2 * SGI UltraViolet TLB flush routines.
4 * (c) 2008-2014 Cliff Wickman <cpw@sgi.com>, SGI.
6 * This code is released under the GNU General Public License version 2 or
9 #include <linux/seq_file.h>
10 #include <linux/proc_fs.h>
11 #include <linux/debugfs.h>
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/delay.h>
16 #include <asm/mmu_context.h>
17 #include <asm/uv/uv.h>
18 #include <asm/uv/uv_mmrs.h>
19 #include <asm/uv/uv_hub.h>
20 #include <asm/uv/uv_bau.h>
24 #include <asm/irq_vectors.h>
25 #include <asm/timer.h>
27 static struct bau_operations ops;
29 static struct bau_operations uv123_bau_ops = {
30 .bau_gpa_to_offset = uv_gpa_to_offset,
31 .read_l_sw_ack = read_mmr_sw_ack,
32 .read_g_sw_ack = read_gmmr_sw_ack,
33 .write_l_sw_ack = write_mmr_sw_ack,
34 .write_g_sw_ack = write_gmmr_sw_ack,
35 .write_payload_first = write_mmr_payload_first,
36 .write_payload_last = write_mmr_payload_last,
39 static struct bau_operations uv4_bau_ops = {
40 .bau_gpa_to_offset = uv_gpa_to_soc_phys_ram,
41 .read_l_sw_ack = read_mmr_proc_sw_ack,
42 .read_g_sw_ack = read_gmmr_proc_sw_ack,
43 .write_l_sw_ack = write_mmr_proc_sw_ack,
44 .write_g_sw_ack = write_gmmr_proc_sw_ack,
45 .write_payload_first = write_mmr_proc_payload_first,
46 .write_payload_last = write_mmr_proc_payload_last,
50 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
51 static int timeout_base_ns[] = {
62 static int timeout_us;
63 static bool nobau = true;
64 static int nobau_perm;
65 static cycles_t congested_cycles;
68 static int max_concurr = MAX_BAU_CONCURRENT;
69 static int max_concurr_const = MAX_BAU_CONCURRENT;
70 static int plugged_delay = PLUGGED_DELAY;
71 static int plugsb4reset = PLUGSB4RESET;
72 static int giveup_limit = GIVEUP_LIMIT;
73 static int timeoutsb4reset = TIMEOUTSB4RESET;
74 static int ipi_reset_limit = IPI_RESET_LIMIT;
75 static int complete_threshold = COMPLETE_THRESHOLD;
76 static int congested_respns_us = CONGESTED_RESPONSE_US;
77 static int congested_reps = CONGESTED_REPS;
78 static int disabled_period = DISABLED_PERIOD;
80 static struct tunables tunables[] = {
81 {&max_concurr, MAX_BAU_CONCURRENT}, /* must be [0] */
82 {&plugged_delay, PLUGGED_DELAY},
83 {&plugsb4reset, PLUGSB4RESET},
84 {&timeoutsb4reset, TIMEOUTSB4RESET},
85 {&ipi_reset_limit, IPI_RESET_LIMIT},
86 {&complete_threshold, COMPLETE_THRESHOLD},
87 {&congested_respns_us, CONGESTED_RESPONSE_US},
88 {&congested_reps, CONGESTED_REPS},
89 {&disabled_period, DISABLED_PERIOD},
90 {&giveup_limit, GIVEUP_LIMIT}
93 static struct dentry *tunables_dir;
94 static struct dentry *tunables_file;
96 /* these correspond to the statistics printed by ptc_seq_show() */
97 static char *stat_description[] = {
98 "sent: number of shootdown messages sent",
99 "stime: time spent sending messages",
100 "numuvhubs: number of hubs targeted with shootdown",
101 "numuvhubs16: number times 16 or more hubs targeted",
102 "numuvhubs8: number times 8 or more hubs targeted",
103 "numuvhubs4: number times 4 or more hubs targeted",
104 "numuvhubs2: number times 2 or more hubs targeted",
105 "numuvhubs1: number times 1 hub targeted",
106 "numcpus: number of cpus targeted with shootdown",
107 "dto: number of destination timeouts",
108 "retries: destination timeout retries sent",
109 "rok: : destination timeouts successfully retried",
110 "resetp: ipi-style resource resets for plugs",
111 "resett: ipi-style resource resets for timeouts",
112 "giveup: fall-backs to ipi-style shootdowns",
113 "sto: number of source timeouts",
114 "bz: number of stay-busy's",
115 "throt: number times spun in throttle",
116 "swack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE",
117 "recv: shootdown messages received",
118 "rtime: time spent processing messages",
119 "all: shootdown all-tlb messages",
120 "one: shootdown one-tlb messages",
121 "mult: interrupts that found multiple messages",
122 "none: interrupts that found no messages",
123 "retry: number of retry messages processed",
124 "canc: number messages canceled by retries",
125 "nocan: number retries that found nothing to cancel",
126 "reset: number of ipi-style reset requests processed",
127 "rcan: number messages canceled by reset requests",
128 "disable: number times use of the BAU was disabled",
129 "enable: number times use of the BAU was re-enabled"
132 static int __init setup_bau(char *arg)
139 result = strtobool(arg, &nobau);
143 /* we need to flip the logic here, so that bau=y sets nobau to false */
147 pr_info("UV BAU Enabled\n");
149 pr_info("UV BAU Disabled\n");
153 early_param("bau", setup_bau);
155 /* base pnode in this partition */
156 static int uv_base_pnode __read_mostly;
158 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
159 static DEFINE_PER_CPU(struct bau_control, bau_control);
160 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
166 struct bau_control *bcp;
169 pr_info("BAU not initialized; cannot be turned on\n");
173 for_each_present_cpu(cpu) {
174 bcp = &per_cpu(bau_control, cpu);
177 pr_info("BAU turned on\n");
185 struct bau_control *bcp;
188 for_each_present_cpu(cpu) {
189 bcp = &per_cpu(bau_control, cpu);
192 pr_info("BAU turned off\n");
197 * Determine the first node on a uvhub. 'Nodes' are used for kernel
200 static int __init uvhub_to_first_node(int uvhub)
204 for_each_online_node(node) {
205 b = uv_node_to_blade_id(node);
213 * Determine the apicid of the first cpu on a uvhub.
215 static int __init uvhub_to_first_apicid(int uvhub)
219 for_each_present_cpu(cpu)
220 if (uvhub == uv_cpu_to_blade_id(cpu))
221 return per_cpu(x86_cpu_to_apicid, cpu);
226 * Free a software acknowledge hardware resource by clearing its Pending
227 * bit. This will return a reply to the sender.
228 * If the message has timed out, a reply has already been sent by the
229 * hardware but the resource has not been released. In that case our
230 * clear of the Timeout bit (as well) will free the resource. No reply will
231 * be sent (the hardware will only do one reply per message).
233 static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp,
237 struct bau_pq_entry *msg;
240 if (!msg->canceled && do_acknowledge) {
241 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
242 ops.write_l_sw_ack(dw);
249 * Process the receipt of a RETRY message
251 static void bau_process_retry_msg(struct msg_desc *mdp,
252 struct bau_control *bcp)
255 int cancel_count = 0;
256 unsigned long msg_res;
257 unsigned long mmr = 0;
258 struct bau_pq_entry *msg = mdp->msg;
259 struct bau_pq_entry *msg2;
260 struct ptc_stats *stat = bcp->statp;
264 * cancel any message from msg+1 to the retry itself
266 for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
267 if (msg2 > mdp->queue_last)
268 msg2 = mdp->queue_first;
272 /* same conditions for cancellation as do_reset */
273 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
274 (msg2->swack_vec) && ((msg2->swack_vec &
275 msg->swack_vec) == 0) &&
276 (msg2->sending_cpu == msg->sending_cpu) &&
277 (msg2->msg_type != MSG_NOOP)) {
278 mmr = ops.read_l_sw_ack();
279 msg_res = msg2->swack_vec;
281 * This is a message retry; clear the resources held
282 * by the previous message only if they timed out.
283 * If it has not timed out we have an unexpected
284 * situation to report.
286 if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
289 * Is the resource timed out?
290 * Make everyone ignore the cancelled message.
295 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
296 ops.write_l_sw_ack(mr);
301 stat->d_nocanceled++;
305 * Do all the things a cpu should do for a TLB shootdown message.
306 * Other cpu's may come here at the same time for this message.
308 static void bau_process_message(struct msg_desc *mdp, struct bau_control *bcp,
311 short socket_ack_count = 0;
313 struct atomic_short *asp;
314 struct ptc_stats *stat = bcp->statp;
315 struct bau_pq_entry *msg = mdp->msg;
316 struct bau_control *smaster = bcp->socket_master;
319 * This must be a normal message, or retry of a normal message
321 if (msg->address == TLB_FLUSH_ALL) {
325 __flush_tlb_one(msg->address);
331 * One cpu on each uvhub has the additional job on a RETRY
332 * of releasing the resource held by the message that is
333 * being retried. That message is identified by sending
336 if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
337 bau_process_retry_msg(mdp, bcp);
340 * This is a swack message, so we have to reply to it.
341 * Count each responding cpu on the socket. This avoids
342 * pinging the count's cache line back and forth between
345 sp = &smaster->socket_acknowledge_count[mdp->msg_slot];
346 asp = (struct atomic_short *)sp;
347 socket_ack_count = atom_asr(1, asp);
348 if (socket_ack_count == bcp->cpus_in_socket) {
351 * Both sockets dump their completed count total into
352 * the message's count.
355 asp = (struct atomic_short *)&msg->acknowledge_count;
356 msg_ack_count = atom_asr(socket_ack_count, asp);
358 if (msg_ack_count == bcp->cpus_in_uvhub) {
360 * All cpus in uvhub saw it; reply
361 * (unless we are in the UV2 workaround)
363 reply_to_message(mdp, bcp, do_acknowledge);
371 * Determine the first cpu on a pnode.
373 static int pnode_to_first_cpu(int pnode, struct bau_control *smaster)
376 struct hub_and_pnode *hpp;
378 for_each_present_cpu(cpu) {
379 hpp = &smaster->thp[cpu];
380 if (pnode == hpp->pnode)
387 * Last resort when we get a large number of destination timeouts is
388 * to clear resources held by a given cpu.
389 * Do this with IPI so that all messages in the BAU message queue
390 * can be identified by their nonzero swack_vec field.
392 * This is entered for a single cpu on the uvhub.
393 * The sender want's this uvhub to free a specific message's
396 static void do_reset(void *ptr)
399 struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id());
400 struct reset_args *rap = (struct reset_args *)ptr;
401 struct bau_pq_entry *msg;
402 struct ptc_stats *stat = bcp->statp;
406 * We're looking for the given sender, and
407 * will free its swack resource.
408 * If all cpu's finally responded after the timeout, its
409 * message 'replied_to' was set.
411 for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
412 unsigned long msg_res;
413 /* do_reset: same conditions for cancellation as
414 bau_process_retry_msg() */
415 if ((msg->replied_to == 0) &&
416 (msg->canceled == 0) &&
417 (msg->sending_cpu == rap->sender) &&
419 (msg->msg_type != MSG_NOOP)) {
423 * make everyone else ignore this message
427 * only reset the resource if it is still pending
429 mmr = ops.read_l_sw_ack();
430 msg_res = msg->swack_vec;
431 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
434 ops.write_l_sw_ack(mr);
442 * Use IPI to get all target uvhubs to release resources held by
443 * a given sending cpu number.
445 static void reset_with_ipi(struct pnmask *distribution, struct bau_control *bcp)
450 int sender = bcp->cpu;
451 cpumask_t *mask = bcp->uvhub_master->cpumask;
452 struct bau_control *smaster = bcp->socket_master;
453 struct reset_args reset_args;
455 reset_args.sender = sender;
457 /* find a single cpu for each uvhub in this distribution mask */
458 maskbits = sizeof(struct pnmask) * BITSPERBYTE;
459 /* each bit is a pnode relative to the partition base pnode */
460 for (pnode = 0; pnode < maskbits; pnode++) {
462 if (!bau_uvhub_isset(pnode, distribution))
464 apnode = pnode + bcp->partition_base_pnode;
465 cpu = pnode_to_first_cpu(apnode, smaster);
466 cpumask_set_cpu(cpu, mask);
469 /* IPI all cpus; preemption is already disabled */
470 smp_call_function_many(mask, do_reset, (void *)&reset_args, 1);
475 * Not to be confused with cycles_2_ns() from tsc.c; this gives a relative
476 * number, not an absolute. It converts a duration in cycles to a duration in
479 static inline unsigned long long cycles_2_ns(unsigned long long cyc)
481 struct cyc2ns_data *data = cyc2ns_read_begin();
482 unsigned long long ns;
484 ns = mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift);
486 cyc2ns_read_end(data);
491 * The reverse of the above; converts a duration in ns to a duration in cycles.
493 static inline unsigned long long ns_2_cycles(unsigned long long ns)
495 struct cyc2ns_data *data = cyc2ns_read_begin();
496 unsigned long long cyc;
498 cyc = (ns << data->cyc2ns_shift) / data->cyc2ns_mul;
500 cyc2ns_read_end(data);
504 static inline unsigned long cycles_2_us(unsigned long long cyc)
506 return cycles_2_ns(cyc) / NSEC_PER_USEC;
509 static inline cycles_t sec_2_cycles(unsigned long sec)
511 return ns_2_cycles(sec * NSEC_PER_SEC);
514 static inline unsigned long long usec_2_cycles(unsigned long usec)
516 return ns_2_cycles(usec * NSEC_PER_USEC);
520 * wait for all cpus on this hub to finish their sends and go quiet
521 * leaves uvhub_quiesce set so that no new broadcasts are started by
522 * bau_flush_send_and_wait()
524 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
526 atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
530 * mark this quiet-requestor as done
532 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
534 atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
537 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
539 unsigned long descriptor_status;
541 descriptor_status = uv_read_local_mmr(mmr_offset);
542 descriptor_status >>= right_shift;
543 descriptor_status &= UV_ACT_STATUS_MASK;
544 return descriptor_status;
548 * Wait for completion of a broadcast software ack message
549 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
551 static int uv1_wait_completion(struct bau_desc *bau_desc,
552 unsigned long mmr_offset, int right_shift,
553 struct bau_control *bcp, long try)
555 unsigned long descriptor_status;
557 struct ptc_stats *stat = bcp->statp;
559 descriptor_status = uv1_read_status(mmr_offset, right_shift);
560 /* spin on the status MMR, waiting for it to go idle */
561 while ((descriptor_status != DS_IDLE)) {
563 * Our software ack messages may be blocked because
564 * there are no swack resources available. As long
565 * as none of them has timed out hardware will NACK
566 * our message and its state will stay IDLE.
