qman_test_stash.c

  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 */
 85struct bstrap {
 86	int (*fn)(void);
 87	atomic_t started;
 88};
 89static 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}
102static int on_all_cpus(int (*fn)(void))
103{
104	int cpu;
105
106	for_each_online_cpu(cpu) {
107		struct bstrap bstrap = {
108			.fn = fn,
109			.started = ATOMIC_INIT(0)
110		};
111		struct task_struct *k = kthread_run_on_cpu(bstrap_fn, &bstrap,
112							   cpu, "hotpotato%d");
113		int ret;
114
115		if (IS_ERR(k))
116			return -ENOMEM;
117		/*
118		 * If we call kthread_stop() before the "wake up" has had an
119		 * effect, then the thread may exit with -EINTR without ever
120		 * running the function. So poll until it's started before
121		 * requesting it to stop.
122		 */
123		while (!atomic_read(&bstrap.started))
124			msleep(20);
125		ret = kthread_stop(k);
126		if (ret)
127			return ret;
128	}
129	return 0;
130}
131
132struct hp_handler {
133
134	/* The following data is stashed when 'rx' is dequeued; */
135	/* -------------- */
136	/* The Rx FQ, dequeues of which will stash the entire hp_handler */
137	struct qman_fq rx;
138	/* The Tx FQ we should forward to */
139	struct qman_fq tx;
140	/* The value we XOR post-dequeue, prior to validating */
141	u32 rx_mixer;
142	/* The value we XOR pre-enqueue, after validating */
143	u32 tx_mixer;
144	/* what the hotpotato address should be on dequeue */
145	dma_addr_t addr;
146	u32 *frame_ptr;
147
148	/* The following data isn't (necessarily) stashed on dequeue; */
149	/* -------------- */
150	u32 fqid_rx, fqid_tx;
151	/* list node for linking us into 'hp_cpu' */
152	struct list_head node;
153	/* Just to check ... */
154	unsigned int processor_id;
155} ____cacheline_aligned;
156
157struct hp_cpu {
158	/* identify the cpu we run on; */
159	unsigned int processor_id;
160	/* root node for the per-cpu list of handlers */
161	struct list_head handlers;
162	/* list node for linking us into 'hp_cpu_list' */
163	struct list_head node;
164	/*
165	 * when repeatedly scanning 'hp_list', each time linking the n'th
166	 * handlers together, this is used as per-cpu iterator state
167	 */
168	struct hp_handler *iterator;
169};
170
171/* Each cpu has one of these */
172static DEFINE_PER_CPU(struct hp_cpu, hp_cpus);
173
174/* links together the hp_cpu structs, in first-come first-serve order. */
175static LIST_HEAD(hp_cpu_list);
176static DEFINE_SPINLOCK(hp_lock);
177
178static unsigned int hp_cpu_list_length;
179
180/* the "special" handler, that starts and terminates the test. */
181static struct hp_handler *special_handler;
182static int loop_counter;
183
184/* handlers are allocated out of this, so they're properly aligned. */
185static struct kmem_cache *hp_handler_slab;
186
187/* this is the frame data */
188static void *__frame_ptr;
189static u32 *frame_ptr;
190static dma_addr_t frame_dma;
191
192/* needed for dma_map*() */
193static const struct qm_portal_config *pcfg;
194
195/* the main function waits on this */
196static DECLARE_WAIT_QUEUE_HEAD(queue);
197
198#define HP_PER_CPU	2
199#define HP_LOOPS	8
200/* 80 bytes, like a small ethernet frame, and bleeds into a second cacheline */
201#define HP_NUM_WORDS	80
202/* First word of the LFSR-based frame data */
203#define HP_FIRST_WORD	0xabbaf00d
204
205static inline u32 do_lfsr(u32 prev)
206{
207	return (prev >> 1) ^ (-(prev & 1u) & 0xd0000001u);
208}
209
210static int allocate_frame_data(void)
211{
212	u32 lfsr = HP_FIRST_WORD;
213	int loop;
214
215	if (!qman_dma_portal) {
216		pr_crit("portal not available\n");
217		return -EIO;
218	}
219
