2 * Copyright(c) 2015, 2016 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/spinlock.h>
49 #include <linux/seqlock.h>
50 #include <linux/netdevice.h>
51 #include <linux/moduleparam.h>
52 #include <linux/bitops.h>
53 #include <linux/timer.h>
54 #include <linux/vmalloc.h>
55 #include <linux/highmem.h>
64 /* must be a power of 2 >= 64 <= 32768 */
65 #define SDMA_DESCQ_CNT 2048
66 #define SDMA_DESC_INTR 64
67 #define INVALID_TAIL 0xffff
69 static uint sdma_descq_cnt = SDMA_DESCQ_CNT;
70 module_param(sdma_descq_cnt, uint, S_IRUGO);
71 MODULE_PARM_DESC(sdma_descq_cnt, "Number of SDMA descq entries");
73 static uint sdma_idle_cnt = 250;
74 module_param(sdma_idle_cnt, uint, S_IRUGO);
75 MODULE_PARM_DESC(sdma_idle_cnt, "sdma interrupt idle delay (ns,default 250)");
78 module_param_named(num_sdma, mod_num_sdma, uint, S_IRUGO);
79 MODULE_PARM_DESC(num_sdma, "Set max number SDMA engines to use");
81 static uint sdma_desct_intr = SDMA_DESC_INTR;
82 module_param_named(desct_intr, sdma_desct_intr, uint, S_IRUGO | S_IWUSR);
83 MODULE_PARM_DESC(desct_intr, "Number of SDMA descriptor before interrupt");
85 #define SDMA_WAIT_BATCH_SIZE 20
86 /* max wait time for a SDMA engine to indicate it has halted */
87 #define SDMA_ERR_HALT_TIMEOUT 10 /* ms */
88 /* all SDMA engine errors that cause a halt */
90 #define SD(name) SEND_DMA_##name
91 #define ALL_SDMA_ENG_HALT_ERRS \
92 (SD(ENG_ERR_STATUS_SDMA_WRONG_DW_ERR_SMASK) \
93 | SD(ENG_ERR_STATUS_SDMA_GEN_MISMATCH_ERR_SMASK) \
94 | SD(ENG_ERR_STATUS_SDMA_TOO_LONG_ERR_SMASK) \
95 | SD(ENG_ERR_STATUS_SDMA_TAIL_OUT_OF_BOUNDS_ERR_SMASK) \
96 | SD(ENG_ERR_STATUS_SDMA_FIRST_DESC_ERR_SMASK) \
97 | SD(ENG_ERR_STATUS_SDMA_MEM_READ_ERR_SMASK) \
98 | SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK) \
99 | SD(ENG_ERR_STATUS_SDMA_LENGTH_MISMATCH_ERR_SMASK) \
100 | SD(ENG_ERR_STATUS_SDMA_PACKET_DESC_OVERFLOW_ERR_SMASK) \
101 | SD(ENG_ERR_STATUS_SDMA_HEADER_SELECT_ERR_SMASK) \
102 | SD(ENG_ERR_STATUS_SDMA_HEADER_ADDRESS_ERR_SMASK) \
103 | SD(ENG_ERR_STATUS_SDMA_HEADER_LENGTH_ERR_SMASK) \
104 | SD(ENG_ERR_STATUS_SDMA_TIMEOUT_ERR_SMASK) \
105 | SD(ENG_ERR_STATUS_SDMA_DESC_TABLE_UNC_ERR_SMASK) \
106 | SD(ENG_ERR_STATUS_SDMA_ASSEMBLY_UNC_ERR_SMASK) \
107 | SD(ENG_ERR_STATUS_SDMA_PACKET_TRACKING_UNC_ERR_SMASK) \
108 | SD(ENG_ERR_STATUS_SDMA_HEADER_STORAGE_UNC_ERR_SMASK) \
109 | SD(ENG_ERR_STATUS_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SMASK))
111 /* sdma_sendctrl operations */
112 #define SDMA_SENDCTRL_OP_ENABLE BIT(0)
113 #define SDMA_SENDCTRL_OP_INTENABLE BIT(1)
114 #define SDMA_SENDCTRL_OP_HALT BIT(2)
115 #define SDMA_SENDCTRL_OP_CLEANUP BIT(3)
117 /* handle long defines */
118 #define SDMA_EGRESS_PACKET_OCCUPANCY_SMASK \
119 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SMASK
120 #define SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT \
121 SEND_EGRESS_SEND_DMA_STATUS_SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT
123 static const char * const sdma_state_names[] = {
124 [sdma_state_s00_hw_down] = "s00_HwDown",
125 [sdma_state_s10_hw_start_up_halt_wait] = "s10_HwStartUpHaltWait",
126 [sdma_state_s15_hw_start_up_clean_wait] = "s15_HwStartUpCleanWait",
127 [sdma_state_s20_idle] = "s20_Idle",
128 [sdma_state_s30_sw_clean_up_wait] = "s30_SwCleanUpWait",
129 [sdma_state_s40_hw_clean_up_wait] = "s40_HwCleanUpWait",
130 [sdma_state_s50_hw_halt_wait] = "s50_HwHaltWait",
131 [sdma_state_s60_idle_halt_wait] = "s60_IdleHaltWait",
132 [sdma_state_s80_hw_freeze] = "s80_HwFreeze",
133 [sdma_state_s82_freeze_sw_clean] = "s82_FreezeSwClean",
134 [sdma_state_s99_running] = "s99_Running",
137 #ifdef CONFIG_SDMA_VERBOSITY
138 static const char * const sdma_event_names[] = {
139 [sdma_event_e00_go_hw_down] = "e00_GoHwDown",
140 [sdma_event_e10_go_hw_start] = "e10_GoHwStart",
141 [sdma_event_e15_hw_halt_done] = "e15_HwHaltDone",
142 [sdma_event_e25_hw_clean_up_done] = "e25_HwCleanUpDone",
143 [sdma_event_e30_go_running] = "e30_GoRunning",
144 [sdma_event_e40_sw_cleaned] = "e40_SwCleaned",
145 [sdma_event_e50_hw_cleaned] = "e50_HwCleaned",
146 [sdma_event_e60_hw_halted] = "e60_HwHalted",
147 [sdma_event_e70_go_idle] = "e70_GoIdle",
148 [sdma_event_e80_hw_freeze] = "e80_HwFreeze",
149 [sdma_event_e81_hw_frozen] = "e81_HwFrozen",
150 [sdma_event_e82_hw_unfreeze] = "e82_HwUnfreeze",
151 [sdma_event_e85_link_down] = "e85_LinkDown",
152 [sdma_event_e90_sw_halted] = "e90_SwHalted",
156 static const struct sdma_set_state_action sdma_action_table[] = {
157 [sdma_state_s00_hw_down] = {
158 .go_s99_running_tofalse = 1,
164 [sdma_state_s10_hw_start_up_halt_wait] = {
170 [sdma_state_s15_hw_start_up_clean_wait] = {
176 [sdma_state_s20_idle] = {
182 [sdma_state_s30_sw_clean_up_wait] = {
188 [sdma_state_s40_hw_clean_up_wait] = {
194 [sdma_state_s50_hw_halt_wait] = {
200 [sdma_state_s60_idle_halt_wait] = {
201 .go_s99_running_tofalse = 1,
207 [sdma_state_s80_hw_freeze] = {
213 [sdma_state_s82_freeze_sw_clean] = {
219 [sdma_state_s99_running] = {
224 .go_s99_running_totrue = 1,
228 #define SDMA_TAIL_UPDATE_THRESH 0x1F
230 /* declare all statics here rather than keep sorting */
231 static void sdma_complete(struct kref *);
232 static void sdma_finalput(struct sdma_state *);
233 static void sdma_get(struct sdma_state *);
234 static void sdma_hw_clean_up_task(unsigned long);
235 static void sdma_put(struct sdma_state *);
236 static void sdma_set_state(struct sdma_engine *, enum sdma_states);
237 static void sdma_start_hw_clean_up(struct sdma_engine *);
238 static void sdma_sw_clean_up_task(unsigned long);
239 static void sdma_sendctrl(struct sdma_engine *, unsigned);
240 static void init_sdma_regs(struct sdma_engine *, u32, uint);
241 static void sdma_process_event(
242 struct sdma_engine *sde,
243 enum sdma_events event);
244 static void __sdma_process_event(
245 struct sdma_engine *sde,
246 enum sdma_events event);
247 static void dump_sdma_state(struct sdma_engine *sde);
248 static void sdma_make_progress(struct sdma_engine *sde, u64 status);
249 static void sdma_desc_avail(struct sdma_engine *sde, uint avail);
250 static void sdma_flush_descq(struct sdma_engine *sde);
253 * sdma_state_name() - return state string from enum
256 static const char *sdma_state_name(enum sdma_states state)
258 return sdma_state_names[state];
261 static void sdma_get(struct sdma_state *ss)
266 static void sdma_complete(struct kref *kref)
268 struct sdma_state *ss =
269 container_of(kref, struct sdma_state, kref);
274 static void sdma_put(struct sdma_state *ss)
276 kref_put(&ss->kref, sdma_complete);
279 static void sdma_finalput(struct sdma_state *ss)
282 wait_for_completion(&ss->comp);
285 static inline void write_sde_csr(
286 struct sdma_engine *sde,
290 write_kctxt_csr(sde->dd, sde->this_idx, offset0, value);
293 static inline u64 read_sde_csr(
294 struct sdma_engine *sde,
297 return read_kctxt_csr(sde->dd, sde->this_idx, offset0);
301 * sdma_wait_for_packet_egress() - wait for the VL FIFO occupancy for
302 * sdma engine 'sde' to drop to 0.
304 static void sdma_wait_for_packet_egress(struct sdma_engine *sde,
307 u64 off = 8 * sde->this_idx;
308 struct hfi1_devdata *dd = sde->dd;
315 reg = read_csr(dd, off + SEND_EGRESS_SEND_DMA_STATUS);
317 reg &= SDMA_EGRESS_PACKET_OCCUPANCY_SMASK;
318 reg >>= SDMA_EGRESS_PACKET_OCCUPANCY_SHIFT;
321 /* counter is reest if accupancy count changes */
325 /* timed out - bounce the link */
326 dd_dev_err(dd, "%s: engine %u timeout waiting for packets to egress, remaining count %u, bouncing link\n",
327 __func__, sde->this_idx, (u32)reg);
328 queue_work(dd->pport->link_wq,
329 &dd->pport->link_bounce_work);
337 * sdma_wait() - wait for packet egress to complete for all SDMA engines,
338 * and pause for credit return.
340 void sdma_wait(struct hfi1_devdata *dd)
344 for (i = 0; i < dd->num_sdma; i++) {
345 struct sdma_engine *sde = &dd->per_sdma[i];
347 sdma_wait_for_packet_egress(sde, 0);
351 static inline void sdma_set_desc_cnt(struct sdma_engine *sde, unsigned cnt)
355 if (!(sde->dd->flags & HFI1_HAS_SDMA_TIMEOUT))
358 reg &= SD(DESC_CNT_CNT_MASK);
359 reg <<= SD(DESC_CNT_CNT_SHIFT);
360 write_sde_csr(sde, SD(DESC_CNT), reg);
363 static inline void complete_tx(struct sdma_engine *sde,
364 struct sdma_txreq *tx,
367 /* protect against complete modifying */
368 struct iowait *wait = tx->wait;
369 callback_t complete = tx->complete;
371 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
372 trace_hfi1_sdma_out_sn(sde, tx->sn);
373 if (WARN_ON_ONCE(sde->head_sn != tx->sn))
374 dd_dev_err(sde->dd, "expected %llu got %llu\n",
375 sde->head_sn, tx->sn);
378 __sdma_txclean(sde->dd, tx);
380 (*complete)(tx, res);
381 if (wait && iowait_sdma_dec(wait))
382 iowait_drain_wakeup(wait);
386 * Complete all the sdma requests with a SDMA_TXREQ_S_ABORTED status
388 * Depending on timing there can be txreqs in two places:
389 * - in the descq ring
390 * - in the flush list
392 * To avoid ordering issues the descq ring needs to be flushed
393 * first followed by the flush list.
395 * This routine is called from two places
396 * - From a work queue item
397 * - Directly from the state machine just before setting the
400 * Must be called with head_lock held
403 static void sdma_flush(struct sdma_engine *sde)
405 struct sdma_txreq *txp, *txp_next;
406 LIST_HEAD(flushlist);
409 /* flush from head to tail */
410 sdma_flush_descq(sde);
411 spin_lock_irqsave(&sde->flushlist_lock, flags);
412 /* copy flush list */
413 list_for_each_entry_safe(txp, txp_next, &sde->flushlist, list) {
414 list_del_init(&txp->list);
415 list_add_tail(&txp->list, &flushlist);
417 spin_unlock_irqrestore(&sde->flushlist_lock, flags);
418 /* flush from flush list */
419 list_for_each_entry_safe(txp, txp_next, &flushlist, list)
420 complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
424 * Fields a work request for flushing the descq ring
427 * If the engine has been brought to running during
428 * the scheduling delay, the flush is ignored, assuming
429 * that the process of bringing the engine to running
430 * would have done this flush prior to going to running.
