IB/ipath: Fix up error handling
authorRalph Campbell <ralph.campbell@qlogic.com>
Thu, 17 Apr 2008 04:09:30 +0000 (21:09 -0700)
committerRoland Dreier <rolandd@cisco.com>
Thu, 17 Apr 2008 04:09:30 +0000 (21:09 -0700)
This patch makes chip reset more robust and reduces lock contention
between user and kernel TID register updates.

Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_iba6120.c
drivers/infiniband/hw/ipath/ipath_init_chip.c
drivers/infiniband/hw/ipath/ipath_kernel.h

index 907b61b5975a0f471a45bf50706fcece7c92a50b..c8d8f1a2c8fbf6f4bc23380e12b5ad14a328303c 100644 (file)
@@ -558,12 +558,40 @@ static void ipath_pe_handle_hwerrors(struct ipath_devdata *dd, char *msg,
                                 dd->ipath_hwerrmask);
        }
 
-       if (*msg)
+       if (hwerrs) {
+               /*
+                * if any set that we aren't ignoring; only
+                * make the complaint once, in case it's stuck
+                * or recurring, and we get here multiple
+                * times.
+                */
                ipath_dev_err(dd, "%s hardware error\n", msg);
-       if (isfatal && !ipath_diag_inuse && dd->ipath_freezemsg) {
+               if (dd->ipath_flags & IPATH_INITTED) {
+                       ipath_set_linkstate(dd, IPATH_IB_LINKDOWN);
+                       ipath_setup_pe_setextled(dd,
+                               INFINIPATH_IBCS_L_STATE_DOWN,
+                               INFINIPATH_IBCS_LT_STATE_DISABLED);
+                       ipath_dev_err(dd, "Fatal Hardware Error (freeze "
+                                         "mode), no longer usable, SN %.16s\n",
+                                         dd->ipath_serial);
+                       isfatal = 1;
+               }
+               *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
+               /* mark as having had error */
+               *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
+               /*
+                * mark as not usable, at a minimum until driver
+                * is reloaded, probably until reboot, since no
+                * other reset is possible.
+                */
+               dd->ipath_flags &= ~IPATH_INITTED;
+       } else
+               *msg = 0; /* recovered from all of them */
+
+       if (isfatal && !ipath_diag_inuse && dd->ipath_freezemsg && msg) {
                /*
-                * for /sys status file ; if no trailing } is copied, we'll
-                * know it was truncated.
+                * for /sys status file ; if no trailing brace is copied,
+                * we'll know it was truncated.
                 */
                snprintf(dd->ipath_freezemsg, dd->ipath_freezelen,
                         "{%s}", msg);
@@ -1127,10 +1155,7 @@ static void ipath_init_pe_variables(struct ipath_devdata *dd)
                INFINIPATH_HWE_RXEMEMPARITYERR_MASK <<
                INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT;
 
-       dd->ipath_eep_st_masks[2].errs_to_log =
-               INFINIPATH_E_INVALIDADDR | INFINIPATH_E_RESET;
-
-
+       dd->ipath_eep_st_masks[2].errs_to_log = INFINIPATH_E_RESET;
        dd->delay_mult = 2; /* SDR, 4X, can't change */
 }
 
@@ -1204,6 +1229,9 @@ static int ipath_setup_pe_reset(struct ipath_devdata *dd)
        u64 val;
        int i;
        int ret;
+       u16 cmdval;
+
+       pci_read_config_word(dd->pcidev, PCI_COMMAND, &cmdval);
 
        /* Use ERROR so it shows up in logs, etc. */
        ipath_dev_err(dd, "Resetting InfiniPath unit %u\n", dd->ipath_unit);
@@ -1231,10 +1259,14 @@ static int ipath_setup_pe_reset(struct ipath_devdata *dd)
                        ipath_dev_err(dd, "rewrite of BAR1 failed: %d\n",
                                      r);
                /* now re-enable memory access */
+               pci_write_config_word(dd->pcidev, PCI_COMMAND, cmdval);
                if ((r = pci_enable_device(dd->pcidev)))
                        ipath_dev_err(dd, "pci_enable_device failed after "
                                      "reset: %d\n", r);
-               /* whether it worked or not, mark as present, again */
+               /*
+                * whether it fully enabled or not, mark as present,
+                * again (but not INITTED)
+                */
                dd->ipath_flags |= IPATH_PRESENT;
                val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
                if (val == dd->ipath_revision) {
@@ -1273,6 +1305,11 @@ static void ipath_pe_put_tid(struct ipath_devdata *dd, u64 __iomem *tidptr,
 {
        u32 __iomem *tidp32 = (u32 __iomem *)tidptr;
        unsigned long flags = 0; /* keep gcc quiet */
+       int tidx;
+       spinlock_t *tidlockp;
+
+       if (!dd->ipath_kregbase)
+               return;
 
