1 /*
2  *  IBM eServer eHCA Infiniband device driver for Linux on POWER
3  *
4  *  Functions for EQs, NEQs and interrupts
5  *
6  *  Authors: Heiko J Schick <schickhj@de.ibm.com>
7  *           Khadija Souissi <souissi@de.ibm.com>
8  *           Hoang-Nam Nguyen <hnguyen@de.ibm.com>
9  *           Joachim Fenkes <fenkes@de.ibm.com>
10  *
11  *  Copyright (c) 2005 IBM Corporation
12  *
13  *  All rights reserved.
14  *
15  *  This source code is distributed under a dual license of GPL v2.0 and OpenIB
16  *  BSD.
17  *
18  * OpenIB BSD License
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions are met:
22  *
23  * Redistributions of source code must retain the above copyright notice, this
24  * list of conditions and the following disclaimer.
25  *
26  * Redistributions in binary form must reproduce the above copyright notice,
27  * this list of conditions and the following disclaimer in the documentation
28  * and/or other materials
29  * provided with the distribution.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
32  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
35  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
36  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
37  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
38  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER
39  * IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
40  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
41  * POSSIBILITY OF SUCH DAMAGE.
42  */
43 
44 #include <linux/slab.h>
45 
46 #include "ehca_classes.h"
47 #include "ehca_irq.h"
48 #include "ehca_iverbs.h"
49 #include "ehca_tools.h"
50 #include "hcp_if.h"
51 #include "hipz_fns.h"
52 #include "ipz_pt_fn.h"
53 
54 #define EQE_COMPLETION_EVENT   EHCA_BMASK_IBM( 1,  1)
55 #define EQE_CQ_QP_NUMBER       EHCA_BMASK_IBM( 8, 31)
56 #define EQE_EE_IDENTIFIER      EHCA_BMASK_IBM( 2,  7)
57 #define EQE_CQ_NUMBER          EHCA_BMASK_IBM( 8, 31)
58 #define EQE_QP_NUMBER          EHCA_BMASK_IBM( 8, 31)
59 #define EQE_QP_TOKEN           EHCA_BMASK_IBM(32, 63)
60 #define EQE_CQ_TOKEN           EHCA_BMASK_IBM(32, 63)
61 
62 #define NEQE_COMPLETION_EVENT  EHCA_BMASK_IBM( 1,  1)
63 #define NEQE_EVENT_CODE        EHCA_BMASK_IBM( 2,  7)
64 #define NEQE_PORT_NUMBER       EHCA_BMASK_IBM( 8, 15)
65 #define NEQE_PORT_AVAILABILITY EHCA_BMASK_IBM(16, 16)
66 #define NEQE_DISRUPTIVE        EHCA_BMASK_IBM(16, 16)
67 #define NEQE_SPECIFIC_EVENT    EHCA_BMASK_IBM(16, 23)
68 
69 #define ERROR_DATA_LENGTH      EHCA_BMASK_IBM(52, 63)
70 #define ERROR_DATA_TYPE        EHCA_BMASK_IBM( 0,  7)
71 
72 static void queue_comp_task(struct ehca_cq *__cq);
73 
74 static struct ehca_comp_pool *pool;
75 
comp_event_callback(struct ehca_cq * cq)76 static inline void comp_event_callback(struct ehca_cq *cq)
77 {
78 	if (!cq->ib_cq.comp_handler)
79 		return;
80 
81 	spin_lock(&cq->cb_lock);
82 	cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
83 	spin_unlock(&cq->cb_lock);
84 
85 	return;
86 }
87 
print_error_data(struct ehca_shca * shca,void * data,u64 * rblock,int length)88 static void print_error_data(struct ehca_shca *shca, void *data,
89 			     u64 *rblock, int length)
90 {
91 	u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
92 	u64 resource = rblock[1];
93 
94 	switch (type) {
95 	case 0x1: /* Queue Pair */
96 	{
97 		struct ehca_qp *qp = (struct ehca_qp *)data;
98 
99 		/* only print error data if AER is set */
100 		if (rblock[6] == 0)
101 			return;
102 
103 		ehca_err(&shca->ib_device,
104 			 "QP 0x%x (resource=%llx) has errors.",
105 			 qp->ib_qp.qp_num, resource);
106 		break;
107 	}
108 	case 0x4: /* Completion Queue */
109 	{
