2 * IBM eServer eHCA Infiniband device driver for Linux on POWER
4 * Functions for EQs, NEQs and interrupts
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>
11 * Copyright (c) 2005 IBM Corporation
13 * All rights reserved.
15 * This source code is distributed under a dual license of GPL v2.0 and OpenIB
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions are met:
23 * Redistributions of source code must retain the above copyright notice, this
24 * list of conditions and the following disclaimer.
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.
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.
44 #include "ehca_classes.h"
46 #include "ehca_iverbs.h"
47 #include "ehca_tools.h"
50 #include "ipz_pt_fn.h"
52 #define EQE_COMPLETION_EVENT EHCA_BMASK_IBM( 1, 1)
53 #define EQE_CQ_QP_NUMBER EHCA_BMASK_IBM( 8, 31)
54 #define EQE_EE_IDENTIFIER EHCA_BMASK_IBM( 2, 7)
55 #define EQE_CQ_NUMBER EHCA_BMASK_IBM( 8, 31)
56 #define EQE_QP_NUMBER EHCA_BMASK_IBM( 8, 31)
57 #define EQE_QP_TOKEN EHCA_BMASK_IBM(32, 63)
58 #define EQE_CQ_TOKEN EHCA_BMASK_IBM(32, 63)
60 #define NEQE_COMPLETION_EVENT EHCA_BMASK_IBM( 1, 1)
61 #define NEQE_EVENT_CODE EHCA_BMASK_IBM( 2, 7)
62 #define NEQE_PORT_NUMBER EHCA_BMASK_IBM( 8, 15)
63 #define NEQE_PORT_AVAILABILITY EHCA_BMASK_IBM(16, 16)
64 #define NEQE_DISRUPTIVE EHCA_BMASK_IBM(16, 16)
65 #define NEQE_SPECIFIC_EVENT EHCA_BMASK_IBM(16, 23)
67 #define ERROR_DATA_LENGTH EHCA_BMASK_IBM(52, 63)
68 #define ERROR_DATA_TYPE EHCA_BMASK_IBM( 0, 7)
70 static void queue_comp_task(struct ehca_cq *__cq);
72 static struct ehca_comp_pool *pool;
74 static inline void comp_event_callback(struct ehca_cq *cq)
76 if (!cq->ib_cq.comp_handler)
79 spin_lock(&cq->cb_lock);
80 cq->ib_cq.comp_handler(&cq->ib_cq, cq->ib_cq.cq_context);
81 spin_unlock(&cq->cb_lock);
86 static void print_error_data(struct ehca_shca *shca, void *data,
87 u64 *rblock, int length)
89 u64 type = EHCA_BMASK_GET(ERROR_DATA_TYPE, rblock[2]);
90 u64 resource = rblock[1];
93 case 0x1: /* Queue Pair */
95 struct ehca_qp *qp = (struct ehca_qp *)data;
97 /* only print error data if AER is set */
101 ehca_err(&shca->ib_device,
102 "QP 0x%x (resource=%llx) has errors.",
103 qp->ib_qp.qp_num, resource);
106 case 0x4: /* Completion Queue */
108 struct ehca_cq *cq = (struct ehca_cq *)data;
110 ehca_err(&shca->ib_device,
111 "CQ 0x%x (resource=%llx) has errors.",
112 cq->cq_number, resource);
116 ehca_err(&shca->ib_device,
117 "Unknown error type: %llx on %s.",
118 type, shca->ib_device.name);
122 ehca_err(&shca->ib_device, "Error data is available: %llx.", resource);
123 ehca_err(&shca->ib_device, "EHCA ----- error data begin "
124 "---------------------------------------------------");
125 ehca_dmp(rblock, length, "resource=%llx", resource);
126 ehca_err(&shca->ib_device, "EHCA ----- error data end "
127 "----------------------------------------------------");
132 int ehca_error_data(struct ehca_shca *shca, void *data,
138 unsigned long block_count;
140 rblock = ehca_alloc_fw_ctrlblock(GFP_ATOMIC);
142 ehca_err(&shca->ib_device, "Cannot allocate rblock memory.");
