drivers/infiniband/hw/ipath/ipath_stats.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
   3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include "ipath_kernel.h"
  35
  36 struct infinipath_stats ipath_stats;
  37
  38 /**
  39  * ipath_snap_cntr - snapshot a chip counter
  40  * @dd: the infinipath device
  41  * @creg: the counter to snapshot
  42  *
  43  * called from add_timer and user counter read calls, to deal with
  44  * counters that wrap in "human time".  The words sent and received, and
  45  * the packets sent and received are all that we worry about.  For now,
  46  * at least, we don't worry about error counters, because if they wrap
  47  * that quickly, we probably don't care.  We may eventually just make this
  48  * handle all the counters.  word counters can wrap in about 20 seconds
  49  * of full bandwidth traffic, packet counters in a few hours.
  50  */
  51
  52 u64 ipath_snap_cntr(struct ipath_devdata *dd, ipath_creg creg)
  53 {
  54         u32 val, reg64 = 0;
  55         u64 val64;
  56         unsigned long t0, t1;
  57         u64 ret;
  58
  59         t0 = jiffies;
  60         /* If fast increment counters are only 32 bits, snapshot them,
  61          * and maintain them as 64bit values in the driver */
  62         if (!(dd->ipath_flags & IPATH_32BITCOUNTERS) &&
  63             (creg == dd->ipath_cregs->cr_wordsendcnt ||
  64              creg == dd->ipath_cregs->cr_wordrcvcnt ||
  65              creg == dd->ipath_cregs->cr_pktsendcnt ||
  66              creg == dd->ipath_cregs->cr_pktrcvcnt)) {
  67                 val64 = ipath_read_creg(dd, creg);
  68                 val = val64 == ~0ULL ? ~0U : 0;
  69                 reg64 = 1;
  70         } else                  /* val64 just to keep gcc quiet... */
  71                 val64 = val = ipath_read_creg32(dd, creg);
  72         /*
  73          * See if a second has passed.  This is just a way to detect things
  74          * that are quite broken.  Normally this should take just a few
  75          * cycles (the check is for long enough that we don't care if we get
  76          * pre-empted.)  An Opteron HT O read timeout is 4 seconds with
  77          * normal NB values
  78          */
  79         t1 = jiffies;
  80         if (time_before(t0 + HZ, t1) && val == -1) {
  81                 ipath_dev_err(dd, "Error!  Read counter 0x%x timed out\n",
  82                               creg);
  83                 ret = 0ULL;
  84                 goto bail;
  85         }
  86         if (reg64) {
  87                 ret = val64;
  88                 goto bail;
  89         }
  90
  91         if (creg == dd->ipath_cregs->cr_wordsendcnt) {
  92                 if (val != dd->ipath_lastsword) {
  93                         dd->ipath_sword += val - dd->ipath_lastsword;
  94                         dd->ipath_lastsword = val;
  95                 }
  96                 val64 = dd->ipath_sword;
  97         } else if (creg == dd->ipath_cregs->cr_wordrcvcnt) {
  98                 if (val != dd->ipath_lastrword) {
  99                         dd->ipath_rword += val - dd->ipath_lastrword;
 100                         dd->ipath_lastrword = val;
 101                 }
 102                 val64 = dd->ipath_rword;
 103         } else if (creg == dd->ipath_cregs->cr_pktsendcnt) {
 104                 if (val != dd->ipath_lastspkts) {
 105                         dd->ipath_spkts += val - dd->ipath_lastspkts;
 106                         dd->ipath_lastspkts = val;
 107                 }
 108                 val64 = dd->ipath_spkts;
 109         } else if (creg == dd->ipath_cregs->cr_pktrcvcnt) {
 110                 if (val != dd->ipath_lastrpkts) {
 111                         dd->ipath_rpkts += val - dd->ipath_lastrpkts;
 112                         dd->ipath_lastrpkts = val;
 113                 }
 114                 val64 = dd->ipath_rpkts;
 115         } else
 116                 val64 = (u64) val;
 117
 118         ret = val64;
 119
 120 bail:
 121         return ret;
 122 }
 123
 124 /**
 125  * ipath_qcheck - print delta of egrfull/hdrqfull errors for kernel ports
 126  * @dd: the infinipath device
 127  *
 128  * print the delta of egrfull/hdrqfull errors for kernel ports no more than
 129  * every 5 seconds.  User processes are printed at close, but kernel doesn't
 130  * close, so...  Separate routine so may call from other places someday, and
 131  * so function name when printed by _IPATH_INFO is meaningfull
 132  */
 133 static void ipath_qcheck(struct ipath_devdata *dd)
 134 {
 135         static u64 last_tot_hdrqfull;
 136         struct ipath_portdata *pd = dd->ipath_pd[0];
 137         size_t blen = 0;
 138         char buf[128];
 139         u32 hdrqtail;
 140
 141         *buf = 0;
 142         if (pd->port_hdrqfull != dd->ipath_p0_hdrqfull) {
 143                 blen = snprintf(buf, sizeof buf, "port 0 hdrqfull %u",
 144                                 pd->port_hdrqfull -
 145                                 dd->ipath_p0_hdrqfull);
 146                 dd->ipath_p0_hdrqfull = pd->port_hdrqfull;
 147         }
