2 * SGI UltraViolet TLB flush routines.
4 * (c) 2008-2011 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
[] = {
39 static int timeout_us
;
41 static int baudisabled
;
42 static spinlock_t disable_lock
;
43 static cycles_t congested_cycles
;
46 static int max_concurr
= MAX_BAU_CONCURRENT
;
47 static int max_concurr_const
= MAX_BAU_CONCURRENT
;
48 static int plugged_delay
= PLUGGED_DELAY
;
49 static int plugsb4reset
= PLUGSB4RESET
;
50 static int timeoutsb4reset
= TIMEOUTSB4RESET
;
51 static int ipi_reset_limit
= IPI_RESET_LIMIT
;
52 static int complete_threshold
= COMPLETE_THRESHOLD
;
53 static int congested_respns_us
= CONGESTED_RESPONSE_US
;
54 static int congested_reps
= CONGESTED_REPS
;
55 static int congested_period
= CONGESTED_PERIOD
;
57 static struct tunables tunables
[] = {
58 {&max_concurr
, MAX_BAU_CONCURRENT
}, /* must be [0] */
59 {&plugged_delay
, PLUGGED_DELAY
},
60 {&plugsb4reset
, PLUGSB4RESET
},
61 {&timeoutsb4reset
, TIMEOUTSB4RESET
},
62 {&ipi_reset_limit
, IPI_RESET_LIMIT
},
63 {&complete_threshold
, COMPLETE_THRESHOLD
},
64 {&congested_respns_us
, CONGESTED_RESPONSE_US
},
65 {&congested_reps
, CONGESTED_REPS
},
66 {&congested_period
, CONGESTED_PERIOD
}
69 static struct dentry
*tunables_dir
;
70 static struct dentry
*tunables_file
;
72 /* these correspond to the statistics printed by ptc_seq_show() */
73 static char *stat_description
[] = {
74 "sent: number of shootdown messages sent",
75 "stime: time spent sending messages",
76 "numuvhubs: number of hubs targeted with shootdown",
77 "numuvhubs16: number times 16 or more hubs targeted",
78 "numuvhubs8: number times 8 or more hubs targeted",
79 "numuvhubs4: number times 4 or more hubs targeted",
80 "numuvhubs2: number times 2 or more hubs targeted",
81 "numuvhubs1: number times 1 hub targeted",
82 "numcpus: number of cpus targeted with shootdown",
83 "dto: number of destination timeouts",
84 "retries: destination timeout retries sent",
85 "rok: : destination timeouts successfully retried",
86 "resetp: ipi-style resource resets for plugs",
87 "resett: ipi-style resource resets for timeouts",
88 "giveup: fall-backs to ipi-style shootdowns",
89 "sto: number of source timeouts",
90 "bz: number of stay-busy's",
91 "throt: number times spun in throttle",
92 "swack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE",
93 "recv: shootdown messages received",
94 "rtime: time spent processing messages",
95 "all: shootdown all-tlb messages",
96 "one: shootdown one-tlb messages",
97 "mult: interrupts that found multiple messages",
98 "none: interrupts that found no messages",
99 "retry: number of retry messages processed",
100 "canc: number messages canceled by retries",
101 "nocan: number retries that found nothing to cancel",
102 "reset: number of ipi-style reset requests processed",
103 "rcan: number messages canceled by reset requests",
104 "disable: number times use of the BAU was disabled",
105 "enable: number times use of the BAU was re-enabled"
109 setup_nobau(char *arg
)
114 early_param("nobau", setup_nobau
);
116 /* base pnode in this partition */
117 static int uv_base_pnode __read_mostly
;
119 static DEFINE_PER_CPU(struct ptc_stats
, ptcstats
);
120 static DEFINE_PER_CPU(struct bau_control
, bau_control
);
121 static DEFINE_PER_CPU(cpumask_var_t
, uv_flush_tlb_mask
);
124 * Determine the first node on a uvhub. 'Nodes' are used for kernel
127 static int __init
uvhub_to_first_node(int uvhub
)
131 for_each_online_node(node
) {
132 b
= uv_node_to_blade_id(node
);
140 * Determine the apicid of the first cpu on a uvhub.
142 static int __init
uvhub_to_first_apicid(int uvhub
)
146 for_each_present_cpu(cpu
)
147 if (uvhub
== uv_cpu_to_blade_id(cpu
))
148 return per_cpu(x86_cpu_to_apicid
, cpu
);
153 * Free a software acknowledge hardware resource by clearing its Pending
154 * bit. This will return a reply to the sender.
155 * If the message has timed out, a reply has already been sent by the
156 * hardware but the resource has not been released. In that case our
157 * clear of the Timeout bit (as well) will free the resource. No reply will
158 * be sent (the hardware will only do one reply per message).
160 static void reply_to_message(struct msg_desc
*mdp
, struct bau_control
*bcp
)
163 struct bau_pq_entry
*msg
;
166 if (!msg
->canceled
) {
167 dw
= (msg
->swack_vec
<< UV_SW_ACK_NPENDING
) | msg
->swack_vec
;
168 write_mmr_sw_ack(dw
);
175 * Process the receipt of a RETRY message
177 static void bau_process_retry_msg(struct msg_desc
*mdp
,
178 struct bau_control
*bcp
)
181 int cancel_count
= 0;
182 unsigned long msg_res
;
183 unsigned long mmr
= 0;
184 struct bau_pq_entry
*msg
= mdp
->msg
;
185 struct bau_pq_entry
*msg2
;
186 struct ptc_stats
*stat
= bcp
->statp
;
190 * cancel any message from msg+1 to the retry itself
192 for (msg2
= msg
+1, i
= 0; i
< DEST_Q_SIZE
; msg2
++, i
++) {
193 if (msg2
> mdp
->queue_last
)
194 msg2
= mdp
->queue_first
;
198 /* same conditions for cancellation as do_reset */
199 if ((msg2
->replied_to
== 0) && (msg2
->canceled
== 0) &&
200 (msg2
->swack_vec
) && ((msg2
->swack_vec
&
201 msg
->swack_vec
) == 0) &&
202 (msg2
->sending_cpu
== msg
->sending_cpu
) &&
203 (msg2
->msg_type
!= MSG_NOOP
)) {
204 mmr
= read_mmr_sw_ack();
205 msg_res
= msg2
->swack_vec
;
207 * This is a message retry; clear the resources held
208 * by the previous message only if they timed out.
209 * If it has not timed out we have an unexpected
210 * situation to report.
212 if (mmr
& (msg_res
<< UV_SW_ACK_NPENDING
)) {
215 * is the resource timed out?
216 * make everyone ignore the cancelled message.
221 mr
= (msg_res
<< UV_SW_ACK_NPENDING
) | msg_res
;
222 write_mmr_sw_ack(mr
);
227 stat
->d_nocanceled
++;
231 * Do all the things a cpu should do for a TLB shootdown message.
232 * Other cpu's may come here at the same time for this message.
234 static void bau_process_message(struct msg_desc
*mdp
,
235 struct bau_control
*bcp
)
237 short socket_ack_count
= 0;
239 struct atomic_short
*asp
;
240 struct ptc_stats
*stat
= bcp
->statp
;
241 struct bau_pq_entry
*msg
= mdp
->msg
;
242 struct bau_control
*smaster
= bcp
->socket_master
;
245 * This must be a normal message, or retry of a normal message
247 if (msg
->address
== TLB_FLUSH_ALL
) {
251 __flush_tlb_one(msg
->address
);
257 * One cpu on each uvhub has the additional job on a RETRY
258 * of releasing the resource held by the message that is
259 * being retried. That message is identified by sending
262 if (msg
->msg_type
== MSG_RETRY
&& bcp
== bcp
->uvhub_master
)
263 bau_process_retry_msg(mdp
, bcp
);
266 * This is a swack message, so we have to reply to it.