147 /* rblock must be 4K aligned and should be 4K large */
148 ret = hipz_h_error_data(shca->ipz_hca_handle,
153 if (ret == H_R_STATE)
154 ehca_err(&shca->ib_device,
155 "No error data is available: %llx.", resource);
156 else if (ret == H_SUCCESS) {
159 length = EHCA_BMASK_GET(ERROR_DATA_LENGTH, rblock[0]);
161 if (length > EHCA_PAGESIZE)
162 length = EHCA_PAGESIZE;
164 print_error_data(shca, data, rblock, length);
166 ehca_err(&shca->ib_device,
167 "Error data could not be fetched: %llx", resource);
169 ehca_free_fw_ctrlblock(rblock);
176 static void dispatch_qp_event(struct ehca_shca *shca, struct ehca_qp *qp,
177 enum ib_event_type event_type)
179 struct ib_event event;
181 /* PATH_MIG without the QP ever having been armed is false alarm */
182 if (event_type == IB_EVENT_PATH_MIG && !qp->mig_armed)
185 event.device = &shca->ib_device;
186 event.event = event_type;
188 if (qp->ext_type == EQPT_SRQ) {
189 if (!qp->ib_srq.event_handler)
192 event.element.srq = &qp->ib_srq;
193 qp->ib_srq.event_handler(&event, qp->ib_srq.srq_context);
195 if (!qp->ib_qp.event_handler)
198 event.element.qp = &qp->ib_qp;
199 qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
203 static void qp_event_callback(struct ehca_shca *shca, u64 eqe,
204 enum ib_event_type event_type, int fatal)
207 u32 token = EHCA_BMASK_GET(EQE_QP_TOKEN, eqe);
209 read_lock(&ehca_qp_idr_lock);
210 qp = idr_find(&ehca_qp_idr, token);
212 atomic_inc(&qp->nr_events);
213 read_unlock(&ehca_qp_idr_lock);
219 ehca_error_data(shca, qp, qp->ipz_qp_handle.handle);
221 dispatch_qp_event(shca, qp, fatal && qp->ext_type == EQPT_SRQ ?
222 IB_EVENT_SRQ_ERR : event_type);
225 * eHCA only processes one WQE at a time for SRQ base QPs,
226 * so the last WQE has been processed as soon as the QP enters
229 if (fatal && qp->ext_type == EQPT_SRQBASE)
230 dispatch_qp_event(shca, qp, IB_EVENT_QP_LAST_WQE_REACHED);
232 if (atomic_dec_and_test(&qp->nr_events))
233 wake_up(&qp->wait_completion);
237 static void cq_event_callback(struct ehca_shca *shca,
241 u32 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe);
243 read_lock(&ehca_cq_idr_lock);
244 cq = idr_find(&ehca_cq_idr, token);
246 atomic_inc(&cq->nr_events);
247 read_unlock(&ehca_cq_idr_lock);
252 ehca_error_data(shca, cq, cq->ipz_cq_handle.handle);
254 if (atomic_dec_and_test(&cq->nr_events))
255 wake_up(&cq->wait_completion);
260 static void parse_identifier(struct ehca_shca *shca, u64 eqe)
262 u8 identifier = EHCA_BMASK_GET(EQE_EE_IDENTIFIER, eqe);
264 switch (identifier) {
265 case 0x02: /* path migrated */
266 qp_event_callback(shca, eqe, IB_EVENT_PATH_MIG, 0);
268 case 0x03: /* communication established */
269 qp_event_callback(shca, eqe, IB_EVENT_COMM_EST, 0);
271 case 0x04: /* send queue drained */
272 qp_event_callback(shca, eqe, IB_EVENT_SQ_DRAINED, 0);
274 case 0x05: /* QP error */
275 case 0x06: /* QP error */
276 qp_event_callback(shca, eqe, IB_EVENT_QP_FATAL, 1);
278 case 0x07: /* CQ error */
279 case 0x08: /* CQ error */
280 cq_event_callback(shca, eqe);
282 case 0x09: /* MRMWPTE error */
283 ehca_err(&shca->ib_device, "MRMWPTE error.");
