aboutsummaryrefslogtreecommitdiffstats
path: root/arch/x86/platform/uv/tlb_uv.c
diff options
context:
space:
mode:
Diffstat (limited to 'arch/x86/platform/uv/tlb_uv.c')
-rw-r--r--arch/x86/platform/uv/tlb_uv.c1484
1 files changed, 816 insertions, 668 deletions
diff --git a/arch/x86/platform/uv/tlb_uv.c b/arch/x86/platform/uv/tlb_uv.c
index c58e0ea39ef..68e467f69fe 100644
--- a/arch/x86/platform/uv/tlb_uv.c
+++ b/arch/x86/platform/uv/tlb_uv.c
@@ -1,7 +1,7 @@
1/* 1/*
2 * SGI UltraViolet TLB flush routines. 2 * SGI UltraViolet TLB flush routines.
3 * 3 *
4 * (c) 2008-2010 Cliff Wickman <cpw@sgi.com>, SGI. 4 * (c) 2008-2011 Cliff Wickman <cpw@sgi.com>, SGI.
5 * 5 *
6 * This code is released under the GNU General Public License version 2 or 6 * This code is released under the GNU General Public License version 2 or
7 * later. 7 * later.
@@ -35,6 +35,7 @@ static int timeout_base_ns[] = {
35 5242880, 35 5242880,
36 167772160 36 167772160
37}; 37};
38
38static int timeout_us; 39static int timeout_us;
39static int nobau; 40static int nobau;
40static int baudisabled; 41static int baudisabled;
@@ -42,20 +43,70 @@ static spinlock_t disable_lock;
42static cycles_t congested_cycles; 43static cycles_t congested_cycles;
43 44
44/* tunables: */ 45/* tunables: */
45static int max_bau_concurrent = MAX_BAU_CONCURRENT; 46static int max_concurr = MAX_BAU_CONCURRENT;
46static int max_bau_concurrent_constant = MAX_BAU_CONCURRENT; 47static int max_concurr_const = MAX_BAU_CONCURRENT;
47static int plugged_delay = PLUGGED_DELAY; 48static int plugged_delay = PLUGGED_DELAY;
48static int plugsb4reset = PLUGSB4RESET; 49static int plugsb4reset = PLUGSB4RESET;
49static int timeoutsb4reset = TIMEOUTSB4RESET; 50static int timeoutsb4reset = TIMEOUTSB4RESET;
50static int ipi_reset_limit = IPI_RESET_LIMIT; 51static int ipi_reset_limit = IPI_RESET_LIMIT;
51static int complete_threshold = COMPLETE_THRESHOLD; 52static int complete_threshold = COMPLETE_THRESHOLD;
52static int congested_response_us = CONGESTED_RESPONSE_US; 53static int congested_respns_us = CONGESTED_RESPONSE_US;
53static int congested_reps = CONGESTED_REPS; 54static int congested_reps = CONGESTED_REPS;
54static int congested_period = CONGESTED_PERIOD; 55static int congested_period = CONGESTED_PERIOD;
56
57static 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
55static struct dentry *tunables_dir; 69static struct dentry *tunables_dir;
56static struct dentry *tunables_file; 70static struct dentry *tunables_file;
57 71
58static int __init setup_nobau(char *arg) 72/* these correspond to the statistics printed by ptc_seq_show() */
73static 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
108static int __init
109setup_nobau(char *arg)
59{ 110{
60 nobau = 1; 111 nobau = 1;
61 return 0; 112 return 0;
@@ -63,7 +114,7 @@ static int __init setup_nobau(char *arg)
63early_param("nobau", setup_nobau); 114early_param("nobau", setup_nobau);
64 115
65/* base pnode in this partition */ 116/* base pnode in this partition */
66static int uv_partition_base_pnode __read_mostly; 117static int uv_base_pnode __read_mostly;
67/* position of pnode (which is nasid>>1): */ 118/* position of pnode (which is nasid>>1): */
68static int uv_nshift __read_mostly; 119static int uv_nshift __read_mostly;
69static unsigned long uv_mmask __read_mostly; 120static unsigned long uv_mmask __read_mostly;
@@ -109,60 +160,52 @@ static int __init uvhub_to_first_apicid(int uvhub)
109 * clear of the Timeout bit (as well) will free the resource. No reply will 160 * clear of the Timeout bit (as well) will free the resource. No reply will
110 * be sent (the hardware will only do one reply per message). 161 * be sent (the hardware will only do one reply per message).
111 */ 162 */
112static inline void uv_reply_to_message(struct msg_desc *mdp, 163static void reply_to_message(struct msg_desc *mdp, struct bau_control *bcp)
113 struct bau_control *bcp)
114{ 164{
115 unsigned long dw; 165 unsigned long dw;
116 struct bau_payload_queue_entry *msg; 166 struct bau_pq_entry *msg;
117 167
118 msg = mdp->msg; 168 msg = mdp->msg;
119 if (!msg->canceled) { 169 if (!msg->canceled) {
120 dw = (msg->sw_ack_vector << UV_SW_ACK_NPENDING) | 170 dw = (msg->swack_vec << UV_SW_ACK_NPENDING) | msg->swack_vec;
121 msg->sw_ack_vector; 171 write_mmr_sw_ack(dw);
122 uv_write_local_mmr(
123 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw);
124 } 172 }
125 msg->replied_to = 1; 173 msg->replied_to = 1;
126 msg->sw_ack_vector = 0; 174 msg->swack_vec = 0;
127} 175}
128 176
129/* 177/*
130 * Process the receipt of a RETRY message 178 * Process the receipt of a RETRY message
131 */ 179 */
132static inline void uv_bau_process_retry_msg(struct msg_desc *mdp, 180static void bau_process_retry_msg(struct msg_desc *mdp,
133 struct bau_control *bcp) 181 struct bau_control *bcp)
134{ 182{
135 int i; 183 int i;
136 int cancel_count = 0; 184 int cancel_count = 0;
137 int slot2;
138 unsigned long msg_res; 185 unsigned long msg_res;
139 unsigned long mmr = 0; 186 unsigned long mmr = 0;
140 struct bau_payload_queue_entry *msg; 187 struct bau_pq_entry *msg = mdp->msg;
141 struct bau_payload_queue_entry *msg2; 188 struct bau_pq_entry *msg2;
142 struct ptc_stats *stat; 189 struct ptc_stats *stat = bcp->statp;
143 190
144 msg = mdp->msg;
145 stat = bcp->statp;
146 stat->d_retries++; 191 stat->d_retries++;
147 /* 192 /*
148 * cancel any message from msg+1 to the retry itself 193 * cancel any message from msg+1 to the retry itself
149 */ 194 */
150 for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) { 195 for (msg2 = msg+1, i = 0; i < DEST_Q_SIZE; msg2++, i++) {
151 if (msg2 > mdp->va_queue_last) 196 if (msg2 > mdp->queue_last)
152 msg2 = mdp->va_queue_first; 197 msg2 = mdp->queue_first;
153 if (msg2 == msg) 198 if (msg2 == msg)
154 break; 199 break;
155 200
156 /* same conditions for cancellation as uv_do_reset */ 201 /* same conditions for cancellation as do_reset */
157 if ((msg2->replied_to == 0) && (msg2->canceled == 0) && 202 if ((msg2->replied_to == 0) && (msg2->canceled == 0) &&
158 (msg2->sw_ack_vector) && ((msg2->sw_ack_vector & 203 (msg2->swack_vec) && ((msg2->swack_vec &
159 msg->sw_ack_vector) == 0) && 204 msg->swack_vec) == 0) &&
160 (msg2->sending_cpu == msg->sending_cpu) && 205 (msg2->sending_cpu == msg->sending_cpu) &&
161 (msg2->msg_type != MSG_NOOP)) { 206 (msg2->msg_type != MSG_NOOP)) {
162 slot2 = msg2 - mdp->va_queue_first; 207 mmr = read_mmr_sw_ack();
163 mmr = uv_read_local_mmr 208 msg_res = msg2->swack_vec;
164 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE);
165 msg_res = msg2->sw_ack_vector;
166 /* 209 /*
167 * This is a message retry; clear the resources held 210 * This is a message retry; clear the resources held
168 * by the previous message only if they timed out. 211 * by the previous message only if they timed out.
@@ -170,6 +213,7 @@ static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
170 * situation to report. 213 * situation to report.
171 */ 214 */
172 if (mmr & (msg_res << UV_SW_ACK_NPENDING)) { 215 if (mmr & (msg_res << UV_SW_ACK_NPENDING)) {
216 unsigned long mr;
173 /* 217 /*
174 * is the resource timed out? 218 * is the resource timed out?
175 * make everyone ignore the cancelled message. 219 * make everyone ignore the cancelled message.
@@ -177,10 +221,8 @@ static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
177 msg2->canceled = 1; 221 msg2->canceled = 1;
178 stat->d_canceled++; 222 stat->d_canceled++;
179 cancel_count++; 223 cancel_count++;
180 uv_write_local_mmr( 224 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
181 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, 225 write_mmr_sw_ack(mr);
182 (msg_res << UV_SW_ACK_NPENDING) |
183 msg_res);
184 } 226 }
185 } 227 }
186 } 228 }
@@ -192,20 +234,19 @@ static inline void uv_bau_process_retry_msg(struct msg_desc *mdp,
192 * Do all the things a cpu should do for a TLB shootdown message. 234 * Do all the things a cpu should do for a TLB shootdown message.
193 * Other cpu's may come here at the same time for this message. 235 * Other cpu's may come here at the same time for this message.
194 */ 236 */
195static void uv_bau_process_message(struct msg_desc *mdp, 237static void bau_process_message(struct msg_desc *mdp,
196 struct bau_control *bcp) 238 struct bau_control *bcp)
197{ 239{
198 int msg_ack_count;
199 short socket_ack_count = 0; 240 short socket_ack_count = 0;
200 struct ptc_stats *stat; 241 short *sp;
201 struct bau_payload_queue_entry *msg; 242 struct atomic_short *asp;
243 struct ptc_stats *stat = bcp->statp;
244 struct bau_pq_entry *msg = mdp->msg;
202 struct bau_control *smaster = bcp->socket_master; 245 struct bau_control *smaster = bcp->socket_master;
203 246
204 /* 247 /*
205 * This must be a normal message, or retry of a normal message 248 * This must be a normal message, or retry of a normal message
206 */ 249 */
207 msg = mdp->msg;
208 stat = bcp->statp;
209 if (msg->address == TLB_FLUSH_ALL) { 250 if (msg->address == TLB_FLUSH_ALL) {
210 local_flush_tlb(); 251 local_flush_tlb();
211 stat->d_alltlb++; 252 stat->d_alltlb++;
@@ -222,30 +263,32 @@ static void uv_bau_process_message(struct msg_desc *mdp,
222 * cpu number. 263 * cpu number.
223 */ 264 */
224 if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master) 265 if (msg->msg_type == MSG_RETRY && bcp == bcp->uvhub_master)
225 uv_bau_process_retry_msg(mdp, bcp); 266 bau_process_retry_msg(mdp, bcp);
226 267
227 /* 268 /*
228 * This is a sw_ack message, so we have to reply to it. 269 * This is a swack message, so we have to reply to it.
229 * Count each responding cpu on the socket. This avoids 270 * Count each responding cpu on the socket. This avoids
230 * pinging the count's cache line back and forth between 271 * pinging the count's cache line back and forth between
231 * the sockets. 272 * the sockets.
232 */ 273 */
233 socket_ack_count = atomic_add_short_return(1, (struct atomic_short *) 274 sp = &smaster->socket_acknowledge_count[mdp->msg_slot];
234 &smaster->socket_acknowledge_count[mdp->msg_slot]); 275 asp = (struct atomic_short *)sp;
276 socket_ack_count = atom_asr(1, asp);
235 if (socket_ack_count == bcp->cpus_in_socket) { 277 if (socket_ack_count == bcp->cpus_in_socket) {
278 int msg_ack_count;
236 /* 279 /*
237 * Both sockets dump their completed count total into 280 * Both sockets dump their completed count total into
238 * the message's count. 281 * the message's count.
239 */ 282 */
240 smaster->socket_acknowledge_count[mdp->msg_slot] = 0; 283 smaster->socket_acknowledge_count[mdp->msg_slot] = 0;
241 msg_ack_count = atomic_add_short_return(socket_ack_count, 284 asp = (struct atomic_short *)&msg->acknowledge_count;
242 (struct atomic_short *)&msg->acknowledge_count); 285 msg_ack_count = atom_asr(socket_ack_count, asp);
243 286
244 if (msg_ack_count == bcp->cpus_in_uvhub) { 287 if (msg_ack_count == bcp->cpus_in_uvhub) {
245 /* 288 /*
246 * All cpus in uvhub saw it; reply 289 * All cpus in uvhub saw it; reply
247 */ 290 */
248 uv_reply_to_message(mdp, bcp); 291 reply_to_message(mdp, bcp);
249 } 292 }
250 } 293 }
251 294
@@ -268,62 +311,51 @@ static int uvhub_to_first_cpu(int uvhub)
268 * Last resort when we get a large number of destination timeouts is 311 * Last resort when we get a large number of destination timeouts is
269 * to clear resources held by a given cpu. 312 * to clear resources held by a given cpu.
270 * Do this with IPI so that all messages in the BAU message queue 313 * Do this with IPI so that all messages in the BAU message queue
271 * can be identified by their nonzero sw_ack_vector field. 314 * can be identified by their nonzero swack_vec field.
272 * 315 *
273 * This is entered for a single cpu on the uvhub. 316 * This is entered for a single cpu on the uvhub.
274 * The sender want's this uvhub to free a specific message's 317 * The sender want's this uvhub to free a specific message's
275 * sw_ack resources. 318 * swack resources.
276 */ 319 */
277static void 320static void do_reset(void *ptr)
278uv_do_reset(void *ptr)
279{ 321{
280 int i; 322 int i;
281 int slot; 323 struct bau_control *bcp = &per_cpu(bau_control, smp_processor_id());
282 int count = 0; 324 struct reset_args *rap = (struct reset_args *)ptr;
283 unsigned long mmr; 325 struct bau_pq_entry *msg;
284 unsigned long msg_res; 326 struct ptc_stats *stat = bcp->statp;
285 struct bau_control *bcp;
286 struct reset_args *rap;
287 struct bau_payload_queue_entry *msg;
288 struct ptc_stats *stat;
289 327
290 bcp = &per_cpu(bau_control, smp_processor_id());
291 rap = (struct reset_args *)ptr;
292 stat = bcp->statp;
293 stat->d_resets++; 328 stat->d_resets++;
294
295 /* 329 /*
296 * We're looking for the given sender, and 330 * We're looking for the given sender, and
297 * will free its sw_ack resource. 331 * will free its swack resource.
298 * If all cpu's finally responded after the timeout, its 332 * If all cpu's finally responded after the timeout, its
299 * message 'replied_to' was set. 333 * message 'replied_to' was set.
300 */ 334 */
301 for (msg = bcp->va_queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) { 335 for (msg = bcp->queue_first, i = 0; i < DEST_Q_SIZE; msg++, i++) {
302 /* uv_do_reset: same conditions for cancellation as 336 unsigned long msg_res;
303 uv_bau_process_retry_msg() */ 337 /* do_reset: same conditions for cancellation as
338 bau_process_retry_msg() */
304 if ((msg->replied_to == 0) && 339 if ((msg->replied_to == 0) &&
305 (msg->canceled == 0) && 340 (msg->canceled == 0) &&
306 (msg->sending_cpu == rap->sender) && 341 (msg->sending_cpu == rap->sender) &&
307 (msg->sw_ack_vector) && 342 (msg->swack_vec) &&
308 (msg->msg_type != MSG_NOOP)) { 343 (msg->msg_type != MSG_NOOP)) {
344 unsigned long mmr;
345 unsigned long mr;
309 /* 346 /*
310 * make everyone else ignore this message 347 * make everyone else ignore this message
311 */ 348 */
312 msg->canceled = 1; 349 msg->canceled = 1;
313 slot = msg - bcp->va_queue_first;
314 count++;
315 /* 350 /*
316 * only reset the resource if it is still pending 351 * only reset the resource if it is still pending
317 */ 352 */
318 mmr = uv_read_local_mmr 353 mmr = read_mmr_sw_ack();
319 (UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE); 354 msg_res = msg->swack_vec;
320 msg_res = msg->sw_ack_vector; 355 mr = (msg_res << UV_SW_ACK_NPENDING) | msg_res;
321 if (mmr & msg_res) { 356 if (mmr & msg_res) {
322 stat->d_rcanceled++; 357 stat->d_rcanceled++;
323 uv_write_local_mmr( 358 write_mmr_sw_ack(mr);
324 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS,
325 (msg_res << UV_SW_ACK_NPENDING) |
326 msg_res);
327 } 359 }
328 } 360 }
329 } 361 }
@@ -334,39 +366,38 @@ uv_do_reset(void *ptr)
334 * Use IPI to get all target uvhubs to release resources held by 366 * Use IPI to get all target uvhubs to release resources held by
335 * a given sending cpu number. 367 * a given sending cpu number.
336 */ 368 */
337static void uv_reset_with_ipi(struct bau_target_uvhubmask *distribution, 369static void reset_with_ipi(struct bau_targ_hubmask *distribution, int sender)
338 int sender)
339{ 370{
340 int uvhub; 371 int uvhub;
341 int cpu; 372 int maskbits;
342 cpumask_t mask; 373 cpumask_t mask;
343 struct reset_args reset_args; 374 struct reset_args reset_args;
344 375
345 reset_args.sender = sender; 376 reset_args.sender = sender;
346
347 cpus_clear(mask); 377 cpus_clear(mask);
348 /* find a single cpu for each uvhub in this distribution mask */ 378 /* find a single cpu for each uvhub in this distribution mask */
349 for (uvhub = 0; 379 maskbits = sizeof(struct bau_targ_hubmask) * BITSPERBYTE;
350 uvhub < sizeof(struct bau_target_uvhubmask) * BITSPERBYTE; 380 for (uvhub = 0; uvhub < maskbits; uvhub++) {
351 uvhub++) { 381 int cpu;
352 if (!bau_uvhub_isset(uvhub, distribution)) 382 if (!bau_uvhub_isset(uvhub, distribution))
353 continue; 383 continue;
354 /* find a cpu for this uvhub */ 384 /* find a cpu for this uvhub */
355 cpu = uvhub_to_first_cpu(uvhub); 385 cpu = uvhub_to_first_cpu(uvhub);
356 cpu_set(cpu, mask); 386 cpu_set(cpu, mask);
357 } 387 }
358 /* IPI all cpus; Preemption is already disabled */ 388
359 smp_call_function_many(&mask, uv_do_reset, (void *)&reset_args, 1); 389 /* IPI all cpus; preemption is already disabled */
390 smp_call_function_many(&mask, do_reset, (void *)&reset_args, 1);
360 return; 391 return;
361} 392}
362 393
363static inline unsigned long 394static inline unsigned long cycles_2_us(unsigned long long cyc)
364cycles_2_us(unsigned long long cyc)
365{ 395{
366 unsigned long long ns; 396 unsigned long long ns;
367 unsigned long us; 397 unsigned long us;
368 ns = (cyc * per_cpu(cyc2ns, smp_processor_id())) 398 int cpu = smp_processor_id();
369 >> CYC2NS_SCALE_FACTOR; 399
400 ns = (cyc * per_cpu(cyc2ns, cpu)) >> CYC2NS_SCALE_FACTOR;
370 us = ns / 1000; 401 us = ns / 1000;
371 return us; 402 return us;
372} 403}
@@ -376,56 +407,56 @@ cycles_2_us(unsigned long long cyc)
376 * leaves uvhub_quiesce set so that no new broadcasts are started by 407 * leaves uvhub_quiesce set so that no new broadcasts are started by
377 * bau_flush_send_and_wait() 408 * bau_flush_send_and_wait()
378 */ 409 */
379static inline void 410static inline void quiesce_local_uvhub(struct bau_control *hmaster)
380quiesce_local_uvhub(struct bau_control *hmaster)
381{ 411{
382 atomic_add_short_return(1, (struct atomic_short *) 412 atom_asr(1, (struct atomic_short *)&hmaster->uvhub_quiesce);
383 &hmaster->uvhub_quiesce);
384} 413}
385 414
386/* 415/*
387 * mark this quiet-requestor as done 416 * mark this quiet-requestor as done
388 */ 417 */
389static inline void 418static inline void end_uvhub_quiesce(struct bau_control *hmaster)
390end_uvhub_quiesce(struct bau_control *hmaster)
391{ 419{
392 atomic_add_short_return(-1, (struct atomic_short *) 420 atom_asr(-1, (struct atomic_short *)&hmaster->uvhub_quiesce);
393 &hmaster->uvhub_quiesce); 421}
422
423static unsigned long uv1_read_status(unsigned long mmr_offset, int right_shift)
424{
425 unsigned long descriptor_status;
426
427 descriptor_status = uv_read_local_mmr(mmr_offset);
428 descriptor_status >>= right_shift;
429 descriptor_status &= UV_ACT_STATUS_MASK;
430 return descriptor_status;
394} 431}
395 432
396/* 433/*
397 * Wait for completion of a broadcast software ack message 434 * Wait for completion of a broadcast software ack message
398 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP 435 * return COMPLETE, RETRY(PLUGGED or TIMEOUT) or GIVEUP
399 */ 436 */
400static int uv_wait_completion(struct bau_desc *bau_desc, 437static int uv1_wait_completion(struct bau_desc *bau_desc,
401 unsigned long mmr_offset, int right_shift, int this_cpu, 438 unsigned long mmr_offset, int right_shift,
402 struct bau_control *bcp, struct bau_control *smaster, long try) 439 struct bau_control *bcp, long try)
403{ 440{
404 unsigned long descriptor_status; 441 unsigned long descriptor_status;
405 cycles_t ttime; 442 cycles_t ttm;
406 struct ptc_stats *stat = bcp->statp; 443 struct ptc_stats *stat = bcp->statp;
407 struct bau_control *hmaster;
408
409 hmaster = bcp->uvhub_master;
410 444
445 descriptor_status = uv1_read_status(mmr_offset, right_shift);
411 /* spin on the status MMR, waiting for it to go idle */ 446 /* spin on the status MMR, waiting for it to go idle */
412 while ((descriptor_status = (((unsigned long) 447 while ((descriptor_status != DS_IDLE)) {
413 uv_read_local_mmr(mmr_offset) >>
414 right_shift) & UV_ACT_STATUS_MASK)) !=
415 DESC_STATUS_IDLE) {
416 /* 448 /*
417 * Our software ack messages may be blocked because there are 449 * Our software ack messages may be blocked because
418 * no swack resources available. As long as none of them 450 * there are no swack resources available. As long
419 * has timed out hardware will NACK our message and its 451 * as none of them has timed out hardware will NACK
420 * state will stay IDLE. 452 * our message and its state will stay IDLE.
421 */ 453 */
422 if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) { 454 if (descriptor_status == DS_SOURCE_TIMEOUT) {
423 stat->s_stimeout++; 455 stat->s_stimeout++;
424 return FLUSH_GIVEUP; 456 return FLUSH_GIVEUP;
425 } else if (descriptor_status == 457 } else if (descriptor_status == DS_DESTINATION_TIMEOUT) {
426 DESC_STATUS_DESTINATION_TIMEOUT) {
427 stat->s_dtimeout++; 458 stat->s_dtimeout++;
428 ttime = get_cycles(); 459 ttm = get_cycles();
429 460
430 /* 461 /*
431 * Our retries may be blocked by all destination 462 * Our retries may be blocked by all destination
@@ -433,8 +464,7 @@ static int uv_wait_completion(struct bau_desc *bau_desc,
433 * pending. In that case hardware returns the 464 * pending. In that case hardware returns the
434 * ERROR that looks like a destination timeout. 465 * ERROR that looks like a destination timeout.
435 */ 466 */
436 if (cycles_2_us(ttime - bcp->send_message) < 467 if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
437 timeout_us) {
438 bcp->conseccompletes = 0; 468 bcp->conseccompletes = 0;
439 return FLUSH_RETRY_PLUGGED; 469 return FLUSH_RETRY_PLUGGED;
440 } 470 }
@@ -447,80 +477,160 @@ static int uv_wait_completion(struct bau_desc *bau_desc,
447 */ 477 */
448 cpu_relax(); 478 cpu_relax();
449 } 479 }
480 descriptor_status = uv1_read_status(mmr_offset, right_shift);
450 } 481 }
451 bcp->conseccompletes++; 482 bcp->conseccompletes++;
452 return FLUSH_COMPLETE; 483 return FLUSH_COMPLETE;
453} 484}
454 485
455static inline cycles_t 486/*
456sec_2_cycles(unsigned long sec) 487 * UV2 has an extra bit of status in the ACTIVATION_STATUS_2 register.
488 */
489static unsigned long uv2_read_status(unsigned long offset, int rshft, int cpu)
457{ 490{
458 unsigned long ns; 491 unsigned long descriptor_status;
459 cycles_t cyc; 492 unsigned long descriptor_status2;
460 493
461 ns = sec * 1000000000; 494 descriptor_status = ((read_lmmr(offset) >> rshft) & UV_ACT_STATUS_MASK);
462 cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id())); 495 descriptor_status2 = (read_mmr_uv2_status() >> cpu) & 0x1UL;
463 return cyc; 496 descriptor_status = (descriptor_status << 1) | descriptor_status2;
497 return descriptor_status;
498}
499
500static int uv2_wait_completion(struct bau_desc *bau_desc,
501 unsigned long mmr_offset, int right_shift,
502 struct bau_control *bcp, long try)
503{
504 unsigned long descriptor_stat;
505 cycles_t ttm;
506 int cpu = bcp->uvhub_cpu;
507 struct ptc_stats *stat = bcp->statp;
508
509 descriptor_stat = uv2_read_status(mmr_offset, right_shift, cpu);
510
511 /* spin on the status MMR, waiting for it to go idle */
512 while (descriptor_stat != UV2H_DESC_IDLE) {
513 /*
514 * Our software ack messages may be blocked because
515 * there are no swack resources available. As long
516 * as none of them has timed out hardware will NACK
517 * our message and its state will stay IDLE.
518 */
519 if ((descriptor_stat == UV2H_DESC_SOURCE_TIMEOUT) ||
520 (descriptor_stat == UV2H_DESC_DEST_STRONG_NACK) ||
521 (descriptor_stat == UV2H_DESC_DEST_PUT_ERR)) {
522 stat->s_stimeout++;
523 return FLUSH_GIVEUP;
524 } else if (descriptor_stat == UV2H_DESC_DEST_TIMEOUT) {
525 stat->s_dtimeout++;
526 ttm = get_cycles();
527 /*
528 * Our retries may be blocked by all destination
529 * swack resources being consumed, and a timeout
530 * pending. In that case hardware returns the
531 * ERROR that looks like a destination timeout.
532 */
533 if (cycles_2_us(ttm - bcp->send_message) < timeout_us) {
534 bcp->conseccompletes = 0;
535 return FLUSH_RETRY_PLUGGED;
536 }
537 bcp->conseccompletes = 0;
538 return FLUSH_RETRY_TIMEOUT;
539 } else {
540 /*
541 * descriptor_stat is still BUSY
542 */
543 cpu_relax();
544 }
545 descriptor_stat = uv2_read_status(mmr_offset, right_shift, cpu);
546 }
547 bcp->conseccompletes++;
548 return FLUSH_COMPLETE;
464} 549}
465 550
466/* 551/*
467 * conditionally add 1 to *v, unless *v is >= u 552 * There are 2 status registers; each and array[32] of 2 bits. Set up for
468 * return 0 if we cannot add 1 to *v because it is >= u 553 * which register to read and position in that register based on cpu in
469 * return 1 if we can add 1 to *v because it is < u 554 * current hub.
470 * the add is atomic
471 *
472 * This is close to atomic_add_unless(), but this allows the 'u' value
473 * to be lowered below the current 'v'. atomic_add_unless can only stop
474 * on equal.
475 */ 555 */
476static inline int atomic_inc_unless_ge(spinlock_t *lock, atomic_t *v, int u) 556static int wait_completion(struct bau_desc *bau_desc,
557 struct bau_control *bcp, long try)
477{ 558{
478 spin_lock(lock); 559 int right_shift;
479 if (atomic_read(v) >= u) { 560 unsigned long mmr_offset;
480 spin_unlock(lock); 561 int cpu = bcp->uvhub_cpu;
481 return 0; 562
563 if (cpu < UV_CPUS_PER_AS) {
564 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
565 right_shift = cpu * UV_ACT_STATUS_SIZE;
566 } else {
567 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
568 right_shift = ((cpu - UV_CPUS_PER_AS) * UV_ACT_STATUS_SIZE);
482 } 569 }
483 atomic_inc(v); 570
484 spin_unlock(lock); 571 if (is_uv1_hub())
485 return 1; 572 return uv1_wait_completion(bau_desc, mmr_offset, right_shift,
573 bcp, try);
574 else
575 return uv2_wait_completion(bau_desc, mmr_offset, right_shift,
576 bcp, try);
577}
578
579static inline cycles_t sec_2_cycles(unsigned long sec)
580{
581 unsigned long ns;
582 cycles_t cyc;
583
584 ns = sec * 1000000000;
585 cyc = (ns << CYC2NS_SCALE_FACTOR)/(per_cpu(cyc2ns, smp_processor_id()));
586 return cyc;
486} 587}
487 588
488/* 589/*
489 * Our retries are blocked by all destination swack resources being 590 * Our retries are blocked by all destination sw ack resources being
490 * in use, and a timeout is pending. In that case hardware immediately 591 * in use, and a timeout is pending. In that case hardware immediately
491 * returns the ERROR that looks like a destination timeout. 592 * returns the ERROR that looks like a destination timeout.
492 */ 593 */
493static void 594static void destination_plugged(struct bau_desc *bau_desc,
494destination_plugged(struct bau_desc *bau_desc, struct bau_control *bcp, 595 struct bau_control *bcp,
495 struct bau_control *hmaster, struct ptc_stats *stat) 596 struct bau_control *hmaster, struct ptc_stats *stat)
496{ 597{
497 udelay(bcp->plugged_delay); 598 udelay(bcp->plugged_delay);
498 bcp->plugged_tries++; 599 bcp->plugged_tries++;
600
499 if (bcp->plugged_tries >= bcp->plugsb4reset) { 601 if (bcp->plugged_tries >= bcp->plugsb4reset) {
500 bcp->plugged_tries = 0; 602 bcp->plugged_tries = 0;
603
501 quiesce_local_uvhub(hmaster); 604 quiesce_local_uvhub(hmaster);
605
502 spin_lock(&hmaster->queue_lock); 606 spin_lock(&hmaster->queue_lock);
503 uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu); 607 reset_with_ipi(&bau_desc->distribution, bcp->cpu);
504 spin_unlock(&hmaster->queue_lock); 608 spin_unlock(&hmaster->queue_lock);
609
505 end_uvhub_quiesce(hmaster); 610 end_uvhub_quiesce(hmaster);
611
506 bcp->ipi_attempts++; 612 bcp->ipi_attempts++;
507 stat->s_resets_plug++; 613 stat->s_resets_plug++;
508 } 614 }
509} 615}
510 616
511static void 617static void destination_timeout(struct bau_desc *bau_desc,
512destination_timeout(struct bau_desc *bau_desc, struct bau_control *bcp, 618 struct bau_control *bcp, struct bau_control *hmaster,
513 struct bau_control *hmaster, struct ptc_stats *stat) 619 struct ptc_stats *stat)
514{ 620{
515 hmaster->max_bau_concurrent = 1; 621 hmaster->max_concurr = 1;
516 bcp->timeout_tries++; 622 bcp->timeout_tries++;
517 if (bcp->timeout_tries >= bcp->timeoutsb4reset) { 623 if (bcp->timeout_tries >= bcp->timeoutsb4reset) {
518 bcp->timeout_tries = 0; 624 bcp->timeout_tries = 0;
625
519 quiesce_local_uvhub(hmaster); 626 quiesce_local_uvhub(hmaster);
627
520 spin_lock(&hmaster->queue_lock); 628 spin_lock(&hmaster->queue_lock);
521 uv_reset_with_ipi(&bau_desc->distribution, bcp->cpu); 629 reset_with_ipi(&bau_desc->distribution, bcp->cpu);
522 spin_unlock(&hmaster->queue_lock); 630 spin_unlock(&hmaster->queue_lock);
631
523 end_uvhub_quiesce(hmaster); 632 end_uvhub_quiesce(hmaster);
633
524 bcp->ipi_attempts++; 634 bcp->ipi_attempts++;
525 stat->s_resets_timeout++; 635 stat->s_resets_timeout++;
526 } 636 }
@@ -530,34 +640,104 @@ destination_timeout(struct bau_desc *bau_desc, struct bau_control *bcp,
530 * Completions are taking a very long time due to a congested numalink 640 * Completions are taking a very long time due to a congested numalink
531 * network. 641 * network.
532 */ 642 */
533static void 643static void disable_for_congestion(struct bau_control *bcp,
534disable_for_congestion(struct bau_control *bcp, struct ptc_stats *stat) 644 struct ptc_stats *stat)
535{ 645{
536 int tcpu;
537 struct bau_control *tbcp;
538
539 /* let only one cpu do this disabling */ 646 /* let only one cpu do this disabling */
540 spin_lock(&disable_lock); 647 spin_lock(&disable_lock);
648
541 if (!baudisabled && bcp->period_requests && 649 if (!baudisabled && bcp->period_requests &&
542 ((bcp->period_time / bcp->period_requests) > congested_cycles)) { 650 ((bcp->period_time / bcp->period_requests) > congested_cycles)) {
651 int tcpu;
652 struct bau_control *tbcp;
543 /* it becomes this cpu's job to turn on the use of the 653 /* it becomes this cpu's job to turn on the use of the
544 BAU again */ 654 BAU again */
545 baudisabled = 1; 655 baudisabled = 1;
546 bcp->set_bau_off = 1; 656 bcp->set_bau_off = 1;
547 bcp->set_bau_on_time = get_cycles() + 657 bcp->set_bau_on_time = get_cycles();
548 sec_2_cycles(bcp->congested_period); 658 bcp->set_bau_on_time += sec_2_cycles(bcp->cong_period);
549 stat->s_bau_disabled++; 659 stat->s_bau_disabled++;
550 for_each_present_cpu(tcpu) { 660 for_each_present_cpu(tcpu) {
551 tbcp = &per_cpu(bau_control, tcpu); 661 tbcp = &per_cpu(bau_control, tcpu);
552 tbcp->baudisabled = 1; 662 tbcp->baudisabled = 1;
553 } 663 }
554 } 664 }
665
555 spin_unlock(&disable_lock); 666 spin_unlock(&disable_lock);
556} 667}
557 668
558/** 669static void count_max_concurr(int stat, struct bau_control *bcp,
559 * uv_flush_send_and_wait 670 struct bau_control *hmaster)
560 * 671{
672 bcp->plugged_tries = 0;
673 bcp->timeout_tries = 0;
674 if (stat != FLUSH_COMPLETE)
675 return;
676 if (bcp->conseccompletes <= bcp->complete_threshold)
677 return;
678 if (hmaster->max_concurr >= hmaster->max_concurr_const)
679 return;
680 hmaster->max_concurr++;
681}
682
683static void record_send_stats(cycles_t time1, cycles_t time2,
684 struct bau_control *bcp, struct ptc_stats *stat,
685 int completion_status, int try)
686{
687 cycles_t elapsed;
688
689 if (time2 > time1) {
690 elapsed = time2 - time1;
691 stat->s_time += elapsed;
692
693 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) {
694 bcp->period_requests++;
695 bcp->period_time += elapsed;
696 if ((elapsed > congested_cycles) &&
697 (bcp->period_requests > bcp->cong_reps))
698 disable_for_congestion(bcp, stat);
699 }
700 } else
701 stat->s_requestor--;
702
703 if (completion_status == FLUSH_COMPLETE && try > 1)
704 stat->s_retriesok++;
705 else if (completion_status == FLUSH_GIVEUP)
706 stat->s_giveup++;
707}
708
709/*
710 * Because of a uv1 hardware bug only a limited number of concurrent
711 * requests can be made.
712 */
713static void uv1_throttle(struct bau_control *hmaster, struct ptc_stats *stat)
714{
715 spinlock_t *lock = &hmaster->uvhub_lock;
716 atomic_t *v;
717
718 v = &hmaster->active_descriptor_count;
719 if (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr)) {
720 stat->s_throttles++;
721 do {
722 cpu_relax();
723 } while (!atomic_inc_unless_ge(lock, v, hmaster->max_concurr));
724 }
725}
726
727/*
728 * Handle the completion status of a message send.
729 */
730static void handle_cmplt(int completion_status, struct bau_desc *bau_desc,
731 struct bau_control *bcp, struct bau_control *hmaster,
732 struct ptc_stats *stat)
733{
734 if (completion_status == FLUSH_RETRY_PLUGGED)
735 destination_plugged(bau_desc, bcp, hmaster, stat);
736 else if (completion_status == FLUSH_RETRY_TIMEOUT)
737 destination_timeout(bau_desc, bcp, hmaster, stat);
738}
739
740/*
561 * Send a broadcast and wait for it to complete. 741 * Send a broadcast and wait for it to complete.
562 * 742 *
563 * The flush_mask contains the cpus the broadcast is to be sent to including 743 * The flush_mask contains the cpus the broadcast is to be sent to including
@@ -568,44 +748,23 @@ disable_for_congestion(struct bau_control *bcp, struct ptc_stats *stat)
568 * returned to the kernel. 748 * returned to the kernel.
569 */ 749 */
570int uv_flush_send_and_wait(struct bau_desc *bau_desc, 750int uv_flush_send_and_wait(struct bau_desc *bau_desc,
571 struct cpumask *flush_mask, struct bau_control *bcp) 751 struct cpumask *flush_mask, struct bau_control *bcp)
572{ 752{
573 int right_shift;
574 int completion_status = 0;
575 int seq_number = 0; 753 int seq_number = 0;
754 int completion_stat = 0;
576 long try = 0; 755 long try = 0;
577 int cpu = bcp->uvhub_cpu;
578 int this_cpu = bcp->cpu;
579 unsigned long mmr_offset;
580 unsigned long index; 756 unsigned long index;
581 cycles_t time1; 757 cycles_t time1;
582 cycles_t time2; 758 cycles_t time2;
583 cycles_t elapsed;
584 struct ptc_stats *stat = bcp->statp; 759 struct ptc_stats *stat = bcp->statp;
585 struct bau_control *smaster = bcp->socket_master;
586 struct bau_control *hmaster = bcp->uvhub_master; 760 struct bau_control *hmaster = bcp->uvhub_master;
587 761
588 if (!atomic_inc_unless_ge(&hmaster->uvhub_lock, 762 if (is_uv1_hub())
589 &hmaster->active_descriptor_count, 763 uv1_throttle(hmaster, stat);
590 hmaster->max_bau_concurrent)) { 764
591 stat->s_throttles++;
592 do {
593 cpu_relax();
594 } while (!atomic_inc_unless_ge(&hmaster->uvhub_lock,
595 &hmaster->active_descriptor_count,
596 hmaster->max_bau_concurrent));
597 }
598 while (hmaster->uvhub_quiesce) 765 while (hmaster->uvhub_quiesce)
599 cpu_relax(); 766 cpu_relax();
600 767
601 if (cpu < UV_CPUS_PER_ACT_STATUS) {
602 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0;
603 right_shift = cpu * UV_ACT_STATUS_SIZE;
604 } else {
605 mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1;
606 right_shift =
607 ((cpu - UV_CPUS_PER_ACT_STATUS) * UV_ACT_STATUS_SIZE);
608 }
609 time1 = get_cycles(); 768 time1 = get_cycles();
610 do { 769 do {
611 if (try == 0) { 770 if (try == 0) {
@@ -615,64 +774,134 @@ int uv_flush_send_and_wait(struct bau_desc *bau_desc,
615 bau_desc->header.msg_type = MSG_RETRY; 774 bau_desc->header.msg_type = MSG_RETRY;
616 stat->s_retry_messages++; 775 stat->s_retry_messages++;
617 } 776 }
777
618 bau_desc->header.sequence = seq_number; 778 bau_desc->header.sequence = seq_number;
619 index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) | 779 index = (1UL << AS_PUSH_SHIFT) | bcp->uvhub_cpu;
620 bcp->uvhub_cpu;
621 bcp->send_message = get_cycles(); 780 bcp->send_message = get_cycles();
622 uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index); 781
782 write_mmr_activation(index);
783
623 try++; 784 try++;
624 completion_status = uv_wait_completion(bau_desc, mmr_offset, 785 completion_stat = wait_completion(bau_desc, bcp, try);
625 right_shift, this_cpu, bcp, smaster, try); 786
787 handle_cmplt(completion_stat, bau_desc, bcp, hmaster, stat);
626 788
627 if (completion_status == FLUSH_RETRY_PLUGGED) {
628 destination_plugged(bau_desc, bcp, hmaster, stat);
629 } else if (completion_status == FLUSH_RETRY_TIMEOUT) {
630 destination_timeout(bau_desc, bcp, hmaster, stat);
631 }
632 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) { 789 if (bcp->ipi_attempts >= bcp->ipi_reset_limit) {
633 bcp->ipi_attempts = 0; 790 bcp->ipi_attempts = 0;
634 completion_status = FLUSH_GIVEUP; 791 completion_stat = FLUSH_GIVEUP;
635 break; 792 break;
636 } 793 }
637 cpu_relax(); 794 cpu_relax();
638 } while ((completion_status == FLUSH_RETRY_PLUGGED) || 795 } while ((completion_stat == FLUSH_RETRY_PLUGGED) ||
639 (completion_status == FLUSH_RETRY_TIMEOUT)); 796 (completion_stat == FLUSH_RETRY_TIMEOUT));
797
640 time2 = get_cycles(); 798 time2 = get_cycles();
641 bcp->plugged_tries = 0; 799
642 bcp->timeout_tries = 0; 800 count_max_concurr(completion_stat, bcp, hmaster);
643 if ((completion_status == FLUSH_COMPLETE) && 801
644 (bcp->conseccompletes > bcp->complete_threshold) &&
645 (hmaster->max_bau_concurrent <
646 hmaster->max_bau_concurrent_constant))
647 hmaster->max_bau_concurrent++;
648 while (hmaster->uvhub_quiesce) 802 while (hmaster->uvhub_quiesce)
649 cpu_relax(); 803 cpu_relax();
804
650 atomic_dec(&hmaster->active_descriptor_count); 805 atomic_dec(&hmaster->active_descriptor_count);
651 if (time2 > time1) { 806
652 elapsed = time2 - time1; 807 record_send_stats(time1, time2, bcp, stat, completion_stat, try);
653 stat->s_time += elapsed; 808
654 if ((completion_status == FLUSH_COMPLETE) && (try == 1)) { 809 if (completion_stat == FLUSH_GIVEUP)
655 bcp->period_requests++; 810 return 1;
656 bcp->period_time += elapsed; 811 return 0;
657 if ((elapsed > congested_cycles) && 812}
658 (bcp->period_requests > bcp->congested_reps)) { 813
659 disable_for_congestion(bcp, stat); 814/*
815 * The BAU is disabled. When the disabled time period has expired, the cpu
816 * that disabled it must re-enable it.
