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
,
164 struct bau_pq_entry
*msg
;
167 if (!msg
->canceled
&& do_acknowledge
) {
168 dw
= (msg
->swack_vec
<< UV_SW_ACK_NPENDING
) | msg
->swack_vec
;
169 write_mmr_sw_ack(dw
);
176 * Process the receipt of a RETRY message
178 static void bau_process_retry_msg(struct msg_desc
*mdp
,
179 struct bau_control
*bcp
)
182 int cancel_count
= 0;
183 unsigned long msg_res
;
184 unsigned long mmr
= 0;
185 struct bau_pq_entry
*msg
= mdp
->msg
;
186 struct bau_pq_entry
*msg2
;
187 struct ptc_stats
*stat
= bcp
->statp
;
191 * cancel any message from msg+1 to the retry itself
193 for (msg2
= msg
+1, i
= 0; i
< DEST_Q_SIZE
; msg2
++, i
++) {
194 if (msg2
> mdp
->queue_last
)
195 msg2
= mdp
->queue_first
;
199 /* same conditions for cancellation as do_reset */
200 if ((msg2
->replied_to
== 0) && (msg2
->canceled
== 0) &&
201 (msg2
->swack_vec
) && ((msg2
->swack_vec
&
202 msg
->swack_vec
) == 0) &&
203 (msg2
->sending_cpu
== msg
->sending_cpu
) &&
204 (msg2
->msg_type
!= MSG_NOOP
)) {
205 mmr
= read_mmr_sw_ack();
206 msg_res
= msg2
->swack_vec
;
208 * This is a message retry; clear the resources held
209 * by the previous message only if they timed out.
210 * If it has not timed out we have an unexpected
211 * situation to report.
213 if (mmr
& (msg_res
<< UV_SW_ACK_NPENDING
)) {
216 * Is the resource timed out?
217 * Make everyone ignore the cancelled message.
222 mr
= (msg_res
<< UV_SW_ACK_NPENDING
) | msg_res
;
223 write_mmr_sw_ack(mr
);
228 stat
->d_nocanceled
++;
232 * Do all the things a cpu should do for a TLB shootdown message.
233 * Other cpu's may come here at the same time for this message.
235 static void bau_process_message(struct msg_desc
*mdp
, struct bau_control
*bcp
,
238 short socket_ack_count
= 0;
240 struct atomic_short
*asp
;
241 struct ptc_stats
*stat
= bcp
->statp
;
242 struct bau_pq_entry
*msg
= mdp
->msg
;
243 struct bau_control
*smaster
= bcp
->socket_master
;
246 * This must be a normal message, or retry of a normal message
248 if (msg
->address
== TLB_FLUSH_ALL
) {
252 __flush_tlb_one(msg
->address
);
258 * One cpu on each uvhub has the additional job on a RETRY
259 * of releasing the resource held by the message that is
260 * being retried. That message is identified by sending
263 if (msg
->msg_type
== MSG_RETRY
&& bcp
== bcp
->uvhub_master
)
264 bau_process_retry_msg(mdp
, bcp
);
267 * This is a swack message, so we have to reply to it.
268 * Count each responding cpu on the socket. This avoids
269 * pinging the count's cache line back and forth between
272 sp
= &smaster
->socket_acknowledge_count
[mdp
->msg_slot
];
273 asp
= (struct atomic_short
*)sp
;
274 socket_ack_count
= atom_asr(1, asp
);
275 if (socket_ack_count
== bcp
->cpus_in_socket
) {
278 * Both sockets dump their completed count total into
279 * the message's count.
281 smaster
->socket_acknowledge_count
[mdp
->msg_slot
] = 0;
282 asp
= (struct atomic_short
*)&msg
->acknowledge_count
;
283 msg_ack_count
= atom_asr(socket_ack_count
, asp
);
285 if (msg_ack_count
== bcp
->cpus_in_uvhub
) {
287 * All cpus in uvhub saw it; reply
288 * (unless we are in the UV2 workaround)
290 reply_to_message(mdp
, bcp
, do_acknowledge
);
298 * Determine the first cpu on a pnode.
300 static int pnode_to_first_cpu(int pnode
, struct bau_control
*smaster
)
303 struct hub_and_pnode
*hpp
;
305 for_each_present_cpu(cpu
) {
306 hpp
= &smaster
->thp
[cpu
];
307 if (pnode
== hpp
->pnode
)
314 * Last resort when we get a large number of destination timeouts is
315 * to clear resources held by a given cpu.
316 * Do this with IPI so that all messages in the BAU message queue
317 * can be identified by their nonzero swack_vec field.
319 * This is entered for a single cpu on the uvhub.
320 * The sender want's this uvhub to free a specific message's
323 static void do_reset(void *ptr
)
326 struct bau_control
*bcp
= &per_cpu(bau_control
, smp_processor_id());
327 struct reset_args
*rap
= (struct reset_args
*)ptr
;
328 struct bau_pq_entry
*msg
;
329 struct ptc_stats
*stat
= bcp
->statp
;
333 * We're looking for the given sender, and
334 * will free its swack resource.
335 * If all cpu's finally responded after the timeout, its
336 * message 'replied_to' was set.
338 for (msg
= bcp
->queue_first
, i
= 0; i
< DEST_Q_SIZE
; msg
++, i
++) {
339 unsigned long msg_res
;
340 /* do_reset: same conditions for cancellation as
341 bau_process_retry_msg() */
342 if ((msg
->replied_to
== 0) &&
343 (msg
->canceled
== 0) &&
344 (msg
->sending_cpu
== rap
->sender
) &&
346 (msg
->msg_type
!= MSG_NOOP
)) {
350 * make everyone else ignore this message
354 * only reset the resource if it is still pending
356 mmr
= read_mmr_sw_ack();
357 msg_res
= msg
->swack_vec
;
358 mr
= (msg_res
<< UV_SW_ACK_NPENDING
) | msg_res
;
361 write_mmr_sw_ack(mr
);
369 * Use IPI to get all target uvhubs to release resources held by
370 * a given sending cpu number.
372 static void reset_with_ipi(struct pnmask
*distribution
, struct bau_control
*bcp
)
377 int sender
= bcp
->cpu
;
378 cpumask_t
*mask
= bcp
->uvhub_master
->cpumask
;
379 struct bau_control
*smaster
= bcp
->socket_master
;
380 struct reset_args reset_args
;
382 reset_args
.sender
= sender
;
384 /* find a single cpu for each uvhub in this distribution mask */
385 maskbits
= sizeof(struct pnmask
) * BITSPERBYTE
;
386 /* each bit is a pnode relative to the partition base pnode */
387 for (pnode
= 0; pnode
< maskbits
; pnode
++) {
389 if (!bau_uvhub_isset(pnode
, distribution
))
391 apnode
= pnode
+ bcp
->partition_base_pnode
;
392 cpu
= pnode_to_first_cpu(apnode
, smaster
);
396 /* IPI all cpus; preemption is already disabled */
397 smp_call_function_many(mask
, do_reset
, (void *)&reset_args
, 1);
401 static inline unsigned long cycles_2_us(unsigned long long cyc
)
403 unsigned long long ns
;
405 int cpu
= smp_processor_id();
407 ns
= (cyc
* per_cpu(cyc2ns
, cpu
)) >> CYC2NS_SCALE_FACTOR
;
413 * wait for all cpus on this hub to finish their sends and go quiet
414 * leaves uvhub_quiesce set so that no new broadcasts are started by
415 * bau_flush_send_and_wait()
417 static inline void quiesce_local_uvhub(struct bau_control
*hmaster
)
419 atom_asr(1, (struct atomic_short
*)&hmaster
->uvhub_quiesce
);
423 * mark this quiet-requestor as done
425 static inline void end_uvhub_quiesce(struct bau_control
*hmaster
)
427 atom_asr(-1, (struct atomic_short
*)&hmaster
->uvhub_quiesce
);
430 static unsigned long uv1_read_status(unsigned long mmr_offset
, int right_shift
)
432 unsigned long descriptor_status
;
434 descriptor_status
= uv_read_local_mmr(mmr_offset
);
435 descriptor_status
>>= right_shift
;
436 descriptor_status
&= UV_ACT_STATUS_MASK
;
437 return descriptor_status
;
441 * Wait for completion of a broadcast software ack message
442 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
444 static int uv1_wait_completion(struct bau_desc
*bau_desc
,
445 unsigned long mmr_offset
, int right_shift
,
446 struct bau_control
*bcp
, long try)
448 unsigned long descriptor_status
;
450 struct ptc_stats
*stat
= bcp
->statp
;
452 descriptor_status
= uv1_read_status(mmr_offset
, right_shift
);
453 /* spin on the status MMR, waiting for it to go idle */
454 while ((descriptor_status
!= DS_IDLE
)) {
456 * Our software ack messages may be blocked because
457 * there are no swack resources available. As long
458 * as none of them has timed out hardware will NACK
459 * our message and its state will stay IDLE.