285 case 0x0A: /* port event */
286 ehca_err(&shca->ib_device, "Port event.");
288 case 0x0B: /* MR access error */
289 ehca_err(&shca->ib_device, "MR access error.");
291 case 0x0C: /* EQ error */
292 ehca_err(&shca->ib_device, "EQ error.");
294 case 0x0D: /* P/Q_Key mismatch */
295 ehca_err(&shca->ib_device, "P/Q_Key mismatch.");
297 case 0x10: /* sampling complete */
298 ehca_err(&shca->ib_device, "Sampling complete.");
300 case 0x11: /* unaffiliated access error */
301 ehca_err(&shca->ib_device, "Unaffiliated access error.");
303 case 0x12: /* path migrating */
304 ehca_err(&shca->ib_device, "Path migrating.");
306 case 0x13: /* interface trace stopped */
307 ehca_err(&shca->ib_device, "Interface trace stopped.");
309 case 0x14: /* first error capture info available */
310 ehca_info(&shca->ib_device, "First error capture available");
312 case 0x15: /* SRQ limit reached */
313 qp_event_callback(shca, eqe, IB_EVENT_SRQ_LIMIT_REACHED, 0);
316 ehca_err(&shca->ib_device, "Unknown identifier: %x on %s.",
317 identifier, shca->ib_device.name);
324 static void dispatch_port_event(struct ehca_shca *shca, int port_num,
325 enum ib_event_type type, const char *msg)
327 struct ib_event event;
329 ehca_info(&shca->ib_device, "port %d %s.", port_num, msg);
330 event.device = &shca->ib_device;
332 event.element.port_num = port_num;
333 ib_dispatch_event(&event);
336 static void notify_port_conf_change(struct ehca_shca *shca, int port_num)
338 struct ehca_sma_attr new_attr;
339 struct ehca_sma_attr *old_attr = &shca->sport[port_num - 1].saved_attr;
341 ehca_query_sma_attr(shca, port_num, &new_attr);
343 if (new_attr.sm_sl != old_attr->sm_sl ||
344 new_attr.sm_lid != old_attr->sm_lid)
345 dispatch_port_event(shca, port_num, IB_EVENT_SM_CHANGE,
348 if (new_attr.lid != old_attr->lid ||
349 new_attr.lmc != old_attr->lmc)
350 dispatch_port_event(shca, port_num, IB_EVENT_LID_CHANGE,
353 if (new_attr.pkey_tbl_len != old_attr->pkey_tbl_len ||
354 memcmp(new_attr.pkeys, old_attr->pkeys,
355 sizeof(u16) * new_attr.pkey_tbl_len))
356 dispatch_port_event(shca, port_num, IB_EVENT_PKEY_CHANGE,
359 *old_attr = new_attr;
362 /* replay modify_qp for sqps -- return 0 if all is well, 1 if AQP1 destroyed */
363 static int replay_modify_qp(struct ehca_sport *sport)
368 spin_lock_irqsave(&sport->mod_sqp_lock, flags);
370 aqp1_destroyed = !sport->ibqp_sqp[IB_QPT_GSI];
372 if (sport->ibqp_sqp[IB_QPT_SMI])
373 ehca_recover_sqp(sport->ibqp_sqp[IB_QPT_SMI]);
375 ehca_recover_sqp(sport->ibqp_sqp[IB_QPT_GSI]);
377 spin_unlock_irqrestore(&sport->mod_sqp_lock, flags);
379 return aqp1_destroyed;
382 static void parse_ec(struct ehca_shca *shca, u64 eqe)
384 u8 ec = EHCA_BMASK_GET(NEQE_EVENT_CODE, eqe);
385 u8 port = EHCA_BMASK_GET(NEQE_PORT_NUMBER, eqe);
387 struct ehca_sport *sport = &shca->sport[port - 1];
390 case 0x30: /* port availability change */
391 if (EHCA_BMASK_GET(NEQE_PORT_AVAILABILITY, eqe)) {
392 /* only replay modify_qp calls in autodetect mode;
393 * if AQP1 was destroyed, the port is already down
394 * again and we can drop the event.
