1 /* Copyright 2009 - 2016 Freescale Semiconductor, Inc.
2  *
3  * Redistribution and use in source and binary forms, with or without
4  * modification, are permitted provided that the following conditions are met:
5  *     * Redistributions of source code must retain the above copyright
6  *	 notice, this list of conditions and the following disclaimer.
7  *     * Redistributions in binary form must reproduce the above copyright
8  *	 notice, this list of conditions and the following disclaimer in the
9  *	 documentation and/or other materials provided with the distribution.
10  *     * Neither the name of Freescale Semiconductor nor the
11  *	 names of its contributors may be used to endorse or promote products
12  *	 derived from this software without specific prior written permission.
13  *
14  * ALTERNATIVELY, this software may be distributed under the terms of the
15  * GNU General Public License ("GPL") as published by the Free Software
16  * Foundation, either version 2 of that License or (at your option) any
17  * later version.
18  *
19  * THIS SOFTWARE IS PROVIDED BY Freescale Semiconductor ``AS IS'' AND ANY
20  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22  * DISCLAIMED. IN NO EVENT SHALL Freescale Semiconductor BE LIABLE FOR ANY
23  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
24  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
25  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
26  * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
28  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 
31 #include "qman_test.h"
32 
33 #include <linux/dma-mapping.h>
34 #include <linux/delay.h>
35 
36 /*
37  * Algorithm:
38  *
39  * Each cpu will have HP_PER_CPU "handlers" set up, each of which incorporates
40  * an rx/tx pair of FQ objects (both of which are stashed on dequeue). The
41  * organisation of FQIDs is such that the HP_PER_CPU*NUM_CPUS handlers will
42  * shuttle a "hot potato" frame around them such that every forwarding action
43  * moves it from one cpu to another. (The use of more than one handler per cpu
44  * is to allow enough handlers/FQs to truly test the significance of caching -
45  * ie. when cache-expiries are occurring.)
46  *
47  * The "hot potato" frame content will be HP_NUM_WORDS*4 bytes in size, and the
48  * first and last words of the frame data will undergo a transformation step on
49  * each forwarding action. To achieve this, each handler will be assigned a
50  * 32-bit "mixer", that is produced using a 32-bit LFSR. When a frame is
51  * received by a handler, the mixer of the expected sender is XOR'd into all
52  * words of the entire frame, which is then validated against the original
53  * values. Then, before forwarding, the entire frame is XOR'd with the mixer of
54  * the current handler. Apart from validating that the frame is taking the
55  * expected path, this also provides some quasi-realistic overheads to each
56  * forwarding action - dereferencing *all* the frame data, computation, and
57  * conditional branching. There is a "special" handler designated to act as the
58  * instigator of the test by creating an enqueuing the "hot potato" frame, and
59  * to determine when the test has completed by counting HP_LOOPS iterations.
60  *
61  * Init phases:
62  *
63  * 1. prepare each cpu's 'hp_cpu' struct using on_each_cpu(,,1) and link them
64  *    into 'hp_cpu_list'. Specifically, set processor_id, allocate HP_PER_CPU
65  *    handlers and link-list them (but do no other handler setup).
66  *
67  * 2. scan over 'hp_cpu_list' HP_PER_CPU times, the first time sets each
68  *    hp_cpu's 'iterator' to point to its first handler. With each loop,
69  *    allocate rx/tx FQIDs and mixer values to the hp_cpu's iterator handler
70  *    and advance the iterator for the next loop. This includes a final fixup,
71  *    which connects the last handler to the first (and which is why phase 2
72  *    and 3 are separate).
73  *
74  * 3. scan over 'hp_cpu_list' HP_PER_CPU times, the first time sets each
75  *    hp_cpu's 'iterator' to point to its first handler. With each loop,
76  *    initialise FQ objects and advance the iterator for the next loop.
77  *    Moreover, do this initialisation on the cpu it applies to so that Rx FQ
78  *    initialisation targets the correct cpu.
79  */
80 
81 /*
82  * helper to run something on all cpus (can't use on_each_cpu(), as that invokes
83  * the fn from irq context, which is too restrictive).
84  */
85 struct bstrap {
86 	int (*fn)(void);
87 	atomic_t started;
88 };
89 static int bstrap_fn(void *bs)
90 {
91 	struct bstrap *bstrap = bs;
92 	int err;
93 
94 	atomic_inc(&bstrap->started);
95 	err = bstrap->fn();
96 	if (err)
97 		return err;
98 	while (!kthread_should_stop())
99 		msleep(20);
100 	return 0;
101 }
102 static int on_all_cpus(int (*fn)(void))
103 {
104 	int cpu;
105 
106 	for_each_cpu(cpu, cpu_online_mask) {
107 		struct bstrap bstrap = {
108 			.fn = fn,
109 			.started = ATOMIC_INIT(0)
110 		};
111 		struct task_struct *k = kthread_create(bstrap_fn, &bstrap,
112 			"hotpotato%d", cpu);
113 		int ret;
114 
115 		if (IS_ERR(k))
116 			return -ENOMEM;
117 		kthread_bind(k, cpu);
118 		wake_up_process(k);
119 		/*
120 		 * If we call kthread_stop() before the "wake up" has had an
121 		 * effect, then the thread may exit with -EINTR without ever
122 		 * running the function. So poll until it's started before
123 		 * requesting it to stop.
124 		 */
125 		while (!atomic_read(&bstrap.started))
126 			msleep(20);
127 		ret = kthread_stop(k);
128 		if (ret)
129 			return ret;
130 	}
131 	return 0;
132 }
133 
134 struct hp_handler {
135 
136 	/* The following data is stashed when 'rx' is dequeued; */
137 	/* -------------- */
138 	/* The Rx FQ, dequeues of which will stash the entire hp_handler */
139 	struct qman_fq rx;
140 	/* The Tx FQ we should forward to */
141 	struct qman_fq tx;
142 	/* The value we XOR post-dequeue, prior to validating */
143 	u32 rx_mixer;
144 	/* The value we XOR pre-enqueue, after validating */
145 	u32 tx_mixer;
146 	/* what the hotpotato address should be on dequeue */
147 	dma_addr_t addr;
148 	u32 *frame_ptr;
149 
150 	/* The following data isn't (necessarily) stashed on dequeue; */
151 	/* -------------- */
152 	u32 fqid_rx, fqid_tx;
153 	/* list node for linking us into 'hp_cpu' */
154 	struct list_head node;
155 	/* Just to check ... */
156 	unsigned int processor_id;
157 } ____cacheline_aligned;
158 
159 struct hp_cpu {
160 	/* identify the cpu we run on; */
161 	unsigned int processor_id;
162 	/* root node for the per-cpu list of handlers */
163 	struct list_head handlers;
164 	/* list node for linking us into 'hp_cpu_list' */
165 	struct list_head node;
166 	/*
167 	 * when repeatedly scanning 'hp_list', each time linking the n'th
168 	 * handlers together, this is used as per-cpu iterator state
169 	 */
170 	struct hp_handler *iterator;
171 };
172 
173 /* Each cpu has one of these */
174 static DEFINE_PER_CPU(struct hp_cpu, hp_cpus);
175 
176 /* links together the hp_cpu structs, in first-come first-serve order. */
177 static LIST_HEAD(hp_cpu_list);
178 static DEFINE_SPINLOCK(hp_lock);
179 
180 static unsigned int hp_cpu_list_length;
181 
182 /* the "special" handler, that starts and terminates the test. */
183 static struct hp_handler *special_handler;
184 static int loop_counter;
185 
186 /* handlers are allocated out of this, so they're properly aligned. */
187 static struct kmem_cache *hp_handler_slab;
188 