220	pcfg = qman_get_qm_portal_config(qman_dma_portal);
221
222	__frame_ptr = kmalloc_array(4, HP_NUM_WORDS, GFP_KERNEL);
223	if (!__frame_ptr)
224		return -ENOMEM;
225
226	frame_ptr = PTR_ALIGN(__frame_ptr, 64);
227	for (loop = 0; loop < HP_NUM_WORDS; loop++) {
228		frame_ptr[loop] = lfsr;
229		lfsr = do_lfsr(lfsr);
230	}
231
232	frame_dma = dma_map_single(pcfg->dev, frame_ptr, 4 * HP_NUM_WORDS,
233				   DMA_BIDIRECTIONAL);
234	if (dma_mapping_error(pcfg->dev, frame_dma)) {
235		pr_crit("dma mapping failure\n");
236		kfree(__frame_ptr);
237		return -EIO;
238	}
239
240	return 0;
241}
242
243static void deallocate_frame_data(void)
244{
245	dma_unmap_single(pcfg->dev, frame_dma, 4 * HP_NUM_WORDS,
246			 DMA_BIDIRECTIONAL);
247	kfree(__frame_ptr);
248}
249
250static inline int process_frame_data(struct hp_handler *handler,
251				     const struct qm_fd *fd)
252{
253	u32 *p = handler->frame_ptr;
254	u32 lfsr = HP_FIRST_WORD;
255	int loop;
256
257	if (qm_fd_addr_get64(fd) != handler->addr) {
258		pr_crit("bad frame address, [%llX != %llX]\n",
259			qm_fd_addr_get64(fd), handler->addr);
260		return -EIO;
261	}
262	for (loop = 0; loop < HP_NUM_WORDS; loop++, p++) {
263		*p ^= handler->rx_mixer;
264		if (*p != lfsr) {
265			pr_crit("corrupt frame data");
266			return -EIO;
267		}
268		*p ^= handler->tx_mixer;
269		lfsr = do_lfsr(lfsr);
270	}
271	return 0;
272}
273
274static enum qman_cb_dqrr_result normal_dqrr(struct qman_portal *portal,
275					    struct qman_fq *fq,
276					    const struct qm_dqrr_entry *dqrr,
277					    bool sched_napi)
278{
279	struct hp_handler *handler = (struct hp_handler *)fq;
280
281	if (process_frame_data(handler, &dqrr->fd)) {
282		WARN_ON(1);
283		goto skip;
284	}
285	if (qman_enqueue(&handler->tx, &dqrr->fd)) {
286		pr_crit("qman_enqueue() failed");
287		WARN_ON(1);
288	}
289skip:
290	return qman_cb_dqrr_consume;
291}
292
293static enum qman_cb_dqrr_result special_dqrr(struct qman_portal *portal,
294					     struct qman_fq *fq,
295					     const struct qm_dqrr_entry *dqrr,
296					     bool sched_napi)
297{
298	struct hp_handler *handler = (struct hp_handler *)fq;
299
300	process_frame_data(handler, &dqrr->fd);
301	if (++loop_counter < HP_LOOPS) {
302		if (qman_enqueue(&handler->tx, &dqrr->fd)) {
303			pr_crit("qman_enqueue() failed");
304			WARN_ON(1);
305			goto skip;
306		}
307	} else {
308		pr_info("Received final (%dth) frame\n", loop_counter);
309		wake_up(&queue);
310	}
311skip:
312	return qman_cb_dqrr_consume;
313}
314
315static int create_per_cpu_handlers(void)
316{
317	struct hp_handler *handler;
318	int loop;
319	struct hp_cpu *hp_cpu = this_cpu_ptr(&hp_cpus);
320
321	hp_cpu->processor_id = smp_processor_id();
322	spin_lock(&hp_lock);
323	list_add_tail(&hp_cpu->node, &hp_cpu_list);
324	hp_cpu_list_length++;
325	spin_unlock(&hp_lock);
326	INIT_LIST_HEAD(&hp_cpu->handlers);
327	for (loop = 0; loop < HP_PER_CPU; loop++) {
328		handler = kmem_cache_alloc(hp_handler_slab, GFP_KERNEL);
329		if (!handler) {
330			pr_crit("kmem_cache_alloc() failed");
331			WARN_ON(1);
332			return -EIO;
333		}
334		handler->processor_id = hp_cpu->processor_id;
335		handler->addr = frame_dma;
336		handler->frame_ptr = frame_ptr;
337		list_add_tail(&handler->node, &hp_cpu->handlers);
338	}
339	return 0;
340}
341
342static int destroy_per_cpu_handlers(void)
343{
344	struct list_head *loop, *tmp;
345	struct hp_cpu *hp_cpu = this_cpu_ptr(&hp_cpus);
346
347	spin_lock(&hp_lock);
348	list_del(&hp_cpu->node);
349	spin_unlock(&hp_lock);
350	list_for_each_safe(loop, tmp, &hp_cpu->handlers) {
351		u32 flags = 0;
352		struct hp_handler *handler = list_entry(loop, struct hp_handler,