267 * Count each responding cpu on the socket. This avoids
268 * pinging the count's cache line back and forth between
271 sp
= &smaster
->socket_acknowledge_count
[mdp
->msg_slot
];
272 asp
= (struct atomic_short
*)sp
;
273 socket_ack_count
= atom_asr(1, asp
);
274 if (socket_ack_count
== bcp
->cpus_in_socket
) {
277 * Both sockets dump their completed count total into
278 * the message's count.
280 smaster
->socket_acknowledge_count
[mdp
->msg_slot
] = 0;
281 asp
= (struct atomic_short
*)&msg
->acknowledge_count
;
282 msg_ack_count
= atom_asr(socket_ack_count
, asp
);
284 if (msg_ack_count
== bcp
->cpus_in_uvhub
) {
286 * All cpus in uvhub saw it; reply
288 reply_to_message(mdp
, bcp
);
296 * Determine the first cpu on a pnode.
298 static int pnode_to_first_cpu(int pnode
, struct bau_control
*smaster
)
301 struct hub_and_pnode
*hpp
;
303 for_each_present_cpu(cpu
) {
304 hpp
= &smaster
->thp
[cpu
];
305 if (pnode
== hpp
->pnode
)
312 * Last resort when we get a large number of destination timeouts is
313 * to clear resources held by a given cpu.
314 * Do this with IPI so that all messages in the BAU message queue
315 * can be identified by their nonzero swack_vec field.
317 * This is entered for a single cpu on the uvhub.
318 * The sender want's this uvhub to free a specific message's
321 static void do_reset(void *ptr
)
324 struct bau_control
*bcp
= &per_cpu(bau_control
, smp_processor_id());
325 struct reset_args
*rap
= (struct reset_args
*)ptr
;
326 struct bau_pq_entry
*msg
;
327 struct ptc_stats
*stat
= bcp
->statp
;
331 * We're looking for the given sender, and
332 * will free its swack resource.
333 * If all cpu's finally responded after the timeout, its
334 * message 'replied_to' was set.
336 for (msg
= bcp
->queue_first
, i
= 0; i
< DEST_Q_SIZE
; msg
++, i
++) {
337 unsigned long msg_res
;
338 /* do_reset: same conditions for cancellation as
339 bau_process_retry_msg() */
340 if ((msg
->replied_to
== 0) &&
341 (msg
->canceled
== 0) &&
342 (msg
->sending_cpu
== rap
->sender
) &&
344 (msg
->msg_type
!= MSG_NOOP
)) {
348 * make everyone else ignore this message
352 * only reset the resource if it is still pending
354 mmr
= read_mmr_sw_ack();
355 msg_res
= msg
->swack_vec
;
356 mr
= (msg_res
<< UV_SW_ACK_NPENDING
) | msg_res
;
359 write_mmr_sw_ack(mr
);
367 * Use IPI to get all target uvhubs to release resources held by
368 * a given sending cpu number.
370 static void reset_with_ipi(struct pnmask
*distribution
, struct bau_control
*bcp
)
375 int sender
= bcp
->cpu
;
376 cpumask_t
*mask
= bcp
->uvhub_master
->cpumask
;
377 struct bau_control
*smaster
= bcp
->socket_master
;
378 struct reset_args reset_args
;
380 reset_args
.sender
= sender
;
382 /* find a single cpu for each uvhub in this distribution mask */
383 maskbits
= sizeof(struct pnmask
) * BITSPERBYTE
;
384 /* each bit is a pnode relative to the partition base pnode */
385 for (pnode
= 0; pnode
< maskbits
; pnode
++) {
387 if (!bau_uvhub_isset(pnode
, distribution
))
389 apnode
= pnode
+ bcp
->partition_base_pnode
;
390 cpu
= pnode_to_first_cpu(apnode
, smaster
);
394 /* IPI all cpus; preemption is already disabled */
395 smp_call_function_many(mask
, do_reset
, (void *)&reset_args
, 1);
399 static inline unsigned long cycles_2_us(unsigned long long cyc
)
401 unsigned long long ns
;
403 int cpu
= smp_processor_id();
405 ns
= (cyc
* per_cpu(cyc2ns
, cpu
)) >> CYC2NS_SCALE_FACTOR
;
411 * wait for all cpus on this hub to finish their sends and go quiet
412 * leaves uvhub_quiesce set so that no new broadcasts are started by
413 * bau_flush_send_and_wait()
415 static inline void quiesce_local_uvhub(struct bau_control
*hmaster
)
417 atom_asr(1, (struct atomic_short
*)&hmaster
->uvhub_quiesce
);
421 * mark this quiet-requestor as done
423 static inline void end_uvhub_quiesce(struct bau_control
*hmaster
)
425 atom_asr(-1, (struct atomic_short
*)&hmaster
->uvhub_quiesce
);
428 static unsigned long uv1_read_status(unsigned long mmr_offset
, int right_shift
)
430 unsigned long descriptor_status
;
432 descriptor_status
= uv_read_local_mmr(mmr_offset
);
433 descriptor_status
>>= right_shift
;
434 descriptor_status
&= UV_ACT_STATUS_MASK
;
435 return descriptor_status
;
439 * Wait for completion of a broadcast software ack message
440 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
442 static int uv1_wait_completion(struct bau_desc
*bau_desc
,
443 unsigned long mmr_offset
, int right_shift
,
444 struct bau_control
*bcp
, long try)
446 unsigned long descriptor_status
;
448 struct ptc_stats
*stat
= bcp
->statp
;
450 descriptor_status
= uv1_read_status(mmr_offset
, right_shift
);
451 /* spin on the status MMR, waiting for it to go idle */
452 while ((descriptor_status
!= DS_IDLE
)) {
454 * Our software ack messages may be blocked because
455 * there are no swack resources available. As long
456 * as none of them has timed out hardware will NACK
457 * our message and its state will stay IDLE.
459 if (descriptor_status
== DS_SOURCE_TIMEOUT
) {
462 } else if (descriptor_status
== DS_DESTINATION_TIMEOUT
) {
467 * Our retries may be blocked by all destination
468 * swack resources being consumed, and a timeout
469 * pending. In that case hardware returns the
470 * ERROR that looks like a destination timeout.
472 if (cycles_2_us(ttm
- bcp
->send_message
) < timeout_us
) {
473 bcp
->conseccompletes
= 0;
474 return FLUSH_RETRY_PLUGGED
;
477 bcp
->conseccompletes
= 0;
478 return FLUSH_RETRY_TIMEOUT
;
481 * descriptor_status is still BUSY
485 descriptor_status
= uv1_read_status(mmr_offset
, right_shift
);
487 bcp
->conseccompletes
++;
488 return FLUSH_COMPLETE
;
492 * UV2 has an extra bit of status in the ACTIVATION_STATUS_2 register.