189 /* this is the frame data */
190 static void *__frame_ptr;
191 static u32 *frame_ptr;
192 static dma_addr_t frame_dma;
193 
194 /* needed for dma_map*() */
195 static const struct qm_portal_config *pcfg;
196 
197 /* the main function waits on this */
198 static DECLARE_WAIT_QUEUE_HEAD(queue);
199 
200 #define HP_PER_CPU	2
201 #define HP_LOOPS	8
202 /* 80 bytes, like a small ethernet frame, and bleeds into a second cacheline */
203 #define HP_NUM_WORDS	80
204 /* First word of the LFSR-based frame data */
205 #define HP_FIRST_WORD	0xabbaf00d
206 
207 static inline u32 do_lfsr(u32 prev)
208 {
209 	return (prev >> 1) ^ (-(prev & 1u) & 0xd0000001u);
210 }
211 
212 static int allocate_frame_data(void)
213 {
214 	u32 lfsr = HP_FIRST_WORD;
215 	int loop;
216 
217 	if (!qman_dma_portal) {
218 		pr_crit("portal not available\n");
219 		return -EIO;
220 	}
221 
222 	pcfg = qman_get_qm_portal_config(qman_dma_portal);
223 
224 	__frame_ptr = kmalloc(4 * HP_NUM_WORDS, GFP_KERNEL);
225 	if (!__frame_ptr)
226 		return -ENOMEM;
227 
228 	frame_ptr = PTR_ALIGN(__frame_ptr, 64);
229 	for (loop = 0; loop < HP_NUM_WORDS; loop++) {
230 		frame_ptr[loop] = lfsr;
231 		lfsr = do_lfsr(lfsr);
232 	}
233 
234 	frame_dma = dma_map_single(pcfg->dev, frame_ptr, 4 * HP_NUM_WORDS,
235 				   DMA_BIDIRECTIONAL);
236 	if (dma_mapping_error(pcfg->dev, frame_dma)) {
237 		pr_crit("dma mapping failure\n");
238 		kfree(__frame_ptr);
239 		return -EIO;
240 	}
241 
242 	return 0;
243 }
244 
245 static void deallocate_frame_data(void)
246 {
247 	dma_unmap_single(pcfg->dev, frame_dma, 4 * HP_NUM_WORDS,
248 			 DMA_BIDIRECTIONAL);
249 	kfree(__frame_ptr);
250 }
251 
252 static inline int process_frame_data(struct hp_handler *handler,
253 				     const struct qm_fd *fd)
254 {
255 	u32 *p = handler->frame_ptr;
256 	u32 lfsr = HP_FIRST_WORD;
257 	int loop;
258 
259 	if (qm_fd_addr_get64(fd) != handler->addr) {
260 		pr_crit("bad frame address, [%llX != %llX]\n",
261 			qm_fd_addr_get64(fd), handler->addr);
262 		return -EIO;
263 	}
264 	for (loop = 0; loop < HP_NUM_WORDS; loop++, p++) {
265 		*p ^= handler->rx_mixer;
266 		if (*p != lfsr) {
267 			pr_crit("corrupt frame data");
268 			return -EIO;
269 		}
270 		*p ^= handler->tx_mixer;
271 		lfsr = do_lfsr(lfsr);
272 	}
273 	return 0;
274 }
275 
276 static enum qman_cb_dqrr_result normal_dqrr(struct qman_portal *portal,
277 					    struct qman_fq *fq,
278 					    const struct qm_dqrr_entry *dqrr,
279 					    bool sched_napi)
280 {
281 	struct hp_handler *handler = (struct hp_handler *)fq;
282 
283 	if (process_frame_data(handler, &dqrr->fd)) {
284 		WARN_ON(1);
285 		goto skip;
286 	}
287 	if (qman_enqueue(&handler->tx, &dqrr->fd)) {
288 		pr_crit("qman_enqueue() failed");
289 		WARN_ON(1);
290 	}
291 skip:
292 	return qman_cb_dqrr_consume;
293 }
294 
295 static enum qman_cb_dqrr_result special_dqrr(struct qman_portal *portal,
296 					     struct qman_fq *fq,
297 					     const struct qm_dqrr_entry *dqrr,
298 					     bool sched_napi)
299 {
300 	struct hp_handler *handler = (struct hp_handler *)fq;