110 		struct ehca_cq *cq = (struct ehca_cq *)data;
111 
112 		ehca_err(&shca->ib_device,
113 			 "CQ 0x%x (resource=%llx) has errors.",
114 			 cq->cq_number, resource);
115 		break;
116 	}
117 	default:
118 		ehca_err(&shca->ib_device,
119 			 "Unknown error type: %llx on %s.",
120 			 type, shca->ib_device.name);
121 		break;
122 	}
123 
124 	ehca_err(&shca->ib_device, "Error data is available: %llx.", resource);
125 	ehca_err(&shca->ib_device, "EHCA ----- error data begin "
126 		 "---------------------------------------------------");
127 	ehca_dmp(rblock, length, "resource=%llx", resource);
128 	ehca_err(&shca->ib_device, "EHCA ----- error data end "
129 		 "----------------------------------------------------");
130 
131 	return;
132 }
133 
ehca_error_data(struct ehca_shca * shca,void * data,u64 resource)134 int ehca_error_data(struct ehca_shca *shca, void *data,
135 		    u64 resource)
136 {
137 
138 	unsigned long ret;
139 	u64 *rblock;
140 	unsigned long block_count;
141 
142 	rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
143 	if (!rblock) {
144 		ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
145 		ret = -ENOMEM;
146 		goto error_data1;
147 	}
148 
149 	/* rblock must be 4K aligned and should be 4K large */
150 	ret = hipz_h_error_data(shca->ipz_hca_handle,
151 				resource,
152 				rblock,
153 				&block_count);
154 
155 	if (ret == H_R_STATE)
156 		ehca_err(&shca->ib_device,
157 			 "No error data is available: %llx.", resource);
158 	else if (ret == H_SUCCESS) {
159 		int length;
160 
161 		length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
162 
163 		if (length > EHCA_PAGESIZE)
164 			length = EHCA_PAGESIZE;
165 
166 		print_error_data(shca, data, rblock, length);
167 	} else
168 		ehca_err(&shca->ib_device,
169 			 "Error data could not be fetched: %llx", resource);
170 
171 	ehca_free_fw_ctrlblock(rblock);
172 
173 error_data1:
174 	return ret;
175 
176 }
177 
dispatch_qp_event(struct ehca_shca * shca,struct ehca_qp * qp,enum ib_event_type event_type)178 static void dispatch_qp_event(struct ehca_shca *shca, struct ehca_qp *qp,
179 			      enum ib_event_type event_type)
180 {
181 	struct ib_event event;
182 
183 	/* PATH_MIG without the QP ever having been armed is false alarm */
184 	if (event_type == IB_EVENT_PATH_MIG && !qp->mig_armed)
185 		return;
186 
187 	event.device = &shca->ib_device;
188 	event.event = event_type;
189 
190 	if (qp->ext_type == EQPT_SRQ) {
191 		if (!qp->ib_srq.event_handler)
192 			return;
193 
194 		event.element.srq = &qp->ib_srq;
195 		qp->ib_srq.event_handler(&event, qp->ib_srq.srq_context);
196 	} else {
197 		if (!qp->ib_qp.event_handler)
198 			return;
199 
200 		event.element.qp = &qp->ib_qp;
201 		qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
202 	}
203 }
204 
qp_event_callback(struct ehca_shca * shca,u64 eqe,enum ib_event_type event_type,int fatal)205 static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
206 			      enum ib_event_type event_type, int fatal)
207 {
208 	struct ehca_qp *qp;
209 	u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
210 
211 	read_lock(&ehca_qp_idr_lock);
212 	qp = idr_find(&ehca_qp_idr, token);
213 	if (qp)
214 		atomic_inc(&qp->nr_events);
215 	read_unlock(&ehca_qp_idr_lock);
216 
217 	if (!qp)
218 		return;
219 
220 	if (fatal)
221 		ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
222 
223 	dispatch_qp_event(shca, qp, fatal && qp->ext_type == EQPT_SRQ ?
224 			  IB_EVENT_SRQ_ERR : event_type);
225 
226 	/*
227 	 * eHCA only processes one WQE at a time for SRQ base QPs,
228 	 * so the last WQE has been processed as soon as the QP enters
229 	 * error state.
