xref: /f-stack/freebsd/mips/nlm/cms.c (revision 22ce4aff)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright 2003-2011 Netlogic Microsystems (Netlogic). All rights
5  * reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions are
9  * met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY Netlogic Microsystems ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
21  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NETLOGIC OR CONTRIBUTORS BE
22  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
23  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
24  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
25  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
26  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
27  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
28  * THE POSSIBILITY OF SUCH DAMAGE.
29  *
30  * NETLOGIC_BSD */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 #include <sys/types.h>
35 #include <sys/systm.h>
36 #include <sys/param.h>
37 #include <sys/lock.h>
38 #include <sys/mutex.h>
39 #include <sys/proc.h>
40 #include <sys/limits.h>
41 #include <sys/bus.h>
42 #include <sys/sbuf.h>
43 
44 #include <sys/ktr.h>
45 #include <sys/kernel.h>
46 #include <sys/kthread.h>
47 #include <sys/proc.h>
48 #include <sys/resourcevar.h>
49 #include <sys/sched.h>
50 #include <sys/unistd.h>
51 #include <sys/sysctl.h>
52 #include <sys/malloc.h>
53 
54 #include <machine/reg.h>
55 #include <machine/cpu.h>
56 #include <machine/hwfunc.h>
57 #include <machine/mips_opcode.h>
58 #include <machine/intr_machdep.h>
59 
60 #include <mips/nlm/hal/mips-extns.h>
61 #include <mips/nlm/hal/haldefs.h>
62 #include <mips/nlm/hal/iomap.h>
63 #include <mips/nlm/hal/cop2.h>
64 #include <mips/nlm/hal/fmn.h>
65 #include <mips/nlm/hal/pic.h>
66 
67 #include <mips/nlm/msgring.h>
68 #include <mips/nlm/interrupt.h>
69 #include <mips/nlm/xlp.h>
70 
71 #define	MSGRNG_NSTATIONS	1024
72 /*
73  * Keep track of our message ring handler threads, each core has a
74  * different message station. Ideally we will need to start a few
75  * message handling threads every core, and wake them up depending on
76  * load
77  */
78 struct msgring_thread {
79 	struct thread	*thread;	/* msgring handler threads */
80 	int	needed;			/* thread needs to wake up */
81 };
82 static struct msgring_thread msgring_threads[XLP_MAX_CORES * XLP_MAX_THREADS];
83 static struct proc *msgring_proc;	/* all threads are under a proc */
84 
85 /*
86  * The device drivers can register a handler for the messages sent
87  * from a station (corresponding to the device).
88  */
89 struct tx_stn_handler {
90 	msgring_handler action;
91 	void *arg;
92 };
93 static struct tx_stn_handler msgmap[MSGRNG_NSTATIONS];
94 static struct mtx	msgmap_lock;
95 uint32_t xlp_msg_thread_mask;
96 static int xlp_msg_threads_per_core = XLP_MAX_THREADS;
97 
98 static void create_msgring_thread(int hwtid);
99 static int msgring_process_fast_intr(void *arg);
100 
101 /* Debug counters */
102 static int msgring_nintr[XLP_MAX_CORES * XLP_MAX_THREADS];
103 static int msgring_wakeup_sleep[XLP_MAX_CORES * XLP_MAX_THREADS];
104 static int msgring_wakeup_nosleep[XLP_MAX_CORES * XLP_MAX_THREADS];
105 static int fmn_msgcount[XLP_MAX_CORES * XLP_MAX_THREADS][4];
106 static int fmn_loops[XLP_MAX_CORES * XLP_MAX_THREADS];
107 
108 /* Whether polled driver implementation */
109 static int polled = 0;
110 
111 /* We do only i/o device credit setup here. CPU credit setup is now
112  * moved to xlp_msgring_cpu_init() so that the credits get setup
113  * only if the CPU exists. xlp_msgring_cpu_init() gets called from
114  * platform_init_ap; and this makes it easy for us to setup CMS
115  * credits for various types of XLP chips, with varying number of
116  * cpu's and cores.
117  */
118 static void
xlp_cms_credit_setup(int credit)119 xlp_cms_credit_setup(int credit)
120 {
121 	uint64_t cmspcibase, cmsbase, pcibase;
122 	uint32_t devoffset;
123 	int dev, fn, maxqid;
124 	int src, qid, i;
125 
126 	for (i = 0; i < XLP_MAX_NODES; i++) {
127 		cmspcibase = nlm_get_cms_pcibase(i);
128 		if (!nlm_dev_exists(XLP_IO_CMS_OFFSET(i)))
129 			continue;
130 		cmsbase = nlm_get_cms_regbase(i);
131 		maxqid = nlm_read_reg(cmspcibase, XLP_PCI_DEVINFO_REG0);
132 		for (dev = 0; dev < 8; dev++) {
133 			for (fn = 0; fn < 8; fn++) {
134 				devoffset = XLP_HDR_OFFSET(i, 0, dev, fn);
135 				if (nlm_dev_exists(devoffset) == 0)
136 					continue;
137 				pcibase = nlm_pcicfg_base(devoffset);
138 				src = nlm_qidstart(pcibase);
139 				if (src == 0)
140 					continue;
141 #if 0 /* Debug */
142 				printf("Setup CMS credits for queues ");
143 				printf("[%d to %d] from src %d\n", 0,
144 				    maxqid, src);
145 #endif
146 				for (qid = 0; qid < maxqid; qid++)
147 					nlm_cms_setup_credits(cmsbase, qid,
148 					    src, credit);
149 			}
150 		}
151 	}
152 }
153 
154 void
xlp_msgring_cpu_init(int node,int cpu,int credit)155 xlp_msgring_cpu_init(int node, int cpu, int credit)
156 {
157 	uint64_t cmspcibase = nlm_get_cms_pcibase(node);
158 	uint64_t cmsbase = nlm_get_cms_regbase(node);
159 	int qid, maxqid, src;
160 
161 	maxqid = nlm_read_reg(cmspcibase, XLP_PCI_DEVINFO_REG0);
162 
163 	/* cpu credit setup is done only from thread-0 of each core */
164 	if((cpu % 4) == 0) {
165 		src = cpu << 2; /* each thread has 4 vc's */
166 		for (qid = 0; qid < maxqid; qid++)
167 			nlm_cms_setup_credits(cmsbase, qid, src, credit);
168 	}
169 }
170 
171 /*
172  * Drain out max_messages for the buckets set in the bucket mask.
173  * Use max_msgs = 0 to drain out all messages.
174  */
175 int
xlp_handle_msg_vc(u_int vcmask,int max_msgs)176 xlp_handle_msg_vc(u_int vcmask, int max_msgs)
177 {
178 	struct nlm_fmn_msg msg;
179 	int srcid = 0, size = 0, code = 0;
180 	struct tx_stn_handler *he;
181 	uint32_t mflags, status;
182 	int n_msgs = 0, vc, m, hwtid;
183 	u_int msgmask;
184 
185 	hwtid = nlm_cpuid();
186 	for (;;) {
187 		/* check if VC empty */
188 		mflags = nlm_save_flags_cop2();
189 		status = nlm_read_c2_msgstatus1();
190 		nlm_restore_flags(mflags);
191 
192 		msgmask = ((status >> 24) & 0xf) ^ 0xf;
193 		msgmask &= vcmask;
194 		if (msgmask == 0)
195 			    break;
196 		m = 0;
197 		for (vc = 0; vc < 4; vc++) {
198 			if ((msgmask & (1 << vc)) == 0)
199 				continue;
200 
201 			mflags = nlm_save_flags_cop2();
202 			status = nlm_fmn_msgrcv(vc, &srcid, &size, &code,
203 			    &msg);
204 			nlm_restore_flags(mflags);
205 			if (status != 0)	/*  no msg or error */
206 				continue;
207 			if (srcid < 0 || srcid >= 1024) {
208 				printf("[%s]: bad src id %d\n", __func__,
209 				    srcid);
210 				continue;
211 			}
212 			he = &msgmap[srcid];
213 			if(he->action != NULL)
214 				(he->action)(vc, size, code, srcid, &msg,
215 				he->arg);
216 #if 0
217 			else
218 				printf("[%s]: No Handler for msg from stn %d,"
