IB/hfi1: Fix softlockup issue
authorTadeusz Struk <tadeusz.struk@intel.com>
Fri, 28 Apr 2017 17:40:02 +0000 (10:40 -0700)
committerDoug Ledford <dledford@redhat.com>
Fri, 28 Apr 2017 17:56:15 +0000 (13:56 -0400)
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@intel.com>
Reviewed-by: Mike Marciszyn <mike.marciniszyn@intel.com>
Signed-off-by: Tadeusz Struk <tadeusz.struk@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/hfi1/chip.c
drivers/infiniband/hw/hfi1/hfi.h
drivers/infiniband/hw/hfi1/init.c

index e520929ac5019d4afaf139873b835c25ff6c60b5..07aa76a74f642832560c23462b09e62e938bb1b9 100644 (file)
@@ -6410,18 +6410,17 @@ 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);
        /*
@@ -6432,35 +6431,45 @@ 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);
 }
 
 /*
@@ -8513,16 +8522,11 @@ 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) {
@@ -8548,10 +8552,8 @@ 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);
        }
 
        /*
@@ -8632,8 +8634,7 @@ 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;
 }
 
@@ -12007,6 +12008,10 @@ 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,
@@ -12162,7 +12167,7 @@ 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;
@@ -12171,8 +12176,8 @@ 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
@@ -12255,7 +12260,13 @@ 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);
 }
 
@@ -12491,6 +12502,13 @@ 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:
index 550116fd9d48dd3a8071b02e59d8c53977d93c09..d253ea2521a0960e757d3071ab492fd9bfed1d1e 100644 (file)
@@ -484,7 +484,7 @@ struct rvt_sge_state;
 #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 */
@@ -962,8 +962,9 @@ 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 */
index b4c7e04f45780143dfd09c591e44ffb7f0abb930..21dca7ac059ca77313b694db5cbf3bdefa252e25 100644 (file)
@@ -1081,11 +1081,11 @@ 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);