From: Tadeusz Struk <tadeusz.struk@xxxxxxxxx>
Soft lockups can occur because the mad processing on different CPUs acquire
the spin lock dc8051_lock:
[534552.835870] [<ffffffffa026f993>] ? read_dev_port_cntr.isra.37+0x23/0x160 [hfi1]
[534552.835880] [<ffffffffa02775af>] read_dev_cntr+0x4f/0x60 [hfi1]
[534552.835893] [<ffffffffa028d7cd>] pma_get_opa_portstatus+0x64d/0x8c0 [hfi1]
[534552.835904] [<ffffffffa0290e7d>] hfi1_process_mad+0x48d/0x18c0 [hfi1]
[534552.835908] [<ffffffff811dc1f1>] ? __slab_free+0x81/0x2f0
[534552.835936] [<ffffffffa024c34e>] ? ib_mad_recv_done+0x21e/0xa30 [ib_core]
[534552.835939] [<ffffffff811dd153>] ? __kmalloc+0x1f3/0x240
[534552.835947] [<ffffffffa024c3fb>] ib_mad_recv_done+0x2cb/0xa30 [ib_core]
[534552.835955] [<ffffffffa0237c85>] __ib_process_cq+0x55/0xd0 [ib_core]
[534552.835962] [<ffffffffa0237d70>] ib_cq_poll_work+0x20/0x60 [ib_core]
[534552.835964] [<ffffffff810a7f3b>] process_one_work+0x17b/0x470
[534552.835966] [<ffffffff810a8d76>] worker_thread+0x126/0x410
[534552.835969] [<ffffffff810a8c50>] ? rescuer_thread+0x460/0x460
[534552.835971] [<ffffffff810b052f>] kthread+0xcf/0xe0
[534552.835974] [<ffffffff810b0460>] ? kthread_create_on_node+0x140/0x140
[534552.835977] [<ffffffff81696418>] ret_from_fork+0x58/0x90
[534552.835980] [<ffffffff810b0460>] ? kthread_create_on_node+0x140/0x140
This issue is made worse when the 8051 is busy and the reads take longer.
Fix by using a non-spinning lock procure.
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@xxxxxxxxx>
Reviewed-by: Mike Marciszyn <mike.marciniszyn@xxxxxxxxx>
Signed-off-by: Tadeusz Struk <tadeusz.struk@xxxxxxxxx>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxx>
---
drivers/infiniband/hw/hfi1/chip.c | 86 ++++++++++++++++++++++---------------
drivers/infiniband/hw/hfi1/hfi.h | 7 ++-
drivers/infiniband/hw/hfi1/init.c | 2 -
3 files changed, 57 insertions(+), 38 deletions(-)
diff --git a/drivers/infiniband/hw/hfi1/chip.c b/drivers/infiniband/hw/hfi1/chip.c
index 79a316a..6b70d1d 100644
--- a/drivers/infiniband/hw/hfi1/chip.c
+++ b/drivers/infiniband/hw/hfi1/chip.c
@@ -6381,18 +6381,17 @@ static void lcb_shutdown(struct hfi1_devdata *dd, int abort)
*
* The expectation is that the caller of this routine would have taken
* care of properly transitioning the link into the correct state.
+ * NOTE: the caller needs to acquire the dd->dc8051_lock lock
+ * before calling this function.
*/
-static void dc_shutdown(struct hfi1_devdata *dd)
+static void _dc_shutdown(struct hfi1_devdata *dd)
{
- unsigned long flags;
+ lockdep_assert_held(&dd->dc8051_lock);
- spin_lock_irqsave(&dd->dc8051_lock, flags);
- if (dd->dc_shutdown) {
- spin_unlock_irqrestore(&dd->dc8051_lock, flags);
+ if (dd->dc_shutdown)
return;
- }
+
dd->dc_shutdown = 1;
- spin_unlock_irqrestore(&dd->dc8051_lock, flags);
/* Shutdown the LCB */
lcb_shutdown(dd, 1);
/*
@@ -6403,35 +6402,45 @@ static void dc_shutdown(struct hfi1_devdata *dd)
write_csr(dd, DC_DC8051_CFG_RST, 0x1);
}
+static void dc_shutdown(struct hfi1_devdata *dd)
+{
+ mutex_lock(&dd->dc8051_lock);
+ _dc_shutdown(dd);
+ mutex_unlock(&dd->dc8051_lock);
+}
+
/*
* Calling this after the DC has been brought out of reset should not
* do any damage.
