defs.h"
#include "t3_cpl.h"
#include "firmware_exports.h"
+#include "cxgb3_offload.h"
#define USE_GTS 0
@@ -2828,9 +2829,17 @@ void t3_sge_err_intr_handler(struct adapter *adapter)
"(0x%x)\n", (v >> S_RSPQ0DISABLED) & 0xff);
}
- if (status & (F_HIPIODRBDROPERR | F_LOPIODRBDROPERR))
- CH_ALERT(adapter, "SGE dropped %s priority doorbell\n",
- status & F_HIPIODRBDROPERR ? "high" : "lo");
+ if (status & (F_HIPIODRBDROPERR | F_LOPIODRBDROPERR)) {
+ queue_work(cxgb3_wq, &adapter->db_drop_task);
+ }
+
+ if (status & (F_HIPRIORITYDBFULL | F_LOPRIORITYDBFULL)) {
+ queue_work(cxgb3_wq, &adapter->db_full_task);
+ }
+
+ if (status & (F_HIPRIORITYDBEMPTY | F_LOPRIORITYDBEMPTY)) {
+ queue_work(cxgb3_wq, &adapter->db_empty_task);
+ }
t3_write_reg(adapter, A_SG_INT_CAUSE, status);
if (status & SGE_FATALERR)
diff --git a/drivers/net/cxgb3/t3_hw.c b/drivers/net/cxgb3/t3_hw.c
index 032cfe0..c38fc71 100644
--- a/drivers/net/cxgb3/t3_hw.c
+++ b/drivers/net/cxgb3/t3_hw.c
@@ -1432,7 +1432,10 @@ static int t3_handle_intr_status(struct adapter *adapter, unsigned int reg,
F_IRPARITYERROR | V_ITPARITYERROR(M_ITPARITYERROR) | \
V_FLPARITYERROR(M_FLPARITYERROR) | F_LODRBPARITYERROR | \
F_HIDRBPARITYERROR | F_LORCQPARITYERROR | \
- F_HIRCQPARITYERROR)
+ F_HIRCQPARITYERROR | F_LOPRIORITYDBFULL | \
+ F_HIPRIORITYDBFULL | F_LOPRIORITYDBEMPTY | \
+ F_HIPRIORITYDBEMPTY | F_HIPIODRBDROPERR | \
+ F_LOPIODRBDROPERR)
#define MC5_INTR_MASK (F_PARITYERR | F_ACTRGNFULL | F_UNKNOWNCMD | \
F_REQQPARERR | F_DISPQPARERR | F_DELACTEMPTY | \
F_NFASRCHFAIL)
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists