Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/cjb/mmc
[pandora-kernel.git] / arch / x86 / platform / uv / tlb_uv.c
1 /*
2  *      SGI UltraViolet TLB flush routines.
3  *
4  *      (c) 2008-2011 Cliff Wickman <cpw@sgi.com>, SGI.
5  *
6  *      This code is released under the GNU General Public License version 2 or
7  *      later.
8  */
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>
15
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>
21 #include <asm/apic.h>
22 #include <asm/idle.h>
23 #include <asm/tsc.h>
24 #include <asm/irq_vectors.h>
25 #include <asm/timer.h>
26
27 /* timeouts in nanoseconds (indexed by UVH_AGING_PRESCALE_SEL urgency7 30:28) */
28 static int timeout_base_ns[] = {
29                 20,
30                 160,
31                 1280,
32                 10240,
33                 81920,
34                 655360,
35                 5242880,
36                 167772160
37 };
38
39 static int timeout_us;
40 static int nobau;
41 static int baudisabled;
42 static spinlock_t disable_lock;
43 static cycles_t congested_cycles;
44
45 /* tunables: */
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;
56
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}
67 };
68
69 static struct dentry *tunables_dir;
70 static struct dentry *tunables_file;
71
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"
106 };
107
108 static int __init
109 setup_nobau(char *arg)
110 {
111         nobau = 1;
112         return 0;
113 }
114 early_param("nobau", setup_nobau);
115
116 /* base pnode in this partition */
117 static int uv_base_pnode __read_mostly;
118
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);
122
123 /*
124  * Determine the first node on a uvhub. 'Nodes' are used for kernel
125  * memory allocation.
126  */
127 static int __init uvhub_to_first_node(int uvhub)
128 {
129         int node, b;
130
131         for_each_online_node(node) {
132                 b = uv_node_to_blade_id(node);
133                 if (uvhub == b)
134                         return node;
135         }
136         return -1;
137 }
138
139 /*
140  * Determine the apicid of the first cpu on a uvhub.
141  */
142 static int __init uvhub_to_first_apicid(int uvhub)
143 {
144         int cpu;
145
146         for_each_present_cpu(cpu)
147                 if (uvhub == uv_cpu_to_blade_id(cpu))
148                         return per_cpu(x86_cpu_to_apicid, cpu);
149         return -1;
150 }
151
152 /*
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).
159  */
160 static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp)
161 {
162         unsigned long dw;
163         struct bau_pq_entry *msg;
164
165         msg = mdp->msg;
166         if (!msg->canceled) {
167                 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
168                 write_mmr_sw_ack(dw);
169         }
170         msg->replied_to = 1;
171         msg->swack_vec = 0;
172 }
173
174 /*
175  * Process the receipt of a RETRY message
176  */
177 static void bau_process_retry_msg(struct msg_desc *mdp,
178                                         struct bau_control *bcp)
179 {
180         int i;
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;
187
188         stat->d_retries++;
189         /*
190          * cancel any message from msg+1 to the retry itself
191          */
192         for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
193                 if (msg2 > mdp->queue_last)
194                         msg2 = mdp->queue_first;
195                 if (msg2 == msg)
196                         break;
197
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;
206                         /*
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.
211                          */
212                         if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
213                                 unsigned long mr;
214                                 /*
215                                  * is the resource timed out?
216                                  * make everyone ignore the cancelled message.
217                                  */
218                                 msg2->canceled = 1;
219                                 stat->d_canceled++;
220                                 cancel_count++;
221                                 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
222                                 write_mmr_sw_ack(mr);
223                         }
224                 }
225         }
226         if (!cancel_count)
227                 stat->d_nocanceled++;
228 }
229
230 /*
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.
233  */
234 static void bau_process_message(struct msg_desc *mdp,
235                                         struct bau_control *bcp)
236 {
237         short socket_ack_count = 0;
238         short *sp;
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;
243
244         /*
245          * This must be a normal message, or retry of a normal message
246          */
247         if (msg->address == TLB_FLUSH_ALL) {
248                 local_flush_tlb();
249                 stat->d_alltlb++;
250         } else {
251                 __flush_tlb_one(msg->address);
252                 stat->d_onetlb++;
253         }
254         stat->d_requestee++;
255
256         /*
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
260          * cpu number.
261          */
262         if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
263                 bau_process_retry_msg(mdp, bcp);
264
265         /*
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
269          * the sockets.
270          */
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) {
275                 int msg_ack_count;
276                 /*
277                  * Both sockets dump their completed count total into
278                  * the message's count.
279                  */
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);
283
284                 if (msg_ack_count == bcp->cpus_in_uvhub) {
285                         /*
286                          * All cpus in uvhub saw it; reply
287                          */
288                         reply_to_message(mdp, bcp);
289                 }
290         }
291
292         return;
293 }
294
295 /*
296  * Determine the first cpu on a pnode.
297  */
298 static int pnode_to_first_cpu(int pnode, struct bau_control *smaster)
299 {
300         int cpu;
301         struct hub_and_pnode *hpp;
302
303         for_each_present_cpu(cpu) {
304                 hpp = &smaster->thp[cpu];
305                 if (pnode == hpp->pnode)
306                         return cpu;
307         }
308         return -1;
309 }
310
311 /*
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.
316  *
317  * This is entered for a single cpu on the uvhub.
318  * The sender want's this uvhub to free a specific message's
319  * swack resources.
320  */
321 static void do_reset(void *ptr)
322 {
323         int i;
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;
328
329         stat->d_resets++;
330         /*
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.
335          */
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) &&
343                     (msg->swack_vec) &&
344                     (msg->msg_type != MSG_NOOP)) {
345                         unsigned long mmr;
346                         unsigned long mr;
347                         /*
348                          * make everyone else ignore this message
349                          */
350                         msg->canceled = 1;
351                         /*
352                          * only reset the resource if it is still pending
353                          */
354                         mmr = read_mmr_sw_ack();
355                         msg_res = msg->swack_vec;
356                         mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
357                         if (mmr & msg_res) {
358                                 stat->d_rcanceled++;
359                                 write_mmr_sw_ack(mr);
360                         }
361                 }
362         }
363         return;
364 }
365
366 /*
367  * Use IPI to get all target uvhubs to release resources held by
368  * a given sending cpu number.