494 static unsigned long uv2_read_status(unsigned long offset
, int rshft
, int cpu
)
496 unsigned long descriptor_status
;
497 unsigned long descriptor_status2
;
499 descriptor_status
= ((read_lmmr(offset
) >> rshft
) & UV_ACT_STATUS_MASK
);
500 descriptor_status2
= (read_mmr_uv2_status() >> cpu
) & 0x1UL
;
501 descriptor_status
= (descriptor_status
<< 1) | descriptor_status2
;
502 return descriptor_status
;
505 static int uv2_wait_completion(struct bau_desc
*bau_desc
,
506 unsigned long mmr_offset
, int right_shift
,
507 struct bau_control
*bcp
, long try)
509 unsigned long descriptor_stat
;
511 int cpu
= bcp
->uvhub_cpu
;
512 struct ptc_stats
*stat
= bcp
->statp
;
514 descriptor_stat
= uv2_read_status(mmr_offset
, right_shift
, cpu
);
516 /* spin on the status MMR, waiting for it to go idle */
517 while (descriptor_stat
!= UV2H_DESC_IDLE
) {
519 * Our software ack messages may be blocked because
520 * there are no swack resources available. As long
521 * as none of them has timed out hardware will NACK
522 * our message and its state will stay IDLE.
524 if ((descriptor_stat
== UV2H_DESC_SOURCE_TIMEOUT
) ||
525 (descriptor_stat
== UV2H_DESC_DEST_STRONG_NACK
) ||
526 (descriptor_stat
== UV2H_DESC_DEST_PUT_ERR
)) {
529 } else if (descriptor_stat
== UV2H_DESC_DEST_TIMEOUT
) {
533 * Our retries may be blocked by all destination
534 * swack resources being consumed, and a timeout
535 * pending. In that case hardware returns the
536 * ERROR that looks like a destination timeout.
538 if (cycles_2_us(ttm
- bcp
->send_message
) < timeout_us
) {
539 bcp
->conseccompletes
= 0;
540 return FLUSH_RETRY_PLUGGED
;
542 bcp
->conseccompletes
= 0;
543 return FLUSH_RETRY_TIMEOUT
;
546 * descriptor_stat is still BUSY
550 descriptor_stat
= uv2_read_status(mmr_offset
, right_shift
, cpu
);
552 bcp
->conseccompletes
++;
553 return FLUSH_COMPLETE
;
557 * There are 2 status registers; each and array[32] of 2 bits. Set up for
558 * which register to read and position in that register based on cpu in
561 static int wait_completion(struct bau_desc
*bau_desc
,
562 struct bau_control
*bcp
, long try)
565 unsigned long mmr_offset
;
566 int cpu
= bcp
->uvhub_cpu
;
568 if (cpu
< UV_CPUS_PER_AS
) {
569 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_0
;
570 right_shift
= cpu
* UV_ACT_STATUS_SIZE
;
572 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_1
;
573 right_shift
= ((cpu
- UV_CPUS_PER_AS
) * UV_ACT_STATUS_SIZE
);
576 if (bcp
->uvhub_version
== 1)
577 return uv1_wait_completion(bau_desc
, mmr_offset
, right_shift
,
580 return uv2_wait_completion(bau_desc
, mmr_offset
, right_shift
,
584 static inline cycles_t
sec_2_cycles(unsigned long sec
)
589 ns
= sec
* 1000000000;
590 cyc
= (ns
<< CYC2NS_SCALE_FACTOR
)/(per_cpu(cyc2ns
, smp_processor_id()));
595 * Our retries are blocked by all destination sw ack resources being
596 * in use, and a timeout is pending. In that case hardware immediately
597 * returns the ERROR that looks like a destination timeout.
599 static void destination_plugged(struct bau_desc
*bau_desc
,
600 struct bau_control
*bcp
,
601 struct bau_control
*hmaster
, struct ptc_stats
*stat
)
603 udelay(bcp
->plugged_delay
);
604 bcp
->plugged_tries
++;
606 if (bcp
->plugged_tries
>= bcp
->plugsb4reset
) {
607 bcp
->plugged_tries
= 0;
609 quiesce_local_uvhub(hmaster
);
611 spin_lock(&hmaster
->queue_lock
);
612 reset_with_ipi(&bau_desc
->distribution
, bcp
);
613 spin_unlock(&hmaster
->queue_lock
);
615 end_uvhub_quiesce(hmaster
);
618 stat
->s_resets_plug
++;
622 static void destination_timeout(struct bau_desc
*bau_desc
,
623 struct bau_control
*bcp
, struct bau_control
*hmaster
,
624 struct ptc_stats
*stat
)
626 hmaster
->max_concurr
= 1;
627 bcp
->timeout_tries
++;
628 if (bcp
->timeout_tries
>= bcp
->timeoutsb4reset
) {
629 bcp
->timeout_tries
= 0;
631 quiesce_local_uvhub(hmaster
);
633 spin_lock(&hmaster
->queue_lock
);
634 reset_with_ipi(&bau_desc
->distribution
, bcp
);
635 spin_unlock(&hmaster
->queue_lock
);
637 end_uvhub_quiesce(hmaster
);
640 stat
->s_resets_timeout
++;
645 * Completions are taking a very long time due to a congested numalink
648 static void disable_for_congestion(struct bau_control
*bcp
,
649 struct ptc_stats
*stat
)
651 /* let only one cpu do this disabling */
652 spin_lock(&disable_lock
);
654 if (!baudisabled
&& bcp
->period_requests
&&
655 ((bcp
->period_time
/ bcp
->period_requests
) > congested_cycles
)) {
657 struct bau_control
*tbcp
;
658 /* it becomes this cpu's job to turn on the use of the
661 bcp
->set_bau_off
= 1;
662 bcp
->set_bau_on_time
= get_cycles();
663 bcp
->set_bau_on_time
+= sec_2_cycles(bcp
->cong_period
);
664 stat
->s_bau_disabled
++;
665 for_each_present_cpu(tcpu
) {
666 tbcp
= &per_cpu(bau_control
, tcpu
);
667 tbcp
->baudisabled
= 1;
671 spin_unlock(&disable_lock
);
674 static void count_max_concurr(int stat
, struct bau_control
*bcp
,
675 struct bau_control
*hmaster
)
677 bcp
->plugged_tries
= 0;
678 bcp
->timeout_tries
= 0;
679 if (stat
!= FLUSH_COMPLETE
)
681 if (bcp
->conseccompletes
<= bcp
->complete_threshold
)
683 if (hmaster
->max_concurr
>= hmaster
->max_concurr_const
)
685 hmaster
->max_concurr
++;
688 static void record_send_stats(cycles_t time1
, cycles_t time2
,
689 struct bau_control
*bcp
, struct ptc_stats
*stat
,
690 int completion_status
, int try)
695 elapsed
= time2
- time1
;
696 stat
->s_time
+= elapsed
;
698 if ((completion_status
== FLUSH_COMPLETE
) && (try == 1)) {
699 bcp
->period_requests
++;
700 bcp
->period_time
+= elapsed
;
701 if ((elapsed
> congested_cycles
) &&
702 (bcp
->period_requests
> bcp
->cong_reps
))
703 disable_for_congestion(bcp
, stat
);
708 if (completion_status
== FLUSH_COMPLETE
&& try > 1)
710 else if (completion_status
== FLUSH_GIVEUP
)
715 * Because of a uv1 hardware bug only a limited number of concurrent
716 * requests can be made.
718 static void uv1_throttle(struct bau_control
*hmaster
, struct ptc_stats
*stat
)
720 spinlock_t
*lock
= &hmaster
->uvhub_lock
;
723 v
= &hmaster
->active_descriptor_count
;
724 if (!atomic_inc_unless_ge(lock
, v
, hmaster
->max_concurr
)) {
728 } while (!atomic_inc_unless_ge(lock
, v
, hmaster
->max_concurr
));
733 * Handle the completion status of a message send.