230 	 */
231 	if (fatal && qp->ext_type == EQPT_SRQBASE)
232 		dispatch_qp_event(shca, qp, IB_EVENT_QP_LAST_WQE_REACHED);
233 
234 	if (atomic_dec_and_test(&qp->nr_events))
235 		wake_up(&qp->wait_completion);
236 	return;
237 }
238 
cq_event_callback(struct ehca_shca * shca,u64 eqe)239 static void cq_event_callback(struct ehca_shca *shca,
240 			      u64 eqe)
241 {
242 	struct ehca_cq *cq;
243 	u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
244 
245 	read_lock(&ehca_cq_idr_lock);
246 	cq = idr_find(&ehca_cq_idr, token);
247 	if (cq)
248 		atomic_inc(&cq->nr_events);
249 	read_unlock(&ehca_cq_idr_lock);
250 
251 	if (!cq)
252 		return;
253 
254 	ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
255 
256 	if (atomic_dec_and_test(&cq->nr_events))
257 		wake_up(&cq->wait_completion);
258 
259 	return;
260 }
261 
parse_identifier(struct ehca_shca * shca,u64 eqe)262 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
263 {
264 	u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
265 
266 	switch (identifier) {
267 	case 0x02: /* path migrated */
268 		qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG, 0);
269 		break;
270 	case 0x03: /* communication established */
271 		qp_event_callback(shca, eqe, IB_EVENT_COMM_EST, 0);
272 		break;
273 	case 0x04: /* send queue drained */
274 		qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED, 0);
275 		break;
276 	case 0x05: /* QP error */
277 	case 0x06: /* QP error */
278 		qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL, 1);
279 		break;
280 	case 0x07: /* CQ error */
281 	case 0x08: /* CQ error */
282 		cq_event_callback(shca, eqe);
283 		break;
284 	case 0x09: /* MRMWPTE error */
285 		ehca_err(&shca->ib_device, "MRMWPTE error.");
286 		break;
287 	case 0x0A: /* port event */
288 		ehca_err(&shca->ib_device, "Port event.");
289 		break;
290 	case 0x0B: /* MR access error */
291 		ehca_err(&shca->ib_device, "MR access error.");
292 		break;
293 	case 0x0C: /* EQ error */
294 		ehca_err(&shca->ib_device, "EQ error.");
295 		break;
296 	case 0x0D: /* P/Q_Key mismatch */
297 		ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
298 		break;
299 	case 0x10: /* sampling complete */
300 		ehca_err(&shca->ib_device, "Sampling complete.");
301 		break;
302 	case 0x11: /* unaffiliated access error */
303 		ehca_err(&shca->ib_device, "Unaffiliated access error.");
304 		break;
305 	case 0x12: /* path migrating */
306 		ehca_err(&shca->ib_device, "Path migrating.");
307 		break;
308 	case 0x13: /* interface trace stopped */
309 		ehca_err(&shca->ib_device, "Interface trace stopped.");
310 		break;
311 	case 0x14: /* first error capture info available */
312 		ehca_info(&shca->ib_device, "First error capture available");
313 		break;
314 	case 0x15: /* SRQ limit reached */
315 		qp_event_callback(shca, eqe, IB_EVENT_SRQ_LIMIT_REACHED, 0);
316 		break;
317 	default:
318 		ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
319 			 identifier, shca->ib_device.name);
320 		break;
321 	}
322 
323 	return;
324 }
325 
dispatch_port_event(struct ehca_shca * shca,int port_num,enum ib_event_type type,const char * msg)326 static void dispatch_port_event(struct ehca_shca *shca, int port_num,
327 				enum ib_event_type type, const char *msg)
328 {
329 	struct ib_event event;
330 
331 	ehca_info(&shca->ib_device, "port %d %s.", port_num, msg);
332 	event.device = &shca->ib_device;
333 	event.event = type;
334 	event.element.port_num = port_num;
335 	ib_dispatch_event(&event);
336 }
337 
notify_port_conf_change(struct ehca_shca * shca,int port_num)338 static void notify_port_conf_change(struct ehca_shca *shca, int port_num)
339 {
340 	struct ehca_sma_attr  new_attr;
