Merge to Fedora kernel-2.6.18-1.2224_FC5 patched with stable patch-2.6.18.1-vs2.0...
[linux-2.6.git] / drivers / infiniband / hw / ipath / ipath_intr.c
index 3e72a1f..280e732 100644 (file)
@@ -1,4 +1,5 @@
 /*
+ * Copyright (c) 2006 QLogic, Inc. All rights reserved.
  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
  *
  * This software is available to you under a choice of one of two
 #include <linux/pci.h>
 
 #include "ipath_kernel.h"
-#include "ips_common.h"
 #include "ipath_layer.h"
+#include "ipath_common.h"
 
+/* These are all rcv-related errors which we want to count for stats */
 #define E_SUM_PKTERRS \
        (INFINIPATH_E_RHDRLEN | INFINIPATH_E_RBADTID | \
         INFINIPATH_E_RBADVERSION | INFINIPATH_E_RHDR | \
@@ -44,6 +46,7 @@
         INFINIPATH_E_RFORMATERR | INFINIPATH_E_RUNSUPVL | \
         INFINIPATH_E_RUNEXPCHAR | INFINIPATH_E_REBP)
 
+/* These are all send-related errors which we want to count for stats */
 #define E_SUM_ERRS \
        (INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM | \
         INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
         INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
         INFINIPATH_E_INVALIDADDR)
 
+/*
+ * these are errors that can occur when the link changes state while
+ * a packet is being sent or received.  This doesn't cover things
+ * like EBP or VCRC that can be the result of a sending having the
+ * link change state, so we receive a "known bad" packet.
+ */
+#define E_SUM_LINK_PKTERRS \
+       (INFINIPATH_E_SDROPPEDDATAPKT | INFINIPATH_E_SDROPPEDSMPPKT | \
+        INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SPKTLEN | \
+        INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RMINPKTLEN | \
+        INFINIPATH_E_RUNEXPCHAR)
+
 static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
 {
        unsigned long sbuf[4];
@@ -100,9 +115,7 @@ static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
                if (ipath_debug & __IPATH_PKTDBG)
                        printk("\n");
        }
-       if ((errs & (INFINIPATH_E_SDROPPEDDATAPKT |
-                    INFINIPATH_E_SDROPPEDSMPPKT |
-                    INFINIPATH_E_SMINPKTLEN)) &&
+       if ((errs & E_SUM_LINK_PKTERRS) &&
            !(dd->ipath_flags & IPATH_LINKACTIVE)) {
                /*
                 * This can happen when SMA is trying to bring the link
@@ -111,11 +124,9 @@ static u64 handle_e_sum_errs(struct ipath_devdata *dd, ipath_err_t errs)
                 * valid.  We don't want to confuse people, so we just
                 * don't print them, except at debug
                 */
-               ipath_dbg("Ignoring pktsend errors %llx, because not "
-                         "yet active\n", (unsigned long long) errs);
-               ignore_this_time = INFINIPATH_E_SDROPPEDDATAPKT |
-                       INFINIPATH_E_SDROPPEDSMPPKT |
-                       INFINIPATH_E_SMINPKTLEN;
+               ipath_dbg("Ignoring packet errors %llx, because link not "
+                         "ACTIVE\n", (unsigned long long) errs);
+               ignore_this_time = errs & E_SUM_LINK_PKTERRS;
        }
 
        return ignore_this_time;
@@ -156,7 +167,29 @@ static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
         */
        val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
        lstate = val & IPATH_IBSTATE_MASK;
-       if (lstate == IPATH_IBSTATE_INIT || lstate == IPATH_IBSTATE_ARM ||
+
+       /*
+        * this is confusing enough when it happens that I want to always put it
+        * on the console and in the logs.  If it was a requested state change,
+        * we'll have already cleared the flags, so we won't print this warning
+        */
+       if ((lstate != IPATH_IBSTATE_ARM && lstate != IPATH_IBSTATE_ACTIVE)
+               && (dd->ipath_flags & (IPATH_LINKARMED | IPATH_LINKACTIVE))) {
+               dev_info(&dd->pcidev->dev, "Link state changed from %s to %s\n",
+                                (dd->ipath_flags & IPATH_LINKARMED) ? "ARM" : "ACTIVE",
+                                ib_linkstate(lstate));
+               /*
+                * Flush all queued sends when link went to DOWN or INIT,
+                * to be sure that they don't block SMA and other MAD packets
+                */
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+                                INFINIPATH_S_ABORT);
+               ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
+                                                       (unsigned)(dd->ipath_piobcnt2k +
+                                       dd->ipath_piobcnt4k) -
+                                       dd->ipath_lastport_piobuf);
+       }
+       else if (lstate == IPATH_IBSTATE_INIT || lstate == IPATH_IBSTATE_ARM ||
            lstate == IPATH_IBSTATE_ACTIVE) {
                /*
                 * only print at SMA if there is a change, debug if not
@@ -229,6 +262,7 @@ static void handle_e_ibstatuschanged(struct ipath_devdata *dd,
                                     | IPATH_LINKACTIVE |
                                     IPATH_LINKARMED);
                *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
+               dd->ipath_lli_counter = 0;
                if (!noprint) {
                        if (((dd->ipath_lastibcstat >>
                              INFINIPATH_IBCS_LINKSTATE_SHIFT) &
@@ -350,7 +384,7 @@ static unsigned handle_frequent_errors(struct ipath_devdata *dd,
        return supp_msgs;
 }
 