735 static void handle_cmplt(int completion_status
, struct bau_desc
*bau_desc
,
736 struct bau_control
*bcp
, struct bau_control
*hmaster
,
737 struct ptc_stats
*stat
)
739 if (completion_status
== FLUSH_RETRY_PLUGGED
)
740 destination_plugged(bau_desc
, bcp
, hmaster
, stat
);
741 else if (completion_status
== FLUSH_RETRY_TIMEOUT
)
742 destination_timeout(bau_desc
, bcp
, hmaster
, stat
);
746 * Send a broadcast and wait for it to complete.
748 * The flush_mask contains the cpus the broadcast is to be sent to including
749 * cpus that are on the local uvhub.
751 * Returns 0 if all flushing represented in the mask was done.
752 * Returns 1 if it gives up entirely and the original cpu mask is to be
753 * returned to the kernel.
755 int uv_flush_send_and_wait(struct bau_desc
*bau_desc
,
756 struct cpumask
*flush_mask
, struct bau_control
*bcp
)
759 int completion_stat
= 0;
765 struct ptc_stats
*stat
= bcp
->statp
;
766 struct bau_control
*hmaster
= bcp
->uvhub_master
;
767 struct uv1_bau_msg_header
*uv1_hdr
= NULL
;
768 struct uv2_bau_msg_header
*uv2_hdr
= NULL
;
770 if (bcp
->uvhub_version
== 1) {
772 uv1_throttle(hmaster
, stat
);
773 uv1_hdr
= &bau_desc
->header
.uv1_hdr
;
775 uv2_hdr
= &bau_desc
->header
.uv2_hdr
;
777 while (hmaster
->uvhub_quiesce
)
780 time1
= get_cycles();
784 uv1_hdr
->msg_type
= MSG_REGULAR
;
786 uv2_hdr
->msg_type
= MSG_REGULAR
;
787 seq_number
= bcp
->message_number
++;
790 uv1_hdr
->msg_type
= MSG_RETRY
;
792 uv2_hdr
->msg_type
= MSG_RETRY
;
793 stat
->s_retry_messages
++;
797 uv1_hdr
->sequence
= seq_number
;
799 uv2_hdr
->sequence
= seq_number
;
800 index
= (1UL << AS_PUSH_SHIFT
) | bcp
->uvhub_cpu
;
801 bcp
->send_message
= get_cycles();
803 write_mmr_activation(index
);
806 completion_stat
= wait_completion(bau_desc
, bcp
, try);
808 handle_cmplt(completion_stat
, bau_desc
, bcp
, hmaster
, stat
);
810 if (bcp
->ipi_attempts
>= bcp
->ipi_reset_limit
) {
811 bcp
->ipi_attempts
= 0;
812 completion_stat
= FLUSH_GIVEUP
;
816 } while ((completion_stat
== FLUSH_RETRY_PLUGGED
) ||
817 (completion_stat
== FLUSH_RETRY_TIMEOUT
));
819 time2
= get_cycles();
821 count_max_concurr(completion_stat
, bcp
, hmaster
);
823 while (hmaster
->uvhub_quiesce
)
826 atomic_dec(&hmaster
->active_descriptor_count
);
828 record_send_stats(time1
, time2
, bcp
, stat
, completion_stat
, try);
830 if (completion_stat
== FLUSH_GIVEUP
)
836 * The BAU is disabled. When the disabled time period has expired, the cpu
837 * that disabled it must re-enable it.
838 * Return 0 if it is re-enabled for all cpus.
840 static int check_enable(struct bau_control
*bcp
, struct ptc_stats
*stat
)
843 struct bau_control
*tbcp
;
845 if (bcp
->set_bau_off
) {
846 if (get_cycles() >= bcp
->set_bau_on_time
) {
847 stat
->s_bau_reenabled
++;
849 for_each_present_cpu(tcpu
) {
850 tbcp
= &per_cpu(bau_control
, tcpu
);
851 tbcp
->baudisabled
= 0;
852 tbcp
->period_requests
= 0;
853 tbcp
->period_time
= 0;
861 static void record_send_statistics(struct ptc_stats
*stat
, int locals
, int hubs
,
862 int remotes
, struct bau_desc
*bau_desc
)
865 stat
->s_ntargcpu
+= remotes
+ locals
;
866 stat
->s_ntargremotes
+= remotes
;
867 stat
->s_ntarglocals
+= locals
;
869 /* uvhub statistics */
870 hubs
= bau_uvhub_weight(&bau_desc
->distribution
);
872 stat
->s_ntarglocaluvhub
++;
873 stat
->s_ntargremoteuvhub
+= (hubs
- 1);
875 stat
->s_ntargremoteuvhub
+= hubs
;
877 stat
->s_ntarguvhub
+= hubs
;
880 stat
->s_ntarguvhub16
++;
882 stat
->s_ntarguvhub8
++;
884 stat
->s_ntarguvhub4
++;
886 stat
->s_ntarguvhub2
++;
888 stat
->s_ntarguvhub1
++;
892 * Translate a cpu mask to the uvhub distribution mask in the BAU
893 * activation descriptor.
895 static int set_distrib_bits(struct cpumask
*flush_mask
, struct bau_control
*bcp
,
896 struct bau_desc
*bau_desc
, int *localsp
, int *remotesp
)
901 struct hub_and_pnode
*hpp
;
903 for_each_cpu(cpu
, flush_mask
) {
905 * The distribution vector is a bit map of pnodes, relative
906 * to the partition base pnode (and the partition base nasid
908 * Translate cpu to pnode and hub using a local memory array.
910 hpp
= &bcp
->socket_master
->thp
[cpu
];
911 pnode
= hpp
->pnode
- bcp
->partition_base_pnode
;
912 bau_uvhub_set(pnode
, &bau_desc
->distribution
);
914 if (hpp
->uvhub
== bcp
->uvhub
)
925 * globally purge translation cache of a virtual address or all TLB's
926 * @cpumask: mask of all cpu's in which the address is to be removed
927 * @mm: mm_struct containing virtual address range
928 * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
929 * @cpu: the current cpu
931 * This is the entry point for initiating any UV global TLB shootdown.
933 * Purges the translation caches of all specified processors of the given
934 * virtual address, or purges all TLB's on specified processors.
936 * The caller has derived the cpumask from the mm_struct. This function
937 * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
939 * The cpumask is converted into a uvhubmask of the uvhubs containing
942 * Note that this function should be called with preemption disabled.
944 * Returns NULL if all remote flushing was done.
945 * Returns pointer to cpumask if some remote flushing remains to be
946 * done. The returned pointer is valid till preemption is re-enabled.
948 const struct cpumask
*uv_flush_tlb_others(const struct cpumask
*cpumask
,
949 struct mm_struct
*mm
, unsigned long va
,
955 struct bau_desc
*bau_desc
;
956 struct cpumask
*flush_mask
;
957 struct ptc_stats
*stat
;
958 struct bau_control
*bcp
;
960 /* kernel was booted 'nobau' */
964 bcp
= &per_cpu(bau_control
, cpu
);
967 /* bau was disabled due to slow response */
968 if (bcp
->baudisabled
) {
969 if (check_enable(bcp
, stat
))
974 * Each sending cpu has a per-cpu mask which it fills from the caller's
975 * cpu mask. All cpus are converted to uvhubs and copied to the
976 * activation descriptor.