369  */
370 static void reset_with_ipi(struct pnmask *distribution, struct bau_control *bcp)
371 {
372         int pnode;
373         int apnode;
374         int maskbits;
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;
379
380         reset_args.sender = sender;
381         cpus_clear(*mask);
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++) {
386                 int cpu;
387                 if (!bau_uvhub_isset(pnode, distribution))
388                         continue;
389                 apnode = pnode + bcp->partition_base_pnode;
390                 cpu = pnode_to_first_cpu(apnode, smaster);
391                 cpu_set(cpu, *mask);
392         }
393
394         /* IPI all cpus; preemption is already disabled */
395         smp_call_function_many(mask, do_reset, (void *)&reset_args, 1);
396         return;
397 }
398
399 static inline unsigned long cycles_2_us(unsigned long long cyc)
400 {
401         unsigned long long ns;
402         unsigned long us;
403         int cpu = smp_processor_id();
404
405         ns =  (cyc * per_cpu(cyc2ns, cpu)) >> CYC2NS_SCALE_FACTOR;
406         us = ns / 1000;
407         return us;
408 }
409
410 /*
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()
414  */
415 static inline void quiesce_local_uvhub(struct bau_control *hmaster)
416 {
417         atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
418 }
419
420 /*
421  * mark this quiet-requestor as done
422  */
423 static inline void end_uvhub_quiesce(struct bau_control *hmaster)
424 {
425         atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
426 }
427
428 static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
429 {
430         unsigned long descriptor_status;
431
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;
436 }
437
438 /*
439  * Wait for completion of a broadcast software ack message
440  * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
441  */
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)
445 {
446         unsigned long descriptor_status;
447         cycles_t ttm;
448         struct ptc_stats *stat = bcp->statp;
449
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)) {
453                 /*
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.
458                  */
459                 if (descriptor_status == DS_SOURCE_TIMEOUT) {
460                         stat->s_stimeout++;
461                         return FLUSH_GIVEUP;
462                 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
463                         stat->s_dtimeout++;
464                         ttm = get_cycles();
465
466                         /*
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.
471                          */
472                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
473                                 bcp->conseccompletes = 0;
474                                 return FLUSH_RETRY_PLUGGED;
475                         }
476
477                         bcp->conseccompletes = 0;
478                         return FLUSH_RETRY_TIMEOUT;
479                 } else {
480                         /*
481                          * descriptor_status is still BUSY
482                          */
483                         cpu_relax();
484                 }
485                 descriptor_status = uv1_read_status(mmr_offset, right_shift);
486         }
487         bcp->conseccompletes++;
488         return FLUSH_COMPLETE;
489 }
490
491 /*
492  * UV2 has an extra bit of status in the ACTIVATION_STATUS_2 register.
493  */
494 static unsigned long uv2_read_status(unsigned long offset, int rshft, int cpu)
495 {
496         unsigned long descriptor_status;
497         unsigned long descriptor_status2;
498
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;
503 }
504
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)
508 {
509         unsigned long descriptor_stat;
510         cycles_t ttm;
511         int cpu = bcp->uvhub_cpu;
512         struct ptc_stats *stat = bcp->statp;
513
514         descriptor_stat = uv2_read_status(mmr_offset, right_shift, cpu);
515
516         /* spin on the status MMR, waiting for it to go idle */
517         while (descriptor_stat != UV2H_DESC_IDLE) {
518                 /*
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.
523                  */
524                 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT) ||
525                     (descriptor_stat == UV2H_DESC_DEST_STRONG_NACK) ||
526                     (descriptor_stat == UV2H_DESC_DEST_PUT_ERR)) {
527                         stat->s_stimeout++;
528                         return FLUSH_GIVEUP;
529                 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
530                         stat->s_dtimeout++;
531                         ttm = get_cycles();
532                         /*
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.
537                          */
538                         if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
539                                 bcp->conseccompletes = 0;
540                                 return FLUSH_RETRY_PLUGGED;
541                         }
542                         bcp->conseccompletes = 0;
543                         return FLUSH_RETRY_TIMEOUT;
544                 } else {
545                         /*
546                          * descriptor_stat is still BUSY
547                          */
548                         cpu_relax();
549                 }
550                 descriptor_stat = uv2_read_status(mmr_offset, right_shift, cpu);
551         }
552         bcp->conseccompletes++;
553         return FLUSH_COMPLETE;
554 }
555
556 /*
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
559  * current hub.
560  */
561 static int wait_completion(struct bau_desc *bau_desc,
562                                 struct bau_control *bcp, long try)
563 {
564         int right_shift;
565         unsigned long mmr_offset;
566         int cpu = bcp->uvhub_cpu;
567
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;
571         } else {
572                 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
573                 right_shift = ((cpu - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
574         }
575
576         if (is_uv1_hub())
577                 return uv1_wait_completion(bau_desc, mmr_offset, right_shift,
578                                                                 bcp, try);
579         else
580                 return uv2_wait_completion(bau_desc, mmr_offset, right_shift,
581                                                                 bcp, try);
582 }
583
584 static inline cycles_t sec_2_cycles(unsigned long sec)
585 {
586         unsigned long ns;
587         cycles_t cyc;
588
589         ns = sec * 1000000000;
590         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
591         return cyc;
592 }
593
594 /*
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.
598  */
599 static void destination_plugged(struct bau_desc *bau_desc,
600                         struct bau_control *bcp,
601                         struct bau_control *hmaster, struct ptc_stats *stat)
602 {
603         udelay(bcp->plugged_delay);
604         bcp->plugged_tries++;
605
606         if (bcp->plugged_tries >= bcp->plugsb4reset) {
607                 bcp->plugged_tries = 0;
608
609                 quiesce_local_uvhub(hmaster);
610
611                 spin_lock(&hmaster->queue_lock);
612                 reset_with_ipi(&bau_desc->distribution, bcp);
613                 spin_unlock(&hmaster->queue_lock);
614
615                 end_uvhub_quiesce(hmaster);
616
617                 bcp->ipi_attempts++;
618                 stat->s_resets_plug++;
619         }
620 }
621
622 static void destination_timeout(struct bau_desc *bau_desc,
623                         struct bau_control *bcp, struct bau_control *hmaster,
624                         struct ptc_stats *stat)
625 {
626         hmaster->max_concurr = 1;
627         bcp->timeout_tries++;
628         if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
629                 bcp->timeout_tries = 0;
630
631                 quiesce_local_uvhub(hmaster);
632
633                 spin_lock(&hmaster->queue_lock);
634                 reset_with_ipi(&bau_desc->distribution, bcp);
635                 spin_unlock(&hmaster->queue_lock);
636
637                 end_uvhub_quiesce(hmaster);
638
639                 bcp->ipi_attempts++;
640                 stat->s_resets_timeout++;
641         }
642 }
643
644 /*
645  * Completions are taking a very long time due to a congested numalink
646  * network.
647  */
648 static void disable_for_congestion(struct bau_control *bcp,
649                                         struct ptc_stats *stat)
650 {
651         /* let only one cpu do this disabling */
652         spin_lock(&disable_lock);
653
654         if (!baudisabled && bcp->period_requests &&
655             ((bcp->period_time / bcp->period_requests) > congested_cycles)) {
656                 int tcpu;
657                 struct bau_control *tbcp;
658                 /* it becomes this cpu's job to turn on the use of the
659                    BAU again */
660                 baudisabled = 1;
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;
668                 }
669         }
670
671         spin_unlock(&disable_lock);
672 }
673
674 static void count_max_concurr(int stat, struct bau_control *bcp,
675                                 struct bau_control *hmaster)
676 {
677         bcp->plugged_tries = 0;
678         bcp->timeout_tries = 0;
679         if (stat != FLUSH_COMPLETE)
680                 return;
681         if (bcp->conseccompletes <= bcp->complete_threshold)
682                 return;
683         if (hmaster->max_concurr >= hmaster->max_concurr_const)
684                 return;
685         hmaster->max_concurr++;
686 }
687
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)
691 {
692         cycles_t elapsed;
693
694         if (time2 > time1) {
695                 elapsed = time2 - time1;
696                 stat->s_time += elapsed;
697
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);
704                 }
705         } else
706                 stat->s_requestor--;
707
708         if (completion_status == FLUSH_COMPLETE && try > 1)
709                 stat->s_retriesok++;
710         else if (completion_status == FLUSH_GIVEUP)
711                 stat->s_giveup++;
712 }
713
714 /*
715  * Because of a uv1 hardware bug only a limited number of concurrent
716  * requests can be made.
717  */
718 static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
719 {
720         spinlock_t *lock = &hmaster->uvhub_lock;
721         atomic_t *v;
722
723         v = &hmaster->active_descriptor_count;
724         if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
725                 stat->s_throttles++;
726                 do {
727                         cpu_relax();
728                 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
729         }
730 }
731
732 /*
733  * Handle the completion status of a message send.
734  */
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)
738 {
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);
743 }
744
745 /*
746  * Send a broadcast and wait for it to complete.
747  *
748  * The flush_mask contains the cpus the broadcast is to be sent to including
749  * cpus that are on the local uvhub.
750  *
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.
754  */
755 int uv_flush_send_and_wait(struct bau_desc *bau_desc,
756                         struct cpumask *flush_mask, struct bau_control *bcp)
757 {
758         int seq_number = 0;
759         int completion_stat = 0;
760         long try = 0;
761         unsigned long index;
762         cycles_t time1;
763         cycles_t time2;
764         struct ptc_stats *stat = bcp->statp;
765         struct bau_control *hmaster = bcp->uvhub_master;
766
767         if (is_uv1_hub())
768                 uv1_throttle(hmaster, stat);
769
770         while (hmaster->uvhub_quiesce)
771                 cpu_relax();
772
773         time1 = get_cycles();
774         do {
775                 if (try == 0) {
776                         bau_desc->header.msg_type = MSG_REGULAR;
777                         seq_number = bcp->message_number++;
778                 } else {
779                         bau_desc->header.msg_type = MSG_RETRY;
780                         stat->s_retry_messages++;
781                 }
782
783                 bau_desc->header.sequence = seq_number;
784                 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
785                 bcp->send_message = get_cycles();
786
787                 write_mmr_activation(index);
788
789                 try++;
790                 completion_stat = wait_completion(bau_desc, bcp, try);
791
792                 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
793
794                 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
795                         bcp->ipi_attempts = 0;
796                         completion_stat = FLUSH_GIVEUP;
797                         break;
798                 }
799                 cpu_relax();
800         } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
801                  (completion_stat == FLUSH_RETRY_TIMEOUT));
802
803         time2 = get_cycles();
804
805         count_max_concurr(completion_stat, bcp, hmaster);
806
807         while (hmaster->uvhub_quiesce)
808                 cpu_relax();
809
810         atomic_dec(&hmaster->active_descriptor_count);
811
812         record_send_stats(time1, time2, bcp, stat, completion_stat, try);
813
814         if (completion_stat == FLUSH_GIVEUP)
815                 return 1;
816         return 0;
817 }
818
819 /*
820  * The BAU is disabled. When the disabled time period has expired, the cpu
821  * that disabled it must re-enable it.
822  * Return 0 if it is re-enabled for all cpus.
823  */
824 static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
825 {
826         int tcpu;
827         struct bau_control *tbcp;
828
829         if (bcp->set_bau_off) {
830                 if (get_cycles() >= bcp->set_bau_on_time) {
831                         stat->s_bau_reenabled++;
832                         baudisabled = 0;
833                         for_each_present_cpu(tcpu) {
834                                 tbcp = &per_cpu(bau_control, tcpu);
835                                 tbcp->baudisabled = 0;
836                                 tbcp->period_requests = 0;
837                                 tbcp->period_time = 0;
838                         }
839                         return 0;
840                 }
841         }
842         return -1;
843 }
844
845 static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
846                                 int remotes, struct bau_desc *bau_desc)
847 {
848         stat->s_requestor++;
849         stat->s_ntargcpu += remotes + locals;
850         stat->s_ntargremotes += remotes;
851         stat->s_ntarglocals += locals;
852
853         /* uvhub statistics */
854         hubs = bau_uvhub_weight(&bau_desc->distribution);
855         if (locals) {
856                 stat->s_ntarglocaluvhub++;
857                 stat->s_ntargremoteuvhub += (hubs - 1);
858         } else
859                 stat->s_ntargremoteuvhub += hubs;
860
861         stat->s_ntarguvhub += hubs;
862
863         if (hubs >= 16)
864                 stat->s_ntarguvhub16++;
865         else if (hubs >= 8)
866                 stat->s_ntarguvhub8++;
867         else if (hubs >= 4)
868                 stat->s_ntarguvhub4++;
869         else if (hubs >= 2)
870                 stat->s_ntarguvhub2++;
871         else
872                 stat->s_ntarguvhub1++;
873 }
874
875 /*
876  * Translate a cpu mask to the uvhub distribution mask in the BAU
877  * activation descriptor.
878  */
879 static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
880                         struct bau_desc *bau_desc, int *localsp, int *remotesp)
881 {
882         int cpu;
883         int pnode;
884         int cnt = 0;
885         struct hub_and_pnode *hpp;
886
887         for_each_cpu(cpu, flush_mask) {
888                 /*
889                  * The distribution vector is a bit map of pnodes, relative
890                  * to the partition base pnode (and the partition base nasid
891                  * in the header).
892                  * Translate cpu to pnode and hub using a local memory array.
