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