301 
302 	process_frame_data(handler, &dqrr->fd);
303 	if (++loop_counter < HP_LOOPS) {
304 		if (qman_enqueue(&handler->tx, &dqrr->fd)) {
305 			pr_crit("qman_enqueue() failed");
306 			WARN_ON(1);
307 			goto skip;
308 		}
309 	} else {
310 		pr_info("Received final (%dth) frame\n", loop_counter);
311 		wake_up(&queue);
312 	}
313 skip:
314 	return qman_cb_dqrr_consume;
315 }
316 
317 static int create_per_cpu_handlers(void)
318 {
319 	struct hp_handler *handler;
320 	int loop;
321 	struct hp_cpu *hp_cpu = this_cpu_ptr(&hp_cpus);
322 
323 	hp_cpu->processor_id = smp_processor_id();
324 	spin_lock(&hp_lock);
325 	list_add_tail(&hp_cpu->node, &hp_cpu_list);
326 	hp_cpu_list_length++;
327 	spin_unlock(&hp_lock);
328 	INIT_LIST_HEAD(&hp_cpu->handlers);
329 	for (loop = 0; loop < HP_PER_CPU; loop++) {
330 		handler = kmem_cache_alloc(hp_handler_slab, GFP_KERNEL);
331 		if (!handler) {
332 			pr_crit("kmem_cache_alloc() failed");
333 			WARN_ON(1);
334 			return -EIO;
335 		}
336 		handler->processor_id = hp_cpu->processor_id;
337 		handler->addr = frame_dma;
338 		handler->frame_ptr = frame_ptr;
339 		list_add_tail(&handler->node, &hp_cpu->handlers);
340 	}
341 	return 0;
342 }
343 
344 static int destroy_per_cpu_handlers(void)
345 {
346 	struct list_head *loop, *tmp;
347 	struct hp_cpu *hp_cpu = this_cpu_ptr(&hp_cpus);
348 
349 	spin_lock(&hp_lock);
350 	list_del(&hp_cpu->node);
351 	spin_unlock(&hp_lock);
352 	list_for_each_safe(loop, tmp, &hp_cpu->handlers) {
353 		u32 flags = 0;
354 		struct hp_handler *handler = list_entry(loop, struct hp_handler,
355 							node);
356 		if (qman_retire_fq(&handler->rx, &flags) ||
357 		    (flags & QMAN_FQ_STATE_BLOCKOOS)) {
358 			pr_crit("qman_retire_fq(rx) failed, flags: %x", flags);
359 			WARN_ON(1);
360 			return -EIO;
361 		}
362 		if (qman_oos_fq(&handler->rx)) {
363 			pr_crit("qman_oos_fq(rx) failed");
364 			WARN_ON(1);
365 			return -EIO;
366 		}
367 		qman_destroy_fq(&handler->rx);
368 		qman_destroy_fq(&handler->tx);
369 		qman_release_fqid(handler->fqid_rx);
370 		list_del(&handler->node);
371 		kmem_cache_free(hp_handler_slab, handler);
372 	}
373 	return 0;
374 }
375 
376 static inline u8 num_cachelines(u32 offset)
377 {
378 	u8 res = (offset + (L1_CACHE_BYTES - 1))
379 			 / (L1_CACHE_BYTES);
380 	if (res > 3)
381 		return 3;
382 	return res;
383 }
384 #define STASH_DATA_CL \
385 	num_cachelines(HP_NUM_WORDS * 4)
386 #define STASH_CTX_CL \
387 	num_cachelines(offsetof(struct hp_handler, fqid_rx))
388 
389 static int init_handler(void *h)
390 {
391 	struct qm_mcc_initfq opts;
392 	struct hp_handler *handler = h;
393 	int err;
394 
395 	if (handler->processor_id != smp_processor_id()) {
396 		err = -EIO;
397 		goto failed;
398 	}
399 	/* Set up rx */
400 	memset(&handler->rx, 0, sizeof(handler->rx));
401 	if (handler == special_handler)
402 		handler->rx.cb.dqrr = special_dqrr;
403 	else
404 		handler->rx.cb.dqrr = normal_dqrr;
405 	err = qman_create_fq(handler->fqid_rx, 0, &handler->rx);
406 	if (err) {
