2 * SGI UltraViolet TLB flush routines.
4 * (c) 2008-2010 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 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
28 static int timeout_base_ns[] = {
38 static int timeout_us;
40 static int baudisabled;
41 static spinlock_t disable_lock;
42 static cycles_t congested_cycles;
45 static int max_bau_concurrent = MAX_BAU_CONCURRENT;
46 static int max_bau_concurrent_constant = MAX_BAU_CONCURRENT;
47 static int plugged_delay = PLUGGED_DELAY;
48 static int plugsb4reset = PLUGSB4RESET;
49 static int timeoutsb4reset = TIMEOUTSB4RESET;
50 static int ipi_reset_limit = IPI_RESET_LIMIT;
51 static int complete_threshold = COMPLETE_THRESHOLD;
52 static int congested_response_us = CONGESTED_RESPONSE_US;
53 static int congested_reps = CONGESTED_REPS;
54 static int congested_period = CONGESTED_PERIOD;
55 static struct dentry *tunables_dir;
56 static struct dentry *tunables_file;
58 static int __init setup_nobau(char *arg)
63 early_param("nobau", setup_nobau);
65 /* base pnode in this partition */
66 static int uv_partition_base_pnode __read_mostly;
67 /* position of pnode (which is nasid>>1): */
68 static int uv_nshift __read_mostly;
69 static unsigned long uv_mmask __read_mostly;
71 static DEFINE_PER_CPU(struct ptc_stats, ptcstats);
72 static DEFINE_PER_CPU(struct bau_control, bau_control);
73 static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
76 * Determine the first node on a uvhub. 'Nodes' are used for kernel
79 static int __init uvhub_to_first_node(int uvhub)
83 for_each_online_node(node) {
84 b = uv_node_to_blade_id(node);
92 * Determine the apicid of the first cpu on a uvhub.
94 static int __init uvhub_to_first_apicid(int uvhub)
98 for_each_present_cpu(cpu)
99 if (uvhub == uv_cpu_to_blade_id(cpu))
100 return per_cpu(x86_cpu_to_apicid, cpu);
105 * Free a software acknowledge hardware resource by clearing its Pending
106 * bit. This will return a reply to the sender.
107 * If the message has timed out, a reply has already been sent by the
108 * hardware but the resource has not been released. In that case our
109 * clear of the Timeout bit (as well) will free the resource. No reply will
110 * be sent (the hardware will only do one reply per message).
112 static inline void uv_reply_to_message(struct msg_desc *mdp,
113 struct bau_control *bcp)
116 struct bau_payload_queue_entry *msg;
119 if (!msg->canceled) {
120 dw = (msg->sw_ack_vector << UV_SW_ACK_NPENDING) |
123 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw);
126 msg->sw_ack_vector = 0;
130 * Process the receipt of a RETRY message
132 static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
133 struct bau_control *bcp)
136 int cancel_count = 0;
138 unsigned long msg_res;
139 unsigned long mmr = 0;
140 struct bau_payload_queue_entry *msg;
141 struct bau_payload_queue_entry *msg2;
142 struct ptc_stats *stat;
148 * cancel any message from msg+1 to the retry itself
150 for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
151 if (msg2 > mdp->va_queue_last)
152 msg2 = mdp->va_queue_first;
156 /* same conditions for cancellation as uv_do_reset */
157 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
158 (msg2->sw_ack_vector) && ((msg2->sw_ack_vector &
159 msg->sw_ack_vector) == 0) &&
160 (msg2->sending_cpu == msg->sending_cpu) &&
161 (msg2->msg_type != MSG_NOOP)) {
162 slot2 = msg2 - mdp->va_queue_first;
163 mmr = uv_read_local_mmr
164 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
165 msg_res = msg2->sw_ack_vector;
167 * This is a message retry; clear the resources held
168 * by the previous message only if they timed out.
169 * If it has not timed out we have an unexpected
170 * situation to report.
172 if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
174 * is the resource timed out?
175 * make everyone ignore the cancelled message.
181 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
182 (msg_res << UV_SW_ACK_NPENDING) |
188 stat->d_nocanceled++;
192 * Do all the things a cpu should do for a TLB shootdown message.
193 * Other cpu's may come here at the same time for this message.
195 static void uv_bau_process_message(struct msg_desc *mdp,
196 struct bau_control *bcp)
199 short socket_ack_count = 0;
200 struct ptc_stats *stat;
201 struct bau_payload_queue_entry *msg;
202 struct bau_control *smaster = bcp->socket_master;
205 * This must be a normal message, or retry of a normal message
209 if (msg->address == TLB_FLUSH_ALL) {
213 __flush_tlb_one(msg->address);
219 * One cpu on each uvhub has the additional job on a RETRY
220 * of releasing the resource held by the message that is
221 * being retried. That message is identified by sending
224 if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
225 uv_bau_process_retry_msg(mdp, bcp);
228 * This is a sw_ack message, so we have to reply to it.
229 * Count each responding cpu on the socket. This avoids
230 * pinging the count's cache line back and forth between
233 socket_ack_count = atomic_add_short_return(1, (struct atomic_short *)
234 &smaster->socket_acknowledge_count[mdp->msg_slot]);
235 if (socket_ack_count == bcp->cpus_in_socket) {
237 * Both sockets dump their completed count total into
238 * the message's count.