+ * NOTE: the caller needs to acquire the dd->dc8051_lock lock
+ * before calling this function.
*/
-static void dc_start(struct hfi1_devdata *dd)
+static void _dc_start(struct hfi1_devdata *dd)
{
- unsigned long flags;
- int ret;
+ lockdep_assert_held(&dd->dc8051_lock);
- spin_lock_irqsave(&dd->dc8051_lock, flags);
if (!dd->dc_shutdown)
- goto done;
- spin_unlock_irqrestore(&dd->dc8051_lock, flags);
+ return;
+
/* Take the 8051 out of reset */
write_csr(dd, DC_DC8051_CFG_RST, 0ull);
/* Wait until 8051 is ready */
- ret = wait_fm_ready(dd, TIMEOUT_8051_START);
- if (ret) {
+ if (wait_fm_ready(dd, TIMEOUT_8051_START))
dd_dev_err(dd, "%s: timeout starting 8051 firmware\n",
__func__);
- }
+
/* Take away reset for LCB and RX FPE (set in lcb_shutdown). */
write_csr(dd, DCC_CFG_RESET, 0x10);
/* lcb_shutdown() with abort=1 does not restore these */
write_csr(dd, DC_LCB_ERR_EN, dd->lcb_err_en);
- spin_lock_irqsave(&dd->dc8051_lock, flags);
dd->dc_shutdown = 0;
-done:
- spin_unlock_irqrestore(&dd->dc8051_lock, flags);
+}
+
+static void dc_start(struct hfi1_devdata *dd)
+{
+ mutex_lock(&dd->dc8051_lock);
+ _dc_start(dd);
+ mutex_unlock(&dd->dc8051_lock);
}
/*
@@ -8475,16 +8484,11 @@ static int do_8051_command(
{
u64 reg, completed;
int return_code;
- unsigned long flags;
unsigned long timeout;
hfi1_cdbg(DC8051, "type %d, data 0x%012llx", type, in_data);
- /*
- * Alternative to holding the lock for a long time:
- * - keep busy wait - have other users bounce off
- */
- spin_lock_irqsave(&dd->dc8051_lock, flags);
+ mutex_lock(&dd->dc8051_lock);
/* We can't send any commands to the 8051 if it's in reset */
if (dd->dc_shutdown) {
@@ -8510,10 +8514,8 @@ static int do_8051_command(
return_code = -ENXIO;
goto fail;
}
- spin_unlock_irqrestore(&dd->dc8051_lock, flags);
- dc_shutdown(dd);
- dc_start(dd);
- spin_lock_irqsave(&dd->dc8051_lock, flags);
+ _dc_shutdown(dd);
+ _dc_start(dd);
}
/*
@@ -8594,8 +8596,7 @@ static int do_8051_command(
write_csr(dd, DC_DC8051_CFG_HOST_CMD_0, 0);
fail:
- spin_unlock_irqrestore(&dd->dc8051_lock, flags);
-
+ mutex_unlock(&dd->dc8051_lock);
return return_code;
}
@@ -11969,6 +11970,10 @@ static void free_cntrs(struct hfi1_devdata *dd)
dd->scntrs = NULL;
kfree(dd->cntrnames);
dd->cntrnames = NULL;
+ if (dd->update_cntr_wq) {
+ destroy_workqueue(dd->update_cntr_wq);
+ dd->update_cntr_wq = NULL;
+ }
}
static u64 read_dev_port_cntr(struct hfi1_devdata *dd, struct cntr_entry *entry,
@@ -12124,7 +12129,7 @@ u64 write_port_cntr(struct hfi1_pportdata *ppd, int index, int vl, u64 data)
return write_dev_port_cntr(ppd->dd, entry, sval, ppd, vl, data);
}
-static void update_synth_timer(unsigned long opaque)
+static void do_update_synth_timer(struct work_struct *work)
{
u64 cur_tx;
u64 cur_rx;
@@ -12133,8 +12138,8 @@ static void update_synth_timer(unsigned long opaque)
