1 /*-
2 * Copyright 2003-2011 Netlogic Microsystems (Netlogic). All rights
3 * reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are
7 * met:
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in
13 * the documentation and/or other materials provided with the
14 * distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY Netlogic Microsystems ``AS IS'' AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
19 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NETLOGIC OR CONTRIBUTORS BE
20 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
21 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
22 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
23 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
24 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
25 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
26 * THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * NETLOGIC_BSD */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD: stable/10/sys/mips/nlm/cms.c 233534 2012-03-27 07:47:13Z jchandra $");
32 #include <sys/types.h>
33 #include <sys/systm.h>
34 #include <sys/param.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/proc.h>
38 #include <sys/limits.h>
39 #include <sys/bus.h>
40
41 #include <sys/ktr.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/proc.h>
45 #include <sys/resourcevar.h>
46 #include <sys/sched.h>
47 #include <sys/unistd.h>
48 #include <sys/sysctl.h>
49 #include <sys/malloc.h>
50
51 #include <machine/reg.h>
52 #include <machine/cpu.h>
53 #include <machine/hwfunc.h>
54 #include <machine/mips_opcode.h>
55 #include <machine/param.h>
56 #include <machine/intr_machdep.h>
57
58 #include <mips/nlm/hal/mips-extns.h>
59 #include <mips/nlm/hal/haldefs.h>
60 #include <mips/nlm/hal/iomap.h>
61 #include <mips/nlm/hal/cop2.h>
62 #include <mips/nlm/hal/fmn.h>
63 #include <mips/nlm/hal/pic.h>
64
65 #include <mips/nlm/msgring.h>
66 #include <mips/nlm/interrupt.h>
67 #include <mips/nlm/xlp.h>
68
69 #define MSGRNG_NSTATIONS 1024
70 /*
71 * Keep track of our message ring handler threads, each core has a
72 * different message station. Ideally we will need to start a few
73 * message handling threads every core, and wake them up depending on
74 * load
75 */
76 struct msgring_thread {
77 struct thread *thread; /* msgring handler threads */
78 int needed; /* thread needs to wake up */
79 };
80 static struct msgring_thread msgring_threads[XLP_MAX_CORES * XLP_MAX_THREADS];
81 static struct proc *msgring_proc; /* all threads are under a proc */
82
83 /*
84 * The device drivers can register a handler for the messages sent
85 * from a station (corresponding to the device).
86 */
87 struct tx_stn_handler {
88 msgring_handler action;
89 void *arg;
90 };
91 static struct tx_stn_handler msgmap[MSGRNG_NSTATIONS];
92 static struct mtx msgmap_lock;
93 uint32_t xlp_msg_thread_mask;
94 static int xlp_msg_threads_per_core = XLP_MAX_THREADS;
95
96 static void create_msgring_thread(int hwtid);
97 static int msgring_process_fast_intr(void *arg);
98
99 /* Debug counters */
100 static int msgring_nintr[XLP_MAX_CORES * XLP_MAX_THREADS];
101 static int msgring_wakeup_sleep[XLP_MAX_CORES * XLP_MAX_THREADS];
102 static int msgring_wakeup_nosleep[XLP_MAX_CORES * XLP_MAX_THREADS];
103 static int fmn_msgcount[XLP_MAX_CORES * XLP_MAX_THREADS][4];
104 static int fmn_loops[XLP_MAX_CORES * XLP_MAX_THREADS];
105
106 /* Whether polled driver implementation */
107 static int polled = 0;
108
109 /* We do only i/o device credit setup here. CPU credit setup is now
110 * moved to xlp_msgring_cpu_init() so that the credits get setup
111 * only if the CPU exists. xlp_msgring_cpu_init() gets called from
112 * platform_init_ap; and this makes it easy for us to setup CMS
113 * credits for various types of XLP chips, with varying number of
114 * cpu's and cores.