240 smaster->socket_acknowledge_count[mdp->msg_slot] = 0;
241 msg_ack_count = atomic_add_short_return(socket_ack_count,
242 (struct atomic_short *)&msg->acknowledge_count);
244 if (msg_ack_count == bcp->cpus_in_uvhub) {
246 * All cpus in uvhub saw it; reply
248 uv_reply_to_message(mdp, bcp);
256 * Determine the first cpu on a uvhub.
258 static int uvhub_to_first_cpu(int uvhub)
261 for_each_present_cpu(cpu)
262 if (uvhub == uv_cpu_to_blade_id(cpu))
268 * Last resort when we get a large number of destination timeouts is
269 * to clear resources held by a given cpu.
270 * Do this with IPI so that all messages in the BAU message queue
271 * can be identified by their nonzero sw_ack_vector field.
273 * This is entered for a single cpu on the uvhub.
274 * The sender want's this uvhub to free a specific message's
278 uv_do_reset(void *ptr)
284 unsigned long msg_res;
285 struct bau_control *bcp;
286 struct reset_args *rap;
287 struct bau_payload_queue_entry *msg;
288 struct ptc_stats *stat;
290 bcp = &per_cpu(bau_control, smp_processor_id());
291 rap = (struct reset_args *)ptr;
296 * We're looking for the given sender, and
297 * will free its sw_ack resource.
298 * If all cpu's finally responded after the timeout, its
299 * message 'replied_to' was set.
301 for (msg = bcp->va_queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
302 /* uv_do_reset: same conditions for cancellation as
303 uv_bau_process_retry_msg() */
304 if ((msg->replied_to == 0) &&
305 (msg->canceled == 0) &&
306 (msg->sending_cpu == rap->sender) &&
307 (msg->sw_ack_vector) &&
308 (msg->msg_type != MSG_NOOP)) {
310 * make everyone else ignore this message
313 slot = msg - bcp->va_queue_first;
316 * only reset the resource if it is still pending
318 mmr = uv_read_local_mmr
319 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
320 msg_res = msg->sw_ack_vector;
324 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
325 (msg_res << UV_SW_ACK_NPENDING) |
334 * Use IPI to get all target uvhubs to release resources held by
335 * a given sending cpu number.
337 static void uv_reset_with_ipi(struct bau_target_uvhubmask *distribution,
343 struct reset_args reset_args;
345 reset_args.sender = sender;
348 /* find a single cpu for each uvhub in this distribution mask */
350 uvhub < sizeof(struct bau_target_uvhubmask) * BITSPERBYTE;
352 if (!bau_uvhub_isset(uvhub, distribution))
354 /* find a cpu for this uvhub */
355 cpu = uvhub_to_first_cpu(uvhub);
358 /* IPI all cpus; Preemption is already disabled */
359 smp_call_function_many(&mask, uv_do_reset, (void *)&reset_args, 1);
363 static inline unsigned long
364 cycles_2_us(unsigned long long cyc)
366 unsigned long long ns;
368 ns = (cyc * per_cpu(cyc2ns, smp_processor_id()))
369 >> CYC2NS_SCALE_FACTOR;
375 * wait for all cpus on this hub to finish their sends and go quiet
376 * leaves uvhub_quiesce set so that no new broadcasts are started by
377 * bau_flush_send_and_wait()
380 quiesce_local_uvhub(struct bau_control *hmaster)
382 atomic_add_short_return(1, (struct atomic_short *)
383 &hmaster->uvhub_quiesce);
387 * mark this quiet-requestor as done
390 end_uvhub_quiesce(struct bau_control *hmaster)
392 atomic_add_short_return(-1, (struct atomic_short *)
393 &hmaster->uvhub_quiesce);
397 * Wait for completion of a broadcast software ack message
398 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
400 static int uv_wait_completion(struct bau_desc *bau_desc,
401 unsigned long mmr_offset, int right_shift, int this_cpu,
402 struct bau_control *bcp, struct bau_control *smaster, long try)
404 unsigned long descriptor_status;
406 struct ptc_stats *stat = bcp->statp;
407 struct bau_control *hmaster;
409 hmaster = bcp->uvhub_master;
411 /* spin on the status MMR, waiting for it to go idle */
412 while ((descriptor_status = (((unsigned long)
413 uv_read_local_mmr(mmr_offset) >>
414 right_shift) & UV_ACT_STATUS_MASK)) !=
417 * Our software ack messages may be blocked because there are
418 * no swack resources available. As long as none of them
419 * has timed out hardware will NACK our message and its
420 * state will stay IDLE.
422 if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) {
425 } else if (descriptor_status ==
426 DESC_STATUS_DESTINATION_TIMEOUT) {
428 ttime = get_cycles();
431 * Our retries may be blocked by all destination
432 * swack resources being consumed, and a timeout
433 * pending. In that case hardware returns the
434 * ERROR that looks like a destination timeout.
436 if (cycles_2_us(ttime - bcp->send_message) <
438 bcp->conseccompletes = 0;
439 return FLUSH_RETRY_PLUGGED;
442 bcp->conseccompletes = 0;
443 return FLUSH_RETRY_TIMEOUT;
446 * descriptor_status is still BUSY
451 bcp->conseccompletes++;
452 return FLUSH_COMPLETE;
455 static inline cycles_t
456 sec_2_cycles(unsigned long sec)
461 ns = sec * 1000000000;
462 cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
467 * conditionally add 1 to *v, unless *v is >= u
468 * return 0 if we cannot add 1 to *v because it is >= u
469 * return 1 if we can add 1 to *v because it is < u
472 * This is close to atomic_add_unless(), but this allows the 'u' value
473 * to be lowered below the current 'v'. atomic_add_unless can only stop
476 static inline int atomic_inc_unless_ge(spinlock_t *lock, atomic_t *v, int u)
479 if (atomic_read(v) >= u) {
489 * Our retries are blocked by all destination swack resources being
490 * in use, and a timeout is pending. In that case hardware immediately
491 * returns the ERROR that looks like a destination timeout.