-static void handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
+static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 {
        char msg[512];
        u64 ignore_this_time = 0;
@@ -379,6 +413,19 @@ static void handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 
        if (errs & E_SUM_ERRS)
                ignore_this_time = handle_e_sum_errs(dd, errs);
+       else if ((errs & E_SUM_LINK_PKTERRS) &&
+           !(dd->ipath_flags & IPATH_LINKACTIVE)) {
+               /*
+                * This can happen when SMA is trying to bring the link
+                * up, but the IB link changes state at the "wrong" time.
+                * The IB logic then complains that the packet isn't
+                * valid.  We don't want to confuse people, so we just
+                * don't print them, except at debug
+                */
+               ipath_dbg("Ignoring packet errors %llx, because link not "
+                         "ACTIVE\n", (unsigned long long) errs);
+               ignore_this_time = errs & E_SUM_LINK_PKTERRS;
+       }
 
        if (supp_msgs == 250000) {
                /*
@@ -397,7 +444,7 @@ static void handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
                if ((dd->ipath_maskederrs & ~dd->ipath_ignorederrs) &
                    ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL))
                        ipath_dev_err(dd, "Disabling error(s) %llx because "
-                                     "occuring too frequently (%s)\n",
+                                     "occurring too frequently (%s)\n",
                                      (unsigned long long)
                                      (dd->ipath_maskederrs &
                                       ~dd->ipath_ignorederrs), msg);
@@ -434,7 +481,7 @@ static void handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
                          INFINIPATH_E_IBSTATUSCHANGED);
        }
        if (!errs)
-               return;
+               return 0;
 
        if (!noprint)
                /*
@@ -493,10 +540,10 @@ static void handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
                                continue;
                        if (hd == (tl + 1) ||
                            (!hd && tl == dd->ipath_hdrqlast)) {
-                               dd->ipath_lastrcvhdrqtails[i] = tl;
-                               pd->port_hdrqfull++;
                                if (i == 0)
                                        chkerrpkts = 1;
+                               dd->ipath_lastrcvhdrqtails[i] = tl;
+                               pd->port_hdrqfull++;
                        }
                }
        }
@@ -558,9 +605,7 @@ static void handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
                wake_up_interruptible(&ipath_sma_state_wait);
        }
 
-       if (chkerrpkts)
-               /* process possible error packets in hdrq */
-               ipath_kreceive(dd);
+       return chkerrpkts;
 }
 
 /* this is separate to allow for better optimization of ipath_intr() */
@@ -678,7 +723,12 @@ set:
                         dd->ipath_sendctrl);
 }
 