115 */
116 static void
xlp_cms_credit_setup(int credit)117 xlp_cms_credit_setup(int credit)
118 {
119 uint64_t cmspcibase, cmsbase, pcibase;
120 uint32_t devoffset;
121 int dev, fn, maxqid;
122 int src, qid, i;
123
124 for (i = 0; i < XLP_MAX_NODES; i++) {
125 cmspcibase = nlm_get_cms_pcibase(i);
126 if (!nlm_dev_exists(XLP_IO_CMS_OFFSET(i)))
127 continue;
128 cmsbase = nlm_get_cms_regbase(i);
129 maxqid = nlm_read_reg(cmspcibase, XLP_PCI_DEVINFO_REG0);
130 for (dev = 0; dev < 8; dev++) {
131 for (fn = 0; fn < 8; fn++) {
132 devoffset = XLP_HDR_OFFSET(i, 0, dev, fn);
133 if (nlm_dev_exists(devoffset) == 0)
134 continue;
135 pcibase = nlm_pcicfg_base(devoffset);
136 src = nlm_qidstart(pcibase);
137 if (src == 0)
138 continue;
139 #if 0 /* Debug */
140 printf("Setup CMS credits for queues ");
141 printf("[%d to %d] from src %d\n", 0,
142 maxqid, src);
143 #endif
144 for (qid = 0; qid < maxqid; qid++)
145 nlm_cms_setup_credits(cmsbase, qid,
146 src, credit);
147 }
148 }
149 }
150 }
151
152 void
xlp_msgring_cpu_init(int node,int cpu,int credit)153 xlp_msgring_cpu_init(int node, int cpu, int credit)
154 {
155 uint64_t cmspcibase = nlm_get_cms_pcibase(node);
156 uint64_t cmsbase = nlm_get_cms_regbase(node);
157 int qid, maxqid, src;
158
159 maxqid = nlm_read_reg(cmspcibase, XLP_PCI_DEVINFO_REG0);
160
161 /* cpu credit setup is done only from thread-0 of each core */
162 if((cpu % 4) == 0) {
163 src = cpu << 2; /* each thread has 4 vc's */
164 for (qid = 0; qid < maxqid; qid++)
165 nlm_cms_setup_credits(cmsbase, qid, src, credit);
166 }
167 }
168
169 /*
170 * Drain out max_messages for the buckets set in the bucket mask.
171 * Use max_msgs = 0 to drain out all messages.
172 */
173 int
xlp_handle_msg_vc(u_int vcmask,int max_msgs)174 xlp_handle_msg_vc(u_int vcmask, int max_msgs)
175 {
176 struct nlm_fmn_msg msg;
177 int srcid = 0, size = 0, code = 0;
178 struct tx_stn_handler *he;
179 uint32_t mflags, status;
180 int n_msgs = 0, vc, m, hwtid;
181 u_int msgmask;
182
183 hwtid = nlm_cpuid();
184 for (;;) {
185 /* check if VC empty */
186 mflags = nlm_save_flags_cop2();
187 status = nlm_read_c2_msgstatus1();
188 nlm_restore_flags(mflags);
189
190 msgmask = ((status >> 24) & 0xf) ^ 0xf;
191 msgmask &= vcmask;
192 if (msgmask == 0)
193 break;
194 m = 0;
195 for (vc = 0; vc < 4; vc++) {
196 if ((msgmask & (1 << vc)) == 0)
197 continue;
198
199 mflags = nlm_save_flags_cop2();
200 status = nlm_fmn_msgrcv(vc, &srcid, &size, &code,
201 &msg);
202 nlm_restore_flags(mflags);
203 if (status != 0) /* no msg or error */
204 continue;
205 if (srcid < 0 && srcid >= 1024) {
206 printf("[%s]: bad src id %d\n", __func__,
207 srcid);
208 continue;
209 }
210 he = &msgmap[srcid];
211 if(he->action != NULL)
212 (he->action)(vc, size, code, srcid, &msg,
213 he->arg);
214 #if 0
215 else
216 printf("[%s]: No Handler for msg from stn %d,"
217 " vc=%d, size=%d, msg0=%jx, droppinge\n",
218 __func__, srcid, vc, size,
219 (uintmax_t)msg.msg[0]);
220 #endif
221 fmn_msgcount[hwtid][vc] += 1;
222 m++; /* msgs handled in this iter */
223 }
224 if (m == 0)
225 break; /* nothing done in this iter */
226 n_msgs += m;
227 if (max_msgs > 0 && n_msgs >= max_msgs)
228 break;
229 }
230
231 return (n_msgs);
232 }
233
234 static void
xlp_discard_msg_vc(u_int vcmask)235 xlp_discard_msg_vc(u_int vcmask)
236 {
237 struct nlm_fmn_msg msg;
238 int srcid = 0, size = 0, code = 0, vc;
239 uint32_t mflags, status;
240
241 for (vc = 0; vc < 4; vc++) {
242 for (;;) {
243 mflags = nlm_save_flags_cop2();
244 status = nlm_fmn_msgrcv(vc, &srcid,
245 &size, &code, &msg);
246 nlm_restore_flags(mflags);
247
248 /* break if there is no msg or error */
249 if (status != 0)
250 break;
251 }
252 }
253 }
254
255 void
xlp_cms_enable_intr(int node,int cpu,int type,int watermark)256 xlp_cms_enable_intr(int node, int cpu, int type, int watermark)
257 {
258 uint64_t cmsbase;
259 int i, qid;
260
261 cmsbase = nlm_get_cms_regbase(node);
262
263 for (i = 0; i < 4; i++) {