407 		pr_crit("qman_create_fq(rx) failed");
408 		goto failed;
409 	}
410 	memset(&opts, 0, sizeof(opts));
411 	opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL |
412 				   QM_INITFQ_WE_CONTEXTA);
413 	opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CTXASTASHING);
414 	qm_fqd_set_stashing(&opts.fqd, 0, STASH_DATA_CL, STASH_CTX_CL);
415 	err = qman_init_fq(&handler->rx, QMAN_INITFQ_FLAG_SCHED |
416 			   QMAN_INITFQ_FLAG_LOCAL, &opts);
417 	if (err) {
418 		pr_crit("qman_init_fq(rx) failed");
419 		goto failed;
420 	}
421 	/* Set up tx */
422 	memset(&handler->tx, 0, sizeof(handler->tx));
423 	err = qman_create_fq(handler->fqid_tx, QMAN_FQ_FLAG_NO_MODIFY,
424 			     &handler->tx);
425 	if (err) {
426 		pr_crit("qman_create_fq(tx) failed");
427 		goto failed;
428 	}
429 
430 	return 0;
431 failed:
432 	return err;
433 }
434 
435 static void init_handler_cb(void *h)
436 {
437 	if (init_handler(h))
438 		WARN_ON(1);
439 }
440 
441 static int init_phase2(void)
442 {
443 	int loop;
444 	u32 fqid = 0;
445 	u32 lfsr = 0xdeadbeef;
446 	struct hp_cpu *hp_cpu;
447 	struct hp_handler *handler;
448 
449 	for (loop = 0; loop < HP_PER_CPU; loop++) {
450 		list_for_each_entry(hp_cpu, &hp_cpu_list, node) {
451 			int err;
452 
453 			if (!loop)
454 				hp_cpu->iterator = list_first_entry(
455 						&hp_cpu->handlers,
456 						struct hp_handler, node);
457 			else
458 				hp_cpu->iterator = list_entry(
459 						hp_cpu->iterator->node.next,
460 						struct hp_handler, node);
461 			/* Rx FQID is the previous handler's Tx FQID */
462 			hp_cpu->iterator->fqid_rx = fqid;
463 			/* Allocate new FQID for Tx */
464 			err = qman_alloc_fqid(&fqid);
465 			if (err) {
466 				pr_crit("qman_alloc_fqid() failed");
467 				return err;
468 			}
469 			hp_cpu->iterator->fqid_tx = fqid;
470 			/* Rx mixer is the previous handler's Tx mixer */
471 			hp_cpu->iterator->rx_mixer = lfsr;
472 			/* Get new mixer for Tx */
473 			lfsr = do_lfsr(lfsr);
474 			hp_cpu->iterator->tx_mixer = lfsr;
475 		}
476 	}
477 	/* Fix up the first handler (fqid_rx==0, rx_mixer=0xdeadbeef) */
478 	hp_cpu = list_first_entry(&hp_cpu_list, struct hp_cpu, node);
479 	handler = list_first_entry(&hp_cpu->handlers, struct hp_handler, node);
480 	if (handler->fqid_rx != 0 || handler->rx_mixer != 0xdeadbeef)
481 		return 1;
482 	handler->fqid_rx = fqid;
483 	handler->rx_mixer = lfsr;
484 	/* and tag it as our "special" handler */
485 	special_handler = handler;
486 	return 0;
487 }
488 
489 static int init_phase3(void)
490 {
491 	int loop, err;
492 	struct hp_cpu *hp_cpu;
493 
494 	for (loop = 0; loop < HP_PER_CPU; loop++) {
495 		list_for_each_entry(hp_cpu, &hp_cpu_list, node) {
496 			if (!loop)
497 				hp_cpu->iterator = list_first_entry(
498 						&hp_cpu->handlers,
499 						struct hp_handler, node);
500 			else
501 				hp_cpu->iterator = list_entry(
502 						hp_cpu->iterator->node.next,
503 						struct hp_handler, node);
504 			preempt_disable();
505 			if (hp_cpu->processor_id == smp_processor_id()) {
506 				err = init_handler(hp_cpu->iterator);
