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
ipath_snap_cntr(struct ipath_devdata * dd,ipath_creg creg)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 if (creg == dd->ipath_cregs->cr_ibsymbolerrcnt) {
116 if (dd->ibdeltainprog)
117 val64 -= val64 - dd->ibsymsnap;
118 val64 -= dd->ibsymdelta;
119 } else if (creg == dd->ipath_cregs->cr_iblinkerrrecovcnt) {
120 if (dd->ibdeltainprog)
121 val64 -= val64 - dd->iblnkerrsnap;
122 val64 -= dd->iblnkerrdelta;
123 } else
124 val64 = (u64) val;
125
126 ret = val64;
127
128 bail:
129 return ret;
130 }
131
132 /**
133 * ipath_qcheck - print delta of egrfull/hdrqfull errors for kernel ports
134 * @dd: the infinipath device
135 *
136 * print the delta of egrfull/hdrqfull errors for kernel ports no more than
137 * every 5 seconds. User processes are printed at close, but kernel doesn't
138 * close, so... Separate routine so may call from other places someday, and
139 * so function name when printed by _IPATH_INFO is meaningfull
140 */
ipath_qcheck(struct ipath_devdata * dd)141 static void ipath_qcheck(struct ipath_devdata *dd)
142 {
143 static u64 last_tot_hdrqfull;
144 struct ipath_portdata *pd = dd->ipath_pd[0];
145 size_t blen = 0;
146 char buf[128];
147 u32 hdrqtail;
148
149 *buf = 0;
150 if (pd->port_hdrqfull != dd->ipath_p0_hdrqfull) {
151 blen = snprintf(buf, sizeof buf, "port 0 hdrqfull %u",
152 pd->port_hdrqfull -
153 dd->ipath_p0_hdrqfull);
154 dd->ipath_p0_hdrqfull = pd->port_hdrqfull;
155 }
156 if (ipath_stats.sps_etidfull != dd->ipath_last_tidfull) {
157 blen += snprintf(buf + blen, sizeof buf - blen,
158 "%srcvegrfull %llu",
159 blen ? ", " : "",
160 (unsigned long long)
161 (ipath_stats.sps_etidfull -
162 dd->ipath_last_tidfull));
163 dd->ipath_last_tidfull = ipath_stats.sps_etidfull;
164 }
165
166 /*
167 * this is actually the number of hdrq full interrupts, not actual
168 * events, but at the moment that's mostly what I'm interested in.
169 * Actual count, etc. is in the counters, if needed. For production
170 * users this won't ordinarily be printed.
171 */
172
173 if ((ipath_debug & (__IPATH_PKTDBG | __IPATH_DBG)) &&
174 ipath_stats.sps_hdrqfull != last_tot_hdrqfull) {
175 blen += snprintf(buf + blen, sizeof buf - blen,
176 "%shdrqfull %llu (all ports)",
177 blen ? ", " : "",
178 (unsigned long long)
179 (ipath_stats.sps_hdrqfull -
180 last_tot_hdrqfull));
181 last_tot_hdrqfull = ipath_stats.sps_hdrqfull;
182 }
183 if (blen)
184 ipath_dbg("%s\n", buf);
185
186 hdrqtail = ipath_get_hdrqtail(pd);
187 if (pd->port_head != hdrqtail) {
188 if (dd->ipath_lastport0rcv_cnt ==
189 ipath_stats.sps_port0pkts) {
190 ipath_cdbg(PKT, "missing rcv interrupts? "
191 "port0 hd=%x tl=%x; port0pkts %llx; write"
192 " hd (w/intr)\n",
193 pd->port_head, hdrqtail,
194 (unsigned long long)
195 ipath_stats.sps_port0pkts);
196 ipath_write_ureg(dd, ur_rcvhdrhead, hdrqtail |
197 dd->ipath_rhdrhead_intr_off, pd->port_port);
198 }
199 dd->ipath_lastport0rcv_cnt = ipath_stats.sps_port0pkts;
200 }
201 }
202
ipath_chk_errormask(struct ipath_devdata * dd)203 static void ipath_chk_errormask(struct ipath_devdata *dd)
204 {
205 static u32 fixed;
206 u32 ctrl;
207 unsigned long errormask;
208 unsigned long hwerrs;
209
210 if (!dd->ipath_errormask || !(dd->ipath_flags & IPATH_INITTED))
211 return;
212
213 errormask = ipath_read_kreg64(dd, dd->ipath_kregs->kr_errormask);
214
215 if (errormask == dd->ipath_errormask)
216 return;
217 fixed++;
218
219 hwerrs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus);
220 ctrl = ipath_read_kreg32(dd, dd->ipath_kregs->kr_control);
221
222 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
223 dd->ipath_errormask);
224
225 if ((hwerrs & dd->ipath_hwerrmask) ||
226 (ctrl & INFINIPATH_C_FREEZEMODE)) {
227 /* force re-interrupt of pending events, just in case */
228 ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear, 0ULL);