353							node);
354		if (qman_retire_fq(&handler->rx, &flags) ||
355		    (flags & QMAN_FQ_STATE_BLOCKOOS)) {
356			pr_crit("qman_retire_fq(rx) failed, flags: %x", flags);
357			WARN_ON(1);
358			return -EIO;
359		}
360		if (qman_oos_fq(&handler->rx)) {
361			pr_crit("qman_oos_fq(rx) failed");
362			WARN_ON(1);
363			return -EIO;
364		}
365		qman_destroy_fq(&handler->rx);
366		qman_destroy_fq(&handler->tx);
367		qman_release_fqid(handler->fqid_rx);
368		list_del(&handler->node);
369		kmem_cache_free(hp_handler_slab, handler);
370	}
371	return 0;
372}
373
374static inline u8 num_cachelines(u32 offset)
375{
376	u8 res = (offset + (L1_CACHE_BYTES - 1))
377			 / (L1_CACHE_BYTES);
378	if (res > 3)
379		return 3;
380	return res;
381}
382#define STASH_DATA_CL \
383	num_cachelines(HP_NUM_WORDS * 4)
384#define STASH_CTX_CL \
385	num_cachelines(offsetof(struct hp_handler, fqid_rx))
386
387static int init_handler(void *h)
388{
389	struct qm_mcc_initfq opts;
390	struct hp_handler *handler = h;
391	int err;
392
393	if (handler->processor_id != smp_processor_id()) {
394		err = -EIO;
395		goto failed;
396	}
397	/* Set up rx */
398	memset(&handler->rx, 0, sizeof(handler->rx));
399	if (handler == special_handler)
400		handler->rx.cb.dqrr = special_dqrr;
401	else
402		handler->rx.cb.dqrr = normal_dqrr;
403	err = qman_create_fq(handler->fqid_rx, 0, &handler->rx);
404	if (err) {
405		pr_crit("qman_create_fq(rx) failed");
406		goto failed;
407	}
408	memset(&opts, 0, sizeof(opts));
409	opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL |
410				   QM_INITFQ_WE_CONTEXTA);
411	opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CTXASTASHING);
412	qm_fqd_set_stashing(&opts.fqd, 0, STASH_DATA_CL, STASH_CTX_CL);
413	err = qman_init_fq(&handler->rx, QMAN_INITFQ_FLAG_SCHED |
414			   QMAN_INITFQ_FLAG_LOCAL, &opts);
415	if (err) {
416		pr_crit("qman_init_fq(rx) failed");
417		goto failed;
418	}
419	/* Set up tx */
420	memset(&handler->tx, 0, sizeof(handler->tx));
421	err = qman_create_fq(handler->fqid_tx, QMAN_FQ_FLAG_NO_MODIFY,
422			     &handler->tx);
423	if (err) {
424		pr_crit("qman_create_fq(tx) failed");
425		goto failed;
426	}
427
428	return 0;
429failed:
430	return err;
431}
432
433static void init_handler_cb(void *h)
434{
435	if (init_handler(h))
436		WARN_ON(1);
437}
438
439static int init_phase2(void)
440{
441	int loop;
442	u32 fqid = 0;
443	u32 lfsr = 0xdeadbeef;
444	struct hp_cpu *hp_cpu;
445	struct hp_handler *handler;
446
447	for (loop = 0; loop < HP_PER_CPU; loop++) {
448		list_for_each_entry(hp_cpu, &hp_cpu_list, node) {
449			int err;
450
451			if (!loop)
452				hp_cpu->iterator = list_first_entry(
453						&hp_cpu->handlers,
454						struct hp_handler, node);
455			else
456				hp_cpu->iterator = list_entry(
457						hp_cpu->iterator->node.next,
458						struct hp_handler, node);
459			/* Rx FQID is the previous handler's Tx FQID */
460			hp_cpu->iterator->fqid_rx = fqid;
461			/* Allocate new FQID for Tx */
462			err = qman_alloc_fqid(&fqid);
463			if (err) {
464				pr_crit("qman_alloc_fqid() failed");
465				return err;
466			}
467			hp_cpu->iterator->fqid_tx = fqid;
468			/* Rx mixer is the previous handler's Tx mixer */
469			hp_cpu->iterator->rx_mixer = lfsr;
470			/* Get new mixer for Tx */
471			lfsr = do_lfsr(lfsr);
472			hp_cpu->iterator->tx_mixer = lfsr;
473		}
474	}
475	/* Fix up the first handler (fqid_rx==0, rx_mixer=0xdeadbeef) */
476	hp_cpu = list_first_entry(&hp_cpu_list, struct hp_cpu, node);
477	handler = list_first_entry(&hp_cpu->handlers, struct hp_handler, node);
478	if (handler->fqid_rx != 0 || handler->rx_mixer != 0xdeadbeef)
