drivers/infiniband/hw/ipath/ipath_stats.c

   1 /*
   2  * Copyright (c) 2006, 2007 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         unsigned long flags;
  59
  60         t0 = jiffies;
  61         /* If fast increment counters are only 32 bits, snapshot them,
  62          * and maintain them as 64bit values in the driver */
  63         if (!(dd->ipath_flags & IPATH_32BITCOUNTERS) &&
  64             (creg == dd->ipath_cregs->cr_wordsendcnt ||
  65              creg == dd->ipath_cregs->cr_wordrcvcnt ||
  66              creg == dd->ipath_cregs->cr_pktsendcnt ||
  67              creg == dd->ipath_cregs->cr_pktrcvcnt)) {
  68                 val64 = ipath_read_creg(dd, creg);
  69                 val = val64 == ~0ULL ? ~0U : 0;
  70                 reg64 = 1;
  71         } else                  /* val64 just to keep gcc quiet... */
  72                 val64 = val = ipath_read_creg32(dd, creg);
  73         /*
  74          * See if a second has passed.  This is just a way to detect things
  75          * that are quite broken.  Normally this should take just a few
  76          * cycles (the check is for long enough that we don't care if we get
  77          * pre-empted.)  An Opteron HT O read timeout is 4 seconds with
  78          * normal NB values
  79          */
  80         t1 = jiffies;
  81         if (time_before(t0 + HZ, t1) && val == -1) {
  82                 ipath_dev_err(dd, "Error!  Read counter 0x%x timed out\n",
  83                               creg);
  84                 ret = 0ULL;
  85                 goto bail;
  86         }
  87         if (reg64) {
  88                 ret = val64;
  89                 goto bail;
  90         }
  91
  92         if (creg == dd->ipath_cregs->cr_wordsendcnt) {
  93                 if (val != dd->ipath_lastsword) {
  94                         dd->ipath_sword += val - dd->ipath_lastsword;
  95                         spin_lock_irqsave(&dd->ipath_eep_st_lock, flags);
  96                         dd->ipath_traffic_wds += val - dd->ipath_lastsword;
  97                         spin_unlock_irqrestore(&dd->ipath_eep_st_lock, flags);
  98                         dd->ipath_lastsword = val;
  99                 }
 100                 val64 = dd->ipath_sword;
 101         } else if (creg == dd->ipath_cregs->cr_wordrcvcnt) {
 102                 if (val != dd->ipath_lastrword) {
 103                         dd->ipath_rword += val - dd->ipath_lastrword;
 104                         spin_lock_irqsave(&dd->ipath_eep_st_lock, flags);
 105                         dd->ipath_traffic_wds += val - dd->ipath_lastrword;
 106                         spin_unlock_irqrestore(&dd->ipath_eep_st_lock, flags);
 107                         dd->ipath_lastrword = val;
 108                 }
 109                 val64 = dd->ipath_rword;
 110         } else if (creg == dd->ipath_cregs->cr_pktsendcnt) {
 111                 if (val != dd->ipath_lastspkts) {
 112                         dd->ipath_spkts += val - dd->ipath_lastspkts;
 113                         dd->ipath_lastspkts = val;
 114                 }
 115                 val64 = dd->ipath_spkts;
 116         } else if (creg == dd->ipath_cregs->cr_pktrcvcnt) {
 117                 if (val != dd->ipath_lastrpkts) {
 118                         dd->ipath_rpkts += val - dd->ipath_lastrpkts;
 119                         dd->ipath_lastrpkts = val;
 120                 }
 121                 val64 = dd->ipath_rpkts;
 122         } else
 123                 val64 = (u64) val;
 124
 125         ret = val64;
 126
 127 bail:
 128         return ret;
 129 }
 130
 131 /**
 132  * ipath_qcheck - print delta of egrfull/hdrqfull errors for kernel ports
 133  * @dd: the infinipath device
 134  *
 135  * print the delta of egrfull/hdrqfull errors for kernel ports no more than
 136  * every 5 seconds.  User processes are printed at close, but kernel doesn't
 137  * close, so...  Separate routine so may call from other places someday, and
 138  * so function name when printed by _IPATH_INFO is meaningfull
 139  */
 140 static void ipath_qcheck(struct ipath_devdata *dd)
 141 {
 142         static u64 last_tot_hdrqfull;
 143         size_t blen = 0;
 144         char buf[128];
 145
 146         *buf = 0;
 147         if (dd->ipath_pd[0]->port_hdrqfull != dd->ipath_p0_hdrqfull) {
 148                 blen = snprintf(buf, sizeof buf, "port 0 hdrqfull %u",
 149                                 dd->ipath_pd[0]->port_hdrqfull -
 150                                 dd->ipath_p0_hdrqfull);
 151                 dd->ipath_p0_hdrqfull = dd->ipath_pd[0]->port_hdrqfull;
 152         }
 153         if (ipath_stats.sps_etidfull != dd->ipath_last_tidfull) {
 154                 blen += snprintf(buf + blen, sizeof buf - blen,
 155                                  "%srcvegrfull %llu",
 156                                  blen ? ", " : "",
 157                                  (unsigned long long)
 158                                  (ipath_stats.sps_etidfull -
 159                                   dd->ipath_last_tidfull));
 160                 dd->ipath_last_tidfull = ipath_stats.sps_etidfull;
 161         }
 162
 163         /*
 164          * this is actually the number of hdrq full interrupts, not actual
 165          * events, but at the moment that's mostly what I'm interested in.