433 static void sdma_field_flush(struct work_struct *work)
436 struct sdma_engine *sde =
437 container_of(work, struct sdma_engine, flush_worker);
439 write_seqlock_irqsave(&sde->head_lock, flags);
440 if (!__sdma_running(sde))
442 write_sequnlock_irqrestore(&sde->head_lock, flags);
445 static void sdma_err_halt_wait(struct work_struct *work)
447 struct sdma_engine *sde = container_of(work, struct sdma_engine,
450 unsigned long timeout;
452 timeout = jiffies + msecs_to_jiffies(SDMA_ERR_HALT_TIMEOUT);
454 statuscsr = read_sde_csr(sde, SD(STATUS));
455 statuscsr &= SD(STATUS_ENG_HALTED_SMASK);
458 if (time_after(jiffies, timeout)) {
460 "SDMA engine %d - timeout waiting for engine to halt\n",
463 * Continue anyway. This could happen if there was
464 * an uncorrectable error in the wrong spot.
468 usleep_range(80, 120);
471 sdma_process_event(sde, sdma_event_e15_hw_halt_done);
474 static void sdma_err_progress_check_schedule(struct sdma_engine *sde)
476 if (!is_bx(sde->dd) && HFI1_CAP_IS_KSET(SDMA_AHG)) {
478 struct hfi1_devdata *dd = sde->dd;
480 for (index = 0; index < dd->num_sdma; index++) {
481 struct sdma_engine *curr_sdma = &dd->per_sdma[index];
483 if (curr_sdma != sde)
484 curr_sdma->progress_check_head =
485 curr_sdma->descq_head;
488 "SDMA engine %d - check scheduled\n",
490 mod_timer(&sde->err_progress_check_timer, jiffies + 10);
494 static void sdma_err_progress_check(unsigned long data)
497 struct sdma_engine *sde = (struct sdma_engine *)data;
499 dd_dev_err(sde->dd, "SDE progress check event\n");
500 for (index = 0; index < sde->dd->num_sdma; index++) {
501 struct sdma_engine *curr_sde = &sde->dd->per_sdma[index];
504 /* check progress on each engine except the current one */
508 * We must lock interrupts when acquiring sde->lock,
509 * to avoid a deadlock if interrupt triggers and spins on
510 * the same lock on same CPU
512 spin_lock_irqsave(&curr_sde->tail_lock, flags);
513 write_seqlock(&curr_sde->head_lock);
515 /* skip non-running queues */
516 if (curr_sde->state.current_state != sdma_state_s99_running) {
517 write_sequnlock(&curr_sde->head_lock);
518 spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
522 if ((curr_sde->descq_head != curr_sde->descq_tail) &&
523 (curr_sde->descq_head ==
524 curr_sde->progress_check_head))
525 __sdma_process_event(curr_sde,
526 sdma_event_e90_sw_halted);
527 write_sequnlock(&curr_sde->head_lock);
528 spin_unlock_irqrestore(&curr_sde->tail_lock, flags);
530 schedule_work(&sde->err_halt_worker);
533 static void sdma_hw_clean_up_task(unsigned long opaque)
535 struct sdma_engine *sde = (struct sdma_engine *)opaque;
539 #ifdef CONFIG_SDMA_VERBOSITY
540 dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
541 sde->this_idx, slashstrip(__FILE__), __LINE__,
544 statuscsr = read_sde_csr(sde, SD(STATUS));
545 statuscsr &= SD(STATUS_ENG_CLEANED_UP_SMASK);
551 sdma_process_event(sde, sdma_event_e25_hw_clean_up_done);
554 static inline struct sdma_txreq *get_txhead(struct sdma_engine *sde)
556 smp_read_barrier_depends(); /* see sdma_update_tail() */
557 return sde->tx_ring[sde->tx_head & sde->sdma_mask];
561 * flush ring for recovery
563 static void sdma_flush_descq(struct sdma_engine *sde)
567 struct sdma_txreq *txp = get_txhead(sde);
569 /* The reason for some of the complexity of this code is that
570 * not all descriptors have corresponding txps. So, we have to
571 * be able to skip over descs until we wander into the range of
572 * the next txp on the list.
574 head = sde->descq_head & sde->sdma_mask;
575 tail = sde->descq_tail & sde->sdma_mask;
576 while (head != tail) {
577 /* advance head, wrap if needed */
578 head = ++sde->descq_head & sde->sdma_mask;
579 /* if now past this txp's descs, do the callback */
580 if (txp && txp->next_descq_idx == head) {
581 /* remove from list */
582 sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
583 complete_tx(sde, txp, SDMA_TXREQ_S_ABORTED);
584 trace_hfi1_sdma_progress(sde, head, tail, txp);
585 txp = get_txhead(sde);
590 sdma_desc_avail(sde, sdma_descq_freecnt(sde));
593 static void sdma_sw_clean_up_task(unsigned long opaque)
595 struct sdma_engine *sde = (struct sdma_engine *)opaque;
598 spin_lock_irqsave(&sde->tail_lock, flags);
599 write_seqlock(&sde->head_lock);
602 * At this point, the following should always be true:
603 * - We are halted, so no more descriptors are getting retired.
604 * - We are not running, so no one is submitting new work.
605 * - Only we can send the e40_sw_cleaned, so we can't start
606 * running again until we say so. So, the active list and
607 * descq are ours to play with.
611 * In the error clean up sequence, software clean must be called
612 * before the hardware clean so we can use the hardware head in
613 * the progress routine. A hardware clean or SPC unfreeze will
614 * reset the hardware head.
616 * Process all retired requests. The progress routine will use the
617 * latest physical hardware head - we are not running so speed does
620 sdma_make_progress(sde, 0);
625 * Reset our notion of head and tail.
626 * Note that the HW registers have been reset via an earlier
631 sde->desc_avail = sdma_descq_freecnt(sde);
634 __sdma_process_event(sde, sdma_event_e40_sw_cleaned);
636 write_sequnlock(&sde->head_lock);
637 spin_unlock_irqrestore(&sde->tail_lock, flags);
640 static void sdma_sw_tear_down(struct sdma_engine *sde)
642 struct sdma_state *ss = &sde->state;
644 /* Releasing this reference means the state machine has stopped. */
647 /* stop waiting for all unfreeze events to complete */
648 atomic_set(&sde->dd->sdma_unfreeze_count, -1);
649 wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
652 static void sdma_start_hw_clean_up(struct sdma_engine *sde)
654 tasklet_hi_schedule(&sde->sdma_hw_clean_up_task);
657 static void sdma_set_state(struct sdma_engine *sde,
658 enum sdma_states next_state)
660 struct sdma_state *ss = &sde->state;
661 const struct sdma_set_state_action *action = sdma_action_table;
664 trace_hfi1_sdma_state(
666 sdma_state_names[ss->current_state],
667 sdma_state_names[next_state]);
669 /* debugging bookkeeping */
670 ss->previous_state = ss->current_state;
671 ss->previous_op = ss->current_op;
672 ss->current_state = next_state;
674 if (ss->previous_state != sdma_state_s99_running &&
675 next_state == sdma_state_s99_running)
678 if (action[next_state].op_enable)
679 op |= SDMA_SENDCTRL_OP_ENABLE;
681 if (action[next_state].op_intenable)
682 op |= SDMA_SENDCTRL_OP_INTENABLE;
684 if (action[next_state].op_halt)
685 op |= SDMA_SENDCTRL_OP_HALT;
687 if (action[next_state].op_cleanup)
688 op |= SDMA_SENDCTRL_OP_CLEANUP;
690 if (action[next_state].go_s99_running_tofalse)
691 ss->go_s99_running = 0;
693 if (action[next_state].go_s99_running_totrue)
694 ss->go_s99_running = 1;
697 sdma_sendctrl(sde, ss->current_op);
701 * sdma_get_descq_cnt() - called when device probed
703 * Return a validated descq count.
705 * This is currently only used in the verbs initialization to build the tx
708 * This will probably be deleted in favor of a more scalable approach to
712 u16 sdma_get_descq_cnt(void)
714 u16 count = sdma_descq_cnt;
717 return SDMA_DESCQ_CNT;
718 /* count must be a power of 2 greater than 64 and less than
719 * 32768. Otherwise return default.
721 if (!is_power_of_2(count))
722 return SDMA_DESCQ_CNT;
723 if (count < 64 || count > 32768)
724 return SDMA_DESCQ_CNT;
729 * sdma_engine_get_vl() - return vl for a given sdma engine
732 * This function returns the vl mapped to a given engine, or an error if
733 * the mapping can't be found. The mapping fields are protected by RCU.
735 int sdma_engine_get_vl(struct sdma_engine *sde)
737 struct hfi1_devdata *dd = sde->dd;
738 struct sdma_vl_map *m;
741 if (sde->this_idx >= TXE_NUM_SDMA_ENGINES)
745 m = rcu_dereference(dd->sdma_map);
750 vl = m->engine_to_vl[sde->this_idx];
757 * sdma_select_engine_vl() - select sdma engine
759 * @selector: a spreading factor
763 * This function returns an engine based on the selector and a vl. The
764 * mapping fields are protected by RCU.
766 struct sdma_engine *sdma_select_engine_vl(
767 struct hfi1_devdata *dd,
771 struct sdma_vl_map *m;
772 struct sdma_map_elem *e;
773 struct sdma_engine *rval;
775 /* NOTE This should only happen if SC->VL changed after the initial
776 * checks on the QP/AH
777 * Default will return engine 0 below
785 m = rcu_dereference(dd->sdma_map);
788 return &dd->per_sdma[0];
790 e = m->map[vl & m->mask];
791 rval = e->sde[selector & e->mask];
795 rval = !rval ? &dd->per_sdma[0] : rval;
796 trace_hfi1_sdma_engine_select(dd, selector, vl, rval->this_idx);
801 * sdma_select_engine_sc() - select sdma engine
803 * @selector: a spreading factor
807 * This function returns an engine based on the selector and an sc.
809 struct sdma_engine *sdma_select_engine_sc(
810 struct hfi1_devdata *dd,
814 u8 vl = sc_to_vlt(dd, sc5);
816 return sdma_select_engine_vl(dd, selector, vl);
819 struct sdma_rht_map_elem {
822 struct sdma_engine *sde[0];
825 struct sdma_rht_node {
826 unsigned long cpu_id;
827 struct sdma_rht_map_elem *map[HFI1_MAX_VLS_SUPPORTED];
828 struct rhash_head node;
831 #define NR_CPUS_HINT 192
833 static const struct rhashtable_params sdma_rht_params = {
834 .nelem_hint = NR_CPUS_HINT,
835 .head_offset = offsetof(struct sdma_rht_node, node),
836 .key_offset = offsetof(struct sdma_rht_node, cpu_id),
837 .key_len = FIELD_SIZEOF(struct sdma_rht_node, cpu_id),
840 .automatic_shrinking = true,
844 * sdma_select_user_engine() - select sdma engine based on user setup
846 * @selector: a spreading factor
849 * This function returns an sdma engine for a user sdma request.
850 * User defined sdma engine affinity setting is honored when applicable,
851 * otherwise system default sdma engine mapping is used. To ensure correct
852 * ordering, the mapping from <selector, vl> to sde must remain unchanged.
854 struct sdma_engine *sdma_select_user_engine(struct hfi1_devdata *dd,
857 struct sdma_rht_node *rht_node;
858 struct sdma_engine *sde = NULL;
859 const struct cpumask *current_mask = ¤t->cpus_allowed;
860 unsigned long cpu_id;
863 * To ensure that always the same sdma engine(s) will be
864 * selected make sure the process is pinned to this CPU only.