568 if (descriptor_status == DS_SOURCE_TIMEOUT) {
571 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
576 * Our retries may be blocked by all destination
577 * swack resources being consumed, and a timeout
578 * pending. In that case hardware returns the
579 * ERROR that looks like a destination timeout.
581 if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
582 bcp->conseccompletes = 0;
583 return FLUSH_RETRY_PLUGGED;
586 bcp->conseccompletes = 0;
587 return FLUSH_RETRY_TIMEOUT;
590 * descriptor_status is still BUSY
594 descriptor_status = uv1_read_status(mmr_offset, right_shift);
596 bcp->conseccompletes++;
597 return FLUSH_COMPLETE;
601 * UV2 could have an extra bit of status in the ACTIVATION_STATUS_2 register.
602 * But not currently used.
604 static unsigned long uv2_3_read_status(unsigned long offset, int rshft, int desc)
606 unsigned long descriptor_status;
609 ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK) << 1;
610 return descriptor_status;
614 * Return whether the status of the descriptor that is normally used for this
615 * cpu (the one indexed by its hub-relative cpu number) is busy.
616 * The status of the original 32 descriptors is always reflected in the 64
617 * bits of UVH_LB_BAU_SB_ACTIVATION_STATUS_0.
618 * The bit provided by the activation_status_2 register is irrelevant to
619 * the status if it is only being tested for busy or not busy.
621 int normal_busy(struct bau_control *bcp)
623 int cpu = bcp->uvhub_cpu;
627 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
628 right_shift = cpu * UV_ACT_STATUS_SIZE;
629 return (((((read_lmmr(mmr_offset) >> right_shift) &
630 UV_ACT_STATUS_MASK)) << 1) == UV2H_DESC_BUSY);
634 * Entered when a bau descriptor has gone into a permanent busy wait because
636 * Workaround the bug.
638 int handle_uv2_busy(struct bau_control *bcp)
640 struct ptc_stats *stat = bcp->statp;
647 static int uv2_3_wait_completion(struct bau_desc *bau_desc,
648 unsigned long mmr_offset, int right_shift,
649 struct bau_control *bcp, long try)
651 unsigned long descriptor_stat;
653 int desc = bcp->uvhub_cpu;
655 struct ptc_stats *stat = bcp->statp;
657 descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
659 /* spin on the status MMR, waiting for it to go idle */
660 while (descriptor_stat != UV2H_DESC_IDLE) {
661 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT)) {
663 * A h/w bug on the destination side may
664 * have prevented the message being marked
665 * pending, thus it doesn't get replied to
666 * and gets continually nacked until it times
667 * out with a SOURCE_TIMEOUT.
671 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
675 * Our retries may be blocked by all destination
676 * swack resources being consumed, and a timeout
677 * pending. In that case hardware returns the
678 * ERROR that looks like a destination timeout.
679 * Without using the extended status we have to
680 * deduce from the short time that this was a
683 if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
684 bcp->conseccompletes = 0;
686 /* FLUSH_RETRY_PLUGGED causes hang on boot */
690 bcp->conseccompletes = 0;
691 /* FLUSH_RETRY_TIMEOUT causes hang on boot */
695 if (busy_reps > 1000000) {
696 /* not to hammer on the clock */
699 if ((ttm - bcp->send_message) > bcp->timeout_interval)
700 return handle_uv2_busy(bcp);
703 * descriptor_stat is still BUSY
707 descriptor_stat = uv2_3_read_status(mmr_offset, right_shift, desc);
709 bcp->conseccompletes++;
710 return FLUSH_COMPLETE;
714 * There are 2 status registers; each and array[32] of 2 bits. Set up for
715 * which register to read and position in that register based on cpu in
718 static int wait_completion(struct bau_desc *bau_desc, struct bau_control *bcp, long try)
721 unsigned long mmr_offset;
722 int desc = bcp->uvhub_cpu;
724 if (desc < UV_CPUS_PER_AS) {
725 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
726 right_shift = desc * UV_ACT_STATUS_SIZE;
728 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
729 right_shift = ((desc - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
732 if (bcp->uvhub_version == 1)
733 return uv1_wait_completion(bau_desc, mmr_offset, right_shift, bcp, try);
735 return uv2_3_wait_completion(bau_desc, mmr_offset, right_shift, bcp, try);
739 * Our retries are blocked by all destination sw ack resources being
740 * in use, and a timeout is pending. In that case hardware immediately
741 * returns the ERROR that looks like a destination timeout.
743 static void destination_plugged(struct bau_desc *bau_desc,
744 struct bau_control *bcp,
745 struct bau_control *hmaster, struct ptc_stats *stat)
747 udelay(bcp->plugged_delay);
748 bcp->plugged_tries++;
750 if (bcp->plugged_tries >= bcp->plugsb4reset) {
751 bcp->plugged_tries = 0;
753 quiesce_local_uvhub(hmaster);
755 spin_lock(&hmaster->queue_lock);
756 reset_with_ipi(&bau_desc->distribution, bcp);
757 spin_unlock(&hmaster->queue_lock);
759 end_uvhub_quiesce(hmaster);
762 stat->s_resets_plug++;
766 static void destination_timeout(struct bau_desc *bau_desc,
767 struct bau_control *bcp, struct bau_control *hmaster,
768 struct ptc_stats *stat)
770 hmaster->max_concurr = 1;
771 bcp->timeout_tries++;
772 if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
773 bcp->timeout_tries = 0;
775 quiesce_local_uvhub(hmaster);
777 spin_lock(&hmaster->queue_lock);
778 reset_with_ipi(&bau_desc->distribution, bcp);
779 spin_unlock(&hmaster->queue_lock);
781 end_uvhub_quiesce(hmaster);
784 stat->s_resets_timeout++;
789 * Stop all cpus on a uvhub from using the BAU for a period of time.
790 * This is reversed by check_enable.