479		return 1;
480	handler->fqid_rx = fqid;
481	handler->rx_mixer = lfsr;
482	/* and tag it as our "special" handler */
483	special_handler = handler;
484	return 0;
485}
486
487static int init_phase3(void)
488{
489	int loop, err;
490	struct hp_cpu *hp_cpu;
491
492	for (loop = 0; loop < HP_PER_CPU; loop++) {
493		list_for_each_entry(hp_cpu, &hp_cpu_list, node) {
494			if (!loop)
495				hp_cpu->iterator = list_first_entry(
496						&hp_cpu->handlers,
497						struct hp_handler, node);
498			else
499				hp_cpu->iterator = list_entry(
500						hp_cpu->iterator->node.next,
501						struct hp_handler, node);
502			preempt_disable();
503			if (hp_cpu->processor_id == smp_processor_id()) {
504				err = init_handler(hp_cpu->iterator);
505				if (err)
506					return err;
507			} else {
508				smp_call_function_single(hp_cpu->processor_id,
509					init_handler_cb, hp_cpu->iterator, 1);
510			}
511			preempt_enable();
512		}
513	}
514	return 0;
515}
516
517static int send_first_frame(void *ignore)
518{
519	u32 *p = special_handler->frame_ptr;
520	u32 lfsr = HP_FIRST_WORD;
521	int loop, err;
522	struct qm_fd fd;
523
524	if (special_handler->processor_id != smp_processor_id()) {
525		err = -EIO;
526		goto failed;
527	}
528	memset(&fd, 0, sizeof(fd));
529	qm_fd_addr_set64(&fd, special_handler->addr);
530	qm_fd_set_contig_big(&fd, HP_NUM_WORDS * 4);
531	for (loop = 0; loop < HP_NUM_WORDS; loop++, p++) {
532		if (*p != lfsr) {
533			err = -EIO;
534			pr_crit("corrupt frame data");
535			goto failed;
536		}
537		*p ^= special_handler->tx_mixer;
538		lfsr = do_lfsr(lfsr);
539	}
540	pr_info("Sending first frame\n");
541	err = qman_enqueue(&special_handler->tx, &fd);
542	if (err) {
543		pr_crit("qman_enqueue() failed");
544		goto failed;
545	}
546
547	return 0;
548failed:
549	return err;
550}
551
552static void send_first_frame_cb(void *ignore)
553{
554	if (send_first_frame(NULL))
555		WARN_ON(1);
556}
557
558int qman_test_stash(void)
559{
560	int err;
561
562	if (cpumask_weight(cpu_online_mask) < 2) {
563		pr_info("%s(): skip - only 1 CPU\n", __func__);
564		return 0;
565	}
566
567	pr_info("%s(): Starting\n", __func__);
568
569	hp_cpu_list_length = 0;
570	loop_counter = 0;
571	hp_handler_slab = kmem_cache_create("hp_handler_slab",
572			sizeof(struct hp_handler), L1_CACHE_BYTES,
573			SLAB_HWCACHE_ALIGN, NULL);
574	if (!hp_handler_slab) {
575		err = -EIO;
576		pr_crit("kmem_cache_create() failed");
577		goto failed;
578	}
579
580	err = allocate_frame_data();
581	if (err)
582		goto failed;
583
584	/* Init phase 1 */
585	pr_info("Creating %d handlers per cpu...\n", HP_PER_CPU);
586	if (on_all_cpus(create_per_cpu_handlers)) {
587		err = -EIO;
588		pr_crit("on_each_cpu() failed");
589		goto failed;
590	}
591	pr_info("Number of cpus: %d, total of %d handlers\n",
592		hp_cpu_list_length, hp_cpu_list_length * HP_PER_CPU);
593
594	err = init_phase2();
595	if (err)
596		goto failed;
597
598	err = init_phase3();
599	if (err)
600		goto failed;
601
602	preempt_disable();
603	if (special_handler->processor_id == smp_processor_id()) {
604		err = send_first_frame(NULL);
605		if (err)
606			goto failed;
607	} else {
608		smp_call_function_single(special_handler->processor_id,
609					 send_first_frame_cb, NULL, 1);
610	}
611	preempt_enable();
612
613	wait_event(queue, loop_counter == HP_LOOPS);
614	deallocate_frame_data();
615	if (on_all_cpus(destroy_per_cpu_handlers)) {
616		err = -EIO;
617		pr_crit("on_each_cpu() failed");
618		goto failed;
619	}
620	kmem_cache_destroy(hp_handler_slab);
621	pr_info("%s(): Finished\n", __func__);
622
623	return 0;
624failed:
625	WARN_ON(1);
626	return err;
627}