978 flush_mask
= (struct cpumask
*)per_cpu(uv_flush_tlb_mask
, cpu
);
979 /* don't actually do a shootdown of the local cpu */
980 cpumask_andnot(flush_mask
, cpumask
, cpumask_of(cpu
));
982 if (cpu_isset(cpu
, *cpumask
))
985 bau_desc
= bcp
->descriptor_base
;
986 bau_desc
+= (ITEMS_PER_DESC
* bcp
->uvhub_cpu
);
987 bau_uvhubs_clear(&bau_desc
->distribution
, UV_DISTRIBUTION_SIZE
);
988 if (set_distrib_bits(flush_mask
, bcp
, bau_desc
, &locals
, &remotes
))
991 record_send_statistics(stat
, locals
, hubs
, remotes
, bau_desc
);
993 bau_desc
->payload
.address
= va
;
994 bau_desc
->payload
.sending_cpu
= cpu
;
996 * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
997 * or 1 if it gave up and the original cpumask should be returned.
999 if (!uv_flush_send_and_wait(bau_desc
, flush_mask
, bcp
))
1006 * The BAU message interrupt comes here. (registered by set_intr_gate)
1009 * We received a broadcast assist message.
1011 * Interrupts are disabled; this interrupt could represent
1012 * the receipt of several messages.
1014 * All cores/threads on this hub get this interrupt.
1015 * The last one to see it does the software ack.
1016 * (the resource will not be freed until noninterruptable cpus see this
1017 * interrupt; hardware may timeout the s/w ack and reply ERROR)
1019 void uv_bau_message_interrupt(struct pt_regs
*regs
)
1022 cycles_t time_start
;
1023 struct bau_pq_entry
*msg
;
1024 struct bau_control
*bcp
;
1025 struct ptc_stats
*stat
;
1026 struct msg_desc msgdesc
;
1028 time_start
= get_cycles();
1030 bcp
= &per_cpu(bau_control
, smp_processor_id());
1033 msgdesc
.queue_first
= bcp
->queue_first
;
1034 msgdesc
.queue_last
= bcp
->queue_last
;
1036 msg
= bcp
->bau_msg_head
;
1037 while (msg
->swack_vec
) {
1040 msgdesc
.msg_slot
= msg
- msgdesc
.queue_first
;
1041 msgdesc
.swack_slot
= ffs(msg
->swack_vec
) - 1;
1043 bau_process_message(&msgdesc
, bcp
);
1046 if (msg
> msgdesc
.queue_last
)
1047 msg
= msgdesc
.queue_first
;
1048 bcp
->bau_msg_head
= msg
;
1050 stat
->d_time
+= (get_cycles() - time_start
);
1060 * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1061 * shootdown message timeouts enabled. The timeout does not cause
1062 * an interrupt, but causes an error message to be returned to
1065 static void __init
enable_timeouts(void)
1070 unsigned long mmr_image
;
1072 nuvhubs
= uv_num_possible_blades();
1074 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
1075 if (!uv_blade_nr_possible_cpus(uvhub
))
1078 pnode
= uv_blade_to_pnode(uvhub
);
1079 mmr_image
= read_mmr_misc_control(pnode
);
1081 * Set the timeout period and then lock it in, in three
1082 * steps; captures and locks in the period.
1084 * To program the period, the SOFT_ACK_MODE must be off.
1086 mmr_image
&= ~(1L << SOFTACK_MSHIFT
);
1087 write_mmr_misc_control(pnode
, mmr_image
);
1089 * Set the 4-bit period.
1091 mmr_image
&= ~((unsigned long)0xf << SOFTACK_PSHIFT
);
1092 mmr_image
|= (SOFTACK_TIMEOUT_PERIOD
<< SOFTACK_PSHIFT
);
1093 write_mmr_misc_control(pnode
, mmr_image
);
1096 * Subsequent reversals of the timebase bit (3) cause an
1097 * immediate timeout of one or all INTD resources as
1098 * indicated in bits 2:0 (7 causes all of them to timeout).
1100 mmr_image
|= (1L << SOFTACK_MSHIFT
);
1102 mmr_image
&= ~(1L << UV2_LEG_SHFT
);
1103 mmr_image
|= (1L << UV2_EXT_SHFT
);
1105 write_mmr_misc_control(pnode
, mmr_image
);
1109 static void *ptc_seq_start(struct seq_file
*file
, loff_t
*offset
)
1111 if (*offset
< num_possible_cpus())
1116 static void *ptc_seq_next(struct seq_file
*file
, void *data
, loff_t
*offset
)
1119 if (*offset
< num_possible_cpus())
1124 static void ptc_seq_stop(struct seq_file
*file
, void *data
)
1128 static inline unsigned long long usec_2_cycles(unsigned long microsec
)
1131 unsigned long long cyc
;
1133 ns
= microsec
* 1000;
1134 cyc
= (ns
<< CYC2NS_SCALE_FACTOR
)/(per_cpu(cyc2ns
, smp_processor_id()));
1139 * Display the statistics thru /proc/sgi_uv/ptc_statistics
1140 * 'data' points to the cpu number
1141 * Note: see the descriptions in stat_description[].
1143 static int ptc_seq_show(struct seq_file
*file
, void *data
)
1145 struct ptc_stats
*stat
;
1148 cpu
= *(loff_t
*)data
;
1151 "# cpu sent stime self locals remotes ncpus localhub ");
1153 "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1155 "numuvhubs4 numuvhubs2 numuvhubs1 dto retries rok ");
1157 "resetp resett giveup sto bz throt swack recv rtime ");
1159 "all one mult none retry canc nocan reset rcan ");
1161 "disable enable\n");
1163 if (cpu
< num_possible_cpus() && cpu_online(cpu
)) {
1164 stat
= &per_cpu(ptcstats
, cpu
);
1165 /* source side statistics */
1167 "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1168 cpu
, stat
->s_requestor
, cycles_2_us(stat
->s_time
),
1169 stat
->s_ntargself
, stat
->s_ntarglocals
,
1170 stat
->s_ntargremotes
, stat
->s_ntargcpu
,
1171 stat
->s_ntarglocaluvhub
, stat
->s_ntargremoteuvhub
,
1172 stat
->s_ntarguvhub
, stat
->s_ntarguvhub16
);
1173 seq_printf(file
, "%ld %ld %ld %ld %ld ",
1174 stat
->s_ntarguvhub8
, stat
->s_ntarguvhub4
,
1175 stat
->s_ntarguvhub2
, stat
->s_ntarguvhub1
,
1177 seq_printf(file
, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1178 stat
->s_retry_messages
, stat
->s_retriesok
,
1179 stat
->s_resets_plug
, stat
->s_resets_timeout
,
1180 stat
->s_giveup
, stat
->s_stimeout
,
1181 stat
->s_busy
, stat
->s_throttles
);
1183 /* destination side statistics */
1185 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1186 read_gmmr_sw_ack(uv_cpu_to_pnode(cpu
)),
1187 stat
->d_requestee
, cycles_2_us(stat
->d_time
),
1188 stat
->d_alltlb
, stat
->d_onetlb
, stat
->d_multmsg
,
1189 stat
->d_nomsg
, stat
->d_retries
, stat
->d_canceled
,
1190 stat
->d_nocanceled
, stat
->d_resets
,
1192 seq_printf(file
, "%ld %ld\n",
1193 stat
->s_bau_disabled
, stat
->s_bau_reenabled
);
1199 * Display the tunables thru debugfs
1201 static ssize_t
tunables_read(struct file
*file
, char __user
*userbuf
,
1202 size_t count
, loff_t
*ppos
)
1207 buf
= kasprintf(GFP_KERNEL
, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1208 "max_concur plugged_delay plugsb4reset",
1209 "timeoutsb4reset ipi_reset_limit complete_threshold",
1210 "congested_response_us congested_reps congested_period",
1211 max_concurr
, plugged_delay
, plugsb4reset
,
1212 timeoutsb4reset
, ipi_reset_limit
, complete_threshold
,
1213 congested_respns_us
, congested_reps
, congested_period
);
1218 ret
= simple_read_from_buffer(userbuf
, count
, ppos
, buf
, strlen(buf
));
1224 * handle a write to /proc/sgi_uv/ptc_statistics
1225 * -1: reset the statistics
1226 * 0: display meaning of the statistics
1228 static ssize_t
ptc_proc_write(struct file
*file
, const char __user
*user
,
1229 size_t count
, loff_t
*data
)
1236 struct ptc_stats
*stat
;
1238 if (count
== 0 || count
> sizeof(optstr
))
1240 if (copy_from_user(optstr
, user
, count
))
1242 optstr
[count
- 1] = '\0';
1244 if (strict_strtol(optstr
, 10, &input_arg
) < 0) {
1245 printk(KERN_DEBUG
"%s is invalid\n", optstr
);
1249 if (input_arg
== 0) {
1250 elements
= sizeof(stat_description
)/sizeof(*stat_description
);
1251 printk(KERN_DEBUG
"# cpu: cpu number\n");
1252 printk(KERN_DEBUG
"Sender statistics:\n");
1253 for (i
= 0; i
< elements
; i
++)
1254 printk(KERN_DEBUG
"%s\n", stat_description
[i
]);
1255 } else if (input_arg
== -1) {
1256 for_each_present_cpu(cpu
) {
1257 stat
= &per_cpu(ptcstats
, cpu
);
1258 memset(stat
, 0, sizeof(struct ptc_stats
));
1265 static int local_atoi(const char *name
)
1272 val
= 10*val
+(*name
-'0');
1281 * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1282 * Zero values reset them to defaults.