792 static void disable_for_period(struct bau_control *bcp, struct ptc_stats *stat)
795 struct bau_control *tbcp;
796 struct bau_control *hmaster;
799 hmaster = bcp->uvhub_master;
800 spin_lock(&hmaster->disable_lock);
801 if (!bcp->baudisabled) {
802 stat->s_bau_disabled++;
804 for_each_present_cpu(tcpu) {
805 tbcp = &per_cpu(bau_control, tcpu);
806 if (tbcp->uvhub_master == hmaster) {
807 tbcp->baudisabled = 1;
808 tbcp->set_bau_on_time =
809 tm1 + bcp->disabled_period;
813 spin_unlock(&hmaster->disable_lock);
816 static void count_max_concurr(int stat, struct bau_control *bcp,
817 struct bau_control *hmaster)
819 bcp->plugged_tries = 0;
820 bcp->timeout_tries = 0;
821 if (stat != FLUSH_COMPLETE)
823 if (bcp->conseccompletes <= bcp->complete_threshold)
825 if (hmaster->max_concurr >= hmaster->max_concurr_const)
827 hmaster->max_concurr++;
830 static void record_send_stats(cycles_t time1, cycles_t time2,
831 struct bau_control *bcp, struct ptc_stats *stat,
832 int completion_status, int try)
837 elapsed = time2 - time1;
838 stat->s_time += elapsed;
840 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
841 bcp->period_requests++;
842 bcp->period_time += elapsed;
843 if ((elapsed > congested_cycles) &&
844 (bcp->period_requests > bcp->cong_reps) &&
845 ((bcp->period_time / bcp->period_requests) >
848 disable_for_period(bcp, stat);
854 if (completion_status == FLUSH_COMPLETE && try > 1)
856 else if (completion_status == FLUSH_GIVEUP) {
858 if (get_cycles() > bcp->period_end)
859 bcp->period_giveups = 0;
860 bcp->period_giveups++;
861 if (bcp->period_giveups == 1)
862 bcp->period_end = get_cycles() + bcp->disabled_period;
863 if (bcp->period_giveups > bcp->giveup_limit) {
864 disable_for_period(bcp, stat);
865 stat->s_giveuplimit++;
871 * Because of a uv1 hardware bug only a limited number of concurrent
872 * requests can be made.
874 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
876 spinlock_t *lock = &hmaster->uvhub_lock;
879 v = &hmaster->active_descriptor_count;
880 if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
884 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
889 * Handle the completion status of a message send.
891 static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
892 struct bau_control *bcp, struct bau_control *hmaster,
893 struct ptc_stats *stat)
895 if (completion_status == FLUSH_RETRY_PLUGGED)
896 destination_plugged(bau_desc, bcp, hmaster, stat);
897 else if (completion_status == FLUSH_RETRY_TIMEOUT)
898 destination_timeout(bau_desc, bcp, hmaster, stat);
902 * Send a broadcast and wait for it to complete.
904 * The flush_mask contains the cpus the broadcast is to be sent to including
905 * cpus that are on the local uvhub.
907 * Returns 0 if all flushing represented in the mask was done.
908 * Returns 1 if it gives up entirely and the original cpu mask is to be
909 * returned to the kernel.
911 int uv_flush_send_and_wait(struct cpumask *flush_mask, struct bau_control *bcp,
912 struct bau_desc *bau_desc)
915 int completion_stat = 0;
921 struct ptc_stats *stat = bcp->statp;
922 struct bau_control *hmaster = bcp->uvhub_master;
923 struct uv1_bau_msg_header *uv1_hdr = NULL;
924 struct uv2_3_bau_msg_header *uv2_3_hdr = NULL;
926 if (bcp->uvhub_version == 1) {
928 uv1_throttle(hmaster, stat);
931 while (hmaster->uvhub_quiesce)
934 time1 = get_cycles();
936 uv1_hdr = &bau_desc->header.uv1_hdr;
939 uv2_3_hdr = &bau_desc->header.uv2_3_hdr;
944 uv1_hdr->msg_type = MSG_REGULAR;
946 uv2_3_hdr->msg_type = MSG_REGULAR;
947 seq_number = bcp->message_number++;
950 uv1_hdr->msg_type = MSG_RETRY;
952 uv2_3_hdr->msg_type = MSG_RETRY;
953 stat->s_retry_messages++;
957 uv1_hdr->sequence = seq_number;
959 uv2_3_hdr->sequence = seq_number;
960 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
961 bcp->send_message = get_cycles();
963 write_mmr_activation(index);
966 completion_stat = wait_completion(bau_desc, bcp, try);
968 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
970 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
971 bcp->ipi_attempts = 0;
972 stat->s_overipilimit++;
973 completion_stat = FLUSH_GIVEUP;
977 } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
978 (completion_stat == FLUSH_RETRY_TIMEOUT));
980 time2 = get_cycles();
982 count_max_concurr(completion_stat, bcp, hmaster);
984 while (hmaster->uvhub_quiesce)
987 atomic_dec(&hmaster->active_descriptor_count);
989 record_send_stats(time1, time2, bcp, stat, completion_stat, try);
991 if (completion_stat == FLUSH_GIVEUP)
992 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
998 * The BAU is disabled for this uvhub. When the disabled time period has
999 * expired re-enable it.
1000 * Return 0 if it is re-enabled for all cpus on this uvhub.
1002 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
1005 struct bau_control *tbcp;
1006 struct bau_control *hmaster;
1008 hmaster = bcp->uvhub_master;
1009 spin_lock(&hmaster->disable_lock);
1010 if (bcp->baudisabled && (get_cycles() >= bcp->set_bau_on_time)) {
1011 stat->s_bau_reenabled++;
1012 for_each_present_cpu(tcpu) {
1013 tbcp = &per_cpu(bau_control, tcpu);
1014 if (tbcp->uvhub_master == hmaster) {
1015 tbcp->baudisabled = 0;
1016 tbcp->period_requests = 0;
1017 tbcp->period_time = 0;
1018 tbcp->period_giveups = 0;
1021 spin_unlock(&hmaster->disable_lock);
1024 spin_unlock(&hmaster->disable_lock);
1028 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
1029 int remotes, struct bau_desc *bau_desc)
1031 stat->s_requestor++;
1032 stat->s_ntargcpu += remotes + locals;
1033 stat->s_ntargremotes += remotes;
1034 stat->s_ntarglocals += locals;
1036 /* uvhub statistics */
1037 hubs = bau_uvhub_weight(&bau_desc->distribution);
1039 stat->s_ntarglocaluvhub++;
1040 stat->s_ntargremoteuvhub += (hubs - 1);
1042 stat->s_ntargremoteuvhub += hubs;
1044 stat->s_ntarguvhub += hubs;
1047 stat->s_ntarguvhub16++;
1049 stat->s_ntarguvhub8++;
1051 stat->s_ntarguvhub4++;
1053 stat->s_ntarguvhub2++;
1055 stat->s_ntarguvhub1++;
1059 * Translate a cpu mask to the uvhub distribution mask in the BAU
1060 * activation descriptor.
1062 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
1063 struct bau_desc *bau_desc, int *localsp, int *remotesp)
1068 struct hub_and_pnode *hpp;
1070 for_each_cpu(cpu, flush_mask) {
1072 * The distribution vector is a bit map of pnodes, relative
1073 * to the partition base pnode (and the partition base nasid
1075 * Translate cpu to pnode and hub using a local memory array.
1077 hpp = &bcp->socket_master->thp[cpu];
1078 pnode = hpp->pnode - bcp->partition_base_pnode;
1079 bau_uvhub_set(pnode, &bau_desc->distribution);
1081 if (hpp->uvhub == bcp->uvhub)
1092 * globally purge translation cache of a virtual address or all TLB's
1093 * @cpumask: mask of all cpu's in which the address is to be removed
1094 * @mm: mm_struct containing virtual address range
1095 * @start: start virtual address to be removed from TLB
1096 * @end: end virtual address to be remove from TLB
1097 * @cpu: the current cpu
1099 * This is the entry point for initiating any UV global TLB shootdown.