893                  */
894                 hpp = &bcp->socket_master->thp[cpu];
895                 pnode = hpp->pnode - bcp->partition_base_pnode;
896                 bau_uvhub_set(pnode, &bau_desc->distribution);
897                 cnt++;
898                 if (hpp->uvhub == bcp->uvhub)
899                         (*localsp)++;
900                 else
901                         (*remotesp)++;
902         }
903         if (!cnt)
904                 return 1;
905         return 0;
906 }
907
908 /*
909  * globally purge translation cache of a virtual address or all TLB's
910  * @cpumask: mask of all cpu's in which the address is to be removed
911  * @mm: mm_struct containing virtual address range
912  * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
913  * @cpu: the current cpu
914  *
915  * This is the entry point for initiating any UV global TLB shootdown.
916  *
917  * Purges the translation caches of all specified processors of the given
918  * virtual address, or purges all TLB's on specified processors.
919  *
920  * The caller has derived the cpumask from the mm_struct.  This function
921  * is called only if there are bits set in the mask. (e.g. flush_tlb_page())
922  *
923  * The cpumask is converted into a uvhubmask of the uvhubs containing
924  * those cpus.
925  *
926  * Note that this function should be called with preemption disabled.
927  *
928  * Returns NULL if all remote flushing was done.
929  * Returns pointer to cpumask if some remote flushing remains to be
930  * done.  The returned pointer is valid till preemption is re-enabled.
931  */
932 const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
933                                 struct mm_struct *mm, unsigned long va,
934                                 unsigned int cpu)
935 {
936         int locals = 0;
937         int remotes = 0;
938         int hubs = 0;
939         struct bau_desc *bau_desc;
940         struct cpumask *flush_mask;
941         struct ptc_stats *stat;
942         struct bau_control *bcp;
943
944         /* kernel was booted 'nobau' */
945         if (nobau)
946                 return cpumask;
947
948         bcp = &per_cpu(bau_control, cpu);
949         stat = bcp->statp;
950
951         /* bau was disabled due to slow response */
952         if (bcp->baudisabled) {
953                 if (check_enable(bcp, stat))
954                         return cpumask;
955         }
956
957         /*
958          * Each sending cpu has a per-cpu mask which it fills from the caller's
959          * cpu mask.  All cpus are converted to uvhubs and copied to the
960          * activation descriptor.
961          */
962         flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
963         /* don't actually do a shootdown of the local cpu */
964         cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
965
966         if (cpu_isset(cpu, *cpumask))
967                 stat->s_ntargself++;
968
969         bau_desc = bcp->descriptor_base;
970         bau_desc += ITEMS_PER_DESC * bcp->uvhub_cpu;
971         bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
972         if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
973                 return NULL;
974
975         record_send_statistics(stat, locals, hubs, remotes, bau_desc);
976
977         bau_desc->payload.address = va;
978         bau_desc->payload.sending_cpu = cpu;
979         /*
980          * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
981          * or 1 if it gave up and the original cpumask should be returned.
982          */
983         if (!uv_flush_send_and_wait(bau_desc, flush_mask, bcp))
984                 return NULL;
985         else
986                 return cpumask;
987 }
988
989 /*
990  * The BAU message interrupt comes here. (registered by set_intr_gate)
991  * See entry_64.S
992  *
993  * We received a broadcast assist message.
994  *
995  * Interrupts are disabled; this interrupt could represent
996  * the receipt of several messages.
997  *
998  * All cores/threads on this hub get this interrupt.
999  * The last one to see it does the software ack.
1000  * (the resource will not be freed until noninterruptable cpus see this
1001  *  interrupt; hardware may timeout the s/w ack and reply ERROR)
1002  */
1003 void uv_bau_message_interrupt(struct pt_regs *regs)
1004 {
1005         int count = 0;
1006         cycles_t time_start;
1007         struct bau_pq_entry *msg;
1008         struct bau_control *bcp;
1009         struct ptc_stats *stat;
1010         struct msg_desc msgdesc;
1011
1012         time_start = get_cycles();
1013
1014         bcp = &per_cpu(bau_control, smp_processor_id());
1015         stat = bcp->statp;
1016
1017         msgdesc.queue_first = bcp->queue_first;
1018         msgdesc.queue_last = bcp->queue_last;
1019
1020         msg = bcp->bau_msg_head;
1021         while (msg->swack_vec) {
1022                 count++;
1023
1024                 msgdesc.msg_slot = msg - msgdesc.queue_first;
1025                 msgdesc.swack_slot = ffs(msg->swack_vec) - 1;
1026                 msgdesc.msg = msg;
1027                 bau_process_message(&msgdesc, bcp);
1028
1029                 msg++;
1030                 if (msg > msgdesc.queue_last)
1031                         msg = msgdesc.queue_first;
1032                 bcp->bau_msg_head = msg;
1033         }
1034         stat->d_time += (get_cycles() - time_start);
1035         if (!count)
1036                 stat->d_nomsg++;
1037         else if (count > 1)
1038                 stat->d_multmsg++;
1039
1040         ack_APIC_irq();
1041 }
1042
1043 /*
1044  * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
1045  * shootdown message timeouts enabled.  The timeout does not cause
1046  * an interrupt, but causes an error message to be returned to
1047  * the sender.
1048  */
1049 static void __init enable_timeouts(void)
1050 {
1051         int uvhub;
1052         int nuvhubs;
1053         int pnode;
1054         unsigned long mmr_image;
1055
1056         nuvhubs = uv_num_possible_blades();
1057
1058         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1059                 if (!uv_blade_nr_possible_cpus(uvhub))
1060                         continue;
1061
1062                 pnode = uv_blade_to_pnode(uvhub);
1063                 mmr_image = read_mmr_misc_control(pnode);
1064                 /*
1065                  * Set the timeout period and then lock it in, in three
1066                  * steps; captures and locks in the period.
1067                  *
1068                  * To program the period, the SOFT_ACK_MODE must be off.
1069                  */
1070                 mmr_image &= ~(1L << SOFTACK_MSHIFT);
1071                 write_mmr_misc_control(pnode, mmr_image);
1072                 /*
1073                  * Set the 4-bit period.
1074                  */
1075                 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
1076                 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
1077                 write_mmr_misc_control(pnode, mmr_image);
1078                 /*
1079                  * UV1:
1080                  * Subsequent reversals of the timebase bit (3) cause an
1081                  * immediate timeout of one or all INTD resources as
1082                  * indicated in bits 2:0 (7 causes all of them to timeout).