461 if (descriptor_status
== DS_SOURCE_TIMEOUT
) {
464 } else if (descriptor_status
== DS_DESTINATION_TIMEOUT
) {
469 * Our retries may be blocked by all destination
470 * swack resources being consumed, and a timeout
471 * pending. In that case hardware returns the
472 * ERROR that looks like a destination timeout.
474 if (cycles_2_us(ttm
- bcp
->send_message
) < timeout_us
) {
475 bcp
->conseccompletes
= 0;
476 return FLUSH_RETRY_PLUGGED
;
479 bcp
->conseccompletes
= 0;
480 return FLUSH_RETRY_TIMEOUT
;
483 * descriptor_status is still BUSY
487 descriptor_status
= uv1_read_status(mmr_offset
, right_shift
);
489 bcp
->conseccompletes
++;
490 return FLUSH_COMPLETE
;
494 * UV2 has an extra bit of status in the ACTIVATION_STATUS_2 register.
496 static unsigned long uv2_read_status(unsigned long offset
, int rshft
, int desc
)
498 unsigned long descriptor_status
;
499 unsigned long descriptor_status2
;
501 descriptor_status
= ((read_lmmr(offset
) >> rshft
) & UV_ACT_STATUS_MASK
);
502 descriptor_status2
= (read_mmr_uv2_status() >> desc
) & 0x1UL
;
503 descriptor_status
= (descriptor_status
<< 1) | descriptor_status2
;
504 return descriptor_status
;
508 * Return whether the status of the descriptor that is normally used for this
509 * cpu (the one indexed by its hub-relative cpu number) is busy.
510 * The status of the original 32 descriptors is always reflected in the 64
511 * bits of UVH_LB_BAU_SB_ACTIVATION_STATUS_0.
512 * The bit provided by the activation_status_2 register is irrelevant to
513 * the status if it is only being tested for busy or not busy.
515 int normal_busy(struct bau_control
*bcp
)
517 int cpu
= bcp
->uvhub_cpu
;
521 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_0
;
522 right_shift
= cpu
* UV_ACT_STATUS_SIZE
;
523 return (((((read_lmmr(mmr_offset
) >> right_shift
) &
524 UV_ACT_STATUS_MASK
)) << 1) == UV2H_DESC_BUSY
);
528 * Entered when a bau descriptor has gone into a permanent busy wait because
530 * Workaround the bug.
532 int handle_uv2_busy(struct bau_control
*bcp
)
534 int busy_one
= bcp
->using_desc
;
535 int normal
= bcp
->uvhub_cpu
;
538 unsigned long descriptor_status
;
539 unsigned long status
;
541 struct bau_desc
*bau_desc_old
;
542 struct bau_desc
*bau_desc_new
;
543 struct bau_control
*hmaster
= bcp
->uvhub_master
;
544 struct ptc_stats
*stat
= bcp
->statp
;
548 spin_lock(&hmaster
->uvhub_lock
);
549 /* try for the original first */
550 if (busy_one
!= normal
) {
551 if (!normal_busy(bcp
))
555 /* can't use the normal, select an alternate */
556 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_1
;
557 descriptor_status
= read_lmmr(mmr_offset
);
559 /* scan available descriptors 32-63 */
560 for (i
= 0; i
< UV_CPUS_PER_AS
; i
++) {
561 if ((hmaster
->inuse_map
& (1 << i
)) == 0) {
562 status
= ((descriptor_status
>>
563 (i
* UV_ACT_STATUS_SIZE
)) &
564 UV_ACT_STATUS_MASK
) << 1;
565 if (status
!= UV2H_DESC_BUSY
) {
566 selected
= i
+ UV_CPUS_PER_AS
;
573 if (busy_one
!= normal
)
574 /* mark the busy alternate as not in-use */
575 hmaster
->inuse_map
&= ~(1 << (busy_one
- UV_CPUS_PER_AS
));
578 /* switch to the selected descriptor */
579 if (selected
!= normal
) {
580 /* set the selected alternate as in-use */
581 hmaster
->inuse_map
|=
582 (1 << (selected
- UV_CPUS_PER_AS
));
583 if (selected
> stat
->s_uv2_wars_hw
)
584 stat
->s_uv2_wars_hw
= selected
;
586 bau_desc_old
= bcp
->descriptor_base
;
587 bau_desc_old
+= (ITEMS_PER_DESC
* busy_one
);
588 bcp
->using_desc
= selected
;
589 bau_desc_new
= bcp
->descriptor_base
;
590 bau_desc_new
+= (ITEMS_PER_DESC
* selected
);
591 *bau_desc_new
= *bau_desc_old
;
594 * All are busy. Wait for the normal one for this cpu to
597 stat
->s_uv2_war_waits
++;
598 spin_unlock(&hmaster
->uvhub_lock
);
602 } while (normal_busy(bcp
));
603 spin_lock(&hmaster
->uvhub_lock
);
604 /* switch to the original descriptor */
605 bcp
->using_desc
= normal
;
606 bau_desc_old
= bcp
->descriptor_base
;
607 bau_desc_old
+= (ITEMS_PER_DESC
* bcp
->using_desc
);
608 bcp
->using_desc
= (ITEMS_PER_DESC
* normal
);
609 bau_desc_new
= bcp
->descriptor_base
;
610 bau_desc_new
+= (ITEMS_PER_DESC
* normal
);
611 *bau_desc_new
= *bau_desc_old
; /* copy the entire descriptor */
613 spin_unlock(&hmaster
->uvhub_lock
);
614 return FLUSH_RETRY_BUSYBUG
;
617 static int uv2_wait_completion(struct bau_desc
*bau_desc
,
618 unsigned long mmr_offset
, int right_shift
,
619 struct bau_control
*bcp
, long try)
621 unsigned long descriptor_stat
;
623 int desc
= bcp
->using_desc
;
625 struct ptc_stats
*stat
= bcp
->statp
;
627 descriptor_stat
= uv2_read_status(mmr_offset
, right_shift
, desc
);
629 /* spin on the status MMR, waiting for it to go idle */
630 while (descriptor_stat
!= UV2H_DESC_IDLE
) {
632 * Our software ack messages may be blocked because
633 * there are no swack resources available. As long
634 * as none of them has timed out hardware will NACK
635 * our message and its state will stay IDLE.
637 if ((descriptor_stat
== UV2H_DESC_SOURCE_TIMEOUT
) ||
638 (descriptor_stat
== UV2H_DESC_DEST_STRONG_NACK
) ||
639 (descriptor_stat
== UV2H_DESC_DEST_PUT_ERR
)) {
642 } else if (descriptor_stat
== UV2H_DESC_DEST_TIMEOUT
) {
646 * Our retries may be blocked by all destination
647 * swack resources being consumed, and a timeout
648 * pending. In that case hardware returns the
649 * ERROR that looks like a destination timeout.
