[PATCH] IB/ipath: reduce overhead on receive interrupts
authorBryan O'Sullivan <bos@pathscale.com>
Sat, 1 Jul 2006 11:36:04 +0000 (04:36 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Sat, 1 Jul 2006 16:56:00 +0000 (09:56 -0700)
Also count the number of interrupts where that works (fastrcvint).  On any
interrupt where the port0 head and tail registers are not equal, just call the
ipath_kreceive code without reading the interrupt status, thus saving the
approximately 0.25usec processor stall waiting for the read to return.  If any
other interrupt bits are set, or head==tail, take the normal path, but that
has been reordered to handle read ahead of pioavail.  Also no longer call
ipath_kreceive() from ipath_qcheck(), because that just seems to make things
worse, and isn't really buying us anything, these days.

Also no longer loop in ipath_kreceive(); better to not hold things off too
long (I saw many cases where we would loop 4-8 times, and handle thousands (up
to 3500) in a single call).

Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Cc: "Michael S. Tsirkin" <mst@mellanox.co.il>
Cc: Roland Dreier <rolandd@cisco.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
drivers/infiniband/hw/ipath/ipath_common.h
drivers/infiniband/hw/ipath/ipath_driver.c
drivers/infiniband/hw/ipath/ipath_intr.c
drivers/infiniband/hw/ipath/ipath_stats.c

index 3a849e7..7a8a086 100644 (file)
@@ -97,8 +97,8 @@ struct infinipath_stats {
        __u64 sps_hwerrs;
        /* number of times IB link changed state unexpectedly */
        __u64 sps_iblink;
-       /* no longer used; left for compatibility */
-       __u64 sps_unused3;
+       /* kernel receive interrupts that didn't read intstat */
+       __u64 sps_fastrcvint;
        /* number of kernel (port0) packets received */
        __u64 sps_port0pkts;
        /* number of "ethernet" packets sent by driver */
index 0b88642..eb74523 100644 (file)
@@ -888,12 +888,7 @@ void ipath_kreceive(struct ipath_devdata *dd)
            (u32)le64_to_cpu(*dd->ipath_hdrqtailptr))
                goto done;
 
-gotmore:
-       /*
-        * read only once at start.  If in flood situation, this helps
-        * performance slightly.  If more arrive while we are processing,
-        * we'll come back here and do them
-        */
+       /* read only once at start for performance */
        hdrqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
 
        for (i = 0, l = dd->ipath_port0head; l != hdrqtail; i++) {
@@ -1022,10 +1017,6 @@ gotmore:
 
        dd->ipath_port0head = l;
 
-       if (hdrqtail != (u32)le64_to_cpu(*dd->ipath_hdrqtailptr))
-               /* more arrived while we handled first batch */
-               goto gotmore;
-
        if (pkttot > ipath_stats.sps_maxpkts_call)
                ipath_stats.sps_maxpkts_call = pkttot;
        ipath_stats.sps_port0pkts += pkttot;
index 8a4d732..e8c8a25 100644 (file)
@@ -539,10 +539,10 @@ 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++;
                        }
                }
        }
@@ -724,7 +724,12 @@ 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;
@@ -734,22 +739,17 @@ 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) {
@@ -767,14 +767,17 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
        struct ipath_devdata *dd = data;
        u32 istat;
        ipath_err_t estat = 0;
-       static unsigned unexpected = 0;
        irqreturn_t ret;
+       u32 p0bits;
+       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.
@@ -782,6 +785,50 @@ 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.
+        */
+       if (dd->ipath_port0head !=
+               (u32)le64_to_cpu(*dd->ipath_hdrqtailptr)) {
+               u32 oldhead = dd->ipath_port0head;
+               if (dd->ipath_flags & IPATH_GPIO_INTR) {
+                       ipath_write_kreg(dd, dd->ipath_kregs->kr_gpio_clear,
+                                        (u64) (1 << 2));
+                       p0bits = port0rbits | INFINIPATH_I_GPIO;
+               }
+               else
+                       p0bits = port0rbits;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, p0bits);
+               ipath_kreceive(dd);
+               if (oldhead != dd->ipath_port0head) {
+                       ipath_stats.sps_fastrcvint++;
+                       goto done;
+               }
+               istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
+       }
+
        istat = ipath_read_kreg32(dd, dd->ipath_kregs->kr_intstatus);
        if (unlikely(!istat)) {
                ipath_stats.sps_nullintr++;
@@ -795,31 +842,17 @@ 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);
@@ -837,7 +870,14 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                        handle_errors(dd, estat);
        }
 
+       p0bits = port0rbits;
        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(
@@ -851,14 +891,7 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
                        /* 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);
+                       p0bits |= INFINIPATH_I_GPIO;
                }
        }
 
@@ -871,6 +904,25 @@ irqreturn_t ipath_intr(int irq, void *data, struct pt_regs *regs)
         */
        ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 
+       /*
+        * we check for both transition from empty to non-empty, and urgent
+        * packets (those with the interrupt bit set in the header), and
+        * if enabled, the GPIO bit 2 interrupt used for port0 on some
+        * HT-400 boards.
+        * Do this before checking for pio buffers available, since
+        * receives can overflow; piobuf waiters can afford a few
+        * extra cycles, since they were waiting anyway.
+        */
+       if (istat & p0bits) {
+               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,
@@ -882,17 +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:
index 2651701..70351b7 100644 (file)
@@ -186,7 +186,6 @@ static void ipath_qcheck(struct ipath_devdata *dd)
                                   dd->ipath_port0head,
                                   (unsigned long long)
                                   ipath_stats.sps_port0pkts);
-                       ipath_kreceive(dd);
                }
                dd->ipath_lastport0rcv_cnt = ipath_stats.sps_port0pkts;
        }