219 				    " vc=%d, size=%d, msg0=%jx, droppinge\n",
220 				    __func__, srcid, vc, size,
221 				    (uintmax_t)msg.msg[0]);
222 #endif
223 			fmn_msgcount[hwtid][vc] += 1;
224 			m++;	/* msgs handled in this iter */
225 		}
226 		if (m == 0)
227 			break;	/* nothing done in this iter */
228 		n_msgs += m;
229 		if (max_msgs > 0 && n_msgs >= max_msgs)
230 			break;
231 	}
232 
233 	return (n_msgs);
234 }
235 
236 static void
xlp_discard_msg_vc(u_int vcmask)237 xlp_discard_msg_vc(u_int vcmask)
238 {
239 	struct nlm_fmn_msg msg;
240 	int srcid = 0, size = 0, code = 0, vc;
241 	uint32_t mflags, status;
242 
243 	for (vc = 0; vc < 4; vc++) {
244 		for (;;) {
245 			mflags = nlm_save_flags_cop2();
246 			status = nlm_fmn_msgrcv(vc, &srcid,
247 			    &size, &code, &msg);
248 			nlm_restore_flags(mflags);
249 
250 			/* break if there is no msg or error */
251 			if (status != 0)
252 				break;
253 		}
254 	}
255 }
256 
257 void
xlp_cms_enable_intr(int node,int cpu,int type,int watermark)258 xlp_cms_enable_intr(int node, int cpu, int type, int watermark)
259 {
260 	uint64_t cmsbase;
261 	int i, qid;
262 
263 	cmsbase = nlm_get_cms_regbase(node);
264 
265 	for (i = 0; i < 4; i++) {
266 		qid = (i + (cpu * 4)) & 0x7f;
267 		nlm_cms_per_queue_level_intr(cmsbase, qid, type, watermark);
268 		nlm_cms_per_queue_timer_intr(cmsbase, qid, 0x1, 0);
269 	}
270 }
271 
272 static int
msgring_process_fast_intr(void * arg)273 msgring_process_fast_intr(void *arg)
274 {
275 	struct msgring_thread *mthd;
276 	struct thread *td;
277 	int	cpu;
278 
279 	cpu = nlm_cpuid();
280 	mthd = &msgring_threads[cpu];
281 	msgring_nintr[cpu]++;
282 	td = mthd->thread;
283 
284 	/* clear pending interrupts */
285 	nlm_write_c0_eirr(1ULL << IRQ_MSGRING);
286 
287 	/* wake up the target thread */
288 	mthd->needed = 1;
289 	thread_lock(td);
290 	if (TD_AWAITING_INTR(td)) {
291 		msgring_wakeup_sleep[cpu]++;
292 		TD_CLR_IWAIT(td);
293 		sched_add(td, SRQ_INTR);
294 	} else {
295 		thread_unlock(td);
296 		msgring_wakeup_nosleep[cpu]++;
297 	}
298 
299 	return (FILTER_HANDLED);
300 }
301 
302 static void
msgring_process(void * arg)303 msgring_process(void * arg)
304 {
305 	volatile struct msgring_thread *mthd;
306 	struct thread *td;
307 	uint32_t mflags, msgstatus1;
308 	int hwtid, nmsgs;
309 
310 	hwtid = (intptr_t)arg;
311 	mthd = &msgring_threads[hwtid];
312 	td = mthd->thread;
313 	KASSERT(curthread == td,
314 	    ("%s:msg_ithread and proc linkage out of sync", __func__));
315 
316 	/* First bind this thread to the right CPU */
317 	thread_lock(td);
318 	sched_bind(td, xlp_hwtid_to_cpuid[hwtid]);
319 	thread_unlock(td);
320 
321 	if (hwtid != nlm_cpuid())
322 		printf("Misscheduled hwtid %d != cpuid %d\n", hwtid,
323 		    nlm_cpuid());
324 
325 	xlp_discard_msg_vc(0xf);
326 	xlp_msgring_cpu_init(nlm_nodeid(), nlm_cpuid(), CMS_DEFAULT_CREDIT);
327 	if (polled == 0) {
328 		mflags = nlm_save_flags_cop2();
329 		nlm_fmn_cpu_init(IRQ_MSGRING, 0, 0, 0, 0, 0);
330 		nlm_restore_flags(mflags);
331 		xlp_cms_enable_intr(nlm_nodeid(), nlm_cpuid(), 0x2, 0);
332 		/* clear pending interrupts.
