2 * SGI UltraViolet TLB flush routines.
4 * (c) 2008-2010 Cliff Wickman <cpw@sgi.com>, SGI.
6 * This code is released under the GNU General Public License version 2 or
9 #include <linux/seq_file.h>
10 #include <linux/proc_fs.h>
11 #include <linux/kernel.h>
12 #include <linux/slab.h>
14 #include <asm/mmu_context.h>
15 #include <asm/uv/uv.h>
16 #include <asm/uv/uv_mmrs.h>
17 #include <asm/uv/uv_hub.h>
18 #include <asm/uv/uv_bau.h>
22 #include <asm/irq_vectors.h>
23 #include <asm/timer.h>
26 struct bau_payload_queue_entry
*msg
;
29 struct bau_payload_queue_entry
*va_queue_first
;
30 struct bau_payload_queue_entry
*va_queue_last
;
33 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
34 static int timeout_base_ns
[] = {
44 static int timeout_us
;
46 #define UV_INTD_SOFT_ACK_TIMEOUT_PERIOD 0x000000000bUL
48 static int uv_bau_max_concurrent __read_mostly
;
51 static int __init
setup_nobau(char *arg
)
56 early_param("nobau", setup_nobau
);
58 /* base pnode in this partition */
59 static int uv_partition_base_pnode __read_mostly
;
60 /* position of pnode (which is nasid>>1): */
61 static int uv_nshift __read_mostly
;
62 static unsigned long uv_mmask __read_mostly
;
64 static DEFINE_PER_CPU(struct ptc_stats
, ptcstats
);
65 static DEFINE_PER_CPU(struct bau_control
, bau_control
);
66 static DEFINE_PER_CPU(cpumask_var_t
, uv_flush_tlb_mask
);
73 * Determine the first node on a uvhub. 'Nodes' are used for kernel
76 static int __init
uvhub_to_first_node(int uvhub
)
80 for_each_online_node(node
) {
81 b
= uv_node_to_blade_id(node
);
89 * Determine the apicid of the first cpu on a uvhub.
91 static int __init
uvhub_to_first_apicid(int uvhub
)
95 for_each_present_cpu(cpu
)
96 if (uvhub
== uv_cpu_to_blade_id(cpu
))
97 return per_cpu(x86_cpu_to_apicid
, cpu
);
102 * Free a software acknowledge hardware resource by clearing its Pending
103 * bit. This will return a reply to the sender.
104 * If the message has timed out, a reply has already been sent by the
105 * hardware but the resource has not been released. In that case our
106 * clear of the Timeout bit (as well) will free the resource. No reply will
107 * be sent (the hardware will only do one reply per message).
109 static inline void uv_reply_to_message(struct msg_desc
*mdp
,
110 struct bau_control
*bcp
)
113 struct bau_payload_queue_entry
*msg
;
116 if (!msg
->canceled
) {
117 dw
= (msg
->sw_ack_vector
<< UV_SW_ACK_NPENDING
) |
120 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS
, dw
);
123 msg
->sw_ack_vector
= 0;
127 * Process the receipt of a RETRY message
129 static inline void uv_bau_process_retry_msg(struct msg_desc
*mdp
,
130 struct bau_control
*bcp
)
133 int cancel_count
= 0;
135 unsigned long msg_res
;
136 unsigned long mmr
= 0;
137 struct bau_payload_queue_entry
*msg
;
138 struct bau_payload_queue_entry
*msg2
;
139 struct ptc_stats
*stat
;
142 stat
= &per_cpu(ptcstats
, bcp
->cpu
);
145 * cancel any message from msg+1 to the retry itself
147 for (msg2
= msg
+1, i
= 0; i
< DEST_Q_SIZE
; msg2
++, i
++) {
148 if (msg2
> mdp
->va_queue_last
)
149 msg2
= mdp
->va_queue_first
;
153 /* same conditions for cancellation as uv_do_reset */
154 if ((msg2
->replied_to
== 0) && (msg2
->canceled
== 0) &&
155 (msg2
->sw_ack_vector
) && ((msg2
->sw_ack_vector
&
156 msg
->sw_ack_vector
) == 0) &&
157 (msg2
->sending_cpu
== msg
->sending_cpu
) &&
158 (msg2
->msg_type
!= MSG_NOOP
)) {
159 slot2
= msg2
- mdp
->va_queue_first
;
160 mmr
= uv_read_local_mmr
161 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE
);
162 msg_res
= ((msg2
->sw_ack_vector
<< 8) |
163 msg2
->sw_ack_vector
);
165 * This is a message retry; clear the resources held
166 * by the previous message only if they timed out.
167 * If it has not timed out we have an unexpected
168 * situation to report.
170 if (mmr
& (msg_res
<< 8)) {
172 * is the resource timed out?
173 * make everyone ignore the cancelled message.
179 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS
,
180 (msg_res
<< 8) | msg_res
);
182 printk(KERN_INFO
"note bau retry: no effect\n");
186 stat
->d_nocanceled
++;
190 * Do all the things a cpu should do for a TLB shootdown message.
191 * Other cpu's may come here at the same time for this message.
193 static void uv_bau_process_message(struct msg_desc
*mdp
,
194 struct bau_control
*bcp
)
197 short socket_ack_count
= 0;
198 struct ptc_stats
*stat
;
199 struct bau_payload_queue_entry
*msg
;
200 struct bau_control
*smaster
= bcp
->socket_master
;
203 * This must be a normal message, or retry of a normal message
206 stat
= &per_cpu(ptcstats
, bcp
->cpu
);
207 if (msg
->address
== TLB_FLUSH_ALL
) {
211 __flush_tlb_one(msg
->address
);
217 * One cpu on each uvhub has the additional job on a RETRY
218 * of releasing the resource held by the message that is
219 * being retried. That message is identified by sending
222 if (msg
->msg_type
== MSG_RETRY
&& bcp
== bcp
->uvhub_master
)
223 uv_bau_process_retry_msg(mdp
, bcp
);
226 * This is a sw_ack message, so we have to reply to it.
227 * Count each responding cpu on the socket. This avoids
228 * pinging the count's cache line back and forth between
231 socket_ack_count
= atomic_add_short_return(1, (struct atomic_short
*)
232 &smaster
->socket_acknowledge_count
[mdp
->msg_slot
]);
233 if (socket_ack_count
== bcp
->cpus_in_socket
) {
235 * Both sockets dump their completed count total into
236 * the message's count.
238 smaster
->socket_acknowledge_count
[mdp
->msg_slot
] = 0;
239 msg_ack_count
= atomic_add_short_return(socket_ack_count
,
240 (struct atomic_short
*)&msg
->acknowledge_count
);
242 if (msg_ack_count
== bcp
->cpus_in_uvhub
) {
244 * All cpus in uvhub saw it; reply
246 uv_reply_to_message(mdp
, bcp
);
254 * Determine the first cpu on a uvhub.
256 static int uvhub_to_first_cpu(int uvhub
)
259 for_each_present_cpu(cpu
)
260 if (uvhub
== uv_cpu_to_blade_id(cpu
))
266 * Last resort when we get a large number of destination timeouts is
267 * to clear resources held by a given cpu.
268 * Do this with IPI so that all messages in the BAU message queue
269 * can be identified by their nonzero sw_ack_vector field.
271 * This is entered for a single cpu on the uvhub.
272 * The sender want's this uvhub to free a specific message's
276 uv_do_reset(void *ptr
)
282 unsigned long msg_res
;
283 struct bau_control
*bcp
;
284 struct reset_args
*rap
;
285 struct bau_payload_queue_entry
*msg
;
286 struct ptc_stats
*stat
;
288 bcp
= &per_cpu(bau_control
, smp_processor_id());
289 rap
= (struct reset_args
*)ptr
;
290 stat
= &per_cpu(ptcstats
, bcp
->cpu
);
294 * We're looking for the given sender, and
295 * will free its sw_ack resource.
296 * If all cpu's finally responded after the timeout, its
297 * message 'replied_to' was set.
299 for (msg
= bcp
->va_queue_first
, i
= 0; i
< DEST_Q_SIZE
; msg
++, i
++) {
300 /* uv_do_reset: same conditions for cancellation as
301 uv_bau_process_retry_msg() */
302 if ((msg
->replied_to
== 0) &&
303 (msg
->canceled
== 0) &&
304 (msg
->sending_cpu
== rap
->sender
) &&
305 (msg
->sw_ack_vector
) &&
306 (msg
->msg_type
!= MSG_NOOP
)) {
308 * make everyone else ignore this message
311 slot
= msg
- bcp
->va_queue_first
;
314 * only reset the resource if it is still pending
316 mmr
= uv_read_local_mmr
317 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE
);
318 msg_res
= ((msg
->sw_ack_vector
<< 8) |
323 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS
,
332 * Use IPI to get all target uvhubs to release resources held by
333 * a given sending cpu number.
335 static void uv_reset_with_ipi(struct bau_target_uvhubmask
*distribution
,
341 struct reset_args reset_args
;
343 reset_args
.sender
= sender
;
346 /* find a single cpu for each uvhub in this distribution mask */
348 uvhub
< sizeof(struct bau_target_uvhubmask
) * BITSPERBYTE
;
350 if (!bau_uvhub_isset(uvhub
, distribution
))
352 /* find a cpu for this uvhub */
353 cpu
= uvhub_to_first_cpu(uvhub
);
356 /* IPI all cpus; Preemption is already disabled */
357 smp_call_function_many(&mask
, uv_do_reset
, (void *)&reset_args
, 1);
361 static inline unsigned long
362 cycles_2_us(unsigned long long cyc
)
364 unsigned long long ns
;
366 ns
= (cyc
* per_cpu(cyc2ns
, smp_processor_id()))
367 >> CYC2NS_SCALE_FACTOR
;
373 * wait for all cpus on this hub to finish their sends and go quiet
374 * leaves uvhub_quiesce set so that no new broadcasts are started by
375 * bau_flush_send_and_wait()
378 quiesce_local_uvhub(struct bau_control
*hmaster
)
380 atomic_add_short_return(1, (struct atomic_short
*)
381 &hmaster
->uvhub_quiesce
);
385 * mark this quiet-requestor as done
388 end_uvhub_quiesce(struct bau_control
*hmaster
)
390 atomic_add_short_return(-1, (struct atomic_short
*)
391 &hmaster
->uvhub_quiesce
);
395 * Wait for completion of a broadcast software ack message
396 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
398 static int uv_wait_completion(struct bau_desc
*bau_desc
,
399 unsigned long mmr_offset
, int right_shift
, int this_cpu
,
400 struct bau_control
*bcp
, struct bau_control
*smaster
, long try)
403 unsigned long descriptor_status
;
407 cycles_t timeout_time
;
408 struct ptc_stats
*stat
= &per_cpu(ptcstats
, this_cpu
);
409 struct bau_control
*hmaster
;
411 hmaster
= bcp
->uvhub_master
;
412 timeout_time
= get_cycles() + bcp
->timeout_interval
;
414 /* spin on the status MMR, waiting for it to go idle */
415 while ((descriptor_status
= (((unsigned long)
416 uv_read_local_mmr(mmr_offset
) >>
417 right_shift
) & UV_ACT_STATUS_MASK
)) !=
420 * Our software ack messages may be blocked because there are
421 * no swack resources available. As long as none of them
422 * has timed out hardware will NACK our message and its
423 * state will stay IDLE.
425 if (descriptor_status
== DESC_STATUS_SOURCE_TIMEOUT
) {
428 } else if (descriptor_status
==
429 DESC_STATUS_DESTINATION_TIMEOUT
) {
431 ttime
= get_cycles();
434 * Our retries may be blocked by all destination
435 * swack resources being consumed, and a timeout
436 * pending. In that case hardware returns the
437 * ERROR that looks like a destination timeout.
439 if (cycles_2_us(ttime
- bcp
->send_message
) <
441 bcp
->conseccompletes
= 0;
442 return FLUSH_RETRY_PLUGGED
;
445 bcp
->conseccompletes
= 0;
446 return FLUSH_RETRY_TIMEOUT
;
449 * descriptor_status is still BUSY
453 if (relaxes
>= 10000) {
455 if (get_cycles() > timeout_time
) {
456 quiesce_local_uvhub(hmaster
);
458 /* single-thread the register change */
459 spin_lock(&hmaster
->masks_lock
);
460 mmr
= uv_read_local_mmr(mmr_offset
);
462 mask
|= (3UL < right_shift
);
465 uv_write_local_mmr(mmr_offset
, mmr
);
466 spin_unlock(&hmaster
->masks_lock
);
467 end_uvhub_quiesce(hmaster
);
474 bcp
->conseccompletes
++;
475 return FLUSH_COMPLETE
;
478 static inline cycles_t
479 sec_2_cycles(unsigned long sec
)
484 ns
= sec
* 1000000000;
485 cyc
= (ns
<< CYC2NS_SCALE_FACTOR
)/(per_cpu(cyc2ns
, smp_processor_id()));
490 * conditionally add 1 to *v, unless *v is >= u
491 * return 0 if we cannot add 1 to *v because it is >= u
492 * return 1 if we can add 1 to *v because it is < u
495 * This is close to atomic_add_unless(), but this allows the 'u' value
496 * to be lowered below the current 'v'. atomic_add_unless can only stop
499 static inline int atomic_inc_unless_ge(spinlock_t
*lock
, atomic_t
*v
, int u
)
502 if (atomic_read(v
) >= u
) {
512 * uv_flush_send_and_wait
514 * Send a broadcast and wait for it to complete.
516 * The flush_mask contains the cpus the broadcast is to be sent to, plus
517 * cpus that are on the local uvhub.
519 * Returns NULL if all flushing represented in the mask was done. The mask
521 * Returns @flush_mask if some remote flushing remains to be done. The
522 * mask will have some bits still set, representing any cpus on the local
523 * uvhub (not current cpu) and any on remote uvhubs if the broadcast failed.
525 const struct cpumask
*uv_flush_send_and_wait(struct bau_desc
*bau_desc
,
526 struct cpumask
*flush_mask
,
527 struct bau_control
*bcp
)
532 int completion_status
= 0;
535 int cpu
= bcp
->uvhub_cpu
;
536 int this_cpu
= bcp
->cpu
;
537 int this_uvhub
= bcp
->uvhub
;
538 unsigned long mmr_offset
;
542 struct ptc_stats
*stat
= &per_cpu(ptcstats
, bcp
->cpu
);
543 struct bau_control
*smaster
= bcp
->socket_master
;
544 struct bau_control
*hmaster
= bcp
->uvhub_master
;
547 * Spin here while there are hmaster->max_concurrent or more active
548 * descriptors. This is the per-uvhub 'throttle'.
550 if (!atomic_inc_unless_ge(&hmaster
->uvhub_lock
,
551 &hmaster
->active_descriptor_count
,
552 hmaster
->max_concurrent
)) {
556 } while (!atomic_inc_unless_ge(&hmaster
->uvhub_lock
,
557 &hmaster
->active_descriptor_count
,
558 hmaster
->max_concurrent
));
561 while (hmaster
->uvhub_quiesce
)
564 if (cpu
< UV_CPUS_PER_ACT_STATUS
) {
565 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_0
;
566 right_shift
= cpu
* UV_ACT_STATUS_SIZE
;
568 mmr_offset
= UVH_LB_BAU_SB_ACTIVATION_STATUS_1
;
570 ((cpu
- UV_CPUS_PER_ACT_STATUS
) * UV_ACT_STATUS_SIZE
);
572 time1
= get_cycles();
575 * Every message from any given cpu gets a unique message
576 * sequence number. But retries use that same number.
577 * Our message may have timed out at the destination because
578 * all sw-ack resources are in use and there is a timeout
579 * pending there. In that case, our last send never got
580 * placed into the queue and we need to persist until it
583 * Make any retry a type MSG_RETRY so that the destination will
584 * free any resource held by a previous message from this cpu.
587 /* use message type set by the caller the first time */
588 seq_number
= bcp
->message_number
++;
590 /* use RETRY type on all the rest; same sequence */
591 bau_desc
->header
.msg_type
= MSG_RETRY
;
592 stat
->s_retry_messages
++;
594 bau_desc
->header
.sequence
= seq_number
;
595 index
= (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT
) |
597 bcp
->send_message
= get_cycles();
599 uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL
, index
);
602 completion_status
= uv_wait_completion(bau_desc
, mmr_offset
,
603 right_shift
, this_cpu
, bcp
, smaster
, try);
605 if (completion_status
== FLUSH_RETRY_PLUGGED
) {
607 * Our retries may be blocked by all destination swack
608 * resources being consumed, and a timeout pending. In
609 * that case hardware immediately returns the ERROR
610 * that looks like a destination timeout.
612 udelay(TIMEOUT_DELAY
);
613 bcp
->plugged_tries
++;
614 if (bcp
->plugged_tries
>= PLUGSB4RESET
) {
615 bcp
->plugged_tries
= 0;
616 quiesce_local_uvhub(hmaster
);
617 spin_lock(&hmaster
->queue_lock
);
618 uv_reset_with_ipi(&bau_desc
->distribution
,
620 spin_unlock(&hmaster
->queue_lock
);
621 end_uvhub_quiesce(hmaster
);
623 stat
->s_resets_plug
++;
625 } else if (completion_status
== FLUSH_RETRY_TIMEOUT
) {
626 hmaster
->max_concurrent
= 1;
627 bcp
->timeout_tries
++;
628 udelay(TIMEOUT_DELAY
);
629 if (bcp
->timeout_tries
>= TIMEOUTSB4RESET
) {
630 bcp
->timeout_tries
= 0;
631 quiesce_local_uvhub(hmaster
);
632 spin_lock(&hmaster
->queue_lock
);
633 uv_reset_with_ipi(&bau_desc
->distribution
,
635 spin_unlock(&hmaster
->queue_lock
);
636 end_uvhub_quiesce(hmaster
);
638 stat
->s_resets_timeout
++;
641 if (bcp
->ipi_attempts
>= 3) {
642 bcp
->ipi_attempts
= 0;
643 completion_status
= FLUSH_GIVEUP
;
647 } while ((completion_status
== FLUSH_RETRY_PLUGGED
) ||
648 (completion_status
== FLUSH_RETRY_TIMEOUT
));
649 time2
= get_cycles();
651 if ((completion_status
== FLUSH_COMPLETE
) && (bcp
->conseccompletes
> 5)
652 && (hmaster
->max_concurrent
< hmaster
->max_concurrent_constant
))
653 hmaster
->max_concurrent
++;
656 * hold any cpu not timing out here; no other cpu currently held by
657 * the 'throttle' should enter the activation code
659 while (hmaster
->uvhub_quiesce
)
661 atomic_dec(&hmaster
->active_descriptor_count
);
663 /* guard against cycles wrap */
665 stat
->s_time
+= (time2
- time1
);
667 stat
->s_requestor
--; /* don't count this one */
668 if (completion_status
== FLUSH_COMPLETE
&& try > 1)
670 else if (completion_status
== FLUSH_GIVEUP
) {
672 * Cause the caller to do an IPI-style TLB shootdown on
673 * the target cpu's, all of which are still in the mask.
680 * Success, so clear the remote cpu's from the mask so we don't
681 * use the IPI method of shootdown on them.
683 for_each_cpu(bit
, flush_mask
) {
684 uvhub
= uv_cpu_to_blade_id(bit
);
685 if (uvhub
== this_uvhub
)
687 cpumask_clear_cpu(bit
, flush_mask
);
689 if (!cpumask_empty(flush_mask
))
696 * uv_flush_tlb_others - globally purge translation cache of a virtual
697 * address or all TLB's
698 * @cpumask: mask of all cpu's in which the address is to be removed
699 * @mm: mm_struct containing virtual address range
700 * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
701 * @cpu: the current cpu
703 * This is the entry point for initiating any UV global TLB shootdown.
705 * Purges the translation caches of all specified processors of the given
706 * virtual address, or purges all TLB's on specified processors.
708 * The caller has derived the cpumask from the mm_struct. This function
709 * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
711 * The cpumask is converted into a uvhubmask of the uvhubs containing
714 * Note that this function should be called with preemption disabled.
716 * Returns NULL if all remote flushing was done.
717 * Returns pointer to cpumask if some remote flushing remains to be
718 * done. The returned pointer is valid till preemption is re-enabled.
720 const struct cpumask
*uv_flush_tlb_others(const struct cpumask
*cpumask
,
721 struct mm_struct
*mm
,
722 unsigned long va
, unsigned int cpu
)
728 struct bau_desc
*bau_desc
;
729 struct cpumask
*flush_mask
;
730 struct ptc_stats
*stat
;
731 struct bau_control
*bcp
;
736 bcp
= &per_cpu(bau_control
, cpu
);
738 * Each sending cpu has a per-cpu mask which it fills from the caller's
739 * cpu mask. Only remote cpus are converted to uvhubs and copied.
741 flush_mask
= (struct cpumask
*)per_cpu(uv_flush_tlb_mask
, cpu
);
743 * copy cpumask to flush_mask, removing current cpu
744 * (current cpu should already have been flushed by the caller and
745 * should never be returned if we return flush_mask)
747 cpumask_andnot(flush_mask
, cpumask
, cpumask_of(cpu
));
748 if (cpu_isset(cpu
, *cpumask
))
749 locals
++; /* current cpu was targeted */
751 bau_desc
= bcp
->descriptor_base
;
752 bau_desc
+= UV_ITEMS_PER_DESCRIPTOR
* bcp
->uvhub_cpu
;
754 bau_uvhubs_clear(&bau_desc
->distribution
, UV_DISTRIBUTION_SIZE
);
756 for_each_cpu(tcpu
, flush_mask
) {
757 uvhub
= uv_cpu_to_blade_id(tcpu
);
758 if (uvhub
== bcp
->uvhub
) {
762 bau_uvhub_set(uvhub
, &bau_desc
->distribution
);
767 * No off_hub flushing; return status for local hub.
768 * Return the caller's mask if all were local (the current
769 * cpu may be in that mask).