494 destination_plugged(struct bau_desc *bau_desc, struct bau_control *bcp,
495 struct bau_control *hmaster, struct ptc_stats *stat)
497 udelay(bcp->plugged_delay);
498 bcp->plugged_tries++;
499 if (bcp->plugged_tries >= bcp->plugsb4reset) {
500 bcp->plugged_tries = 0;
501 quiesce_local_uvhub(hmaster);
502 spin_lock(&hmaster->queue_lock);
503 uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu);
504 spin_unlock(&hmaster->queue_lock);
505 end_uvhub_quiesce(hmaster);
507 stat->s_resets_plug++;
512 destination_timeout(struct bau_desc *bau_desc, struct bau_control *bcp,
513 struct bau_control *hmaster, struct ptc_stats *stat)
515 hmaster->max_bau_concurrent = 1;
516 bcp->timeout_tries++;
517 if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
518 bcp->timeout_tries = 0;
519 quiesce_local_uvhub(hmaster);
520 spin_lock(&hmaster->queue_lock);
521 uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu);
522 spin_unlock(&hmaster->queue_lock);
523 end_uvhub_quiesce(hmaster);
525 stat->s_resets_timeout++;
530 * Completions are taking a very long time due to a congested numalink
534 disable_for_congestion(struct bau_control *bcp, struct ptc_stats *stat)
537 struct bau_control *tbcp;
539 /* let only one cpu do this disabling */
540 spin_lock(&disable_lock);
541 if (!baudisabled && bcp->period_requests &&
542 ((bcp->period_time / bcp->period_requests) > congested_cycles)) {
543 /* it becomes this cpu's job to turn on the use of the
546 bcp->set_bau_off = 1;
547 bcp->set_bau_on_time = get_cycles() +
548 sec_2_cycles(bcp->congested_period);
549 stat->s_bau_disabled++;
550 for_each_present_cpu(tcpu) {
551 tbcp = &per_cpu(bau_control, tcpu);
552 tbcp->baudisabled = 1;
555 spin_unlock(&disable_lock);
559 * uv_flush_send_and_wait
561 * Send a broadcast and wait for it to complete.
563 * The flush_mask contains the cpus the broadcast is to be sent to including
564 * cpus that are on the local uvhub.
566 * Returns 0 if all flushing represented in the mask was done.
567 * Returns 1 if it gives up entirely and the original cpu mask is to be
568 * returned to the kernel.
570 int uv_flush_send_and_wait(struct bau_desc *bau_desc,
571 struct cpumask *flush_mask, struct bau_control *bcp)
574 int completion_status = 0;
577 int cpu = bcp->uvhub_cpu;
578 int this_cpu = bcp->cpu;
579 unsigned long mmr_offset;
584 struct ptc_stats *stat = bcp->statp;
585 struct bau_control *smaster = bcp->socket_master;
586 struct bau_control *hmaster = bcp->uvhub_master;
588 if (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
589 &hmaster->active_descriptor_count,
590 hmaster->max_bau_concurrent)) {
594 } while (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
595 &hmaster->active_descriptor_count,
596 hmaster->max_bau_concurrent));
598 while (hmaster->uvhub_quiesce)
601 if (cpu < UV_CPUS_PER_ACT_STATUS) {
602 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
603 right_shift = cpu * UV_ACT_STATUS_SIZE;
605 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
607 ((cpu - UV_CPUS_PER_ACT_STATUS) * UV_ACT_STATUS_SIZE);
609 time1 = get_cycles();
612 bau_desc->header.msg_type = MSG_REGULAR;
613 seq_number = bcp->message_number++;
615 bau_desc->header.msg_type = MSG_RETRY;
616 stat->s_retry_messages++;
618 bau_desc->header.sequence = seq_number;
619 index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) |
621 bcp->send_message = get_cycles();
622 uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index);
624 completion_status = uv_wait_completion(bau_desc, mmr_offset,
625 right_shift, this_cpu, bcp, smaster, try);
627 if (completion_status == FLUSH_RETRY_PLUGGED) {
628 destination_plugged(bau_desc, bcp, hmaster, stat);
629 } else if (completion_status == FLUSH_RETRY_TIMEOUT) {
630 destination_timeout(bau_desc, bcp, hmaster, stat);
632 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
633 bcp->ipi_attempts = 0;
634 completion_status = FLUSH_GIVEUP;
638 } while ((completion_status == FLUSH_RETRY_PLUGGED) ||
639 (completion_status == FLUSH_RETRY_TIMEOUT));
640 time2 = get_cycles();
641 bcp->plugged_tries = 0;
642 bcp->timeout_tries = 0;
643 if ((completion_status == FLUSH_COMPLETE) &&
644 (bcp->conseccompletes > bcp->complete_threshold) &&
645 (hmaster->max_bau_concurrent <
646 hmaster->max_bau_concurrent_constant))
647 hmaster->max_bau_concurrent++;
648 while (hmaster->uvhub_quiesce)
650 atomic_dec(&hmaster->active_descriptor_count);
652 elapsed = time2 - time1;
653 stat->s_time += elapsed;
654 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
655 bcp->period_requests++;
656 bcp->period_time += elapsed;
657 if ((elapsed > congested_cycles) &&
658 (bcp->period_requests > bcp->congested_reps)) {
659 disable_for_congestion(bcp, stat);
664 if (completion_status == FLUSH_COMPLETE && try > 1)
666 else if (completion_status == FLUSH_GIVEUP) {
674 * uv_flush_tlb_others - globally purge translation cache of a virtual
675 * address or all TLB's
676 * @cpumask: mask of all cpu's in which the address is to be removed
677 * @mm: mm_struct containing virtual address range
678 * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
679 * @cpu: the current cpu
681 * This is the entry point for initiating any UV global TLB shootdown.