341 	struct ehca_sma_attr *old_attr = &shca->sport[port_num - 1].saved_attr;
342 
343 	ehca_query_sma_attr(shca, port_num, &new_attr);
344 
345 	if (new_attr.sm_sl  != old_attr->sm_sl ||
346 	    new_attr.sm_lid != old_attr->sm_lid)
347 		dispatch_port_event(shca, port_num, IB_EVENT_SM_CHANGE,
348 				    "SM changed");
349 
350 	if (new_attr.lid != old_attr->lid ||
351 	    new_attr.lmc != old_attr->lmc)
352 		dispatch_port_event(shca, port_num, IB_EVENT_LID_CHANGE,
353 				    "LID changed");
354 
355 	if (new_attr.pkey_tbl_len != old_attr->pkey_tbl_len ||
356 	    memcmp(new_attr.pkeys, old_attr->pkeys,
357 		   sizeof(u16) * new_attr.pkey_tbl_len))
358 		dispatch_port_event(shca, port_num, IB_EVENT_PKEY_CHANGE,
359 				    "P_Key changed");
360 
361 	*old_attr = new_attr;
362 }
363 
364 /* replay modify_qp for sqps -- return 0 if all is well, 1 if AQP1 destroyed */
replay_modify_qp(struct ehca_sport * sport)365 static int replay_modify_qp(struct ehca_sport *sport)
366 {
367 	int aqp1_destroyed;
368 	unsigned long flags;
369 
370 	spin_lock_irqsave(&sport->mod_sqp_lock, flags);
371 
372 	aqp1_destroyed = !sport->ibqp_sqp[IB_QPT_GSI];
373 
374 	if (sport->ibqp_sqp[IB_QPT_SMI])
375 		ehca_recover_sqp(sport->ibqp_sqp[IB_QPT_SMI]);
376 	if (!aqp1_destroyed)
377 		ehca_recover_sqp(sport->ibqp_sqp[IB_QPT_GSI]);
378 
379 	spin_unlock_irqrestore(&sport->mod_sqp_lock, flags);
380 
381 	return aqp1_destroyed;
382 }
383 
parse_ec(struct ehca_shca * shca,u64 eqe)384 static void parse_ec(struct ehca_shca *shca, u64 eqe)
385 {
386 	u8 ec   = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
387 	u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
388 	u8 spec_event;
389 	struct ehca_sport *sport = &shca->sport[port - 1];
390 
391 	switch (ec) {
392 	case 0x30: /* port availability change */
393 		if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
394 			/* only replay modify_qp calls in autodetect mode;
395 			 * if AQP1 was destroyed, the port is already down
396 			 * again and we can drop the event.
397 			 */
398 			if (ehca_nr_ports < 0)
399 				if (replay_modify_qp(sport))
400 					break;
401 
402 			sport->port_state = IB_PORT_ACTIVE;
403 			dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
404 					    "is active");
405 			ehca_query_sma_attr(shca, port, &sport->saved_attr);
406 		} else {
407 			sport->port_state = IB_PORT_DOWN;
408 			dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
409 					    "is inactive");
410 		}
411 		break;
412 	case 0x31:
413 		/* port configuration change
414 		 * disruptive change is caused by
415 		 * LID, PKEY or SM change
416 		 */
417 		if (EHCA_BMASK_GET(NEQE_DISRUPTIVE, eqe)) {
418 			ehca_warn(&shca->ib_device, "disruptive port "
419 				  "%d configuration change", port);
420 
421 			sport->port_state = IB_PORT_DOWN;
422 			dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
423 					    "is inactive");
424 
425 			sport->port_state = IB_PORT_ACTIVE;
426 			dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
427 					    "is active");
428 			ehca_query_sma_attr(shca, port,
429 					    &sport->saved_attr);
430 		} else
431 			notify_port_conf_change(shca, port);
432 		break;
433 	case 0x32: /* adapter malfunction */
434 		ehca_err(&shca->ib_device, "Adapter malfunction.");
435 		break;
436 	case 0x33:  /* trace stopped */
437 		ehca_err(&shca->ib_device, "Traced stopped.");
438 		break;
439 	case 0x34: /* util async event */
440 		spec_event = EHCA_BMASK_GET(NEQE_SPECIFIC_EVENT, eqe);
441 		if (spec_event == 0x80) /* client reregister required */
442 			dispatch_port_event(shca, port,
443 					    IB_EVENT_CLIENT_REREGISTER,
444 					    "client reregister req.");