651 if (cycles_2_us(ttm
- bcp
->send_message
) < timeout_us
) {
652 bcp
->conseccompletes
= 0;
653 return FLUSH_RETRY_PLUGGED
;
655 bcp
->conseccompletes
= 0;
656 return FLUSH_RETRY_TIMEOUT
;
659 if (busy_reps
> 1000000) {
660 /* not to hammer on the clock */
663 if ((ttm
- bcp
->send_message
) >
664 (bcp
->clocks_per_100_usec
)) {
665 return handle_uv2_busy(bcp
);
669 * descriptor_stat is still BUSY
673 descriptor_stat
= uv2_read_status(mmr_offset
, right_shift
,
676 bcp
->conseccompletes
++;
677 return FLUSH_COMPLETE
;
681 * There are 2 status registers; each and array[32] of 2 bits. Set up for
682 * which register to read and position in that register based on cpu in
685 static int wait_completion(struct bau_desc
*bau_desc
,
686 struct bau_control
*bcp
, long try)
689 unsigned long mmr_offset
;
690 int desc
= bcp
->using_desc
;
692 if (desc
< UV_CPUS_PER_AS
) {
693 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_0
;
694 right_shift
= desc
* UV_ACT_STATUS_SIZE
;
696 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_1
;
697 right_shift
= ((desc
- UV_CPUS_PER_AS
) * UV_ACT_STATUS_SIZE
);
700 if (bcp
->uvhub_version
== 1)
701 return uv1_wait_completion(bau_desc
, mmr_offset
, right_shift
,
704 return uv2_wait_completion(bau_desc
, mmr_offset
, right_shift
,
708 static inline cycles_t
sec_2_cycles(unsigned long sec
)
713 ns
= sec
* 1000000000;
714 cyc
= (ns
<< CYC2NS_SCALE_FACTOR
)/(per_cpu(cyc2ns
, smp_processor_id()));
719 * Our retries are blocked by all destination sw ack resources being
720 * in use, and a timeout is pending. In that case hardware immediately
721 * returns the ERROR that looks like a destination timeout.
723 static void destination_plugged(struct bau_desc
*bau_desc
,
724 struct bau_control
*bcp
,
725 struct bau_control
*hmaster
, struct ptc_stats
*stat
)
727 udelay(bcp
->plugged_delay
);
728 bcp
->plugged_tries
++;
730 if (bcp
->plugged_tries
>= bcp
->plugsb4reset
) {
731 bcp
->plugged_tries
= 0;
733 quiesce_local_uvhub(hmaster
);
735 spin_lock(&hmaster
->queue_lock
);
736 reset_with_ipi(&bau_desc
->distribution
, bcp
);
737 spin_unlock(&hmaster
->queue_lock
);
739 end_uvhub_quiesce(hmaster
);
742 stat
->s_resets_plug
++;
746 static void destination_timeout(struct bau_desc
*bau_desc
,
747 struct bau_control
*bcp
, struct bau_control
*hmaster
,
748 struct ptc_stats
*stat
)
750 hmaster
->max_concurr
= 1;
751 bcp
->timeout_tries
++;
752 if (bcp
->timeout_tries
>= bcp
->timeoutsb4reset
) {
753 bcp
->timeout_tries
= 0;
755 quiesce_local_uvhub(hmaster
);
757 spin_lock(&hmaster
->queue_lock
);
758 reset_with_ipi(&bau_desc
->distribution
, bcp
);
759 spin_unlock(&hmaster
->queue_lock
);
761 end_uvhub_quiesce(hmaster
);
764 stat
->s_resets_timeout
++;
769 * Completions are taking a very long time due to a congested numalink
772 static void disable_for_congestion(struct bau_control
*bcp
,
773 struct ptc_stats
*stat
)
775 /* let only one cpu do this disabling */
776 spin_lock(&disable_lock
);
778 if (!baudisabled
&& bcp
->period_requests
&&
779 ((bcp
->period_time
/ bcp
->period_requests
) > congested_cycles
)) {
781 struct bau_control
*tbcp
;
782 /* it becomes this cpu's job to turn on the use of the
785 bcp
->set_bau_off
= 1;
786 bcp
->set_bau_on_time
= get_cycles();
787 bcp
->set_bau_on_time
+= sec_2_cycles(bcp
->cong_period
);
788 stat
->s_bau_disabled
++;
789 for_each_present_cpu(tcpu
) {
790 tbcp
= &per_cpu(bau_control
, tcpu
);
791 tbcp
->baudisabled
= 1;
795 spin_unlock(&disable_lock
);
798 static void count_max_concurr(int stat
, struct bau_control
*bcp
,
799 struct bau_control
*hmaster
)
801 bcp
->plugged_tries
= 0;
802 bcp
->timeout_tries
= 0;
803 if (stat
!= FLUSH_COMPLETE
)
805 if (bcp
->conseccompletes
<= bcp
->complete_threshold
)
807 if (hmaster
->max_concurr
>= hmaster
->max_concurr_const
)
809 hmaster
->max_concurr
++;
812 static void record_send_stats(cycles_t time1
, cycles_t time2
,
813 struct bau_control
*bcp
, struct ptc_stats
*stat
,
814 int completion_status
, int try)
819 elapsed
= time2
- time1
;
820 stat
->s_time
+= elapsed
;
822 if ((completion_status
== FLUSH_COMPLETE
) && (try == 1)) {
823 bcp
->period_requests
++;
824 bcp
->period_time
+= elapsed
;
825 if ((elapsed
> congested_cycles
) &&
826 (bcp
->period_requests
> bcp
->cong_reps
))
827 disable_for_congestion(bcp
, stat
);
832 if (completion_status
== FLUSH_COMPLETE
&& try > 1)
834 else if (completion_status
== FLUSH_GIVEUP
)
839 * Because of a uv1 hardware bug only a limited number of concurrent
840 * requests can be made.
842 static void uv1_throttle(struct bau_control
*hmaster
, struct ptc_stats
*stat
)
844 spinlock_t
*lock
= &hmaster
->uvhub_lock
;
847 v
= &hmaster
->active_descriptor_count
;
848 if (!atomic_inc_unless_ge(lock
, v
, hmaster
->max_concurr
)) {
852 } while (!atomic_inc_unless_ge(lock
, v
, hmaster
->max_concurr
));
857 * Handle the completion status of a message send.
859 static void handle_cmplt(int completion_status
, struct bau_desc
*bau_desc
,
860 struct bau_control
*bcp
, struct bau_control
*hmaster
,
861 struct ptc_stats
*stat
)
863 if (completion_status
== FLUSH_RETRY_PLUGGED
)
864 destination_plugged(bau_desc
, bcp
, hmaster
, stat
);
865 else if (completion_status
== FLUSH_RETRY_TIMEOUT
)
866 destination_timeout(bau_desc
, bcp
, hmaster
, stat
);
870 * Send a broadcast and wait for it to complete.
872 * The flush_mask contains the cpus the broadcast is to be sent to including
873 * cpus that are on the local uvhub.
875 * Returns 0 if all flushing represented in the mask was done.
876 * Returns 1 if it gives up entirely and the original cpu mask is to be
877 * returned to the kernel.