817 * Return 0 if it is re-enabled for all cpus.
818 */
819static int check_enable(struct bau_control *bcp, struct ptc_stats *stat)
820{
821 int tcpu;
822 struct bau_control *tbcp;
823
824 if (bcp->set_bau_off) {
825 if (get_cycles() >= bcp->set_bau_on_time) {
826 stat->s_bau_reenabled++;
827 baudisabled = 0;
828 for_each_present_cpu(tcpu) {
829 tbcp = &per_cpu(bau_control, tcpu);
830 tbcp->baudisabled = 0;
831 tbcp->period_requests = 0;
832 tbcp->period_time = 0;
660 } 833 }
834 return 0;
661 } 835 }
836 }
837 return -1;
838}
839
840static void record_send_statistics(struct ptc_stats *stat, int locals, int hubs,
841 int remotes, struct bau_desc *bau_desc)
842{
843 stat->s_requestor++;
844 stat->s_ntargcpu += remotes + locals;
845 stat->s_ntargremotes += remotes;
846 stat->s_ntarglocals += locals;
847
848 /* uvhub statistics */
849 hubs = bau_uvhub_weight(&bau_desc->distribution);
850 if (locals) {
851 stat->s_ntarglocaluvhub++;
852 stat->s_ntargremoteuvhub += (hubs - 1);
662 } else 853 } else
663 stat->s_requestor--; 854 stat->s_ntargremoteuvhub += hubs;
664 if (completion_status == FLUSH_COMPLETE && try > 1) 855
665 stat->s_retriesok++; 856 stat->s_ntarguvhub += hubs;
666 else if (completion_status == FLUSH_GIVEUP) { 857
667 stat->s_giveup++; 858 if (hubs >= 16)
668 return 1; 859 stat->s_ntarguvhub16++;
860 else if (hubs >= 8)
861 stat->s_ntarguvhub8++;
862 else if (hubs >= 4)
863 stat->s_ntarguvhub4++;
864 else if (hubs >= 2)
865 stat->s_ntarguvhub2++;
866 else
867 stat->s_ntarguvhub1++;
868}
869
870/*
871 * Translate a cpu mask to the uvhub distribution mask in the BAU
872 * activation descriptor.
873 */
874static int set_distrib_bits(struct cpumask *flush_mask, struct bau_control *bcp,
875 struct bau_desc *bau_desc, int *localsp, int *remotesp)
876{
877 int cpu;
878 int pnode;
879 int cnt = 0;
880 struct hub_and_pnode *hpp;
881
882 for_each_cpu(cpu, flush_mask) {
883 /*
884 * The distribution vector is a bit map of pnodes, relative
885 * to the partition base pnode (and the partition base nasid
886 * in the header).
887 * Translate cpu to pnode and hub using a local memory array.
888 */
889 hpp = &bcp->socket_master->thp[cpu];
890 pnode = hpp->pnode - bcp->partition_base_pnode;
891 bau_uvhub_set(pnode, &bau_desc->distribution);
892 cnt++;
893 if (hpp->uvhub == bcp->uvhub)
894 (*localsp)++;
895 else
896 (*remotesp)++;
669 } 897 }
898 if (!cnt)
899 return 1;
670 return 0; 900 return 0;
671} 901}
672 902
673/** 903/*
674 * uv_flush_tlb_others - globally purge translation cache of a virtual 904 * globally purge translation cache of a virtual address or all TLB's
675 * address or all TLB's
676 * @cpumask: mask of all cpu's in which the address is to be removed 905 * @cpumask: mask of all cpu's in which the address is to be removed
677 * @mm: mm_struct containing virtual address range 906 * @mm: mm_struct containing virtual address range
678 * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu) 907 * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu)
@@ -696,20 +925,16 @@ int uv_flush_send_and_wait(struct bau_desc *bau_desc,
696 * done. The returned pointer is valid till preemption is re-enabled. 925 * done. The returned pointer is valid till preemption is re-enabled.
697 */ 926 */
698const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask, 927const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
699 struct mm_struct *mm, 928 struct mm_struct *mm, unsigned long va,
700 unsigned long va, unsigned int cpu) 929 unsigned int cpu)
701{ 930{
702 int locals = 0; 931 int locals = 0;
703 int remotes = 0; 932 int remotes = 0;
704 int hubs = 0; 933 int hubs = 0;
705 int tcpu;
706 int tpnode;
707 struct bau_desc *bau_desc; 934 struct bau_desc *bau_desc;
708 struct cpumask *flush_mask; 935 struct cpumask *flush_mask;
709 struct ptc_stats *stat; 936 struct ptc_stats *stat;
710 struct bau_control *bcp; 937 struct bau_control *bcp;
711 struct bau_control *tbcp;
712 struct hub_and_pnode *hpp;
713 938
714 /* kernel was booted 'nobau' */ 939 /* kernel was booted 'nobau' */
715 if (nobau) 940 if (nobau)
@@ -720,20 +945,8 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
720 945
721 /* bau was disabled due to slow response */ 946 /* bau was disabled due to slow response */
722 if (bcp->baudisabled) { 947 if (bcp->baudisabled) {
723 /* the cpu that disabled it must re-enable it */ 948 if (check_enable(bcp, stat))
724 if (bcp->set_bau_off) { 949 return cpumask;
725 if (get_cycles() >= bcp->set_bau_on_time) {
726 stat->s_bau_reenabled++;
727 baudisabled = 0;
728 for_each_present_cpu(tcpu) {
729 tbcp = &per_cpu(bau_control, tcpu);
730 tbcp->baudisabled = 0;
731 tbcp->period_requests = 0;
732 tbcp->period_time = 0;
733 }
734 }
735 }
736 return cpumask;
737 } 950 }
738 951
739 /* 952 /*
@@ -744,59 +957,20 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
744 flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu); 957 flush_mask = (struct cpumask *)per_cpu(uv_flush_tlb_mask, cpu);
745 /* don't actually do a shootdown of the local cpu */ 958 /* don't actually do a shootdown of the local cpu */
746 cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu)); 959 cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu));
960
747 if (cpu_isset(cpu, *cpumask)) 961 if (cpu_isset(cpu, *cpumask))
748 stat->s_ntargself++; 962 stat->s_ntargself++;
749 963
750 bau_desc = bcp->descriptor_base; 964 bau_desc = bcp->descriptor_base;
751 bau_desc += UV_ITEMS_PER_DESCRIPTOR * bcp->uvhub_cpu; 965 bau_desc += ITEMS_PER_DESC * bcp->uvhub_cpu;
752 bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE); 966 bau_uvhubs_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE);
753 967 if (set_distrib_bits(flush_mask, bcp, bau_desc, &locals, &remotes))
754 for_each_cpu(tcpu, flush_mask) {
755 /*
756 * The distribution vector is a bit map of pnodes, relative
757 * to the partition base pnode (and the partition base nasid
758 * in the header).
759 * Translate cpu to pnode and hub using an array stored
760 * in local memory.
761 */
762 hpp = &bcp->socket_master->target_hub_and_pnode[tcpu];
763 tpnode = hpp->pnode - bcp->partition_base_pnode;
764 bau_uvhub_set(tpnode, &bau_desc->distribution);
765 if (hpp->uvhub == bcp->uvhub)
766 locals++;
767 else
768 remotes++;
769 }
770 if ((locals + remotes) == 0)
771 return NULL; 968 return NULL;
772 stat->s_requestor++;
773 stat->s_ntargcpu += remotes + locals;
774 stat->s_ntargremotes += remotes;
775 stat->s_ntarglocals += locals;
776 remotes = bau_uvhub_weight(&bau_desc->distribution);
777 969
778 /* uvhub statistics */ 970 record_send_statistics(stat, locals, hubs, remotes, bau_desc);
779 hubs = bau_uvhub_weight(&bau_desc->distribution);
780 if (locals) {
781 stat->s_ntarglocaluvhub++;
782 stat->s_ntargremoteuvhub += (hubs - 1);
783 } else
784 stat->s_ntargremoteuvhub += hubs;
785 stat->s_ntarguvhub += hubs;
786 if (hubs >= 16)
787 stat->s_ntarguvhub16++;
788 else if (hubs >= 8)
789 stat->s_ntarguvhub8++;
790 else if (hubs >= 4)
791 stat->s_ntarguvhub4++;
792 else if (hubs >= 2)
793 stat->s_ntarguvhub2++;
794 else
795 stat->s_ntarguvhub1++;
796 971
797 bau_desc->payload.address = va; 972 bau_desc->payload.address = va;
798 bau_desc->payload.sending_cpu = cpu; 973 bau_desc->payload.sending_cpu = cpu;
799
800 /* 974 /*
801 * uv_flush_send_and_wait returns 0 if all cpu's were messaged, 975 * uv_flush_send_and_wait returns 0 if all cpu's were messaged,
802 * or 1 if it gave up and the original cpumask should be returned. 976 * or 1 if it gave up and the original cpumask should be returned.
@@ -825,26 +999,31 @@ void uv_bau_message_interrupt(struct pt_regs *regs)
825{ 999{
826 int count = 0; 1000 int count = 0;
827 cycles_t time_start; 1001 cycles_t time_start;
828 struct bau_payload_queue_entry *msg; 1002 struct bau_pq_entry *msg;
829 struct bau_control *bcp; 1003 struct bau_control *bcp;
830 struct ptc_stats *stat; 1004 struct ptc_stats *stat;
831 struct msg_desc msgdesc; 1005 struct msg_desc msgdesc;
832 1006
833 time_start = get_cycles(); 1007 time_start = get_cycles();
1008
834 bcp = &per_cpu(bau_control, smp_processor_id()); 1009 bcp = &per_cpu(bau_control, smp_processor_id());
835 stat = bcp->statp; 1010 stat = bcp->statp;
836 msgdesc.va_queue_first = bcp->va_queue_first; 1011
837 msgdesc.va_queue_last = bcp->va_queue_last; 1012 msgdesc.queue_first = bcp->queue_first;
1013 msgdesc.queue_last = bcp->queue_last;
1014
838 msg = bcp->bau_msg_head; 1015 msg = bcp->bau_msg_head;
839 while (msg->sw_ack_vector) { 1016 while (msg->swack_vec) {
840 count++; 1017 count++;
841 msgdesc.msg_slot = msg - msgdesc.va_queue_first; 1018
842 msgdesc.sw_ack_slot = ffs(msg->sw_ack_vector) - 1; 1019 msgdesc.msg_slot = msg - msgdesc.queue_first;
1020 msgdesc.swack_slot = ffs(msg->swack_vec) - 1;
843 msgdesc.msg = msg; 1021 msgdesc.msg = msg;
844 uv_bau_process_message(&msgdesc, bcp); 1022 bau_process_message(&msgdesc, bcp);
1023
845 msg++; 1024 msg++;
846 if (msg > msgdesc.va_queue_last) 1025 if (msg > msgdesc.queue_last)
847 msg = msgdesc.va_queue_first; 1026 msg = msgdesc.queue_first;
848 bcp->bau_msg_head = msg; 1027 bcp->bau_msg_head = msg;
849 } 1028 }
850 stat->d_time += (get_cycles() - time_start); 1029 stat->d_time += (get_cycles() - time_start);
@@ -852,18 +1031,17 @@ void uv_bau_message_interrupt(struct pt_regs *regs)
852 stat->d_nomsg++; 1031 stat->d_nomsg++;
853 else if (count > 1) 1032 else if (count > 1)
854 stat->d_multmsg++; 1033 stat->d_multmsg++;
1034
855 ack_APIC_irq(); 1035 ack_APIC_irq();
856} 1036}
857 1037
858/* 1038/*
859 * uv_enable_timeouts 1039 * Each target uvhub (i.e. a uvhub that has cpu's) needs to have
860 *
861 * Each target uvhub (i.e. a uvhub that has no cpu's) needs to have
862 * shootdown message timeouts enabled. The timeout does not cause 1040 * shootdown message timeouts enabled. The timeout does not cause
863 * an interrupt, but causes an error message to be returned to 1041 * an interrupt, but causes an error message to be returned to
864 * the sender. 1042 * the sender.