 166          * Actual count, etc. is in the counters, if needed.  For production
 167          * users this won't ordinarily be printed.
 168          */
 169
 170         if ((ipath_debug & (__IPATH_PKTDBG | __IPATH_DBG)) &&
 171             ipath_stats.sps_hdrqfull != last_tot_hdrqfull) {
 172                 blen += snprintf(buf + blen, sizeof buf - blen,
 173                                  "%shdrqfull %llu (all ports)",
 174                                  blen ? ", " : "",
 175                                  (unsigned long long)
 176                                  (ipath_stats.sps_hdrqfull -
 177                                   last_tot_hdrqfull));
 178                 last_tot_hdrqfull = ipath_stats.sps_hdrqfull;
 179         }
 180         if (blen)
 181                 ipath_dbg("%s\n", buf);
 182
 183         if (dd->ipath_port0head != (u32)
 184             le64_to_cpu(*dd->ipath_hdrqtailptr)) {
 185                 if (dd->ipath_lastport0rcv_cnt ==
 186                     ipath_stats.sps_port0pkts) {
 187                         ipath_cdbg(PKT, "missing rcv interrupts? "
 188                                    "port0 hd=%llx tl=%x; port0pkts %llx\n",
 189                                    (unsigned long long)
 190                                    le64_to_cpu(*dd->ipath_hdrqtailptr),
 191                                    dd->ipath_port0head,
 192                                    (unsigned long long)
 193                                    ipath_stats.sps_port0pkts);
 194                 }
 195                 dd->ipath_lastport0rcv_cnt = ipath_stats.sps_port0pkts;
 196         }
 197 }
 198
 199 static void ipath_chk_errormask(struct ipath_devdata *dd)
 200 {
 201         static u32 fixed;
 202         u32 ctrl;
 203         unsigned long errormask;
 204         unsigned long hwerrs;
 205
 206         if (!dd->ipath_errormask || !(dd->ipath_flags & IPATH_INITTED))
 207                 return;
 208
 209         errormask = ipath_read_kreg64(dd, dd->ipath_kregs->kr_errormask);
 210
 211         if (errormask == dd->ipath_errormask)
 212                 return;
 213         fixed++;
 214
 215         hwerrs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus);
 216         ctrl = ipath_read_kreg32(dd, dd->ipath_kregs->kr_control);
 217
 218         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 219                 dd->ipath_errormask);
 220
 221         if ((hwerrs & dd->ipath_hwerrmask) ||
 222                 (ctrl & INFINIPATH_C_FREEZEMODE)) {
 223                 /* force re-interrupt of pending events, just in case */
 224                 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
 225                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, 0ULL);
 226                 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
 227                 dev_info(&dd->pcidev->dev,
 228                         "errormask fixed(%u) %lx -> %lx, ctrl %x hwerr %lx\n",
 229                         fixed, errormask, (unsigned long)dd->ipath_errormask,
 230                         ctrl, hwerrs);
 231         } else
 232                 ipath_dbg("errormask fixed(%u) %lx -> %lx, no freeze\n",
 233                         fixed, errormask,
 234                         (unsigned long)dd->ipath_errormask);
 235 }
 236
 237
 238 /**
 239  * ipath_get_faststats - get word counters from chip before they overflow
 240  * @opaque - contains a pointer to the infinipath device ipath_devdata
 241  *
 242  * called from add_timer
 243  */
 244 void ipath_get_faststats(unsigned long opaque)
 245 {
 246         struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
 247         u32 val;
 248         static unsigned cnt;
 249         unsigned long flags;
 250
 251         /*
 252          * don't access the chip while running diags, or memory diags can
 253          * fail
 254          */
 255         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_INITTED) ||
 256             ipath_diag_inuse)
 257                 /* but re-arm the timer, for diags case; won't hurt other */
 258                 goto done;
 259
 260         /*
 261          * We now try to maintain a "active timer", based on traffic
 262          * exceeding a threshold, so we need to check the word-counts
 263          * even if they are 64-bit.