683 * Purges the translation caches of all specified processors of the given
684 * virtual address, or purges all TLB's on specified processors.
686 * The caller has derived the cpumask from the mm_struct. This function
687 * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
689 * The cpumask is converted into a uvhubmask of the uvhubs containing
692 * Note that this function should be called with preemption disabled.
694 * Returns NULL if all remote flushing was done.
695 * Returns pointer to cpumask if some remote flushing remains to be
696 * done. The returned pointer is valid till preemption is re-enabled.
698 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
699 struct mm_struct *mm,
700 unsigned long va, unsigned int cpu)
707 struct bau_desc *bau_desc;
708 struct cpumask *flush_mask;
709 struct ptc_stats *stat;
710 struct bau_control *bcp;
711 struct bau_control *tbcp;
712 struct hub_and_pnode *hpp;
714 /* kernel was booted 'nobau' */
718 bcp = &per_cpu(bau_control, cpu);
721 /* bau was disabled due to slow response */
722 if (bcp->baudisabled) {
723 /* the cpu that disabled it must re-enable it */
724 if (bcp->set_bau_off) {
725 if (get_cycles() >= bcp->set_bau_on_time) {
726 stat->s_bau_reenabled++;
728 for_each_present_cpu(tcpu) {
729 tbcp = &per_cpu(bau_control, tcpu);
730 tbcp->baudisabled = 0;
731 tbcp->period_requests = 0;
732 tbcp->period_time = 0;
740 * Each sending cpu has a per-cpu mask which it fills from the caller's
741 * cpu mask. All cpus are converted to uvhubs and copied to the
742 * activation descriptor.
744 flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
745 /* don't actually do a shootdown of the local cpu */
746 cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
747 if (cpu_isset(cpu, *cpumask))
750 bau_desc = bcp->descriptor_base;
751 bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu;
752 bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
754 for_each_cpu(tcpu, flush_mask) {
756 * The distribution vector is a bit map of pnodes, relative
757 * to the partition base pnode (and the partition base nasid
759 * Translate cpu to pnode and hub using an array stored
762 hpp = &bcp->socket_master->target_hub_and_pnode[tcpu];
763 tpnode = hpp->pnode - bcp->partition_base_pnode;
764 bau_uvhub_set(tpnode, &bau_desc->distribution);
765 if (hpp->uvhub == bcp->uvhub)
770 if ((locals + remotes) == 0)
773 stat->s_ntargcpu += remotes + locals;
774 stat->s_ntargremotes += remotes;
775 stat->s_ntarglocals += locals;
776 remotes = bau_uvhub_weight(&bau_desc->distribution);
778 /* uvhub statistics */
779 hubs = bau_uvhub_weight(&bau_desc->distribution);
781 stat->s_ntarglocaluvhub++;
782 stat->s_ntargremoteuvhub += (hubs - 1);
784 stat->s_ntargremoteuvhub += hubs;
785 stat->s_ntarguvhub += hubs;
787 stat->s_ntarguvhub16++;
789 stat->s_ntarguvhub8++;
791 stat->s_ntarguvhub4++;
793 stat->s_ntarguvhub2++;
795 stat->s_ntarguvhub1++;
797 bau_desc->payload.address = va;
798 bau_desc->payload.sending_cpu = cpu;
801 * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
802 * or 1 if it gave up and the original cpumask should be returned.
804 if (!uv_flush_send_and_wait(bau_desc, flush_mask, bcp))
811 * The BAU message interrupt comes here. (registered by set_intr_gate)
814 * We received a broadcast assist message.
816 * Interrupts are disabled; this interrupt could represent
817 * the receipt of several messages.
819 * All cores/threads on this hub get this interrupt.
820 * The last one to see it does the software ack.
821 * (the resource will not be freed until noninterruptable cpus see this
822 * interrupt; hardware may timeout the s/w ack and reply ERROR)
824 void uv_bau_message_interrupt(struct pt_regs *regs)
828 struct bau_payload_queue_entry *msg;
829 struct bau_control *bcp;
830 struct ptc_stats *stat;
831 struct msg_desc msgdesc;
833 time_start = get_cycles();
834 bcp = &per_cpu(bau_control, smp_processor_id());
836 msgdesc.va_queue_first = bcp->va_queue_first;
837 msgdesc.va_queue_last = bcp->va_queue_last;
838 msg = bcp->bau_msg_head;
839 while (msg->sw_ack_vector) {
841 msgdesc.msg_slot = msg - msgdesc.va_queue_first;
842 msgdesc.sw_ack_slot = ffs(msg->sw_ack_vector) - 1;
844 uv_bau_process_message(&msgdesc, bcp);
846 if (msg > msgdesc.va_queue_last)
847 msg = msgdesc.va_queue_first;
848 bcp->bau_msg_head = msg;
850 stat->d_time += (get_cycles() - time_start);
861 * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
862 * shootdown message timeouts enabled. The timeout does not cause
863 * an interrupt, but causes an error message to be returned to
866 static void __init uv_enable_timeouts(void)
871 unsigned long mmr_image;
873 nuvhubs = uv_num_possible_blades();
875 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
876 if (!uv_blade_nr_possible_cpus(uvhub))
879 pnode = uv_blade_to_pnode(uvhub);
881 uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL);
883 * Set the timeout period and then lock it in, in three
884 * steps; captures and locks in the period.