1083                  */
1084                 mmr_image |= (1L << SOFTACK_MSHIFT);
1085                 if (is_uv2_hub()) {
1086                         mmr_image |= (1L << UV2_LEG_SHFT);
1087                         mmr_image |= (1L << UV2_EXT_SHFT);
1088                 }
1089                 write_mmr_misc_control(pnode, mmr_image);
1090         }
1091 }
1092
1093 static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
1094 {
1095         if (*offset < num_possible_cpus())
1096                 return offset;
1097         return NULL;
1098 }
1099
1100 static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
1101 {
1102         (*offset)++;
1103         if (*offset < num_possible_cpus())
1104                 return offset;
1105         return NULL;
1106 }
1107
1108 static void ptc_seq_stop(struct seq_file *file, void *data)
1109 {
1110 }
1111
1112 static inline unsigned long long usec_2_cycles(unsigned long microsec)
1113 {
1114         unsigned long ns;
1115         unsigned long long cyc;
1116
1117         ns = microsec * 1000;
1118         cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
1119         return cyc;
1120 }
1121
1122 /*
1123  * Display the statistics thru /proc/sgi_uv/ptc_statistics
1124  * 'data' points to the cpu number
1125  * Note: see the descriptions in stat_description[].
1126  */
1127 static int ptc_seq_show(struct seq_file *file, void *data)
1128 {
1129         struct ptc_stats *stat;
1130         int cpu;
1131
1132         cpu = *(loff_t *)data;
1133         if (!cpu) {
1134                 seq_printf(file,
1135                         "# cpu sent stime self locals remotes ncpus localhub ");
1136                 seq_printf(file,
1137                         "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
1138                 seq_printf(file,
1139                         "numuvhubs4 numuvhubs2 numuvhubs1 dto retries rok ");
1140                 seq_printf(file,
1141                         "resetp resett giveup sto bz throt swack recv rtime ");
1142                 seq_printf(file,
1143                         "all one mult none retry canc nocan reset rcan ");
1144                 seq_printf(file,
1145                         "disable enable\n");
1146         }
1147         if (cpu < num_possible_cpus() && cpu_online(cpu)) {
1148                 stat = &per_cpu(ptcstats, cpu);
1149                 /* source side statistics */
1150                 seq_printf(file,
1151                         "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1152                            cpu, stat->s_requestor, cycles_2_us(stat->s_time),
1153                            stat->s_ntargself, stat->s_ntarglocals,
1154                            stat->s_ntargremotes, stat->s_ntargcpu,
1155                            stat->s_ntarglocaluvhub, stat->s_ntargremoteuvhub,
1156                            stat->s_ntarguvhub, stat->s_ntarguvhub16);
1157                 seq_printf(file, "%ld %ld %ld %ld %ld ",
1158                            stat->s_ntarguvhub8, stat->s_ntarguvhub4,
1159                            stat->s_ntarguvhub2, stat->s_ntarguvhub1,
1160                            stat->s_dtimeout);
1161                 seq_printf(file, "%ld %ld %ld %ld %ld %ld %ld %ld ",
1162                            stat->s_retry_messages, stat->s_retriesok,
1163                            stat->s_resets_plug, stat->s_resets_timeout,
1164                            stat->s_giveup, stat->s_stimeout,
1165                            stat->s_busy, stat->s_throttles);
1166
1167                 /* destination side statistics */
1168                 seq_printf(file,
1169                            "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
1170                            read_gmmr_sw_ack(uv_cpu_to_pnode(cpu)),
1171                            stat->d_requestee, cycles_2_us(stat->d_time),
1172                            stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
1173                            stat->d_nomsg, stat->d_retries, stat->d_canceled,
1174                            stat->d_nocanceled, stat->d_resets,
1175                            stat->d_rcanceled);
1176                 seq_printf(file, "%ld %ld\n",
1177                         stat->s_bau_disabled, stat->s_bau_reenabled);
1178         }
1179         return 0;
1180 }
1181
1182 /*
1183  * Display the tunables thru debugfs
1184  */
1185 static ssize_t tunables_read(struct file *file, char __user *userbuf,
1186                                 size_t count, loff_t *ppos)
1187 {
1188         char *buf;
1189         int ret;
1190
1191         buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1192                 "max_concur plugged_delay plugsb4reset",
1193                 "timeoutsb4reset ipi_reset_limit complete_threshold",
1194                 "congested_response_us congested_reps congested_period",
1195                 max_concurr, plugged_delay, plugsb4reset,
1196                 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1197                 congested_respns_us, congested_reps, congested_period);
1198
1199         if (!buf)
1200                 return -ENOMEM;
1201
1202         ret = simple_read_from_buffer(userbuf, count, ppos, buf, strlen(buf));
1203         kfree(buf);
1204         return ret;
1205 }
1206
1207 /*
1208  * handle a write to /proc/sgi_uv/ptc_statistics
1209  * -1: reset the statistics
1210  *  0: display meaning of the statistics
1211  */
1212 static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1213                                 size_t count, loff_t *data)
1214 {
1215         int cpu;
1216         int i;
1217         int elements;
1218         long input_arg;
1219         char optstr[64];
1220         struct ptc_stats *stat;
1221
1222         if (count == 0 || count > sizeof(optstr))
1223                 return -EINVAL;
1224         if (copy_from_user(optstr, user, count))
1225                 return -EFAULT;
1226         optstr[count - 1] = '\0';
1227
1228         if (strict_strtol(optstr, 10, &input_arg) < 0) {
1229                 printk(KERN_DEBUG "%s is invalid\n", optstr);
1230                 return -EINVAL;
1231         }
1232
1233         if (input_arg == 0) {
1234                 elements = sizeof(stat_description)/sizeof(*stat_description);
1235                 printk(KERN_DEBUG "# cpu:      cpu number\n");
1236                 printk(KERN_DEBUG "Sender statistics:\n");
1237                 for (i = 0; i < elements; i++)
1238                         printk(KERN_DEBUG "%s\n", stat_description[i]);
1239         } else if (input_arg == -1) {
1240                 for_each_present_cpu(cpu) {
1241                         stat = &per_cpu(ptcstats, cpu);
1242                         memset(stat, 0, sizeof(struct ptc_stats));
1243                 }
1244         }
1245
1246         return count;
1247 }
1248
1249 static int local_atoi(const char *name)
1250 {
1251         int val = 0;
1252
1253         for (;; name++) {
1254                 switch (*name) {
1255                 case '0' ... '9':
1256                         val = 10*val+(*name-'0');
1257                         break;
1258                 default:
1259                         return val;
1260                 }
1261         }
1262 }
1263
1264 /*
1265  * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1266  * Zero values reset them to defaults.