int i, j, vl;
struct hfi1_pportdata *ppd;
struct cntr_entry *entry;
-
- struct hfi1_devdata *dd = (struct hfi1_devdata *)opaque;
+ struct hfi1_devdata *dd = container_of(work, struct hfi1_devdata,
+ update_cntr_work);
/*
* Rather than keep beating on the CSRs pick a minimal set that we can
@@ -12217,7 +12222,13 @@ static void update_synth_timer(unsigned long opaque)
} else {
hfi1_cdbg(CNTR, "[%d] No update necessary", dd->unit);
}
+}
+static void update_synth_timer(unsigned long opaque)
+{
+ struct hfi1_devdata *dd = (struct hfi1_devdata *)opaque;
+
+ queue_work(dd->update_cntr_wq, &dd->update_cntr_work);
mod_timer(&dd->synth_stats_timer, jiffies + HZ * SYNTH_CNT_TIME);
}
@@ -12453,6 +12464,13 @@ static int init_cntrs(struct hfi1_devdata *dd)
if (init_cpu_counters(dd))
goto bail;
+ dd->update_cntr_wq = alloc_ordered_workqueue("hfi1_update_cntr_%d",
+ WQ_MEM_RECLAIM, dd->unit);
+ if (!dd->update_cntr_wq)
+ goto bail;
+
+ INIT_WORK(&dd->update_cntr_work, do_update_synth_timer);
+
mod_timer(&dd->synth_stats_timer, jiffies + HZ * SYNTH_CNT_TIME);
return 0;
bail:
diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h
index 79c4289..c6b6853 100644
--- a/drivers/infiniband/hw/hfi1/hfi.h
+++ b/drivers/infiniband/hw/hfi1/hfi.h
@@ -474,7 +474,7 @@ struct hfi1_packet {
#define HFI1_PART_ENFORCE_OUT 0x2
/* how often we check for synthetic counter wrap around */
-#define SYNTH_CNT_TIME 2
+#define SYNTH_CNT_TIME 3
/* Counter flags */
#define CNTR_NORMAL 0x0 /* Normal counters, just read register */
@@ -926,8 +926,9 @@ struct hfi1_devdata {
spinlock_t rcvctrl_lock; /* protect changes to RcvCtrl */
/* around rcd and (user ctxts) ctxt_cnt use (intr vs free) */
spinlock_t uctxt_lock; /* rcd and user context changes */
- /* exclusive access to 8051 */
- spinlock_t dc8051_lock;
+ struct mutex dc8051_lock; /* exclusive access to 8051 */
+ struct workqueue_struct *update_cntr_wq;
+ struct work_struct update_cntr_work;
/* exclusive access to 8051 memory */
spinlock_t dc8051_memlock;
int dc8051_timed_out; /* remember if the 8051 timed out */
diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c
index 9bfb8eb..55a11a6 100644
--- a/drivers/infiniband/hw/hfi1/init.c
+++ b/drivers/infiniband/hw/hfi1/init.c
@@ -1078,11 +1078,11 @@ struct hfi1_devdata *hfi1_alloc_devdata(struct pci_dev *pdev, size_t extra)
spin_lock_init(&dd->uctxt_lock);
spin_lock_init(&dd->hfi1_diag_trans_lock);
spin_lock_init(&dd->sc_init_lock);
- spin_lock_init(&dd->dc8051_lock);
spin_lock_init(&dd->dc8051_memlock);
seqlock_init(&dd->sc2vl_lock);
spin_lock_init(&dd->sde_map_lock);
spin_lock_init(&dd->pio_map_lock);
+ mutex_init(&dd->dc8051_lock);
init_waitqueue_head(&dd->event_queue);
dd->int_counter = alloc_percpu(u64);
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