396 if (ehca_nr_ports < 0)
397 if (replay_modify_qp(sport))
400 sport->port_state = IB_PORT_ACTIVE;
401 dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
403 ehca_query_sma_attr(shca, port, &sport->saved_attr);
405 sport->port_state = IB_PORT_DOWN;
406 dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
411 /* port configuration change
412 * disruptive change is caused by
413 * LID, PKEY or SM change
415 if (EHCA_BMASK_GET(NEQE_DISRUPTIVE, eqe)) {
416 ehca_warn(&shca->ib_device, "disruptive port "
417 "%d configuration change", port);
419 sport->port_state = IB_PORT_DOWN;
420 dispatch_port_event(shca, port, IB_EVENT_PORT_ERR,
423 sport->port_state = IB_PORT_ACTIVE;
424 dispatch_port_event(shca, port, IB_EVENT_PORT_ACTIVE,
426 ehca_query_sma_attr(shca, port,
429 notify_port_conf_change(shca, port);
431 case 0x32: /* adapter malfunction */
432 ehca_err(&shca->ib_device, "Adapter malfunction.");
434 case 0x33: /* trace stopped */
435 ehca_err(&shca->ib_device, "Traced stopped.");
437 case 0x34: /* util async event */
438 spec_event = EHCA_BMASK_GET(NEQE_SPECIFIC_EVENT, eqe);
439 if (spec_event == 0x80) /* client reregister required */
440 dispatch_port_event(shca, port,
441 IB_EVENT_CLIENT_REREGISTER,
442 "client reregister req.");
444 ehca_warn(&shca->ib_device, "Unknown util async "
445 "event %x on port %x", spec_event, port);
448 ehca_err(&shca->ib_device, "Unknown event code: %x on %s.",
449 ec, shca->ib_device.name);
456 static inline void reset_eq_pending(struct ehca_cq *cq)
459 struct h_galpa gal = cq->galpas.kernel;
461 hipz_galpa_store_cq(gal, cqx_ep, 0x0);
462 CQx_EP = hipz_galpa_load(gal, CQTEMM_OFFSET(cqx_ep));
467 irqreturn_t ehca_interrupt_neq(int irq, void *dev_id)
469 struct ehca_shca *shca = (struct ehca_shca*)dev_id;
471 tasklet_hi_schedule(&shca->neq.interrupt_task);
476 void ehca_tasklet_neq(unsigned long data)
478 struct ehca_shca *shca = (struct ehca_shca*)data;
479 struct ehca_eqe *eqe;
482 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
485 if (!EHCA_BMASK_GET(NEQE_COMPLETION_EVENT, eqe->entry))
486 parse_ec(shca, eqe->entry);
488 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->neq);
491 ret = hipz_h_reset_event(shca->ipz_hca_handle,
492 shca->neq.ipz_eq_handle, 0xFFFFFFFFFFFFFFFFL);
494 if (ret != H_SUCCESS)
495 ehca_err(&shca->ib_device, "Can't clear notification events.");
500 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
502 struct ehca_shca *shca = (struct ehca_shca*)dev_id;
504 tasklet_hi_schedule(&shca->eq.interrupt_task);
510 static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
516 eqe_value = eqe->entry;
517 ehca_dbg(&shca->ib_device, "eqe_value=%llx", eqe_value);
518 if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
519 ehca_dbg(&shca->ib_device, "Got completion event");
520 token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
521 read_lock(&ehca_cq_idr_lock);
522 cq = idr_find(&ehca_cq_idr, token);
524 atomic_inc(&cq->nr_events);
525 read_unlock(&ehca_cq_idr_lock);
527 ehca_err(&shca->ib_device,
528 "Invalid eqe for non-existing cq token=%x",
532 reset_eq_pending(cq);
533 if (ehca_scaling_code)
536 comp_event_callback(cq);
537 if (atomic_dec_and_test(&cq->nr_events))
538 wake_up(&cq->wait_completion);