1101 * Purges the translation caches of all specified processors of the given
1102 * virtual address, or purges all TLB's on specified processors.
1104 * The caller has derived the cpumask from the mm_struct. This function
1105 * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1107 * The cpumask is converted into a uvhubmask of the uvhubs containing
1110 * Note that this function should be called with preemption disabled.
1112 * Returns NULL if all remote flushing was done.
1113 * Returns pointer to cpumask if some remote flushing remains to be
1114 * done. The returned pointer is valid till preemption is re-enabled.
1116 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
1117 struct mm_struct *mm,
1118 unsigned long start,
1125 struct bau_desc *bau_desc;
1126 struct cpumask *flush_mask;
1127 struct ptc_stats *stat;
1128 struct bau_control *bcp;
1129 unsigned long descriptor_status;
1130 unsigned long status;
1132 bcp = &per_cpu(bau_control, cpu);
1142 read_lmmr(UVH_LB_BAU_SB_ACTIVATION_STATUS_0);
1143 status = ((descriptor_status >> (bcp->uvhub_cpu *
1144 UV_ACT_STATUS_SIZE)) & UV_ACT_STATUS_MASK) << 1;
1145 if (status == UV2H_DESC_BUSY)
1150 /* bau was disabled due to slow response */
1151 if (bcp->baudisabled) {
1152 if (check_enable(bcp, stat)) {
1153 stat->s_ipifordisabled++;
1159 * Each sending cpu has a per-cpu mask which it fills from the caller's
1160 * cpu mask. All cpus are converted to uvhubs and copied to the
1161 * activation descriptor.
1163 flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
1164 /* don't actually do a shootdown of the local cpu */
1165 cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
1167 if (cpumask_test_cpu(cpu, cpumask))
1168 stat->s_ntargself++;
1170 bau_desc = bcp->descriptor_base;
1171 bau_desc += (ITEMS_PER_DESC * bcp->uvhub_cpu);
1172 bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
1173 if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
1176 record_send_statistics(stat, locals, hubs, remotes, bau_desc);
1178 if (!end || (end - start) <= PAGE_SIZE)
1179 bau_desc->payload.address = start;
1181 bau_desc->payload.address = TLB_FLUSH_ALL;
1182 bau_desc->payload.sending_cpu = cpu;
1184 * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1185 * or 1 if it gave up and the original cpumask should be returned.
1187 if (!uv_flush_send_and_wait(flush_mask, bcp, bau_desc))
1194 * Search the message queue for any 'other' unprocessed message with the
1195 * same software acknowledge resource bit vector as the 'msg' message.
1197 struct bau_pq_entry *find_another_by_swack(struct bau_pq_entry *msg,
1198 struct bau_control *bcp)
1200 struct bau_pq_entry *msg_next = msg + 1;
1201 unsigned char swack_vec = msg->swack_vec;
1203 if (msg_next > bcp->queue_last)
1204 msg_next = bcp->queue_first;
1205 while (msg_next != msg) {
1206 if ((msg_next->canceled == 0) && (msg_next->replied_to == 0) &&
1207 (msg_next->swack_vec == swack_vec))
1210 if (msg_next > bcp->queue_last)
1211 msg_next = bcp->queue_first;
1217 * UV2 needs to work around a bug in which an arriving message has not
1218 * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1219 * Such a message must be ignored.
1221 void process_uv2_message(struct msg_desc *mdp, struct bau_control *bcp)
1223 unsigned long mmr_image;
1224 unsigned char swack_vec;
1225 struct bau_pq_entry *msg = mdp->msg;
1226 struct bau_pq_entry *other_msg;
1228 mmr_image = ops.read_l_sw_ack();
1229 swack_vec = msg->swack_vec;
1231 if ((swack_vec & mmr_image) == 0) {
1233 * This message was assigned a swack resource, but no
1234 * reserved acknowlegment is pending.
1235 * The bug has prevented this message from setting the MMR.
1238 * Some message has set the MMR 'pending' bit; it might have
1239 * been another message. Look for that message.
1241 other_msg = find_another_by_swack(msg, bcp);
1244 * There is another. Process this one but do not
1247 bau_process_message(mdp, bcp, 0);
1249 * Let the natural processing of that other message
1250 * acknowledge it. Don't get the processing of sw_ack's
1258 * Either the MMR shows this one pending a reply or there is no
1259 * other message using this sw_ack, so it is safe to acknowledge it.
1261 bau_process_message(mdp, bcp, 1);
1267 * The BAU message interrupt comes here. (registered by set_intr_gate)
1270 * We received a broadcast assist message.
1272 * Interrupts are disabled; this interrupt could represent
1273 * the receipt of several messages.
1275 * All cores/threads on this hub get this interrupt.
1276 * The last one to see it does the software ack.
1277 * (the resource will not be freed until noninterruptable cpus see this
1278 * interrupt; hardware may timeout the s/w ack and reply ERROR)
1280 void uv_bau_message_interrupt(struct pt_regs *regs)
1283 cycles_t time_start;
1284 struct bau_pq_entry *msg;
1285 struct bau_control *bcp;
1286 struct ptc_stats *stat;
1287 struct msg_desc msgdesc;
1290 time_start = get_cycles();
1292 bcp = &per_cpu(bau_control, smp_processor_id());
1295 msgdesc.queue_first = bcp->queue_first;
1296 msgdesc.queue_last = bcp->queue_last;
1298 msg = bcp->bau_msg_head;
1299 while (msg->swack_vec) {
1302 msgdesc.msg_slot = msg - msgdesc.queue_first;
1304 if (bcp->uvhub_version == 2)
1305 process_uv2_message(&msgdesc, bcp);
1307 /* no error workaround for uv1 or uv3 */
1308 bau_process_message(&msgdesc, bcp, 1);
1311 if (msg > msgdesc.queue_last)
1312 msg = msgdesc.queue_first;
1313 bcp->bau_msg_head = msg;
1315 stat->d_time += (get_cycles() - time_start);
1323 * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1324 * shootdown message timeouts enabled. The timeout does not cause
1325 * an interrupt, but causes an error message to be returned to
1328 static void __init enable_timeouts(void)
1333 unsigned long mmr_image;
1335 nuvhubs = uv_num_possible_blades();
1337 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1338 if (!uv_blade_nr_possible_cpus(uvhub))
1341 pnode = uv_blade_to_pnode(uvhub);
1342 mmr_image = read_mmr_misc_control(pnode);
1344 * Set the timeout period and then lock it in, in three
1345 * steps; captures and locks in the period.
1347 * To program the period, the SOFT_ACK_MODE must be off.
1349 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1350 write_mmr_misc_control(pnode, mmr_image);
1352 * Set the 4-bit period.
1354 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1355 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1356 write_mmr_misc_control(pnode, mmr_image);
1359 * Subsequent reversals of the timebase bit (3) cause an
1360 * immediate timeout of one or all INTD resources as
1361 * indicated in bits 2:0 (7 causes all of them to timeout).