776 stat
= &per_cpu(ptcstats
, cpu
);
778 stat
->s_ntargcpu
+= remotes
;
779 remotes
= bau_uvhub_weight(&bau_desc
->distribution
);
780 stat
->s_ntarguvhub
+= remotes
;
782 stat
->s_ntarguvhub16
++;
783 else if (remotes
>= 8)
784 stat
->s_ntarguvhub8
++;
785 else if (remotes
>= 4)
786 stat
->s_ntarguvhub4
++;
787 else if (remotes
>= 2)
788 stat
->s_ntarguvhub2
++;
790 stat
->s_ntarguvhub1
++;
792 bau_desc
->payload
.address
= va
;
793 bau_desc
->payload
.sending_cpu
= cpu
;
796 * uv_flush_send_and_wait returns null if all cpu's were messaged, or
797 * the adjusted flush_mask if any cpu's were not messaged.
799 return uv_flush_send_and_wait(bau_desc
, flush_mask
, bcp
);
803 * The BAU message interrupt comes here. (registered by set_intr_gate)
806 * We received a broadcast assist message.
808 * Interrupts are disabled; this interrupt could represent
809 * the receipt of several messages.
811 * All cores/threads on this hub get this interrupt.
812 * The last one to see it does the software ack.
813 * (the resource will not be freed until noninterruptable cpus see this
814 * interrupt; hardware may timeout the s/w ack and reply ERROR)
816 void uv_bau_message_interrupt(struct pt_regs
*regs
)
820 struct bau_payload_queue_entry
*msg
;
821 struct bau_control
*bcp
;
822 struct ptc_stats
*stat
;
823 struct msg_desc msgdesc
;
825 time_start
= get_cycles();
826 bcp
= &per_cpu(bau_control
, smp_processor_id());
827 stat
= &per_cpu(ptcstats
, smp_processor_id());
828 msgdesc
.va_queue_first
= bcp
->va_queue_first
;
829 msgdesc
.va_queue_last
= bcp
->va_queue_last
;
830 msg
= bcp
->bau_msg_head
;
831 while (msg
->sw_ack_vector
) {
833 msgdesc
.msg_slot
= msg
- msgdesc
.va_queue_first
;
834 msgdesc
.sw_ack_slot
= ffs(msg
->sw_ack_vector
) - 1;
836 uv_bau_process_message(&msgdesc
, bcp
);
838 if (msg
> msgdesc
.va_queue_last
)
839 msg
= msgdesc
.va_queue_first
;
840 bcp
->bau_msg_head
= msg
;
842 stat
->d_time
+= (get_cycles() - time_start
);
853 * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
854 * shootdown message timeouts enabled. The timeout does not cause
855 * an interrupt, but causes an error message to be returned to
858 static void uv_enable_timeouts(void)
863 unsigned long mmr_image
;
865 nuvhubs
= uv_num_possible_blades();
867 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++) {
868 if (!uv_blade_nr_possible_cpus(uvhub
))
871 pnode
= uv_blade_to_pnode(uvhub
);
873 uv_read_global_mmr64(pnode
, UVH_LB_BAU_MISC_CONTROL
);
875 * Set the timeout period and then lock it in, in three
876 * steps; captures and locks in the period.
878 * To program the period, the SOFT_ACK_MODE must be off.
880 mmr_image
&= ~((unsigned long)1 <<
881 UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT
);
882 uv_write_global_mmr64
883 (pnode
, UVH_LB_BAU_MISC_CONTROL
, mmr_image
);
885 * Set the 4-bit period.
887 mmr_image
&= ~((unsigned long)0xf <<
888 UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT
);
889 mmr_image
|= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD
<<
890 UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT
);
891 uv_write_global_mmr64
892 (pnode
, UVH_LB_BAU_MISC_CONTROL
, mmr_image
);
894 * Subsequent reversals of the timebase bit (3) cause an
895 * immediate timeout of one or all INTD resources as
896 * indicated in bits 2:0 (7 causes all of them to timeout).
898 mmr_image
|= ((unsigned long)1 <<
899 UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT
);
900 uv_write_global_mmr64
901 (pnode
, UVH_LB_BAU_MISC_CONTROL
, mmr_image
);
905 static void *uv_ptc_seq_start(struct seq_file
*file
, loff_t
*offset
)
907 if (*offset
< num_possible_cpus())
912 static void *uv_ptc_seq_next(struct seq_file
*file
, void *data
, loff_t
*offset
)
915 if (*offset
< num_possible_cpus())
920 static void uv_ptc_seq_stop(struct seq_file
*file
, void *data
)
924 static inline unsigned long long
925 microsec_2_cycles(unsigned long microsec
)
928 unsigned long long cyc
;
930 ns
= microsec
* 1000;
931 cyc
= (ns
<< CYC2NS_SCALE_FACTOR
)/(per_cpu(cyc2ns
, smp_processor_id()));
936 * Display the statistics thru /proc.