445 		else
446 			ehca_warn(&shca->ib_device, "Unknown util async "
447 				  "event %x on port %x", spec_event, port);
448 		break;
449 	default:
450 		ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
451 			 ec, shca->ib_device.name);
452 		break;
453 	}
454 
455 	return;
456 }
457 
reset_eq_pending(struct ehca_cq * cq)458 static inline void reset_eq_pending(struct ehca_cq *cq)
459 {
460 	u64 CQx_EP;
461 	struct h_galpa gal = cq->galpas.kernel;
462 
463 	hipz_galpa_store_cq(gal, cqx_ep, 0x0);
464 	CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
465 
466 	return;
467 }
468 
ehca_interrupt_neq(int irq,void * dev_id)469 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
470 {
471 	struct ehca_shca *shca = (struct ehca_shca*)dev_id;
472 
473 	tasklet_hi_schedule(&shca->neq.interrupt_task);
474 
475 	return IRQ_HANDLED;
476 }
477 
ehca_tasklet_neq(unsigned long data)478 void ehca_tasklet_neq(unsigned long data)
479 {
480 	struct ehca_shca *shca = (struct ehca_shca*)data;
481 	struct ehca_eqe *eqe;
482 	u64 ret;
483 
484 	eqe = ehca_poll_eq(shca, &shca->neq);
485 
486 	while (eqe) {
487 		if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
488 			parse_ec(shca, eqe->entry);
489 
490 		eqe = ehca_poll_eq(shca, &shca->neq);
491 	}
492 
493 	ret = hipz_h_reset_event(shca->ipz_hca_handle,
494 				 shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
495 
496 	if (ret != H_SUCCESS)
497 		ehca_err(&shca->ib_device, "Can't clear notification events.");
498 
499 	return;
500 }
501 
ehca_interrupt_eq(int irq,void * dev_id)502 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
503 {
504 	struct ehca_shca *shca = (struct ehca_shca*)dev_id;
505 
506 	tasklet_hi_schedule(&shca->eq.interrupt_task);
507 
508 	return IRQ_HANDLED;
509 }
510 
511 
process_eqe(struct ehca_shca * shca,struct ehca_eqe * eqe)512 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
513 {
514 	u64 eqe_value;
515 	u32 token;
516 	struct ehca_cq *cq;
517 
518 	eqe_value = eqe->entry;
519 	ehca_dbg(&shca->ib_device, "eqe_value=%llx", eqe_value);
520 	if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
521 		ehca_dbg(&shca->ib_device, "Got completion event");
522 		token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
523 		read_lock(&ehca_cq_idr_lock);
524 		cq = idr_find(&ehca_cq_idr, token);
525 		if (cq)
526 			atomic_inc(&cq->nr_events);
527 		read_unlock(&ehca_cq_idr_lock);
528 		if (cq == NULL) {
529 			ehca_err(&shca->ib_device,
530 				 "Invalid eqe for non-existing cq token=%x",
531 				 token);
532 			return;
533 		}
534 		reset_eq_pending(cq);
535 		if (ehca_scaling_code)
536 			queue_comp_task(cq);
537 		else {
538 			comp_event_callback(cq);
539 			if (atomic_dec_and_test(&cq->nr_events))
540 				wake_up(&cq->wait_completion);
541 		}
542 	} else {
543 		ehca_dbg(&shca->ib_device, "Got non completion event");
544 		parse_identifier(shca, eqe_value);
545 	}
546 }
547 
ehca_process_eq(struct ehca_shca * shca,int is_irq)548 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
549 {
550 	struct ehca_eq *eq = &shca->eq;
551 	struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
552 	u64 eqe_value, ret;
553 	int eqe_cnt, i;
554 	int eq_empty = 0;
555 
556 	spin_lock(&eq->irq_spinlock);
557 	if (is_irq) {
558 		const int max_query_cnt = 100;
559 		int query_cnt = 0;
560 		int int_state = 1;
561 		do {
562 			int_state = hipz_h_query_int_state(
563 				shca->ipz_hca_handle, eq->ist);
564 			query_cnt++;
565 			iosync();
566 		} while (int_state && query_cnt < max_query_cnt);
567 		if (unlikely((query_cnt == max_query_cnt)))
568 			ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