264 qid = (i + (cpu * 4)) & 0x7f;
265 nlm_cms_per_queue_level_intr(cmsbase, qid, type, watermark);
266 nlm_cms_per_queue_timer_intr(cmsbase, qid, 0x1, 0);
267 }
268 }
269
270 static int
msgring_process_fast_intr(void * arg)271 msgring_process_fast_intr(void *arg)
272 {
273 struct msgring_thread *mthd;
274 struct thread *td;
275 int cpu;
276
277 cpu = nlm_cpuid();
278 mthd = &msgring_threads[cpu];
279 msgring_nintr[cpu]++;
280 td = mthd->thread;
281
282 /* clear pending interrupts */
283 nlm_write_c0_eirr(1ULL << IRQ_MSGRING);
284
285 /* wake up the target thread */
286 mthd->needed = 1;
287 thread_lock(td);
288 if (TD_AWAITING_INTR(td)) {
289 msgring_wakeup_sleep[cpu]++;
290 TD_CLR_IWAIT(td);
291 sched_add(td, SRQ_INTR);
292 } else
293 msgring_wakeup_nosleep[cpu]++;
294
295 thread_unlock(td);
296
297 return (FILTER_HANDLED);
298 }
299
300 static void
msgring_process(void * arg)301 msgring_process(void * arg)
302 {
303 volatile struct msgring_thread *mthd;
304 struct thread *td;
305 uint32_t mflags, msgstatus1;
306 int hwtid, nmsgs;
307
308 hwtid = (intptr_t)arg;
309 mthd = &msgring_threads[hwtid];
310 td = mthd->thread;
311 KASSERT(curthread == td,
312 ("%s:msg_ithread and proc linkage out of sync", __func__));
313
314 /* First bind this thread to the right CPU */
315 thread_lock(td);
316 sched_bind(td, xlp_hwtid_to_cpuid[hwtid]);
317 thread_unlock(td);
318
319 if (hwtid != nlm_cpuid())
320 printf("Misscheduled hwtid %d != cpuid %d\n", hwtid,
321 nlm_cpuid());
322
323 xlp_discard_msg_vc(0xf);
324 xlp_msgring_cpu_init(nlm_nodeid(), nlm_cpuid(), CMS_DEFAULT_CREDIT);
325 if (polled == 0) {
326 mflags = nlm_save_flags_cop2();
327 nlm_fmn_cpu_init(IRQ_MSGRING, 0, 0, 0, 0, 0);
328 nlm_restore_flags(mflags);
329 xlp_cms_enable_intr(nlm_nodeid(), nlm_cpuid(), 0x2, 0);
330 /* clear pending interrupts.
331 * they will get re-raised if still valid */
332 nlm_write_c0_eirr(1ULL << IRQ_MSGRING);
333 }
334
335 /* start processing messages */
336 for (;;) {
337 atomic_store_rel_int(&mthd->needed, 0);
338 nmsgs = xlp_handle_msg_vc(0xf, 0);
339
340 /* sleep */
341 if (polled == 0) {
342 /* clear VC-pend bits */
343 mflags = nlm_save_flags_cop2();
344 msgstatus1 = nlm_read_c2_msgstatus1();
345 msgstatus1 |= (0xf << 16);
346 nlm_write_c2_msgstatus1(msgstatus1);
347 nlm_restore_flags(mflags);
348
349 thread_lock(td);
350 if (mthd->needed) {
351 thread_unlock(td);
352 continue;
353 }
354 sched_class(td, PRI_ITHD);
355 TD_SET_IWAIT(td);
356 mi_switch(SW_VOL, NULL);
357 thread_unlock(td);
358 } else
359 pause("wmsg", 1);
360
361 fmn_loops[hwtid]++;
362 }
363 }
364
365 static void
create_msgring_thread(int hwtid)366 create_msgring_thread(int hwtid)
367 {
368 struct msgring_thread *mthd;
369 struct thread *td;
370 int error;
371
372 mthd = &msgring_threads[hwtid];
373 error = kproc_kthread_add(msgring_process, (void *)(uintptr_t)hwtid,
374 &msgring_proc, &td, RFSTOPPED, 2, "msgrngproc",
375 "msgthr%d", hwtid);
376 if (error)
377 panic("kproc_kthread_add() failed with %d", error);
378 mthd->thread = td;
379
380 thread_lock(td);
381 sched_class(td, PRI_ITHD);
382 sched_add(td, SRQ_INTR);
383 thread_unlock(td);
384 }
385
386 int
register_msgring_handler(int startb,int endb,msgring_handler action,void * arg)387 register_msgring_handler(int startb, int endb, msgring_handler action,
388 void *arg)
389 {
390 int i;
391
392 if (bootverbose)
393 printf("Register handler %d-%d %p(%p)\n",
394 startb, endb, action, arg);
395 KASSERT(startb >= 0 && startb <= endb && endb < MSGRNG_NSTATIONS,
396 ("Invalid value for bucket range %d,%d", startb, endb));
397
398 mtx_lock_spin(&msgmap_lock);
399 for (i = startb; i <= endb; i++) {
400 KASSERT(msgmap[i].action == NULL,
401 ("Bucket %d already used [action %p]", i, msgmap[i].action));
402 msgmap[i].action = action;
403 msgmap[i].arg = arg;
404 }
405 mtx_unlock_spin(&msgmap_lock);
406 return (0);
407 }
408
409 /*
410 * Initialize the messaging subsystem.