1267  */
1268 static int parse_tunables_write(struct bau_control *bcp, char *instr,
1269                                 int count)
1270 {
1271         char *p;
1272         char *q;
1273         int cnt = 0;
1274         int val;
1275         int e = sizeof(tunables) / sizeof(*tunables);
1276
1277         p = instr + strspn(instr, WHITESPACE);
1278         q = p;
1279         for (; *p; p = q + strspn(q, WHITESPACE)) {
1280                 q = p + strcspn(p, WHITESPACE);
1281                 cnt++;
1282                 if (q == p)
1283                         break;
1284         }
1285         if (cnt != e) {
1286                 printk(KERN_INFO "bau tunable error: should be %d values\n", e);
1287                 return -EINVAL;
1288         }
1289
1290         p = instr + strspn(instr, WHITESPACE);
1291         q = p;
1292         for (cnt = 0; *p; p = q + strspn(q, WHITESPACE), cnt++) {
1293                 q = p + strcspn(p, WHITESPACE);
1294                 val = local_atoi(p);
1295                 switch (cnt) {
1296                 case 0:
1297                         if (val == 0) {
1298                                 max_concurr = MAX_BAU_CONCURRENT;
1299                                 max_concurr_const = MAX_BAU_CONCURRENT;
1300                                 continue;
1301                         }
1302                         if (val < 1 || val > bcp->cpus_in_uvhub) {
1303                                 printk(KERN_DEBUG
1304                                 "Error: BAU max concurrent %d is invalid\n",
1305                                 val);
1306                                 return -EINVAL;
1307                         }
1308                         max_concurr = val;
1309                         max_concurr_const = val;
1310                         continue;
1311                 default:
1312                         if (val == 0)
1313                                 *tunables[cnt].tunp = tunables[cnt].deflt;
1314                         else
1315                                 *tunables[cnt].tunp = val;
1316                         continue;
1317                 }
1318                 if (q == p)
1319                         break;
1320         }
1321         return 0;
1322 }
1323
1324 /*
1325  * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1326  */
1327 static ssize_t tunables_write(struct file *file, const char __user *user,
1328                                 size_t count, loff_t *data)
1329 {
1330         int cpu;
1331         int ret;
1332         char instr[100];
1333         struct bau_control *bcp;
1334
1335         if (count == 0 || count > sizeof(instr)-1)
1336                 return -EINVAL;
1337         if (copy_from_user(instr, user, count))
1338                 return -EFAULT;
1339
1340         instr[count] = '\0';
1341
1342         cpu = get_cpu();
1343         bcp = &per_cpu(bau_control, cpu);
1344         ret = parse_tunables_write(bcp, instr, count);
1345         put_cpu();
1346         if (ret)
1347                 return ret;
1348
1349         for_each_present_cpu(cpu) {
1350                 bcp = &per_cpu(bau_control, cpu);
1351                 bcp->max_concurr =              max_concurr;
1352                 bcp->max_concurr_const =        max_concurr;
1353                 bcp->plugged_delay =            plugged_delay;
1354                 bcp->plugsb4reset =             plugsb4reset;
1355                 bcp->timeoutsb4reset =          timeoutsb4reset;
1356                 bcp->ipi_reset_limit =          ipi_reset_limit;
1357                 bcp->complete_threshold =       complete_threshold;
1358                 bcp->cong_response_us =         congested_respns_us;
1359                 bcp->cong_reps =                congested_reps;
1360                 bcp->cong_period =              congested_period;
1361         }
1362         return count;
1363 }
1364
1365 static const struct seq_operations uv_ptc_seq_ops = {
1366         .start          = ptc_seq_start,
1367         .next           = ptc_seq_next,
1368         .stop           = ptc_seq_stop,
1369         .show           = ptc_seq_show
1370 };
1371
1372 static int ptc_proc_open(struct inode *inode, struct file *file)
1373 {
1374         return seq_open(file, &uv_ptc_seq_ops);
1375 }
1376
1377 static int tunables_open(struct inode *inode, struct file *file)
1378 {
1379         return 0;
1380 }
1381
1382 static const struct file_operations proc_uv_ptc_operations = {
1383         .open           = ptc_proc_open,
1384         .read           = seq_read,
1385         .write          = ptc_proc_write,
1386         .llseek         = seq_lseek,
1387         .release        = seq_release,
1388 };
1389
1390 static const struct file_operations tunables_fops = {
1391         .open           = tunables_open,
1392         .read           = tunables_read,
1393         .write          = tunables_write,
1394         .llseek         = default_llseek,
1395 };
1396
1397 static int __init uv_ptc_init(void)
1398 {
1399         struct proc_dir_entry *proc_uv_ptc;
1400
1401         if (!is_uv_system())
1402                 return 0;
1403
1404         proc_uv_ptc = proc_create(UV_PTC_BASENAME, 0444, NULL,
1405                                   &proc_uv_ptc_operations);
1406         if (!proc_uv_ptc) {
1407                 printk(KERN_ERR "unable to create %s proc entry\n",
1408                        UV_PTC_BASENAME);
1409                 return -EINVAL;
1410         }
1411
1412         tunables_dir = debugfs_create_dir(UV_BAU_TUNABLES_DIR, NULL);
1413         if (!tunables_dir) {
1414                 printk(KERN_ERR "unable to create debugfs directory %s\n",
1415                        UV_BAU_TUNABLES_DIR);
1416                 return -EINVAL;
1417         }
1418         tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1419                                         tunables_dir, NULL, &tunables_fops);
1420         if (!tunables_file) {
1421                 printk(KERN_ERR "unable to create debugfs file %s\n",
1422                        UV_BAU_TUNABLES_FILE);
1423                 return -EINVAL;
1424         }
1425         return 0;
1426 }
1427
1428 /*
1429  * Initialize the sending side's sending buffers.
1430  */
1431 static void activation_descriptor_init(int node, int pnode, int base_pnode)
1432 {
1433         int i;
1434         int cpu;
1435         unsigned long gpa;
1436         unsigned long m;
1437         unsigned long n;
1438         size_t dsize;
1439         struct bau_desc *bau_desc;
1440         struct bau_desc *bd2;
1441         struct bau_control *bcp;
1442
1443         /*
1444          * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1445          * per cpu; and one per cpu on the uvhub (ADP_SZ)
1446          */
1447         dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1448         bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1449         BUG_ON(!bau_desc);
1450
1451         gpa = uv_gpa(bau_desc);
1452         n = uv_gpa_to_gnode(gpa);
1453         m = uv_gpa_to_offset(gpa);
1454
1455         /* the 14-bit pnode */
1456         write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1457         /*
1458          * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1459          * cpu even though we only use the first one; one descriptor can
1460          * describe a broadcast to 256 uv hubs.
1461          */
1462         for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1463                 memset(bd2, 0, sizeof(struct bau_desc));
1464                 bd2->header.swack_flag =        1;
1465                 /*
1466                  * The base_dest_nasid set in the message header is the nasid
1467                  * of the first uvhub in the partition. The bit map will
1468                  * indicate destination pnode numbers relative to that base.
1469                  * They may not be consecutive if nasid striding is being used.