        if (pa != dd->ipath_tidinvalid) {
                if (pa & ((1U << 11) - 1)) {
@@ -1302,14 +1339,22 @@ static void ipath_pe_put_tid(struct ipath_devdata *dd, u64 __iomem *tidptr,
         * call can be done from interrupt level for the port 0 eager TIDs,
         * so we have to use irqsave locks.
         */
-       spin_lock_irqsave(&dd->ipath_tid_lock, flags);
+       /*
+        * Assumes tidptr always > ipath_egrtidbase
+        * if type == RCVHQ_RCV_TYPE_EAGER.
+        */
+       tidx = tidptr - dd->ipath_egrtidbase;
+
+       tidlockp = (type == RCVHQ_RCV_TYPE_EAGER && tidx < dd->ipath_rcvegrcnt)
+               ? &dd->ipath_kernel_tid_lock : &dd->ipath_user_tid_lock;
+       spin_lock_irqsave(tidlockp, flags);
        ipath_write_kreg(dd, dd->ipath_kregs->kr_scratch, 0xfeeddeaf);
-       if (dd->ipath_kregbase)
-               writel(pa, tidp32);
+       writel(pa, tidp32);
        ipath_write_kreg(dd, dd->ipath_kregs->kr_scratch, 0xdeadbeef);
        mmiowb();
-       spin_unlock_irqrestore(&dd->ipath_tid_lock, flags);
+       spin_unlock_irqrestore(tidlockp, flags);
 }
+
 /**
  * ipath_pe_put_tid_2 - write a TID in chip, Revision 2 or higher
  * @dd: the infinipath device
@@ -1325,6 +1370,10 @@ static void ipath_pe_put_tid_2(struct ipath_devdata *dd, u64 __iomem *tidptr,
                             u32 type, unsigned long pa)
 {
        u32 __iomem *tidp32 = (u32 __iomem *)tidptr;
+       u32 tidx;
+
+       if (!dd->ipath_kregbase)
+               return;
 
        if (pa != dd->ipath_tidinvalid) {
                if (pa & ((1U << 11) - 1)) {
@@ -1344,8 +1393,8 @@ static void ipath_pe_put_tid_2(struct ipath_devdata *dd, u64 __iomem *tidptr,
                else /* for now, always full 4KB page */
                        pa |= 2 << 29;
        }
-       if (dd->ipath_kregbase)
-               writel(pa, tidp32);
+       tidx = tidptr - dd->ipath_egrtidbase;
+       writel(pa, tidp32);
        mmiowb();
 }
 
index 0db19c1b45c3e2e638f74e956d1e299b8f8e64aa..8d8e572baf6ef25159bd2ead48a5ba1ce38befe7 100644 (file)
@@ -319,7 +319,7 @@ static int init_chip_first(struct ipath_devdata *dd)
        else ipath_dbg("%u 2k piobufs @ %p\n",
                       dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
 
-       spin_lock_init(&dd->ipath_tid_lock);
+       spin_lock_init(&dd->ipath_user_tid_lock);
        spin_lock_init(&dd->ipath_sendctrl_lock);
        spin_lock_init(&dd->ipath_gpio_lock);
        spin_lock_init(&dd->ipath_eep_st_lock);
index 56e51cd7d6b9df04c3b99353e5f66bc9db05c7b1..890599621098bb7c07b3e3c28dad6fd4d09684a9 100644 (file)
@@ -407,7 +407,7 @@ struct ipath_devdata {
        u64 __iomem *ipath_egrtidbase;
        /* lock to workaround chip bug 9437 and others */
        spinlock_t ipath_kernel_tid_lock;
-       spinlock_t ipath_tid_lock;
+       spinlock_t ipath_user_tid_lock;
        spinlock_t ipath_sendctrl_lock;
 
        /*