411 *
412 * Message Stations are shared among all threads in a cpu core, this
413 * has to be called once from every core which is online.
414 */
415 static void
xlp_msgring_config(void * arg)416 xlp_msgring_config(void *arg)
417 {
418 void *cookie;
419 unsigned int thrmask, mask;
420 int i;
421
422 /* used polled handler for Ax silion */
423 if (nlm_is_xlp8xx_ax())
424 polled = 1;
425
426 /* Don't poll on all threads, if polled */
427 if (polled)
428 xlp_msg_threads_per_core -= 1;
429
430 mtx_init(&msgmap_lock, "msgring", NULL, MTX_SPIN);
431 if (xlp_threads_per_core < xlp_msg_threads_per_core)
432 xlp_msg_threads_per_core = xlp_threads_per_core;
433 thrmask = ((1 << xlp_msg_threads_per_core) - 1);
434 mask = 0;
435 for (i = 0; i < XLP_MAX_CORES; i++) {
436 mask <<= XLP_MAX_THREADS;
437 mask |= thrmask;
438 }
439 xlp_msg_thread_mask = xlp_hw_thread_mask & mask;
440 #if 0
441 printf("CMS Message handler thread mask %#jx\n",
442 (uintmax_t)xlp_msg_thread_mask);
443 #endif
444 xlp_cms_credit_setup(CMS_DEFAULT_CREDIT);
445 create_msgring_thread(0);
446 cpu_establish_hardintr("msgring", msgring_process_fast_intr, NULL,
447 NULL, IRQ_MSGRING, INTR_TYPE_NET, &cookie);
448 }
449
450 /*
451 * Start message ring processing threads on other CPUs, after SMP start
452 */
453 static void
start_msgring_threads(void * arg)454 start_msgring_threads(void *arg)
455 {
456 int hwt;
457
458 for (hwt = 1; hwt < XLP_MAX_CORES * XLP_MAX_THREADS; hwt++) {
459 if ((xlp_msg_thread_mask & (1 << hwt)) == 0)
460 continue;
461 create_msgring_thread(hwt);
462 }
463 }
464
465 SYSINIT(xlp_msgring_config, SI_SUB_DRIVERS, SI_ORDER_FIRST,
466 xlp_msgring_config, NULL);
467 SYSINIT(start_msgring_threads, SI_SUB_SMP, SI_ORDER_MIDDLE,
468 start_msgring_threads, NULL);
469
470 /*
471 * DEBUG support, XXX: static buffer, not locked
472 */
473 static int
sys_print_debug(SYSCTL_HANDLER_ARGS)474 sys_print_debug(SYSCTL_HANDLER_ARGS)
475 {
476 int error, nb, i, fs;
477 static char xprintb[4096], *buf;
478
479 buf = xprintb;
480 fs = sizeof(xprintb);
481 nb = snprintf(buf, fs,
482 "\nID vc0 vc1 vc2 vc3 loops\n");
483 buf += nb;
484 fs -= nb;
485 for (i = 0; i < 32; i++) {
486 if ((xlp_hw_thread_mask & (1 << i)) == 0)
487 continue;
488 nb = snprintf(buf, fs,
489 "%2d: %8d %8d %8d %8d %8d\n", i,
490 fmn_msgcount[i][0], fmn_msgcount[i][1],
491 fmn_msgcount[i][2], fmn_msgcount[i][3],
492 fmn_loops[i]);
493 buf += nb;
494 fs -= nb;
495 }
496 error = SYSCTL_OUT(req, xprintb, buf - xprintb);
497 return (error);
498 }
499
500 SYSCTL_PROC(_debug, OID_AUTO, msgring, CTLTYPE_STRING | CTLFLAG_RD, 0, 0,
501 sys_print_debug, "A", "msgring debug info");
502