879 int uv_flush_send_and_wait(struct cpumask
*flush_mask
, struct bau_control
*bcp
)
882 int completion_stat
= 0;
888 struct ptc_stats
*stat
= bcp
->statp
;
889 struct bau_control
*hmaster
= bcp
->uvhub_master
;
890 struct uv1_bau_msg_header
*uv1_hdr
= NULL
;
891 struct uv2_bau_msg_header
*uv2_hdr
= NULL
;
892 struct bau_desc
*bau_desc
;
894 if (bcp
->uvhub_version
== 1)
895 uv1_throttle(hmaster
, stat
);
897 while (hmaster
->uvhub_quiesce
)
900 time1
= get_cycles();
902 bau_desc
= bcp
->descriptor_base
;
903 bau_desc
+= (ITEMS_PER_DESC
* bcp
->using_desc
);
904 if (bcp
->uvhub_version
== 1) {
906 uv1_hdr
= &bau_desc
->header
.uv1_hdr
;
908 uv2_hdr
= &bau_desc
->header
.uv2_hdr
;
909 if ((try == 0) || (completion_stat
== FLUSH_RETRY_BUSYBUG
)) {
911 uv1_hdr
->msg_type
= MSG_REGULAR
;
913 uv2_hdr
->msg_type
= MSG_REGULAR
;
914 seq_number
= bcp
->message_number
++;
917 uv1_hdr
->msg_type
= MSG_RETRY
;
919 uv2_hdr
->msg_type
= MSG_RETRY
;
920 stat
->s_retry_messages
++;
924 uv1_hdr
->sequence
= seq_number
;
926 uv2_hdr
->sequence
= seq_number
;
927 index
= (1UL << AS_PUSH_SHIFT
) | bcp
->using_desc
;
928 bcp
->send_message
= get_cycles();
930 write_mmr_activation(index
);
933 completion_stat
= wait_completion(bau_desc
, bcp
, try);
934 /* UV2: wait_completion() may change the bcp->using_desc */
936 handle_cmplt(completion_stat
, bau_desc
, bcp
, hmaster
, stat
);
938 if (bcp
->ipi_attempts
>= bcp
->ipi_reset_limit
) {
939 bcp
->ipi_attempts
= 0;
940 completion_stat
= FLUSH_GIVEUP
;
944 } while ((completion_stat
== FLUSH_RETRY_PLUGGED
) ||
945 (completion_stat
== FLUSH_RETRY_BUSYBUG
) ||
946 (completion_stat
== FLUSH_RETRY_TIMEOUT
));
948 time2
= get_cycles();
950 count_max_concurr(completion_stat
, bcp
, hmaster
);
952 while (hmaster
->uvhub_quiesce
)
955 atomic_dec(&hmaster
->active_descriptor_count
);
957 record_send_stats(time1
, time2
, bcp
, stat
, completion_stat
, try);
959 if (completion_stat
== FLUSH_GIVEUP
)
960 /* FLUSH_GIVEUP will fall back to using IPI's for tlb flush */
966 * The BAU is disabled. When the disabled time period has expired, the cpu
967 * that disabled it must re-enable it.
968 * Return 0 if it is re-enabled for all cpus.
970 static int check_enable(struct bau_control
*bcp
, struct ptc_stats
*stat
)
973 struct bau_control
*tbcp
;
975 if (bcp
->set_bau_off
) {
976 if (get_cycles() >= bcp
->set_bau_on_time
) {
977 stat
->s_bau_reenabled
++;
979 for_each_present_cpu(tcpu
) {
980 tbcp
= &per_cpu(bau_control
, tcpu
);
981 tbcp
->baudisabled
= 0;
982 tbcp
->period_requests
= 0;
983 tbcp
->period_time
= 0;
991 static void record_send_statistics(struct ptc_stats
*stat
, int locals
, int hubs
,
992 int remotes
, struct bau_desc
*bau_desc
)
995 stat
->s_ntargcpu
+= remotes
+ locals
;
996 stat
->s_ntargremotes
+= remotes
;
997 stat
->s_ntarglocals
+= locals
;
999 /* uvhub statistics */
1000 hubs
= bau_uvhub_weight(&bau_desc
->distribution
);
1002 stat
->s_ntarglocaluvhub
++;
1003 stat
->s_ntargremoteuvhub
+= (hubs
- 1);
1005 stat
->s_ntargremoteuvhub
+= hubs
;
1007 stat
->s_ntarguvhub
+= hubs
;
1010 stat
->s_ntarguvhub16
++;
1012 stat
->s_ntarguvhub8
++;
1014 stat
->s_ntarguvhub4
++;
1016 stat
->s_ntarguvhub2
++;
1018 stat
->s_ntarguvhub1
++;
1022 * Translate a cpu mask to the uvhub distribution mask in the BAU
1023 * activation descriptor.
1025 static int set_distrib_bits(struct cpumask
*flush_mask
, struct bau_control
*bcp
,
1026 struct bau_desc
*bau_desc
, int *localsp
, int *remotesp
)
1031 struct hub_and_pnode
*hpp
;
1033 for_each_cpu(cpu
, flush_mask
) {
1035 * The distribution vector is a bit map of pnodes, relative
1036 * to the partition base pnode (and the partition base nasid
1038 * Translate cpu to pnode and hub using a local memory array.
1040 hpp
= &bcp
->socket_master
->thp
[cpu
];
1041 pnode
= hpp
->pnode
- bcp
->partition_base_pnode
;
1042 bau_uvhub_set(pnode
, &bau_desc
->distribution
);
1044 if (hpp
->uvhub
== bcp
->uvhub
)
1055 * globally purge translation cache of a virtual address or all TLB's
1056 * @cpumask: mask of all cpu's in which the address is to be removed
1057 * @mm: mm_struct containing virtual address range
1058 * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
1059 * @cpu: the current cpu
1061 * This is the entry point for initiating any UV global TLB shootdown.
1063 * Purges the translation caches of all specified processors of the given
1064 * virtual address, or purges all TLB's on specified processors.
1066 * The caller has derived the cpumask from the mm_struct. This function
1067 * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
1069 * The cpumask is converted into a uvhubmask of the uvhubs containing
1072 * Note that this function should be called with preemption disabled.
1074 * Returns NULL if all remote flushing was done.
1075 * Returns pointer to cpumask if some remote flushing remains to be
1076 * done. The returned pointer is valid till preemption is re-enabled.
1078 const struct cpumask
*uv_flush_tlb_others(const struct cpumask
*cpumask
,
1079 struct mm_struct
*mm
, unsigned long va
,
1085 struct bau_desc
*bau_desc
;
1086 struct cpumask
*flush_mask
;
1087 struct ptc_stats
*stat
;
1088 struct bau_control
*bcp
;
1090 /* kernel was booted 'nobau' */
1094 bcp
= &per_cpu(bau_control
, cpu
);
1097 /* bau was disabled due to slow response */
1098 if (bcp
->baudisabled
) {
1099 if (check_enable(bcp
, stat
))
1104 * Each sending cpu has a per-cpu mask which it fills from the caller's
1105 * cpu mask. All cpus are converted to uvhubs and copied to the
1106 * activation descriptor.
1108 flush_mask
= (struct cpumask
*)per_cpu(uv_flush_tlb_mask
, cpu
);
1109 /* don't actually do a shootdown of the local cpu */
1110 cpumask_andnot(flush_mask
, cpumask
, cpumask_of(cpu
));
1112 if (cpu_isset(cpu
, *cpumask
))
1113 stat
->s_ntargself
++;
1115 bau_desc
= bcp
->descriptor_base
;
1116 bau_desc
+= (ITEMS_PER_DESC
* bcp
->using_desc
);
1117 bau_uvhubs_clear(&bau_desc
->distribution
, UV_DISTRIBUTION_SIZE
);
1118 if (set_distrib_bits(flush_mask
, bcp
, bau_desc
, &locals
, &remotes
))
1121 record_send_statistics(stat
, locals
, hubs
, remotes
, bau_desc
);
1123 bau_desc
->payload
.address
= va
;
1124 bau_desc
->payload
.sending_cpu
= cpu
;
1126 * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
1127 * or 1 if it gave up and the original cpumask should be returned.
1129 if (!uv_flush_send_and_wait(flush_mask
, bcp
))
1136 * Search the message queue for any 'other' message with the same software
1137 * acknowledge resource bit vector.
1139 struct bau_pq_entry
*find_another_by_swack(struct bau_pq_entry
*msg
,
1140 struct bau_control
*bcp
, unsigned char swack_vec
)
1142 struct bau_pq_entry
*msg_next
= msg
+ 1;
1144 if (msg_next
> bcp
->queue_last
)
1145 msg_next
= bcp
->queue_first
;
1146 while ((msg_next
->swack_vec
!= 0) && (msg_next
!= msg
)) {
1147 if (msg_next
->swack_vec
== swack_vec
)
1150 if (msg_next
> bcp
->queue_last
)
1151 msg_next
= bcp
->queue_first
;
1157 * UV2 needs to work around a bug in which an arriving message has not
1158 * set a bit in the UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE register.