569 				 int_state, query_cnt);
570 	}
571 
572 	/* read out all eqes */
573 	eqe_cnt = 0;
574 	do {
575 		u32 token;
576 		eqe_cache[eqe_cnt].eqe = ehca_poll_eq(shca, eq);
577 		if (!eqe_cache[eqe_cnt].eqe)
578 			break;
579 		eqe_value = eqe_cache[eqe_cnt].eqe->entry;
580 		if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
581 			token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
582 			read_lock(&ehca_cq_idr_lock);
583 			eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
584 			if (eqe_cache[eqe_cnt].cq)
585 				atomic_inc(&eqe_cache[eqe_cnt].cq->nr_events);
586 			read_unlock(&ehca_cq_idr_lock);
587 			if (!eqe_cache[eqe_cnt].cq) {
588 				ehca_err(&shca->ib_device,
589 					 "Invalid eqe for non-existing cq "
590 					 "token=%x", token);
591 				continue;
592 			}
593 		} else
594 			eqe_cache[eqe_cnt].cq = NULL;
595 		eqe_cnt++;
596 	} while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
597 	if (!eqe_cnt) {
598 		if (is_irq)
599 			ehca_dbg(&shca->ib_device,
600 				 "No eqe found for irq event");
601 		goto unlock_irq_spinlock;
602 	} else if (!is_irq) {
603 		ret = hipz_h_eoi(eq->ist);
604 		if (ret != H_SUCCESS)
605 			ehca_err(&shca->ib_device,
606 				 "bad return code EOI -rc = %lld\n", ret);
607 		ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
608 	}
609 	if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
610 		ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
611 	/* enable irq for new packets */
612 	for (i = 0; i < eqe_cnt; i++) {
613 		if (eq->eqe_cache[i].cq)
614 			reset_eq_pending(eq->eqe_cache[i].cq);
615 	}
616 	/* check eq */
617 	spin_lock(&eq->spinlock);
618 	eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
619 	spin_unlock(&eq->spinlock);
620 	/* call completion handler for cached eqes */
621 	for (i = 0; i < eqe_cnt; i++)
622 		if (eq->eqe_cache[i].cq) {
623 			if (ehca_scaling_code)
624 				queue_comp_task(eq->eqe_cache[i].cq);
625 			else {
626 				struct ehca_cq *cq = eq->eqe_cache[i].cq;
627 				comp_event_callback(cq);
628 				if (atomic_dec_and_test(&cq->nr_events))
629 					wake_up(&cq->wait_completion);
630 			}
631 		} else {
632 			ehca_dbg(&shca->ib_device, "Got non completion event");
633 			parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
634 		}
635 	/* poll eq if not empty */
636 	if (eq_empty)
637 		goto unlock_irq_spinlock;
638 	do {
639 		struct ehca_eqe *eqe;
640 		eqe = ehca_poll_eq(shca, &shca->eq);
641 		if (!eqe)
642 			break;
643 		process_eqe(shca, eqe);
644 	} while (1);
645 
646 unlock_irq_spinlock:
647 	spin_unlock(&eq->irq_spinlock);
648 }
649 
ehca_tasklet_eq(unsigned long data)650 void ehca_tasklet_eq(unsigned long data)
651 {
652 	ehca_process_eq((struct ehca_shca*)data, 1);
653 }
654 
find_next_online_cpu(struct ehca_comp_pool * pool)655 static inline int find_next_online_cpu(struct ehca_comp_pool *pool)
656 {
657 	int cpu;
658 	unsigned long flags;
659 
660 	WARN_ON_ONCE(!in_interrupt());
661 	if (ehca_debug_level >= 3)
662 		ehca_dmp(cpu_online_mask, cpumask_size(), "");
663 
664 	spin_lock_irqsave(&pool->last_cpu_lock, flags);
665 	cpu = cpumask_next(pool->last_cpu, cpu_online_mask);
666 	if (cpu >= nr_cpu_ids)
667 		cpu = cpumask_first(cpu_online_mask);
668 	pool->last_cpu = cpu;
669 	spin_unlock_irqrestore(&pool->last_cpu_lock, flags);
670 
671 	return cpu;
672 }
673 
__queue_comp_task(struct ehca_cq * __cq,struct ehca_cpu_comp_task * cct)674 static void __queue_comp_task(struct ehca_cq *__cq,
675 			      struct ehca_cpu_comp_task *cct)
676 {
677 	unsigned long flags;
678 
679 	spin_lock_irqsave(&cct->task_lock, flags);
680 	spin_lock(&__cq->task_lock);
681 
682 	if (__cq->nr_callbacks == 0) {
683 		__cq->nr_callbacks++;