866 if (cpumask_weight(current_mask) != 1)
869 cpu_id = smp_processor_id();
871 rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu_id,
874 if (rht_node && rht_node->map[vl]) {
875 struct sdma_rht_map_elem *map = rht_node->map[vl];
877 sde = map->sde[selector & map->mask];
885 return sdma_select_engine_vl(dd, selector, vl);
888 static void sdma_populate_sde_map(struct sdma_rht_map_elem *map)
892 for (i = 0; i < roundup_pow_of_two(map->ctr ? : 1) - map->ctr; i++)
893 map->sde[map->ctr + i] = map->sde[i];
896 static void sdma_cleanup_sde_map(struct sdma_rht_map_elem *map,
897 struct sdma_engine *sde)
901 /* only need to check the first ctr entries for a match */
902 for (i = 0; i < map->ctr; i++) {
903 if (map->sde[i] == sde) {
904 memmove(&map->sde[i], &map->sde[i + 1],
905 (map->ctr - i - 1) * sizeof(map->sde[0]));
907 pow = roundup_pow_of_two(map->ctr ? : 1);
909 sdma_populate_sde_map(map);
916 * Prevents concurrent reads and writes of the sdma engine cpu_mask
918 static DEFINE_MUTEX(process_to_sde_mutex);
920 ssize_t sdma_set_cpu_to_sde_map(struct sdma_engine *sde, const char *buf,
923 struct hfi1_devdata *dd = sde->dd;
924 cpumask_var_t mask, new_mask;
928 vl = sdma_engine_get_vl(sde);
929 if (unlikely(vl < 0))
932 ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
936 ret = zalloc_cpumask_var(&new_mask, GFP_KERNEL);
938 free_cpumask_var(mask);
941 ret = cpulist_parse(buf, mask);
945 if (!cpumask_subset(mask, cpu_online_mask)) {
946 dd_dev_warn(sde->dd, "Invalid CPU mask\n");
951 sz = sizeof(struct sdma_rht_map_elem) +
952 (TXE_NUM_SDMA_ENGINES * sizeof(struct sdma_engine *));
954 mutex_lock(&process_to_sde_mutex);
956 for_each_cpu(cpu, mask) {
957 struct sdma_rht_node *rht_node;
959 /* Check if we have this already mapped */
960 if (cpumask_test_cpu(cpu, &sde->cpu_mask)) {
961 cpumask_set_cpu(cpu, new_mask);
965 if (vl >= ARRAY_SIZE(rht_node->map)) {
970 rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
973 rht_node = kzalloc(sizeof(*rht_node), GFP_KERNEL);
979 rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
980 if (!rht_node->map[vl]) {
985 rht_node->cpu_id = cpu;
986 rht_node->map[vl]->mask = 0;
987 rht_node->map[vl]->ctr = 1;
988 rht_node->map[vl]->sde[0] = sde;
990 ret = rhashtable_insert_fast(dd->sdma_rht,
994 kfree(rht_node->map[vl]);
996 dd_dev_err(sde->dd, "Failed to set process to sde affinity for cpu %lu\n",
1004 /* Add new user mappings */
1005 if (!rht_node->map[vl])
1006 rht_node->map[vl] = kzalloc(sz, GFP_KERNEL);
1008 if (!rht_node->map[vl]) {
1013 rht_node->map[vl]->ctr++;
1014 ctr = rht_node->map[vl]->ctr;
1015 rht_node->map[vl]->sde[ctr - 1] = sde;
1016 pow = roundup_pow_of_two(ctr);
1017 rht_node->map[vl]->mask = pow - 1;
1019 /* Populate the sde map table */
1020 sdma_populate_sde_map(rht_node->map[vl]);
1022 cpumask_set_cpu(cpu, new_mask);
1025 /* Clean up old mappings */
1026 for_each_cpu(cpu, cpu_online_mask) {
1027 struct sdma_rht_node *rht_node;
1029 /* Don't cleanup sdes that are set in the new mask */
1030 if (cpumask_test_cpu(cpu, mask))
1033 rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpu,
1039 /* Remove mappings for old sde */
1040 for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1041 if (rht_node->map[i])
1042 sdma_cleanup_sde_map(rht_node->map[i],
1045 /* Free empty hash table entries */
1046 for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1047 if (!rht_node->map[i])
1050 if (rht_node->map[i]->ctr) {
1057 ret = rhashtable_remove_fast(dd->sdma_rht,
1062 for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1063 kfree(rht_node->map[i]);
1070 cpumask_copy(&sde->cpu_mask, new_mask);
1072 mutex_unlock(&process_to_sde_mutex);
1074 free_cpumask_var(mask);
1075 free_cpumask_var(new_mask);
1076 return ret ? : strnlen(buf, PAGE_SIZE);
1079 ssize_t sdma_get_cpu_to_sde_map(struct sdma_engine *sde, char *buf)
1081 mutex_lock(&process_to_sde_mutex);
1082 if (cpumask_empty(&sde->cpu_mask))
1083 snprintf(buf, PAGE_SIZE, "%s\n", "empty");
1085 cpumap_print_to_pagebuf(true, buf, &sde->cpu_mask);
1086 mutex_unlock(&process_to_sde_mutex);
1087 return strnlen(buf, PAGE_SIZE);
1090 static void sdma_rht_free(void *ptr, void *arg)
1092 struct sdma_rht_node *rht_node = ptr;
1095 for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++)
1096 kfree(rht_node->map[i]);
1102 * sdma_seqfile_dump_cpu_list() - debugfs dump the cpu to sdma mappings
1107 * This routine dumps the process to sde mappings per cpu
1109 void sdma_seqfile_dump_cpu_list(struct seq_file *s,
1110 struct hfi1_devdata *dd,
1111 unsigned long cpuid)
1113 struct sdma_rht_node *rht_node;
1116 rht_node = rhashtable_lookup_fast(dd->sdma_rht, &cpuid,
1121 seq_printf(s, "cpu%3lu: ", cpuid);
1122 for (i = 0; i < HFI1_MAX_VLS_SUPPORTED; i++) {
1123 if (!rht_node->map[i] || !rht_node->map[i]->ctr)
1126 seq_printf(s, " vl%d: [", i);
1128 for (j = 0; j < rht_node->map[i]->ctr; j++) {
1129 if (!rht_node->map[i]->sde[j])
1135 seq_printf(s, " sdma%2d",
1136 rht_node->map[i]->sde[j]->this_idx);
1145 * Free the indicated map struct
1147 static void sdma_map_free(struct sdma_vl_map *m)
1151 for (i = 0; m && i < m->actual_vls; i++)
1157 * Handle RCU callback
1159 static void sdma_map_rcu_callback(struct rcu_head *list)
1161 struct sdma_vl_map *m = container_of(list, struct sdma_vl_map, list);
1167 * sdma_map_init - called when # vls change
1169 * @port: port number
1170 * @num_vls: number of vls
1171 * @vl_engines: per vl engine mapping (optional)
1173 * This routine changes the mapping based on the number of vls.
1175 * vl_engines is used to specify a non-uniform vl/engine loading. NULL
1176 * implies auto computing the loading and giving each VLs a uniform
1177 * distribution of engines per VL.
1179 * The auto algorithm computes the sde_per_vl and the number of extra
1180 * engines. Any extra engines are added from the last VL on down.
1182 * rcu locking is used here to control access to the mapping fields.
1184 * If either the num_vls or num_sdma are non-power of 2, the array sizes
1185 * in the struct sdma_vl_map and the struct sdma_map_elem are rounded
1186 * up to the next highest power of 2 and the first entry is reused
1187 * in a round robin fashion.
1189 * If an error occurs the map change is not done and the mapping is
1193 int sdma_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_engines)
1196 int extra, sde_per_vl;
1198 u8 lvl_engines[OPA_MAX_VLS];
1199 struct sdma_vl_map *oldmap, *newmap;
1201 if (!(dd->flags & HFI1_HAS_SEND_DMA))
1205 /* truncate divide */
1206 sde_per_vl = dd->num_sdma / num_vls;
1208 extra = dd->num_sdma % num_vls;
1209 vl_engines = lvl_engines;
1210 /* add extras from last vl down */
1211 for (i = num_vls - 1; i >= 0; i--, extra--)
1212 vl_engines[i] = sde_per_vl + (extra > 0 ? 1 : 0);
1216 sizeof(struct sdma_vl_map) +
1217 roundup_pow_of_two(num_vls) *
1218 sizeof(struct sdma_map_elem *),
1222 newmap->actual_vls = num_vls;
1223 newmap->vls = roundup_pow_of_two(num_vls);
1224 newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1225 /* initialize back-map */
1226 for (i = 0; i < TXE_NUM_SDMA_ENGINES; i++)
1227 newmap->engine_to_vl[i] = -1;
1228 for (i = 0; i < newmap->vls; i++) {
1229 /* save for wrap around */
1230 int first_engine = engine;
1232 if (i < newmap->actual_vls) {
1233 int sz = roundup_pow_of_two(vl_engines[i]);
1235 /* only allocate once */
1236 newmap->map[i] = kzalloc(
1237 sizeof(struct sdma_map_elem) +
1238 sz * sizeof(struct sdma_engine *),
1240 if (!newmap->map[i])
1242 newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1243 /* assign engines */
1244 for (j = 0; j < sz; j++) {
1245 newmap->map[i]->sde[j] =
1246 &dd->per_sdma[engine];
1247 if (++engine >= first_engine + vl_engines[i])
1248 /* wrap back to first engine */
1249 engine = first_engine;
1251 /* assign back-map */
1252 for (j = 0; j < vl_engines[i]; j++)
1253 newmap->engine_to_vl[first_engine + j] = i;
1255 /* just re-use entry without allocating */
1256 newmap->map[i] = newmap->map[i % num_vls];
1258 engine = first_engine + vl_engines[i];
1260 /* newmap in hand, save old map */
1261 spin_lock_irq(&dd->sde_map_lock);
1262 oldmap = rcu_dereference_protected(dd->sdma_map,
1263 lockdep_is_held(&dd->sde_map_lock));
1265 /* publish newmap */
1266 rcu_assign_pointer(dd->sdma_map, newmap);
1268 spin_unlock_irq(&dd->sde_map_lock);
1269 /* success, free any old map after grace period */
1271 call_rcu(&oldmap->list, sdma_map_rcu_callback);
1274 /* free any partial allocation */
1275 sdma_map_free(newmap);
1280 * Clean up allocated memory.
1282 * This routine is can be called regardless of the success of sdma_init()
1285 static void sdma_clean(struct hfi1_devdata *dd, size_t num_engines)
1288 struct sdma_engine *sde;
1290 if (dd->sdma_pad_dma) {
1291 dma_free_coherent(&dd->pcidev->dev, 4,
1292 (void *)dd->sdma_pad_dma,
1294 dd->sdma_pad_dma = NULL;
1295 dd->sdma_pad_phys = 0;
1297 if (dd->sdma_heads_dma) {
1298 dma_free_coherent(&dd->pcidev->dev, dd->sdma_heads_size,
1299 (void *)dd->sdma_heads_dma,
1300 dd->sdma_heads_phys);
1301 dd->sdma_heads_dma = NULL;
1302 dd->sdma_heads_phys = 0;
1304 for (i = 0; dd->per_sdma && i < num_engines; ++i) {
1305 sde = &dd->per_sdma[i];
1307 sde->head_dma = NULL;
1313 sde->descq_cnt * sizeof(u64[2]),
1318 sde->descq_phys = 0;
1320 kvfree(sde->tx_ring);
1321 sde->tx_ring = NULL;
1323 spin_lock_irq(&dd->sde_map_lock);
1324 sdma_map_free(rcu_access_pointer(dd->sdma_map));
1325 RCU_INIT_POINTER(dd->sdma_map, NULL);
1326 spin_unlock_irq(&dd->sde_map_lock);
1328 kfree(dd->per_sdma);
1329 dd->per_sdma = NULL;
1332 rhashtable_free_and_destroy(dd->sdma_rht, sdma_rht_free, NULL);
1333 kfree(dd->sdma_rht);
1334 dd->sdma_rht = NULL;
1339 * sdma_init() - called when device probed
1341 * @port: port number (currently only zero)
1343 * Initializes each sde and its csrs.
1344 * Interrupts are not required to be enabled.
1347 * 0 - success, -errno on failure
1349 int sdma_init(struct hfi1_devdata *dd, u8 port)
1352 struct sdma_engine *sde;
1353 struct rhashtable *tmp_sdma_rht;
1356 struct hfi1_pportdata *ppd = dd->pport + port;
1357 u32 per_sdma_credits;
1358 uint idle_cnt = sdma_idle_cnt;
1359 size_t num_engines = dd->chip_sdma_engines;
1362 if (!HFI1_CAP_IS_KSET(SDMA)) {
1363 HFI1_CAP_CLEAR(SDMA_AHG);
1367 /* can't exceed chip support */
1368 mod_num_sdma <= dd->chip_sdma_engines &&
1369 /* count must be >= vls */
1370 mod_num_sdma >= num_vls)
1371 num_engines = mod_num_sdma;
1373 dd_dev_info(dd, "SDMA mod_num_sdma: %u\n", mod_num_sdma);
1374 dd_dev_info(dd, "SDMA chip_sdma_engines: %u\n", dd->chip_sdma_engines);
1375 dd_dev_info(dd, "SDMA chip_sdma_mem_size: %u\n",
1376 dd->chip_sdma_mem_size);
1379 dd->chip_sdma_mem_size / (num_engines * SDMA_BLOCK_SIZE);
1381 /* set up freeze waitqueue */
1382 init_waitqueue_head(&dd->sdma_unfreeze_wq);
1383 atomic_set(&dd->sdma_unfreeze_count, 0);
1385 descq_cnt = sdma_get_descq_cnt();
1386 dd_dev_info(dd, "SDMA engines %zu descq_cnt %u\n",
1387 num_engines, descq_cnt);
1389 /* alloc memory for array of send engines */
1390 dd->per_sdma = kcalloc(num_engines, sizeof(*dd->per_sdma), GFP_KERNEL);
1394 idle_cnt = ns_to_cclock(dd, idle_cnt);
1395 if (!sdma_desct_intr)
1396 sdma_desct_intr = SDMA_DESC_INTR;
1398 /* Allocate memory for SendDMA descriptor FIFOs */
1399 for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1400 sde = &dd->per_sdma[this_idx];
1403 sde->this_idx = this_idx;
1404 sde->descq_cnt = descq_cnt;
1405 sde->desc_avail = sdma_descq_freecnt(sde);
1406 sde->sdma_shift = ilog2(descq_cnt);
1407 sde->sdma_mask = (1 << sde->sdma_shift) - 1;
1409 /* Create a mask specifically for each interrupt source */
1410 sde->int_mask = (u64)1 << (0 * TXE_NUM_SDMA_ENGINES +
1412 sde->progress_mask = (u64)1 << (1 * TXE_NUM_SDMA_ENGINES +
1414 sde->idle_mask = (u64)1 << (2 * TXE_NUM_SDMA_ENGINES +
1416 /* Create a combined mask to cover all 3 interrupt sources */
1417 sde->imask = sde->int_mask | sde->progress_mask |
1420 spin_lock_init(&sde->tail_lock);
1421 seqlock_init(&sde->head_lock);
1422 spin_lock_init(&sde->senddmactrl_lock);
1423 spin_lock_init(&sde->flushlist_lock);
1424 /* insure there is always a zero bit */
1425 sde->ahg_bits = 0xfffffffe00000000ULL;
1427 sdma_set_state(sde, sdma_state_s00_hw_down);
1429 /* set up reference counting */
1430 kref_init(&sde->state.kref);
1431 init_completion(&sde->state.comp);
1433 INIT_LIST_HEAD(&sde->flushlist);
1434 INIT_LIST_HEAD(&sde->dmawait);
1437 get_kctxt_csr_addr(dd, this_idx, SD(TAIL));
1441 SDMA_DESC1_HEAD_TO_HOST_FLAG;
1444 SDMA_DESC1_INT_REQ_FLAG;
1446 tasklet_init(&sde->sdma_hw_clean_up_task, sdma_hw_clean_up_task,
1447 (unsigned long)sde);
1449 tasklet_init(&sde->sdma_sw_clean_up_task, sdma_sw_clean_up_task,
1450 (unsigned long)sde);
1451 INIT_WORK(&sde->err_halt_worker, sdma_err_halt_wait);
1452 INIT_WORK(&sde->flush_worker, sdma_field_flush);
1454 sde->progress_check_head = 0;
1456 setup_timer(&sde->err_progress_check_timer,
1457 sdma_err_progress_check, (unsigned long)sde);
1459 sde->descq = dma_zalloc_coherent(
1461 descq_cnt * sizeof(u64[2]),
1468 kcalloc(descq_cnt, sizeof(struct sdma_txreq *),
1473 sizeof(struct sdma_txreq *) *
1479 dd->sdma_heads_size = L1_CACHE_BYTES * num_engines;
1480 /* Allocate memory for DMA of head registers to memory */
1481 dd->sdma_heads_dma = dma_zalloc_coherent(
1483 dd->sdma_heads_size,
1484 &dd->sdma_heads_phys,
1487 if (!dd->sdma_heads_dma) {
1488 dd_dev_err(dd, "failed to allocate SendDMA head memory\n");
1492 /* Allocate memory for pad */
1493 dd->sdma_pad_dma = dma_zalloc_coherent(
1499 if (!dd->sdma_pad_dma) {
1500 dd_dev_err(dd, "failed to allocate SendDMA pad memory\n");
1504 /* assign each engine to different cacheline and init registers */
1505 curr_head = (void *)dd->sdma_heads_dma;
1506 for (this_idx = 0; this_idx < num_engines; ++this_idx) {
1507 unsigned long phys_offset;
1509 sde = &dd->per_sdma[this_idx];
1511 sde->head_dma = curr_head;
1512 curr_head += L1_CACHE_BYTES;
1513 phys_offset = (unsigned long)sde->head_dma -
1514 (unsigned long)dd->sdma_heads_dma;
1515 sde->head_phys = dd->sdma_heads_phys + phys_offset;
1516 init_sdma_regs(sde, per_sdma_credits, idle_cnt);
1518 dd->flags |= HFI1_HAS_SEND_DMA;
1519 dd->flags |= idle_cnt ? HFI1_HAS_SDMA_TIMEOUT : 0;
1520 dd->num_sdma = num_engines;
1521 ret = sdma_map_init(dd, port, ppd->vls_operational, NULL);
1525 tmp_sdma_rht = kzalloc(sizeof(*tmp_sdma_rht), GFP_KERNEL);
1526 if (!tmp_sdma_rht) {
1531 ret = rhashtable_init(tmp_sdma_rht, &sdma_rht_params);
1533 kfree(tmp_sdma_rht);
1537 dd->sdma_rht = tmp_sdma_rht;
1539 dd_dev_info(dd, "SDMA num_sdma: %u\n", dd->num_sdma);
1543 sdma_clean(dd, num_engines);
1548 * sdma_all_running() - called when the link goes up
1551 * This routine moves all engines to the running state.
1553 void sdma_all_running(struct hfi1_devdata *dd)
1555 struct sdma_engine *sde;
1558 /* move all engines to running */
1559 for (i = 0; i < dd->num_sdma; ++i) {
1560 sde = &dd->per_sdma[i];
1561 sdma_process_event(sde, sdma_event_e30_go_running);
1566 * sdma_all_idle() - called when the link goes down
1569 * This routine moves all engines to the idle state.
1571 void sdma_all_idle(struct hfi1_devdata *dd)
1573 struct sdma_engine *sde;
1576 /* idle all engines */
1577 for (i = 0; i < dd->num_sdma; ++i) {
1578 sde = &dd->per_sdma[i];
1579 sdma_process_event(sde, sdma_event_e70_go_idle);
1584 * sdma_start() - called to kick off state processing for all engines
1587 * This routine is for kicking off the state processing for all required
1588 * sdma engines. Interrupts need to be working at this point.
1591 void sdma_start(struct hfi1_devdata *dd)
1594 struct sdma_engine *sde;
1596 /* kick off the engines state processing */
1597 for (i = 0; i < dd->num_sdma; ++i) {
1598 sde = &dd->per_sdma[i];
1599 sdma_process_event(sde, sdma_event_e10_go_hw_start);
1604 * sdma_exit() - used when module is removed
1607 void sdma_exit(struct hfi1_devdata *dd)
1610 struct sdma_engine *sde;
1612 for (this_idx = 0; dd->per_sdma && this_idx < dd->num_sdma;
1614 sde = &dd->per_sdma[this_idx];
1615 if (!list_empty(&sde->dmawait))
1616 dd_dev_err(dd, "sde %u: dmawait list not empty!\n",
1618 sdma_process_event(sde, sdma_event_e00_go_hw_down);
1620 del_timer_sync(&sde->err_progress_check_timer);
1623 * This waits for the state machine to exit so it is not
1624 * necessary to kill the sdma_sw_clean_up_task to make sure
1625 * it is not running.
1627 sdma_finalput(&sde->state);
1629 sdma_clean(dd, dd->num_sdma);
1633 * unmap the indicated descriptor
1635 static inline void sdma_unmap_desc(
1636 struct hfi1_devdata *dd,
1637 struct sdma_desc *descp)
1639 switch (sdma_mapping_type(descp)) {
1640 case SDMA_MAP_SINGLE:
1643 sdma_mapping_addr(descp),
1644 sdma_mapping_len(descp),
1650 sdma_mapping_addr(descp),
1651 sdma_mapping_len(descp),
1658 * return the mode as indicated by the first
1659 * descriptor in the tx.
1661 static inline u8 ahg_mode(struct sdma_txreq *tx)
1663 return (tx->descp[0].qw[1] & SDMA_DESC1_HEADER_MODE_SMASK)
1664 >> SDMA_DESC1_HEADER_MODE_SHIFT;
1668 * __sdma_txclean() - clean tx of mappings, descp *kmalloc's
1669 * @dd: hfi1_devdata for unmapping
1670 * @tx: tx request to clean
1672 * This is used in the progress routine to clean the tx or
1673 * by the ULP to toss an in-process tx build.
1675 * The code can be called multiple times without issue.
1678 void __sdma_txclean(
1679 struct hfi1_devdata *dd,
1680 struct sdma_txreq *tx)
1685 u8 skip = 0, mode = ahg_mode(tx);
1688 sdma_unmap_desc(dd, &tx->descp[0]);
1689 /* determine number of AHG descriptors to skip */
1690 if (mode > SDMA_AHG_APPLY_UPDATE1)
1692 for (i = 1 + skip; i < tx->num_desc; i++)
1693 sdma_unmap_desc(dd, &tx->descp[i]);
1696 kfree(tx->coalesce_buf);
1697 tx->coalesce_buf = NULL;
1698 /* kmalloc'ed descp */
1699 if (unlikely(tx->desc_limit > ARRAY_SIZE(tx->descs))) {
1700 tx->desc_limit = ARRAY_SIZE(tx->descs);
1705 static inline u16 sdma_gethead(struct sdma_engine *sde)
1707 struct hfi1_devdata *dd = sde->dd;
1711 #ifdef CONFIG_SDMA_VERBOSITY
1712 dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1713 sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
1717 use_dmahead = HFI1_CAP_IS_KSET(USE_SDMA_HEAD) && __sdma_running(sde) &&
1718 (dd->flags & HFI1_HAS_SDMA_TIMEOUT);
1719 hwhead = use_dmahead ?
1720 (u16)le64_to_cpu(*sde->head_dma) :
1721 (u16)read_sde_csr(sde, SD(HEAD));
1723 if (unlikely(HFI1_CAP_IS_KSET(SDMA_HEAD_CHECK))) {
1729 swhead = sde->descq_head & sde->sdma_mask;
1730 /* this code is really bad for cache line trading */
1731 swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
1732 cnt = sde->descq_cnt;
1734 if (swhead < swtail)
1736 sane = (hwhead >= swhead) & (hwhead <= swtail);
1737 else if (swhead > swtail)
1738 /* wrapped around */
1739 sane = ((hwhead >= swhead) && (hwhead < cnt)) ||
1743 sane = (hwhead == swhead);
1745 if (unlikely(!sane)) {
1746 dd_dev_err(dd, "SDMA(%u) bad head (%s) hwhd=%hu swhd=%hu swtl=%hu cnt=%hu\n",
1748 use_dmahead ? "dma" : "kreg",
1749 hwhead, swhead, swtail, cnt);
1751 /* try one more time, using csr */
1755 /* proceed as if no progress */
1763 * This is called when there are send DMA descriptors that might be
1766 * This is called with head_lock held.
1768 static void sdma_desc_avail(struct sdma_engine *sde, uint avail)
1770 struct iowait *wait, *nw;
1771 struct iowait *waits[SDMA_WAIT_BATCH_SIZE];
1772 uint i, n = 0, seq, max_idx = 0;
1773 struct sdma_txreq *stx;
1774 struct hfi1_ibdev *dev = &sde->dd->verbs_dev;
1775 u8 max_starved_cnt = 0;
1777 #ifdef CONFIG_SDMA_VERBOSITY
1778 dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n", sde->this_idx,
1779 slashstrip(__FILE__), __LINE__, __func__);
1780 dd_dev_err(sde->dd, "avail: %u\n", avail);
1784 seq = read_seqbegin(&dev->iowait_lock);
1785 if (!list_empty(&sde->dmawait)) {
1786 /* at least one item */
1787 write_seqlock(&dev->iowait_lock);
1788 /* Harvest waiters wanting DMA descriptors */
1789 list_for_each_entry_safe(
1798 if (n == ARRAY_SIZE(waits))
1800 if (!list_empty(&wait->tx_head)) {
1801 stx = list_first_entry(
1805 num_desc = stx->num_desc;
1807 if (num_desc > avail)
1810 /* Find the most starved wait memeber */
1811 iowait_starve_find_max(wait, &max_starved_cnt,
1813 list_del_init(&wait->list);
1816 write_sequnlock(&dev->iowait_lock);
1819 } while (read_seqretry(&dev->iowait_lock, seq));
1821 /* Schedule the most starved one first */
1823 waits[max_idx]->wakeup(waits[max_idx], SDMA_AVAIL_REASON);
1825 for (i = 0; i < n; i++)
1827 waits[i]->wakeup(waits[i], SDMA_AVAIL_REASON);
1830 /* head_lock must be held */
1831 static void sdma_make_progress(struct sdma_engine *sde, u64 status)
1833 struct sdma_txreq *txp = NULL;
1836 int idle_check_done = 0;
1838 hwhead = sdma_gethead(sde);
1840 /* The reason for some of the complexity of this code is that
1841 * not all descriptors have corresponding txps. So, we have to
1842 * be able to skip over descs until we wander into the range of
1843 * the next txp on the list.
1847 txp = get_txhead(sde);
1848 swhead = sde->descq_head & sde->sdma_mask;
1849 trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1850 while (swhead != hwhead) {
1851 /* advance head, wrap if needed */
1852 swhead = ++sde->descq_head & sde->sdma_mask;
1854 /* if now past this txp's descs, do the callback */
1855 if (txp && txp->next_descq_idx == swhead) {
1856 /* remove from list */
1857 sde->tx_ring[sde->tx_head++ & sde->sdma_mask] = NULL;
1858 complete_tx(sde, txp, SDMA_TXREQ_S_OK);
1859 /* see if there is another txp */
1860 txp = get_txhead(sde);
1862 trace_hfi1_sdma_progress(sde, hwhead, swhead, txp);
1867 * The SDMA idle interrupt is not guaranteed to be ordered with respect
1868 * to updates to the the dma_head location in host memory. The head
1869 * value read might not be fully up to date. If there are pending
1870 * descriptors and the SDMA idle interrupt fired then read from the
1871 * CSR SDMA head instead to get the latest value from the hardware.
1872 * The hardware SDMA head should be read at most once in this invocation
1873 * of sdma_make_progress(..) which is ensured by idle_check_done flag
1875 if ((status & sde->idle_mask) && !idle_check_done) {
1878 swtail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
1879 if (swtail != hwhead) {
1880 hwhead = (u16)read_sde_csr(sde, SD(HEAD));
1881 idle_check_done = 1;
1886 sde->last_status = status;
1888 sdma_desc_avail(sde, sdma_descq_freecnt(sde));
1892 * sdma_engine_interrupt() - interrupt handler for engine
1894 * @status: sdma interrupt reason
1896 * Status is a mask of the 3 possible interrupts for this engine. It will
1897 * contain bits _only_ for this SDMA engine. It will contain at least one
1898 * bit, it may contain more.
1900 void sdma_engine_interrupt(struct sdma_engine *sde, u64 status)
1902 trace_hfi1_sdma_engine_interrupt(sde, status);
1903 write_seqlock(&sde->head_lock);
1904 sdma_set_desc_cnt(sde, sdma_desct_intr);
1905 if (status & sde->idle_mask)
1906 sde->idle_int_cnt++;
1907 else if (status & sde->progress_mask)
1908 sde->progress_int_cnt++;
1909 else if (status & sde->int_mask)
1910 sde->sdma_int_cnt++;
1911 sdma_make_progress(sde, status);
1912 write_sequnlock(&sde->head_lock);
1916 * sdma_engine_error() - error handler for engine
1918 * @status: sdma interrupt reason
1920 void sdma_engine_error(struct sdma_engine *sde, u64 status)
1922 unsigned long flags;
1924 #ifdef CONFIG_SDMA_VERBOSITY
1925 dd_dev_err(sde->dd, "CONFIG SDMA(%u) error status 0x%llx state %s\n",
1927 (unsigned long long)status,
1928 sdma_state_names[sde->state.current_state]);
1930 spin_lock_irqsave(&sde->tail_lock, flags);
1931 write_seqlock(&sde->head_lock);
1932 if (status & ALL_SDMA_ENG_HALT_ERRS)
1933 __sdma_process_event(sde, sdma_event_e60_hw_halted);
1934 if (status & ~SD(ENG_ERR_STATUS_SDMA_HALT_ERR_SMASK)) {
1936 "SDMA (%u) engine error: 0x%llx state %s\n",
1938 (unsigned long long)status,
1939 sdma_state_names[sde->state.current_state]);
1940 dump_sdma_state(sde);
1942 write_sequnlock(&sde->head_lock);
1943 spin_unlock_irqrestore(&sde->tail_lock, flags);
1946 static void sdma_sendctrl(struct sdma_engine *sde, unsigned op)
1948 u64 set_senddmactrl = 0;
1949 u64 clr_senddmactrl = 0;
1950 unsigned long flags;
1952 #ifdef CONFIG_SDMA_VERBOSITY
1953 dd_dev_err(sde->dd, "CONFIG SDMA(%u) senddmactrl E=%d I=%d H=%d C=%d\n",
1955 (op & SDMA_SENDCTRL_OP_ENABLE) ? 1 : 0,
1956 (op & SDMA_SENDCTRL_OP_INTENABLE) ? 1 : 0,
1957 (op & SDMA_SENDCTRL_OP_HALT) ? 1 : 0,
1958 (op & SDMA_SENDCTRL_OP_CLEANUP) ? 1 : 0);
1961 if (op & SDMA_SENDCTRL_OP_ENABLE)
1962 set_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1964 clr_senddmactrl |= SD(CTRL_SDMA_ENABLE_SMASK);
1966 if (op & SDMA_SENDCTRL_OP_INTENABLE)
1967 set_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1969 clr_senddmactrl |= SD(CTRL_SDMA_INT_ENABLE_SMASK);
1971 if (op & SDMA_SENDCTRL_OP_HALT)
1972 set_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1974 clr_senddmactrl |= SD(CTRL_SDMA_HALT_SMASK);
1976 spin_lock_irqsave(&sde->senddmactrl_lock, flags);
1978 sde->p_senddmactrl |= set_senddmactrl;
1979 sde->p_senddmactrl &= ~clr_senddmactrl;
1981 if (op & SDMA_SENDCTRL_OP_CLEANUP)
1982 write_sde_csr(sde, SD(CTRL),
1983 sde->p_senddmactrl |
1984 SD(CTRL_SDMA_CLEANUP_SMASK));
1986 write_sde_csr(sde, SD(CTRL), sde->p_senddmactrl);
1988 spin_unlock_irqrestore(&sde->senddmactrl_lock, flags);
1990 #ifdef CONFIG_SDMA_VERBOSITY
1991 sdma_dumpstate(sde);
1995 static void sdma_setlengen(struct sdma_engine *sde)
1997 #ifdef CONFIG_SDMA_VERBOSITY
1998 dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
1999 sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2003 * Set SendDmaLenGen and clear-then-set the MSB of the generation
2004 * count to enable generation checking and load the internal
2005 * generation counter.
2007 write_sde_csr(sde, SD(LEN_GEN),
2008 (sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT));
2009 write_sde_csr(sde, SD(LEN_GEN),
2010 ((sde->descq_cnt / 64) << SD(LEN_GEN_LENGTH_SHIFT)) |
2011 (4ULL << SD(LEN_GEN_GENERATION_SHIFT)));
2014 static inline void sdma_update_tail(struct sdma_engine *sde, u16 tail)
2016 /* Commit writes to memory and advance the tail on the chip */
2017 smp_wmb(); /* see get_txhead() */
2018 writeq(tail, sde->tail_csr);
2022 * This is called when changing to state s10_hw_start_up_halt_wait as
2023 * a result of send buffer errors or send DMA descriptor errors.
2025 static void sdma_hw_start_up(struct sdma_engine *sde)
2029 #ifdef CONFIG_SDMA_VERBOSITY
2030 dd_dev_err(sde->dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2031 sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2034 sdma_setlengen(sde);
2035 sdma_update_tail(sde, 0); /* Set SendDmaTail */
2038 reg = SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_MASK) <<
2039 SD(ENG_ERR_CLEAR_SDMA_HEADER_REQUEST_FIFO_UNC_ERR_SHIFT);
2040 write_sde_csr(sde, SD(ENG_ERR_CLEAR), reg);
2044 * set_sdma_integrity
2046 * Set the SEND_DMA_CHECK_ENABLE register for send DMA engine 'sde'.
2048 static void set_sdma_integrity(struct sdma_engine *sde)
2050 struct hfi1_devdata *dd = sde->dd;
2052 write_sde_csr(sde, SD(CHECK_ENABLE),
2053 hfi1_pkt_base_sdma_integrity(dd));
2056 static void init_sdma_regs(
2057 struct sdma_engine *sde,
2062 #ifdef CONFIG_SDMA_VERBOSITY
2063 struct hfi1_devdata *dd = sde->dd;
2065 dd_dev_err(dd, "CONFIG SDMA(%u) %s:%d %s()\n",
2066 sde->this_idx, slashstrip(__FILE__), __LINE__, __func__);
2069 write_sde_csr(sde, SD(BASE_ADDR), sde->descq_phys);
2070 sdma_setlengen(sde);
2071 sdma_update_tail(sde, 0); /* Set SendDmaTail */
2072 write_sde_csr(sde, SD(RELOAD_CNT), idle_cnt);
2073 write_sde_csr(sde, SD(DESC_CNT), 0);
2074 write_sde_csr(sde, SD(HEAD_ADDR), sde->head_phys);
2075 write_sde_csr(sde, SD(MEMORY),
2076 ((u64)credits << SD(MEMORY_SDMA_MEMORY_CNT_SHIFT)) |
2077 ((u64)(credits * sde->this_idx) <<
2078 SD(MEMORY_SDMA_MEMORY_INDEX_SHIFT)));
2079 write_sde_csr(sde, SD(ENG_ERR_MASK), ~0ull);
2080 set_sdma_integrity(sde);
2081 opmask = OPCODE_CHECK_MASK_DISABLED;
2082 opval = OPCODE_CHECK_VAL_DISABLED;
2083 write_sde_csr(sde, SD(CHECK_OPCODE),
2084 (opmask << SEND_CTXT_CHECK_OPCODE_MASK_SHIFT) |
2085 (opval << SEND_CTXT_CHECK_OPCODE_VALUE_SHIFT));
2088 #ifdef CONFIG_SDMA_VERBOSITY
2090 #define sdma_dumpstate_helper0(reg) do { \
2091 csr = read_csr(sde->dd, reg); \
2092 dd_dev_err(sde->dd, "%36s 0x%016llx\n", #reg, csr); \
2095 #define sdma_dumpstate_helper(reg) do { \
2096 csr = read_sde_csr(sde, reg); \
2097 dd_dev_err(sde->dd, "%36s[%02u] 0x%016llx\n", \
2098 #reg, sde->this_idx, csr); \
2101 #define sdma_dumpstate_helper2(reg) do { \
2102 csr = read_csr(sde->dd, reg + (8 * i)); \
2103 dd_dev_err(sde->dd, "%33s_%02u 0x%016llx\n", \
2107 void sdma_dumpstate(struct sdma_engine *sde)
2112 sdma_dumpstate_helper(SD(CTRL));
2113 sdma_dumpstate_helper(SD(STATUS));
2114 sdma_dumpstate_helper0(SD(ERR_STATUS));
2115 sdma_dumpstate_helper0(SD(ERR_MASK));
2116 sdma_dumpstate_helper(SD(ENG_ERR_STATUS));
2117 sdma_dumpstate_helper(SD(ENG_ERR_MASK));
2119 for (i = 0; i < CCE_NUM_INT_CSRS; ++i) {
2120 sdma_dumpstate_helper2(CCE_INT_STATUS);
2121 sdma_dumpstate_helper2(CCE_INT_MASK);
2122 sdma_dumpstate_helper2(CCE_INT_BLOCKED);
2125 sdma_dumpstate_helper(SD(TAIL));
2126 sdma_dumpstate_helper(SD(HEAD));
2127 sdma_dumpstate_helper(SD(PRIORITY_THLD));
2128 sdma_dumpstate_helper(SD(IDLE_CNT));
2129 sdma_dumpstate_helper(SD(RELOAD_CNT));
2130 sdma_dumpstate_helper(SD(DESC_CNT));
2131 sdma_dumpstate_helper(SD(DESC_FETCHED_CNT));
2132 sdma_dumpstate_helper(SD(MEMORY));
2133 sdma_dumpstate_helper0(SD(ENGINES));
2134 sdma_dumpstate_helper0(SD(MEM_SIZE));
2135 /* sdma_dumpstate_helper(SEND_EGRESS_SEND_DMA_STATUS); */
2136 sdma_dumpstate_helper(SD(BASE_ADDR));
2137 sdma_dumpstate_helper(SD(LEN_GEN));
2138 sdma_dumpstate_helper(SD(HEAD_ADDR));
2139 sdma_dumpstate_helper(SD(CHECK_ENABLE));
2140 sdma_dumpstate_helper(SD(CHECK_VL));
2141 sdma_dumpstate_helper(SD(CHECK_JOB_KEY));
2142 sdma_dumpstate_helper(SD(CHECK_PARTITION_KEY));
2143 sdma_dumpstate_helper(SD(CHECK_SLID));
2144 sdma_dumpstate_helper(SD(CHECK_OPCODE));
2148 static void dump_sdma_state(struct sdma_engine *sde)
2150 struct hw_sdma_desc *descq;
2151 struct hw_sdma_desc *descqp;
2156 u16 head, tail, cnt;
2158 head = sde->descq_head & sde->sdma_mask;
2159 tail = sde->descq_tail & sde->sdma_mask;
2160 cnt = sdma_descq_freecnt(sde);
2164 "SDMA (%u) descq_head: %u descq_tail: %u freecnt: %u FLE %d\n",
2165 sde->this_idx, head, tail, cnt,
2166 !list_empty(&sde->flushlist));
2168 /* print info for each entry in the descriptor queue */
2169 while (head != tail) {
2170 char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2172 descqp = &sde->descq[head];
2173 desc[0] = le64_to_cpu(descqp->qw[0]);
2174 desc[1] = le64_to_cpu(descqp->qw[1]);
2175 flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2176 flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2178 flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2179 flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2180 addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2181 & SDMA_DESC0_PHY_ADDR_MASK;
2182 gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2183 & SDMA_DESC1_GENERATION_MASK;
2184 len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2185 & SDMA_DESC0_BYTE_COUNT_MASK;
2187 "SDMA sdmadesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2188 head, flags, addr, gen, len);
2190 "\tdesc0:0x%016llx desc1 0x%016llx\n",
2192 if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2194 "\taidx: %u amode: %u alen: %u\n",
2196 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2197 SDMA_DESC1_HEADER_INDEX_SHIFT),
2199 SDMA_DESC1_HEADER_MODE_SMASK) >>
2200 SDMA_DESC1_HEADER_MODE_SHIFT),
2202 SDMA_DESC1_HEADER_DWS_SMASK) >>
2203 SDMA_DESC1_HEADER_DWS_SHIFT));
2205 head &= sde->sdma_mask;
2210 "SDE %u CPU %d STE %s C 0x%llx S 0x%016llx E 0x%llx T(HW) 0x%llx T(SW) 0x%x H(HW) 0x%llx H(SW) 0x%x H(D) 0x%llx DM 0x%llx GL 0x%llx R 0x%llx LIS 0x%llx AHGI 0x%llx TXT %u TXH %u DT %u DH %u FLNE %d DQF %u SLC 0x%llx\n"
2212 * sdma_seqfile_dump_sde() - debugfs dump of sde
2214 * @sde: send dma engine to dump
2216 * This routine dumps the sde to the indicated seq file.