1470                  */
1471                 bd2->header.base_dest_nasid =   UV_PNODE_TO_NASID(base_pnode);
1472                 bd2->header.dest_subnodeid =    UV_LB_SUBNODEID;
1473                 bd2->header.command =           UV_NET_ENDPOINT_INTD;
1474                 bd2->header.int_both =          1;
1475                 /*
1476                  * all others need to be set to zero:
1477                  *   fairness chaining multilevel count replied_to
1478                  */
1479         }
1480         for_each_present_cpu(cpu) {
1481                 if (pnode != uv_blade_to_pnode(uv_cpu_to_blade_id(cpu)))
1482                         continue;
1483                 bcp = &per_cpu(bau_control, cpu);
1484                 bcp->descriptor_base = bau_desc;
1485         }
1486 }
1487
1488 /*
1489  * initialize the destination side's receiving buffers
1490  * entered for each uvhub in the partition
1491  * - node is first node (kernel memory notion) on the uvhub
1492  * - pnode is the uvhub's physical identifier
1493  */
1494 static void pq_init(int node, int pnode)
1495 {
1496         int cpu;
1497         size_t plsize;
1498         char *cp;
1499         void *vp;
1500         unsigned long pn;
1501         unsigned long first;
1502         unsigned long pn_first;
1503         unsigned long last;
1504         struct bau_pq_entry *pqp;
1505         struct bau_control *bcp;
1506
1507         plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1508         vp = kmalloc_node(plsize, GFP_KERNEL, node);
1509         pqp = (struct bau_pq_entry *)vp;
1510         BUG_ON(!pqp);
1511
1512         cp = (char *)pqp + 31;
1513         pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1514
1515         for_each_present_cpu(cpu) {
1516                 if (pnode != uv_cpu_to_pnode(cpu))
1517                         continue;
1518                 /* for every cpu on this pnode: */
1519                 bcp = &per_cpu(bau_control, cpu);
1520                 bcp->queue_first        = pqp;
1521                 bcp->bau_msg_head       = pqp;
1522                 bcp->queue_last         = pqp + (DEST_Q_SIZE - 1);
1523         }
1524         /*
1525          * need the gnode of where the memory was really allocated
1526          */
1527         pn = uv_gpa_to_gnode(uv_gpa(pqp));
1528         first = uv_physnodeaddr(pqp);
1529         pn_first = ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | first;
1530         last = uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1));
1531         write_mmr_payload_first(pnode, pn_first);
1532         write_mmr_payload_tail(pnode, first);
1533         write_mmr_payload_last(pnode, last);
1534
1535         /* in effect, all msg_type's are set to MSG_NOOP */
1536         memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1537 }
1538
1539 /*
1540  * Initialization of each UV hub's structures
1541  */
1542 static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1543 {
1544         int node;
1545         int pnode;
1546         unsigned long apicid;
1547
1548         node = uvhub_to_first_node(uvhub);
1549         pnode = uv_blade_to_pnode(uvhub);
1550
1551         activation_descriptor_init(node, pnode, base_pnode);
1552
1553         pq_init(node, pnode);
1554         /*
1555          * The below initialization can't be in firmware because the
1556          * messaging IRQ will be determined by the OS.
1557          */
1558         apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1559         write_mmr_data_config(pnode, ((apicid << 32) | vector));
1560 }
1561
1562 /*
1563  * We will set BAU_MISC_CONTROL with a timeout period.
1564  * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1565  * So the destination timeout period has to be calculated from them.
1566  */
1567 static int calculate_destination_timeout(void)
1568 {
1569         unsigned long mmr_image;
1570         int mult1;
1571         int mult2;
1572         int index;
1573         int base;
1574         int ret;
1575         unsigned long ts_ns;
1576
1577         if (is_uv1_hub()) {
1578                 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1579                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1580                 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1581                 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1582                 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1583                 base = timeout_base_ns[index];
1584                 ts_ns = base * mult1 * mult2;
1585                 ret = ts_ns / 1000;
1586         } else {
1587                 /* 4 bits  0/1 for 10/80us, 3 bits of multiplier */
1588                 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1589                 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1590                 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1591                         mult1 = 80;
1592                 else
1593                         mult1 = 10;
1594                 base = mmr_image & UV2_ACK_MASK;
1595                 ret = mult1 * base;
1596         }
1597         return ret;
1598 }
1599
1600 static void __init init_per_cpu_tunables(void)
1601 {
1602         int cpu;
1603         struct bau_control *bcp;
1604
1605         for_each_present_cpu(cpu) {
1606                 bcp = &per_cpu(bau_control, cpu);
1607                 bcp->baudisabled                = 0;
1608                 bcp->statp                      = &per_cpu(ptcstats, cpu);
1609                 /* time interval to catch a hardware stay-busy bug */
1610                 bcp->timeout_interval           = usec_2_cycles(2*timeout_us);
1611                 bcp->max_concurr                = max_concurr;
1612                 bcp->max_concurr_const          = max_concurr;
1613                 bcp->plugged_delay              = plugged_delay;
1614                 bcp->plugsb4reset               = plugsb4reset;
1615                 bcp->timeoutsb4reset            = timeoutsb4reset;
1616                 bcp->ipi_reset_limit            = ipi_reset_limit;
1617                 bcp->complete_threshold         = complete_threshold;
1618                 bcp->cong_response_us           = congested_respns_us;
1619                 bcp->cong_reps                  = congested_reps;
1620                 bcp->cong_period                = congested_period;
1621         }
1622 }
1623
1624 /*
1625  * Scan all cpus to collect blade and socket summaries.
1626  */
1627 static int __init get_cpu_topology(int base_pnode,
1628                                         struct uvhub_desc *uvhub_descs,
1629                                         unsigned char *uvhub_mask)
1630 {
1631         int cpu;
1632         int pnode;
1633         int uvhub;
1634         int socket;
1635         struct bau_control *bcp;
1636         struct uvhub_desc *bdp;
1637         struct socket_desc *sdp;
1638
1639         for_each_present_cpu(cpu) {
1640                 bcp = &per_cpu(bau_control, cpu);
1641
1642                 memset(bcp, 0, sizeof(struct bau_control));
1643
1644                 pnode = uv_cpu_hub_info(cpu)->pnode;
1645                 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1646                         printk(KERN_EMERG
1647                                 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1648                                 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1649                         return 1;
1650                 }
1651
1652                 bcp->osnode = cpu_to_node(cpu);
1653                 bcp->partition_base_pnode = base_pnode;
1654
1655                 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1656                 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1657                 bdp = &uvhub_descs[uvhub];
1658
1659                 bdp->num_cpus++;
1660                 bdp->uvhub = uvhub;
1661                 bdp->pnode = pnode;
1662
1663                 /* kludge: 'assuming' one node per socket, and assuming that
1664                    disabling a socket just leaves a gap in node numbers */
1665                 socket = bcp->osnode & 1;
1666                 bdp->socket_mask |= (1 << socket);
1667                 sdp = &bdp->socket[socket];
1668                 sdp->cpu_number[sdp->num_cpus] = cpu;
1669                 sdp->num_cpus++;
1670                 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1671                         printk(KERN_EMERG "%d cpus per socket invalid\n",
1672                                 sdp->num_cpus);
1673                         return 1;
1674                 }
1675         }
1676         return 0;
1677 }
1678
1679 /*
1680  * Each socket is to get a local array of pnodes/hubs.