1284 static int parse_tunables_write(struct bau_control
*bcp
, char *instr
,
1291 int e
= sizeof(tunables
) / sizeof(*tunables
);
1293 p
= instr
+ strspn(instr
, WHITESPACE
);
1295 for (; *p
; p
= q
+ strspn(q
, WHITESPACE
)) {
1296 q
= p
+ strcspn(p
, WHITESPACE
);
1302 printk(KERN_INFO
"bau tunable error: should be %d values\n", e
);
1306 p
= instr
+ strspn(instr
, WHITESPACE
);
1308 for (cnt
= 0; *p
; p
= q
+ strspn(q
, WHITESPACE
), cnt
++) {
1309 q
= p
+ strcspn(p
, WHITESPACE
);
1310 val
= local_atoi(p
);
1314 max_concurr
= MAX_BAU_CONCURRENT
;
1315 max_concurr_const
= MAX_BAU_CONCURRENT
;
1318 if (val
< 1 || val
> bcp
->cpus_in_uvhub
) {
1320 "Error: BAU max concurrent %d is invalid\n",
1325 max_concurr_const
= val
;
1329 *tunables
[cnt
].tunp
= tunables
[cnt
].deflt
;
1331 *tunables
[cnt
].tunp
= val
;
1341 * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1343 static ssize_t
tunables_write(struct file
*file
, const char __user
*user
,
1344 size_t count
, loff_t
*data
)
1349 struct bau_control
*bcp
;
1351 if (count
== 0 || count
> sizeof(instr
)-1)
1353 if (copy_from_user(instr
, user
, count
))
1356 instr
[count
] = '\0';
1359 bcp
= &per_cpu(bau_control
, cpu
);
1360 ret
= parse_tunables_write(bcp
, instr
, count
);
1365 for_each_present_cpu(cpu
) {
1366 bcp
= &per_cpu(bau_control
, cpu
);
1367 bcp
->max_concurr
= max_concurr
;
1368 bcp
->max_concurr_const
= max_concurr
;
1369 bcp
->plugged_delay
= plugged_delay
;
1370 bcp
->plugsb4reset
= plugsb4reset
;
1371 bcp
->timeoutsb4reset
= timeoutsb4reset
;
1372 bcp
->ipi_reset_limit
= ipi_reset_limit
;
1373 bcp
->complete_threshold
= complete_threshold
;
1374 bcp
->cong_response_us
= congested_respns_us
;
1375 bcp
->cong_reps
= congested_reps
;
1376 bcp
->cong_period
= congested_period
;
1381 static const struct seq_operations uv_ptc_seq_ops
= {
1382 .start
= ptc_seq_start
,
1383 .next
= ptc_seq_next
,
1384 .stop
= ptc_seq_stop
,
1385 .show
= ptc_seq_show
1388 static int ptc_proc_open(struct inode
*inode
, struct file
*file
)
1390 return seq_open(file
, &uv_ptc_seq_ops
);
1393 static int tunables_open(struct inode
*inode
, struct file
*file
)
1398 static const struct file_operations proc_uv_ptc_operations
= {
1399 .open
= ptc_proc_open
,
1401 .write
= ptc_proc_write
,
1402 .llseek
= seq_lseek
,
1403 .release
= seq_release
,
1406 static const struct file_operations tunables_fops
= {
1407 .open
= tunables_open
,
1408 .read
= tunables_read
,
1409 .write
= tunables_write
,
1410 .llseek
= default_llseek
,
1413 static int __init
uv_ptc_init(void)
1415 struct proc_dir_entry
*proc_uv_ptc
;
1417 if (!is_uv_system())
1420 proc_uv_ptc
= proc_create(UV_PTC_BASENAME
, 0444, NULL
,
1421 &proc_uv_ptc_operations
);
1423 printk(KERN_ERR
"unable to create %s proc entry\n",
1428 tunables_dir
= debugfs_create_dir(UV_BAU_TUNABLES_DIR
, NULL
);
1429 if (!tunables_dir
) {
1430 printk(KERN_ERR
"unable to create debugfs directory %s\n",
1431 UV_BAU_TUNABLES_DIR
);
1434 tunables_file
= debugfs_create_file(UV_BAU_TUNABLES_FILE
, 0600,
1435 tunables_dir
, NULL
, &tunables_fops
);
1436 if (!tunables_file
) {
1437 printk(KERN_ERR
"unable to create debugfs file %s\n",
1438 UV_BAU_TUNABLES_FILE
);
1445 * Initialize the sending side's sending buffers.
1447 static void activation_descriptor_init(int node
, int pnode
, int base_pnode
)
1456 struct bau_desc
*bau_desc
;
1457 struct bau_desc
*bd2
;
1458 struct uv1_bau_msg_header
*uv1_hdr
;
1459 struct uv2_bau_msg_header
*uv2_hdr
;
1460 struct bau_control
*bcp
;
1463 * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1464 * per cpu; and one per cpu on the uvhub (ADP_SZ)
1466 dsize
= sizeof(struct bau_desc
) * ADP_SZ
* ITEMS_PER_DESC
;
1467 bau_desc
= kmalloc_node(dsize
, GFP_KERNEL
, node
);
1470 gpa
= uv_gpa(bau_desc
);
1471 n
= uv_gpa_to_gnode(gpa
);
1472 m
= uv_gpa_to_offset(gpa
);
1476 /* the 14-bit pnode */
1477 write_mmr_descriptor_base(pnode
, (n
<< UV_DESC_PSHIFT
| m
));
1479 * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1480 * cpu even though we only use the first one; one descriptor can
1481 * describe a broadcast to 256 uv hubs.