865 */ 1043 */
866static void __init uv_enable_timeouts(void) 1044static void __init enable_timeouts(void)
867{ 1045{
868 int uvhub; 1046 int uvhub;
869 int nuvhubs; 1047 int nuvhubs;
@@ -877,47 +1055,44 @@ static void __init uv_enable_timeouts(void)
877 continue; 1055 continue;
878 1056
879 pnode = uv_blade_to_pnode(uvhub); 1057 pnode = uv_blade_to_pnode(uvhub);
880 mmr_image = 1058 mmr_image = read_mmr_misc_control(pnode);
881 uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL);
882 /* 1059 /*
883 * Set the timeout period and then lock it in, in three 1060 * Set the timeout period and then lock it in, in three
884 * steps; captures and locks in the period. 1061 * steps; captures and locks in the period.
885 * 1062 *
886 * To program the period, the SOFT_ACK_MODE must be off. 1063 * To program the period, the SOFT_ACK_MODE must be off.
887 */ 1064 */
888 mmr_image &= ~((unsigned long)1 << 1065 mmr_image &= ~(1L << SOFTACK_MSHIFT);
889 UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT); 1066 write_mmr_misc_control(pnode, mmr_image);
890 uv_write_global_mmr64
891 (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
892 /* 1067 /*
893 * Set the 4-bit period. 1068 * Set the 4-bit period.
894 */ 1069 */
895 mmr_image &= ~((unsigned long)0xf << 1070 mmr_image &= ~((unsigned long)0xf << SOFTACK_PSHIFT);
896 UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT); 1071 mmr_image |= (SOFTACK_TIMEOUT_PERIOD << SOFTACK_PSHIFT);
897 mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD << 1072 write_mmr_misc_control(pnode, mmr_image);
898 UVH_LB_BAU_MISC_CONTROL_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHFT);
899 uv_write_global_mmr64
900 (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image);
901 /* 1073 /*
1074 * UV1:
902 * Subsequent reversals of the timebase bit (3) cause an 1075 * Subsequent reversals of the timebase bit (3) cause an
903 * immediate timeout of one or all INTD resources as 1076 * immediate timeout of one or all INTD resources as
904 * indicated in bits 2:0 (7 causes all of them to timeout). 1077 * indicated in bits 2:0 (7 causes all of them to timeout).
905 */ 1078 */
906 mmr_image |= ((unsigned long)1 << 1079 mmr_image |= (1L << SOFTACK_MSHIFT);
907 UVH_LB_BAU_MISC_CONTROL_ENABLE_INTD_SOFT_ACK_MODE_SHFT); 1080 if (is_uv2_hub()) {
908 uv_write_global_mmr64 1081 mmr_image |= (1L << UV2_LEG_SHFT);
909 (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image); 1082 mmr_image |= (1L << UV2_EXT_SHFT);
1083 }
1084 write_mmr_misc_control(pnode, mmr_image);
910 } 1085 }
911} 1086}
912 1087
913static void *uv_ptc_seq_start(struct seq_file *file, loff_t *offset) 1088static void *ptc_seq_start(struct seq_file *file, loff_t *offset)
914{ 1089{
915 if (*offset < num_possible_cpus()) 1090 if (*offset < num_possible_cpus())
916 return offset; 1091 return offset;
917 return NULL; 1092 return NULL;
918} 1093}
919 1094
920static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset) 1095static void *ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
921{ 1096{
922 (*offset)++; 1097 (*offset)++;
923 if (*offset < num_possible_cpus()) 1098 if (*offset < num_possible_cpus())
@@ -925,12 +1100,11 @@ static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset)
925 return NULL; 1100 return NULL;
926} 1101}
927 1102
928static void uv_ptc_seq_stop(struct seq_file *file, void *data) 1103static void ptc_seq_stop(struct seq_file *file, void *data)
929{ 1104{
930} 1105}
931 1106
932static inline unsigned long long 1107static inline unsigned long long usec_2_cycles(unsigned long microsec)
933microsec_2_cycles(unsigned long microsec)
934{ 1108{
935 unsigned long ns; 1109 unsigned long ns;
936 unsigned long long cyc; 1110 unsigned long long cyc;
@@ -941,29 +1115,27 @@ microsec_2_cycles(unsigned long microsec)
941} 1115}
942 1116
943/* 1117/*
944 * Display the statistics thru /proc. 1118 * Display the statistics thru /proc/sgi_uv/ptc_statistics
945 * 'data' points to the cpu number 1119 * 'data' points to the cpu number
1120 * Note: see the descriptions in stat_description[].
946 */ 1121 */
947static int uv_ptc_seq_show(struct seq_file *file, void *data) 1122static int ptc_seq_show(struct seq_file *file, void *data)
948{ 1123{
949 struct ptc_stats *stat; 1124 struct ptc_stats *stat;
950 int cpu; 1125 int cpu;
951 1126
952 cpu = *(loff_t *)data; 1127 cpu = *(loff_t *)data;
953
954 if (!cpu) { 1128 if (!cpu) {
955 seq_printf(file, 1129 seq_printf(file,
956 "# cpu sent stime self locals remotes ncpus localhub "); 1130 "# cpu sent stime self locals remotes ncpus localhub ");
957 seq_printf(file, 1131 seq_printf(file,
958 "remotehub numuvhubs numuvhubs16 numuvhubs8 "); 1132 "remotehub numuvhubs numuvhubs16 numuvhubs8 ");
959 seq_printf(file, 1133 seq_printf(file,
960 "numuvhubs4 numuvhubs2 numuvhubs1 dto "); 1134 "numuvhubs4 numuvhubs2 numuvhubs1 dto retries rok ");
961 seq_printf(file,
962 "retries rok resetp resett giveup sto bz throt ");
963 seq_printf(file, 1135 seq_printf(file,
964 "sw_ack recv rtime all "); 1136 "resetp resett giveup sto bz throt swack recv rtime ");
965 seq_printf(file, 1137 seq_printf(file,
966 "one mult none retry canc nocan reset rcan "); 1138 "all one mult none retry canc nocan reset rcan ");
967 seq_printf(file, 1139 seq_printf(file,
968 "disable enable\n"); 1140 "disable enable\n");
969 } 1141 }
@@ -990,8 +1162,7 @@ static int uv_ptc_seq_show(struct seq_file *file, void *data)
990 /* destination side statistics */ 1162 /* destination side statistics */
991 seq_printf(file, 1163 seq_printf(file,
992 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ", 1164 "%lx %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld ",
993 uv_read_global_mmr64(uv_cpu_to_pnode(cpu), 1165 read_gmmr_sw_ack(uv_cpu_to_pnode(cpu)),
994 UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE),
995 stat->d_requestee, cycles_2_us(stat->d_time), 1166 stat->d_requestee, cycles_2_us(stat->d_time),
996 stat->d_alltlb, stat->d_onetlb, stat->d_multmsg, 1167 stat->d_alltlb, stat->d_onetlb, stat->d_multmsg,
997 stat->d_nomsg, stat->d_retries, stat->d_canceled, 1168 stat->d_nomsg, stat->d_retries, stat->d_canceled,
@@ -1000,7 +1171,6 @@ static int uv_ptc_seq_show(struct seq_file *file, void *data)
1000 seq_printf(file, "%ld %ld\n", 1171 seq_printf(file, "%ld %ld\n",
1001 stat->s_bau_disabled, stat->s_bau_reenabled); 1172 stat->s_bau_disabled, stat->s_bau_reenabled);
1002 } 1173 }
1003
1004 return 0; 1174 return 0;
1005} 1175}
1006 1176
@@ -1008,18 +1178,18 @@ static int uv_ptc_seq_show(struct seq_file *file, void *data)
1008 * Display the tunables thru debugfs 1178 * Display the tunables thru debugfs
1009 */ 1179 */
1010static ssize_t tunables_read(struct file *file, char __user *userbuf, 1180static ssize_t tunables_read(struct file *file, char __user *userbuf,
1011 size_t count, loff_t *ppos) 1181 size_t count, loff_t *ppos)
1012{ 1182{
1013 char *buf; 1183 char *buf;
1014 int ret; 1184 int ret;
1015 1185
1016 buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n", 1186 buf = kasprintf(GFP_KERNEL, "%s %s %s\n%d %d %d %d %d %d %d %d %d\n",
1017 "max_bau_concurrent plugged_delay plugsb4reset", 1187 "max_concur plugged_delay plugsb4reset",
1018 "timeoutsb4reset ipi_reset_limit complete_threshold", 1188 "timeoutsb4reset ipi_reset_limit complete_threshold",
1019 "congested_response_us congested_reps congested_period", 1189 "congested_response_us congested_reps congested_period",
1020 max_bau_concurrent, plugged_delay, plugsb4reset, 1190 max_concurr, plugged_delay, plugsb4reset,
1021 timeoutsb4reset, ipi_reset_limit, complete_threshold, 1191 timeoutsb4reset, ipi_reset_limit, complete_threshold,
1022 congested_response_us, congested_reps, congested_period); 1192 congested_respns_us, congested_reps, congested_period);
1023 1193
1024 if (!buf) 1194 if (!buf)
1025 return -ENOMEM; 1195 return -ENOMEM;
@@ -1030,13 +1200,16 @@ static ssize_t tunables_read(struct file *file, char __user *userbuf,
1030} 1200}
1031 1201
1032/* 1202/*
1033 * -1: resetf the statistics 1203 * handle a write to /proc/sgi_uv/ptc_statistics
1204 * -1: reset the statistics
1034 * 0: display meaning of the statistics 1205 * 0: display meaning of the statistics
1035 */ 1206 */
1036static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user, 1207static ssize_t ptc_proc_write(struct file *file, const char __user *user,
1037 size_t count, loff_t *data) 1208 size_t count, loff_t *data)
1038{ 1209{
1039 int cpu; 1210 int cpu;
1211 int i;
1212 int elements;
1040 long input_arg; 1213 long input_arg;
1041 char optstr[64]; 1214 char optstr[64];
1042 struct ptc_stats *stat; 1215 struct ptc_stats *stat;
@@ -1046,79 +1219,18 @@ static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user,
1046 if (copy_from_user(optstr, user, count)) 1219 if (copy_from_user(optstr, user, count))
1047 return -EFAULT; 1220 return -EFAULT;
1048 optstr[count - 1] = '\0'; 1221 optstr[count - 1] = '\0';
1222
1049 if (strict_strtol(optstr, 10, &input_arg) < 0) { 1223 if (strict_strtol(optstr, 10, &input_arg) < 0) {
1050 printk(KERN_DEBUG "%s is invalid\n", optstr); 1224 printk(KERN_DEBUG "%s is invalid\n", optstr);
1051 return -EINVAL; 1225 return -EINVAL;
1052 } 1226 }
1053 1227
1054 if (input_arg == 0) { 1228 if (input_arg == 0) {
1229 elements = sizeof(stat_description)/sizeof(*stat_description);
1055 printk(KERN_DEBUG "# cpu: cpu number\n"); 1230 printk(KERN_DEBUG "# cpu: cpu number\n");
1056 printk(KERN_DEBUG "Sender statistics:\n"); 1231 printk(KERN_DEBUG "Sender statistics:\n");
1057 printk(KERN_DEBUG 1232 for (i = 0; i < elements; i++)
1058 "sent: number of shootdown messages sent\n"); 1233 printk(KERN_DEBUG "%s\n", stat_description[i]);
1059 printk(KERN_DEBUG
1060 "stime: time spent sending messages\n");
1061 printk(KERN_DEBUG
1062 "numuvhubs: number of hubs targeted with shootdown\n");
1063 printk(KERN_DEBUG
1064 "numuvhubs16: number times 16 or more hubs targeted\n");
1065 printk(KERN_DEBUG
1066 "numuvhubs8: number times 8 or more hubs targeted\n");
1067 printk(KERN_DEBUG
1068 "numuvhubs4: number times 4 or more hubs targeted\n");
1069 printk(KERN_DEBUG
1070 "numuvhubs2: number times 2 or more hubs targeted\n");
1071 printk(KERN_DEBUG
1072 "numuvhubs1: number times 1 hub targeted\n");
1073 printk(KERN_DEBUG
1074 "numcpus: number of cpus targeted with shootdown\n");
1075 printk(KERN_DEBUG
1076 "dto: number of destination timeouts\n");
1077 printk(KERN_DEBUG
1078 "retries: destination timeout retries sent\n");
1079 printk(KERN_DEBUG
1080 "rok: : destination timeouts successfully retried\n");
1081 printk(KERN_DEBUG
1082 "resetp: ipi-style resource resets for plugs\n");
1083 printk(KERN_DEBUG
1084 "resett: ipi-style resource resets for timeouts\n");
1085 printk(KERN_DEBUG
1086 "giveup: fall-backs to ipi-style shootdowns\n");
1087 printk(KERN_DEBUG
1088 "sto: number of source timeouts\n");
1089 printk(KERN_DEBUG
1090 "bz: number of stay-busy's\n");
1091 printk(KERN_DEBUG
1092 "throt: number times spun in throttle\n");
1093 printk(KERN_DEBUG "Destination side statistics:\n");
1094 printk(KERN_DEBUG
1095 "sw_ack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n");
1096 printk(KERN_DEBUG
1097 "recv: shootdown messages received\n");
1098 printk(KERN_DEBUG
1099 "rtime: time spent processing messages\n");
1100 printk(KERN_DEBUG
1101 "all: shootdown all-tlb messages\n");
1102 printk(KERN_DEBUG
1103 "one: shootdown one-tlb messages\n");
1104 printk(KERN_DEBUG
1105 "mult: interrupts that found multiple messages\n");
1106 printk(KERN_DEBUG
1107 "none: interrupts that found no messages\n");
1108 printk(KERN_DEBUG
1109 "retry: number of retry messages processed\n");
1110 printk(KERN_DEBUG
1111 "canc: number messages canceled by retries\n");
1112 printk(KERN_DEBUG
1113 "nocan: number retries that found nothing to cancel\n");
1114 printk(KERN_DEBUG
1115 "reset: number of ipi-style reset requests processed\n");
1116 printk(KERN_DEBUG
1117 "rcan: number messages canceled by reset requests\n");
1118 printk(KERN_DEBUG
1119 "disable: number times use of the BAU was disabled\n");
1120 printk(KERN_DEBUG
1121 "enable: number times use of the BAU was re-enabled\n");
1122 } else if (input_arg == -1) { 1234 } else if (input_arg == -1) {
1123 for_each_present_cpu(cpu) { 1235 for_each_present_cpu(cpu) {
1124 stat = &per_cpu(ptcstats, cpu); 1236 stat = &per_cpu(ptcstats, cpu);
@@ -1145,27 +1257,18 @@ static int local_atoi(const char *name)
1145} 1257}
1146 1258
1147/* 1259/*
1148 * set the tunables 1260 * Parse the values written to /sys/kernel/debug/sgi_uv/bau_tunables.