1159 * Such a message must be ignored.
1161 void process_uv2_message(struct msg_desc
*mdp
, struct bau_control
*bcp
)
1163 unsigned long mmr_image
;
1164 unsigned char swack_vec
;
1165 struct bau_pq_entry
*msg
= mdp
->msg
;
1166 struct bau_pq_entry
*other_msg
;
1168 mmr_image
= read_mmr_sw_ack();
1169 swack_vec
= msg
->swack_vec
;
1171 if ((swack_vec
& mmr_image
) == 0) {
1173 * This message was assigned a swack resource, but no
1174 * reserved acknowlegment is pending.
1175 * The bug has prevented this message from setting the MMR.
1176 * And no other message has used the same sw_ack resource.
1177 * Do the requested shootdown but do not reply to the msg.
1178 * (the 0 means make no acknowledge)
1180 bau_process_message(mdp
, bcp
, 0);
1185 * Some message has set the MMR 'pending' bit; it might have been
1186 * another message. Look for that message.
1188 other_msg
= find_another_by_swack(msg
, bcp
, msg
->swack_vec
);
1190 /* There is another. Do not ack the current one. */
1191 bau_process_message(mdp
, bcp
, 0);
1193 * Let the natural processing of that message acknowledge
1194 * it. Don't get the processing of sw_ack's out of order.
1200 * There is no other message using this sw_ack, so it is safe to
1203 bau_process_message(mdp
, bcp
, 1);
1209 * The BAU message interrupt comes here. (registered by set_intr_gate)
1212 * We received a broadcast assist message.
1214 * Interrupts are disabled; this interrupt could represent
1215 * the receipt of several messages.
1217 * All cores/threads on this hub get this interrupt.
1218 * The last one to see it does the software ack.
1219 * (the resource will not be freed until noninterruptable cpus see this
1220 * interrupt; hardware may timeout the s/w ack and reply ERROR)
1222 void uv_bau_message_interrupt(struct pt_regs
*regs
)
1225 cycles_t time_start
;
1226 struct bau_pq_entry
*msg
;
1227 struct bau_control
*bcp
;
1228 struct ptc_stats
*stat
;
1229 struct msg_desc msgdesc
;
1231 time_start
= get_cycles();
1233 bcp
= &per_cpu(bau_control
, smp_processor_id());
1236 msgdesc
.queue_first
= bcp
->queue_first
;
1237 msgdesc
.queue_last
= bcp
->queue_last
;
1239 msg
= bcp
->bau_msg_head
;
1240 while (msg
->swack_vec
) {
1243 msgdesc
.msg_slot
= msg
- msgdesc
.queue_first
;
1245 if (bcp
->uvhub_version
== 2)
1246 process_uv2_message(&msgdesc
, bcp
);
1248 bau_process_message(&msgdesc
, bcp
, 1);
1251 if (msg
> msgdesc
.queue_last
)
1252 msg
= msgdesc
.queue_first
;
1253 bcp
->bau_msg_head
= msg
;
1255 stat
->d_time
+= (get_cycles() - time_start
);
1265 * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1266 * shootdown message timeouts enabled. The timeout does not cause
1267 * an interrupt, but causes an error message to be returned to
1270 static void __init
enable_timeouts(void)
1275 unsigned long mmr_image
;
1277 nuvhubs
= uv_num_possible_blades();
1279 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
1280 if (!uv_blade_nr_possible_cpus(uvhub
))
1283 pnode
= uv_blade_to_pnode(uvhub
);
1284 mmr_image
= read_mmr_misc_control(pnode
);
1286 * Set the timeout period and then lock it in, in three
1287 * steps; captures and locks in the period.
1289 * To program the period, the SOFT_ACK_MODE must be off.
1291 mmr_image
&= ~(1L << SOFTACK_MSHIFT
);
1292 write_mmr_misc_control(pnode
, mmr_image
);
1294 * Set the 4-bit period.
1296 mmr_image
&= ~((unsigned long)0xf << SOFTACK_PSHIFT
);
1297 mmr_image
|= (SOFTACK_TIMEOUT_PERIOD
<< SOFTACK_PSHIFT
);
1298 write_mmr_misc_control(pnode
, mmr_image
);
1301 * Subsequent reversals of the timebase bit (3) cause an
1302 * immediate timeout of one or all INTD resources as
1303 * indicated in bits 2:0 (7 causes all of them to timeout).
1305 mmr_image
|= (1L << SOFTACK_MSHIFT
);
1307 mmr_image
&= ~(1L << UV2_LEG_SHFT
);
1308 mmr_image
|= (1L << UV2_EXT_SHFT
);
1310 write_mmr_misc_control(pnode
, mmr_image
);
1314 static void *ptc_seq_start(struct seq_file
*file
, loff_t
*offset
)
1316 if (*offset
< num_possible_cpus())
1321 static void *ptc_seq_next(struct seq_file
*file
, void *data
, loff_t
*offset
)
1324 if (*offset
< num_possible_cpus())
1329 static void ptc_seq_stop(struct seq_file
*file
, void *data
)
1333 static inline unsigned long long usec_2_cycles(unsigned long microsec
)
1336 unsigned long long cyc
;
1338 ns
= microsec
* 1000;
1339 cyc
= (ns
<< CYC2NS_SCALE_FACTOR
)/(per_cpu(cyc2ns
, smp_processor_id()));
1344 * Display the statistics thru /proc/sgi_uv/ptc_statistics
1345 * 'data' points to the cpu number
1346 * Note: see the descriptions in stat_description[].