1363 mmr_image |= (1L << SOFTACK_MSHIFT);
1365 /* do not touch the legacy mode bit */
1366 /* hw bug workaround; do not use extended status */
1367 mmr_image &= ~(1L << UV2_EXT_SHFT);
1368 } else if (is_uv3_hub()) {
1369 mmr_image &= ~(1L << PREFETCH_HINT_SHFT);
1370 mmr_image |= (1L << SB_STATUS_SHFT);
1372 write_mmr_misc_control(pnode, mmr_image);
1376 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1378 if (*offset < num_possible_cpus())
1383 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1386 if (*offset < num_possible_cpus())
1391 static void ptc_seq_stop(struct seq_file *file, void *data)
1396 * Display the statistics thru /proc/sgi_uv/ptc_statistics
1397 * 'data' points to the cpu number
1398 * Note: see the descriptions in stat_description[].
1400 static int ptc_seq_show(struct seq_file *file, void *data)
1402 struct ptc_stats *stat;
1403 struct bau_control *bcp;
1406 cpu = *(loff_t *)data;
1409 "# cpu bauoff sent stime self locals remotes ncpus localhub ");
1410 seq_puts(file, "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1412 "numuvhubs4 numuvhubs2 numuvhubs1 dto snacks retries ");
1414 "rok resetp resett giveup sto bz throt disable ");
1416 "enable wars warshw warwaits enters ipidis plugged ");
1418 "ipiover glim cong swack recv rtime all one mult ");
1419 seq_puts(file, "none retry canc nocan reset rcan\n");
1421 if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1422 bcp = &per_cpu(bau_control, cpu);
1424 seq_printf(file, "cpu %d bau disabled\n", cpu);
1428 /* source side statistics */
1430 "cpu %d %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1431 cpu, bcp->nobau, stat->s_requestor,
1432 cycles_2_us(stat->s_time),
1433 stat->s_ntargself, stat->s_ntarglocals,
1434 stat->s_ntargremotes, stat->s_ntargcpu,
1435 stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1436 stat->s_ntarguvhub, stat->s_ntarguvhub16);
1437 seq_printf(file, "%ld %ld %ld %ld %ld %ld ",
1438 stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1439 stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1440 stat->s_dtimeout, stat->s_strongnacks);
1441 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1442 stat->s_retry_messages, stat->s_retriesok,
1443 stat->s_resets_plug, stat->s_resets_timeout,
1444 stat->s_giveup, stat->s_stimeout,
1445 stat->s_busy, stat->s_throttles);
1446 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1447 stat->s_bau_disabled, stat->s_bau_reenabled,
1448 stat->s_uv2_wars, stat->s_uv2_wars_hw,
1449 stat->s_uv2_war_waits, stat->s_enters,
1450 stat->s_ipifordisabled, stat->s_plugged,
1451 stat->s_overipilimit, stat->s_giveuplimit,
1454 /* destination side statistics */
1456 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
1457 ops.read_g_sw_ack(uv_cpu_to_pnode(cpu)),
1458 stat->d_requestee, cycles_2_us(stat->d_time),
1459 stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1460 stat->d_nomsg, stat->d_retries, stat->d_canceled,
1461 stat->d_nocanceled, stat->d_resets,
1468 * Display the tunables thru debugfs
1470 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1471 size_t count, loff_t *ppos)
1476 buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d %d\n",
1477 "max_concur plugged_delay plugsb4reset timeoutsb4reset",
1478 "ipi_reset_limit complete_threshold congested_response_us",
1479 "congested_reps disabled_period giveup_limit",
1480 max_concurr, plugged_delay, plugsb4reset,
1481 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1482 congested_respns_us, congested_reps, disabled_period,
1488 ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1494 * handle a write to /proc/sgi_uv/ptc_statistics
1495 * -1: reset the statistics
1496 * 0: display meaning of the statistics
1498 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1499 size_t count, loff_t *data)
1506 struct ptc_stats *stat;
1508 if (count == 0 || count > sizeof(optstr))
1510 if (copy_from_user(optstr, user, count))
1512 optstr[count - 1] = '\0';
1514 if (!strcmp(optstr, "on")) {
1517 } else if (!strcmp(optstr, "off")) {
1522 if (kstrtol(optstr, 10, &input_arg) < 0) {
1523 pr_debug("%s is invalid\n", optstr);
1527 if (input_arg == 0) {
1528 elements = ARRAY_SIZE(stat_description);
1529 pr_debug("# cpu: cpu number\n");
1530 pr_debug("Sender statistics:\n");
1531 for (i = 0; i < elements; i++)
1532 pr_debug("%s\n", stat_description[i]);
1533 } else if (input_arg == -1) {
1534 for_each_present_cpu(cpu) {
1535 stat = &per_cpu(ptcstats, cpu);
1536 memset(stat, 0, sizeof(struct ptc_stats));
1543 static int local_atoi(const char *name)
1550 val = 10*val+(*name-'0');
1559 * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1560 * Zero values reset them to defaults.
1562 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1569 int e = ARRAY_SIZE(tunables);
1571 p = instr + strspn(instr, WHITESPACE);
1573 for (; *p; p = q + strspn(q, WHITESPACE)) {
1574 q = p + strcspn(p, WHITESPACE);
1580 pr_info("bau tunable error: should be %d values\n", e);
1584 p = instr + strspn(instr, WHITESPACE);
1586 for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1587 q = p + strcspn(p, WHITESPACE);
1588 val = local_atoi(p);
1592 max_concurr = MAX_BAU_CONCURRENT;
1593 max_concurr_const = MAX_BAU_CONCURRENT;
1596 if (val < 1 || val > bcp->cpus_in_uvhub) {
1598 "Error: BAU max concurrent %d is invalid\n",
1603 max_concurr_const = val;
1607 *tunables[cnt].tunp = tunables[cnt].deflt;
1609 *tunables[cnt].tunp = val;
1619 * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1621 static ssize_t tunables_write(struct file *file, const char __user *user,
1622 size_t count, loff_t *data)
1627 struct bau_control *bcp;
1629 if (count == 0 || count > sizeof(instr)-1)
1631 if (copy_from_user(instr, user, count))
1634 instr[count] = '\0';
1637 bcp = &per_cpu(bau_control, cpu);
1638 ret = parse_tunables_write(bcp, instr, count);
1643 for_each_present_cpu(cpu) {
1644 bcp = &per_cpu(bau_control, cpu);
1645 bcp->max_concurr = max_concurr;
1646 bcp->max_concurr_const = max_concurr;
1647 bcp->plugged_delay = plugged_delay;
1648 bcp->plugsb4reset = plugsb4reset;
1649 bcp->timeoutsb4reset = timeoutsb4reset;
1650 bcp->ipi_reset_limit = ipi_reset_limit;
1651 bcp->complete_threshold = complete_threshold;
1652 bcp->cong_response_us = congested_respns_us;
1653 bcp->cong_reps = congested_reps;
1654 bcp->disabled_period = sec_2_cycles(disabled_period);
1655 bcp->giveup_limit = giveup_limit;
1660 static const struct seq_operations uv_ptc_seq_ops = {
1661 .start = ptc_seq_start,
1662 .next = ptc_seq_next,
1663 .stop = ptc_seq_stop,
1664 .show = ptc_seq_show
1667 static int ptc_proc_open(struct inode *inode, struct file *file)
1669 return seq_open(file, &uv_ptc_seq_ops);
1672 static int tunables_open(struct inode *inode, struct file *file)
1677 static const struct file_operations proc_uv_ptc_operations = {
1678 .open = ptc_proc_open,
1680 .write = ptc_proc_write,
1681 .llseek = seq_lseek,
1682 .release = seq_release,
1685 static const struct file_operations tunables_fops = {
1686 .open = tunables_open,
1687 .read = tunables_read,
1688 .write = tunables_write,
1689 .llseek = default_llseek,
1692 static int __init uv_ptc_init(void)
1694 struct proc_dir_entry *proc_uv_ptc;
1696 if (!is_uv_system())
1699 proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1700 &proc_uv_ptc_operations);
1702 pr_err("unable to create %s proc entry\n",
1707 tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1708 if (!tunables_dir) {
1709 pr_err("unable to create debugfs directory %s\n",
1710 UV_BAU_TUNABLES_DIR);
1713 tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1714 tunables_dir, NULL, &tunables_fops);
1715 if (!tunables_file) {
1716 pr_err("unable to create debugfs file %s\n",
1717 UV_BAU_TUNABLES_FILE);
1724 * Initialize the sending side's sending buffers.