507 				if (err)
508 					return err;
509 			} else {
510 				smp_call_function_single(hp_cpu->processor_id,
511 					init_handler_cb, hp_cpu->iterator, 1);
512 			}
513 			preempt_enable();
514 		}
515 	}
516 	return 0;
517 }
518 
519 static int send_first_frame(void *ignore)
520 {
521 	u32 *p = special_handler->frame_ptr;
522 	u32 lfsr = HP_FIRST_WORD;
523 	int loop, err;
524 	struct qm_fd fd;
525 
526 	if (special_handler->processor_id != smp_processor_id()) {
527 		err = -EIO;
528 		goto failed;
529 	}
530 	memset(&fd, 0, sizeof(fd));
531 	qm_fd_addr_set64(&fd, special_handler->addr);
532 	qm_fd_set_contig_big(&fd, HP_NUM_WORDS * 4);
533 	for (loop = 0; loop < HP_NUM_WORDS; loop++, p++) {
534 		if (*p != lfsr) {
535 			err = -EIO;
536 			pr_crit("corrupt frame data");
537 			goto failed;
538 		}
539 		*p ^= special_handler->tx_mixer;
540 		lfsr = do_lfsr(lfsr);
541 	}
542 	pr_info("Sending first frame\n");
543 	err = qman_enqueue(&special_handler->tx, &fd);
544 	if (err) {
545 		pr_crit("qman_enqueue() failed");
546 		goto failed;
547 	}
548 
549 	return 0;
550 failed:
551 	return err;
552 }
553 
554 static void send_first_frame_cb(void *ignore)
555 {
556 	if (send_first_frame(NULL))
557 		WARN_ON(1);
558 }
559 
560 int qman_test_stash(void)
561 {
562 	int err;
563 
564 	if (cpumask_weight(cpu_online_mask) < 2) {
565 		pr_info("%s(): skip - only 1 CPU\n", __func__);
566 		return 0;
567 	}
568 
569 	pr_info("%s(): Starting\n", __func__);
570 
571 	hp_cpu_list_length = 0;
572 	loop_counter = 0;
573 	hp_handler_slab = kmem_cache_create("hp_handler_slab",
574 			sizeof(struct hp_handler), L1_CACHE_BYTES,
575 			SLAB_HWCACHE_ALIGN, NULL);
576 	if (!hp_handler_slab) {
577 		err = -EIO;
578 		pr_crit("kmem_cache_create() failed");
579 		goto failed;
580 	}
581 
582 	err = allocate_frame_data();
583 	if (err)
584 		goto failed;
585 
586 	/* Init phase 1 */
587 	pr_info("Creating %d handlers per cpu...\n", HP_PER_CPU);
588 	if (on_all_cpus(create_per_cpu_handlers)) {
589 		err = -EIO;
590 		pr_crit("on_each_cpu() failed");
591 		goto failed;
592 	}
593 	pr_info("Number of cpus: %d, total of %d handlers\n",
594 		hp_cpu_list_length, hp_cpu_list_length * HP_PER_CPU);
595 
596 	err = init_phase2();
597 	if (err)
598 		goto failed;
599 
600 	err = init_phase3();
601 	if (err)
602 		goto failed;
603 
604 	preempt_disable();
605 	if (special_handler->processor_id == smp_processor_id()) {
606 		err = send_first_frame(NULL);
607 		if (err)
608 			goto failed;
609 	} else {
610 		smp_call_function_single(special_handler->processor_id,
611 					 send_first_frame_cb, NULL, 1);
612 	}
613 	preempt_enable();
614 
615 	wait_event(queue, loop_counter == HP_LOOPS);
616 	deallocate_frame_data();
617 	if (on_all_cpus(destroy_per_cpu_handlers)) {
618 		err = -EIO;
619 		pr_crit("on_each_cpu() failed");
620 		goto failed;
621 	}
622 	kmem_cache_destroy(hp_handler_slab);
623 	pr_info("%s(): Finished\n", __func__);
624 
625 	return 0;
626 failed:
627 	WARN_ON(1);
628 	return err;
629 }
630