1348 static int ptc_seq_show(struct seq_file
*file
, void *data
)
1350 struct ptc_stats
*stat
;
1353 cpu
= *(loff_t
*)data
;
1356 "# cpu sent stime self locals remotes ncpus localhub ");
1358 "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1360 "numuvhubs4 numuvhubs2 numuvhubs1 dto retries rok ");
1362 "resetp resett giveup sto bz throt swack recv rtime ");
1364 "all one mult none retry canc nocan reset rcan ");
1366 "disable enable wars warshw warwaits\n");
1368 if (cpu
< num_possible_cpus() && cpu_online(cpu
)) {
1369 stat
= &per_cpu(ptcstats
, cpu
);
1370 /* source side statistics */
1372 "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1373 cpu
, stat
->s_requestor
, cycles_2_us(stat
->s_time
),
1374 stat
->s_ntargself
, stat
->s_ntarglocals
,
1375 stat
->s_ntargremotes
, stat
->s_ntargcpu
,
1376 stat
->s_ntarglocaluvhub
, stat
->s_ntargremoteuvhub
,
1377 stat
->s_ntarguvhub
, stat
->s_ntarguvhub16
);
1378 seq_printf(file
, "%ld %ld %ld %ld %ld ",
1379 stat
->s_ntarguvhub8
, stat
->s_ntarguvhub4
,
1380 stat
->s_ntarguvhub2
, stat
->s_ntarguvhub1
,
1382 seq_printf(file
, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1383 stat
->s_retry_messages
, stat
->s_retriesok
,
1384 stat
->s_resets_plug
, stat
->s_resets_timeout
,
1385 stat
->s_giveup
, stat
->s_stimeout
,
1386 stat
->s_busy
, stat
->s_throttles
);
1388 /* destination side statistics */
1390 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1391 read_gmmr_sw_ack(uv_cpu_to_pnode(cpu
)),
1392 stat
->d_requestee
, cycles_2_us(stat
->d_time
),
1393 stat
->d_alltlb
, stat
->d_onetlb
, stat
->d_multmsg
,
1394 stat
->d_nomsg
, stat
->d_retries
, stat
->d_canceled
,
1395 stat
->d_nocanceled
, stat
->d_resets
,
1397 seq_printf(file
, "%ld %ld %ld %ld %ld\n",
1398 stat
->s_bau_disabled
, stat
->s_bau_reenabled
,
1399 stat
->s_uv2_wars
, stat
->s_uv2_wars_hw
,
1400 stat
->s_uv2_war_waits
);
1406 * Display the tunables thru debugfs
1408 static ssize_t
tunables_read(struct file
*file
, char __user
*userbuf
,
1409 size_t count
, loff_t
*ppos
)
1414 buf
= kasprintf(GFP_KERNEL
, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1415 "max_concur plugged_delay plugsb4reset",
1416 "timeoutsb4reset ipi_reset_limit complete_threshold",
1417 "congested_response_us congested_reps congested_period",
1418 max_concurr
, plugged_delay
, plugsb4reset
,
1419 timeoutsb4reset
, ipi_reset_limit
, complete_threshold
,
1420 congested_respns_us
, congested_reps
, congested_period
);
1425 ret
= simple_read_from_buffer(userbuf
, count
, ppos
, buf
, strlen(buf
));
1431 * handle a write to /proc/sgi_uv/ptc_statistics
1432 * -1: reset the statistics
1433 * 0: display meaning of the statistics
1435 static ssize_t
ptc_proc_write(struct file
*file
, const char __user
*user
,
1436 size_t count
, loff_t
*data
)
1443 struct ptc_stats
*stat
;
1445 if (count
== 0 || count
> sizeof(optstr
))
1447 if (copy_from_user(optstr
, user
, count
))
1449 optstr
[count
- 1] = '\0';
1451 if (strict_strtol(optstr
, 10, &input_arg
) < 0) {
1452 printk(KERN_DEBUG
"%s is invalid\n", optstr
);
1456 if (input_arg
== 0) {
1457 elements
= sizeof(stat_description
)/sizeof(*stat_description
);
1458 printk(KERN_DEBUG
"# cpu: cpu number\n");
1459 printk(KERN_DEBUG
"Sender statistics:\n");
1460 for (i
= 0; i
< elements
; i
++)
1461 printk(KERN_DEBUG
"%s\n", stat_description
[i
]);
1462 } else if (input_arg
== -1) {
1463 for_each_present_cpu(cpu
) {
1464 stat
= &per_cpu(ptcstats
, cpu
);
1465 memset(stat
, 0, sizeof(struct ptc_stats
));
1472 static int local_atoi(const char *name
)
1479 val
= 10*val
+(*name
-'0');
1488 * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1489 * Zero values reset them to defaults.
1491 static int parse_tunables_write(struct bau_control
*bcp
, char *instr
,
1498 int e
= sizeof(tunables
) / sizeof(*tunables
);
1500 p
= instr
+ strspn(instr
, WHITESPACE
);
1502 for (; *p
; p
= q
+ strspn(q
, WHITESPACE
)) {
1503 q
= p
+ strcspn(p
, WHITESPACE
);
1509 printk(KERN_INFO
"bau tunable error: should be %d values\n", e
);
1513 p
= instr
+ strspn(instr
, WHITESPACE
);
1515 for (cnt
= 0; *p
; p
= q
+ strspn(q
, WHITESPACE
), cnt
++) {
1516 q
= p
+ strcspn(p
, WHITESPACE
);
1517 val
= local_atoi(p
);
1521 max_concurr
= MAX_BAU_CONCURRENT
;
1522 max_concurr_const
= MAX_BAU_CONCURRENT
;
1525 if (val
< 1 || val
> bcp
->cpus_in_uvhub
) {
1527 "Error: BAU max concurrent %d is invalid\n",
1532 max_concurr_const
= val
;
1536 *tunables
[cnt
].tunp
= tunables
[cnt
].deflt
;
1538 *tunables
[cnt
].tunp
= val
;
1548 * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1550 static ssize_t
tunables_write(struct file
*file
, const char __user
*user
,
1551 size_t count
, loff_t
*data
)
1556 struct bau_control
*bcp
;
1558 if (count
== 0 || count
> sizeof(instr
)-1)
1560 if (copy_from_user(instr
, user
, count
))
1563 instr
[count
] = '\0';
1566 bcp
= &per_cpu(bau_control
, cpu
);
1567 ret
= parse_tunables_write(bcp
, instr
, count
);
1572 for_each_present_cpu(cpu
) {
1573 bcp
= &per_cpu(bau_control
, cpu
);
1574 bcp
->max_concurr
= max_concurr
;
1575 bcp
->max_concurr_const
= max_concurr
;
1576 bcp
->plugged_delay
= plugged_delay
;
1577 bcp
->plugsb4reset
= plugsb4reset
;
1578 bcp
->timeoutsb4reset
= timeoutsb4reset
;
1579 bcp
->ipi_reset_limit
= ipi_reset_limit
;
1580 bcp
->complete_threshold
= complete_threshold
;
1581 bcp
->cong_response_us
= congested_respns_us
;
1582 bcp
->cong_reps
= congested_reps
;
1583 bcp
->cong_period
= congested_period
;
1588 static const struct seq_operations uv_ptc_seq_ops
= {
1589 .start
= ptc_seq_start
,
1590 .next
= ptc_seq_next
,
1591 .stop
= ptc_seq_stop
,
1592 .show
= ptc_seq_show
1595 static int ptc_proc_open(struct inode
*inode
, struct file
*file
)
1597 return seq_open(file
, &uv_ptc_seq_ops
);
1600 static int tunables_open(struct inode
*inode
, struct file
*file
)
1605 static const struct file_operations proc_uv_ptc_operations
= {
1606 .open
= ptc_proc_open
,
1608 .write
= ptc_proc_write
,
1609 .llseek
= seq_lseek
,
1610 .release
= seq_release
,
1613 static const struct file_operations tunables_fops
= {
1614 .open
= tunables_open
,
1615 .read
= tunables_read
,
1616 .write
= tunables_write
,
1617 .llseek
= default_llseek
,
1620 static int __init
uv_ptc_init(void)
1622 struct proc_dir_entry
*proc_uv_ptc
;
1624 if (!is_uv_system())
1627 proc_uv_ptc
= proc_create(UV_PTC_BASENAME
, 0444, NULL
,
1628 &proc_uv_ptc_operations
);
1630 printk(KERN_ERR
"unable to create %s proc entry\n",
1635 tunables_dir
= debugfs_create_dir(UV_BAU_TUNABLES_DIR
, NULL
);
1636 if (!tunables_dir
) {
1637 printk(KERN_ERR
"unable to create debugfs directory %s\n",
1638 UV_BAU_TUNABLES_DIR
);
1641 tunables_file
= debugfs_create_file(UV_BAU_TUNABLES_FILE
, 0600,
1642 tunables_dir
, NULL
, &tunables_fops
);
1643 if (!tunables_file
) {
1644 printk(KERN_ERR
"unable to create debugfs file %s\n",
1645 UV_BAU_TUNABLES_FILE
);
1652 * Initialize the sending side's sending buffers.
1654 static void activation_descriptor_init(int node
, int pnode
, int base_pnode
)
1663 struct bau_desc
*bau_desc
;
1664 struct bau_desc
*bd2
;
1665 struct uv1_bau_msg_header
*uv1_hdr
;
1666 struct uv2_bau_msg_header
*uv2_hdr
;
1667 struct bau_control
*bcp
;
1670 * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1671 * per cpu; and one per cpu on the uvhub (ADP_SZ)
1673 dsize
= sizeof(struct bau_desc
) * ADP_SZ
* ITEMS_PER_DESC
;
1674 bau_desc
= kmalloc_node(dsize
, GFP_KERNEL
, node
);
1677 gpa
= uv_gpa(bau_desc
);
1678 n
= uv_gpa_to_gnode(gpa
);
1679 m
= uv_gpa_to_offset(gpa
);
1683 /* the 14-bit pnode */
1684 write_mmr_descriptor_base(pnode
, (n
<< UV_DESC_PSHIFT
| m
));
1686 * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1687 * cpu even though we only use the first one; one descriptor can
1688 * describe a broadcast to 256 uv hubs.