-static void handle_rcv(struct ipath_devdata *dd, u32 istat)
+/*
+ * Handle receive interrupts for user ports; this means a user
+ * process was waiting for a packet to arrive, and didn't want
+ * to poll
+ */
+static void handle_urcv(struct ipath_devdata *dd, u32 istat)
 {
        u64 portr;
        int i;
@@ -688,22 +738,17 @@ static void handle_rcv(struct ipath_devdata *dd, u32 istat)
                 infinipath_i_rcvavail_mask)
                | ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
                   infinipath_i_rcvurg_mask);
-       for (i = 0; i < dd->ipath_cfgports; i++) {
+       for (i = 1; i < dd->ipath_cfgports; i++) {
                struct ipath_portdata *pd = dd->ipath_pd[i];
-               if (portr & (1 << i) && pd &&
-                   pd->port_cnt) {
-                       if (i == 0)
-                               ipath_kreceive(dd);
-                       else if (test_bit(IPATH_PORT_WAITING_RCV,
-                                         &pd->port_flag)) {
-                               int rcbit;
-                               clear_bit(IPATH_PORT_WAITING_RCV,
-                                         &pd->port_flag);
-                               rcbit = i + INFINIPATH_R_INTRAVAIL_SHIFT;
-                               clear_bit(1UL << rcbit, &dd->ipath_rcvctrl);
-                               wake_up_interruptible(&pd->port_wait);
-                               rcvdint = 1;
-                       }
+               if (portr & (1 << i) && pd && pd->port_cnt &&
+                       test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
+                       int rcbit;
+                       clear_bit(IPATH_PORT_WAITING_RCV,
+                                 &pd->port_flag);
+                       rcbit = i + INFINIPATH_R_INTRAVAIL_SHIFT;
+                       clear_bit(1UL << rcbit, &dd->ipath_rcvctrl);
+                       wake_up_interruptible(&pd->port_wait);
+                       rcvdint = 1;
                }
        }
        if (rcvdint) {
@@ -719,16 +764,19 @@ static void handle_rcv(struct ipath_devdata *dd, u32 istat)
 irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
 {
        struct ipath_devdata *dd = data;
-       u32 istat;
+       u32 istat, chk0rcv = 0;
        ipath_err_t estat = 0;
-       static unsigned unexpected = 0;
        irqreturn_t ret;
+       u32 oldhead, curtail;
+       static unsigned unexpected = 0;
+       static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
+                (1U<<INFINIPATH_I_RCVURG_SHIFT);
+
+       ipath_stats.sps_ints++;
 
-       if(!(dd->ipath_flags & IPATH_PRESENT)) {
-               /* this is mostly so we don't try to touch the chip while
-                * it is being reset */
+       if (!(dd->ipath_flags & IPATH_PRESENT)) {
                /*
-                * This return value is perhaps odd, but we do not want the
+                * This return value is not great, but we do not want the
                 * interrupt core code to remove our interrupt handler
                 * because we don't appear to be handling an interrupt
                 * during a chip reset.
@@ -736,7 +784,51 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                return IRQ_HANDLED;
        }
 
+       /*
+        * this needs to be flags&initted, not statusp, so we keep
+        * taking interrupts even after link goes down, etc.
+        * Also, we *must* clear the interrupt at some point, or we won't
+        * take it again, which can be real bad for errors, etc...
+        */
+
+       if (!(dd->ipath_flags & IPATH_INITTED)) {
+               ipath_bad_intr(dd, &unexpected);
+               ret = IRQ_NONE;
+               goto bail;
+       }
+
+       /*
+        * We try to avoid reading the interrupt status register, since
+        * that's a PIO read, and stalls the processor for up to about
+        * ~0.25 usec. The idea is that if we processed a port0 packet,
+        * we blindly clear the  port 0 receive interrupt bits, and nothing
+        * else, then return.  If other interrupts are pending, the chip
+        * will re-interrupt us as soon as we write the intclear register.
+        * We then won't process any more kernel packets (if not the 2nd
+        * time, then the 3rd or 4th) and we'll then handle the other
+        * interrupts.   We clear the interrupts first so that we don't
+        * lose intr for later packets that arrive while we are processing.
+        */
+       oldhead = dd->ipath_port0head;
+       curtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
+       if (oldhead != curtail) {
+               if (dd->ipath_flags & IPATH_GPIO_INTR) {
+                       ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
+                                        (u64) (1 << 2));
+                       istat = port0rbits | INFINIPATH_I_GPIO;
+               }
+               else
+                       istat = port0rbits;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
+               ipath_kreceive(dd);
+               if (oldhead != dd->ipath_port0head) {
+                       ipath_stats.sps_fastrcvint++;
+                       goto done;
+               }
+       }
+
        istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
+
        if (unlikely(!istat)) {
                ipath_stats.sps_nullintr++;
                ret = IRQ_NONE; /* not our interrupt, or already handled */
@@ -749,31 +841,17 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                goto bail;
        }
 
-       ipath_stats.sps_ints++;
-
-       /*
-        * this needs to be flags&initted, not statusp, so we keep
-        * taking interrupts even after link goes down, etc.
-        * Also, we *must* clear the interrupt at some point, or we won't
-        * take it again, which can be real bad for errors, etc...
-        */
-
-       if (!(dd->ipath_flags & IPATH_INITTED)) {
-               ipath_bad_intr(dd, &unexpected);
-               ret = IRQ_NONE;
-               goto bail;
-       }
        if (unexpected)
                unexpected = 0;
 