1149 * 0 values reset them to defaults 1261 * Zero values reset them to defaults.
1150 */ 1262 */
1151static ssize_t tunables_write(struct file *file, const char __user *user, 1263static int parse_tunables_write(struct bau_control *bcp, char *instr,
1152 size_t count, loff_t *data) 1264 int count)
1153{ 1265{
1154 int cpu;
1155 int cnt = 0;
1156 int val;
1157 char *p; 1266 char *p;
1158 char *q; 1267 char *q;
1159 char instr[64]; 1268 int cnt = 0;
1160 struct bau_control *bcp; 1269 int val;
1161 1270 int e = sizeof(tunables) / sizeof(*tunables);
1162 if (count == 0 || count > sizeof(instr)-1)
1163 return -EINVAL;
1164 if (copy_from_user(instr, user, count))
1165 return -EFAULT;
1166 1271
1167 instr[count] = '\0';
1168 /* count the fields */
1169 p = instr + strspn(instr, WHITESPACE); 1272 p = instr + strspn(instr, WHITESPACE);
1170 q = p; 1273 q = p;
1171 for (; *p; p = q + strspn(q, WHITESPACE)) { 1274 for (; *p; p = q + strspn(q, WHITESPACE)) {
@@ -1174,8 +1277,8 @@ static ssize_t tunables_write(struct file *file, const char __user *user,
1174 if (q == p) 1277 if (q == p)
1175 break; 1278 break;
1176 } 1279 }
1177 if (cnt != 9) { 1280 if (cnt != e) {
1178 printk(KERN_INFO "bau tunable error: should be 9 numbers\n"); 1281 printk(KERN_INFO "bau tunable error: should be %d values\n", e);
1179 return -EINVAL; 1282 return -EINVAL;
1180 } 1283 }
1181 1284
@@ -1187,97 +1290,80 @@ static ssize_t tunables_write(struct file *file, const char __user *user,
1187 switch (cnt) { 1290 switch (cnt) {
1188 case 0: 1291 case 0:
1189 if (val == 0) { 1292 if (val == 0) {
1190 max_bau_concurrent = MAX_BAU_CONCURRENT; 1293 max_concurr = MAX_BAU_CONCURRENT;
1191 max_bau_concurrent_constant = 1294 max_concurr_const = MAX_BAU_CONCURRENT;
1192 MAX_BAU_CONCURRENT;
1193 continue; 1295 continue;
1194 } 1296 }
1195 bcp = &per_cpu(bau_control, smp_processor_id());
1196 if (val < 1 || val > bcp->cpus_in_uvhub) { 1297 if (val < 1 || val > bcp->cpus_in_uvhub) {
1197 printk(KERN_DEBUG 1298 printk(KERN_DEBUG
1198 "Error: BAU max concurrent %d is invalid\n", 1299 "Error: BAU max concurrent %d is invalid\n",
1199 val); 1300 val);
1200 return -EINVAL; 1301 return -EINVAL;
1201 } 1302 }
1202 max_bau_concurrent = val; 1303 max_concurr = val;
1203 max_bau_concurrent_constant = val; 1304 max_concurr_const = val;
1204 continue;
1205 case 1:
1206 if (val == 0)
1207 plugged_delay = PLUGGED_DELAY;
1208 else
1209 plugged_delay = val;
1210 continue;
1211 case 2:
1212 if (val == 0)
1213 plugsb4reset = PLUGSB4RESET;
1214 else
1215 plugsb4reset = val;
1216 continue;
1217 case 3:
1218 if (val == 0)
1219 timeoutsb4reset = TIMEOUTSB4RESET;
1220 else
1221 timeoutsb4reset = val;
1222 continue;
1223 case 4:
1224 if (val == 0)
1225 ipi_reset_limit = IPI_RESET_LIMIT;
1226 else
1227 ipi_reset_limit = val;
1228 continue;
1229 case 5:
1230 if (val == 0)
1231 complete_threshold = COMPLETE_THRESHOLD;
1232 else
1233 complete_threshold = val;
1234 continue;
1235 case 6:
1236 if (val == 0)
1237 congested_response_us = CONGESTED_RESPONSE_US;
1238 else
1239 congested_response_us = val;
1240 continue;
1241 case 7:
1242 if (val == 0)
1243 congested_reps = CONGESTED_REPS;
1244 else
1245 congested_reps = val;
1246 continue; 1305 continue;
1247 case 8: 1306 default:
1248 if (val == 0) 1307 if (val == 0)
1249 congested_period = CONGESTED_PERIOD; 1308 *tunables[cnt].tunp = tunables[cnt].deflt;
1250 else 1309 else
1251 congested_period = val; 1310 *tunables[cnt].tunp = val;
1252 continue; 1311 continue;
1253 } 1312 }
1254 if (q == p) 1313 if (q == p)
1255 break; 1314 break;
1256 } 1315 }
1316 return 0;
1317}
1318
1319/*
1320 * Handle a write to debugfs. (/sys/kernel/debug/sgi_uv/bau_tunables)
1321 */
1322static ssize_t tunables_write(struct file *file, const char __user *user,
1323 size_t count, loff_t *data)
1324{
1325 int cpu;
1326 int ret;
1327 char instr[100];
1328 struct bau_control *bcp;
1329
1330 if (count == 0 || count > sizeof(instr)-1)
1331 return -EINVAL;
1332 if (copy_from_user(instr, user, count))
1333 return -EFAULT;
1334
1335 instr[count] = '\0';
1336
1337 bcp = &per_cpu(bau_control, smp_processor_id());
1338
1339 ret = parse_tunables_write(bcp, instr, count);
1340 if (ret)
1341 return ret;
1342
1257 for_each_present_cpu(cpu) { 1343 for_each_present_cpu(cpu) {
1258 bcp = &per_cpu(bau_control, cpu); 1344 bcp = &per_cpu(bau_control, cpu);
1259 bcp->max_bau_concurrent = max_bau_concurrent; 1345 bcp->max_concurr = max_concurr;
1260 bcp->max_bau_concurrent_constant = max_bau_concurrent; 1346 bcp->max_concurr_const = max_concurr;
1261 bcp->plugged_delay = plugged_delay; 1347 bcp->plugged_delay = plugged_delay;
1262 bcp->plugsb4reset = plugsb4reset; 1348 bcp->plugsb4reset = plugsb4reset;
1263 bcp->timeoutsb4reset = timeoutsb4reset; 1349 bcp->timeoutsb4reset = timeoutsb4reset;
1264 bcp->ipi_reset_limit = ipi_reset_limit; 1350 bcp->ipi_reset_limit = ipi_reset_limit;
1265 bcp->complete_threshold = complete_threshold; 1351 bcp->complete_threshold = complete_threshold;
1266 bcp->congested_response_us = congested_response_us; 1352 bcp->cong_response_us = congested_respns_us;
1267 bcp->congested_reps = congested_reps; 1353 bcp->cong_reps = congested_reps;
1268 bcp->congested_period = congested_period; 1354 bcp->cong_period = congested_period;
1269 } 1355 }
1270 return count; 1356 return count;
1271} 1357}
1272 1358
1273static const struct seq_operations uv_ptc_seq_ops = { 1359static const struct seq_operations uv_ptc_seq_ops = {
1274 .start = uv_ptc_seq_start, 1360 .start = ptc_seq_start,
1275 .next = uv_ptc_seq_next, 1361 .next = ptc_seq_next,
1276 .stop = uv_ptc_seq_stop, 1362 .stop = ptc_seq_stop,
1277 .show = uv_ptc_seq_show 1363 .show = ptc_seq_show
1278}; 1364};
1279 1365
1280static int uv_ptc_proc_open(struct inode *inode, struct file *file) 1366static int ptc_proc_open(struct inode *inode, struct file *file)
1281{ 1367{
1282 return seq_open(file, &uv_ptc_seq_ops); 1368 return seq_open(file, &uv_ptc_seq_ops);
1283} 1369}
@@ -1288,9 +1374,9 @@ static int tunables_open(struct inode *inode, struct file *file)
1288} 1374}
1289 1375
1290static const struct file_operations proc_uv_ptc_operations = { 1376static const struct file_operations proc_uv_ptc_operations = {
1291 .open = uv_ptc_proc_open, 1377 .open = ptc_proc_open,
1292 .read = seq_read, 1378 .read = seq_read,
1293 .write = uv_ptc_proc_write, 1379 .write = ptc_proc_write,
1294 .llseek = seq_lseek, 1380 .llseek = seq_lseek,
1295 .release = seq_release, 1381 .release = seq_release,
1296}; 1382};
@@ -1324,7 +1410,7 @@ static int __init uv_ptc_init(void)
1324 return -EINVAL; 1410 return -EINVAL;
1325 } 1411 }
1326 tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600, 1412 tunables_file = debugfs_create_file(UV_BAU_TUNABLES_FILE, 0600,
1327 tunables_dir, NULL, &tunables_fops); 1413 tunables_dir, NULL, &tunables_fops);
1328 if (!tunables_file) { 1414 if (!tunables_file) {
1329 printk(KERN_ERR "unable to create debugfs file %s\n", 1415 printk(KERN_ERR "unable to create debugfs file %s\n",
1330 UV_BAU_TUNABLES_FILE); 1416 UV_BAU_TUNABLES_FILE);
@@ -1336,24 +1422,24 @@ static int __init uv_ptc_init(void)
1336/* 1422/*
1337 * Initialize the sending side's sending buffers. 1423 * Initialize the sending side's sending buffers.
1338 */ 1424 */
1339static void 1425static void activation_descriptor_init(int node, int pnode, int base_pnode)
1340uv_activation_descriptor_init(int node, int pnode, int base_pnode)
1341{ 1426{
1342 int i; 1427 int i;
1343 int cpu; 1428 int cpu;
1344 unsigned long pa; 1429 unsigned long pa;
1345 unsigned long m; 1430 unsigned long m;
1346 unsigned long n; 1431 unsigned long n;
1432 size_t dsize;
1347 struct bau_desc *bau_desc; 1433 struct bau_desc *bau_desc;
1348 struct bau_desc *bd2; 1434 struct bau_desc *bd2;
1349 struct bau_control *bcp; 1435 struct bau_control *bcp;
1350 1436
1351 /* 1437 /*
1352 * each bau_desc is 64 bytes; there are 8 (UV_ITEMS_PER_DESCRIPTOR) 1438 * each bau_desc is 64 bytes; there are 8 (ITEMS_PER_DESC)
1353 * per cpu; and one per cpu on the uvhub (UV_ADP_SIZE) 1439 * per cpu; and one per cpu on the uvhub (ADP_SZ)
1354 */ 1440 */
1355 bau_desc = kmalloc_node(sizeof(struct bau_desc) * UV_ADP_SIZE 1441 dsize = sizeof(struct bau_desc) * ADP_SZ * ITEMS_PER_DESC;
1356 * UV_ITEMS_PER_DESCRIPTOR, GFP_KERNEL, node); 1442 bau_desc = kmalloc_node(dsize, GFP_KERNEL, node);
1357 BUG_ON(!bau_desc); 1443 BUG_ON(!bau_desc);
1358 1444
1359 pa = uv_gpa(bau_desc); /* need the real nasid*/ 1445 pa = uv_gpa(bau_desc); /* need the real nasid*/
@@ -1361,27 +1447,25 @@ uv_activation_descriptor_init(int node, int pnode, int base_pnode)
1361 m = pa & uv_mmask; 1447 m = pa & uv_mmask;
1362 1448
1363 /* the 14-bit pnode */ 1449 /* the 14-bit pnode */
1364 uv_write_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE, 1450 write_mmr_descriptor_base(pnode, (n << UV_DESC_PSHIFT | m));
1365 (n << UV_DESC_BASE_PNODE_SHIFT | m));
1366 /* 1451 /*
1367 * Initializing all 8 (UV_ITEMS_PER_DESCRIPTOR) descriptors for each 1452 * Initializing all 8 (ITEMS_PER_DESC) descriptors for each
1368 * cpu even though we only use the first one; one descriptor can 1453 * cpu even though we only use the first one; one descriptor can
1369 * describe a broadcast to 256 uv hubs. 1454 * describe a broadcast to 256 uv hubs.
1370 */ 1455 */
1371 for (i = 0, bd2 = bau_desc; i < (UV_ADP_SIZE*UV_ITEMS_PER_DESCRIPTOR); 1456 for (i = 0, bd2 = bau_desc; i < (ADP_SZ * ITEMS_PER_DESC); i++, bd2++) {
1372 i++, bd2++) {
1373 memset(bd2, 0, sizeof(struct bau_desc)); 1457 memset(bd2, 0, sizeof(struct bau_desc));
1374 bd2->header.sw_ack_flag = 1; 1458 bd2->header.swack_flag = 1;
1375 /* 1459 /*
1376 * The base_dest_nasid set in the message header is the nasid 1460 * The base_dest_nasid set in the message header is the nasid
1377 * of the first uvhub in the partition. The bit map will 1461 * of the first uvhub in the partition. The bit map will
1378 * indicate destination pnode numbers relative to that base. 1462 * indicate destination pnode numbers relative to that base.