937 * 'data' points to the cpu number
939 static int uv_ptc_seq_show(struct seq_file
*file
, void *data
)
941 struct ptc_stats
*stat
;
944 cpu
= *(loff_t
*)data
;
948 "# cpu sent stime numuvhubs numuvhubs16 numuvhubs8 ");
950 "numuvhubs4 numuvhubs2 numuvhubs1 numcpus dto ");
952 "retries rok resetp resett giveup sto bz throt ");
954 "sw_ack recv rtime all ");
956 "one mult none retry canc nocan reset rcan\n");
958 if (cpu
< num_possible_cpus() && cpu_online(cpu
)) {
959 stat
= &per_cpu(ptcstats
, cpu
);
960 /* source side statistics */
962 "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
963 cpu
, stat
->s_requestor
, cycles_2_us(stat
->s_time
),
964 stat
->s_ntarguvhub
, stat
->s_ntarguvhub16
,
965 stat
->s_ntarguvhub8
, stat
->s_ntarguvhub4
,
966 stat
->s_ntarguvhub2
, stat
->s_ntarguvhub1
,
967 stat
->s_ntargcpu
, stat
->s_dtimeout
);
968 seq_printf(file
, "%ld %ld %ld %ld %ld %ld %ld %ld ",
969 stat
->s_retry_messages
, stat
->s_retriesok
,
970 stat
->s_resets_plug
, stat
->s_resets_timeout
,
971 stat
->s_giveup
, stat
->s_stimeout
,
972 stat
->s_busy
, stat
->s_throttles
);
973 /* destination side statistics */
975 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n",
976 uv_read_global_mmr64(uv_cpu_to_pnode(cpu
),
977 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE
),
978 stat
->d_requestee
, cycles_2_us(stat
->d_time
),
979 stat
->d_alltlb
, stat
->d_onetlb
, stat
->d_multmsg
,
980 stat
->d_nomsg
, stat
->d_retries
, stat
->d_canceled
,
981 stat
->d_nocanceled
, stat
->d_resets
,
989 * -1: resetf the statistics
990 * 0: display meaning of the statistics
991 * >0: maximum concurrent active descriptors per uvhub (throttle)
993 static ssize_t
uv_ptc_proc_write(struct file
*file
, const char __user
*user
,
994 size_t count
, loff_t
*data
)
999 struct ptc_stats
*stat
;
1000 struct bau_control
*bcp
;
1002 if (count
== 0 || count
> sizeof(optstr
))
1004 if (copy_from_user(optstr
, user
, count
))
1006 optstr
[count
- 1] = '\0';
1007 if (strict_strtol(optstr
, 10, &input_arg
) < 0) {
1008 printk(KERN_DEBUG
"%s is invalid\n", optstr
);
1012 if (input_arg
== 0) {
1013 printk(KERN_DEBUG
"# cpu: cpu number\n");
1014 printk(KERN_DEBUG
"Sender statistics:\n");
1016 "sent: number of shootdown messages sent\n");
1018 "stime: time spent sending messages\n");
1020 "numuvhubs: number of hubs targeted with shootdown\n");
1022 "numuvhubs16: number times 16 or more hubs targeted\n");
1024 "numuvhubs8: number times 8 or more hubs targeted\n");
1026 "numuvhubs4: number times 4 or more hubs targeted\n");
1028 "numuvhubs2: number times 2 or more hubs targeted\n");
1030 "numuvhubs1: number times 1 hub targeted\n");
1032 "numcpus: number of cpus targeted with shootdown\n");
1034 "dto: number of destination timeouts\n");
1036 "retries: destination timeout retries sent\n");
1038 "rok: : destination timeouts successfully retried\n");
1040 "resetp: ipi-style resource resets for plugs\n");
1042 "resett: ipi-style resource resets for timeouts\n");
1044 "giveup: fall-backs to ipi-style shootdowns\n");
1046 "sto: number of source timeouts\n");
1048 "bz: number of stay-busy's\n");
1050 "throt: number times spun in throttle\n");
1051 printk(KERN_DEBUG
"Destination side statistics:\n");
1053 "sw_ack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
1055 "recv: shootdown messages received\n");
1057 "rtime: time spent processing messages\n");
1059 "all: shootdown all-tlb messages\n");
1061 "one: shootdown one-tlb messages\n");
1063 "mult: interrupts that found multiple messages\n");
1065 "none: interrupts that found no messages\n");
1067 "retry: number of retry messages processed\n");
1069 "canc: number messages canceled by retries\n");
1071 "nocan: number retries that found nothing to cancel\n");
1073 "reset: number of ipi-style reset requests processed\n");
1075 "rcan: number messages canceled by reset requests\n");
1076 } else if (input_arg
== -1) {
1077 for_each_present_cpu(cpu
) {
1078 stat
= &per_cpu(ptcstats
, cpu
);
1079 memset(stat
, 0, sizeof(struct ptc_stats
));
1082 uv_bau_max_concurrent
= input_arg
;
1083 bcp
= &per_cpu(bau_control
, smp_processor_id());
1084 if (uv_bau_max_concurrent
< 1 ||
1085 uv_bau_max_concurrent
> bcp
->cpus_in_uvhub
) {
1087 "Error: BAU max concurrent %d; %d is invalid\n",
1088 bcp
->max_concurrent
, uv_bau_max_concurrent
);
1091 printk(KERN_DEBUG
"Set BAU max concurrent:%d\n",
1092 uv_bau_max_concurrent
);
1093 for_each_present_cpu(cpu
) {
1094 bcp
= &per_cpu(bau_control
, cpu
);
1095 bcp
->max_concurrent
= uv_bau_max_concurrent
;
1102 static const struct seq_operations uv_ptc_seq_ops
= {
1103 .start
= uv_ptc_seq_start
,
1104 .next
= uv_ptc_seq_next
,
1105 .stop
= uv_ptc_seq_stop
,
1106 .show
= uv_ptc_seq_show
1109 static int uv_ptc_proc_open(struct inode
*inode
, struct file
*file
)
1111 return seq_open(file
, &uv_ptc_seq_ops
);
1114 static const struct file_operations proc_uv_ptc_operations
= {
1115 .open
= uv_ptc_proc_open
,
1117 .write
= uv_ptc_proc_write
,
1118 .llseek
= seq_lseek
,
1119 .release
= seq_release
,
1122 static int __init
uv_ptc_init(void)
1124 struct proc_dir_entry
*proc_uv_ptc
;
1126 if (!is_uv_system())
1129 proc_uv_ptc
= proc_create(UV_PTC_BASENAME
, 0444, NULL
,
1130 &proc_uv_ptc_operations
);
1132 printk(KERN_ERR
"unable to create %s proc entry\n",
1140 * initialize the sending side's sending buffers
1143 uv_activation_descriptor_init(int node
, int pnode
)
1150 struct bau_desc
*bau_desc
;
1151 struct bau_desc
*bd2
;
1152 struct bau_control
*bcp
;
1155 * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR)
1156 * per cpu; and up to 32 (UV_ADP_SIZE) cpu's per uvhub
1158 bau_desc
= (struct bau_desc
*)kmalloc_node(sizeof(struct bau_desc
)*
1159 UV_ADP_SIZE
*UV_ITEMS_PER_DESCRIPTOR
, GFP_KERNEL
, node
);
1162 pa
= uv_gpa(bau_desc
); /* need the real nasid*/
1163 n
= pa
>> uv_nshift
;
1166 uv_write_global_mmr64(pnode
, UVH_LB_BAU_SB_DESCRIPTOR_BASE
,
1167 (n
<< UV_DESC_BASE_PNODE_SHIFT
| m
));
1170 * initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each
1171 * cpu even though we only use the first one; one descriptor can
1172 * describe a broadcast to 256 uv hubs.