 264          */
 265         ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt);
 266         ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
 267         spin_lock_irqsave(&dd->ipath_eep_st_lock, flags);
 268         if (dd->ipath_traffic_wds  >= IPATH_TRAFFIC_ACTIVE_THRESHOLD)
 269                 atomic_add(5, &dd->ipath_active_time); /* S/B #define */
 270         dd->ipath_traffic_wds = 0;
 271         spin_unlock_irqrestore(&dd->ipath_eep_st_lock, flags);
 272
 273         if (dd->ipath_flags & IPATH_32BITCOUNTERS) {
 274                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
 275                 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
 276         }
 277
 278         ipath_qcheck(dd);
 279
 280         /*
 281          * deal with repeat error suppression.  Doesn't really matter if
 282          * last error was almost a full interval ago, or just a few usecs
 283          * ago; still won't get more than 2 per interval.  We may want
 284          * longer intervals for this eventually, could do with mod, counter
 285          * or separate timer.  Also see code in ipath_handle_errors() and
 286          * ipath_handle_hwerrors().
 287          */
 288
 289         if (dd->ipath_lasterror)
 290                 dd->ipath_lasterror = 0;
 291         if (dd->ipath_lasthwerror)
 292                 dd->ipath_lasthwerror = 0;
 293         if (dd->ipath_maskederrs
 294             && time_after(jiffies, dd->ipath_unmasktime)) {
 295                 char ebuf[256];
 296                 int iserr;
 297                 iserr = ipath_decode_err(ebuf, sizeof ebuf,
 298                         dd->ipath_maskederrs);
 299                 if (dd->ipath_maskederrs &
 300                                 ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
 301                                 INFINIPATH_E_PKTERRS ))
 302                         ipath_dev_err(dd, "Re-enabling masked errors "
 303                                       "(%s)\n", ebuf);
 304                 else {
 305                         /*
 306                          * rcvegrfull and rcvhdrqfull are "normal", for some
 307                          * types of processes (mostly benchmarks) that send
 308                          * huge numbers of messages, while not processing
 309                          * them.  So only complain about these at debug
 310                          * level.
 311                          */
 312                         if (iserr)
 313                                         ipath_dbg("Re-enabling queue full errors (%s)\n",
 314                                                         ebuf);
 315                         else
 316                                 ipath_cdbg(ERRPKT, "Re-enabling packet"
 317                                                 " problem interrupt (%s)\n", ebuf);
 318                 }
 319
 320                 /* re-enable masked errors */
 321                 dd->ipath_errormask |= dd->ipath_maskederrs;
 322                 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
 323                         dd->ipath_errormask);
 324                 dd->ipath_maskederrs = 0;
 325         }
 326
 327         /* limit qfull messages to ~one per minute per port */
 328         if ((++cnt & 0x10)) {
 329                 for (val = dd->ipath_cfgports - 1; ((int)val) >= 0;
 330                      val--) {
 331                         if (dd->ipath_lastegrheads[val] != -1)
 332                                 dd->ipath_lastegrheads[val] = -1;
 333                         if (dd->ipath_lastrcvhdrqtails[val] != -1)
 334                                 dd->ipath_lastrcvhdrqtails[val] = -1;
 335                 }
 336         }
 337
 338         ipath_chk_errormask(dd);
 339 done:
 340         mod_timer(&dd->ipath_stats_timer, jiffies + HZ * 5);
 341 }