886 * To program the period, the SOFT_ACK_MODE must be off.
888 mmr_image &= ~((unsigned long)1 <<
889 UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
890 uv_write_global_mmr64
891 (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
893 * Set the 4-bit period.
895 mmr_image &= ~((unsigned long)0xf <<
896 UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
897 mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD <<
898 UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
899 uv_write_global_mmr64
900 (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
902 * Subsequent reversals of the timebase bit (3) cause an
903 * immediate timeout of one or all INTD resources as
904 * indicated in bits 2:0 (7 causes all of them to timeout).
906 mmr_image |= ((unsigned long)1 <<
907 UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT);
908 uv_write_global_mmr64
909 (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
913 static void *uv_ptc_seq_start(struct seq_file *file, loff_t *offset)
915 if (*offset < num_possible_cpus())
920 static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
923 if (*offset < num_possible_cpus())
928 static void uv_ptc_seq_stop(struct seq_file *file, void *data)
932 static inline unsigned long long
933 microsec_2_cycles(unsigned long microsec)
936 unsigned long long cyc;
938 ns = microsec * 1000;
939 cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
944 * Display the statistics thru /proc.
945 * 'data' points to the cpu number
947 static int uv_ptc_seq_show(struct seq_file *file, void *data)
949 struct ptc_stats *stat;
952 cpu = *(loff_t *)data;
956 "# cpu sent stime self locals remotes ncpus localhub ");
958 "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
960 "numuvhubs4 numuvhubs2 numuvhubs1 dto ");
962 "retries rok resetp resett giveup sto bz throt ");
964 "sw_ack recv rtime all ");
966 "one mult none retry canc nocan reset rcan ");
970 if (cpu < num_possible_cpus() && cpu_online(cpu)) {
971 stat = &per_cpu(ptcstats, cpu);
972 /* source side statistics */
974 "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
975 cpu, stat->s_requestor, cycles_2_us(stat->s_time),
976 stat->s_ntargself, stat->s_ntarglocals,
977 stat->s_ntargremotes, stat->s_ntargcpu,
978 stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
979 stat->s_ntarguvhub, stat->s_ntarguvhub16);
980 seq_printf(file, "%ld %ld %ld %ld %ld ",
981 stat->s_ntarguvhub8, stat->s_ntarguvhub4,
982 stat->s_ntarguvhub2, stat->s_ntarguvhub1,
984 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
985 stat->s_retry_messages, stat->s_retriesok,
986 stat->s_resets_plug, stat->s_resets_timeout,
987 stat->s_giveup, stat->s_stimeout,
988 stat->s_busy, stat->s_throttles);
990 /* destination side statistics */
992 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
993 uv_read_global_mmr64(uv_cpu_to_pnode(cpu),
994 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE),
995 stat->d_requestee, cycles_2_us(stat->d_time),
996 stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
997 stat->d_nomsg, stat->d_retries, stat->d_canceled,
998 stat->d_nocanceled, stat->d_resets,
1000 seq_printf(file, "%ld %ld\n",
1001 stat->s_bau_disabled, stat->s_bau_reenabled);
1008 * Display the tunables thru debugfs
1010 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1011 size_t count, loff_t *ppos)
1016 buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1017 "max_bau_concurrent plugged_delay plugsb4reset",
1018 "timeoutsb4reset ipi_reset_limit complete_threshold",
1019 "congested_response_us congested_reps congested_period",
1020 max_bau_concurrent, plugged_delay, plugsb4reset,
1021 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1022 congested_response_us, congested_reps, congested_period);
1027 ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1033 * -1: resetf the statistics
1034 * 0: display meaning of the statistics
1036 static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user,
1037 size_t count, loff_t *data)
1042 struct ptc_stats *stat;
1044 if (count == 0 || count > sizeof(optstr))
1046 if (copy_from_user(optstr, user, count))
1048 optstr[count - 1] = '\0';
1049 if (strict_strtol(optstr, 10, &input_arg) < 0) {
1050 printk(KERN_DEBUG "%s is invalid\n", optstr);
1054 if (input_arg == 0) {
1055 printk(KERN_DEBUG "# cpu: cpu number\n");
1056 printk(KERN_DEBUG "Sender statistics:\n");
1058 "sent: number of shootdown messages sent\n");
1060 "stime: time spent sending messages\n");
1062 "numuvhubs: number of hubs targeted with shootdown\n");
1064 "numuvhubs16: number times 16 or more hubs targeted\n");
1066 "numuvhubs8: number times 8 or more hubs targeted\n");
1068 "numuvhubs4: number times 4 or more hubs targeted\n");
1070 "numuvhubs2: number times 2 or more hubs targeted\n");
1072 "numuvhubs1: number times 1 hub targeted\n");
1074 "numcpus: number of cpus targeted with shootdown\n");
1076 "dto: number of destination timeouts\n");