684 		list_add_tail(&__cq->entry, &cct->cq_list);
685 		cct->cq_jobs++;
686 		wake_up(&cct->wait_queue);
687 	} else
688 		__cq->nr_callbacks++;
689 
690 	spin_unlock(&__cq->task_lock);
691 	spin_unlock_irqrestore(&cct->task_lock, flags);
692 }
693 
queue_comp_task(struct ehca_cq * __cq)694 static void queue_comp_task(struct ehca_cq *__cq)
695 {
696 	int cpu_id;
697 	struct ehca_cpu_comp_task *cct;
698 	int cq_jobs;
699 	unsigned long flags;
700 
701 	cpu_id = find_next_online_cpu(pool);
702 	BUG_ON(!cpu_online(cpu_id));
703 
704 	cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
705 	BUG_ON(!cct);
706 
707 	spin_lock_irqsave(&cct->task_lock, flags);
708 	cq_jobs = cct->cq_jobs;
709 	spin_unlock_irqrestore(&cct->task_lock, flags);
710 	if (cq_jobs > 0) {
711 		cpu_id = find_next_online_cpu(pool);
712 		cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
713 		BUG_ON(!cct);
714 	}
715 
716 	__queue_comp_task(__cq, cct);
717 }
718 
run_comp_task(struct ehca_cpu_comp_task * cct)719 static void run_comp_task(struct ehca_cpu_comp_task *cct)
720 {
721 	struct ehca_cq *cq;
722 	unsigned long flags;
723 
724 	spin_lock_irqsave(&cct->task_lock, flags);
725 
726 	while (!list_empty(&cct->cq_list)) {
727 		cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
728 		spin_unlock_irqrestore(&cct->task_lock, flags);
729 
730 		comp_event_callback(cq);
731 		if (atomic_dec_and_test(&cq->nr_events))
732 			wake_up(&cq->wait_completion);
733 
734 		spin_lock_irqsave(&cct->task_lock, flags);
735 		spin_lock(&cq->task_lock);
736 		cq->nr_callbacks--;
737 		if (!cq->nr_callbacks) {
738 			list_del_init(cct->cq_list.next);
739 			cct->cq_jobs--;
740 		}
741 		spin_unlock(&cq->task_lock);
742 	}
743 
744 	spin_unlock_irqrestore(&cct->task_lock, flags);
745 }
746 
comp_task(void * __cct)747 static int comp_task(void *__cct)
748 {
749 	struct ehca_cpu_comp_task *cct = __cct;
750 	int cql_empty;
751 	DECLARE_WAITQUEUE(wait, current);
752 
753 	set_current_state(TASK_INTERRUPTIBLE);
754 	while (!kthread_should_stop()) {
755 		add_wait_queue(&cct->wait_queue, &wait);
756 
757 		spin_lock_irq(&cct->task_lock);
758 		cql_empty = list_empty(&cct->cq_list);
759 		spin_unlock_irq(&cct->task_lock);
760 		if (cql_empty)
761 			schedule();
762 		else
763 			__set_current_state(TASK_RUNNING);
764 
765 		remove_wait_queue(&cct->wait_queue, &wait);
766 
767 		spin_lock_irq(&cct->task_lock);
768 		cql_empty = list_empty(&cct->cq_list);
769 		spin_unlock_irq(&cct->task_lock);
770 		if (!cql_empty)
771 			run_comp_task(__cct);
772 
773 		set_current_state(TASK_INTERRUPTIBLE);
774 	}
775 	__set_current_state(TASK_RUNNING);
776 
777 	return 0;
778 }
779 
create_comp_task(struct ehca_comp_pool * pool,int cpu)780 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
781 					    int cpu)
782 {
783 	struct ehca_cpu_comp_task *cct;
784 
785 	cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
786 	spin_lock_init(&cct->task_lock);
787 	INIT_LIST_HEAD(&cct->cq_list);
788 	init_waitqueue_head(&cct->wait_queue);
789 	cct->task = kthread_create_on_node(comp_task, cct, cpu_to_node(cpu),
790 					   "ehca_comp/%d", cpu);
791 
792 	return cct->task;
793 }
794 
destroy_comp_task(struct ehca_comp_pool * pool,int cpu)795 static void destroy_comp_task(struct ehca_comp_pool *pool,
796 			      int cpu)
797 {
798 	struct ehca_cpu_comp_task *cct;
799 	struct task_struct *task;
800 	unsigned long flags_cct;
801 
802 	cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
803 
804 	spin_lock_irqsave(&cct->task_lock, flags_cct);
805 
806 	task = cct->task;
807 	cct->task = NULL;
808 	cct->cq_jobs = 0;
809 
810 	spin_unlock_irqrestore(&cct->task_lock, flags_cct);