1379 * They may not be consecutive if nasid striding is being used. 1463 * They may not be consecutive if nasid striding is being used.
1380 */ 1464 */
1381 bd2->header.base_dest_nasid = UV_PNODE_TO_NASID(base_pnode); 1465 bd2->header.base_dest_nasid = UV_PNODE_TO_NASID(base_pnode);
1382 bd2->header.dest_subnodeid = UV_LB_SUBNODEID; 1466 bd2->header.dest_subnodeid = UV_LB_SUBNODEID;
1383 bd2->header.command = UV_NET_ENDPOINT_INTD; 1467 bd2->header.command = UV_NET_ENDPOINT_INTD;
1384 bd2->header.int_both = 1; 1468 bd2->header.int_both = 1;
1385 /* 1469 /*
1386 * all others need to be set to zero: 1470 * all others need to be set to zero:
1387 * fairness chaining multilevel count replied_to 1471 * fairness chaining multilevel count replied_to
@@ -1401,57 +1485,55 @@ uv_activation_descriptor_init(int node, int pnode, int base_pnode)
1401 * - node is first node (kernel memory notion) on the uvhub 1485 * - node is first node (kernel memory notion) on the uvhub
1402 * - pnode is the uvhub's physical identifier 1486 * - pnode is the uvhub's physical identifier
1403 */ 1487 */
1404static void 1488static void pq_init(int node, int pnode)
1405uv_payload_queue_init(int node, int pnode)
1406{ 1489{
1407 int pn;
1408 int cpu; 1490 int cpu;
1491 size_t plsize;
1409 char *cp; 1492 char *cp;
1410 unsigned long pa; 1493 void *vp;
1411 struct bau_payload_queue_entry *pqp; 1494 unsigned long pn;
1412 struct bau_payload_queue_entry *pqp_malloc; 1495 unsigned long first;
1496 unsigned long pn_first;
1497 unsigned long last;
1498 struct bau_pq_entry *pqp;
1413 struct bau_control *bcp; 1499 struct bau_control *bcp;
1414 1500
1415 pqp = kmalloc_node((DEST_Q_SIZE + 1) 1501 plsize = (DEST_Q_SIZE + 1) * sizeof(struct bau_pq_entry);
1416 * sizeof(struct bau_payload_queue_entry), 1502 vp = kmalloc_node(plsize, GFP_KERNEL, node);
1417 GFP_KERNEL, node); 1503 pqp = (struct bau_pq_entry *)vp;
1418 BUG_ON(!pqp); 1504 BUG_ON(!pqp);
1419 pqp_malloc = pqp;
1420 1505
1421 cp = (char *)pqp + 31; 1506 cp = (char *)pqp + 31;
1422 pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5); 1507 pqp = (struct bau_pq_entry *)(((unsigned long)cp >> 5) << 5);
1423 1508
1424 for_each_present_cpu(cpu) { 1509 for_each_present_cpu(cpu) {
1425 if (pnode != uv_cpu_to_pnode(cpu)) 1510 if (pnode != uv_cpu_to_pnode(cpu))
1426 continue; 1511 continue;
1427 /* for every cpu on this pnode: */ 1512 /* for every cpu on this pnode: */
1428 bcp = &per_cpu(bau_control, cpu); 1513 bcp = &per_cpu(bau_control, cpu);
1429 bcp->va_queue_first = pqp; 1514 bcp->queue_first = pqp;
1430 bcp->bau_msg_head = pqp; 1515 bcp->bau_msg_head = pqp;
1431 bcp->va_queue_last = pqp + (DEST_Q_SIZE - 1); 1516 bcp->queue_last = pqp + (DEST_Q_SIZE - 1);
1432 } 1517 }
1433 /* 1518 /*
1434 * need the pnode of where the memory was really allocated 1519 * need the pnode of where the memory was really allocated
1435 */ 1520 */
1436 pa = uv_gpa(pqp); 1521 pn = uv_gpa(pqp) >> uv_nshift;
1437 pn = pa >> uv_nshift; 1522 first = uv_physnodeaddr(pqp);
1438 uv_write_global_mmr64(pnode, 1523 pn_first = ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | first;
1439 UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST, 1524 last = uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1));
1440 ((unsigned long)pn << UV_PAYLOADQ_PNODE_SHIFT) | 1525 write_mmr_payload_first(pnode, pn_first);
1441 uv_physnodeaddr(pqp)); 1526 write_mmr_payload_tail(pnode, first);
1442 uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL, 1527 write_mmr_payload_last(pnode, last);
1443 uv_physnodeaddr(pqp)); 1528
1444 uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST,
1445 (unsigned long)
1446 uv_physnodeaddr(pqp + (DEST_Q_SIZE - 1)));
1447 /* in effect, all msg_type's are set to MSG_NOOP */ 1529 /* in effect, all msg_type's are set to MSG_NOOP */
1448 memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE); 1530 memset(pqp, 0, sizeof(struct bau_pq_entry) * DEST_Q_SIZE);
1449} 1531}
1450 1532
1451/* 1533/*
1452 * Initialization of each UV hub's structures 1534 * Initialization of each UV hub's structures
1453 */ 1535 */
1454static void __init uv_init_uvhub(int uvhub, int vector, int base_pnode) 1536static void __init init_uvhub(int uvhub, int vector, int base_pnode)
1455{ 1537{
1456 int node; 1538 int node;
1457 int pnode; 1539 int pnode;
@@ -1459,24 +1541,24 @@ static void __init uv_init_uvhub(int uvhub, int vector, int base_pnode)
1459 1541
1460 node = uvhub_to_first_node(uvhub); 1542 node = uvhub_to_first_node(uvhub);
1461 pnode = uv_blade_to_pnode(uvhub); 1543 pnode = uv_blade_to_pnode(uvhub);
1462 uv_activation_descriptor_init(node, pnode, base_pnode); 1544
1463 uv_payload_queue_init(node, pnode); 1545 activation_descriptor_init(node, pnode, base_pnode);
1546
1547 pq_init(node, pnode);
1464 /* 1548 /*
1465 * The below initialization can't be in firmware because the 1549 * The below initialization can't be in firmware because the
1466 * messaging IRQ will be determined by the OS. 1550 * messaging IRQ will be determined by the OS.
1467 */ 1551 */
1468 apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits; 1552 apicid = uvhub_to_first_apicid(uvhub) | uv_apicid_hibits;
1469 uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG, 1553 write_mmr_data_config(pnode, ((apicid << 32) | vector));
1470 ((apicid << 32) | vector));
1471} 1554}
1472 1555
1473/* 1556/*
1474 * We will set BAU_MISC_CONTROL with a timeout period. 1557 * We will set BAU_MISC_CONTROL with a timeout period.
1475 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT. 1558 * But the BIOS has set UVH_AGING_PRESCALE_SEL and UVH_TRANSACTION_TIMEOUT.
1476 * So the destination timeout period has be be calculated from them. 1559 * So the destination timeout period has to be calculated from them.
1477 */ 1560 */
1478static int 1561static int calculate_destination_timeout(void)
1479calculate_destination_timeout(void)
1480{ 1562{
1481 unsigned long mmr_image; 1563 unsigned long mmr_image;
1482 int mult1; 1564 int mult1;
@@ -1486,73 +1568,92 @@ calculate_destination_timeout(void)
1486 int ret; 1568 int ret;
1487 unsigned long ts_ns; 1569 unsigned long ts_ns;
1488 1570
1489 mult1 = UV_INTD_SOFT_ACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK; 1571 if (is_uv1_hub()) {
1490 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL); 1572 mult1 = SOFTACK_TIMEOUT_PERIOD & BAU_MISC_CONTROL_MULT_MASK;
1491 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK; 1573 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1492 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT); 1574 index = (mmr_image >> BAU_URGENCY_7_SHIFT) & BAU_URGENCY_7_MASK;
1493 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK; 1575 mmr_image = uv_read_local_mmr(UVH_TRANSACTION_TIMEOUT);
1494 base = timeout_base_ns[index]; 1576 mult2 = (mmr_image >> BAU_TRANS_SHIFT) & BAU_TRANS_MASK;
1495 ts_ns = base * mult1 * mult2; 1577 base = timeout_base_ns[index];
1496 ret = ts_ns / 1000; 1578 ts_ns = base * mult1 * mult2;
1579 ret = ts_ns / 1000;
1580 } else {
1581 /* 4 bits 0/1 for 10/80us, 3 bits of multiplier */
1582 mmr_image = uv_read_local_mmr(UVH_AGING_PRESCALE_SEL);
1583 mmr_image = (mmr_image & UV_SA_MASK) >> UV_SA_SHFT;
1584 if (mmr_image & (1L << UV2_ACK_UNITS_SHFT))
1585 mult1 = 80;
1586 else
1587 mult1 = 10;
1588 base = mmr_image & UV2_ACK_MASK;
1589 ret = mult1 * base;
1590 }
1497 return ret; 1591 return ret;
1498} 1592}
1499 1593
1594static void __init init_per_cpu_tunables(void)
1595{
1596 int cpu;
1597 struct bau_control *bcp;
1598
1599 for_each_present_cpu(cpu) {
1600 bcp = &per_cpu(bau_control, cpu);
1601 bcp->baudisabled = 0;
1602 bcp->statp = &per_cpu(ptcstats, cpu);
1603 /* time interval to catch a hardware stay-busy bug */
1604 bcp->timeout_interval = usec_2_cycles(2*timeout_us);
1605 bcp->max_concurr = max_concurr;
1606 bcp->max_concurr_const = max_concurr;
1607 bcp->plugged_delay = plugged_delay;
1608 bcp->plugsb4reset = plugsb4reset;
1609 bcp->timeoutsb4reset = timeoutsb4reset;
1610 bcp->ipi_reset_limit = ipi_reset_limit;
1611 bcp->complete_threshold = complete_threshold;
1612 bcp->cong_response_us = congested_respns_us;
1613 bcp->cong_reps = congested_reps;
1614 bcp->cong_period = congested_period;
1615 }
1616}
1617
1500/* 1618/*
1501 * initialize the bau_control structure for each cpu 1619 * Scan all cpus to collect blade and socket summaries.
1502 */ 1620 */
1503static int __init uv_init_per_cpu(int nuvhubs, int base_part_pnode) 1621static int __init get_cpu_topology(int base_pnode,
1622 struct uvhub_desc *uvhub_descs,
1623 unsigned char *uvhub_mask)
1504{ 1624{
1505 int i;
1506 int cpu; 1625 int cpu;
1507 int tcpu;
1508 int pnode; 1626 int pnode;
1509 int uvhub; 1627 int uvhub;
1510 int have_hmaster; 1628 int socket;
1511 short socket = 0;
1512 unsigned short socket_mask;
1513 unsigned char *uvhub_mask;
1514 struct bau_control *bcp; 1629 struct bau_control *bcp;
1515 struct uvhub_desc *bdp; 1630 struct uvhub_desc *bdp;
1516 struct socket_desc *sdp; 1631 struct socket_desc *sdp;
1517 struct bau_control *hmaster = NULL;
1518 struct bau_control *smaster = NULL;
1519 struct socket_desc {
1520 short num_cpus;
1521 short cpu_number[MAX_CPUS_PER_SOCKET];
1522 };
1523 struct uvhub_desc {
1524 unsigned short socket_mask;
1525 short num_cpus;
1526 short uvhub;
1527 short pnode;
1528 struct socket_desc socket[2];
1529 };
1530 struct uvhub_desc *uvhub_descs;
1531
1532 timeout_us = calculate_destination_timeout();
1533 1632
1534 uvhub_descs = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1535 memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1536 uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
1537 for_each_present_cpu(cpu) { 1633 for_each_present_cpu(cpu) {
1538 bcp = &per_cpu(bau_control, cpu); 1634 bcp = &per_cpu(bau_control, cpu);
1635
1539 memset(bcp, 0, sizeof(struct bau_control)); 1636 memset(bcp, 0, sizeof(struct bau_control));
1637
1540 pnode = uv_cpu_hub_info(cpu)->pnode; 1638 pnode = uv_cpu_hub_info(cpu)->pnode;
1541 if ((pnode - base_part_pnode) >= UV_DISTRIBUTION_SIZE) { 1639 if ((pnode - base_pnode) >= UV_DISTRIBUTION_SIZE) {
1542 printk(KERN_EMERG 1640 printk(KERN_EMERG
1543 "cpu %d pnode %d-%d beyond %d; BAU disabled\n", 1641 "cpu %d pnode %d-%d beyond %d; BAU disabled\n",
1544 cpu, pnode, base_part_pnode, 1642 cpu, pnode, base_pnode, UV_DISTRIBUTION_SIZE);
1545 UV_DISTRIBUTION_SIZE);
1546 return 1; 1643 return 1;
1547 } 1644 }
1645
1548 bcp->osnode = cpu_to_node(cpu); 1646 bcp->osnode = cpu_to_node(cpu);
1549 bcp->partition_base_pnode = uv_partition_base_pnode; 1647 bcp->partition_base_pnode = base_pnode;
1648
1550 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id; 1649 uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1551 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8)); 1650 *(uvhub_mask + (uvhub/8)) |= (1 << (uvhub%8));
1552 bdp = &uvhub_descs[uvhub]; 1651 bdp = &uvhub_descs[uvhub];
1652
1553 bdp->num_cpus++; 1653 bdp->num_cpus++;
1554 bdp->uvhub = uvhub; 1654 bdp->uvhub = uvhub;
1555 bdp->pnode = pnode; 1655 bdp->pnode = pnode;
1656
1556 /* kludge: 'assuming' one node per socket, and assuming that 1657 /* kludge: 'assuming' one node per socket, and assuming that
1557 disabling a socket just leaves a gap in node numbers */ 1658 disabling a socket just leaves a gap in node numbers */
1558 socket = bcp->osnode & 1; 1659 socket = bcp->osnode & 1;
@@ -1561,84 +1662,129 @@ static int __init uv_init_per_cpu(int nuvhubs, int base_part_pnode)
1561 sdp->cpu_number[sdp->num_cpus] = cpu; 1662 sdp->cpu_number[sdp->num_cpus] = cpu;
1562 sdp->num_cpus++; 1663 sdp->num_cpus++;
1563 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) { 1664 if (sdp->num_cpus > MAX_CPUS_PER_SOCKET) {
1564 printk(KERN_EMERG "%d cpus per socket invalid\n", sdp->num_cpus); 1665 printk(KERN_EMERG "%d cpus per socket invalid\n",
1666 sdp->num_cpus);
1565 return 1; 1667 return 1;
1566 } 1668 }
1567 } 1669 }
1670 return 0;
1671}
1672
1673/*
1674 * Each socket is to get a local array of pnodes/hubs.