 148         if (ipath_stats.sps_etidfull != dd->ipath_last_tidfull) {
 149                 blen += snprintf(buf + blen, sizeof buf - blen,
 150                                  "%srcvegrfull %llu",
 151                                  blen ? ", " : "",
 152                                  (unsigned long long)
 153                                  (ipath_stats.sps_etidfull -
 154                                   dd->ipath_last_tidfull));
 155                 dd->ipath_last_tidfull = ipath_stats.sps_etidfull;
 156         }
 157
 158         /*
 159          * this is actually the number of hdrq full interrupts, not actual
 160          * events, but at the moment that's mostly what I'm interested in.
 161          * Actual count, etc. is in the counters, if needed.  For production
 162          * users this won't ordinarily be printed.
 163          */
 164
 165         if ((ipath_debug & (__IPATH_PKTDBG | __IPATH_DBG)) &&
 166             ipath_stats.sps_hdrqfull != last_tot_hdrqfull) {
 167                 blen += snprintf(buf + blen, sizeof buf - blen,
 168                                  "%shdrqfull %llu (all ports)",
 169                                  blen ? ", " : "",
 170                                  (unsigned long long)
 171                                  (ipath_stats.sps_hdrqfull -
 172                                   last_tot_hdrqfull));
 173                 last_tot_hdrqfull = ipath_stats.sps_hdrqfull;
 174         }
 175         if (blen)
 176                 ipath_dbg("%s\n", buf);
 177
 178         hdrqtail = ipath_get_hdrqtail(pd);
 179         if (pd->port_head != hdrqtail) {
 180                 if (dd->ipath_lastport0rcv_cnt ==
 181                     ipath_stats.sps_port0pkts) {
 182                         ipath_cdbg(PKT, "missing rcv interrupts? "
 183                                    "port0 hd=%x tl=%x; port0pkts %llx; write"
 184                                    " hd (w/intr)\n",
 185                                    pd->port_head, hdrqtail,
 186                                    (unsigned long long)
 187                                    ipath_stats.sps_port0pkts);
 188                         ipath_write_ureg(dd, ur_rcvhdrhead, hdrqtail |
 189                                 dd->ipath_rhdrhead_intr_off, pd->port_port);
 190                 }
 191                 dd->ipath_lastport0rcv_cnt = ipath_stats.sps_port0pkts;
 192         }
 193 }
 194
 195 static void ipath_chk_errormask(struct ipath_devdata *dd)
 196 {
 197         static u32 fixed;
 198         u32 ctrl;
 199         unsigned long errormask;
 200         unsigned long hwerrs;
 201
 202         if (!dd->ipath_errormask || !(dd->ipath_flags & IPATH_INITTED))
 203                 return;
 204
 205         errormask = ipath_read_kreg64(dd, dd->ipath_kregs->kr_errormask);
 206
 207         if (errormask == dd->ipath_errormask)
 208                 return;
 209         fixed++;
 210
 211         hwerrs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus);
 212         ctrl = ipath_read_kreg32(dd, dd->ipath_kregs->kr_control);
 213
 214         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 215                 dd->ipath_errormask);
 216
 217         if ((hwerrs & dd->ipath_hwerrmask) ||
 218                 (ctrl & INFINIPATH_C_FREEZEMODE)) {
 219                 /* force re-interrupt of pending events, just in case */
 220                 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
 221                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, 0ULL);
 222                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 223                 dev_info(&dd->pcidev->dev,
 224                         "errormask fixed(%u) %lx -> %lx, ctrl %x hwerr %lx\n",
 225                         fixed, errormask, (unsigned long)dd->ipath_errormask,
 226                         ctrl, hwerrs);
 227         } else
 228                 ipath_dbg("errormask fixed(%u) %lx -> %lx, no freeze\n",
 229                         fixed, errormask,
 230                         (unsigned long)dd->ipath_errormask);
 231 }
 232
 233
 234 /**
 235  * ipath_get_faststats - get word counters from chip before they overflow
 236  * @opaque - contains a pointer to the infinipath device ipath_devdata
 237  *
 238  * called from add_timer
 239  */
 240 void ipath_get_faststats(unsigned long opaque)
 241 {
 242         struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
 243         int i;
 244         static unsigned cnt;
 245         unsigned long flags;
 246         u64 traffic_wds;
 247
 248         /*
 249          * don't access the chip while running diags, or memory diags can
 250          * fail
 251          */
 252         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_INITTED) ||
 253             ipath_diag_inuse)
 254                 /* but re-arm the timer, for diags case; won't hurt other */
 255                 goto done;
 256
 257         /*
 258          * We now try to maintain a "active timer", based on traffic
 259          * exceeding a threshold, so we need to check the word-counts
 260          * even if they are 64-bit.
