1 // SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause
3 * Copyright(c) 2015-2018 Intel Corporation.
6 #include <linux/delay.h>
11 #define SC(name) SEND_CTXT_##name
13 * Send Context functions
15 static void sc_wait_for_packet_egress(struct send_context *sc, int pause);
18 * Set the CM reset bit and wait for it to clear. Use the provided
19 * sendctrl register. This routine has no locking.
21 void __cm_reset(struct hfi1_devdata *dd, u64 sendctrl)
23 write_csr(dd, SEND_CTRL, sendctrl | SEND_CTRL_CM_RESET_SMASK);
26 sendctrl = read_csr(dd, SEND_CTRL);
27 if ((sendctrl & SEND_CTRL_CM_RESET_SMASK) == 0)
32 /* global control of PIO send */
33 void pio_send_control(struct hfi1_devdata *dd, int op)
37 int write = 1; /* write sendctrl back */
38 int flush = 0; /* re-read sendctrl to make sure it is flushed */
41 spin_lock_irqsave(&dd->sendctrl_lock, flags);
43 reg = read_csr(dd, SEND_CTRL);
45 case PSC_GLOBAL_ENABLE:
46 reg |= SEND_CTRL_SEND_ENABLE_SMASK;
48 case PSC_DATA_VL_ENABLE:
50 for (i = 0; i < ARRAY_SIZE(dd->vld); i++)
53 /* Disallow sending on VLs not enabled */
54 mask = (mask & SEND_CTRL_UNSUPPORTED_VL_MASK) <<
55 SEND_CTRL_UNSUPPORTED_VL_SHIFT;
56 reg = (reg & ~SEND_CTRL_UNSUPPORTED_VL_SMASK) | mask;
58 case PSC_GLOBAL_DISABLE:
59 reg &= ~SEND_CTRL_SEND_ENABLE_SMASK;
61 case PSC_GLOBAL_VLARB_ENABLE:
62 reg |= SEND_CTRL_VL_ARBITER_ENABLE_SMASK;
64 case PSC_GLOBAL_VLARB_DISABLE:
65 reg &= ~SEND_CTRL_VL_ARBITER_ENABLE_SMASK;
69 write = 0; /* CSR already written (and flushed) */
71 case PSC_DATA_VL_DISABLE:
72 reg |= SEND_CTRL_UNSUPPORTED_VL_SMASK;
76 dd_dev_err(dd, "%s: invalid control %d\n", __func__, op);
81 write_csr(dd, SEND_CTRL, reg);
83 (void)read_csr(dd, SEND_CTRL); /* flush write */
86 spin_unlock_irqrestore(&dd->sendctrl_lock, flags);
89 /* number of send context memory pools */
90 #define NUM_SC_POOLS 2
92 /* Send Context Size (SCS) wildcards */
96 /* Send Context Count (SCC) wildcards */
98 #define SCC_PER_CPU -2
99 #define SCC_PER_KRCVQ -3
101 /* Send Context Size (SCS) constants */
102 #define SCS_ACK_CREDITS 32
103 #define SCS_VL15_CREDITS 102 /* 3 pkts of 2048B data + 128B header */
105 #define PIO_THRESHOLD_CEILING 4096
107 #define PIO_WAIT_BATCH_SIZE 5
109 /* default send context sizes */
110 static struct sc_config_sizes sc_config_sizes[SC_MAX] = {
111 [SC_KERNEL] = { .size = SCS_POOL_0, /* even divide, pool 0 */
112 .count = SCC_PER_VL }, /* one per NUMA */
113 [SC_ACK] = { .size = SCS_ACK_CREDITS,
114 .count = SCC_PER_KRCVQ },
115 [SC_USER] = { .size = SCS_POOL_0, /* even divide, pool 0 */
116 .count = SCC_PER_CPU }, /* one per CPU */
117 [SC_VL15] = { .size = SCS_VL15_CREDITS,
122 /* send context memory pool configuration */
123 struct mem_pool_config {
124 int centipercent; /* % of memory, in 100ths of 1% */
125 int absolute_blocks; /* absolute block count */
128 /* default memory pool configuration: 100% in pool 0 */
129 static struct mem_pool_config sc_mem_pool_config[NUM_SC_POOLS] = {
130 /* centi%, abs blocks */
131 { 10000, -1 }, /* pool 0 */
132 { 0, -1 }, /* pool 1 */
135 /* memory pool information, used when calculating final sizes */
136 struct mem_pool_info {
138 * 100th of 1% of memory to use, -1 if blocks
141 int count; /* count of contexts in the pool */
142 int blocks; /* block size of the pool */
143 int size; /* context size, in blocks */
147 * Convert a pool wildcard to a valid pool index. The wildcards
148 * start at -1 and increase negatively. Map them as:
153 * Return -1 on non-wildcard input, otherwise convert to a pool number.
155 static int wildcard_to_pool(int wc)
158 return -1; /* non-wildcard */
162 static const char *sc_type_names[SC_MAX] = {
169 static const char *sc_type_name(int index)
171 if (index < 0 || index >= SC_MAX)
173 return sc_type_names[index];
177 * Read the send context memory pool configuration and send context
178 * size configuration. Replace any wildcards and come up with final
179 * counts and sizes for the send context types.
181 int init_sc_pools_and_sizes(struct hfi1_devdata *dd)
183 struct mem_pool_info mem_pool_info[NUM_SC_POOLS] = { { 0 } };
184 int total_blocks = (chip_pio_mem_size(dd) / PIO_BLOCK_SIZE) - 1;
185 int total_contexts = 0;
189 int cp_total; /* centipercent total */
190 int ab_total; /* absolute block total */
195 * When SDMA is enabled, kernel context pio packet size is capped by
196 * "piothreshold". Reduce pio buffer allocation for kernel context by
197 * setting it to a fixed size. The allocation allows 3-deep buffering
198 * of the largest pio packets plus up to 128 bytes header, sufficient
199 * to maintain verbs performance.
201 * When SDMA is disabled, keep the default pooling allocation.
203 if (HFI1_CAP_IS_KSET(SDMA)) {
204 u16 max_pkt_size = (piothreshold < PIO_THRESHOLD_CEILING) ?
205 piothreshold : PIO_THRESHOLD_CEILING;
206 sc_config_sizes[SC_KERNEL].size =
207 3 * (max_pkt_size + 128) / PIO_BLOCK_SIZE;
212 * - copy the centipercents/absolute sizes from the pool config
213 * - sanity check these values
214 * - add up centipercents, then later check for full value
215 * - add up absolute blocks, then later check for over-commit
219 for (i = 0; i < NUM_SC_POOLS; i++) {
220 int cp = sc_mem_pool_config[i].centipercent;
221 int ab = sc_mem_pool_config[i].absolute_blocks;
224 * A negative value is "unused" or "invalid". Both *can*
225 * be valid, but centipercent wins, so check that first
227 if (cp >= 0) { /* centipercent valid */
229 } else if (ab >= 0) { /* absolute blocks valid */
231 } else { /* neither valid */
234 "Send context memory pool %d: both the block count and centipercent are invalid\n",
239 mem_pool_info[i].centipercent = cp;
240 mem_pool_info[i].blocks = ab;
243 /* do not use both % and absolute blocks for different pools */
244 if (cp_total != 0 && ab_total != 0) {
247 "All send context memory pools must be described as either centipercent or blocks, no mixing between pools\n");
251 /* if any percentages are present, they must add up to 100% x 100 */
252 if (cp_total != 0 && cp_total != 10000) {
255 "Send context memory pool centipercent is %d, expecting 10000\n",
260 /* the absolute pool total cannot be more than the mem total */
261 if (ab_total > total_blocks) {
264 "Send context memory pool absolute block count %d is larger than the memory size %d\n",
265 ab_total, total_blocks);
271 * - copy from the context size config
272 * - replace context type wildcard counts with real values
273 * - add up non-memory pool block sizes
274 * - add up memory pool user counts
277 for (i = 0; i < SC_MAX; i++) {
278 int count = sc_config_sizes[i].count;
279 int size = sc_config_sizes[i].size;
283 * Sanity check count: Either a positive value or
284 * one of the expected wildcards is valid. The positive
285 * value is checked later when we compare against total
289 count = dd->n_krcv_queues;
290 } else if (i == SC_KERNEL) {
291 count = INIT_SC_PER_VL * num_vls;
292 } else if (count == SCC_PER_CPU) {
293 count = dd->num_rcv_contexts - dd->n_krcv_queues;
294 } else if (count < 0) {
297 "%s send context invalid count wildcard %d\n",
298 sc_type_name(i), count);
301 if (total_contexts + count > chip_send_contexts(dd))
302 count = chip_send_contexts(dd) - total_contexts;
304 total_contexts += count;
307 * Sanity check pool: The conversion will return a pool
308 * number or -1 if a fixed (non-negative) value. The fixed
309 * value is checked later when we compare against
310 * total memory available.
312 pool = wildcard_to_pool(size);
313 if (pool == -1) { /* non-wildcard */
314 fixed_blocks += size * count;
315 } else if (pool < NUM_SC_POOLS) { /* valid wildcard */
316 mem_pool_info[pool].count += count;
317 } else { /* invalid wildcard */
320 "%s send context invalid pool wildcard %d\n",
321 sc_type_name(i), size);
325 dd->sc_sizes[i].count = count;
326 dd->sc_sizes[i].size = size;
328 if (fixed_blocks > total_blocks) {
331 "Send context fixed block count, %u, larger than total block count %u\n",
332 fixed_blocks, total_blocks);
336 /* step 3: calculate the blocks in the pools, and pool context sizes */
337 pool_blocks = total_blocks - fixed_blocks;
338 if (ab_total > pool_blocks) {
341 "Send context fixed pool sizes, %u, larger than pool block count %u\n",
342 ab_total, pool_blocks);
345 /* subtract off the fixed pool blocks */
346 pool_blocks -= ab_total;
348 for (i = 0; i < NUM_SC_POOLS; i++) {
349 struct mem_pool_info *pi = &mem_pool_info[i];
351 /* % beats absolute blocks */
352 if (pi->centipercent >= 0)
353 pi->blocks = (pool_blocks * pi->centipercent) / 10000;
355 if (pi->blocks == 0 && pi->count != 0) {
358 "Send context memory pool %d has %u contexts, but no blocks\n",
362 if (pi->count == 0) {
363 /* warn about wasted blocks */
367 "Send context memory pool %d has %u blocks, but zero contexts\n",
371 pi->size = pi->blocks / pi->count;
375 /* step 4: fill in the context type sizes from the pool sizes */
377 for (i = 0; i < SC_MAX; i++) {
378 if (dd->sc_sizes[i].size < 0) {
379 unsigned pool = wildcard_to_pool(dd->sc_sizes[i].size);
381 WARN_ON_ONCE(pool >= NUM_SC_POOLS);
382 dd->sc_sizes[i].size = mem_pool_info[pool].size;
384 /* make sure we are not larger than what is allowed by the HW */
385 #define PIO_MAX_BLOCKS 1024
386 if (dd->sc_sizes[i].size > PIO_MAX_BLOCKS)
387 dd->sc_sizes[i].size = PIO_MAX_BLOCKS;
389 /* calculate our total usage */
390 used_blocks += dd->sc_sizes[i].size * dd->sc_sizes[i].count;
392 extra = total_blocks - used_blocks;
394 dd_dev_info(dd, "unused send context blocks: %d\n", extra);
396 return total_contexts;
399 int init_send_contexts(struct hfi1_devdata *dd)
402 int ret, i, j, context;
404 ret = init_credit_return(dd);
408 dd->hw_to_sw = kmalloc_array(TXE_NUM_CONTEXTS, sizeof(u8),
410 dd->send_contexts = kcalloc(dd->num_send_contexts,
411 sizeof(struct send_context_info),
413 if (!dd->send_contexts || !dd->hw_to_sw) {
415 kfree(dd->send_contexts);
416 free_credit_return(dd);
420 /* hardware context map starts with invalid send context indices */
421 for (i = 0; i < TXE_NUM_CONTEXTS; i++)
422 dd->hw_to_sw[i] = INVALID_SCI;
425 * All send contexts have their credit sizes. Allocate credits
426 * for each context one after another from the global space.
430 for (i = 0; i < SC_MAX; i++) {
431 struct sc_config_sizes *scs = &dd->sc_sizes[i];
433 for (j = 0; j < scs->count; j++) {
434 struct send_context_info *sci =
435 &dd->send_contexts[context];
438 sci->credits = scs->size;
449 * Allocate a software index and hardware context of the given type.
451 * Must be called with dd->sc_lock held.
453 static int sc_hw_alloc(struct hfi1_devdata *dd, int type, u32 *sw_index,
456 struct send_context_info *sci;
460 for (index = 0, sci = &dd->send_contexts[0];
461 index < dd->num_send_contexts; index++, sci++) {
462 if (sci->type == type && sci->allocated == 0) {
464 /* use a 1:1 mapping, but make them non-equal */
465 context = chip_send_contexts(dd) - index - 1;
466 dd->hw_to_sw[context] = index;
468 *hw_context = context;
469 return 0; /* success */
472 dd_dev_err(dd, "Unable to locate a free type %d send context\n", type);
477 * Free the send context given by its software index.
479 * Must be called with dd->sc_lock held.
481 static void sc_hw_free(struct hfi1_devdata *dd, u32 sw_index, u32 hw_context)
483 struct send_context_info *sci;
485 sci = &dd->send_contexts[sw_index];
486 if (!sci->allocated) {
487 dd_dev_err(dd, "%s: sw_index %u not allocated? hw_context %u\n",
488 __func__, sw_index, hw_context);
491 dd->hw_to_sw[hw_context] = INVALID_SCI;
494 /* return the base context of a context in a group */
495 static inline u32 group_context(u32 context, u32 group)
497 return (context >> group) << group;
500 /* return the size of a group */
501 static inline u32 group_size(u32 group)
507 * Obtain the credit return addresses, kernel virtual and bus, for the
510 * To understand this routine:
511 * o va and dma are arrays of struct credit_return. One for each physical
512 * send context, per NUMA.
513 * o Each send context always looks in its relative location in a struct
514 * credit_return for its credit return.
515 * o Each send context in a group must have its return address CSR programmed
516 * with the same value. Use the address of the first send context in the
519 static void cr_group_addresses(struct send_context *sc, dma_addr_t *dma)
521 u32 gc = group_context(sc->hw_context, sc->group);
522 u32 index = sc->hw_context & 0x7;
524 sc->hw_free = &sc->dd->cr_base[sc->node].va[gc].cr[index];
525 *dma = (unsigned long)
526 &((struct credit_return *)sc->dd->cr_base[sc->node].dma)[gc];
530 * Work queue function triggered in error interrupt routine for
533 static void sc_halted(struct work_struct *work)
535 struct send_context *sc;
537 sc = container_of(work, struct send_context, halt_work);
542 * Calculate PIO block threshold for this send context using the given MTU.
543 * Trigger a return when one MTU plus optional header of credits remain.
545 * Parameter mtu is in bytes.
546 * Parameter hdrqentsize is in DWORDs.
548 * Return value is what to write into the CSR: trigger return when
549 * unreturned credits pass this count.
551 u32 sc_mtu_to_threshold(struct send_context *sc, u32 mtu, u32 hdrqentsize)
556 /* add in the header size, then divide by the PIO block size */
557 mtu += hdrqentsize << 2;
558 release_credits = DIV_ROUND_UP(mtu, PIO_BLOCK_SIZE);
560 /* check against this context's credits */
561 if (sc->credits <= release_credits)
564 threshold = sc->credits - release_credits;
570 * Calculate credit threshold in terms of percent of the allocated credits.
571 * Trigger when unreturned credits equal or exceed the percentage of the whole.
573 * Return value is what to write into the CSR: trigger return when
574 * unreturned credits pass this count.
576 u32 sc_percent_to_threshold(struct send_context *sc, u32 percent)
578 return (sc->credits * percent) / 100;
582 * Set the credit return threshold.
584 void sc_set_cr_threshold(struct send_context *sc, u32 new_threshold)
588 int force_return = 0;
590 spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
592 old_threshold = (sc->credit_ctrl >>
593 SC(CREDIT_CTRL_THRESHOLD_SHIFT))
594 & SC(CREDIT_CTRL_THRESHOLD_MASK);
596 if (new_threshold != old_threshold) {
599 & ~SC(CREDIT_CTRL_THRESHOLD_SMASK))
601 & SC(CREDIT_CTRL_THRESHOLD_MASK))
602 << SC(CREDIT_CTRL_THRESHOLD_SHIFT));
603 write_kctxt_csr(sc->dd, sc->hw_context,
604 SC(CREDIT_CTRL), sc->credit_ctrl);
606 /* force a credit return on change to avoid a possible stall */
610 spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
613 sc_return_credits(sc);
619 * Set the CHECK_ENABLE register for the send context 'sc'.
621 void set_pio_integrity(struct send_context *sc)
623 struct hfi1_devdata *dd = sc->dd;
624 u32 hw_context = sc->hw_context;
627 write_kctxt_csr(dd, hw_context,
629 hfi1_pkt_default_send_ctxt_mask(dd, type));
632 static u32 get_buffers_allocated(struct send_context *sc)
637 for_each_possible_cpu(cpu)
638 ret += *per_cpu_ptr(sc->buffers_allocated, cpu);
642 static void reset_buffers_allocated(struct send_context *sc)
646 for_each_possible_cpu(cpu)
647 (*per_cpu_ptr(sc->buffers_allocated, cpu)) = 0;
651 * Allocate a NUMA relative send context structure of the given type along
654 struct send_context *sc_alloc(struct hfi1_devdata *dd, int type,
655 uint hdrqentsize, int numa)
657 struct send_context_info *sci;
658 struct send_context *sc = NULL;
668 /* do not allocate while frozen */
669 if (dd->flags & HFI1_FROZEN)
672 sc = kzalloc_node(sizeof(*sc), GFP_KERNEL, numa);
676 sc->buffers_allocated = alloc_percpu(u32);
677 if (!sc->buffers_allocated) {
680 "Cannot allocate buffers_allocated per cpu counters\n"
685 spin_lock_irqsave(&dd->sc_lock, flags);
686 ret = sc_hw_alloc(dd, type, &sw_index, &hw_context);
688 spin_unlock_irqrestore(&dd->sc_lock, flags);
689 free_percpu(sc->buffers_allocated);
694 sci = &dd->send_contexts[sw_index];
700 spin_lock_init(&sc->alloc_lock);
701 spin_lock_init(&sc->release_lock);
702 spin_lock_init(&sc->credit_ctrl_lock);
703 seqlock_init(&sc->waitlock);
704 INIT_LIST_HEAD(&sc->piowait);
705 INIT_WORK(&sc->halt_work, sc_halted);
706 init_waitqueue_head(&sc->halt_wait);
708 /* grouping is always single context for now */
711 sc->sw_index = sw_index;
712 sc->hw_context = hw_context;
713 cr_group_addresses(sc, &dma);
714 sc->credits = sci->credits;
715 sc->size = sc->credits * PIO_BLOCK_SIZE;
717 /* PIO Send Memory Address details */
718 #define PIO_ADDR_CONTEXT_MASK 0xfful
719 #define PIO_ADDR_CONTEXT_SHIFT 16
720 sc->base_addr = dd->piobase + ((hw_context & PIO_ADDR_CONTEXT_MASK)
721 << PIO_ADDR_CONTEXT_SHIFT);
723 /* set base and credits */
724 reg = ((sci->credits & SC(CTRL_CTXT_DEPTH_MASK))
725 << SC(CTRL_CTXT_DEPTH_SHIFT))
726 | ((sci->base & SC(CTRL_CTXT_BASE_MASK))
727 << SC(CTRL_CTXT_BASE_SHIFT));
728 write_kctxt_csr(dd, hw_context, SC(CTRL), reg);
730 set_pio_integrity(sc);
732 /* unmask all errors */
733 write_kctxt_csr(dd, hw_context, SC(ERR_MASK), (u64)-1);
735 /* set the default partition key */
736 write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY),
737 (SC(CHECK_PARTITION_KEY_VALUE_MASK) &
739 SC(CHECK_PARTITION_KEY_VALUE_SHIFT));
741 /* per context type checks */
742 if (type == SC_USER) {
743 opval = USER_OPCODE_CHECK_VAL;
744 opmask = USER_OPCODE_CHECK_MASK;
746 opval = OPCODE_CHECK_VAL_DISABLED;
747 opmask = OPCODE_CHECK_MASK_DISABLED;
750 /* set the send context check opcode mask and value */
751 write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE),
752 ((u64)opmask << SC(CHECK_OPCODE_MASK_SHIFT)) |
753 ((u64)opval << SC(CHECK_OPCODE_VALUE_SHIFT)));
755 /* set up credit return */
756 reg = dma & SC(CREDIT_RETURN_ADDR_ADDRESS_SMASK);
757 write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), reg);
760 * Calculate the initial credit return threshold.
762 * For Ack contexts, set a threshold for half the credits.
763 * For User contexts use the given percentage. This has been
764 * sanitized on driver start-up.
765 * For Kernel contexts, use the default MTU plus a header
766 * or half the credits, whichever is smaller. This should
767 * work for both the 3-deep buffering allocation and the
768 * pooling allocation.
770 if (type == SC_ACK) {
771 thresh = sc_percent_to_threshold(sc, 50);
772 } else if (type == SC_USER) {
773 thresh = sc_percent_to_threshold(sc,
774 user_credit_return_threshold);
775 } else { /* kernel */
776 thresh = min(sc_percent_to_threshold(sc, 50),
777 sc_mtu_to_threshold(sc, hfi1_max_mtu,
780 reg = thresh << SC(CREDIT_CTRL_THRESHOLD_SHIFT);
781 /* add in early return */
782 if (type == SC_USER && HFI1_CAP_IS_USET(EARLY_CREDIT_RETURN))
783 reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
784 else if (HFI1_CAP_IS_KSET(EARLY_CREDIT_RETURN)) /* kernel, ack */
785 reg |= SC(CREDIT_CTRL_EARLY_RETURN_SMASK);
787 /* set up write-through credit_ctrl */
788 sc->credit_ctrl = reg;
789 write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), reg);
791 /* User send contexts should not allow sending on VL15 */
792 if (type == SC_USER) {
794 write_kctxt_csr(dd, hw_context, SC(CHECK_VL), reg);
797 spin_unlock_irqrestore(&dd->sc_lock, flags);
800 * Allocate shadow ring to track outstanding PIO buffers _after_
801 * unlocking. We don't know the size until the lock is held and
802 * we can't allocate while the lock is held. No one is using
803 * the context yet, so allocate it now.
805 * User contexts do not get a shadow ring.
807 if (type != SC_USER) {
809 * Size the shadow ring 1 larger than the number of credits
810 * so head == tail can mean empty.
812 sc->sr_size = sci->credits + 1;
813 sc->sr = kcalloc_node(sc->sr_size,
814 sizeof(union pio_shadow_ring),
823 "Send context %u(%u) %s group %u credits %u credit_ctrl 0x%llx threshold %u\n",
835 /* free a per-NUMA send context structure */
836 void sc_free(struct send_context *sc)
838 struct hfi1_devdata *dd;
846 sc->flags |= SCF_IN_FREE; /* ensure no restarts */
848 if (!list_empty(&sc->piowait))
849 dd_dev_err(dd, "piowait list not empty!\n");
850 sw_index = sc->sw_index;
851 hw_context = sc->hw_context;
852 sc_disable(sc); /* make sure the HW is disabled */
853 flush_work(&sc->halt_work);
855 spin_lock_irqsave(&dd->sc_lock, flags);
856 dd->send_contexts[sw_index].sc = NULL;
858 /* clear/disable all registers set in sc_alloc */
859 write_kctxt_csr(dd, hw_context, SC(CTRL), 0);
860 write_kctxt_csr(dd, hw_context, SC(CHECK_ENABLE), 0);
861 write_kctxt_csr(dd, hw_context, SC(ERR_MASK), 0);
862 write_kctxt_csr(dd, hw_context, SC(CHECK_PARTITION_KEY), 0);
863 write_kctxt_csr(dd, hw_context, SC(CHECK_OPCODE), 0);
864 write_kctxt_csr(dd, hw_context, SC(CREDIT_RETURN_ADDR), 0);
865 write_kctxt_csr(dd, hw_context, SC(CREDIT_CTRL), 0);
867 /* release the index and context for re-use */
868 sc_hw_free(dd, sw_index, hw_context);
869 spin_unlock_irqrestore(&dd->sc_lock, flags);
872 free_percpu(sc->buffers_allocated);
876 /* disable the context */
877 void sc_disable(struct send_context *sc)
880 struct pio_buf *pbuf;
881 LIST_HEAD(wake_list);
886 /* do all steps, even if already disabled */
887 spin_lock_irq(&sc->alloc_lock);
888 reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL));
889 reg &= ~SC(CTRL_CTXT_ENABLE_SMASK);
890 sc->flags &= ~SCF_ENABLED;
891 sc_wait_for_packet_egress(sc, 1);
892 write_kctxt_csr(sc->dd, sc->hw_context, SC(CTRL), reg);
895 * Flush any waiters. Once the context is disabled,
896 * credit return interrupts are stopped (although there
897 * could be one in-process when the context is disabled).
898 * Wait one microsecond for any lingering interrupts, then
899 * proceed with the flush.
902 spin_lock(&sc->release_lock);
903 if (sc->sr) { /* this context has a shadow ring */
904 while (sc->sr_tail != sc->sr_head) {
905 pbuf = &sc->sr[sc->sr_tail].pbuf;
907 (*pbuf->cb)(pbuf->arg, PRC_SC_DISABLE);
909 if (sc->sr_tail >= sc->sr_size)
913 spin_unlock(&sc->release_lock);
915 write_seqlock(&sc->waitlock);
916 if (!list_empty(&sc->piowait))
917 list_move(&sc->piowait, &wake_list);
918 write_sequnlock(&sc->waitlock);
919 while (!list_empty(&wake_list)) {
922 struct hfi1_qp_priv *priv;
924 wait = list_first_entry(&wake_list, struct iowait, list);
925 qp = iowait_to_qp(wait);
927 list_del_init(&priv->s_iowait.list);
928 priv->s_iowait.lock = NULL;
929 hfi1_qp_wakeup(qp, RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
932 spin_unlock_irq(&sc->alloc_lock);
935 /* return SendEgressCtxtStatus.PacketOccupancy */
936 static u64 packet_occupancy(u64 reg)
939 SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SMASK)
940 >> SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_PACKET_OCCUPANCY_SHIFT;
943 /* is egress halted on the context? */
944 static bool egress_halted(u64 reg)
946 return !!(reg & SEND_EGRESS_CTXT_STATUS_CTXT_EGRESS_HALT_STATUS_SMASK);
949 /* is the send context halted? */
950 static bool is_sc_halted(struct hfi1_devdata *dd, u32 hw_context)
952 return !!(read_kctxt_csr(dd, hw_context, SC(STATUS)) &
953 SC(STATUS_CTXT_HALTED_SMASK));
957 * sc_wait_for_packet_egress - wait for packet
958 * @sc: valid send context
959 * @pause: wait for credit return
961 * Wait for packet egress, optionally pause for credit return
963 * Egress halt and Context halt are not necessarily the same thing, so
966 * NOTE: The context halt bit may not be set immediately. Because of this,
967 * it is necessary to check the SW SFC_HALTED bit (set in the IRQ) and the HW
968 * context bit to determine if the context is halted.
970 static void sc_wait_for_packet_egress(struct send_context *sc, int pause)
972 struct hfi1_devdata *dd = sc->dd;
979 reg = read_csr(dd, sc->hw_context * 8 +
980 SEND_EGRESS_CTXT_STATUS);
981 /* done if any halt bits, SW or HW are set */
982 if (sc->flags & SCF_HALTED ||
983 is_sc_halted(dd, sc->hw_context) || egress_halted(reg))
985 reg = packet_occupancy(reg);
988 /* counter is reset if occupancy count changes */
992 /* timed out - bounce the link */
994 "%s: context %u(%u) timeout waiting for packets to egress, remaining count %u, bouncing link\n",
995 __func__, sc->sw_index,
996 sc->hw_context, (u32)reg);
997 queue_work(dd->pport->link_wq,
998 &dd->pport->link_bounce_work);
1006 /* Add additional delay to ensure chip returns all credits */
1007 pause_for_credit_return(dd);
1010 void sc_wait(struct hfi1_devdata *dd)
1014 for (i = 0; i < dd->num_send_contexts; i++) {
1015 struct send_context *sc = dd->send_contexts[i].sc;
1019 sc_wait_for_packet_egress(sc, 0);
1024 * Restart a context after it has been halted due to error.
1026 * If the first step fails - wait for the halt to be asserted, return early.
1027 * Otherwise complain about timeouts but keep going.
1029 * It is expected that allocations (enabled flag bit) have been shut off
1030 * already (only applies to kernel contexts).
1032 int sc_restart(struct send_context *sc)
1034 struct hfi1_devdata *dd = sc->dd;
1039 /* bounce off if not halted, or being free'd */
1040 if (!(sc->flags & SCF_HALTED) || (sc->flags & SCF_IN_FREE))
1043 dd_dev_info(dd, "restarting send context %u(%u)\n", sc->sw_index,
1047 * Step 1: Wait for the context to actually halt.
1049 * The error interrupt is asynchronous to actually setting halt
1054 reg = read_kctxt_csr(dd, sc->hw_context, SC(STATUS));
1055 if (reg & SC(STATUS_CTXT_HALTED_SMASK))
1058 dd_dev_err(dd, "%s: context %u(%u) not halting, skipping\n",
1059 __func__, sc->sw_index, sc->hw_context);
1067 * Step 2: Ensure no users are still trying to write to PIO.
1069 * For kernel contexts, we have already turned off buffer allocation.
1070 * Now wait for the buffer count to go to zero.
1072 * For user contexts, the user handling code has cut off write access
1073 * to the context's PIO pages before calling this routine and will
1074 * restore write access after this routine returns.
1076 if (sc->type != SC_USER) {
1077 /* kernel context */
1080 count = get_buffers_allocated(sc);
1085 "%s: context %u(%u) timeout waiting for PIO buffers to zero, remaining %d\n",
1086 __func__, sc->sw_index,
1087 sc->hw_context, count);
1095 * Step 3: Wait for all packets to egress.
1096 * This is done while disabling the send context
1098 * Step 4: Disable the context
1100 * This is a superset of the halt. After the disable, the
1101 * errors can be cleared.
1106 * Step 5: Enable the context
1108 * This enable will clear the halted flag and per-send context
1111 return sc_enable(sc);
1115 * PIO freeze processing. To be called after the TXE block is fully frozen.
1116 * Go through all frozen send contexts and disable them. The contexts are
1117 * already stopped by the freeze.
1119 void pio_freeze(struct hfi1_devdata *dd)
1121 struct send_context *sc;
1124 for (i = 0; i < dd->num_send_contexts; i++) {
1125 sc = dd->send_contexts[i].sc;
1127 * Don't disable unallocated, unfrozen, or user send contexts.
1128 * User send contexts will be disabled when the process
1129 * calls into the driver to reset its context.
1131 if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1134 /* only need to disable, the context is already stopped */
1140 * Unfreeze PIO for kernel send contexts. The precondition for calling this
1141 * is that all PIO send contexts have been disabled and the SPC freeze has
1142 * been cleared. Now perform the last step and re-enable each kernel context.
1143 * User (PSM) processing will occur when PSM calls into the kernel to
1144 * acknowledge the freeze.
1146 void pio_kernel_unfreeze(struct hfi1_devdata *dd)
1148 struct send_context *sc;
1151 for (i = 0; i < dd->num_send_contexts; i++) {
1152 sc = dd->send_contexts[i].sc;
1153 if (!sc || !(sc->flags & SCF_FROZEN) || sc->type == SC_USER)
1155 if (sc->flags & SCF_LINK_DOWN)
1158 sc_enable(sc); /* will clear the sc frozen flag */
1163 * pio_kernel_linkup() - Re-enable send contexts after linkup event
1164 * @dd: valid devive data
1166 * When the link goes down, the freeze path is taken. However, a link down
1167 * event is different from a freeze because if the send context is re-enabled
1168 * whowever is sending data will start sending data again, which will hang
1169 * any QP that is sending data.
1171 * The freeze path now looks at the type of event that occurs and takes this
1172 * path for link down event.
1174 void pio_kernel_linkup(struct hfi1_devdata *dd)
1176 struct send_context *sc;
1179 for (i = 0; i < dd->num_send_contexts; i++) {
1180 sc = dd->send_contexts[i].sc;
1181 if (!sc || !(sc->flags & SCF_LINK_DOWN) || sc->type == SC_USER)
1184 sc_enable(sc); /* will clear the sc link down flag */
1189 * Wait for the SendPioInitCtxt.PioInitInProgress bit to clear.
1191 * -ETIMEDOUT - if we wait too long
1192 * -EIO - if there was an error
1194 static int pio_init_wait_progress(struct hfi1_devdata *dd)
1199 /* max is the longest possible HW init time / delay */
1200 max = (dd->icode == ICODE_FPGA_EMULATION) ? 120 : 5;
1202 reg = read_csr(dd, SEND_PIO_INIT_CTXT);
1203 if (!(reg & SEND_PIO_INIT_CTXT_PIO_INIT_IN_PROGRESS_SMASK))
1211 return reg & SEND_PIO_INIT_CTXT_PIO_INIT_ERR_SMASK ? -EIO : 0;
1215 * Reset all of the send contexts to their power-on state. Used
1216 * only during manual init - no lock against sc_enable needed.
1218 void pio_reset_all(struct hfi1_devdata *dd)
1222 /* make sure the init engine is not busy */
1223 ret = pio_init_wait_progress(dd);
1224 /* ignore any timeout */
1226 /* clear the error */
1227 write_csr(dd, SEND_PIO_ERR_CLEAR,
1228 SEND_PIO_ERR_CLEAR_PIO_INIT_SM_IN_ERR_SMASK);
1231 /* reset init all */
1232 write_csr(dd, SEND_PIO_INIT_CTXT,
1233 SEND_PIO_INIT_CTXT_PIO_ALL_CTXT_INIT_SMASK);
1235 ret = pio_init_wait_progress(dd);
1238 "PIO send context init %s while initializing all PIO blocks\n",
1239 ret == -ETIMEDOUT ? "is stuck" : "had an error");
1243 /* enable the context */
1244 int sc_enable(struct send_context *sc)
1246 u64 sc_ctrl, reg, pio;
1247 struct hfi1_devdata *dd;
1248 unsigned long flags;
1256 * Obtain the allocator lock to guard against any allocation
1257 * attempts (which should not happen prior to context being
1258 * enabled). On the release/disable side we don't need to
1259 * worry about locking since the releaser will not do anything
1260 * if the context accounting values have not changed.
1262 spin_lock_irqsave(&sc->alloc_lock, flags);
1263 sc_ctrl = read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1264 if ((sc_ctrl & SC(CTRL_CTXT_ENABLE_SMASK)))
1265 goto unlock; /* already enabled */
1267 /* IMPORTANT: only clear free and fill if transitioning 0 -> 1 */
1277 /* the alloc lock insures no fast path allocation */
1278 reset_buffers_allocated(sc);
1281 * Clear all per-context errors. Some of these will be set when
1282 * we are re-enabling after a context halt. Now that the context
1283 * is disabled, the halt will not clear until after the PIO init
1284 * engine runs below.
1286 reg = read_kctxt_csr(dd, sc->hw_context, SC(ERR_STATUS));
1288 write_kctxt_csr(dd, sc->hw_context, SC(ERR_CLEAR), reg);
1291 * The HW PIO initialization engine can handle only one init
1292 * request at a time. Serialize access to each device's engine.
1294 spin_lock(&dd->sc_init_lock);
1296 * Since access to this code block is serialized and
1297 * each access waits for the initialization to complete
1298 * before releasing the lock, the PIO initialization engine
1299 * should not be in use, so we don't have to wait for the
1300 * InProgress bit to go down.
1302 pio = ((sc->hw_context & SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_MASK) <<
1303 SEND_PIO_INIT_CTXT_PIO_CTXT_NUM_SHIFT) |
1304 SEND_PIO_INIT_CTXT_PIO_SINGLE_CTXT_INIT_SMASK;
1305 write_csr(dd, SEND_PIO_INIT_CTXT, pio);
1307 * Wait until the engine is done. Give the chip the required time
1308 * so, hopefully, we read the register just once.
1311 ret = pio_init_wait_progress(dd);
1312 spin_unlock(&dd->sc_init_lock);
1315 "sctxt%u(%u): Context not enabled due to init failure %d\n",
1316 sc->sw_index, sc->hw_context, ret);
1321 * All is well. Enable the context.
1323 sc_ctrl |= SC(CTRL_CTXT_ENABLE_SMASK);
1324 write_kctxt_csr(dd, sc->hw_context, SC(CTRL), sc_ctrl);
1326 * Read SendCtxtCtrl to force the write out and prevent a timing
1327 * hazard where a PIO write may reach the context before the enable.
1329 read_kctxt_csr(dd, sc->hw_context, SC(CTRL));
1330 sc->flags |= SCF_ENABLED;
1333 spin_unlock_irqrestore(&sc->alloc_lock, flags);
1338 /* force a credit return on the context */
1339 void sc_return_credits(struct send_context *sc)
1344 /* a 0->1 transition schedules a credit return */
1345 write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE),
1346 SC(CREDIT_FORCE_FORCE_RETURN_SMASK));
1348 * Ensure that the write is flushed and the credit return is
1349 * scheduled. We care more about the 0 -> 1 transition.
1351 read_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE));
1352 /* set back to 0 for next time */
1353 write_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_FORCE), 0);
1356 /* allow all in-flight packets to drain on the context */
1357 void sc_flush(struct send_context *sc)
1362 sc_wait_for_packet_egress(sc, 1);
1365 /* drop all packets on the context, no waiting until they are sent */
1366 void sc_drop(struct send_context *sc)
1371 dd_dev_info(sc->dd, "%s: context %u(%u) - not implemented\n",
1372 __func__, sc->sw_index, sc->hw_context);
1376 * Start the software reaction to a context halt or SPC freeze:
1377 * - mark the context as halted or frozen
1378 * - stop buffer allocations
1380 * Called from the error interrupt. Other work is deferred until
1381 * out of the interrupt.
1383 void sc_stop(struct send_context *sc, int flag)
1385 unsigned long flags;
1387 /* stop buffer allocations */
1388 spin_lock_irqsave(&sc->alloc_lock, flags);
1389 /* mark the context */
1391 sc->flags &= ~SCF_ENABLED;
1392 spin_unlock_irqrestore(&sc->alloc_lock, flags);
1393 wake_up(&sc->halt_wait);
1396 #define BLOCK_DWORDS (PIO_BLOCK_SIZE / sizeof(u32))
1397 #define dwords_to_blocks(x) DIV_ROUND_UP(x, BLOCK_DWORDS)
1400 * The send context buffer "allocator".
1402 * @sc: the PIO send context we are allocating from
1403 * @len: length of whole packet - including PBC - in dwords
1404 * @cb: optional callback to call when the buffer is finished sending
1405 * @arg: argument for cb
1407 * Return a pointer to a PIO buffer, NULL if not enough room, -ECOMM
1408 * when link is down.
1410 struct pio_buf *sc_buffer_alloc(struct send_context *sc, u32 dw_len,
1411 pio_release_cb cb, void *arg)
1413 struct pio_buf *pbuf = NULL;
1414 unsigned long flags;
1415 unsigned long avail;
1416 unsigned long blocks = dwords_to_blocks(dw_len);
1421 spin_lock_irqsave(&sc->alloc_lock, flags);
1422 if (!(sc->flags & SCF_ENABLED)) {
1423 spin_unlock_irqrestore(&sc->alloc_lock, flags);
1424 return ERR_PTR(-ECOMM);
1428 avail = (unsigned long)sc->credits - (sc->fill - sc->alloc_free);
1429 if (blocks > avail) {
1430 /* not enough room */
1431 if (unlikely(trycount)) { /* already tried to get more room */
1432 spin_unlock_irqrestore(&sc->alloc_lock, flags);
1435 /* copy from receiver cache line and recalculate */
1436 sc->alloc_free = READ_ONCE(sc->free);
1438 (unsigned long)sc->credits -
1439 (sc->fill - sc->alloc_free);
1440 if (blocks > avail) {
1441 /* still no room, actively update */
1442 sc_release_update(sc);
1443 sc->alloc_free = READ_ONCE(sc->free);
1449 /* there is enough room */
1452 this_cpu_inc(*sc->buffers_allocated);
1454 /* read this once */
1457 /* "allocate" the buffer */
1459 fill_wrap = sc->fill_wrap;
1460 sc->fill_wrap += blocks;
1461 if (sc->fill_wrap >= sc->credits)
1462 sc->fill_wrap = sc->fill_wrap - sc->credits;
1465 * Fill the parts that the releaser looks at before moving the head.
1466 * The only necessary piece is the sent_at field. The credits
1467 * we have just allocated cannot have been returned yet, so the
1468 * cb and arg will not be looked at for a "while". Put them
1469 * on this side of the memory barrier anyway.
1471 pbuf = &sc->sr[head].pbuf;
1472 pbuf->sent_at = sc->fill;
1475 pbuf->sc = sc; /* could be filled in at sc->sr init time */
1476 /* make sure this is in memory before updating the head */
1478 /* calculate next head index, do not store */
1480 if (next >= sc->sr_size)
1483 * update the head - must be last! - the releaser can look at fields
1484 * in pbuf once we move the head
1488 spin_unlock_irqrestore(&sc->alloc_lock, flags);
1490 /* finish filling in the buffer outside the lock */
1491 pbuf->start = sc->base_addr + fill_wrap * PIO_BLOCK_SIZE;
1492 pbuf->end = sc->base_addr + sc->size;
1493 pbuf->qw_written = 0;
1494 pbuf->carry_bytes = 0;
1495 pbuf->carry.val64 = 0;
1501 * There are at least two entities that can turn on credit return
1502 * interrupts and they can overlap. Avoid problems by implementing
1503 * a count scheme that is enforced by a lock. The lock is needed because
1504 * the count and CSR write must be paired.
1508 * Start credit return interrupts. This is managed by a count. If already
1509 * on, just increment the count.
1511 void sc_add_credit_return_intr(struct send_context *sc)
1513 unsigned long flags;
1515 /* lock must surround both the count change and the CSR update */
1516 spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1517 if (sc->credit_intr_count == 0) {
1518 sc->credit_ctrl |= SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1519 write_kctxt_csr(sc->dd, sc->hw_context,
1520 SC(CREDIT_CTRL), sc->credit_ctrl);
1522 sc->credit_intr_count++;
1523 spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1527 * Stop credit return interrupts. This is managed by a count. Decrement the
1528 * count, if the last user, then turn the credit interrupts off.
1530 void sc_del_credit_return_intr(struct send_context *sc)
1532 unsigned long flags;
1534 WARN_ON(sc->credit_intr_count == 0);
1536 /* lock must surround both the count change and the CSR update */
1537 spin_lock_irqsave(&sc->credit_ctrl_lock, flags);
1538 sc->credit_intr_count--;
1539 if (sc->credit_intr_count == 0) {
1540 sc->credit_ctrl &= ~SC(CREDIT_CTRL_CREDIT_INTR_SMASK);
1541 write_kctxt_csr(sc->dd, sc->hw_context,
1542 SC(CREDIT_CTRL), sc->credit_ctrl);
1544 spin_unlock_irqrestore(&sc->credit_ctrl_lock, flags);
1548 * The caller must be careful when calling this. All needint calls
1549 * must be paired with !needint.
1551 void hfi1_sc_wantpiobuf_intr(struct send_context *sc, u32 needint)
1554 sc_add_credit_return_intr(sc);
1556 sc_del_credit_return_intr(sc);
1557 trace_hfi1_wantpiointr(sc, needint, sc->credit_ctrl);
1559 sc_return_credits(sc);
1563 * sc_piobufavail - callback when a PIO buffer is available
1564 * @sc: the send context
1566 * This is called from the interrupt handler when a PIO buffer is
1567 * available after hfi1_verbs_send() returned an error that no buffers were
1568 * available. Disable the interrupt if there are no more QPs waiting.
1570 static void sc_piobufavail(struct send_context *sc)
1572 struct hfi1_devdata *dd = sc->dd;
1573 struct list_head *list;
1574 struct rvt_qp *qps[PIO_WAIT_BATCH_SIZE];
1576 struct hfi1_qp_priv *priv;
1577 unsigned long flags;
1578 uint i, n = 0, top_idx = 0;
1580 if (dd->send_contexts[sc->sw_index].type != SC_KERNEL &&
1581 dd->send_contexts[sc->sw_index].type != SC_VL15)
1583 list = &sc->piowait;
1585 * Note: checking that the piowait list is empty and clearing
1586 * the buffer available interrupt needs to be atomic or we
1587 * could end up with QPs on the wait list with the interrupt
1590 write_seqlock_irqsave(&sc->waitlock, flags);
1591 while (!list_empty(list)) {
1592 struct iowait *wait;
1594 if (n == ARRAY_SIZE(qps))
1596 wait = list_first_entry(list, struct iowait, list);
1597 iowait_get_priority(wait);
1598 qp = iowait_to_qp(wait);
1600 list_del_init(&priv->s_iowait.list);
1601 priv->s_iowait.lock = NULL;
1603 priv = qps[top_idx]->priv;
1604 top_idx = iowait_priority_update_top(wait,
1609 /* refcount held until actual wake up */
1613 * If there had been waiters and there are more
1614 * insure that we redo the force to avoid a potential hang.
1617 hfi1_sc_wantpiobuf_intr(sc, 0);
1618 if (!list_empty(list))
1619 hfi1_sc_wantpiobuf_intr(sc, 1);
1621 write_sequnlock_irqrestore(&sc->waitlock, flags);
1623 /* Wake up the top-priority one first */
1625 hfi1_qp_wakeup(qps[top_idx],
1626 RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
1627 for (i = 0; i < n; i++)
1629 hfi1_qp_wakeup(qps[i],
1630 RVT_S_WAIT_PIO | HFI1_S_WAIT_PIO_DRAIN);
1633 /* translate a send credit update to a bit code of reasons */
1634 static inline int fill_code(u64 hw_free)
1638 if (hw_free & CR_STATUS_SMASK)
1639 code |= PRC_STATUS_ERR;
1640 if (hw_free & CR_CREDIT_RETURN_DUE_TO_PBC_SMASK)
1642 if (hw_free & CR_CREDIT_RETURN_DUE_TO_THRESHOLD_SMASK)
1643 code |= PRC_THRESHOLD;
1644 if (hw_free & CR_CREDIT_RETURN_DUE_TO_ERR_SMASK)
1645 code |= PRC_FILL_ERR;
1646 if (hw_free & CR_CREDIT_RETURN_DUE_TO_FORCE_SMASK)
1647 code |= PRC_SC_DISABLE;
1651 /* use the jiffies compare to get the wrap right */
1652 #define sent_before(a, b) time_before(a, b) /* a < b */
1655 * The send context buffer "releaser".
1657 void sc_release_update(struct send_context *sc)
1659 struct pio_buf *pbuf;
1662 unsigned long old_free;
1664 unsigned long extra;
1665 unsigned long flags;
1671 spin_lock_irqsave(&sc->release_lock, flags);
1673 hw_free = le64_to_cpu(*sc->hw_free); /* volatile read */
1674 old_free = sc->free;
1675 extra = (((hw_free & CR_COUNTER_SMASK) >> CR_COUNTER_SHIFT)
1676 - (old_free & CR_COUNTER_MASK))
1678 free = old_free + extra;
1679 trace_hfi1_piofree(sc, extra);
1681 /* call sent buffer callbacks */
1682 code = -1; /* code not yet set */
1683 head = READ_ONCE(sc->sr_head); /* snapshot the head */
1685 while (head != tail) {
1686 pbuf = &sc->sr[tail].pbuf;
1688 if (sent_before(free, pbuf->sent_at)) {
1693 if (code < 0) /* fill in code on first user */
1694 code = fill_code(hw_free);
1695 (*pbuf->cb)(pbuf->arg, code);
1699 if (tail >= sc->sr_size)
1703 /* make sure tail is updated before free */
1706 spin_unlock_irqrestore(&sc->release_lock, flags);
1711 * Send context group releaser. Argument is the send context that caused
1712 * the interrupt. Called from the send context interrupt handler.
1714 * Call release on all contexts in the group.
1716 * This routine takes the sc_lock without an irqsave because it is only
1717 * called from an interrupt handler. Adjust if that changes.
1719 void sc_group_release_update(struct hfi1_devdata *dd, u32 hw_context)
1721 struct send_context *sc;
1725 spin_lock(&dd->sc_lock);
1726 sw_index = dd->hw_to_sw[hw_context];
1727 if (unlikely(sw_index >= dd->num_send_contexts)) {
1728 dd_dev_err(dd, "%s: invalid hw (%u) to sw (%u) mapping\n",
1729 __func__, hw_context, sw_index);
1732 sc = dd->send_contexts[sw_index].sc;
1736 gc = group_context(hw_context, sc->group);
1737 gc_end = gc + group_size(sc->group);
1738 for (; gc < gc_end; gc++) {
1739 sw_index = dd->hw_to_sw[gc];
1740 if (unlikely(sw_index >= dd->num_send_contexts)) {
1742 "%s: invalid hw (%u) to sw (%u) mapping\n",
1743 __func__, hw_context, sw_index);
1746 sc_release_update(dd->send_contexts[sw_index].sc);
1749 spin_unlock(&dd->sc_lock);
1753 * pio_select_send_context_vl() - select send context
1755 * @selector: a spreading factor
1758 * This function returns a send context based on the selector and a vl.
1759 * The mapping fields are protected by RCU
1761 struct send_context *pio_select_send_context_vl(struct hfi1_devdata *dd,
1762 u32 selector, u8 vl)
1764 struct pio_vl_map *m;
1765 struct pio_map_elem *e;
1766 struct send_context *rval;
1769 * NOTE This should only happen if SC->VL changed after the initial
1770 * checks on the QP/AH
1771 * Default will return VL0's send context below
1773 if (unlikely(vl >= num_vls)) {
1779 m = rcu_dereference(dd->pio_map);
1782 return dd->vld[0].sc;
1784 e = m->map[vl & m->mask];
1785 rval = e->ksc[selector & e->mask];
1789 rval = !rval ? dd->vld[0].sc : rval;
1794 * pio_select_send_context_sc() - select send context
1796 * @selector: a spreading factor
1797 * @sc5: the 5 bit sc
1799 * This function returns an send context based on the selector and an sc
1801 struct send_context *pio_select_send_context_sc(struct hfi1_devdata *dd,
1802 u32 selector, u8 sc5)
1804 u8 vl = sc_to_vlt(dd, sc5);
1806 return pio_select_send_context_vl(dd, selector, vl);
1810 * Free the indicated map struct
1812 static void pio_map_free(struct pio_vl_map *m)
1816 for (i = 0; m && i < m->actual_vls; i++)
1822 * Handle RCU callback
1824 static void pio_map_rcu_callback(struct rcu_head *list)
1826 struct pio_vl_map *m = container_of(list, struct pio_vl_map, list);
1832 * Set credit return threshold for the kernel send context
1834 static void set_threshold(struct hfi1_devdata *dd, int scontext, int i)
1838 thres = min(sc_percent_to_threshold(dd->kernel_send_context[scontext],
1840 sc_mtu_to_threshold(dd->kernel_send_context[scontext],
1842 dd->rcd[0]->rcvhdrqentsize));
1843 sc_set_cr_threshold(dd->kernel_send_context[scontext], thres);
1847 * pio_map_init - called when #vls change
1849 * @port: port number
1850 * @num_vls: number of vls
1851 * @vl_scontexts: per vl send context mapping (optional)
1853 * This routine changes the mapping based on the number of vls.
1855 * vl_scontexts is used to specify a non-uniform vl/send context
1856 * loading. NULL implies auto computing the loading and giving each
1857 * VL an uniform distribution of send contexts per VL.
1859 * The auto algorithm computers the sc_per_vl and the number of extra
1860 * send contexts. Any extra send contexts are added from the last VL
1863 * rcu locking is used here to control access to the mapping fields.
1865 * If either the num_vls or num_send_contexts are non-power of 2, the
1866 * array sizes in the struct pio_vl_map and the struct pio_map_elem are
1867 * rounded up to the next highest power of 2 and the first entry is
1868 * reused in a round robin fashion.
1870 * If an error occurs the map change is not done and the mapping is not
1874 int pio_map_init(struct hfi1_devdata *dd, u8 port, u8 num_vls, u8 *vl_scontexts)
1877 int extra, sc_per_vl;
1879 int num_kernel_send_contexts = 0;
1880 u8 lvl_scontexts[OPA_MAX_VLS];
1881 struct pio_vl_map *oldmap, *newmap;
1883 if (!vl_scontexts) {
1884 for (i = 0; i < dd->num_send_contexts; i++)
1885 if (dd->send_contexts[i].type == SC_KERNEL)
1886 num_kernel_send_contexts++;
1887 /* truncate divide */
1888 sc_per_vl = num_kernel_send_contexts / num_vls;
1890 extra = num_kernel_send_contexts % num_vls;
1891 vl_scontexts = lvl_scontexts;
1892 /* add extras from last vl down */
1893 for (i = num_vls - 1; i >= 0; i--, extra--)
1894 vl_scontexts[i] = sc_per_vl + (extra > 0 ? 1 : 0);
1897 newmap = kzalloc(sizeof(*newmap) +
1898 roundup_pow_of_two(num_vls) *
1899 sizeof(struct pio_map_elem *),
1903 newmap->actual_vls = num_vls;
1904 newmap->vls = roundup_pow_of_two(num_vls);
1905 newmap->mask = (1 << ilog2(newmap->vls)) - 1;
1906 for (i = 0; i < newmap->vls; i++) {
1907 /* save for wrap around */
1908 int first_scontext = scontext;
1910 if (i < newmap->actual_vls) {
1911 int sz = roundup_pow_of_two(vl_scontexts[i]);
1913 /* only allocate once */
1914 newmap->map[i] = kzalloc(sizeof(*newmap->map[i]) +
1918 if (!newmap->map[i])
1920 newmap->map[i]->mask = (1 << ilog2(sz)) - 1;
1922 * assign send contexts and
1923 * adjust credit return threshold
1925 for (j = 0; j < sz; j++) {
1926 if (dd->kernel_send_context[scontext]) {
1927 newmap->map[i]->ksc[j] =
1928 dd->kernel_send_context[scontext];
1929 set_threshold(dd, scontext, i);
1931 if (++scontext >= first_scontext +
1933 /* wrap back to first send context */
1934 scontext = first_scontext;
1937 /* just re-use entry without allocating */
1938 newmap->map[i] = newmap->map[i % num_vls];
1940 scontext = first_scontext + vl_scontexts[i];
1942 /* newmap in hand, save old map */
1943 spin_lock_irq(&dd->pio_map_lock);
1944 oldmap = rcu_dereference_protected(dd->pio_map,
1945 lockdep_is_held(&dd->pio_map_lock));
1947 /* publish newmap */
1948 rcu_assign_pointer(dd->pio_map, newmap);
1950 spin_unlock_irq(&dd->pio_map_lock);
1951 /* success, free any old map after grace period */
1953 call_rcu(&oldmap->list, pio_map_rcu_callback);
1956 /* free any partial allocation */
1957 pio_map_free(newmap);
1961 void free_pio_map(struct hfi1_devdata *dd)
1963 /* Free PIO map if allocated */
1964 if (rcu_access_pointer(dd->pio_map)) {
1965 spin_lock_irq(&dd->pio_map_lock);
1966 pio_map_free(rcu_access_pointer(dd->pio_map));
1967 RCU_INIT_POINTER(dd->pio_map, NULL);
1968 spin_unlock_irq(&dd->pio_map_lock);
1971 kfree(dd->kernel_send_context);
1972 dd->kernel_send_context = NULL;
1975 int init_pervl_scs(struct hfi1_devdata *dd)
1978 u64 mask, all_vl_mask = (u64)0x80ff; /* VLs 0-7, 15 */
1979 u64 data_vls_mask = (u64)0x00ff; /* VLs 0-7 */
1981 struct hfi1_pportdata *ppd = dd->pport;
1983 dd->vld[15].sc = sc_alloc(dd, SC_VL15,
1984 dd->rcd[0]->rcvhdrqentsize, dd->node);
1985 if (!dd->vld[15].sc)
1988 hfi1_init_ctxt(dd->vld[15].sc);
1989 dd->vld[15].mtu = enum_to_mtu(OPA_MTU_2048);
1991 dd->kernel_send_context = kcalloc_node(dd->num_send_contexts,
1992 sizeof(struct send_context *),
1993 GFP_KERNEL, dd->node);
1994 if (!dd->kernel_send_context)
1997 dd->kernel_send_context[0] = dd->vld[15].sc;
1999 for (i = 0; i < num_vls; i++) {
2001 * Since this function does not deal with a specific
2002 * receive context but we need the RcvHdrQ entry size,
2003 * use the size from rcd[0]. It is guaranteed to be
2004 * valid at this point and will remain the same for all
2007 dd->vld[i].sc = sc_alloc(dd, SC_KERNEL,
2008 dd->rcd[0]->rcvhdrqentsize, dd->node);
2011 dd->kernel_send_context[i + 1] = dd->vld[i].sc;
2012 hfi1_init_ctxt(dd->vld[i].sc);
2013 /* non VL15 start with the max MTU */
2014 dd->vld[i].mtu = hfi1_max_mtu;
2016 for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
2017 dd->kernel_send_context[i + 1] =
2018 sc_alloc(dd, SC_KERNEL, dd->rcd[0]->rcvhdrqentsize, dd->node);
2019 if (!dd->kernel_send_context[i + 1])
2021 hfi1_init_ctxt(dd->kernel_send_context[i + 1]);
2024 sc_enable(dd->vld[15].sc);
2025 ctxt = dd->vld[15].sc->hw_context;
2026 mask = all_vl_mask & ~(1LL << 15);
2027 write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2029 "Using send context %u(%u) for VL15\n",
2030 dd->vld[15].sc->sw_index, ctxt);
2032 for (i = 0; i < num_vls; i++) {
2033 sc_enable(dd->vld[i].sc);
2034 ctxt = dd->vld[i].sc->hw_context;
2035 mask = all_vl_mask & ~(data_vls_mask);
2036 write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2038 for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++) {
2039 sc_enable(dd->kernel_send_context[i + 1]);
2040 ctxt = dd->kernel_send_context[i + 1]->hw_context;
2041 mask = all_vl_mask & ~(data_vls_mask);
2042 write_kctxt_csr(dd, ctxt, SC(CHECK_VL), mask);
2045 if (pio_map_init(dd, ppd->port - 1, num_vls, NULL))
2050 for (i = 0; i < num_vls; i++) {
2051 sc_free(dd->vld[i].sc);
2052 dd->vld[i].sc = NULL;
2055 for (i = num_vls; i < INIT_SC_PER_VL * num_vls; i++)
2056 sc_free(dd->kernel_send_context[i + 1]);
2058 kfree(dd->kernel_send_context);
2059 dd->kernel_send_context = NULL;
2062 sc_free(dd->vld[15].sc);
2066 int init_credit_return(struct hfi1_devdata *dd)
2071 dd->cr_base = kcalloc(
2072 node_affinity.num_possible_nodes,
2073 sizeof(struct credit_return_base),
2079 for_each_node_with_cpus(i) {
2080 int bytes = TXE_NUM_CONTEXTS * sizeof(struct credit_return);
2082 set_dev_node(&dd->pcidev->dev, i);
2083 dd->cr_base[i].va = dma_alloc_coherent(&dd->pcidev->dev,
2085 &dd->cr_base[i].dma,
2087 if (!dd->cr_base[i].va) {
2088 set_dev_node(&dd->pcidev->dev, dd->node);
2090 "Unable to allocate credit return DMA range for NUMA %d\n",
2096 set_dev_node(&dd->pcidev->dev, dd->node);
2103 void free_credit_return(struct hfi1_devdata *dd)
2109 for (i = 0; i < node_affinity.num_possible_nodes; i++) {
2110 if (dd->cr_base[i].va) {
2111 dma_free_coherent(&dd->pcidev->dev,
2113 sizeof(struct credit_return),
2115 dd->cr_base[i].dma);
2122 void seqfile_dump_sci(struct seq_file *s, u32 i,
2123 struct send_context_info *sci)
2125 struct send_context *sc = sci->sc;
2128 seq_printf(s, "SCI %u: type %u base %u credits %u\n",
2129 i, sci->type, sci->base, sci->credits);
2130 seq_printf(s, " flags 0x%x sw_inx %u hw_ctxt %u grp %u\n",
2131 sc->flags, sc->sw_index, sc->hw_context, sc->group);
2132 seq_printf(s, " sr_size %u credits %u sr_head %u sr_tail %u\n",
2133 sc->sr_size, sc->credits, sc->sr_head, sc->sr_tail);
2134 seq_printf(s, " fill %lu free %lu fill_wrap %u alloc_free %lu\n",
2135 sc->fill, sc->free, sc->fill_wrap, sc->alloc_free);
2136 seq_printf(s, " credit_intr_count %u credit_ctrl 0x%llx\n",
2137 sc->credit_intr_count, sc->credit_ctrl);
2138 reg = read_kctxt_csr(sc->dd, sc->hw_context, SC(CREDIT_STATUS));
2139 seq_printf(s, " *hw_free %llu CurrentFree %llu LastReturned %llu\n",
2140 (le64_to_cpu(*sc->hw_free) & CR_COUNTER_SMASK) >>
2142 (reg >> SC(CREDIT_STATUS_CURRENT_FREE_COUNTER_SHIFT)) &
2143 SC(CREDIT_STATUS_CURRENT_FREE_COUNTER_MASK),
2144 reg & SC(CREDIT_STATUS_LAST_RETURNED_COUNTER_SMASK));