1726 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1735 struct bau_desc *bau_desc;
1736 struct bau_desc *bd2;
1737 struct uv1_bau_msg_header *uv1_hdr;
1738 struct uv2_3_bau_msg_header *uv2_3_hdr;
1739 struct bau_control *bcp;
1742 * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1743 * per cpu; and one per cpu on the uvhub (ADP_SZ)
1745 dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1746 bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1749 gpa = uv_gpa(bau_desc);
1750 n = uv_gpa_to_gnode(gpa);
1751 m = ops.bau_gpa_to_offset(gpa);
1755 /* the 14-bit pnode */
1756 write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1758 * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1759 * cpu even though we only use the first one; one descriptor can
1760 * describe a broadcast to 256 uv hubs.
1762 for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1763 memset(bd2, 0, sizeof(struct bau_desc));
1765 uv1_hdr = &bd2->header.uv1_hdr;
1766 uv1_hdr->swack_flag = 1;
1768 * The base_dest_nasid set in the message header
1769 * is the nasid of the first uvhub in the partition.
1770 * The bit map will indicate destination pnode numbers
1771 * relative to that base. They may not be consecutive
1772 * if nasid striding is being used.
1774 uv1_hdr->base_dest_nasid =
1775 UV_PNODE_TO_NASID(base_pnode);
1776 uv1_hdr->dest_subnodeid = UV_LB_SUBNODEID;
1777 uv1_hdr->command = UV_NET_ENDPOINT_INTD;
1778 uv1_hdr->int_both = 1;
1780 * all others need to be set to zero:
1781 * fairness chaining multilevel count replied_to
1785 * BIOS uses legacy mode, but uv2 and uv3 hardware always
1786 * uses native mode for selective broadcasts.
1788 uv2_3_hdr = &bd2->header.uv2_3_hdr;
1789 uv2_3_hdr->swack_flag = 1;
1790 uv2_3_hdr->base_dest_nasid =
1791 UV_PNODE_TO_NASID(base_pnode);
1792 uv2_3_hdr->dest_subnodeid = UV_LB_SUBNODEID;
1793 uv2_3_hdr->command = UV_NET_ENDPOINT_INTD;
1796 for_each_present_cpu(cpu) {
1797 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1799 bcp = &per_cpu(bau_control, cpu);
1800 bcp->descriptor_base = bau_desc;
1805 * initialize the destination side's receiving buffers
1806 * entered for each uvhub in the partition
1807 * - node is first node (kernel memory notion) on the uvhub
1808 * - pnode is the uvhub's physical identifier
1810 static void pq_init(int node, int pnode)
1816 unsigned long gnode, first, last, tail;
1817 struct bau_pq_entry *pqp;
1818 struct bau_control *bcp;
1820 plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1821 vp = kmalloc_node(plsize, GFP_KERNEL, node);
1822 pqp = (struct bau_pq_entry *)vp;
1825 cp = (char *)pqp + 31;
1826 pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1828 for_each_present_cpu(cpu) {
1829 if (pnode != uv_cpu_to_pnode(cpu))
1831 /* for every cpu on this pnode: */
1832 bcp = &per_cpu(bau_control, cpu);
1833 bcp->queue_first = pqp;
1834 bcp->bau_msg_head = pqp;
1835 bcp->queue_last = pqp + (DEST_Q_SIZE - 1);
1838 first = ops.bau_gpa_to_offset(uv_gpa(pqp));
1839 last = ops.bau_gpa_to_offset(uv_gpa(pqp + (DEST_Q_SIZE - 1)));
1842 * Pre UV4, the gnode is required to locate the payload queue
1843 * and the payload queue tail must be maintained by the kernel.
1845 bcp = &per_cpu(bau_control, smp_processor_id());
1846 if (bcp->uvhub_version <= 3) {
1848 gnode = uv_gpa_to_gnode(uv_gpa(pqp));
1849 first = (gnode << UV_PAYLOADQ_GNODE_SHIFT) | tail;
1850 write_mmr_payload_tail(pnode, tail);
1853 ops.write_payload_first(pnode, first);
1854 ops.write_payload_last(pnode, last);
1855 ops.write_g_sw_ack(pnode, 0xffffUL);
1857 /* in effect, all msg_type's are set to MSG_NOOP */
1858 memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1862 * Initialization of each UV hub's structures
1864 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1868 unsigned long apicid;
1870 node = uvhub_to_first_node(uvhub);
1871 pnode = uv_blade_to_pnode(uvhub);
1873 activation_descriptor_init(node, pnode, base_pnode);
1875 pq_init(node, pnode);
1877 * The below initialization can't be in firmware because the
1878 * messaging IRQ will be determined by the OS.
1880 apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1881 write_mmr_data_config(pnode, ((apicid << 32) | vector));
1885 * We will set BAU_MISC_CONTROL with a timeout period.
1886 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1887 * So the destination timeout period has to be calculated from them.
1889 static int calculate_destination_timeout(void)
1891 unsigned long mmr_image;
1897 unsigned long ts_ns;
1900 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1901 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1902 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1903 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1904 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1905 ts_ns = timeout_base_ns[index];
1906 ts_ns *= (mult1 * mult2);
1909 /* same destination timeout for uv2 and uv3 */
1910 /* 4 bits 0/1 for 10/80us base, 3 bits of multiplier */
1911 mmr_image = uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL);
1912 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1913 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1917 mult1 = mmr_image & UV2_ACK_MASK;
1923 static void __init init_per_cpu_tunables(void)
1926 struct bau_control *bcp;
1928 for_each_present_cpu(cpu) {
1929 bcp = &per_cpu(bau_control, cpu);
1930 bcp->baudisabled = 0;
1933 bcp->statp = &per_cpu(ptcstats, cpu);
1934 /* time interval to catch a hardware stay-busy bug */
1935 bcp->timeout_interval = usec_2_cycles(2*timeout_us);
1936 bcp->max_concurr = max_concurr;
1937 bcp->max_concurr_const = max_concurr;
1938 bcp->plugged_delay = plugged_delay;
1939 bcp->plugsb4reset = plugsb4reset;
1940 bcp->timeoutsb4reset = timeoutsb4reset;
1941 bcp->ipi_reset_limit = ipi_reset_limit;
1942 bcp->complete_threshold = complete_threshold;
1943 bcp->cong_response_us = congested_respns_us;
1944 bcp->cong_reps = congested_reps;
1945 bcp->disabled_period = sec_2_cycles(disabled_period);
1946 bcp->giveup_limit = giveup_limit;
1947 spin_lock_init(&bcp->queue_lock);
1948 spin_lock_init(&bcp->uvhub_lock);
1949 spin_lock_init(&bcp->disable_lock);
1954 * Scan all cpus to collect blade and socket summaries.