 261          */
 262         traffic_wds = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt) +
 263                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
 264         spin_lock_irqsave(&dd->ipath_eep_st_lock, flags);
 265         traffic_wds -= dd->ipath_traffic_wds;
 266         dd->ipath_traffic_wds += traffic_wds;
 267         if (traffic_wds  >= IPATH_TRAFFIC_ACTIVE_THRESHOLD)
 268                 atomic_add(5, &dd->ipath_active_time); /* S/B #define */
 269         spin_unlock_irqrestore(&dd->ipath_eep_st_lock, flags);
 270
 271         if (dd->ipath_flags & IPATH_32BITCOUNTERS) {
 272                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
 273                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
 274         }
 275
 276         ipath_qcheck(dd);
 277
 278         /*
 279          * deal with repeat error suppression.  Doesn't really matter if
 280          * last error was almost a full interval ago, or just a few usecs
 281          * ago; still won't get more than 2 per interval.  We may want
 282          * longer intervals for this eventually, could do with mod, counter
 283          * or separate timer.  Also see code in ipath_handle_errors() and
 284          * ipath_handle_hwerrors().
 285          */
 286
 287         if (dd->ipath_lasterror)
 288                 dd->ipath_lasterror = 0;
 289         if (dd->ipath_lasthwerror)
 290                 dd->ipath_lasthwerror = 0;
 291         if (dd->ipath_maskederrs
 292             && time_after(jiffies, dd->ipath_unmasktime)) {
 293                 char ebuf[256];
 294                 int iserr;
 295                 iserr = ipath_decode_err(dd, ebuf, sizeof ebuf,
 296                                          dd->ipath_maskederrs);
 297                 if (dd->ipath_maskederrs &
 298                     ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 299                       INFINIPATH_E_PKTERRS))
 300                         ipath_dev_err(dd, "Re-enabling masked errors "
 301                                       "(%s)\n", ebuf);
 302                 else {
 303                         /*
 304                          * rcvegrfull and rcvhdrqfull are "normal", for some
 305                          * types of processes (mostly benchmarks) that send
 306                          * huge numbers of messages, while not processing
 307                          * them.  So only complain about these at debug
 308                          * level.
 309                          */
 310                         if (iserr)
 311                                 ipath_dbg(
 312                                         "Re-enabling queue full errors (%s)\n",
 313                                         ebuf);
 314                         else
 315                                 ipath_cdbg(ERRPKT, "Re-enabling packet"
 316                                         " problem interrupt (%s)\n", ebuf);
 317                 }
 318
 319                 /* re-enable masked errors */
 320                 dd->ipath_errormask |= dd->ipath_maskederrs;
 321                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 322                                  dd->ipath_errormask);
 323                 dd->ipath_maskederrs = 0;
 324         }
 325
 326         /* limit qfull messages to ~one per minute per port */
 327         if ((++cnt & 0x10)) {
 328                 for (i = (int) dd->ipath_cfgports; --i >= 0; ) {
 329                         struct ipath_portdata *pd = dd->ipath_pd[i];
 330
 331                         if (pd && pd->port_lastrcvhdrqtail != -1)
 332                                 pd->port_lastrcvhdrqtail = -1;
 333                 }
 334         }
 335
 336         ipath_chk_errormask(dd);
 337 done:
 338         mod_timer(&dd->ipath_stats_timer, jiffies + HZ * 5);
 339 }