1078 "retries: destination timeout retries sent\n");
1080 "rok: : destination timeouts successfully retried\n");
1082 "resetp: ipi-style resource resets for plugs\n");
1084 "resett: ipi-style resource resets for timeouts\n");
1086 "giveup: fall-backs to ipi-style shootdowns\n");
1088 "sto: number of source timeouts\n");
1090 "bz: number of stay-busy's\n");
1092 "throt: number times spun in throttle\n");
1093 printk(KERN_DEBUG "Destination side statistics:\n");
1095 "sw_ack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
1097 "recv: shootdown messages received\n");
1099 "rtime: time spent processing messages\n");
1101 "all: shootdown all-tlb messages\n");
1103 "one: shootdown one-tlb messages\n");
1105 "mult: interrupts that found multiple messages\n");
1107 "none: interrupts that found no messages\n");
1109 "retry: number of retry messages processed\n");
1111 "canc: number messages canceled by retries\n");
1113 "nocan: number retries that found nothing to cancel\n");
1115 "reset: number of ipi-style reset requests processed\n");
1117 "rcan: number messages canceled by reset requests\n");
1119 "disable: number times use of the BAU was disabled\n");
1121 "enable: number times use of the BAU was re-enabled\n");
1122 } else if (input_arg == -1) {
1123 for_each_present_cpu(cpu) {
1124 stat = &per_cpu(ptcstats, cpu);
1125 memset(stat, 0, sizeof(struct ptc_stats));
1132 static int local_atoi(const char *name)
1139 val = 10*val+(*name-'0');
1149 * 0 values reset them to defaults
1151 static ssize_t tunables_write(struct file *file, const char __user *user,
1152 size_t count, loff_t *data)
1160 struct bau_control *bcp;
1162 if (count == 0 || count > sizeof(instr)-1)
1164 if (copy_from_user(instr, user, count))
1167 instr[count] = '\0';
1168 /* count the fields */
1169 p = instr + strspn(instr, WHITESPACE);
1171 for (; *p; p = q + strspn(q, WHITESPACE)) {
1172 q = p + strcspn(p, WHITESPACE);
1178 printk(KERN_INFO "bau tunable error: should be 9 numbers\n");
1182 p = instr + strspn(instr, WHITESPACE);
1184 for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1185 q = p + strcspn(p, WHITESPACE);
1186 val = local_atoi(p);
1190 max_bau_concurrent = MAX_BAU_CONCURRENT;
1191 max_bau_concurrent_constant =
1195 bcp = &per_cpu(bau_control, smp_processor_id());
1196 if (val < 1 || val > bcp->cpus_in_uvhub) {
1198 "Error: BAU max concurrent %d is invalid\n",
1202 max_bau_concurrent = val;
1203 max_bau_concurrent_constant = val;
1207 plugged_delay = PLUGGED_DELAY;
1209 plugged_delay = val;
1213 plugsb4reset = PLUGSB4RESET;
1219 timeoutsb4reset = TIMEOUTSB4RESET;
1221 timeoutsb4reset = val;
1225 ipi_reset_limit = IPI_RESET_LIMIT;
1227 ipi_reset_limit = val;
1231 complete_threshold = COMPLETE_THRESHOLD;
1233 complete_threshold = val;
1237 congested_response_us = CONGESTED_RESPONSE_US;
1239 congested_response_us = val;
1243 congested_reps = CONGESTED_REPS;
1245 congested_reps = val;
1249 congested_period = CONGESTED_PERIOD;
1251 congested_period = val;
1257 for_each_present_cpu(cpu) {
1258 bcp = &per_cpu(bau_control, cpu);
1259 bcp->max_bau_concurrent = max_bau_concurrent;
1260 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1261 bcp->plugged_delay = plugged_delay;
1262 bcp->plugsb4reset = plugsb4reset;
1263 bcp->timeoutsb4reset = timeoutsb4reset;
1264 bcp->ipi_reset_limit = ipi_reset_limit;
1265 bcp->complete_threshold = complete_threshold;
1266 bcp->congested_response_us = congested_response_us;
1267 bcp->congested_reps = congested_reps;
1268 bcp->congested_period = congested_period;
1273 static const struct seq_operations uv_ptc_seq_ops = {
1274 .start = uv_ptc_seq_start,
1275 .next = uv_ptc_seq_next,
1276 .stop = uv_ptc_seq_stop,
1277 .show = uv_ptc_seq_show
1280 static int uv_ptc_proc_open(struct inode *inode, struct file *file)
1282 return seq_open(file, &uv_ptc_seq_ops);
1285 static int tunables_open(struct inode *inode, struct file *file)
1290 static const struct file_operations proc_uv_ptc_operations = {
1291 .open = uv_ptc_proc_open,
1293 .write = uv_ptc_proc_write,
1294 .llseek = seq_lseek,
1295 .release = seq_release,
1298 static const struct file_operations tunables_fops = {
1299 .open = tunables_open,
1300 .read = tunables_read,
1301 .write = tunables_write,
1302 .llseek = default_llseek,
1305 static int __init uv_ptc_init(void)
1307 struct proc_dir_entry *proc_uv_ptc;
1309 if (!is_uv_system())
1312 proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1313 &proc_uv_ptc_operations);
1315 printk(KERN_ERR "unable to create %s proc entry\n",
1320 tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1321 if (!tunables_dir) {
1322 printk(KERN_ERR "unable to create debugfs directory %s\n",
1323 UV_BAU_TUNABLES_DIR);
1326 tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1327 tunables_dir, NULL, &tunables_fops);
1328 if (!tunables_file) {
1329 printk(KERN_ERR "unable to create debugfs file %s\n",
1330 UV_BAU_TUNABLES_FILE);
1337 * Initialize the sending side's sending buffers.