541 ehca_dbg(&shca->ib_device, "Got non completion event");
542 parse_identifier(shca, eqe_value);
546 void ehca_process_eq(struct ehca_shca *shca, int is_irq)
548 struct ehca_eq *eq = &shca->eq;
549 struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
555 spin_lock_irqsave(&eq->irq_spinlock, flags);
557 const int max_query_cnt = 100;
561 int_state = hipz_h_query_int_state(
562 shca->ipz_hca_handle, eq->ist);
565 } while (int_state && query_cnt < max_query_cnt);
566 if (unlikely((query_cnt == max_query_cnt)))
567 ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
568 int_state, query_cnt);
571 /* read out all eqes */
575 eqe_cache[eqe_cnt].eqe =
576 (struct ehca_eqe *)ehca_poll_eq(shca, eq);
577 if (!eqe_cache[eqe_cnt].eqe)
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 "
594 eqe_cache[eqe_cnt].cq = NULL;
596 } while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
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);
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);
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);
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);
632 ehca_dbg(&shca->ib_device, "Got non completion event");
633 parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
635 /* poll eq if not empty */
637 goto unlock_irq_spinlock;
639 struct ehca_eqe *eqe;
640 eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
643 process_eqe(shca, eqe);
647 spin_unlock_irqrestore(&eq->irq_spinlock, flags);
650 void ehca_tasklet_eq(unsigned long data)
652 ehca_process_eq((struct ehca_shca*)data, 1);
655 static inline int find_next_online_cpu(struct ehca_comp_pool *pool)
660 WARN_ON_ONCE(!in_interrupt());
661 if (ehca_debug_level >= 3)
662 ehca_dmp(cpu_online_mask, cpumask_size(), "");
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);
674 static void __queue_comp_task(struct ehca_cq *__cq,
675 struct ehca_cpu_comp_task *cct)
679 spin_lock_irqsave(&cct->task_lock, flags);
680 spin_lock(&__cq->task_lock);
682 if (__cq->nr_callbacks == 0) {
683 __cq->nr_callbacks++;
684 list_add_tail(&__cq->entry, &cct->cq_list);
686 wake_up(&cct->wait_queue);
688 __cq->nr_callbacks++;
690 spin_unlock(&__cq->task_lock);
691 spin_unlock_irqrestore(&cct->task_lock, flags);
694 static void queue_comp_task(struct ehca_cq *__cq)
697 struct ehca_cpu_comp_task *cct;
701 cpu_id = find_next_online_cpu(pool);
702 BUG_ON(!cpu_online(cpu_id));
704 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
707 spin_lock_irqsave(&cct->task_lock, flags);
708 cq_jobs = cct->cq_jobs;
709 spin_unlock_irqrestore(&cct->task_lock, flags);
711 cpu_id = find_next_online_cpu(pool);
712 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
716 __queue_comp_task(__cq, cct);
719 static void run_comp_task(struct ehca_cpu_comp_task *cct)
724 spin_lock_irqsave(&cct->task_lock, flags);
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);
730 comp_event_callback(cq);
731 if (atomic_dec_and_test(&cq->nr_events))
732 wake_up(&cq->wait_completion);
734 spin_lock_irqsave(&cct->task_lock, flags);
735 spin_lock(&cq->task_lock);
737 if (!cq->nr_callbacks) {
738 list_del_init(cct->cq_list.next);
741 spin_unlock(&cq->task_lock);
744 spin_unlock_irqrestore(&cct->task_lock, flags);
747 static int comp_task(void *__cct)