1690 for (i
= 0, bd2
= bau_desc
; i
< (ADP_SZ
* ITEMS_PER_DESC
); i
++, bd2
++) {
1691 memset(bd2
, 0, sizeof(struct bau_desc
));
1693 uv1_hdr
= &bd2
->header
.uv1_hdr
;
1694 uv1_hdr
->swack_flag
= 1;
1696 * The base_dest_nasid set in the message header
1697 * is the nasid of the first uvhub in the partition.
1698 * The bit map will indicate destination pnode numbers
1699 * relative to that base. They may not be consecutive
1700 * if nasid striding is being used.
1702 uv1_hdr
->base_dest_nasid
=
1703 UV_PNODE_TO_NASID(base_pnode
);
1704 uv1_hdr
->dest_subnodeid
= UV_LB_SUBNODEID
;
1705 uv1_hdr
->command
= UV_NET_ENDPOINT_INTD
;
1706 uv1_hdr
->int_both
= 1;
1708 * all others need to be set to zero:
1709 * fairness chaining multilevel count replied_to
1712 uv2_hdr
= &bd2
->header
.uv2_hdr
;
1713 uv2_hdr
->swack_flag
= 1;
1714 uv2_hdr
->base_dest_nasid
=
1715 UV_PNODE_TO_NASID(base_pnode
);
1716 uv2_hdr
->dest_subnodeid
= UV_LB_SUBNODEID
;
1717 uv2_hdr
->command
= UV_NET_ENDPOINT_INTD
;
1720 for_each_present_cpu(cpu
) {
1721 if (pnode
!= uv_blade_to_pnode(uv_cpu_to_blade_id(cpu
)))
1723 bcp
= &per_cpu(bau_control
, cpu
);
1724 bcp
->descriptor_base
= bau_desc
;
1729 * initialize the destination side's receiving buffers
1730 * entered for each uvhub in the partition
1731 * - node is first node (kernel memory notion) on the uvhub
1732 * - pnode is the uvhub's physical identifier
1734 static void pq_init(int node
, int pnode
)
1741 unsigned long first
;
1742 unsigned long pn_first
;
1744 struct bau_pq_entry
*pqp
;
1745 struct bau_control
*bcp
;
1747 plsize
= (DEST_Q_SIZE
+ 1) * sizeof(struct bau_pq_entry
);
1748 vp
= kmalloc_node(plsize
, GFP_KERNEL
, node
);
1749 pqp
= (struct bau_pq_entry
*)vp
;
1752 cp
= (char *)pqp
+ 31;
1753 pqp
= (struct bau_pq_entry
*)(((unsigned long)cp
>> 5) << 5);
1755 for_each_present_cpu(cpu
) {
1756 if (pnode
!= uv_cpu_to_pnode(cpu
))
1758 /* for every cpu on this pnode: */
1759 bcp
= &per_cpu(bau_control
, cpu
);
1760 bcp
->queue_first
= pqp
;
1761 bcp
->bau_msg_head
= pqp
;
1762 bcp
->queue_last
= pqp
+ (DEST_Q_SIZE
- 1);
1765 * need the gnode of where the memory was really allocated
1767 pn
= uv_gpa_to_gnode(uv_gpa(pqp
));
1768 first
= uv_physnodeaddr(pqp
);
1769 pn_first
= ((unsigned long)pn
<< UV_PAYLOADQ_PNODE_SHIFT
) | first
;
1770 last
= uv_physnodeaddr(pqp
+ (DEST_Q_SIZE
- 1));
1771 write_mmr_payload_first(pnode
, pn_first
);
1772 write_mmr_payload_tail(pnode
, first
);
1773 write_mmr_payload_last(pnode
, last
);
1774 write_gmmr_sw_ack(pnode
, 0xffffUL
);
1776 /* in effect, all msg_type's are set to MSG_NOOP */
1777 memset(pqp
, 0, sizeof(struct bau_pq_entry
) * DEST_Q_SIZE
);
1781 * Initialization of each UV hub's structures
1783 static void __init
init_uvhub(int uvhub
, int vector
, int base_pnode
)
1787 unsigned long apicid
;
1789 node
= uvhub_to_first_node(uvhub
);
1790 pnode
= uv_blade_to_pnode(uvhub
);
1792 activation_descriptor_init(node
, pnode
, base_pnode
);
1794 pq_init(node
, pnode
);
1796 * The below initialization can't be in firmware because the
1797 * messaging IRQ will be determined by the OS.
1799 apicid
= uvhub_to_first_apicid(uvhub
) | uv_apicid_hibits
;
1800 write_mmr_data_config(pnode
, ((apicid
<< 32) | vector
));
1804 * We will set BAU_MISC_CONTROL with a timeout period.
1805 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1806 * So the destination timeout period has to be calculated from them.
1808 static int calculate_destination_timeout(void)
1810 unsigned long mmr_image
;
1816 unsigned long ts_ns
;
1819 mult1
= SOFTACK_TIMEOUT_PERIOD
& BAU_MISC_CONTROL_MULT_MASK
;
1820 mmr_image
= uv_read_local_mmr(UVH_AGING_PRESCALE_SEL
);
1821 index
= (mmr_image
>> BAU_URGENCY_7_SHIFT
) & BAU_URGENCY_7_MASK
;
1822 mmr_image
= uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT
);
1823 mult2
= (mmr_image
>> BAU_TRANS_SHIFT
) & BAU_TRANS_MASK
;
1824 base
= timeout_base_ns
[index
];
1825 ts_ns
= base
* mult1
* mult2
;
1828 /* 4 bits 0/1 for 10/80us base, 3 bits of multiplier */
1829 mmr_image
= uv_read_local_mmr(UVH_LB_BAU_MISC_CONTROL
);
1830 mmr_image
= (mmr_image
& UV_SA_MASK
) >> UV_SA_SHFT
;
1831 if (mmr_image
& (1L << UV2_ACK_UNITS_SHFT
))
1835 mult1
= mmr_image
& UV2_ACK_MASK
;
1841 static void __init
init_per_cpu_tunables(void)
1844 struct bau_control
*bcp
;
1846 for_each_present_cpu(cpu
) {
1847 bcp
= &per_cpu(bau_control
, cpu
);
1848 bcp
->baudisabled
= 0;
1849 bcp
->statp
= &per_cpu(ptcstats
, cpu
);
1850 /* time interval to catch a hardware stay-busy bug */
1851 bcp
->timeout_interval
= usec_2_cycles(2*timeout_us
);
1852 bcp
->max_concurr
= max_concurr
;
1853 bcp
->max_concurr_const
= max_concurr
;
1854 bcp
->plugged_delay
= plugged_delay
;
1855 bcp
->plugsb4reset
= plugsb4reset
;
1856 bcp
->timeoutsb4reset
= timeoutsb4reset
;
1857 bcp
->ipi_reset_limit
= ipi_reset_limit
;
1858 bcp
->complete_threshold
= complete_threshold
;
1859 bcp
->cong_response_us
= congested_respns_us
;
1860 bcp
->cong_reps
= congested_reps
;
1861 bcp
->cong_period
= congested_period
;
1862 bcp
->clocks_per_100_usec
= usec_2_cycles(100);
1863 spin_lock_init(&bcp
->queue_lock
);
1864 spin_lock_init(&bcp
->uvhub_lock
);
1869 * Scan all cpus to collect blade and socket summaries.