333 		 *  they will get re-raised if still valid */
334 		nlm_write_c0_eirr(1ULL << IRQ_MSGRING);
335 	}
336 
337 	/* start processing messages */
338 	for (;;) {
339 		atomic_store_rel_int(&mthd->needed, 0);
340 		nmsgs = xlp_handle_msg_vc(0xf, 0);
341 
342 		/* sleep */
343 		if (polled == 0) {
344 			/* clear VC-pend bits */
345 			mflags = nlm_save_flags_cop2();
346 			msgstatus1 = nlm_read_c2_msgstatus1();
347 			msgstatus1 |= (0xf << 16);
348 			nlm_write_c2_msgstatus1(msgstatus1);
349 			nlm_restore_flags(mflags);
350 
351 			thread_lock(td);
352 			if (mthd->needed) {
353 				thread_unlock(td);
354 				continue;
355 			}
356 			sched_class(td, PRI_ITHD);
357 			TD_SET_IWAIT(td);
358 			mi_switch(SW_VOL);
359 		} else
360 			pause("wmsg", 1);
361 
362 		fmn_loops[hwtid]++;
363 	}
364 }
365 
366 static void
create_msgring_thread(int hwtid)367 create_msgring_thread(int hwtid)
368 {
369 	struct msgring_thread *mthd;
370 	struct thread *td;
371 	int	error;
372 
373 	mthd = &msgring_threads[hwtid];
374 	error = kproc_kthread_add(msgring_process, (void *)(uintptr_t)hwtid,
375 	    &msgring_proc, &td, RFSTOPPED, 2, "msgrngproc",
376 	    "msgthr%d", hwtid);
377 	if (error)
378 		panic("kproc_kthread_add() failed with %d", error);
379 	mthd->thread = td;
380 
381 	thread_lock(td);
382 	sched_class(td, PRI_ITHD);
383 	sched_add(td, SRQ_INTR);
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 	struct sbuf sb;
477 	int error, i;
478 
479 	sbuf_new_for_sysctl(&sb, NULL, 64, req);
480 	sbuf_printf(&sb,
481 	    "\nID     vc0       vc1       vc2     vc3     loops\n");
482 	for (i = 0; i < 32; i++) {
483 		if ((xlp_hw_thread_mask & (1 << i)) == 0)
484 			continue;
485 		sbuf_printf(&sb, "%2d: %8d %8d %8d %8d %8d\n", i,
486 		    fmn_msgcount[i][0], fmn_msgcount[i][1],
487 		    fmn_msgcount[i][2], fmn_msgcount[i][3],
488 		    fmn_loops[i]);
489 	}
490 	error = sbuf_finish(&sb);
491 	sbuf_delete(&sb);
492 	return (error);
493 }
494 
495 SYSCTL_PROC(_debug, OID_AUTO, msgring,
496     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, 0, 0,
497     sys_print_debug, "A",
498     "msgring debug info");
499