-       ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
-
-       if (istat & ~infinipath_i_bitsextant)
+       if (unlikely(istat & ~infinipath_i_bitsextant))
                ipath_dev_err(dd,
                              "interrupt with unknown interrupts %x set\n",
                              istat & (u32) ~ infinipath_i_bitsextant);
+       else
+               ipath_cdbg(VERBOSE, "intr stat=0x%x\n", istat);
 
-       if (istat & INFINIPATH_I_ERROR) {
+       if (unlikely(istat & INFINIPATH_I_ERROR)) {
                ipath_stats.sps_errints++;
                estat = ipath_read_kreg64(dd,
                                          dd->ipath_kregs->kr_errorstatus);
@@ -788,10 +866,18 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                        ipath_dev_err(dd, "Read of error status failed "
                                      "(all bits set); ignoring\n");
                else
-                       handle_errors(dd, estat);
+                       if (handle_errors(dd, estat))
+                               /* force calling ipath_kreceive() */
+                               chk0rcv = 1;
        }
 
        if (istat & INFINIPATH_I_GPIO) {
+               /*
+                * Packets are available in the port 0 rcv queue.
+                * Eventually this needs to be generalized to check
+                * IPATH_GPIO_INTR, and the specific GPIO bit, if
+                * GPIO interrupts are used for anything else.
+                */
                if (unlikely(!(dd->ipath_flags & IPATH_GPIO_INTR))) {
                        u32 gpiostatus;
                        gpiostatus = ipath_read_kreg32(
@@ -804,27 +890,39 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                else {
                        /* Clear GPIO status bit 2 */
                        ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
-                                        (u64) (1 << 2));
-
-                       /*
-                        * Packets are available in the port 0 rcv queue.
-                        * Eventually this needs to be generalized to check
-                        * IPATH_GPIO_INTR, and the specific GPIO bit, if
-                        * GPIO interrupts are used for anything else.
-                        */
-                       ipath_kreceive(dd);
+                                       (u64) (1 << 2));
+                       chk0rcv = 1;
                }
        }
+       chk0rcv |= istat & port0rbits;
 
        /*
-        * clear the ones we will deal with on this round
-        * We clear it early, mostly for receive interrupts, so we
-        * know the chip will have seen this by the time we process
-        * the queue, and will re-interrupt if necessary.  The processor
-        * itself won't take the interrupt again until we return.
+        * Clear the interrupt bits we found set, unless they are receive
+        * related, in which case we already cleared them above, and don't
+        * want to clear them again, because we might lose an interrupt.
+        * Clear it early, so we "know" know the chip will have seen this by
+        * the time we process the queue, and will re-interrupt if necessary.
+        * The processor itself won't take the interrupt again until we return.
         */
        ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 
+       /*
+        * handle port0 receive  before checking for pio buffers available,
+        * since receives can overflow; piobuf waiters can afford a few
+        * extra cycles, since they were waiting anyway, and user's waiting
+        * for receive are at the bottom.
+        */
+       if (chk0rcv) {
+               ipath_kreceive(dd);
+               istat &= ~port0rbits;
+       }
+
+       if (istat & ((infinipath_i_rcvavail_mask <<
+                     INFINIPATH_I_RCVAVAIL_SHIFT)
+                    | (infinipath_i_rcvurg_mask <<
+                       INFINIPATH_I_RCVURG_SHIFT)))
+               handle_urcv(dd, istat);
+
        if (istat & INFINIPATH_I_SPIOBUFAVAIL) {
                clear_bit(IPATH_S_PIOINTBUFAVAIL, &dd->ipath_sendctrl);
                ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
@@ -836,17 +934,7 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                handle_layer_pioavail(dd);
        }
 
-       /*
-        * we check for both transition from empty to non-empty, and urgent
-        * packets (those with the interrupt bit set in the header)
-        */
-
-       if (istat & ((infinipath_i_rcvavail_mask <<
-                     INFINIPATH_I_RCVAVAIL_SHIFT)
-                    | (infinipath_i_rcvurg_mask <<
-                       INFINIPATH_I_RCVURG_SHIFT)))
-               handle_rcv(dd, istat);
-
+done:
        ret = IRQ_HANDLED;
 
 bail: