1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * SGI UltraViolet TLB flush routines.
5 * (c) 2008-2014 Cliff Wickman <cpw@sgi.com>, SGI.
7 #include <linux/seq_file.h>
8 #include <linux/proc_fs.h>
9 #include <linux/debugfs.h>
10 #include <linux/kernel.h>
11 #include <linux/slab.h>
12 #include <linux/delay.h>
14 #include <asm/mmu_context.h>
15 #include <asm/uv/uv.h>
16 #include <asm/uv/uv_mmrs.h>
17 #include <asm/uv/uv_hub.h>
18 #include <asm/uv/uv_bau.h>
21 #include <asm/irq_vectors.h>
22 #include <asm/timer.h>
24 static struct bau_operations ops __ro_after_init;
26 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
27 static const int timeout_base_ns[] = {
38 static int timeout_us;
39 static bool nobau = true;
40 static int nobau_perm;
43 static int max_concurr = MAX_BAU_CONCURRENT;
44 static int max_concurr_const = MAX_BAU_CONCURRENT;
45 static int plugged_delay = PLUGGED_DELAY;
46 static int plugsb4reset = PLUGSB4RESET;
47 static int giveup_limit = GIVEUP_LIMIT;
48 static int timeoutsb4reset = TIMEOUTSB4RESET;
49 static int ipi_reset_limit = IPI_RESET_LIMIT;
50 static int complete_threshold = COMPLETE_THRESHOLD;
51 static int congested_respns_us = CONGESTED_RESPONSE_US;
52 static int congested_reps = CONGESTED_REPS;
53 static int disabled_period = DISABLED_PERIOD;
55 static struct tunables tunables[] = {
56 {&max_concurr, MAX_BAU_CONCURRENT}, /* must be [0] */
57 {&plugged_delay, PLUGGED_DELAY},
58 {&plugsb4reset, PLUGSB4RESET},
59 {&timeoutsb4reset, TIMEOUTSB4RESET},
60 {&ipi_reset_limit, IPI_RESET_LIMIT},
61 {&complete_threshold, COMPLETE_THRESHOLD},
62 {&congested_respns_us, CONGESTED_RESPONSE_US},
63 {&congested_reps, CONGESTED_REPS},
64 {&disabled_period, DISABLED_PERIOD},
65 {&giveup_limit, GIVEUP_LIMIT}
68 static struct dentry *tunables_dir;
69 static struct dentry *tunables_file;
71 /* these correspond to the statistics printed by ptc_seq_show() */
72 static char *stat_description[] = {
73 "sent: number of shootdown messages sent",
74 "stime: time spent sending messages",
75 "numuvhubs: number of hubs targeted with shootdown",
76 "numuvhubs16: number times 16 or more hubs targeted",
77 "numuvhubs8: number times 8 or more hubs targeted",
78 "numuvhubs4: number times 4 or more hubs targeted",
79 "numuvhubs2: number times 2 or more hubs targeted",
80 "numuvhubs1: number times 1 hub targeted",
81 "numcpus: number of cpus targeted with shootdown",
82 "dto: number of destination timeouts",
83 "retries: destination timeout retries sent",
84 "rok: : destination timeouts successfully retried",
85 "resetp: ipi-style resource resets for plugs",
86 "resett: ipi-style resource resets for timeouts",
87 "giveup: fall-backs to ipi-style shootdowns",
88 "sto: number of source timeouts",
89 "bz: number of stay-busy's",
90 "throt: number times spun in throttle",
91 "swack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE",
92 "recv: shootdown messages received",
93 "rtime: time spent processing messages",
94 "all: shootdown all-tlb messages",
95 "one: shootdown one-tlb messages",
96 "mult: interrupts that found multiple messages",
97 "none: interrupts that found no messages",
98 "retry: number of retry messages processed",
99 "canc: number messages canceled by retries",
100 "nocan: number retries that found nothing to cancel",
101 "reset: number of ipi-style reset requests processed",
102 "rcan: number messages canceled by reset requests",
103 "disable: number times use of the BAU was disabled",
104 "enable: number times use of the BAU was re-enabled"
107 static int __init setup_bau(char *arg)
114 result = strtobool(arg, &nobau);
118 /* we need to flip the logic here, so that bau=y sets nobau to false */
122 pr_info("UV BAU Enabled\n");
124 pr_info("UV BAU Disabled\n");
128 early_param("bau", setup_bau);
130 /* base pnode in this partition */
131 static int uv_base_pnode __read_mostly;
133 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
134 static DEFINE_PER_CPU(struct bau_control, bau_control);
135 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
141 struct bau_control *bcp;
144 pr_info("BAU not initialized; cannot be turned on\n");
148 for_each_present_cpu(cpu) {
149 bcp = &per_cpu(bau_control, cpu);
152 pr_info("BAU turned on\n");
160 struct bau_control *bcp;
163 for_each_present_cpu(cpu) {
164 bcp = &per_cpu(bau_control, cpu);
167 pr_info("BAU turned off\n");
172 * Determine the first node on a uvhub. 'Nodes' are used for kernel
175 static int __init uvhub_to_first_node(int uvhub)
179 for_each_online_node(node) {
180 b = uv_node_to_blade_id(node);
188 * Determine the apicid of the first cpu on a uvhub.
190 static int __init uvhub_to_first_apicid(int uvhub)
194 for_each_present_cpu(cpu)
195 if (uvhub == uv_cpu_to_blade_id(cpu))
196 return per_cpu(x86_cpu_to_apicid, cpu);
201 * Free a software acknowledge hardware resource by clearing its Pending
202 * bit. This will return a reply to the sender.
203 * If the message has timed out, a reply has already been sent by the
204 * hardware but the resource has not been released. In that case our
205 * clear of the Timeout bit (as well) will free the resource. No reply will
206 * be sent (the hardware will only do one reply per message).
208 static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp,
212 struct bau_pq_entry *msg;
215 if (!msg->canceled && do_acknowledge) {
216 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
217 ops.write_l_sw_ack(dw);
224 * Process the receipt of a RETRY message
226 static void bau_process_retry_msg(struct msg_desc *mdp,
227 struct bau_control *bcp)
230 int cancel_count = 0;
231 unsigned long msg_res;
232 unsigned long mmr = 0;
233 struct bau_pq_entry *msg = mdp->msg;
234 struct bau_pq_entry *msg2;
235 struct ptc_stats *stat = bcp->statp;
239 * cancel any message from msg+1 to the retry itself
241 for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
242 if (msg2 > mdp->queue_last)
243 msg2 = mdp->queue_first;
247 /* same conditions for cancellation as do_reset */
248 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
249 (msg2->swack_vec) && ((msg2->swack_vec &
250 msg->swack_vec) == 0) &&
251 (msg2->sending_cpu == msg->sending_cpu) &&
252 (msg2->msg_type != MSG_NOOP)) {
253 mmr = ops.read_l_sw_ack();
254 msg_res = msg2->swack_vec;
256 * This is a message retry; clear the resources held
257 * by the previous message only if they timed out.
258 * If it has not timed out we have an unexpected
259 * situation to report.
261 if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
264 * Is the resource timed out?
265 * Make everyone ignore the cancelled message.
270 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
271 ops.write_l_sw_ack(mr);
276 stat->d_nocanceled++;
280 * Do all the things a cpu should do for a TLB shootdown message.
281 * Other cpu's may come here at the same time for this message.
283 static void bau_process_message(struct msg_desc *mdp, struct bau_control *bcp,
286 short socket_ack_count = 0;
288 struct atomic_short *asp;
289 struct ptc_stats *stat = bcp->statp;
290 struct bau_pq_entry *msg = mdp->msg;
291 struct bau_control *smaster = bcp->socket_master;
294 * This must be a normal message, or retry of a normal message
296 if (msg->address == TLB_FLUSH_ALL) {
300 __flush_tlb_one_user(msg->address);
306 * One cpu on each uvhub has the additional job on a RETRY
307 * of releasing the resource held by the message that is
308 * being retried. That message is identified by sending
311 if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
312 bau_process_retry_msg(mdp, bcp);
315 * This is a swack message, so we have to reply to it.
316 * Count each responding cpu on the socket. This avoids
317 * pinging the count's cache line back and forth between
320 sp = &smaster->socket_acknowledge_count[mdp->msg_slot];
321 asp = (struct atomic_short *)sp;
322 socket_ack_count = atom_asr(1, asp);
323 if (socket_ack_count == bcp->cpus_in_socket) {
326 * Both sockets dump their completed count total into
327 * the message's count.
330 asp = (struct atomic_short *)&msg->acknowledge_count;
331 msg_ack_count = atom_asr(socket_ack_count, asp);
333 if (msg_ack_count == bcp->cpus_in_uvhub) {
335 * All cpus in uvhub saw it; reply
336 * (unless we are in the UV2 workaround)
338 reply_to_message(mdp, bcp, do_acknowledge);
346 * Determine the first cpu on a pnode.
348 static int pnode_to_first_cpu(int pnode, struct bau_control *smaster)
351 struct hub_and_pnode *hpp;
353 for_each_present_cpu(cpu) {
354 hpp = &smaster->thp[cpu];
355 if (pnode == hpp->pnode)
362 * Last resort when we get a large number of destination timeouts is
363 * to clear resources held by a given cpu.
364 * Do this with IPI so that all messages in the BAU message queue
365 * can be identified by their nonzero swack_vec field.
367 * This is entered for a single cpu on the uvhub.
368 * The sender want's this uvhub to free a specific message's
371 static void do_reset(void *ptr)
374 struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id());
375 struct reset_args *rap = (struct reset_args *)ptr;
376 struct bau_pq_entry *msg;
377 struct ptc_stats *stat = bcp->statp;
381 * We're looking for the given sender, and
382 * will free its swack resource.
383 * If all cpu's finally responded after the timeout, its
384 * message 'replied_to' was set.
386 for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
387 unsigned long msg_res;
388 /* do_reset: same conditions for cancellation as
389 bau_process_retry_msg() */
390 if ((msg->replied_to == 0) &&
391 (msg->canceled == 0) &&
392 (msg->sending_cpu == rap->sender) &&
394 (msg->msg_type != MSG_NOOP)) {
398 * make everyone else ignore this message
402 * only reset the resource if it is still pending
404 mmr = ops.read_l_sw_ack();
405 msg_res = msg->swack_vec;
406 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
409 ops.write_l_sw_ack(mr);
417 * Use IPI to get all target uvhubs to release resources held by
418 * a given sending cpu number.
420 static void reset_with_ipi(struct pnmask *distribution, struct bau_control *bcp)
425 int sender = bcp->cpu;
426 cpumask_t *mask = bcp->uvhub_master->cpumask;
427 struct bau_control *smaster = bcp->socket_master;
428 struct reset_args reset_args;
430 reset_args.sender = sender;
432 /* find a single cpu for each uvhub in this distribution mask */
433 maskbits = sizeof(struct pnmask) * BITSPERBYTE;
434 /* each bit is a pnode relative to the partition base pnode */
435 for (pnode = 0; pnode < maskbits; pnode++) {
437 if (!bau_uvhub_isset(pnode, distribution))
439 apnode = pnode + bcp->partition_base_pnode;
440 cpu = pnode_to_first_cpu(apnode, smaster);
441 cpumask_set_cpu(cpu, mask);
444 /* IPI all cpus; preemption is already disabled */
445 smp_call_function_many(mask, do_reset, (void *)&reset_args, 1);
450 * Not to be confused with cycles_2_ns() from tsc.c; this gives a relative
451 * number, not an absolute. It converts a duration in cycles to a duration in
454 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
456 struct cyc2ns_data data;
457 unsigned long long ns;
459 cyc2ns_read_begin(&data);
460 ns = mul_u64_u32_shr(cyc, data.cyc2ns_mul, data.cyc2ns_shift);
467 * The reverse of the above; converts a duration in ns to a duration in cycles.
469 static inline unsigned long long ns_2_cycles(unsigned long long ns)
471 struct cyc2ns_data data;
472 unsigned long long cyc;
474 cyc2ns_read_begin(&data);
475 cyc = (ns << data.cyc2ns_shift) / data.cyc2ns_mul;
481 static inline unsigned long cycles_2_us(unsigned long long cyc)
483 return cycles_2_ns(cyc) / NSEC_PER_USEC;
486 static inline cycles_t sec_2_cycles(unsigned long sec)
488 return ns_2_cycles(sec * NSEC_PER_SEC);
491 static inline unsigned long long usec_2_cycles(unsigned long usec)
493 return ns_2_cycles(usec * NSEC_PER_USEC);
497 * wait for all cpus on this hub to finish their sends and go quiet
498 * leaves uvhub_quiesce set so that no new broadcasts are started by
499 * bau_flush_send_and_wait()
501 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
503 atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
507 * mark this quiet-requestor as done
509 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
511 atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
514 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
516 unsigned long descriptor_status;
518 descriptor_status = uv_read_local_mmr(mmr_offset);
519 descriptor_status >>= right_shift;
520 descriptor_status &= UV_ACT_STATUS_MASK;
521 return descriptor_status;
525 * Wait for completion of a broadcast software ack message
526 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
528 static int uv1_wait_completion(struct bau_desc *bau_desc,
529 struct bau_control *bcp, long try)
531 unsigned long descriptor_status;
533 u64 mmr_offset = bcp->status_mmr;
534 int right_shift = bcp->status_index;
535 struct ptc_stats *stat = bcp->statp;
537 descriptor_status = uv1_read_status(mmr_offset, right_shift);
538 /* spin on the status MMR, waiting for it to go idle */
539 while ((descriptor_status != DS_IDLE)) {
541 * Our software ack messages may be blocked because
542 * there are no swack resources available. As long
543 * as none of them has timed out hardware will NACK
544 * our message and its state will stay IDLE.
546 if (descriptor_status == DS_SOURCE_TIMEOUT) {
549 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
554 * Our retries may be blocked by all destination
555 * swack resources being consumed, and a timeout
556 * pending. In that case hardware returns the
557 * ERROR that looks like a destination timeout.
559 if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
560 bcp->conseccompletes = 0;
561 return FLUSH_RETRY_PLUGGED;
564 bcp->conseccompletes = 0;
565 return FLUSH_RETRY_TIMEOUT;
568 * descriptor_status is still BUSY
572 descriptor_status = uv1_read_status(mmr_offset, right_shift);
574 bcp->conseccompletes++;
575 return FLUSH_COMPLETE;
579 * UV2 could have an extra bit of status in the ACTIVATION_STATUS_2 register.
580 * But not currently used.
582 static unsigned long uv2_3_read_status(unsigned long offset, int rshft, int desc)
584 return ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK) << 1;
588 * Entered when a bau descriptor has gone into a permanent busy wait because
590 * Workaround the bug.
592 static int handle_uv2_busy(struct bau_control *bcp)
594 struct ptc_stats *stat = bcp->statp;
601 static int uv2_3_wait_completion(struct bau_desc *bau_desc,
602 struct bau_control *bcp, long try)
604 unsigned long descriptor_stat;
606 u64 mmr_offset = bcp->status_mmr;
607 int right_shift = bcp->status_index;
608 int desc = bcp->uvhub_cpu;
610 struct ptc_stats *stat = bcp->statp;
612 descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
614 /* spin on the status MMR, waiting for it to go idle */
615 while (descriptor_stat != UV2H_DESC_IDLE) {
616 if (descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT) {
618 * A h/w bug on the destination side may
619 * have prevented the message being marked
620 * pending, thus it doesn't get replied to
621 * and gets continually nacked until it times
622 * out with a SOURCE_TIMEOUT.
626 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
630 * Our retries may be blocked by all destination
631 * swack resources being consumed, and a timeout
632 * pending. In that case hardware returns the
633 * ERROR that looks like a destination timeout.
634 * Without using the extended status we have to
635 * deduce from the short time that this was a
638 if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
639 bcp->conseccompletes = 0;
641 /* FLUSH_RETRY_PLUGGED causes hang on boot */
645 bcp->conseccompletes = 0;
646 /* FLUSH_RETRY_TIMEOUT causes hang on boot */
650 if (busy_reps > 1000000) {
651 /* not to hammer on the clock */
654 if ((ttm - bcp->send_message) > bcp->timeout_interval)
655 return handle_uv2_busy(bcp);
658 * descriptor_stat is still BUSY
662 descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
664 bcp->conseccompletes++;
665 return FLUSH_COMPLETE;
669 * Returns the status of current BAU message for cpu desc as a bit field
672 static u64 read_status(u64 status_mmr, int index, int desc)
676 stat = ((read_lmmr(status_mmr) >> index) & UV_ACT_STATUS_MASK) << 1;
677 stat |= (read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_2) >> desc) & 0x1;
682 static int uv4_wait_completion(struct bau_desc *bau_desc,
683 struct bau_control *bcp, long try)
685 struct ptc_stats *stat = bcp->statp;
687 u64 mmr = bcp->status_mmr;
688 int index = bcp->status_index;
689 int desc = bcp->uvhub_cpu;
691 descriptor_stat = read_status(mmr, index, desc);
693 /* spin on the status MMR, waiting for it to go idle */
694 while (descriptor_stat != UV2H_DESC_IDLE) {
695 switch (descriptor_stat) {
696 case UV2H_DESC_SOURCE_TIMEOUT:
700 case UV2H_DESC_DEST_TIMEOUT:
702 bcp->conseccompletes = 0;
703 return FLUSH_RETRY_TIMEOUT;
705 case UV2H_DESC_DEST_STRONG_NACK:
707 bcp->conseccompletes = 0;
708 return FLUSH_RETRY_PLUGGED;
710 case UV2H_DESC_DEST_PUT_ERR:
711 bcp->conseccompletes = 0;
715 /* descriptor_stat is still BUSY */
718 descriptor_stat = read_status(mmr, index, desc);
720 bcp->conseccompletes++;
721 return FLUSH_COMPLETE;
725 * Our retries are blocked by all destination sw ack resources being
726 * in use, and a timeout is pending. In that case hardware immediately
727 * returns the ERROR that looks like a destination timeout.
729 static void destination_plugged(struct bau_desc *bau_desc,
730 struct bau_control *bcp,
731 struct bau_control *hmaster, struct ptc_stats *stat)
733 udelay(bcp->plugged_delay);
734 bcp->plugged_tries++;
736 if (bcp->plugged_tries >= bcp->plugsb4reset) {
737 bcp->plugged_tries = 0;
739 quiesce_local_uvhub(hmaster);
741 spin_lock(&hmaster->queue_lock);
742 reset_with_ipi(&bau_desc->distribution, bcp);
743 spin_unlock(&hmaster->queue_lock);
745 end_uvhub_quiesce(hmaster);
748 stat->s_resets_plug++;
752 static void destination_timeout(struct bau_desc *bau_desc,
753 struct bau_control *bcp, struct bau_control *hmaster,
754 struct ptc_stats *stat)
756 hmaster->max_concurr = 1;
757 bcp->timeout_tries++;
758 if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
759 bcp->timeout_tries = 0;
761 quiesce_local_uvhub(hmaster);
763 spin_lock(&hmaster->queue_lock);
764 reset_with_ipi(&bau_desc->distribution, bcp);
765 spin_unlock(&hmaster->queue_lock);
767 end_uvhub_quiesce(hmaster);
770 stat->s_resets_timeout++;
775 * Stop all cpus on a uvhub from using the BAU for a period of time.
776 * This is reversed by check_enable.
778 static void disable_for_period(struct bau_control *bcp, struct ptc_stats *stat)
781 struct bau_control *tbcp;
782 struct bau_control *hmaster;
785 hmaster = bcp->uvhub_master;
786 spin_lock(&hmaster->disable_lock);
787 if (!bcp->baudisabled) {
788 stat->s_bau_disabled++;
790 for_each_present_cpu(tcpu) {
791 tbcp = &per_cpu(bau_control, tcpu);
792 if (tbcp->uvhub_master == hmaster) {
793 tbcp->baudisabled = 1;
794 tbcp->set_bau_on_time =
795 tm1 + bcp->disabled_period;
799 spin_unlock(&hmaster->disable_lock);
802 static void count_max_concurr(int stat, struct bau_control *bcp,
803 struct bau_control *hmaster)
805 bcp->plugged_tries = 0;
806 bcp->timeout_tries = 0;
807 if (stat != FLUSH_COMPLETE)
809 if (bcp->conseccompletes <= bcp->complete_threshold)
811 if (hmaster->max_concurr >= hmaster->max_concurr_const)
813 hmaster->max_concurr++;
816 static void record_send_stats(cycles_t time1, cycles_t time2,
817 struct bau_control *bcp, struct ptc_stats *stat,
818 int completion_status, int try)
823 elapsed = time2 - time1;
824 stat->s_time += elapsed;
826 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
827 bcp->period_requests++;
828 bcp->period_time += elapsed;
829 if ((elapsed > usec_2_cycles(bcp->cong_response_us)) &&
830 (bcp->period_requests > bcp->cong_reps) &&
831 ((bcp->period_time / bcp->period_requests) >
832 usec_2_cycles(bcp->cong_response_us))) {
834 disable_for_period(bcp, stat);
840 if (completion_status == FLUSH_COMPLETE && try > 1)
842 else if (completion_status == FLUSH_GIVEUP) {
844 if (get_cycles() > bcp->period_end)
845 bcp->period_giveups = 0;
846 bcp->period_giveups++;
847 if (bcp->period_giveups == 1)
848 bcp->period_end = get_cycles() + bcp->disabled_period;
849 if (bcp->period_giveups > bcp->giveup_limit) {
850 disable_for_period(bcp, stat);
851 stat->s_giveuplimit++;
857 * Because of a uv1 hardware bug only a limited number of concurrent
858 * requests can be made.
860 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
862 spinlock_t *lock = &hmaster->uvhub_lock;
865 v = &hmaster->active_descriptor_count;
866 if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
870 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
875 * Handle the completion status of a message send.
877 static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
878 struct bau_control *bcp, struct bau_control *hmaster,
879 struct ptc_stats *stat)
881 if (completion_status == FLUSH_RETRY_PLUGGED)
882 destination_plugged(bau_desc, bcp, hmaster, stat);
883 else if (completion_status == FLUSH_RETRY_TIMEOUT)
884 destination_timeout(bau_desc, bcp, hmaster, stat);
888 * Send a broadcast and wait for it to complete.
890 * The flush_mask contains the cpus the broadcast is to be sent to including
891 * cpus that are on the local uvhub.
893 * Returns 0 if all flushing represented in the mask was done.
894 * Returns 1 if it gives up entirely and the original cpu mask is to be
895 * returned to the kernel.
897 static int uv_flush_send_and_wait(struct cpumask *flush_mask,
898 struct bau_control *bcp,
899 struct bau_desc *bau_desc)
902 int completion_stat = 0;
908 struct ptc_stats *stat = bcp->statp;
909 struct bau_control *hmaster = bcp->uvhub_master;
910 struct uv1_bau_msg_header *uv1_hdr = NULL;
911 struct uv2_3_bau_msg_header *uv2_3_hdr = NULL;
913 if (bcp->uvhub_version == UV_BAU_V1) {
915 uv1_throttle(hmaster, stat);
918 while (hmaster->uvhub_quiesce)
921 time1 = get_cycles();
923 uv1_hdr = &bau_desc->header.uv1_hdr;
926 uv2_3_hdr = &bau_desc->header.uv2_3_hdr;
931 uv1_hdr->msg_type = MSG_REGULAR;
933 uv2_3_hdr->msg_type = MSG_REGULAR;
934 seq_number = bcp->message_number++;
937 uv1_hdr->msg_type = MSG_RETRY;
939 uv2_3_hdr->msg_type = MSG_RETRY;
940 stat->s_retry_messages++;
944 uv1_hdr->sequence = seq_number;
946 uv2_3_hdr->sequence = seq_number;
947 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
948 bcp->send_message = get_cycles();
950 write_mmr_activation(index);
953 completion_stat = ops.wait_completion(bau_desc, bcp, try);
955 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
957 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
958 bcp->ipi_attempts = 0;
959 stat->s_overipilimit++;
960 completion_stat = FLUSH_GIVEUP;
964 } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
965 (completion_stat == FLUSH_RETRY_TIMEOUT));
967 time2 = get_cycles();
969 count_max_concurr(completion_stat, bcp, hmaster);
971 while (hmaster->uvhub_quiesce)
974 atomic_dec(&hmaster->active_descriptor_count);
976 record_send_stats(time1, time2, bcp, stat, completion_stat, try);
978 if (completion_stat == FLUSH_GIVEUP)
979 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
985 * The BAU is disabled for this uvhub. When the disabled time period has
986 * expired re-enable it.
987 * Return 0 if it is re-enabled for all cpus on this uvhub.
989 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
992 struct bau_control *tbcp;
993 struct bau_control *hmaster;
995 hmaster = bcp->uvhub_master;
996 spin_lock(&hmaster->disable_lock);
997 if (bcp->baudisabled && (get_cycles() >= bcp->set_bau_on_time)) {
998 stat->s_bau_reenabled++;
999 for_each_present_cpu(tcpu) {
1000 tbcp = &per_cpu(bau_control, tcpu);
1001 if (tbcp->uvhub_master == hmaster) {
1002 tbcp->baudisabled = 0;
1003 tbcp->period_requests = 0;
1004 tbcp->period_time = 0;
1005 tbcp->period_giveups = 0;
1008 spin_unlock(&hmaster->disable_lock);
1011 spin_unlock(&hmaster->disable_lock);
1015 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
1016 int remotes, struct bau_desc *bau_desc)
1018 stat->s_requestor++;
1019 stat->s_ntargcpu += remotes + locals;
1020 stat->s_ntargremotes += remotes;
1021 stat->s_ntarglocals += locals;
1023 /* uvhub statistics */
1024 hubs = bau_uvhub_weight(&bau_desc->distribution);
1026 stat->s_ntarglocaluvhub++;
1027 stat->s_ntargremoteuvhub += (hubs - 1);
1029 stat->s_ntargremoteuvhub += hubs;
1031 stat->s_ntarguvhub += hubs;
1034 stat->s_ntarguvhub16++;
1036 stat->s_ntarguvhub8++;
1038 stat->s_ntarguvhub4++;
1040 stat->s_ntarguvhub2++;
1042 stat->s_ntarguvhub1++;
1046 * Translate a cpu mask to the uvhub distribution mask in the BAU
1047 * activation descriptor.
1049 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
1050 struct bau_desc *bau_desc, int *localsp, int *remotesp)
1055 struct hub_and_pnode *hpp;
1057 for_each_cpu(cpu, flush_mask) {
1059 * The distribution vector is a bit map of pnodes, relative
1060 * to the partition base pnode (and the partition base nasid
1062 * Translate cpu to pnode and hub using a local memory array.
1064 hpp = &bcp->socket_master->thp[cpu];
1065 pnode = hpp->pnode - bcp->partition_base_pnode;
1066 bau_uvhub_set(pnode, &bau_desc->distribution);
1068 if (hpp->uvhub == bcp->uvhub)
1079 * globally purge translation cache of a virtual address or all TLB's
1080 * @cpumask: mask of all cpu's in which the address is to be removed
1081 * @mm: mm_struct containing virtual address range
1082 * @start: start virtual address to be removed from TLB
1083 * @end: end virtual address to be remove from TLB
1084 * @cpu: the current cpu
1086 * This is the entry point for initiating any UV global TLB shootdown.
1088 * Purges the translation caches of all specified processors of the given
1089 * virtual address, or purges all TLB's on specified processors.
1091 * The caller has derived the cpumask from the mm_struct. This function
1092 * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1094 * The cpumask is converted into a uvhubmask of the uvhubs containing
1097 * Note that this function should be called with preemption disabled.
1099 * Returns NULL if all remote flushing was done.
1100 * Returns pointer to cpumask if some remote flushing remains to be
1101 * done. The returned pointer is valid till preemption is re-enabled.
1103 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1104 const struct flush_tlb_info *info)
1106 unsigned int cpu = smp_processor_id();
1107 int locals = 0, remotes = 0, hubs = 0;
1108 struct bau_desc *bau_desc;
1109 struct cpumask *flush_mask;
1110 struct ptc_stats *stat;
1111 struct bau_control *bcp;
1112 unsigned long descriptor_status, status, address;
1114 bcp = &per_cpu(bau_control, cpu);
1124 read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1125 status = ((descriptor_status >> (bcp->uvhub_cpu *
1126 UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1127 if (status == UV2H_DESC_BUSY)
1132 /* bau was disabled due to slow response */
1133 if (bcp->baudisabled) {
1134 if (check_enable(bcp, stat)) {
1135 stat->s_ipifordisabled++;
1141 * Each sending cpu has a per-cpu mask which it fills from the caller's
1142 * cpu mask. All cpus are converted to uvhubs and copied to the
1143 * activation descriptor.
1145 flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1146 /* don't actually do a shootdown of the local cpu */
1147 cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1149 if (cpumask_test_cpu(cpu, cpumask))
1150 stat->s_ntargself++;
1152 bau_desc = bcp->descriptor_base;
1153 bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1154 bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1155 if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1158 record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1160 if (!info->end || (info->end - info->start) <= PAGE_SIZE)
1161 address = info->start;
1163 address = TLB_FLUSH_ALL;
1165 switch (bcp->uvhub_version) {
1169 bau_desc->payload.uv1_2_3.address = address;
1170 bau_desc->payload.uv1_2_3.sending_cpu = cpu;
1173 bau_desc->payload.uv4.address = address;
1174 bau_desc->payload.uv4.sending_cpu = cpu;
1175 bau_desc->payload.uv4.qualifier = BAU_DESC_QUALIFIER;
1180 * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1181 * or 1 if it gave up and the original cpumask should be returned.
1183 if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1190 * Search the message queue for any 'other' unprocessed message with the
1191 * same software acknowledge resource bit vector as the 'msg' message.
1193 static struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1194 struct bau_control *bcp)
1196 struct bau_pq_entry *msg_next = msg + 1;
1197 unsigned char swack_vec = msg->swack_vec;
1199 if (msg_next > bcp->queue_last)
1200 msg_next = bcp->queue_first;
1201 while (msg_next != msg) {
1202 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1203 (msg_next->swack_vec == swack_vec))
1206 if (msg_next > bcp->queue_last)
1207 msg_next = bcp->queue_first;
1213 * UV2 needs to work around a bug in which an arriving message has not
1214 * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1215 * Such a message must be ignored.
1217 static void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1219 unsigned long mmr_image;
1220 unsigned char swack_vec;
1221 struct bau_pq_entry *msg = mdp->msg;
1222 struct bau_pq_entry *other_msg;
1224 mmr_image = ops.read_l_sw_ack();
1225 swack_vec = msg->swack_vec;
1227 if ((swack_vec & mmr_image) == 0) {
1229 * This message was assigned a swack resource, but no
1230 * reserved acknowlegment is pending.
1231 * The bug has prevented this message from setting the MMR.
1234 * Some message has set the MMR 'pending' bit; it might have
1235 * been another message. Look for that message.
1237 other_msg = find_another_by_swack(msg, bcp);
1240 * There is another. Process this one but do not
1243 bau_process_message(mdp, bcp, 0);
1245 * Let the natural processing of that other message
1246 * acknowledge it. Don't get the processing of sw_ack's
1254 * Either the MMR shows this one pending a reply or there is no
1255 * other message using this sw_ack, so it is safe to acknowledge it.
1257 bau_process_message(mdp, bcp, 1);
1263 * The BAU message interrupt comes here. (registered by set_intr_gate)
1266 * We received a broadcast assist message.
1268 * Interrupts are disabled; this interrupt could represent
1269 * the receipt of several messages.
1271 * All cores/threads on this hub get this interrupt.
1272 * The last one to see it does the software ack.
1273 * (the resource will not be freed until noninterruptable cpus see this
1274 * interrupt; hardware may timeout the s/w ack and reply ERROR)
1276 void uv_bau_message_interrupt(struct pt_regs *regs)
1279 cycles_t time_start;
1280 struct bau_pq_entry *msg;
1281 struct bau_control *bcp;
1282 struct ptc_stats *stat;
1283 struct msg_desc msgdesc;
1286 kvm_set_cpu_l1tf_flush_l1d();
1287 time_start = get_cycles();
1289 bcp = &per_cpu(bau_control, smp_processor_id());
1292 msgdesc.queue_first = bcp->queue_first;
1293 msgdesc.queue_last = bcp->queue_last;
1295 msg = bcp->bau_msg_head;
1296 while (msg->swack_vec) {
1299 msgdesc.msg_slot = msg - msgdesc.queue_first;
1301 if (bcp->uvhub_version == UV_BAU_V2)
1302 process_uv2_message(&msgdesc, bcp);
1304 /* no error workaround for uv1 or uv3 */
1305 bau_process_message(&msgdesc, bcp, 1);
1308 if (msg > msgdesc.queue_last)
1309 msg = msgdesc.queue_first;
1310 bcp->bau_msg_head = msg;
1312 stat->d_time += (get_cycles() - time_start);
1320 * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1321 * shootdown message timeouts enabled. The timeout does not cause
1322 * an interrupt, but causes an error message to be returned to
1325 static void __init enable_timeouts(void)
1330 unsigned long mmr_image;
1332 nuvhubs = uv_num_possible_blades();
1334 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1335 if (!uv_blade_nr_possible_cpus(uvhub))
1338 pnode = uv_blade_to_pnode(uvhub);
1339 mmr_image = read_mmr_misc_control(pnode);
1341 * Set the timeout period and then lock it in, in three
1342 * steps; captures and locks in the period.
1344 * To program the period, the SOFT_ACK_MODE must be off.
1346 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1347 write_mmr_misc_control(pnode, mmr_image);
1349 * Set the 4-bit period.
1351 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1352 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1353 write_mmr_misc_control(pnode, mmr_image);
1356 * Subsequent reversals of the timebase bit (3) cause an
1357 * immediate timeout of one or all INTD resources as
1358 * indicated in bits 2:0 (7 causes all of them to timeout).
1360 mmr_image |= (1L << SOFTACK_MSHIFT);
1362 /* do not touch the legacy mode bit */
1363 /* hw bug workaround; do not use extended status */
1364 mmr_image &= ~(1L << UV2_EXT_SHFT);
1365 } else if (is_uv3_hub()) {
1366 mmr_image &= ~(1L << PREFETCH_HINT_SHFT);
1367 mmr_image |= (1L << SB_STATUS_SHFT);
1369 write_mmr_misc_control(pnode, mmr_image);
1373 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1375 if (*offset < num_possible_cpus())
1380 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1383 if (*offset < num_possible_cpus())
1388 static void ptc_seq_stop(struct seq_file *file, void *data)
1393 * Display the statistics thru /proc/sgi_uv/ptc_statistics
1394 * 'data' points to the cpu number
1395 * Note: see the descriptions in stat_description[].
1397 static int ptc_seq_show(struct seq_file *file, void *data)
1399 struct ptc_stats *stat;
1400 struct bau_control *bcp;
1403 cpu = *(loff_t *)data;
1406 "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1407 seq_puts(file, "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1409 "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1411 "rok resetp resett giveup sto bz throt disable ");
1413 "enable wars warshw warwaits enters ipidis plugged ");
1415 "ipiover glim cong swack recv rtime all one mult ");
1416 seq_puts(file, "none retry canc nocan reset rcan\n");
1418 if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1419 bcp = &per_cpu(bau_control, cpu);
1421 seq_printf(file, "cpu %d bau disabled\n", cpu);
1425 /* source side statistics */
1427 "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1428 cpu, bcp->nobau, stat->s_requestor,
1429 cycles_2_us(stat->s_time),
1430 stat->s_ntargself, stat->s_ntarglocals,
1431 stat->s_ntargremotes, stat->s_ntargcpu,
1432 stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1433 stat->s_ntarguvhub, stat->s_ntarguvhub16);
1434 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1435 stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1436 stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1437 stat->s_dtimeout, stat->s_strongnacks);
1438 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1439 stat->s_retry_messages, stat->s_retriesok,
1440 stat->s_resets_plug, stat->s_resets_timeout,
1441 stat->s_giveup, stat->s_stimeout,
1442 stat->s_busy, stat->s_throttles);
1443 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1444 stat->s_bau_disabled, stat->s_bau_reenabled,
1445 stat->s_uv2_wars, stat->s_uv2_wars_hw,
1446 stat->s_uv2_war_waits, stat->s_enters,
1447 stat->s_ipifordisabled, stat->s_plugged,
1448 stat->s_overipilimit, stat->s_giveuplimit,
1451 /* destination side statistics */
1453 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1454 ops.read_g_sw_ack(uv_cpu_to_pnode(cpu)),
1455 stat->d_requestee, cycles_2_us(stat->d_time),
1456 stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1457 stat->d_nomsg, stat->d_retries, stat->d_canceled,
1458 stat->d_nocanceled, stat->d_resets,
1465 * Display the tunables thru debugfs
1467 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1468 size_t count, loff_t *ppos)
1473 buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1474 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1475 "ipi_reset_limit complete_threshold congested_response_us",
1476 "congested_reps disabled_period giveup_limit",
1477 max_concurr, plugged_delay, plugsb4reset,
1478 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1479 congested_respns_us, congested_reps, disabled_period,
1485 ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1491 * handle a write to /proc/sgi_uv/ptc_statistics
1492 * -1: reset the statistics
1493 * 0: display meaning of the statistics
1495 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1496 size_t count, loff_t *data)
1503 struct ptc_stats *stat;
1505 if (count == 0 || count > sizeof(optstr))
1507 if (copy_from_user(optstr, user, count))
1509 optstr[count - 1] = '\0';
1511 if (!strcmp(optstr, "on")) {
1514 } else if (!strcmp(optstr, "off")) {
1519 if (kstrtol(optstr, 10, &input_arg) < 0) {
1520 pr_debug("%s is invalid\n", optstr);
1524 if (input_arg == 0) {
1525 elements = ARRAY_SIZE(stat_description);
1526 pr_debug("# cpu: cpu number\n");
1527 pr_debug("Sender statistics:\n");
1528 for (i = 0; i < elements; i++)
1529 pr_debug("%s\n", stat_description[i]);
1530 } else if (input_arg == -1) {
1531 for_each_present_cpu(cpu) {
1532 stat = &per_cpu(ptcstats, cpu);
1533 memset(stat, 0, sizeof(struct ptc_stats));
1540 static int local_atoi(const char *name)
1547 val = 10*val+(*name-'0');
1556 * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1557 * Zero values reset them to defaults.
1559 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1566 int e = ARRAY_SIZE(tunables);
1568 p = instr + strspn(instr, WHITESPACE);
1570 for (; *p; p = q + strspn(q, WHITESPACE)) {
1571 q = p + strcspn(p, WHITESPACE);
1577 pr_info("bau tunable error: should be %d values\n", e);
1581 p = instr + strspn(instr, WHITESPACE);
1583 for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1584 q = p + strcspn(p, WHITESPACE);
1585 val = local_atoi(p);
1589 max_concurr = MAX_BAU_CONCURRENT;
1590 max_concurr_const = MAX_BAU_CONCURRENT;
1593 if (val < 1 || val > bcp->cpus_in_uvhub) {
1595 "Error: BAU max concurrent %d is invalid\n",
1600 max_concurr_const = val;
1604 *tunables[cnt].tunp = tunables[cnt].deflt;
1606 *tunables[cnt].tunp = val;
1614 * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1616 static ssize_t tunables_write(struct file *file, const char __user *user,
1617 size_t count, loff_t *data)
1622 struct bau_control *bcp;
1624 if (count == 0 || count > sizeof(instr)-1)
1626 if (copy_from_user(instr, user, count))
1629 instr[count] = '\0';
1632 bcp = &per_cpu(bau_control, cpu);
1633 ret = parse_tunables_write(bcp, instr, count);
1638 for_each_present_cpu(cpu) {
1639 bcp = &per_cpu(bau_control, cpu);
1640 bcp->max_concurr = max_concurr;
1641 bcp->max_concurr_const = max_concurr;
1642 bcp->plugged_delay = plugged_delay;
1643 bcp->plugsb4reset = plugsb4reset;
1644 bcp->timeoutsb4reset = timeoutsb4reset;
1645 bcp->ipi_reset_limit = ipi_reset_limit;
1646 bcp->complete_threshold = complete_threshold;
1647 bcp->cong_response_us = congested_respns_us;
1648 bcp->cong_reps = congested_reps;
1649 bcp->disabled_period = sec_2_cycles(disabled_period);
1650 bcp->giveup_limit = giveup_limit;
1655 static const struct seq_operations uv_ptc_seq_ops = {
1656 .start = ptc_seq_start,
1657 .next = ptc_seq_next,
1658 .stop = ptc_seq_stop,
1659 .show = ptc_seq_show
1662 static int ptc_proc_open(struct inode *inode, struct file *file)
1664 return seq_open(file, &uv_ptc_seq_ops);
1667 static int tunables_open(struct inode *inode, struct file *file)
1672 static const struct file_operations proc_uv_ptc_operations = {
1673 .open = ptc_proc_open,
1675 .write = ptc_proc_write,
1676 .llseek = seq_lseek,
1677 .release = seq_release,
1680 static const struct file_operations tunables_fops = {
1681 .open = tunables_open,
1682 .read = tunables_read,
1683 .write = tunables_write,
1684 .llseek = default_llseek,
1687 static int __init uv_ptc_init(void)
1689 struct proc_dir_entry *proc_uv_ptc;
1691 if (!is_uv_system())
1694 proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1695 &proc_uv_ptc_operations);
1697 pr_err("unable to create %s proc entry\n",
1702 tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1703 if (!tunables_dir) {
1704 pr_err("unable to create debugfs directory %s\n",
1705 UV_BAU_TUNABLES_DIR);
1708 tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1709 tunables_dir, NULL, &tunables_fops);
1710 if (!tunables_file) {
1711 pr_err("unable to create debugfs file %s\n",
1712 UV_BAU_TUNABLES_FILE);
1719 * Initialize the sending side's sending buffers.
1721 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1730 struct bau_desc *bau_desc;
1731 struct bau_desc *bd2;
1732 struct uv1_bau_msg_header *uv1_hdr;
1733 struct uv2_3_bau_msg_header *uv2_3_hdr;
1734 struct bau_control *bcp;
1737 * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1738 * per cpu; and one per cpu on the uvhub (ADP_SZ)
1740 dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1741 bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1744 gpa = uv_gpa(bau_desc);
1745 n = uv_gpa_to_gnode(gpa);
1746 m = ops.bau_gpa_to_offset(gpa);
1750 /* the 14-bit pnode */
1751 write_mmr_descriptor_base(pnode,
1752 (n << UVH_LB_BAU_SB_DESCRIPTOR_BASE_NODE_ID_SHFT | m));
1754 * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1755 * cpu even though we only use the first one; one descriptor can
1756 * describe a broadcast to 256 uv hubs.
1758 for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1759 memset(bd2, 0, sizeof(struct bau_desc));
1761 uv1_hdr = &bd2->header.uv1_hdr;
1762 uv1_hdr->swack_flag = 1;
1764 * The base_dest_nasid set in the message header
1765 * is the nasid of the first uvhub in the partition.
1766 * The bit map will indicate destination pnode numbers
1767 * relative to that base. They may not be consecutive
1768 * if nasid striding is being used.
1770 uv1_hdr->base_dest_nasid =
1771 UV_PNODE_TO_NASID(base_pnode);
1772 uv1_hdr->dest_subnodeid = UV_LB_SUBNODEID;
1773 uv1_hdr->command = UV_NET_ENDPOINT_INTD;
1774 uv1_hdr->int_both = 1;
1776 * all others need to be set to zero:
1777 * fairness chaining multilevel count replied_to
1781 * BIOS uses legacy mode, but uv2 and uv3 hardware always
1782 * uses native mode for selective broadcasts.
1784 uv2_3_hdr = &bd2->header.uv2_3_hdr;
1785 uv2_3_hdr->swack_flag = 1;
1786 uv2_3_hdr->base_dest_nasid =
1787 UV_PNODE_TO_NASID(base_pnode);
1788 uv2_3_hdr->dest_subnodeid = UV_LB_SUBNODEID;
1789 uv2_3_hdr->command = UV_NET_ENDPOINT_INTD;
1792 for_each_present_cpu(cpu) {
1793 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1795 bcp = &per_cpu(bau_control, cpu);
1796 bcp->descriptor_base = bau_desc;
1801 * initialize the destination side's receiving buffers
1802 * entered for each uvhub in the partition
1803 * - node is first node (kernel memory notion) on the uvhub
1804 * - pnode is the uvhub's physical identifier
1806 static void pq_init(int node, int pnode)
1812 unsigned long gnode, first, last, tail;
1813 struct bau_pq_entry *pqp;
1814 struct bau_control *bcp;
1816 plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1817 vp = kmalloc_node(plsize, GFP_KERNEL, node);
1818 pqp = (struct bau_pq_entry *)vp;
1821 cp = (char *)pqp + 31;
1822 pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1824 for_each_present_cpu(cpu) {
1825 if (pnode != uv_cpu_to_pnode(cpu))
1827 /* for every cpu on this pnode: */
1828 bcp = &per_cpu(bau_control, cpu);
1829 bcp->queue_first = pqp;
1830 bcp->bau_msg_head = pqp;
1831 bcp->queue_last = pqp + (DEST_Q_SIZE - 1);
1834 first = ops.bau_gpa_to_offset(uv_gpa(pqp));
1835 last = ops.bau_gpa_to_offset(uv_gpa(pqp + (DEST_Q_SIZE - 1)));
1838 * Pre UV4, the gnode is required to locate the payload queue
1839 * and the payload queue tail must be maintained by the kernel.
1841 bcp = &per_cpu(bau_control, smp_processor_id());
1842 if (bcp->uvhub_version <= UV_BAU_V3) {
1844 gnode = uv_gpa_to_gnode(uv_gpa(pqp));
1845 first = (gnode << UV_PAYLOADQ_GNODE_SHIFT) | tail;
1846 write_mmr_payload_tail(pnode, tail);
1849 ops.write_payload_first(pnode, first);
1850 ops.write_payload_last(pnode, last);
1852 /* in effect, all msg_type's are set to MSG_NOOP */
1853 memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1857 * Initialization of each UV hub's structures
1859 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1863 unsigned long apicid;
1865 node = uvhub_to_first_node(uvhub);
1866 pnode = uv_blade_to_pnode(uvhub);
1868 activation_descriptor_init(node, pnode, base_pnode);
1870 pq_init(node, pnode);
1872 * The below initialization can't be in firmware because the
1873 * messaging IRQ will be determined by the OS.
1875 apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1876 write_mmr_data_config(pnode, ((apicid << 32) | vector));
1880 * We will set BAU_MISC_CONTROL with a timeout period.
1881 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1882 * So the destination timeout period has to be calculated from them.
1884 static int calculate_destination_timeout(void)
1886 unsigned long mmr_image;
1892 unsigned long ts_ns;
1895 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1896 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1897 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1898 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1899 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1900 ts_ns = timeout_base_ns[index];
1901 ts_ns *= (mult1 * mult2);
1904 /* same destination timeout for uv2 and uv3 */
1905 /* 4 bits 0/1 for 10/80us base, 3 bits of multiplier */
1906 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1907 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1908 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1912 mult1 = mmr_image & UV2_ACK_MASK;
1918 static void __init init_per_cpu_tunables(void)
1921 struct bau_control *bcp;
1923 for_each_present_cpu(cpu) {
1924 bcp = &per_cpu(bau_control, cpu);
1925 bcp->baudisabled = 0;
1928 bcp->statp = &per_cpu(ptcstats, cpu);
1929 /* time interval to catch a hardware stay-busy bug */
1930 bcp->timeout_interval = usec_2_cycles(2*timeout_us);
1931 bcp->max_concurr = max_concurr;
1932 bcp->max_concurr_const = max_concurr;
1933 bcp->plugged_delay = plugged_delay;
1934 bcp->plugsb4reset = plugsb4reset;
1935 bcp->timeoutsb4reset = timeoutsb4reset;
1936 bcp->ipi_reset_limit = ipi_reset_limit;
1937 bcp->complete_threshold = complete_threshold;
1938 bcp->cong_response_us = congested_respns_us;
1939 bcp->cong_reps = congested_reps;
1940 bcp->disabled_period = sec_2_cycles(disabled_period);
1941 bcp->giveup_limit = giveup_limit;
1942 spin_lock_init(&bcp->queue_lock);
1943 spin_lock_init(&bcp->uvhub_lock);
1944 spin_lock_init(&bcp->disable_lock);
1949 * Scan all cpus to collect blade and socket summaries.
1951 static int __init get_cpu_topology(int base_pnode,
1952 struct uvhub_desc *uvhub_descs,
1953 unsigned char *uvhub_mask)
1959 struct bau_control *bcp;
1960 struct uvhub_desc *bdp;
1961 struct socket_desc *sdp;
1963 for_each_present_cpu(cpu) {
1964 bcp = &per_cpu(bau_control, cpu);
1966 memset(bcp, 0, sizeof(struct bau_control));
1968 pnode = uv_cpu_hub_info(cpu)->pnode;
1969 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1971 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1972 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1976 bcp->osnode = cpu_to_node(cpu);
1977 bcp->partition_base_pnode = base_pnode;
1979 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1980 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1981 bdp = &uvhub_descs[uvhub];
1987 /* kludge: 'assuming' one node per socket, and assuming that
1988 disabling a socket just leaves a gap in node numbers */
1989 socket = bcp->osnode & 1;
1990 bdp->socket_mask |= (1 << socket);
1991 sdp = &bdp->socket[socket];
1992 sdp->cpu_number[sdp->num_cpus] = cpu;
1994 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1995 pr_emerg("%d cpus per socket invalid\n",
2004 * Each socket is to get a local array of pnodes/hubs.
2006 static void make_per_cpu_thp(struct bau_control *smaster)
2009 size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
2011 smaster->thp = kzalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
2012 for_each_present_cpu(cpu) {
2013 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
2014 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
2019 * Each uvhub is to get a local cpumask.
2021 static void make_per_hub_cpumask(struct bau_control *hmaster)
2023 int sz = sizeof(cpumask_t);
2025 hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
2029 * Initialize all the per_cpu information for the cpu's on a given socket,
2030 * given what has been gathered into the socket_desc struct.
2031 * And reports the chosen hub and socket masters back to the caller.
2033 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
2034 struct bau_control **smasterp,
2035 struct bau_control **hmasterp)
2037 int i, cpu, uvhub_cpu;
2038 struct bau_control *bcp;
2040 for (i = 0; i < sdp->num_cpus; i++) {
2041 cpu = sdp->cpu_number[i];
2042 bcp = &per_cpu(bau_control, cpu);
2049 bcp->cpus_in_uvhub = bdp->num_cpus;
2050 bcp->cpus_in_socket = sdp->num_cpus;
2051 bcp->socket_master = *smasterp;
2052 bcp->uvhub = bdp->uvhub;
2054 bcp->uvhub_version = UV_BAU_V1;
2055 else if (is_uv2_hub())
2056 bcp->uvhub_version = UV_BAU_V2;
2057 else if (is_uv3_hub())
2058 bcp->uvhub_version = UV_BAU_V3;
2059 else if (is_uv4_hub())
2060 bcp->uvhub_version = UV_BAU_V4;
2062 pr_emerg("uvhub version not 1, 2, 3, or 4\n");
2065 bcp->uvhub_master = *hmasterp;
2066 uvhub_cpu = uv_cpu_blade_processor_id(cpu);
2067 bcp->uvhub_cpu = uvhub_cpu;
2070 * The ERROR and BUSY status registers are located pairwise over
2071 * the STATUS_0 and STATUS_1 mmrs; each an array[32] of 2 bits.
2073 if (uvhub_cpu < UV_CPUS_PER_AS) {
2074 bcp->status_mmr = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
2075 bcp->status_index = uvhub_cpu * UV_ACT_STATUS_SIZE;
2077 bcp->status_mmr = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
2078 bcp->status_index = (uvhub_cpu - UV_CPUS_PER_AS)
2079 * UV_ACT_STATUS_SIZE;
2082 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2083 pr_emerg("%d cpus per uvhub invalid\n",
2092 * Summarize the blade and socket topology into the per_cpu structures.
2094 static int __init summarize_uvhub_sockets(int nuvhubs,
2095 struct uvhub_desc *uvhub_descs,
2096 unsigned char *uvhub_mask)
2100 unsigned short socket_mask;
2102 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2103 struct uvhub_desc *bdp;
2104 struct bau_control *smaster = NULL;
2105 struct bau_control *hmaster = NULL;
2107 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2110 bdp = &uvhub_descs[uvhub];
2111 socket_mask = bdp->socket_mask;
2113 while (socket_mask) {
2114 struct socket_desc *sdp;
2115 if ((socket_mask & 1)) {
2116 sdp = &bdp->socket[socket];
2117 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2119 make_per_cpu_thp(smaster);
2122 socket_mask = (socket_mask >> 1);
2124 make_per_hub_cpumask(hmaster);
2130 * initialize the bau_control structure for each cpu
2132 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2134 struct uvhub_desc *uvhub_descs;
2135 unsigned char *uvhub_mask = NULL;
2137 if (is_uv3_hub() || is_uv2_hub() || is_uv1_hub())
2138 timeout_us = calculate_destination_timeout();
2140 uvhub_descs = kcalloc(nuvhubs, sizeof(struct uvhub_desc), GFP_KERNEL);
2144 uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2148 if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2151 if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2156 init_per_cpu_tunables();
2165 static const struct bau_operations uv1_bau_ops __initconst = {
2166 .bau_gpa_to_offset = uv_gpa_to_offset,
2167 .read_l_sw_ack = read_mmr_sw_ack,
2168 .read_g_sw_ack = read_gmmr_sw_ack,
2169 .write_l_sw_ack = write_mmr_sw_ack,
2170 .write_g_sw_ack = write_gmmr_sw_ack,
2171 .write_payload_first = write_mmr_payload_first,
2172 .write_payload_last = write_mmr_payload_last,
2173 .wait_completion = uv1_wait_completion,
2176 static const struct bau_operations uv2_3_bau_ops __initconst = {
2177 .bau_gpa_to_offset = uv_gpa_to_offset,
2178 .read_l_sw_ack = read_mmr_sw_ack,
2179 .read_g_sw_ack = read_gmmr_sw_ack,
2180 .write_l_sw_ack = write_mmr_sw_ack,
2181 .write_g_sw_ack = write_gmmr_sw_ack,
2182 .write_payload_first = write_mmr_payload_first,
2183 .write_payload_last = write_mmr_payload_last,
2184 .wait_completion = uv2_3_wait_completion,
2187 static const struct bau_operations uv4_bau_ops __initconst = {
2188 .bau_gpa_to_offset = uv_gpa_to_soc_phys_ram,
2189 .read_l_sw_ack = read_mmr_proc_sw_ack,
2190 .read_g_sw_ack = read_gmmr_proc_sw_ack,
2191 .write_l_sw_ack = write_mmr_proc_sw_ack,
2192 .write_g_sw_ack = write_gmmr_proc_sw_ack,
2193 .write_payload_first = write_mmr_proc_payload_first,
2194 .write_payload_last = write_mmr_proc_payload_last,
2195 .wait_completion = uv4_wait_completion,
2199 * Initialization of BAU-related structures
2201 static int __init uv_bau_init(void)
2209 cpumask_var_t *mask;
2211 if (!is_uv_system())
2216 else if (is_uv3_hub())
2217 ops = uv2_3_bau_ops;
2218 else if (is_uv2_hub())
2219 ops = uv2_3_bau_ops;
2220 else if (is_uv1_hub())
2223 nuvhubs = uv_num_possible_blades();
2225 pr_crit("UV: BAU disabled - insufficient hub count\n");
2226 goto err_bau_disable;
2229 for_each_possible_cpu(cur_cpu) {
2230 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2231 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2234 uv_base_pnode = 0x7fffffff;
2235 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2236 cpus = uv_blade_nr_possible_cpus(uvhub);
2237 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2238 uv_base_pnode = uv_blade_to_pnode(uvhub);
2241 /* software timeouts are not supported on UV4 */
2242 if (is_uv3_hub() || is_uv2_hub() || is_uv1_hub())
2245 if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2246 pr_crit("UV: BAU disabled - per CPU init failed\n");
2247 goto err_bau_disable;
2250 vector = UV_BAU_MESSAGE;
2251 for_each_possible_blade(uvhub) {
2252 if (uv_blade_nr_possible_cpus(uvhub))
2253 init_uvhub(uvhub, vector, uv_base_pnode);
2256 for_each_possible_blade(uvhub) {
2257 if (uv_blade_nr_possible_cpus(uvhub)) {
2260 pnode = uv_blade_to_pnode(uvhub);
2263 write_gmmr_activation(pnode, val);
2264 mmr = 1; /* should be 1 to broadcast to both sockets */
2266 write_mmr_data_broadcast(pnode, mmr);
2274 for_each_possible_cpu(cur_cpu)
2275 free_cpumask_var(per_cpu(uv_flush_tlb_mask, cur_cpu));
2282 core_initcall(uv_bau_init);
2283 fs_initcall(uv_ptc_init);