1675 */
1676static void make_per_cpu_thp(struct bau_control *smaster)
1677{
1678 int cpu;
1679 size_t hpsz = sizeof(struct hub_and_pnode) * num_possible_cpus();
1680
1681 smaster->thp = kmalloc_node(hpsz, GFP_KERNEL, smaster->osnode);
1682 memset(smaster->thp, 0, hpsz);
1683 for_each_present_cpu(cpu) {
1684 smaster->thp[cpu].pnode = uv_cpu_hub_info(cpu)->pnode;
1685 smaster->thp[cpu].uvhub = uv_cpu_hub_info(cpu)->numa_blade_id;
1686 }
1687}
1688
1689/*
1690 * Initialize all the per_cpu information for the cpu's on a given socket,
1691 * given what has been gathered into the socket_desc struct.
1692 * And reports the chosen hub and socket masters back to the caller.
1693 */
1694static int scan_sock(struct socket_desc *sdp, struct uvhub_desc *bdp,
1695 struct bau_control **smasterp,
1696 struct bau_control **hmasterp)
1697{
1698 int i;
1699 int cpu;
1700 struct bau_control *bcp;
1701
1702 for (i = 0; i < sdp->num_cpus; i++) {
1703 cpu = sdp->cpu_number[i];
1704 bcp = &per_cpu(bau_control, cpu);
1705 bcp->cpu = cpu;
1706 if (i == 0) {
1707 *smasterp = bcp;
1708 if (!(*hmasterp))
1709 *hmasterp = bcp;
1710 }
1711 bcp->cpus_in_uvhub = bdp->num_cpus;
1712 bcp->cpus_in_socket = sdp->num_cpus;
1713 bcp->socket_master = *smasterp;
1714 bcp->uvhub = bdp->uvhub;
1715 bcp->uvhub_master = *hmasterp;
1716 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->blade_processor_id;
1717 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
1718 printk(KERN_EMERG "%d cpus per uvhub invalid\n",
1719 bcp->uvhub_cpu);
1720 return 1;
1721 }
1722 }
1723 return 0;
1724}
1725
1726/*
1727 * Summarize the blade and socket topology into the per_cpu structures.
1728 */
1729static int __init summarize_uvhub_sockets(int nuvhubs,
1730 struct uvhub_desc *uvhub_descs,
1731 unsigned char *uvhub_mask)
1732{
1733 int socket;
1734 int uvhub;
1735 unsigned short socket_mask;
1736
1568 for (uvhub = 0; uvhub < nuvhubs; uvhub++) { 1737 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1738 struct uvhub_desc *bdp;
1739 struct bau_control *smaster = NULL;
1740 struct bau_control *hmaster = NULL;
1741
1569 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8)))) 1742 if (!(*(uvhub_mask + (uvhub/8)) & (1 << (uvhub%8))))
1570 continue; 1743 continue;
1571 have_hmaster = 0; 1744
1572 bdp = &uvhub_descs[uvhub]; 1745 bdp = &uvhub_descs[uvhub];
1573 socket_mask = bdp->socket_mask; 1746 socket_mask = bdp->socket_mask;
1574 socket = 0; 1747 socket = 0;
1575 while (socket_mask) { 1748 while (socket_mask) {
1576 if (!(socket_mask & 1)) 1749 struct socket_desc *sdp;
1577 goto nextsocket; 1750 if ((socket_mask & 1)) {
1578 sdp = &bdp->socket[socket]; 1751 sdp = &bdp->socket[socket];
1579 for (i = 0; i < sdp->num_cpus; i++) { 1752 if (scan_sock(sdp, bdp, &smaster, &hmaster))
1580 cpu = sdp->cpu_number[i];
1581 bcp = &per_cpu(bau_control, cpu);
1582 bcp->cpu = cpu;
1583 if (i == 0) {
1584 smaster = bcp;
1585 if (!have_hmaster) {
1586 have_hmaster++;
1587 hmaster = bcp;
1588 }
1589 }
1590 bcp->cpus_in_uvhub = bdp->num_cpus;
1591 bcp->cpus_in_socket = sdp->num_cpus;
1592 bcp->socket_master = smaster;
1593 bcp->uvhub = bdp->uvhub;
1594 bcp->uvhub_master = hmaster;
1595 bcp->uvhub_cpu = uv_cpu_hub_info(cpu)->
1596 blade_processor_id;
1597 if (bcp->uvhub_cpu >= MAX_CPUS_PER_UVHUB) {
1598 printk(KERN_EMERG
1599 "%d cpus per uvhub invalid\n",
1600 bcp->uvhub_cpu);
1601 return 1; 1753 return 1;
1602 }
1603 } 1754 }
1604nextsocket:
1605 socket++; 1755 socket++;
1606 socket_mask = (socket_mask >> 1); 1756 socket_mask = (socket_mask >> 1);
1607 /* each socket gets a local array of pnodes/hubs */ 1757 make_per_cpu_thp(smaster);
1608 bcp = smaster;
1609 bcp->target_hub_and_pnode = kmalloc_node(
1610 sizeof(struct hub_and_pnode) *
1611 num_possible_cpus(), GFP_KERNEL, bcp->osnode);
1612 memset(bcp->target_hub_and_pnode, 0,
1613 sizeof(struct hub_and_pnode) *
1614 num_possible_cpus());
1615 for_each_present_cpu(tcpu) {
1616 bcp->target_hub_and_pnode[tcpu].pnode =
1617 uv_cpu_hub_info(tcpu)->pnode;
1618 bcp->target_hub_and_pnode[tcpu].uvhub =
1619 uv_cpu_hub_info(tcpu)->numa_blade_id;
1620 }
1621 } 1758 }
1622 } 1759 }
1760 return 0;
1761}
1762
1763/*
1764 * initialize the bau_control structure for each cpu
1765 */
1766static int __init init_per_cpu(int nuvhubs, int base_part_pnode)
1767{
1768 unsigned char *uvhub_mask;
1769 void *vp;
1770 struct uvhub_desc *uvhub_descs;
1771
1772 timeout_us = calculate_destination_timeout();
1773
1774 vp = kmalloc(nuvhubs * sizeof(struct uvhub_desc), GFP_KERNEL);
1775 uvhub_descs = (struct uvhub_desc *)vp;
1776 memset(uvhub_descs, 0, nuvhubs * sizeof(struct uvhub_desc));
1777 uvhub_mask = kzalloc((nuvhubs+7)/8, GFP_KERNEL);
1778
1779 if (get_cpu_topology(base_part_pnode, uvhub_descs, uvhub_mask))
1780 return 1;
1781
1782 if (summarize_uvhub_sockets(nuvhubs, uvhub_descs, uvhub_mask))
1783 return 1;
1784
1623 kfree(uvhub_descs); 1785 kfree(uvhub_descs);
1624 kfree(uvhub_mask); 1786 kfree(uvhub_mask);
1625 for_each_present_cpu(cpu) { 1787 init_per_cpu_tunables();
1626 bcp = &per_cpu(bau_control, cpu);
1627 bcp->baudisabled = 0;
1628 bcp->statp = &per_cpu(ptcstats, cpu);
1629 /* time interval to catch a hardware stay-busy bug */
1630 bcp->timeout_interval = microsec_2_cycles(2*timeout_us);
1631 bcp->max_bau_concurrent = max_bau_concurrent;
1632 bcp->max_bau_concurrent_constant = max_bau_concurrent;
1633 bcp->plugged_delay = plugged_delay;
1634 bcp->plugsb4reset = plugsb4reset;
1635 bcp->timeoutsb4reset = timeoutsb4reset;
1636 bcp->ipi_reset_limit = ipi_reset_limit;
1637 bcp->complete_threshold = complete_threshold;
1638 bcp->congested_response_us = congested_response_us;
1639 bcp->congested_reps = congested_reps;
1640 bcp->congested_period = congested_period;
1641 }
1642 return 0; 1788 return 0;
1643} 1789}
1644 1790
@@ -1651,8 +1797,9 @@ static int __init uv_bau_init(void)
1651 int pnode; 1797 int pnode;
1652 int nuvhubs; 1798 int nuvhubs;
1653 int cur_cpu; 1799 int cur_cpu;
1800 int cpus;
1654 int vector; 1801 int vector;
1655 unsigned long mmr; 1802 cpumask_var_t *mask;
1656 1803
1657 if (!is_uv_system()) 1804 if (!is_uv_system())
1658 return 0; 1805 return 0;
@@ -1660,24 +1807,25 @@ static int __init uv_bau_init(void)
1660 if (nobau) 1807 if (nobau)
1661 return 0; 1808 return 0;
1662 1809
1663 for_each_possible_cpu(cur_cpu) 1810 for_each_possible_cpu(cur_cpu) {
1664 zalloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu), 1811 mask = &per_cpu(uv_flush_tlb_mask, cur_cpu);
1665 GFP_KERNEL, cpu_to_node(cur_cpu)); 1812 zalloc_cpumask_var_node(mask, GFP_KERNEL, cpu_to_node(cur_cpu));
1813 }
1666 1814
1667 uv_nshift = uv_hub_info->m_val; 1815 uv_nshift = uv_hub_info->m_val;
1668 uv_mmask = (1UL << uv_hub_info->m_val) - 1; 1816 uv_mmask = (1UL << uv_hub_info->m_val) - 1;
1669 nuvhubs = uv_num_possible_blades(); 1817 nuvhubs = uv_num_possible_blades();
1670 spin_lock_init(&disable_lock); 1818 spin_lock_init(&disable_lock);
1671 congested_cycles = microsec_2_cycles(congested_response_us); 1819 congested_cycles = usec_2_cycles(congested_respns_us);
1672 1820
1673 uv_partition_base_pnode = 0x7fffffff; 1821 uv_base_pnode = 0x7fffffff;
1674 for (uvhub = 0; uvhub < nuvhubs; uvhub++) { 1822 for (uvhub = 0; uvhub < nuvhubs; uvhub++) {
1675 if (uv_blade_nr_possible_cpus(uvhub) && 1823 cpus = uv_blade_nr_possible_cpus(uvhub);
1676 (uv_blade_to_pnode(uvhub) < uv_partition_base_pnode)) 1824 if (cpus && (uv_blade_to_pnode(uvhub) < uv_base_pnode))
1677 uv_partition_base_pnode = uv_blade_to_pnode(uvhub); 1825 uv_base_pnode = uv_blade_to_pnode(uvhub);
1678 } 1826 }
1679 1827
1680 if (uv_init_per_cpu(nuvhubs, uv_partition_base_pnode)) { 1828 if (init_per_cpu(nuvhubs, uv_base_pnode)) {
1681 nobau = 1; 1829 nobau = 1;
1682 return 0; 1830 return 0;
1683 } 1831 }
@@ -1685,21 +1833,21 @@ static int __init uv_bau_init(void)
1685 vector = UV_BAU_MESSAGE; 1833 vector = UV_BAU_MESSAGE;
1686 for_each_possible_blade(uvhub) 1834 for_each_possible_blade(uvhub)
1687 if (uv_blade_nr_possible_cpus(uvhub)) 1835 if (uv_blade_nr_possible_cpus(uvhub))
1688 uv_init_uvhub(uvhub, vector, uv_partition_base_pnode); 1836 init_uvhub(uvhub, vector, uv_base_pnode);
1689 1837
1690 uv_enable_timeouts(); 1838 enable_timeouts();
1691 alloc_intr_gate(vector, uv_bau_message_intr1); 1839 alloc_intr_gate(vector, uv_bau_message_intr1);
1692 1840
1693 for_each_possible_blade(uvhub) { 1841 for_each_possible_blade(uvhub) {
1694 if (uv_blade_nr_possible_cpus(uvhub)) { 1842 if (uv_blade_nr_possible_cpus(uvhub)) {
1843 unsigned long val;
1844 unsigned long mmr;
1695 pnode = uv_blade_to_pnode(uvhub); 1845 pnode = uv_blade_to_pnode(uvhub);
1696 /* INIT the bau */ 1846 /* INIT the bau */
1697 uv_write_global_mmr64(pnode, 1847 val = 1L << 63;
1698 UVH_LB_BAU_SB_ACTIVATION_CONTROL, 1848 write_gmmr_activation(pnode, val);
1699 ((unsigned long)1 << 63));
1700 mmr = 1; /* should be 1 to broadcast to both sockets */ 1849 mmr = 1; /* should be 1 to broadcast to both sockets */
1701 uv_write_global_mmr64(pnode, UVH_BAU_DATA_BROADCAST, 1850 write_mmr_data_broadcast(pnode, mmr);
1702 mmr);
1703 } 1851 }
1704 } 1852 }
1705 1853
generated by cgit v1.2.2 (git 2.25.0) at 2026-01-07 02:36:52 -0500