749 struct ehca_cpu_comp_task *cct = __cct;
751 DECLARE_WAITQUEUE(wait, current);
753 set_current_state(TASK_INTERRUPTIBLE);
754 while (!kthread_should_stop()) {
755 add_wait_queue(&cct->wait_queue, &wait);
757 spin_lock_irq(&cct->task_lock);
758 cql_empty = list_empty(&cct->cq_list);
759 spin_unlock_irq(&cct->task_lock);
763 __set_current_state(TASK_RUNNING);
765 remove_wait_queue(&cct->wait_queue, &wait);
767 spin_lock_irq(&cct->task_lock);
768 cql_empty = list_empty(&cct->cq_list);
769 spin_unlock_irq(&cct->task_lock);
771 run_comp_task(__cct);
773 set_current_state(TASK_INTERRUPTIBLE);
775 __set_current_state(TASK_RUNNING);
780 static struct task_struct *create_comp_task(struct ehca_comp_pool *pool,
783 struct ehca_cpu_comp_task *cct;
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(comp_task, cct, "ehca_comp/%d", cpu);
794 static void destroy_comp_task(struct ehca_comp_pool *pool,
797 struct ehca_cpu_comp_task *cct;
798 struct task_struct *task;
799 unsigned long flags_cct;
801 cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
803 spin_lock_irqsave(&cct->task_lock, flags_cct);
809 spin_unlock_irqrestore(&cct->task_lock, flags_cct);
815 static void __cpuinit take_over_work(struct ehca_comp_pool *pool, int cpu)
817 struct ehca_cpu_comp_task *cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu);
820 unsigned long flags_cct;
822 spin_lock_irqsave(&cct->task_lock, flags_cct);
824 list_splice_init(&cct->cq_list, &list);
826 while (!list_empty(&list)) {
827 cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
829 list_del(&cq->entry);
830 __queue_comp_task(cq, per_cpu_ptr(pool->cpu_comp_tasks,
831 smp_processor_id()));
834 spin_unlock_irqrestore(&cct->task_lock, flags_cct);
838 static int __cpuinit comp_pool_callback(struct notifier_block *nfb,
839 unsigned long action,
842 unsigned int cpu = (unsigned long)hcpu;
843 struct ehca_cpu_comp_task *cct;
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);
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);
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);
868 case CPU_DOWN_PREPARE:
869 case CPU_DOWN_PREPARE_FROZEN:
870 ehca_gen_dbg("CPU: %x (CPU_DOWN_PREPARE)", cpu);
872 case CPU_DOWN_FAILED:
873 case CPU_DOWN_FAILED_FROZEN:
874 ehca_gen_dbg("CPU: %x (CPU_DOWN_FAILED)", cpu);
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);
887 static struct notifier_block comp_pool_callback_nb __cpuinitdata = {
888 .notifier_call = comp_pool_callback,
892 int ehca_create_comp_pool(void)
895 struct task_struct *task;
897 if (!ehca_scaling_code)
900 pool = kzalloc(sizeof(struct ehca_comp_pool), GFP_KERNEL);
904 spin_lock_init(&pool->last_cpu_lock);
905 pool->last_cpu = cpumask_any(cpu_online_mask);
907 pool->cpu_comp_tasks = alloc_percpu(struct ehca_cpu_comp_task);
908 if (pool->cpu_comp_tasks == NULL) {
913 for_each_online_cpu(cpu) {
914 task = create_comp_task(pool, cpu);
916 kthread_bind(task, cpu);
917 wake_up_process(task);
921 register_hotcpu_notifier(&comp_pool_callback_nb);
923 printk(KERN_INFO "eHCA scaling code enabled\n");
928 void ehca_destroy_comp_pool(void)
932 if (!ehca_scaling_code)
935 unregister_hotcpu_notifier(&comp_pool_callback_nb);
937 for_each_online_cpu(i)
938 destroy_comp_task(pool, i);
940 free_percpu(pool->cpu_comp_tasks);