1956 static int __init get_cpu_topology(int base_pnode,
1957 struct uvhub_desc *uvhub_descs,
1958 unsigned char *uvhub_mask)
1964 struct bau_control *bcp;
1965 struct uvhub_desc *bdp;
1966 struct socket_desc *sdp;
1968 for_each_present_cpu(cpu) {
1969 bcp = &per_cpu(bau_control, cpu);
1971 memset(bcp, 0, sizeof(struct bau_control));
1973 pnode = uv_cpu_hub_info(cpu)->pnode;
1974 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1976 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1977 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1981 bcp->osnode = cpu_to_node(cpu);
1982 bcp->partition_base_pnode = base_pnode;
1984 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1985 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1986 bdp = &uvhub_descs[uvhub];
1992 /* kludge: 'assuming' one node per socket, and assuming that
1993 disabling a socket just leaves a gap in node numbers */
1994 socket = bcp->osnode & 1;
1995 bdp->socket_mask |= (1 << socket);
1996 sdp = &bdp->socket[socket];
1997 sdp->cpu_number[sdp->num_cpus] = cpu;
1999 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
2000 pr_emerg("%d cpus per socket invalid\n",
2009 * Each socket is to get a local array of pnodes/hubs.
2011 static void make_per_cpu_thp(struct bau_control *smaster)
2014 size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
2016 smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
2017 memset(smaster->thp, 0, hpsz);
2018 for_each_present_cpu(cpu) {
2019 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
2020 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
2025 * Each uvhub is to get a local cpumask.
2027 static void make_per_hub_cpumask(struct bau_control *hmaster)
2029 int sz = sizeof(cpumask_t);
2031 hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
2035 * Initialize all the per_cpu information for the cpu's on a given socket,
2036 * given what has been gathered into the socket_desc struct.
2037 * And reports the chosen hub and socket masters back to the caller.
2039 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
2040 struct bau_control **smasterp,
2041 struct bau_control **hmasterp)
2045 struct bau_control *bcp;
2047 for (i = 0; i < sdp->num_cpus; i++) {
2048 cpu = sdp->cpu_number[i];
2049 bcp = &per_cpu(bau_control, cpu);
2056 bcp->cpus_in_uvhub = bdp->num_cpus;
2057 bcp->cpus_in_socket = sdp->num_cpus;
2058 bcp->socket_master = *smasterp;
2059 bcp->uvhub = bdp->uvhub;
2061 bcp->uvhub_version = 1;
2062 else if (is_uv2_hub())
2063 bcp->uvhub_version = 2;
2064 else if (is_uv3_hub())
2065 bcp->uvhub_version = 3;
2066 else if (is_uv4_hub())
2067 bcp->uvhub_version = 4;
2069 pr_emerg("uvhub version not 1, 2, 3, or 4\n");
2072 bcp->uvhub_master = *hmasterp;
2073 bcp->uvhub_cpu = uv_cpu_blade_processor_id(cpu);
2075 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
2076 pr_emerg("%d cpus per uvhub invalid\n",
2085 * Summarize the blade and socket topology into the per_cpu structures.
2087 static int __init summarize_uvhub_sockets(int nuvhubs,
2088 struct uvhub_desc *uvhub_descs,
2089 unsigned char *uvhub_mask)
2093 unsigned short socket_mask;
2095 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2096 struct uvhub_desc *bdp;
2097 struct bau_control *smaster = NULL;
2098 struct bau_control *hmaster = NULL;
2100 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
2103 bdp = &uvhub_descs[uvhub];
2104 socket_mask = bdp->socket_mask;
2106 while (socket_mask) {
2107 struct socket_desc *sdp;
2108 if ((socket_mask & 1)) {
2109 sdp = &bdp->socket[socket];
2110 if (scan_sock(sdp, bdp, &smaster, &hmaster))
2112 make_per_cpu_thp(smaster);
2115 socket_mask = (socket_mask >> 1);
2117 make_per_hub_cpumask(hmaster);
2123 * initialize the bau_control structure for each cpu
2125 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
2127 unsigned char *uvhub_mask;
2129 struct uvhub_desc *uvhub_descs;
2131 if (is_uv3_hub() || is_uv2_hub() || is_uv1_hub())
2132 timeout_us = calculate_destination_timeout();
2134 vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
2135 uvhub_descs = (struct uvhub_desc *)vp;
2136 memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
2137 uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
2139 if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
2142 if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
2147 init_per_cpu_tunables();
2157 * Initialization of BAU-related structures
2159 static int __init uv_bau_init(void)
2167 cpumask_var_t *mask;
2169 if (!is_uv_system())
2174 else if (is_uv3_hub())
2175 ops = uv123_bau_ops;
2176 else if (is_uv2_hub())
2177 ops = uv123_bau_ops;
2178 else if (is_uv1_hub())
2179 ops = uv123_bau_ops;
2181 for_each_possible_cpu(cur_cpu) {
2182 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
2183 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
2186 nuvhubs = uv_num_possible_blades();
2187 congested_cycles = usec_2_cycles(congested_respns_us);
2189 uv_base_pnode = 0x7fffffff;
2190 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
2191 cpus = uv_blade_nr_possible_cpus(uvhub);
2192 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
2193 uv_base_pnode = uv_blade_to_pnode(uvhub);
2196 /* software timeouts are not supported on UV4 */
2197 if (is_uv3_hub() || is_uv2_hub() || is_uv1_hub())
2200 if (init_per_cpu(nuvhubs, uv_base_pnode)) {
2206 vector = UV_BAU_MESSAGE;
2207 for_each_possible_blade(uvhub) {
2208 if (uv_blade_nr_possible_cpus(uvhub))
2209 init_uvhub(uvhub, vector, uv_base_pnode);
2212 alloc_intr_gate(vector, uv_bau_message_intr1);
2214 for_each_possible_blade(uvhub) {
2215 if (uv_blade_nr_possible_cpus(uvhub)) {
2218 pnode = uv_blade_to_pnode(uvhub);
2221 write_gmmr_activation(pnode, val);
2222 mmr = 1; /* should be 1 to broadcast to both sockets */
2224 write_mmr_data_broadcast(pnode, mmr);
2230 core_initcall(uv_bau_init);
2231 fs_initcall(uv_ptc_init);