1871 static int __init
get_cpu_topology(int base_pnode
,
1872 struct uvhub_desc
*uvhub_descs
,
1873 unsigned char *uvhub_mask
)
1879 struct bau_control
*bcp
;
1880 struct uvhub_desc
*bdp
;
1881 struct socket_desc
*sdp
;
1883 for_each_present_cpu(cpu
) {
1884 bcp
= &per_cpu(bau_control
, cpu
);
1886 memset(bcp
, 0, sizeof(struct bau_control
));
1888 pnode
= uv_cpu_hub_info(cpu
)->pnode
;
1889 if ((pnode
- base_pnode
) >= UV_DISTRIBUTION_SIZE
) {
1891 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1892 cpu
, pnode
, base_pnode
, UV_DISTRIBUTION_SIZE
);
1896 bcp
->osnode
= cpu_to_node(cpu
);
1897 bcp
->partition_base_pnode
= base_pnode
;
1899 uvhub
= uv_cpu_hub_info(cpu
)->numa_blade_id
;
1900 *(uvhub_mask
+ (uvhub
/8)) |= (1 << (uvhub
%8));
1901 bdp
= &uvhub_descs
[uvhub
];
1907 /* kludge: 'assuming' one node per socket, and assuming that
1908 disabling a socket just leaves a gap in node numbers */
1909 socket
= bcp
->osnode
& 1;
1910 bdp
->socket_mask
|= (1 << socket
);
1911 sdp
= &bdp
->socket
[socket
];
1912 sdp
->cpu_number
[sdp
->num_cpus
] = cpu
;
1914 if (sdp
->num_cpus
> MAX_CPUS_PER_SOCKET
) {
1915 printk(KERN_EMERG
"%d cpus per socket invalid\n",
1924 * Each socket is to get a local array of pnodes/hubs.
1926 static void make_per_cpu_thp(struct bau_control
*smaster
)
1929 size_t hpsz
= sizeof(struct hub_and_pnode
) * num_possible_cpus();
1931 smaster
->thp
= kmalloc_node(hpsz
, GFP_KERNEL
, smaster
->osnode
);
1932 memset(smaster
->thp
, 0, hpsz
);
1933 for_each_present_cpu(cpu
) {
1934 smaster
->thp
[cpu
].pnode
= uv_cpu_hub_info(cpu
)->pnode
;
1935 smaster
->thp
[cpu
].uvhub
= uv_cpu_hub_info(cpu
)->numa_blade_id
;
1940 * Each uvhub is to get a local cpumask.
1942 static void make_per_hub_cpumask(struct bau_control
*hmaster
)
1944 int sz
= sizeof(cpumask_t
);
1946 hmaster
->cpumask
= kzalloc_node(sz
, GFP_KERNEL
, hmaster
->osnode
);
1950 * Initialize all the per_cpu information for the cpu's on a given socket,
1951 * given what has been gathered into the socket_desc struct.
1952 * And reports the chosen hub and socket masters back to the caller.
1954 static int scan_sock(struct socket_desc
*sdp
, struct uvhub_desc
*bdp
,
1955 struct bau_control
**smasterp
,
1956 struct bau_control
**hmasterp
)
1960 struct bau_control
*bcp
;
1962 for (i
= 0; i
< sdp
->num_cpus
; i
++) {
1963 cpu
= sdp
->cpu_number
[i
];
1964 bcp
= &per_cpu(bau_control
, cpu
);
1971 bcp
->cpus_in_uvhub
= bdp
->num_cpus
;
1972 bcp
->cpus_in_socket
= sdp
->num_cpus
;
1973 bcp
->socket_master
= *smasterp
;
1974 bcp
->uvhub
= bdp
->uvhub
;
1976 bcp
->uvhub_version
= 1;
1977 else if (is_uv2_hub())
1978 bcp
->uvhub_version
= 2;
1980 printk(KERN_EMERG
"uvhub version not 1 or 2\n");
1983 bcp
->uvhub_master
= *hmasterp
;
1984 bcp
->uvhub_cpu
= uv_cpu_hub_info(cpu
)->blade_processor_id
;
1985 bcp
->using_desc
= bcp
->uvhub_cpu
;
1986 if (bcp
->uvhub_cpu
>= MAX_CPUS_PER_UVHUB
) {
1987 printk(KERN_EMERG
"%d cpus per uvhub invalid\n",
1996 * Summarize the blade and socket topology into the per_cpu structures.
1998 static int __init
summarize_uvhub_sockets(int nuvhubs
,
1999 struct uvhub_desc
*uvhub_descs
,
2000 unsigned char *uvhub_mask
)
2004 unsigned short socket_mask
;
2006 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
2007 struct uvhub_desc
*bdp
;
2008 struct bau_control
*smaster
= NULL
;
2009 struct bau_control
*hmaster
= NULL
;
2011 if (!(*(uvhub_mask
+ (uvhub
/8)) & (1 << (uvhub
%8))))
2014 bdp
= &uvhub_descs
[uvhub
];
2015 socket_mask
= bdp
->socket_mask
;
2017 while (socket_mask
) {
2018 struct socket_desc
*sdp
;
2019 if ((socket_mask
& 1)) {
2020 sdp
= &bdp
->socket
[socket
];
2021 if (scan_sock(sdp
, bdp
, &smaster
, &hmaster
))
2023 make_per_cpu_thp(smaster
);
2026 socket_mask
= (socket_mask
>> 1);
2028 make_per_hub_cpumask(hmaster
);
2034 * initialize the bau_control structure for each cpu
2036 static int __init
init_per_cpu(int nuvhubs
, int base_part_pnode
)
2038 unsigned char *uvhub_mask
;
2040 struct uvhub_desc
*uvhub_descs
;
2042 timeout_us
= calculate_destination_timeout();
2044 vp
= kmalloc(nuvhubs
* sizeof(struct uvhub_desc
), GFP_KERNEL
);
2045 uvhub_descs
= (struct uvhub_desc
*)vp
;
2046 memset(uvhub_descs
, 0, nuvhubs
* sizeof(struct uvhub_desc
));
2047 uvhub_mask
= kzalloc((nuvhubs
+7)/8, GFP_KERNEL
);
2049 if (get_cpu_topology(base_part_pnode
, uvhub_descs
, uvhub_mask
))
2052 if (summarize_uvhub_sockets(nuvhubs
, uvhub_descs
, uvhub_mask
))
2057 init_per_cpu_tunables();
2067 * Initialization of BAU-related structures
2069 static int __init
uv_bau_init(void)
2077 cpumask_var_t
*mask
;
2079 if (!is_uv_system())
2085 for_each_possible_cpu(cur_cpu
) {
2086 mask
= &per_cpu(uv_flush_tlb_mask
, cur_cpu
);
2087 zalloc_cpumask_var_node(mask
, GFP_KERNEL
, cpu_to_node(cur_cpu
));
2090 nuvhubs
= uv_num_possible_blades();
2091 spin_lock_init(&disable_lock
);
2092 congested_cycles
= usec_2_cycles(congested_respns_us
);
2094 uv_base_pnode
= 0x7fffffff;
2095 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
2096 cpus
= uv_blade_nr_possible_cpus(uvhub
);
2097 if (cpus
&& (uv_blade_to_pnode(uvhub
) < uv_base_pnode
))
2098 uv_base_pnode
= uv_blade_to_pnode(uvhub
);
2103 if (init_per_cpu(nuvhubs
, uv_base_pnode
)) {
2108 vector
= UV_BAU_MESSAGE
;
2109 for_each_possible_blade(uvhub
)
2110 if (uv_blade_nr_possible_cpus(uvhub
))
2111 init_uvhub(uvhub
, vector
, uv_base_pnode
);
2113 alloc_intr_gate(vector
, uv_bau_message_intr1
);
2115 for_each_possible_blade(uvhub
) {
2116 if (uv_blade_nr_possible_cpus(uvhub
)) {
2119 pnode
= uv_blade_to_pnode(uvhub
);
2122 write_gmmr_activation(pnode
, val
);
2123 mmr
= 1; /* should be 1 to broadcast to both sockets */
2125 write_mmr_data_broadcast(pnode
, mmr
);
2131 core_initcall(uv_bau_init
);
2132 fs_initcall(uv_ptc_init
);