2218 void sdma_seqfile_dump_sde(struct seq_file *s, struct sdma_engine *sde)
2221 struct hw_sdma_desc *descqp;
2227 head = sde->descq_head & sde->sdma_mask;
2228 tail = ACCESS_ONCE(sde->descq_tail) & sde->sdma_mask;
2229 seq_printf(s, SDE_FMT, sde->this_idx,
2231 sdma_state_name(sde->state.current_state),
2232 (unsigned long long)read_sde_csr(sde, SD(CTRL)),
2233 (unsigned long long)read_sde_csr(sde, SD(STATUS)),
2234 (unsigned long long)read_sde_csr(sde, SD(ENG_ERR_STATUS)),
2235 (unsigned long long)read_sde_csr(sde, SD(TAIL)), tail,
2236 (unsigned long long)read_sde_csr(sde, SD(HEAD)), head,
2237 (unsigned long long)le64_to_cpu(*sde->head_dma),
2238 (unsigned long long)read_sde_csr(sde, SD(MEMORY)),
2239 (unsigned long long)read_sde_csr(sde, SD(LEN_GEN)),
2240 (unsigned long long)read_sde_csr(sde, SD(RELOAD_CNT)),
2241 (unsigned long long)sde->last_status,
2242 (unsigned long long)sde->ahg_bits,
2247 !list_empty(&sde->flushlist),
2248 sde->descq_full_count,
2249 (unsigned long long)read_sde_csr(sde, SEND_DMA_CHECK_SLID));
2251 /* print info for each entry in the descriptor queue */
2252 while (head != tail) {
2253 char flags[6] = { 'x', 'x', 'x', 'x', 0 };
2255 descqp = &sde->descq[head];
2256 desc[0] = le64_to_cpu(descqp->qw[0]);
2257 desc[1] = le64_to_cpu(descqp->qw[1]);
2258 flags[0] = (desc[1] & SDMA_DESC1_INT_REQ_FLAG) ? 'I' : '-';
2259 flags[1] = (desc[1] & SDMA_DESC1_HEAD_TO_HOST_FLAG) ?
2261 flags[2] = (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG) ? 'F' : '-';
2262 flags[3] = (desc[0] & SDMA_DESC0_LAST_DESC_FLAG) ? 'L' : '-';
2263 addr = (desc[0] >> SDMA_DESC0_PHY_ADDR_SHIFT)
2264 & SDMA_DESC0_PHY_ADDR_MASK;
2265 gen = (desc[1] >> SDMA_DESC1_GENERATION_SHIFT)
2266 & SDMA_DESC1_GENERATION_MASK;
2267 len = (desc[0] >> SDMA_DESC0_BYTE_COUNT_SHIFT)
2268 & SDMA_DESC0_BYTE_COUNT_MASK;
2270 "\tdesc[%u]: flags:%s addr:0x%016llx gen:%u len:%u bytes\n",
2271 head, flags, addr, gen, len);
2272 if (desc[0] & SDMA_DESC0_FIRST_DESC_FLAG)
2273 seq_printf(s, "\t\tahgidx: %u ahgmode: %u\n",
2275 SDMA_DESC1_HEADER_INDEX_SMASK) >>
2276 SDMA_DESC1_HEADER_INDEX_SHIFT),
2278 SDMA_DESC1_HEADER_MODE_SMASK) >>
2279 SDMA_DESC1_HEADER_MODE_SHIFT));
2280 head = (head + 1) & sde->sdma_mask;
2285 * add the generation number into
2286 * the qw1 and return
2288 static inline u64 add_gen(struct sdma_engine *sde, u64 qw1)
2290 u8 generation = (sde->descq_tail >> sde->sdma_shift) & 3;
2292 qw1 &= ~SDMA_DESC1_GENERATION_SMASK;
2293 qw1 |= ((u64)generation & SDMA_DESC1_GENERATION_MASK)
2294 << SDMA_DESC1_GENERATION_SHIFT;
2299 * This routine submits the indicated tx
2301 * Space has already been guaranteed and
2302 * tail side of ring is locked.
2304 * The hardware tail update is done
2305 * in the caller and that is facilitated
2306 * by returning the new tail.
2308 * There is special case logic for ahg
2309 * to not add the generation number for
2310 * up to 2 descriptors that follow the
2314 static inline u16 submit_tx(struct sdma_engine *sde, struct sdma_txreq *tx)
2318 struct sdma_desc *descp = tx->descp;
2319 u8 skip = 0, mode = ahg_mode(tx);
2321 tail = sde->descq_tail & sde->sdma_mask;
2322 sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2323 sde->descq[tail].qw[1] = cpu_to_le64(add_gen(sde, descp->qw[1]));
2324 trace_hfi1_sdma_descriptor(sde, descp->qw[0], descp->qw[1],
2325 tail, &sde->descq[tail]);
2326 tail = ++sde->descq_tail & sde->sdma_mask;
2328 if (mode > SDMA_AHG_APPLY_UPDATE1)
2330 for (i = 1; i < tx->num_desc; i++, descp++) {
2333 sde->descq[tail].qw[0] = cpu_to_le64(descp->qw[0]);
2335 /* edits don't have generation */
2339 /* replace generation with real one for non-edits */
2340 qw1 = add_gen(sde, descp->qw[1]);
2342 sde->descq[tail].qw[1] = cpu_to_le64(qw1);
2343 trace_hfi1_sdma_descriptor(sde, descp->qw[0], qw1,
2344 tail, &sde->descq[tail]);
2345 tail = ++sde->descq_tail & sde->sdma_mask;
2347 tx->next_descq_idx = tail;
2348 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2349 tx->sn = sde->tail_sn++;
2350 trace_hfi1_sdma_in_sn(sde, tx->sn);
2351 WARN_ON_ONCE(sde->tx_ring[sde->tx_tail & sde->sdma_mask]);
2353 sde->tx_ring[sde->tx_tail++ & sde->sdma_mask] = tx;
2354 sde->desc_avail -= tx->num_desc;
2359 * Check for progress
2361 static int sdma_check_progress(
2362 struct sdma_engine *sde,
2363 struct iowait *wait,
2364 struct sdma_txreq *tx,
2369 sde->desc_avail = sdma_descq_freecnt(sde);
2370 if (tx->num_desc <= sde->desc_avail)
2372 /* pulse the head_lock */
2373 if (wait && wait->sleep) {
2376 seq = raw_seqcount_begin(
2377 (const seqcount_t *)&sde->head_lock.seqcount);
2378 ret = wait->sleep(sde, wait, tx, seq, pkts_sent);
2380 sde->desc_avail = sdma_descq_freecnt(sde);
2388 * sdma_send_txreq() - submit a tx req to ring
2389 * @sde: sdma engine to use
2390 * @wait: wait structure to use when full (may be NULL)
2391 * @tx: sdma_txreq to submit
2392 * @pkts_sent: has any packet been sent yet?
2394 * The call submits the tx into the ring. If a iowait structure is non-NULL
2395 * the packet will be queued to the list in wait.
2398 * 0 - Success, -EINVAL - sdma_txreq incomplete, -EBUSY - no space in
2399 * ring (wait == NULL)
2400 * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2402 int sdma_send_txreq(struct sdma_engine *sde,
2403 struct iowait *wait,
2404 struct sdma_txreq *tx,
2409 unsigned long flags;
2411 /* user should have supplied entire packet */
2412 if (unlikely(tx->tlen))
2415 spin_lock_irqsave(&sde->tail_lock, flags);
2417 if (unlikely(!__sdma_running(sde)))
2419 if (unlikely(tx->num_desc > sde->desc_avail))
2421 tail = submit_tx(sde, tx);
2423 iowait_sdma_inc(wait);
2424 sdma_update_tail(sde, tail);
2426 spin_unlock_irqrestore(&sde->tail_lock, flags);
2430 iowait_sdma_inc(wait);
2431 tx->next_descq_idx = 0;
2432 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2433 tx->sn = sde->tail_sn++;
2434 trace_hfi1_sdma_in_sn(sde, tx->sn);
2436 spin_lock(&sde->flushlist_lock);
2437 list_add_tail(&tx->list, &sde->flushlist);
2438 spin_unlock(&sde->flushlist_lock);
2441 wait->count += tx->num_desc;
2443 schedule_work(&sde->flush_worker);
2447 ret = sdma_check_progress(sde, wait, tx, pkts_sent);
2448 if (ret == -EAGAIN) {
2452 sde->descq_full_count++;
2457 * sdma_send_txlist() - submit a list of tx req to ring
2458 * @sde: sdma engine to use
2459 * @wait: wait structure to use when full (may be NULL)
2460 * @tx_list: list of sdma_txreqs to submit
2461 * @count: pointer to a u32 which, after return will contain the total number of
2462 * sdma_txreqs removed from the tx_list. This will include sdma_txreqs
2463 * whose SDMA descriptors are submitted to the ring and the sdma_txreqs
2464 * which are added to SDMA engine flush list if the SDMA engine state is
2467 * The call submits the list into the ring.
2469 * If the iowait structure is non-NULL and not equal to the iowait list
2470 * the unprocessed part of the list will be appended to the list in wait.
2472 * In all cases, the tx_list will be updated so the head of the tx_list is
2473 * the list of descriptors that have yet to be transmitted.
2475 * The intent of this call is to provide a more efficient
2476 * way of submitting multiple packets to SDMA while holding the tail
2481 * -EINVAL - sdma_txreq incomplete, -EBUSY - no space in ring (wait == NULL)
2482 * -EIOCBQUEUED - tx queued to iowait, -ECOMM bad sdma state
2484 int sdma_send_txlist(struct sdma_engine *sde, struct iowait *wait,
2485 struct list_head *tx_list, u32 *count_out)
2487 struct sdma_txreq *tx, *tx_next;
2489 unsigned long flags;
2490 u16 tail = INVALID_TAIL;
2491 u32 submit_count = 0, flush_count = 0, total_count;
2493 spin_lock_irqsave(&sde->tail_lock, flags);
2495 list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2497 if (unlikely(!__sdma_running(sde)))
2499 if (unlikely(tx->num_desc > sde->desc_avail))
2501 if (unlikely(tx->tlen)) {
2505 list_del_init(&tx->list);
2506 tail = submit_tx(sde, tx);
2508 if (tail != INVALID_TAIL &&
2509 (submit_count & SDMA_TAIL_UPDATE_THRESH) == 0) {
2510 sdma_update_tail(sde, tail);
2511 tail = INVALID_TAIL;
2515 total_count = submit_count + flush_count;
2517 iowait_sdma_add(wait, total_count);
2518 iowait_starve_clear(submit_count > 0, wait);
2520 if (tail != INVALID_TAIL)
2521 sdma_update_tail(sde, tail);
2522 spin_unlock_irqrestore(&sde->tail_lock, flags);
2523 *count_out = total_count;
2526 spin_lock(&sde->flushlist_lock);
2527 list_for_each_entry_safe(tx, tx_next, tx_list, list) {
2529 list_del_init(&tx->list);
2530 tx->next_descq_idx = 0;
2531 #ifdef CONFIG_HFI1_DEBUG_SDMA_ORDER
2532 tx->sn = sde->tail_sn++;
2533 trace_hfi1_sdma_in_sn(sde, tx->sn);
2535 list_add_tail(&tx->list, &sde->flushlist);
2539 wait->count += tx->num_desc;
2542 spin_unlock(&sde->flushlist_lock);
2543 schedule_work(&sde->flush_worker);
2547 ret = sdma_check_progress(sde, wait, tx, submit_count > 0);
2548 if (ret == -EAGAIN) {
2552 sde->descq_full_count++;
2556 static void sdma_process_event(struct sdma_engine *sde, enum sdma_events event)
2558 unsigned long flags;
2560 spin_lock_irqsave(&sde->tail_lock, flags);
2561 write_seqlock(&sde->head_lock);
2563 __sdma_process_event(sde, event);
2565 if (sde->state.current_state == sdma_state_s99_running)
2566 sdma_desc_avail(sde, sdma_descq_freecnt(sde));
2568 write_sequnlock(&sde->head_lock);
2569 spin_unlock_irqrestore(&sde->tail_lock, flags);
2572 static void __sdma_process_event(struct sdma_engine *sde,
2573 enum sdma_events event)
2575 struct sdma_state *ss = &sde->state;
2576 int need_progress = 0;
2578 /* CONFIG SDMA temporary */
2579 #ifdef CONFIG_SDMA_VERBOSITY
2580 dd_dev_err(sde->dd, "CONFIG SDMA(%u) [%s] %s\n", sde->this_idx,
2581 sdma_state_names[ss->current_state],
2582 sdma_event_names[event]);
2585 switch (ss->current_state) {
2586 case sdma_state_s00_hw_down:
2588 case sdma_event_e00_go_hw_down:
2590 case sdma_event_e30_go_running:
2592 * If down, but running requested (usually result
2593 * of link up, then we need to start up.
2594 * This can happen when hw down is requested while
2595 * bringing the link up with traffic active on
2598 ss->go_s99_running = 1;
2599 /* fall through and start dma engine */
2600 case sdma_event_e10_go_hw_start:
2601 /* This reference means the state machine is started */
2602 sdma_get(&sde->state);
2604 sdma_state_s10_hw_start_up_halt_wait);
2606 case sdma_event_e15_hw_halt_done:
2608 case sdma_event_e25_hw_clean_up_done:
2610 case sdma_event_e40_sw_cleaned:
2611 sdma_sw_tear_down(sde);
2613 case sdma_event_e50_hw_cleaned:
2615 case sdma_event_e60_hw_halted:
2617 case sdma_event_e70_go_idle:
2619 case sdma_event_e80_hw_freeze:
2621 case sdma_event_e81_hw_frozen:
2623 case sdma_event_e82_hw_unfreeze:
2625 case sdma_event_e85_link_down:
2627 case sdma_event_e90_sw_halted:
2632 case sdma_state_s10_hw_start_up_halt_wait:
2634 case sdma_event_e00_go_hw_down:
2635 sdma_set_state(sde, sdma_state_s00_hw_down);
2636 sdma_sw_tear_down(sde);
2638 case sdma_event_e10_go_hw_start:
2640 case sdma_event_e15_hw_halt_done:
2642 sdma_state_s15_hw_start_up_clean_wait);
2643 sdma_start_hw_clean_up(sde);
2645 case sdma_event_e25_hw_clean_up_done:
2647 case sdma_event_e30_go_running:
2648 ss->go_s99_running = 1;
2650 case sdma_event_e40_sw_cleaned:
2652 case sdma_event_e50_hw_cleaned:
2654 case sdma_event_e60_hw_halted:
2655 schedule_work(&sde->err_halt_worker);
2657 case sdma_event_e70_go_idle:
2658 ss->go_s99_running = 0;
2660 case sdma_event_e80_hw_freeze:
2662 case sdma_event_e81_hw_frozen:
2664 case sdma_event_e82_hw_unfreeze:
2666 case sdma_event_e85_link_down:
2668 case sdma_event_e90_sw_halted:
2673 case sdma_state_s15_hw_start_up_clean_wait:
2675 case sdma_event_e00_go_hw_down:
2676 sdma_set_state(sde, sdma_state_s00_hw_down);
2677 sdma_sw_tear_down(sde);
2679 case sdma_event_e10_go_hw_start:
2681 case sdma_event_e15_hw_halt_done:
2683 case sdma_event_e25_hw_clean_up_done:
2684 sdma_hw_start_up(sde);
2685 sdma_set_state(sde, ss->go_s99_running ?
2686 sdma_state_s99_running :
2687 sdma_state_s20_idle);
2689 case sdma_event_e30_go_running:
2690 ss->go_s99_running = 1;
2692 case sdma_event_e40_sw_cleaned:
2694 case sdma_event_e50_hw_cleaned:
2696 case sdma_event_e60_hw_halted:
2698 case sdma_event_e70_go_idle:
2699 ss->go_s99_running = 0;
2701 case sdma_event_e80_hw_freeze:
2703 case sdma_event_e81_hw_frozen:
2705 case sdma_event_e82_hw_unfreeze:
2707 case sdma_event_e85_link_down:
2709 case sdma_event_e90_sw_halted:
2714 case sdma_state_s20_idle:
2716 case sdma_event_e00_go_hw_down:
2717 sdma_set_state(sde, sdma_state_s00_hw_down);
2718 sdma_sw_tear_down(sde);
2720 case sdma_event_e10_go_hw_start:
2722 case sdma_event_e15_hw_halt_done:
2724 case sdma_event_e25_hw_clean_up_done:
2726 case sdma_event_e30_go_running:
2727 sdma_set_state(sde, sdma_state_s99_running);
2728 ss->go_s99_running = 1;
2730 case sdma_event_e40_sw_cleaned:
2732 case sdma_event_e50_hw_cleaned:
2734 case sdma_event_e60_hw_halted:
2735 sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
2736 schedule_work(&sde->err_halt_worker);
2738 case sdma_event_e70_go_idle:
2740 case sdma_event_e85_link_down:
2742 case sdma_event_e80_hw_freeze:
2743 sdma_set_state(sde, sdma_state_s80_hw_freeze);
2744 atomic_dec(&sde->dd->sdma_unfreeze_count);
2745 wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2747 case sdma_event_e81_hw_frozen:
2749 case sdma_event_e82_hw_unfreeze:
2751 case sdma_event_e90_sw_halted:
2756 case sdma_state_s30_sw_clean_up_wait:
2758 case sdma_event_e00_go_hw_down:
2759 sdma_set_state(sde, sdma_state_s00_hw_down);
2761 case sdma_event_e10_go_hw_start:
2763 case sdma_event_e15_hw_halt_done:
2765 case sdma_event_e25_hw_clean_up_done:
2767 case sdma_event_e30_go_running:
2768 ss->go_s99_running = 1;
2770 case sdma_event_e40_sw_cleaned:
2771 sdma_set_state(sde, sdma_state_s40_hw_clean_up_wait);
2772 sdma_start_hw_clean_up(sde);
2774 case sdma_event_e50_hw_cleaned:
2776 case sdma_event_e60_hw_halted:
2778 case sdma_event_e70_go_idle:
2779 ss->go_s99_running = 0;
2781 case sdma_event_e80_hw_freeze:
2783 case sdma_event_e81_hw_frozen:
2785 case sdma_event_e82_hw_unfreeze:
2787 case sdma_event_e85_link_down:
2788 ss->go_s99_running = 0;
2790 case sdma_event_e90_sw_halted:
2795 case sdma_state_s40_hw_clean_up_wait:
2797 case sdma_event_e00_go_hw_down:
2798 sdma_set_state(sde, sdma_state_s00_hw_down);
2799 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2801 case sdma_event_e10_go_hw_start:
2803 case sdma_event_e15_hw_halt_done:
2805 case sdma_event_e25_hw_clean_up_done:
2806 sdma_hw_start_up(sde);
2807 sdma_set_state(sde, ss->go_s99_running ?
2808 sdma_state_s99_running :
2809 sdma_state_s20_idle);
2811 case sdma_event_e30_go_running:
2812 ss->go_s99_running = 1;
2814 case sdma_event_e40_sw_cleaned:
2816 case sdma_event_e50_hw_cleaned:
2818 case sdma_event_e60_hw_halted:
2820 case sdma_event_e70_go_idle:
2821 ss->go_s99_running = 0;
2823 case sdma_event_e80_hw_freeze:
2825 case sdma_event_e81_hw_frozen:
2827 case sdma_event_e82_hw_unfreeze:
2829 case sdma_event_e85_link_down:
2830 ss->go_s99_running = 0;
2832 case sdma_event_e90_sw_halted:
2837 case sdma_state_s50_hw_halt_wait:
2839 case sdma_event_e00_go_hw_down:
2840 sdma_set_state(sde, sdma_state_s00_hw_down);
2841 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2843 case sdma_event_e10_go_hw_start:
2845 case sdma_event_e15_hw_halt_done:
2846 sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2847 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2849 case sdma_event_e25_hw_clean_up_done:
2851 case sdma_event_e30_go_running:
2852 ss->go_s99_running = 1;
2854 case sdma_event_e40_sw_cleaned:
2856 case sdma_event_e50_hw_cleaned:
2858 case sdma_event_e60_hw_halted:
2859 schedule_work(&sde->err_halt_worker);
2861 case sdma_event_e70_go_idle:
2862 ss->go_s99_running = 0;
2864 case sdma_event_e80_hw_freeze:
2866 case sdma_event_e81_hw_frozen:
2868 case sdma_event_e82_hw_unfreeze:
2870 case sdma_event_e85_link_down:
2871 ss->go_s99_running = 0;
2873 case sdma_event_e90_sw_halted:
2878 case sdma_state_s60_idle_halt_wait:
2880 case sdma_event_e00_go_hw_down:
2881 sdma_set_state(sde, sdma_state_s00_hw_down);
2882 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2884 case sdma_event_e10_go_hw_start:
2886 case sdma_event_e15_hw_halt_done:
2887 sdma_set_state(sde, sdma_state_s30_sw_clean_up_wait);
2888 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2890 case sdma_event_e25_hw_clean_up_done:
2892 case sdma_event_e30_go_running:
2893 ss->go_s99_running = 1;
2895 case sdma_event_e40_sw_cleaned:
2897 case sdma_event_e50_hw_cleaned:
2899 case sdma_event_e60_hw_halted:
2900 schedule_work(&sde->err_halt_worker);
2902 case sdma_event_e70_go_idle:
2903 ss->go_s99_running = 0;
2905 case sdma_event_e80_hw_freeze:
2907 case sdma_event_e81_hw_frozen:
2909 case sdma_event_e82_hw_unfreeze:
2911 case sdma_event_e85_link_down:
2913 case sdma_event_e90_sw_halted:
2918 case sdma_state_s80_hw_freeze:
2920 case sdma_event_e00_go_hw_down:
2921 sdma_set_state(sde, sdma_state_s00_hw_down);
2922 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2924 case sdma_event_e10_go_hw_start:
2926 case sdma_event_e15_hw_halt_done:
2928 case sdma_event_e25_hw_clean_up_done:
2930 case sdma_event_e30_go_running:
2931 ss->go_s99_running = 1;
2933 case sdma_event_e40_sw_cleaned:
2935 case sdma_event_e50_hw_cleaned:
2937 case sdma_event_e60_hw_halted:
2939 case sdma_event_e70_go_idle:
2940 ss->go_s99_running = 0;
2942 case sdma_event_e80_hw_freeze:
2944 case sdma_event_e81_hw_frozen:
2945 sdma_set_state(sde, sdma_state_s82_freeze_sw_clean);
2946 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2948 case sdma_event_e82_hw_unfreeze:
2950 case sdma_event_e85_link_down:
2952 case sdma_event_e90_sw_halted:
2957 case sdma_state_s82_freeze_sw_clean:
2959 case sdma_event_e00_go_hw_down:
2960 sdma_set_state(sde, sdma_state_s00_hw_down);
2961 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
2963 case sdma_event_e10_go_hw_start:
2965 case sdma_event_e15_hw_halt_done:
2967 case sdma_event_e25_hw_clean_up_done:
2969 case sdma_event_e30_go_running:
2970 ss->go_s99_running = 1;
2972 case sdma_event_e40_sw_cleaned:
2973 /* notify caller this engine is done cleaning */
2974 atomic_dec(&sde->dd->sdma_unfreeze_count);
2975 wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
2977 case sdma_event_e50_hw_cleaned:
2979 case sdma_event_e60_hw_halted:
2981 case sdma_event_e70_go_idle:
2982 ss->go_s99_running = 0;
2984 case sdma_event_e80_hw_freeze:
2986 case sdma_event_e81_hw_frozen:
2988 case sdma_event_e82_hw_unfreeze:
2989 sdma_hw_start_up(sde);
2990 sdma_set_state(sde, ss->go_s99_running ?
2991 sdma_state_s99_running :
2992 sdma_state_s20_idle);
2994 case sdma_event_e85_link_down:
2996 case sdma_event_e90_sw_halted:
3001 case sdma_state_s99_running:
3003 case sdma_event_e00_go_hw_down:
3004 sdma_set_state(sde, sdma_state_s00_hw_down);
3005 tasklet_hi_schedule(&sde->sdma_sw_clean_up_task);
3007 case sdma_event_e10_go_hw_start:
3009 case sdma_event_e15_hw_halt_done:
3011 case sdma_event_e25_hw_clean_up_done:
3013 case sdma_event_e30_go_running:
3015 case sdma_event_e40_sw_cleaned:
3017 case sdma_event_e50_hw_cleaned:
3019 case sdma_event_e60_hw_halted:
3021 sdma_err_progress_check_schedule(sde);
3022 case sdma_event_e90_sw_halted:
3024 * SW initiated halt does not perform engines
3027 sdma_set_state(sde, sdma_state_s50_hw_halt_wait);
3028 schedule_work(&sde->err_halt_worker);
3030 case sdma_event_e70_go_idle:
3031 sdma_set_state(sde, sdma_state_s60_idle_halt_wait);
3033 case sdma_event_e85_link_down:
3034 ss->go_s99_running = 0;
3036 case sdma_event_e80_hw_freeze:
3037 sdma_set_state(sde, sdma_state_s80_hw_freeze);
3038 atomic_dec(&sde->dd->sdma_unfreeze_count);
3039 wake_up_interruptible(&sde->dd->sdma_unfreeze_wq);
3041 case sdma_event_e81_hw_frozen:
3043 case sdma_event_e82_hw_unfreeze:
3049 ss->last_event = event;
3051 sdma_make_progress(sde, 0);
3055 * _extend_sdma_tx_descs() - helper to extend txreq
3057 * This is called once the initial nominal allocation
3058 * of descriptors in the sdma_txreq is exhausted.
3060 * The code will bump the allocation up to the max
3061 * of MAX_DESC (64) descriptors. There doesn't seem
3062 * much point in an interim step. The last descriptor
3063 * is reserved for coalesce buffer in order to support
3064 * cases where input packet has >MAX_DESC iovecs.
3067 static int _extend_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3070 struct sdma_desc *descp;
3072 /* Handle last descriptor */
3073 if (unlikely((tx->num_desc == (MAX_DESC - 1)))) {
3074 /* if tlen is 0, it is for padding, release last descriptor */
3076 tx->desc_limit = MAX_DESC;
3077 } else if (!tx->coalesce_buf) {
3078 /* allocate coalesce buffer with space for padding */
3079 tx->coalesce_buf = kmalloc(tx->tlen + sizeof(u32),
3081 if (!tx->coalesce_buf)
3083 tx->coalesce_idx = 0;
3088 if (unlikely(tx->num_desc == MAX_DESC))
3091 descp = kmalloc_array(MAX_DESC, sizeof(struct sdma_desc), GFP_ATOMIC);
3096 /* reserve last descriptor for coalescing */
3097 tx->desc_limit = MAX_DESC - 1;
3098 /* copy ones already built */
3099 for (i = 0; i < tx->num_desc; i++)
3100 tx->descp[i] = tx->descs[i];
3103 __sdma_txclean(dd, tx);
3108 * ext_coal_sdma_tx_descs() - extend or coalesce sdma tx descriptors
3110 * This is called once the initial nominal allocation of descriptors
3111 * in the sdma_txreq is exhausted.
3113 * This function calls _extend_sdma_tx_descs to extend or allocate
3114 * coalesce buffer. If there is a allocated coalesce buffer, it will
3115 * copy the input packet data into the coalesce buffer. It also adds
3116 * coalesce buffer descriptor once when whole packet is received.
3120 * 0 - coalescing, don't populate descriptor
3121 * 1 - continue with populating descriptor
3123 int ext_coal_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx,
3124 int type, void *kvaddr, struct page *page,
3125 unsigned long offset, u16 len)
3130 rval = _extend_sdma_tx_descs(dd, tx);
3132 __sdma_txclean(dd, tx);
3136 /* If coalesce buffer is allocated, copy data into it */
3137 if (tx->coalesce_buf) {
3138 if (type == SDMA_MAP_NONE) {
3139 __sdma_txclean(dd, tx);
3143 if (type == SDMA_MAP_PAGE) {
3144 kvaddr = kmap(page);
3146 } else if (WARN_ON(!kvaddr)) {
3147 __sdma_txclean(dd, tx);
3151 memcpy(tx->coalesce_buf + tx->coalesce_idx, kvaddr, len);
3152 tx->coalesce_idx += len;
3153 if (type == SDMA_MAP_PAGE)
3156 /* If there is more data, return */
3157 if (tx->tlen - tx->coalesce_idx)
3160 /* Whole packet is received; add any padding */
3161 pad_len = tx->packet_len & (sizeof(u32) - 1);
3163 pad_len = sizeof(u32) - pad_len;
3164 memset(tx->coalesce_buf + tx->coalesce_idx, 0, pad_len);
3165 /* padding is taken care of for coalescing case */
3166 tx->packet_len += pad_len;
3167 tx->tlen += pad_len;
3170 /* dma map the coalesce buffer */
3171 addr = dma_map_single(&dd->pcidev->dev,
3176 if (unlikely(dma_mapping_error(&dd->pcidev->dev, addr))) {
3177 __sdma_txclean(dd, tx);
3181 /* Add descriptor for coalesce buffer */
3182 tx->desc_limit = MAX_DESC;
3183 return _sdma_txadd_daddr(dd, SDMA_MAP_SINGLE, tx,
3190 /* Update sdes when the lmc changes */
3191 void sdma_update_lmc(struct hfi1_devdata *dd, u64 mask, u32 lid)
3193 struct sdma_engine *sde;
3197 sreg = ((mask & SD(CHECK_SLID_MASK_MASK)) <<
3198 SD(CHECK_SLID_MASK_SHIFT)) |
3199 (((lid & mask) & SD(CHECK_SLID_VALUE_MASK)) <<
3200 SD(CHECK_SLID_VALUE_SHIFT));
3202 for (i = 0; i < dd->num_sdma; i++) {
3203 hfi1_cdbg(LINKVERB, "SendDmaEngine[%d].SLID_CHECK = 0x%x",
3205 sde = &dd->per_sdma[i];
3206 write_sde_csr(sde, SD(CHECK_SLID), sreg);
3210 /* tx not dword sized - pad */
3211 int _pad_sdma_tx_descs(struct hfi1_devdata *dd, struct sdma_txreq *tx)
3216 if ((unlikely(tx->num_desc == tx->desc_limit))) {
3217 rval = _extend_sdma_tx_descs(dd, tx);
3219 __sdma_txclean(dd, tx);
3223 /* finish the one just added */
3228 sizeof(u32) - (tx->packet_len & (sizeof(u32) - 1)));
3229 _sdma_close_tx(dd, tx);
3234 * Add ahg to the sdma_txreq
3236 * The logic will consume up to 3
3237 * descriptors at the beginning of
3240 void _sdma_txreq_ahgadd(
3241 struct sdma_txreq *tx,
3247 u32 i, shift = 0, desc = 0;
3250 WARN_ON_ONCE(num_ahg > 9 || (ahg_hlen & 3) || ahg_hlen == 4);
3253 mode = SDMA_AHG_APPLY_UPDATE1;
3254 else if (num_ahg <= 5)
3255 mode = SDMA_AHG_APPLY_UPDATE2;
3257 mode = SDMA_AHG_APPLY_UPDATE3;
3259 /* initialize to consumed descriptors to zero */
3261 case SDMA_AHG_APPLY_UPDATE3:
3263 tx->descs[2].qw[0] = 0;
3264 tx->descs[2].qw[1] = 0;
3266 case SDMA_AHG_APPLY_UPDATE2:
3268 tx->descs[1].qw[0] = 0;
3269 tx->descs[1].qw[1] = 0;
3273 tx->descs[0].qw[1] |=
3274 (((u64)ahg_entry & SDMA_DESC1_HEADER_INDEX_MASK)
3275 << SDMA_DESC1_HEADER_INDEX_SHIFT) |
3276 (((u64)ahg_hlen & SDMA_DESC1_HEADER_DWS_MASK)
3277 << SDMA_DESC1_HEADER_DWS_SHIFT) |
3278 (((u64)mode & SDMA_DESC1_HEADER_MODE_MASK)
3279 << SDMA_DESC1_HEADER_MODE_SHIFT) |
3280 (((u64)ahg[0] & SDMA_DESC1_HEADER_UPDATE1_MASK)
3281 << SDMA_DESC1_HEADER_UPDATE1_SHIFT);
3282 for (i = 0; i < (num_ahg - 1); i++) {
3283 if (!shift && !(i & 2))
3285 tx->descs[desc].qw[!!(i & 2)] |=
3288 shift = (shift + 32) & 63;
3293 * sdma_ahg_alloc - allocate an AHG entry
3294 * @sde: engine to allocate from
3297 * 0-31 when successful, -EOPNOTSUPP if AHG is not enabled,
3298 * -ENOSPC if an entry is not available
3300 int sdma_ahg_alloc(struct sdma_engine *sde)
3306 trace_hfi1_ahg_allocate(sde, -EINVAL);
3310 nr = ffz(ACCESS_ONCE(sde->ahg_bits));
3312 trace_hfi1_ahg_allocate(sde, -ENOSPC);
3315 oldbit = test_and_set_bit(nr, &sde->ahg_bits);
3320 trace_hfi1_ahg_allocate(sde, nr);
3325 * sdma_ahg_free - free an AHG entry
3326 * @sde: engine to return AHG entry
3327 * @ahg_index: index to free
3329 * This routine frees the indicate AHG entry.
3331 void sdma_ahg_free(struct sdma_engine *sde, int ahg_index)
3335 trace_hfi1_ahg_deallocate(sde, ahg_index);
3336 if (ahg_index < 0 || ahg_index > 31)
3338 clear_bit(ahg_index, &sde->ahg_bits);
3342 * SPC freeze handling for SDMA engines. Called when the driver knows
3343 * the SPC is going into a freeze but before the freeze is fully
3344 * settled. Generally an error interrupt.
3346 * This event will pull the engine out of running so no more entries can be
3347 * added to the engine's queue.
3349 void sdma_freeze_notify(struct hfi1_devdata *dd, int link_down)
3352 enum sdma_events event = link_down ? sdma_event_e85_link_down :
3353 sdma_event_e80_hw_freeze;
3355 /* set up the wait but do not wait here */
3356 atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3358 /* tell all engines to stop running and wait */
3359 for (i = 0; i < dd->num_sdma; i++)
3360 sdma_process_event(&dd->per_sdma[i], event);
3362 /* sdma_freeze() will wait for all engines to have stopped */
3366 * SPC freeze handling for SDMA engines. Called when the driver knows
3367 * the SPC is fully frozen.
3369 void sdma_freeze(struct hfi1_devdata *dd)
3375 * Make sure all engines have moved out of the running state before
3378 ret = wait_event_interruptible(dd->sdma_unfreeze_wq,
3379 atomic_read(&dd->sdma_unfreeze_count) <=
3381 /* interrupted or count is negative, then unloading - just exit */
3382 if (ret || atomic_read(&dd->sdma_unfreeze_count) < 0)
3385 /* set up the count for the next wait */
3386 atomic_set(&dd->sdma_unfreeze_count, dd->num_sdma);
3388 /* tell all engines that the SPC is frozen, they can start cleaning */
3389 for (i = 0; i < dd->num_sdma; i++)
3390 sdma_process_event(&dd->per_sdma[i], sdma_event_e81_hw_frozen);
3393 * Wait for everyone to finish software clean before exiting. The
3394 * software clean will read engine CSRs, so must be completed before
3395 * the next step, which will clear the engine CSRs.
3397 (void)wait_event_interruptible(dd->sdma_unfreeze_wq,
3398 atomic_read(&dd->sdma_unfreeze_count) <= 0);
3399 /* no need to check results - done no matter what */
3403 * SPC freeze handling for the SDMA engines. Called after the SPC is unfrozen.
3405 * The SPC freeze acts like a SDMA halt and a hardware clean combined. All
3406 * that is left is a software clean. We could do it after the SPC is fully
3407 * frozen, but then we'd have to add another state to wait for the unfreeze.
3408 * Instead, just defer the software clean until the unfreeze step.
3410 void sdma_unfreeze(struct hfi1_devdata *dd)
3414 /* tell all engines start freeze clean up */
3415 for (i = 0; i < dd->num_sdma; i++)
3416 sdma_process_event(&dd->per_sdma[i],
3417 sdma_event_e82_hw_unfreeze);
3421 * _sdma_engine_progress_schedule() - schedule progress on engine
3422 * @sde: sdma_engine to schedule progress
3425 void _sdma_engine_progress_schedule(
3426 struct sdma_engine *sde)
3428 trace_hfi1_sdma_engine_progress(sde, sde->progress_mask);
3429 /* assume we have selected a good cpu */
3431 CCE_INT_FORCE + (8 * (IS_SDMA_START / 64)),
3432 sde->progress_mask);