1483 for (i
= 0, bd2
= bau_desc
; i
< (ADP_SZ
* ITEMS_PER_DESC
); i
++, bd2
++) {
1484 memset(bd2
, 0, sizeof(struct bau_desc
));
1486 uv1_hdr
= &bd2
->header
.uv1_hdr
;
1487 uv1_hdr
->swack_flag
= 1;
1489 * The base_dest_nasid set in the message header
1490 * is the nasid of the first uvhub in the partition.
1491 * The bit map will indicate destination pnode numbers
1492 * relative to that base. They may not be consecutive
1493 * if nasid striding is being used.
1495 uv1_hdr
->base_dest_nasid
=
1496 UV_PNODE_TO_NASID(base_pnode
);
1497 uv1_hdr
->dest_subnodeid
= UV_LB_SUBNODEID
;
1498 uv1_hdr
->command
= UV_NET_ENDPOINT_INTD
;
1499 uv1_hdr
->int_both
= 1;
1501 * all others need to be set to zero:
1502 * fairness chaining multilevel count replied_to
1505 uv2_hdr
= &bd2
->header
.uv2_hdr
;
1506 uv2_hdr
->swack_flag
= 1;
1507 uv2_hdr
->base_dest_nasid
=
1508 UV_PNODE_TO_NASID(base_pnode
);
1509 uv2_hdr
->dest_subnodeid
= UV_LB_SUBNODEID
;
1510 uv2_hdr
->command
= UV_NET_ENDPOINT_INTD
;
1513 for_each_present_cpu(cpu
) {
1514 if (pnode
!= uv_blade_to_pnode(uv_cpu_to_blade_id(cpu
)))
1516 bcp
= &per_cpu(bau_control
, cpu
);
1517 bcp
->descriptor_base
= bau_desc
;
1522 * initialize the destination side's receiving buffers
1523 * entered for each uvhub in the partition
1524 * - node is first node (kernel memory notion) on the uvhub
1525 * - pnode is the uvhub's physical identifier
1527 static void pq_init(int node
, int pnode
)
1534 unsigned long first
;
1535 unsigned long pn_first
;
1537 struct bau_pq_entry
*pqp
;
1538 struct bau_control
*bcp
;
1540 plsize
= (DEST_Q_SIZE
+ 1) * sizeof(struct bau_pq_entry
);
1541 vp
= kmalloc_node(plsize
, GFP_KERNEL
, node
);
1542 pqp
= (struct bau_pq_entry
*)vp
;
1545 cp
= (char *)pqp
+ 31;
1546 pqp
= (struct bau_pq_entry
*)(((unsigned long)cp
>> 5) << 5);
1548 for_each_present_cpu(cpu
) {
1549 if (pnode
!= uv_cpu_to_pnode(cpu
))
1551 /* for every cpu on this pnode: */
1552 bcp
= &per_cpu(bau_control
, cpu
);
1553 bcp
->queue_first
= pqp
;
1554 bcp
->bau_msg_head
= pqp
;
1555 bcp
->queue_last
= pqp
+ (DEST_Q_SIZE
- 1);
1558 * need the gnode of where the memory was really allocated
1560 pn
= uv_gpa_to_gnode(uv_gpa(pqp
));
1561 first
= uv_physnodeaddr(pqp
);
1562 pn_first
= ((unsigned long)pn
<< UV_PAYLOADQ_PNODE_SHIFT
) | first
;
1563 last
= uv_physnodeaddr(pqp
+ (DEST_Q_SIZE
- 1));
1564 write_mmr_payload_first(pnode
, pn_first
);
1565 write_mmr_payload_tail(pnode
, first
);
1566 write_mmr_payload_last(pnode
, last
);
1568 /* in effect, all msg_type's are set to MSG_NOOP */
1569 memset(pqp
, 0, sizeof(struct bau_pq_entry
) * DEST_Q_SIZE
);
1573 * Initialization of each UV hub's structures
1575 static void __init
init_uvhub(int uvhub
, int vector
, int base_pnode
)
1579 unsigned long apicid
;
1581 node
= uvhub_to_first_node(uvhub
);
1582 pnode
= uv_blade_to_pnode(uvhub
);
1584 activation_descriptor_init(node
, pnode
, base_pnode
);
1586 pq_init(node
, pnode
);
1588 * The below initialization can't be in firmware because the
1589 * messaging IRQ will be determined by the OS.
1591 apicid
= uvhub_to_first_apicid(uvhub
) | uv_apicid_hibits
;
1592 write_mmr_data_config(pnode
, ((apicid
<< 32) | vector
));
1596 * We will set BAU_MISC_CONTROL with a timeout period.
1597 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1598 * So the destination timeout period has to be calculated from them.
1600 static int calculate_destination_timeout(void)
1602 unsigned long mmr_image
;
1608 unsigned long ts_ns
;
1611 mult1
= SOFTACK_TIMEOUT_PERIOD
& BAU_MISC_CONTROL_MULT_MASK
;
1612 mmr_image
= uv_read_local_mmr(UVH_AGING_PRESCALE_SEL
);
1613 index
= (mmr_image
>> BAU_URGENCY_7_SHIFT
) & BAU_URGENCY_7_MASK
;
1614 mmr_image
= uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT
);
1615 mult2
= (mmr_image
>> BAU_TRANS_SHIFT
) & BAU_TRANS_MASK
;
1616 base
= timeout_base_ns
[index
];
1617 ts_ns
= base
* mult1
* mult2
;
1620 /* 4 bits 0/1 for 10/80us base, 3 bits of multiplier */
1621 mmr_image
= uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL
);
1622 mmr_image
= (mmr_image
& UV_SA_MASK
) >> UV_SA_SHFT
;
1623 if (mmr_image
& (1L << UV2_ACK_UNITS_SHFT
))
1627 mult1
= mmr_image
& UV2_ACK_MASK
;
1633 static void __init
init_per_cpu_tunables(void)
1636 struct bau_control
*bcp
;
1638 for_each_present_cpu(cpu
) {
1639 bcp
= &per_cpu(bau_control
, cpu
);
1640 bcp
->baudisabled
= 0;
1641 bcp
->statp
= &per_cpu(ptcstats
, cpu
);
1642 /* time interval to catch a hardware stay-busy bug */
1643 bcp
->timeout_interval
= usec_2_cycles(2*timeout_us
);
1644 bcp
->max_concurr
= max_concurr
;
1645 bcp
->max_concurr_const
= max_concurr
;
1646 bcp
->plugged_delay
= plugged_delay
;
1647 bcp
->plugsb4reset
= plugsb4reset
;
1648 bcp
->timeoutsb4reset
= timeoutsb4reset
;
1649 bcp
->ipi_reset_limit
= ipi_reset_limit
;
1650 bcp
->complete_threshold
= complete_threshold
;
1651 bcp
->cong_response_us
= congested_respns_us
;
1652 bcp
->cong_reps
= congested_reps
;
1653 bcp
->cong_period
= congested_period
;
1658 * Scan all cpus to collect blade and socket summaries.