1340 uv_activation_descriptor_init(int node, int pnode, int base_pnode)
1347 struct bau_desc *bau_desc;
1348 struct bau_desc *bd2;
1349 struct bau_control *bcp;
1352 * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR)
1353 * per cpu; and one per cpu on the uvhub (UV_ADP_SIZE)
1355 bau_desc = kmalloc_node(sizeof(struct bau_desc) * UV_ADP_SIZE
1356 * UV_ITEMS_PER_DESCRIPTOR, GFP_KERNEL, node);
1359 pa = uv_gpa(bau_desc); /* need the real nasid*/
1360 n = pa >> uv_nshift;
1363 /* the 14-bit pnode */
1364 uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE,
1365 (n << UV_DESC_BASE_PNODE_SHIFT | m));
1367 * Initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each
1368 * cpu even though we only use the first one; one descriptor can
1369 * describe a broadcast to 256 uv hubs.
1371 for (i = 0, bd2 = bau_desc; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR);
1373 memset(bd2, 0, sizeof(struct bau_desc));
1374 bd2->header.sw_ack_flag = 1;
1376 * The base_dest_nasid set in the message header is the nasid
1377 * of the first uvhub in the partition. The bit map will
1378 * indicate destination pnode numbers relative to that base.
1379 * They may not be consecutive if nasid striding is being used.
1381 bd2->header.base_dest_nasid = UV_PNODE_TO_NASID(base_pnode);
1382 bd2->header.dest_subnodeid = UV_LB_SUBNODEID;
1383 bd2->header.command = UV_NET_ENDPOINT_INTD;
1384 bd2->header.int_both = 1;
1386 * all others need to be set to zero:
1387 * fairness chaining multilevel count replied_to
1390 for_each_present_cpu(cpu) {
1391 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1393 bcp = &per_cpu(bau_control, cpu);
1394 bcp->descriptor_base = bau_desc;
1399 * initialize the destination side's receiving buffers
1400 * entered for each uvhub in the partition
1401 * - node is first node (kernel memory notion) on the uvhub
1402 * - pnode is the uvhub's physical identifier
1405 uv_payload_queue_init(int node, int pnode)
1411 struct bau_payload_queue_entry *pqp;
1412 struct bau_payload_queue_entry *pqp_malloc;
1413 struct bau_control *bcp;
1415 pqp = kmalloc_node((DEST_Q_SIZE + 1)
1416 * sizeof(struct bau_payload_queue_entry),
1421 cp = (char *)pqp + 31;
1422 pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5);
1424 for_each_present_cpu(cpu) {
1425 if (pnode != uv_cpu_to_pnode(cpu))
1427 /* for every cpu on this pnode: */
1428 bcp = &per_cpu(bau_control, cpu);
1429 bcp->va_queue_first = pqp;
1430 bcp->bau_msg_head = pqp;
1431 bcp->va_queue_last = pqp + (DEST_Q_SIZE - 1);
1434 * need the pnode of where the memory was really allocated
1437 pn = pa >> uv_nshift;
1438 uv_write_global_mmr64(pnode,
1439 UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST,
1440 ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) |
1441 uv_physnodeaddr(pqp));
1442 uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL,
1443 uv_physnodeaddr(pqp));
1444 uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST,
1446 uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1)));
1447 /* in effect, all msg_type's are set to MSG_NOOP */
1448 memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE);
1452 * Initialization of each UV hub's structures
1454 static void __init uv_init_uvhub(int uvhub, int vector, int base_pnode)
1458 unsigned long apicid;
1460 node = uvhub_to_first_node(uvhub);
1461 pnode = uv_blade_to_pnode(uvhub);
1462 uv_activation_descriptor_init(node, pnode, base_pnode);
1463 uv_payload_queue_init(node, pnode);
1465 * The below initialization can't be in firmware because the
1466 * messaging IRQ will be determined by the OS.
1468 apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1469 uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG,
1470 ((apicid << 32) | vector));
1474 * We will set BAU_MISC_CONTROL with a timeout period.
1475 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1476 * So the destination timeout period has be be calculated from them.