811 
812 	if (task)
813 		kthread_stop(task);
814 }
815 
take_over_work(struct ehca_comp_pool * pool,int cpu)816 static void __cpuinit take_over_work(struct ehca_comp_pool *pool, int cpu)
817 {
818 	struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
819 	LIST_HEAD(list);
820 	struct ehca_cq *cq;
821 	unsigned long flags_cct;
822 
823 	spin_lock_irqsave(&cct->task_lock, flags_cct);
824 
825 	list_splice_init(&cct->cq_list, &list);
826 
827 	while (!list_empty(&list)) {
828 		cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
829 
830 		list_del(&cq->entry);
831 		__queue_comp_task(cq, this_cpu_ptr(pool->cpu_comp_tasks));
832 	}
833 
834 	spin_unlock_irqrestore(&cct->task_lock, flags_cct);
835 
836 }
837 
comp_pool_callback(struct notifier_block * nfb,unsigned long action,void * hcpu)838 static int __cpuinit comp_pool_callback(struct notifier_block *nfb,
839 					unsigned long action,
840 					void *hcpu)
841 {
842 	unsigned int cpu = (unsigned long)hcpu;
843 	struct ehca_cpu_comp_task *cct;
844 
845 	switch (action) {
846 	case CPU_UP_PREPARE:
847 	case CPU_UP_PREPARE_FROZEN:
848 		ehca_gen_dbg("CPU: %x (CPU_PREPARE)", cpu);
849 		if (!create_comp_task(pool, cpu)) {
850 			ehca_gen_err("Can't create comp_task for cpu: %x", cpu);
851 			return notifier_from_errno(-ENOMEM);
852 		}
853 		break;
854 	case CPU_UP_CANCELED:
855 	case CPU_UP_CANCELED_FROZEN:
856 		ehca_gen_dbg("CPU: %x (CPU_CANCELED)", cpu);
857 		cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
858 		kthread_bind(cct->task, cpumask_any(cpu_online_mask));
859 		destroy_comp_task(pool, cpu);
860 		break;
861 	case CPU_ONLINE:
862 	case CPU_ONLINE_FROZEN:
863 		ehca_gen_dbg("CPU: %x (CPU_ONLINE)", cpu);
864 		cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
865 		kthread_bind(cct->task, cpu);
866 		wake_up_process(cct->task);
867 		break;
868 	case CPU_DOWN_PREPARE:
869 	case CPU_DOWN_PREPARE_FROZEN:
870 		ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
871 		break;
872 	case CPU_DOWN_FAILED:
873 	case CPU_DOWN_FAILED_FROZEN:
874 		ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
875 		break;
876 	case CPU_DEAD:
877 	case CPU_DEAD_FROZEN:
878 		ehca_gen_dbg("CPU: %x (CPU_DEAD)", cpu);
879 		destroy_comp_task(pool, cpu);
880 		take_over_work(pool, cpu);
881 		break;
882 	}
883 
884 	return NOTIFY_OK;
885 }
886 
887 static struct notifier_block comp_pool_callback_nb __cpuinitdata = {
888 	.notifier_call	= comp_pool_callback,
889 	.priority	= 0,
890 };
891 
ehca_create_comp_pool(void)892 int ehca_create_comp_pool(void)
893 {
894 	int cpu;
895 	struct task_struct *task;
896 
897 	if (!ehca_scaling_code)
898 		return 0;
899 
900 	pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
901 	if (pool == NULL)
902 		return -ENOMEM;
903 
904 	spin_lock_init(&pool->last_cpu_lock);
905 	pool->last_cpu = cpumask_any(cpu_online_mask);
906 
907 	pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
908 	if (pool->cpu_comp_tasks == NULL) {
909 		kfree(pool);
910 		return -EINVAL;
911 	}
912 
913 	for_each_online_cpu(cpu) {
914 		task = create_comp_task(pool, cpu);
915 		if (task) {
916 			kthread_bind(task, cpu);
917 			wake_up_process(task);
918 		}
919 	}
920 
921 	register_hotcpu_notifier(&comp_pool_callback_nb);
922 
923 	printk(KERN_INFO "eHCA scaling code enabled\n");
924 
925 	return 0;
926 }
927 
ehca_destroy_comp_pool(void)928 void ehca_destroy_comp_pool(void)
929 {
930 	int i;
931 
932 	if (!ehca_scaling_code)
933 		return;
934 
935 	unregister_hotcpu_notifier(&comp_pool_callback_nb);
936 
937 	for_each_online_cpu(i)
938 		destroy_comp_task(pool, i);
939 
940 	free_percpu(pool->cpu_comp_tasks);
941 	kfree(pool);
942 }
943