1681  */
1682 static void make_per_cpu_thp(struct bau_control *smaster)
1683 {
1684         int cpu;
1685         size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
1686
1687         smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
1688         memset(smaster->thp, 0, hpsz);
1689         for_each_present_cpu(cpu) {
1690                 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
1691                 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1692         }
1693 }
1694
1695 /*
1696  * Each uvhub is to get a local cpumask.
1697  */
1698 static void make_per_hub_cpumask(struct bau_control *hmaster)
1699 {
1700         int sz = sizeof(cpumask_t);
1701
1702         hmaster->cpumask = kzalloc_node(sz, GFP_KERNEL, hmaster->osnode);
1703 }
1704
1705 /*
1706  * Initialize all the per_cpu information for the cpu's on a given socket,
1707  * given what has been gathered into the socket_desc struct.
1708  * And reports the chosen hub and socket masters back to the caller.
1709  */
1710 static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
1711                         struct bau_control **smasterp,
1712                         struct bau_control **hmasterp)
1713 {
1714         int i;
1715         int cpu;
1716         struct bau_control *bcp;
1717
1718         for (i = 0; i < sdp->num_cpus; i++) {
1719                 cpu = sdp->cpu_number[i];
1720                 bcp = &per_cpu(bau_control, cpu);
1721                 bcp->cpu = cpu;
1722                 if (i == 0) {
1723                         *smasterp = bcp;
1724                         if (!(*hmasterp))
1725                                 *hmasterp = bcp;
1726                 }
1727                 bcp->cpus_in_uvhub = bdp->num_cpus;
1728                 bcp->cpus_in_socket = sdp->num_cpus;
1729                 bcp->socket_master = *smasterp;
1730                 bcp->uvhub = bdp->uvhub;
1731                 bcp->uvhub_master = *hmasterp;
1732                 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->blade_processor_id;
1733                 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
1734                         printk(KERN_EMERG "%d cpus per uvhub invalid\n",
1735                                 bcp->uvhub_cpu);
1736                         return 1;
1737                 }
1738         }
1739         return 0;
1740 }
1741
1742 /*
1743  * Summarize the blade and socket topology into the per_cpu structures.
1744  */
1745 static int __init summarize_uvhub_sockets(int nuvhubs,
1746                         struct uvhub_desc *uvhub_descs,
1747                         unsigned char *uvhub_mask)
1748 {
1749         int socket;
1750         int uvhub;
1751         unsigned short socket_mask;
1752
1753         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1754                 struct uvhub_desc *bdp;
1755                 struct bau_control *smaster = NULL;
1756                 struct bau_control *hmaster = NULL;
1757
1758                 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
1759                         continue;
1760
1761                 bdp = &uvhub_descs[uvhub];
1762                 socket_mask = bdp->socket_mask;
1763                 socket = 0;
1764                 while (socket_mask) {
1765                         struct socket_desc *sdp;
1766                         if ((socket_mask & 1)) {
1767                                 sdp = &bdp->socket[socket];
1768                                 if (scan_sock(sdp, bdp, &smaster, &hmaster))
1769                                         return 1;
1770                                 make_per_cpu_thp(smaster);
1771                         }
1772                         socket++;
1773                         socket_mask = (socket_mask >> 1);
1774                 }
1775                 make_per_hub_cpumask(hmaster);
1776         }
1777         return 0;
1778 }
1779
1780 /*
1781  * initialize the bau_control structure for each cpu
1782  */
1783 static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
1784 {
1785         unsigned char *uvhub_mask;
1786         void *vp;
1787         struct uvhub_desc *uvhub_descs;
1788
1789         timeout_us = calculate_destination_timeout();
1790
1791         vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1792         uvhub_descs = (struct uvhub_desc *)vp;
1793         memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1794         uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
1795
1796         if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
1797                 goto fail;
1798
1799         if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
1800                 goto fail;
1801
1802         kfree(uvhub_descs);
1803         kfree(uvhub_mask);
1804         init_per_cpu_tunables();
1805         return 0;
1806
1807 fail:
1808         kfree(uvhub_descs);
1809         kfree(uvhub_mask);
1810         return 1;
1811 }
1812
1813 /*
1814  * Initialization of BAU-related structures
1815  */
1816 static int __init uv_bau_init(void)
1817 {
1818         int uvhub;
1819         int pnode;
1820         int nuvhubs;
1821         int cur_cpu;
1822         int cpus;
1823         int vector;
1824         cpumask_var_t *mask;
1825
1826         if (!is_uv_system())
1827                 return 0;
1828
1829         if (nobau)
1830                 return 0;
1831
1832         for_each_possible_cpu(cur_cpu) {
1833                 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
1834                 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
1835         }
1836
1837         nuvhubs = uv_num_possible_blades();
1838         spin_lock_init(&disable_lock);
1839         congested_cycles = usec_2_cycles(congested_respns_us);
1840
1841         uv_base_pnode = 0x7fffffff;
1842         for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1843                 cpus = uv_blade_nr_possible_cpus(uvhub);
1844                 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
1845                         uv_base_pnode = uv_blade_to_pnode(uvhub);
1846         }
1847
1848         if (init_per_cpu(nuvhubs, uv_base_pnode)) {
1849                 nobau = 1;
1850                 return 0;
1851         }
1852
1853         vector = UV_BAU_MESSAGE;
1854         for_each_possible_blade(uvhub)
1855                 if (uv_blade_nr_possible_cpus(uvhub))
1856                         init_uvhub(uvhub, vector, uv_base_pnode);
1857
1858         enable_timeouts();
1859         alloc_intr_gate(vector, uv_bau_message_intr1);
1860
1861         for_each_possible_blade(uvhub) {
1862                 if (uv_blade_nr_possible_cpus(uvhub)) {
1863                         unsigned long val;
1864                         unsigned long mmr;
1865                         pnode = uv_blade_to_pnode(uvhub);
1866                         /* INIT the bau */
1867                         val = 1L << 63;
1868                         write_gmmr_activation(pnode, val);
1869                         mmr = 1; /* should be 1 to broadcast to both sockets */
1870                         write_mmr_data_broadcast(pnode, mmr);
1871                 }
1872         }
1873
1874         return 0;
1875 }
1876 core_initcall(uv_bau_init);
1877 fs_initcall(uv_ptc_init);