1479 calculate_destination_timeout(void)
1481 unsigned long mmr_image;
1487 unsigned long ts_ns;
1489 mult1 = UV_INTD_SOFT_ACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1490 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1491 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1492 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1493 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1494 base = timeout_base_ns[index];
1495 ts_ns = base * mult1 * mult2;
1501 * initialize the bau_control structure for each cpu
1503 static int __init uv_init_per_cpu(int nuvhubs, int base_part_pnode)
1512 unsigned short socket_mask;
1513 unsigned char *uvhub_mask;
1514 struct bau_control *bcp;
1515 struct uvhub_desc *bdp;
1516 struct socket_desc *sdp;
1517 struct bau_control *hmaster = NULL;
1518 struct bau_control *smaster = NULL;
1519 struct socket_desc {
1521 short cpu_number[MAX_CPUS_PER_SOCKET];
1524 unsigned short socket_mask;
1528 struct socket_desc socket[2];
1530 struct uvhub_desc *uvhub_descs;
1532 timeout_us = calculate_destination_timeout();
1534 uvhub_descs = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1535 memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1536 uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
1537 for_each_present_cpu(cpu) {
1538 bcp = &per_cpu(bau_control, cpu);
1539 memset(bcp, 0, sizeof(struct bau_control));
1540 pnode = uv_cpu_hub_info(cpu)->pnode;
1541 if ((pnode - base_part_pnode) >= UV_DISTRIBUTION_SIZE) {
1543 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1544 cpu, pnode, base_part_pnode,
1545 UV_DISTRIBUTION_SIZE);
1548 bcp->osnode = cpu_to_node(cpu);
1549 bcp->partition_base_pnode = uv_partition_base_pnode;
1550 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1551 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1552 bdp = &uvhub_descs[uvhub];
1556 /* kludge: 'assuming' one node per socket, and assuming that
1557 disabling a socket just leaves a gap in node numbers */
1558 socket = bcp->osnode & 1;
1559 bdp->socket_mask |= (1 << socket);
1560 sdp = &bdp->socket[socket];
1561 sdp->cpu_number[sdp->num_cpus] = cpu;
1563 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1564 printk(KERN_EMERG "%d cpus per socket invalid\n", sdp->num_cpus);
1568 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1569 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
1572 bdp = &uvhub_descs[uvhub];
1573 socket_mask = bdp->socket_mask;
1575 while (socket_mask) {
1576 if (!(socket_mask & 1))
1578 sdp = &bdp->socket[socket];
1579 for (i = 0; i < sdp->num_cpus; i++) {
1580 cpu = sdp->cpu_number[i];
1581 bcp = &per_cpu(bau_control, cpu);
1585 if (!have_hmaster) {
1590 bcp->cpus_in_uvhub = bdp->num_cpus;
1591 bcp->cpus_in_socket = sdp->num_cpus;
1592 bcp->socket_master = smaster;
1593 bcp->uvhub = bdp->uvhub;
1594 bcp->uvhub_master = hmaster;
1595 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->
1597 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
1599 "%d cpus per uvhub invalid\n",
1606 socket_mask = (socket_mask >> 1);
1607 /* each socket gets a local array of pnodes/hubs */
1609 bcp->target_hub_and_pnode = kmalloc_node(
1610 sizeof(struct hub_and_pnode) *
1611 num_possible_cpus(), GFP_KERNEL, bcp->osnode);
1612 memset(bcp->target_hub_and_pnode, 0,
1613 sizeof(struct hub_and_pnode) *
1614 num_possible_cpus());
1615 for_each_present_cpu(tcpu) {
1616 bcp->target_hub_and_pnode[tcpu].pnode =
1617 uv_cpu_hub_info(tcpu)->pnode;
1618 bcp->target_hub_and_pnode[tcpu].uvhub =
1619 uv_cpu_hub_info(tcpu)->numa_blade_id;
1625 for_each_present_cpu(cpu) {
1626 bcp = &per_cpu(bau_control, cpu);
1627 bcp->baudisabled = 0;
1628 bcp->statp = &per_cpu(ptcstats, cpu);
1629 /* time interval to catch a hardware stay-busy bug */
1630 bcp->timeout_interval = microsec_2_cycles(2*timeout_us);
1631 bcp->max_bau_concurrent = max_bau_concurrent;
1632 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1633 bcp->plugged_delay = plugged_delay;
1634 bcp->plugsb4reset = plugsb4reset;
1635 bcp->timeoutsb4reset = timeoutsb4reset;
1636 bcp->ipi_reset_limit = ipi_reset_limit;
1637 bcp->complete_threshold = complete_threshold;
1638 bcp->congested_response_us = congested_response_us;
1639 bcp->congested_reps = congested_reps;
1640 bcp->congested_period = congested_period;
1646 * Initialization of BAU-related structures
1648 static int __init uv_bau_init(void)
1657 if (!is_uv_system())
1663 for_each_possible_cpu(cur_cpu)
1664 zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu),
1665 GFP_KERNEL, cpu_to_node(cur_cpu));
1667 uv_nshift = uv_hub_info->m_val;
1668 uv_mmask = (1UL << uv_hub_info->m_val) - 1;
1669 nuvhubs = uv_num_possible_blades();
1670 spin_lock_init(&disable_lock);
1671 congested_cycles = microsec_2_cycles(congested_response_us);
1673 uv_partition_base_pnode = 0x7fffffff;
1674 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1675 if (uv_blade_nr_possible_cpus(uvhub) &&
1676 (uv_blade_to_pnode(uvhub) < uv_partition_base_pnode))
1677 uv_partition_base_pnode = uv_blade_to_pnode(uvhub);
1680 if (uv_init_per_cpu(nuvhubs, uv_partition_base_pnode)) {
1685 vector = UV_BAU_MESSAGE;
1686 for_each_possible_blade(uvhub)
1687 if (uv_blade_nr_possible_cpus(uvhub))
1688 uv_init_uvhub(uvhub, vector, uv_partition_base_pnode);
1690 uv_enable_timeouts();
1691 alloc_intr_gate(vector, uv_bau_message_intr1);
1693 for_each_possible_blade(uvhub) {
1694 if (uv_blade_nr_possible_cpus(uvhub)) {
1695 pnode = uv_blade_to_pnode(uvhub);
1697 uv_write_global_mmr64(pnode,
1698 UVH_LB_BAU_SB_ACTIVATION_CONTROL,
1699 ((unsigned long)1 << 63));
1700 mmr = 1; /* should be 1 to broadcast to both sockets */
1701 uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST,
1708 core_initcall(uv_bau_init);
1709 fs_initcall(uv_ptc_init);