1660 static int __init
get_cpu_topology(int base_pnode
,
1661 struct uvhub_desc
*uvhub_descs
,
1662 unsigned char *uvhub_mask
)
1668 struct bau_control
*bcp
;
1669 struct uvhub_desc
*bdp
;
1670 struct socket_desc
*sdp
;
1672 for_each_present_cpu(cpu
) {
1673 bcp
= &per_cpu(bau_control
, cpu
);
1675 memset(bcp
, 0, sizeof(struct bau_control
));
1677 pnode
= uv_cpu_hub_info(cpu
)->pnode
;
1678 if ((pnode
- base_pnode
) >= UV_DISTRIBUTION_SIZE
) {
1680 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1681 cpu
, pnode
, base_pnode
, UV_DISTRIBUTION_SIZE
);
1685 bcp
->osnode
= cpu_to_node(cpu
);
1686 bcp
->partition_base_pnode
= base_pnode
;
1688 uvhub
= uv_cpu_hub_info(cpu
)->numa_blade_id
;
1689 *(uvhub_mask
+ (uvhub
/8)) |= (1 << (uvhub
%8));
1690 bdp
= &uvhub_descs
[uvhub
];
1696 /* kludge: 'assuming' one node per socket, and assuming that
1697 disabling a socket just leaves a gap in node numbers */
1698 socket
= bcp
->osnode
& 1;
1699 bdp
->socket_mask
|= (1 << socket
);
1700 sdp
= &bdp
->socket
[socket
];
1701 sdp
->cpu_number
[sdp
->num_cpus
] = cpu
;
1703 if (sdp
->num_cpus
> MAX_CPUS_PER_SOCKET
) {
1704 printk(KERN_EMERG
"%d cpus per socket invalid\n",
1713 * Each socket is to get a local array of pnodes/hubs.
1715 static void make_per_cpu_thp(struct bau_control
*smaster
)
1718 size_t hpsz
= sizeof(struct hub_and_pnode
) * num_possible_cpus();
1720 smaster
->thp
= kmalloc_node(hpsz
, GFP_KERNEL
, smaster
->osnode
);
1721 memset(smaster
->thp
, 0, hpsz
);
1722 for_each_present_cpu(cpu
) {
1723 smaster
->thp
[cpu
].pnode
= uv_cpu_hub_info(cpu
)->pnode
;
1724 smaster
->thp
[cpu
].uvhub
= uv_cpu_hub_info(cpu
)->numa_blade_id
;
1729 * Each uvhub is to get a local cpumask.
1731 static void make_per_hub_cpumask(struct bau_control
*hmaster
)
1733 int sz
= sizeof(cpumask_t
);
1735 hmaster
->cpumask
= kzalloc_node(sz
, GFP_KERNEL
, hmaster
->osnode
);
1739 * Initialize all the per_cpu information for the cpu's on a given socket,
1740 * given what has been gathered into the socket_desc struct.
1741 * And reports the chosen hub and socket masters back to the caller.
1743 static int scan_sock(struct socket_desc
*sdp
, struct uvhub_desc
*bdp
,
1744 struct bau_control
**smasterp
,
1745 struct bau_control
**hmasterp
)
1749 struct bau_control
*bcp
;
1751 for (i
= 0; i
< sdp
->num_cpus
; i
++) {
1752 cpu
= sdp
->cpu_number
[i
];
1753 bcp
= &per_cpu(bau_control
, cpu
);
1760 bcp
->cpus_in_uvhub
= bdp
->num_cpus
;
1761 bcp
->cpus_in_socket
= sdp
->num_cpus
;
1762 bcp
->socket_master
= *smasterp
;
1763 bcp
->uvhub
= bdp
->uvhub
;
1765 bcp
->uvhub_version
= 1;
1766 else if (is_uv2_hub())
1767 bcp
->uvhub_version
= 2;
1769 printk(KERN_EMERG
"uvhub version not 1 or 2\n");
1772 bcp
->uvhub_master
= *hmasterp
;
1773 bcp
->uvhub_cpu
= uv_cpu_hub_info(cpu
)->blade_processor_id
;
1774 if (bcp
->uvhub_cpu
>= MAX_CPUS_PER_UVHUB
) {
1775 printk(KERN_EMERG
"%d cpus per uvhub invalid\n",
1784 * Summarize the blade and socket topology into the per_cpu structures.
1786 static int __init
summarize_uvhub_sockets(int nuvhubs
,
1787 struct uvhub_desc
*uvhub_descs
,
1788 unsigned char *uvhub_mask
)
1792 unsigned short socket_mask
;
1794 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
1795 struct uvhub_desc
*bdp
;
1796 struct bau_control
*smaster
= NULL
;
1797 struct bau_control
*hmaster
= NULL
;
1799 if (!(*(uvhub_mask
+ (uvhub
/8)) & (1 << (uvhub
%8))))
1802 bdp
= &uvhub_descs
[uvhub
];
1803 socket_mask
= bdp
->socket_mask
;
1805 while (socket_mask
) {
1806 struct socket_desc
*sdp
;
1807 if ((socket_mask
& 1)) {
1808 sdp
= &bdp
->socket
[socket
];
1809 if (scan_sock(sdp
, bdp
, &smaster
, &hmaster
))
1811 make_per_cpu_thp(smaster
);
1814 socket_mask
= (socket_mask
>> 1);
1816 make_per_hub_cpumask(hmaster
);
1822 * initialize the bau_control structure for each cpu
1824 static int __init
init_per_cpu(int nuvhubs
, int base_part_pnode
)
1826 unsigned char *uvhub_mask
;
1828 struct uvhub_desc
*uvhub_descs
;
1830 timeout_us
= calculate_destination_timeout();
1832 vp
= kmalloc(nuvhubs
* sizeof(struct uvhub_desc
), GFP_KERNEL
);
1833 uvhub_descs
= (struct uvhub_desc
*)vp
;
1834 memset(uvhub_descs
, 0, nuvhubs
* sizeof(struct uvhub_desc
));
1835 uvhub_mask
= kzalloc((nuvhubs
+7)/8, GFP_KERNEL
);
1837 if (get_cpu_topology(base_part_pnode
, uvhub_descs
, uvhub_mask
))
1840 if (summarize_uvhub_sockets(nuvhubs
, uvhub_descs
, uvhub_mask
))
1845 init_per_cpu_tunables();
1855 * Initialization of BAU-related structures
1857 static int __init
uv_bau_init(void)
1865 cpumask_var_t
*mask
;
1867 if (!is_uv_system())
1873 for_each_possible_cpu(cur_cpu
) {
1874 mask
= &per_cpu(uv_flush_tlb_mask
, cur_cpu
);
1875 zalloc_cpumask_var_node(mask
, GFP_KERNEL
, cpu_to_node(cur_cpu
));
1878 nuvhubs
= uv_num_possible_blades();
1879 spin_lock_init(&disable_lock
);
1880 congested_cycles
= usec_2_cycles(congested_respns_us
);
1882 uv_base_pnode
= 0x7fffffff;
1883 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
1884 cpus
= uv_blade_nr_possible_cpus(uvhub
);
1885 if (cpus
&& (uv_blade_to_pnode(uvhub
) < uv_base_pnode
))
1886 uv_base_pnode
= uv_blade_to_pnode(uvhub
);
1891 if (init_per_cpu(nuvhubs
, uv_base_pnode
)) {
1896 vector
= UV_BAU_MESSAGE
;
1897 for_each_possible_blade(uvhub
)
1898 if (uv_blade_nr_possible_cpus(uvhub
))
1899 init_uvhub(uvhub
, vector
, uv_base_pnode
);
1901 alloc_intr_gate(vector
, uv_bau_message_intr1
);
1903 for_each_possible_blade(uvhub
) {
1904 if (uv_blade_nr_possible_cpus(uvhub
)) {
1907 pnode
= uv_blade_to_pnode(uvhub
);
1910 write_gmmr_activation(pnode
, val
);
1911 mmr
= 1; /* should be 1 to broadcast to both sockets */
1913 write_mmr_data_broadcast(pnode
, mmr
);
1919 core_initcall(uv_bau_init
);
1920 fs_initcall(uv_ptc_init
);