229 ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, 0ULL);
230 ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
231 dev_info(&dd->pcidev->dev,
232 "errormask fixed(%u) %lx -> %lx, ctrl %x hwerr %lx\n",
233 fixed, errormask, (unsigned long)dd->ipath_errormask,
234 ctrl, hwerrs);
235 } else
236 ipath_dbg("errormask fixed(%u) %lx -> %lx, no freeze\n",
237 fixed, errormask,
238 (unsigned long)dd->ipath_errormask);
239 }
240
241
242 /**
243 * ipath_get_faststats - get word counters from chip before they overflow
244 * @opaque - contains a pointer to the infinipath device ipath_devdata
245 *
246 * called from add_timer
247 */
ipath_get_faststats(unsigned long opaque)248 void ipath_get_faststats(unsigned long opaque)
249 {
250 struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
251 int i;
252 static unsigned cnt;
253 unsigned long flags;
254 u64 traffic_wds;
255
256 /*
257 * don't access the chip while running diags, or memory diags can
258 * fail
259 */
260 if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_INITTED) ||
261 ipath_diag_inuse)
262 /* but re-arm the timer, for diags case; won't hurt other */
263 goto done;
264
265 /*
266 * We now try to maintain a "active timer", based on traffic
267 * exceeding a threshold, so we need to check the word-counts
268 * even if they are 64-bit.
269 */
270 traffic_wds = ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordsendcnt) +
271 ipath_snap_cntr(dd, dd->ipath_cregs->cr_wordrcvcnt);
272 spin_lock_irqsave(&dd->ipath_eep_st_lock, flags);
273 traffic_wds -= dd->ipath_traffic_wds;
274 dd->ipath_traffic_wds += traffic_wds;
275 if (traffic_wds >= IPATH_TRAFFIC_ACTIVE_THRESHOLD)
276 atomic_add(5, &dd->ipath_active_time); /* S/B #define */
277 spin_unlock_irqrestore(&dd->ipath_eep_st_lock, flags);
278
279 if (dd->ipath_flags & IPATH_32BITCOUNTERS) {
280 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktsendcnt);
281 ipath_snap_cntr(dd, dd->ipath_cregs->cr_pktrcvcnt);
282 }
283
284 ipath_qcheck(dd);
285
286 /*
287 * deal with repeat error suppression. Doesn't really matter if
288 * last error was almost a full interval ago, or just a few usecs
289 * ago; still won't get more than 2 per interval. We may want
290 * longer intervals for this eventually, could do with mod, counter
291 * or separate timer. Also see code in ipath_handle_errors() and
292 * ipath_handle_hwerrors().
293 */
294
295 if (dd->ipath_lasterror)
296 dd->ipath_lasterror = 0;
297 if (dd->ipath_lasthwerror)
298 dd->ipath_lasthwerror = 0;
299 if (dd->ipath_maskederrs
300 && time_after(jiffies, dd->ipath_unmasktime)) {
301 char ebuf[256];
302 int iserr;
303 iserr = ipath_decode_err(dd, ebuf, sizeof ebuf,
304 dd->ipath_maskederrs);
305 if (dd->ipath_maskederrs &
306 ~(INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
307 INFINIPATH_E_PKTERRS))
308 ipath_dev_err(dd, "Re-enabling masked errors "
309 "(%s)\n", ebuf);
310 else {
311 /*
312 * rcvegrfull and rcvhdrqfull are "normal", for some
313 * types of processes (mostly benchmarks) that send
314 * huge numbers of messages, while not processing
315 * them. So only complain about these at debug
316 * level.
317 */
318 if (iserr)
319 ipath_dbg(
320 "Re-enabling queue full errors (%s)\n",
321 ebuf);
322 else
323 ipath_cdbg(ERRPKT, "Re-enabling packet"
324 " problem interrupt (%s)\n", ebuf);
325 }
326
327 /* re-enable masked errors */
328 dd->ipath_errormask |= dd->ipath_maskederrs;
329 ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
330 dd->ipath_errormask);
331 dd->ipath_maskederrs = 0;
332 }
333
334 /* limit qfull messages to ~one per minute per port */
335 if ((++cnt & 0x10)) {
336 for (i = (int) dd->ipath_cfgports; --i >= 0; ) {
337 struct ipath_portdata *pd = dd->ipath_pd[i];
338
339 if (pd && pd->port_lastrcvhdrqtail != -1)
340 pd->port_lastrcvhdrqtail = -1;
341 }
342 }
343
344 ipath_chk_errormask(dd);
345 done:
346 mod_timer(&dd->ipath_stats_timer, jiffies + HZ * 5);
347 }
348