1174 for (i
= 0, bd2
= bau_desc
; i
< (UV_ADP_SIZE
*UV_ITEMS_PER_DESCRIPTOR
);
1176 memset(bd2
, 0, sizeof(struct bau_desc
));
1177 bd2
->header
.sw_ack_flag
= 1;
1179 * base_dest_nodeid is the nasid (pnode<<1) of the first uvhub
1180 * in the partition. The bit map will indicate uvhub numbers,
1181 * which are 0-N in a partition. Pnodes are unique system-wide.
1183 bd2
->header
.base_dest_nodeid
= uv_partition_base_pnode
<< 1;
1184 bd2
->header
.dest_subnodeid
= 0x10; /* the LB */
1185 bd2
->header
.command
= UV_NET_ENDPOINT_INTD
;
1186 bd2
->header
.int_both
= 1;
1188 * all others need to be set to zero:
1189 * fairness chaining multilevel count replied_to
1192 for_each_present_cpu(cpu
) {
1193 if (pnode
!= uv_blade_to_pnode(uv_cpu_to_blade_id(cpu
)))
1195 bcp
= &per_cpu(bau_control
, cpu
);
1196 bcp
->descriptor_base
= bau_desc
;
1201 * initialize the destination side's receiving buffers
1202 * entered for each uvhub in the partition
1203 * - node is first node (kernel memory notion) on the uvhub
1204 * - pnode is the uvhub's physical identifier
1207 uv_payload_queue_init(int node
, int pnode
)
1213 struct bau_payload_queue_entry
*pqp
;
1214 struct bau_payload_queue_entry
*pqp_malloc
;
1215 struct bau_control
*bcp
;
1217 pqp
= (struct bau_payload_queue_entry
*) kmalloc_node(
1218 (DEST_Q_SIZE
+ 1) * sizeof(struct bau_payload_queue_entry
),
1223 cp
= (char *)pqp
+ 31;
1224 pqp
= (struct bau_payload_queue_entry
*)(((unsigned long)cp
>> 5) << 5);
1226 for_each_present_cpu(cpu
) {
1227 if (pnode
!= uv_cpu_to_pnode(cpu
))
1229 /* for every cpu on this pnode: */
1230 bcp
= &per_cpu(bau_control
, cpu
);
1231 bcp
->va_queue_first
= pqp
;
1232 bcp
->bau_msg_head
= pqp
;
1233 bcp
->va_queue_last
= pqp
+ (DEST_Q_SIZE
- 1);
1236 * need the pnode of where the memory was really allocated
1239 pn
= pa
>> uv_nshift
;
1240 uv_write_global_mmr64(pnode
,
1241 UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST
,
1242 ((unsigned long)pn
<< UV_PAYLOADQ_PNODE_SHIFT
) |
1243 uv_physnodeaddr(pqp
));
1244 uv_write_global_mmr64(pnode
, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL
,
1245 uv_physnodeaddr(pqp
));
1246 uv_write_global_mmr64(pnode
, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST
,
1248 uv_physnodeaddr(pqp
+ (DEST_Q_SIZE
- 1)));
1249 /* in effect, all msg_type's are set to MSG_NOOP */
1250 memset(pqp
, 0, sizeof(struct bau_payload_queue_entry
) * DEST_Q_SIZE
);
1254 * Initialization of each UV hub's structures
1256 static void __init
uv_init_uvhub(int uvhub
, int vector
)
1260 unsigned long apicid
;
1262 node
= uvhub_to_first_node(uvhub
);
1263 pnode
= uv_blade_to_pnode(uvhub
);
1264 uv_activation_descriptor_init(node
, pnode
);
1265 uv_payload_queue_init(node
, pnode
);
1267 * the below initialization can't be in firmware because the
1268 * messaging IRQ will be determined by the OS
1270 apicid
= uvhub_to_first_apicid(uvhub
);
1271 uv_write_global_mmr64(pnode
, UVH_BAU_DATA_CONFIG
,
1272 ((apicid
<< 32) | vector
));
1276 * We will set BAU_MISC_CONTROL with a timeout period.
1277 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1278 * So the destination timeout period has be be calculated from them.
1281 calculate_destination_timeout(void)
1283 unsigned long mmr_image
;
1289 unsigned long ts_ns
;
1291 mult1
= UV_INTD_SOFT_ACK_TIMEOUT_PERIOD
& BAU_MISC_CONTROL_MULT_MASK
;
1292 mmr_image
= uv_read_local_mmr(UVH_AGING_PRESCALE_SEL
);
1293 index
= (mmr_image
>> BAU_URGENCY_7_SHIFT
) & BAU_URGENCY_7_MASK
;
1294 mmr_image
= uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT
);
1295 mult2
= (mmr_image
>> BAU_TRANS_SHIFT
) & BAU_TRANS_MASK
;
1296 base
= timeout_base_ns
[index
];
1297 ts_ns
= base
* mult1
* mult2
;
1303 * initialize the bau_control structure for each cpu
1305 static void uv_init_per_cpu(int nuvhubs
)
1312 struct bau_control
*bcp
;
1313 struct uvhub_desc
*bdp
;
1314 struct socket_desc
*sdp
;
1315 struct bau_control
*hmaster
= NULL
;
1316 struct bau_control
*smaster
= NULL
;
1317 struct socket_desc
{
1319 short cpu_number
[16];
1326 struct socket_desc socket
[2];
1328 struct uvhub_desc
*uvhub_descs
;
1330 timeout_us
= calculate_destination_timeout();
1332 uvhub_descs
= (struct uvhub_desc
*)
1333 kmalloc(nuvhubs
* sizeof(struct uvhub_desc
), GFP_KERNEL
);
1334 memset(uvhub_descs
, 0, nuvhubs
* sizeof(struct uvhub_desc
));
1335 for_each_present_cpu(cpu
) {
1336 bcp
= &per_cpu(bau_control
, cpu
);
1337 memset(bcp
, 0, sizeof(struct bau_control
));
1338 spin_lock_init(&bcp
->masks_lock
);
1339 bcp
->max_concurrent
= uv_bau_max_concurrent
;
1340 pnode
= uv_cpu_hub_info(cpu
)->pnode
;
1341 uvhub
= uv_cpu_hub_info(cpu
)->numa_blade_id
;
1342 bdp
= &uvhub_descs
[uvhub
];
1346 /* time interval to catch a hardware stay-busy bug */
1347 bcp
->timeout_interval
= microsec_2_cycles(2*timeout_us
);
1348 /* kludge: assume uv_hub.h is constant */
1349 socket
= (cpu_physical_id(cpu
)>>5)&1;
1350 if (socket
>= bdp
->num_sockets
)
1351 bdp
->num_sockets
= socket
+1;
1352 sdp
= &bdp
->socket
[socket
];
1353 sdp
->cpu_number
[sdp
->num_cpus
] = cpu
;
1357 for_each_possible_blade(uvhub
) {
1358 bdp
= &uvhub_descs
[uvhub
];
1359 for (i
= 0; i
< bdp
->num_sockets
; i
++) {
1360 sdp
= &bdp
->socket
[i
];
1361 for (j
= 0; j
< sdp
->num_cpus
; j
++) {
1362 cpu
= sdp
->cpu_number
[j
];
1363 bcp
= &per_cpu(bau_control
, cpu
);
1370 bcp
->cpus_in_uvhub
= bdp
->num_cpus
;
1371 bcp
->cpus_in_socket
= sdp
->num_cpus
;
1372 bcp
->socket_master
= smaster
;
1373 bcp
->uvhub_master
= hmaster
;
1374 for (k
= 0; k
< DEST_Q_SIZE
; k
++)
1375 bcp
->socket_acknowledge_count
[k
] = 0;
1377 uv_cpu_hub_info(cpu
)->blade_processor_id
;
1386 * Initialization of BAU-related structures
1388 static int __init
uv_bau_init(void)
1397 if (!is_uv_system())
1403 for_each_possible_cpu(cur_cpu
)
1404 zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask
, cur_cpu
),
1405 GFP_KERNEL
, cpu_to_node(cur_cpu
));
1407 uv_bau_max_concurrent
= MAX_BAU_CONCURRENT
;
1408 uv_nshift
= uv_hub_info
->m_val
;
1409 uv_mmask
= (1UL << uv_hub_info
->m_val
) - 1;
1410 nuvhubs
= uv_num_possible_blades();
1412 uv_init_per_cpu(nuvhubs
);
1414 uv_partition_base_pnode
= 0x7fffffff;
1415 for (uvhub
= 0; uvhub
< nuvhubs
; uvhub
++)
1416 if (uv_blade_nr_possible_cpus(uvhub
) &&
1417 (uv_blade_to_pnode(uvhub
) < uv_partition_base_pnode
))
1418 uv_partition_base_pnode
= uv_blade_to_pnode(uvhub
);
1420 vector
= UV_BAU_MESSAGE
;
1421 for_each_possible_blade(uvhub
)
1422 if (uv_blade_nr_possible_cpus(uvhub
))
1423 uv_init_uvhub(uvhub
, vector
);
1425 uv_enable_timeouts();
1426 alloc_intr_gate(vector
, uv_bau_message_intr1
);
1428 for_each_possible_blade(uvhub
) {
1429 pnode
= uv_blade_to_pnode(uvhub
);
1431 uv_write_global_mmr64(pnode
, UVH_LB_BAU_SB_ACTIVATION_CONTROL
,
1432 ((unsigned long)1 << 63));
1433 mmr
= 1; /* should be 1 to broadcast to both sockets */
1434 uv_write_global_mmr64(pnode
, UVH_BAU_DATA_BROADCAST
, mmr
);
1439 core_initcall(uv_bau_init
);
1440 core_initcall(uv_ptc_init
);