Commit | Line | Data |
---|---|---|
1812924b CW |
1 | /* |
2 | * SGI UltraViolet TLB flush routines. | |
3 | * | |
4 | * (c) 2008 Cliff Wickman <cpw@sgi.com>, SGI. | |
5 | * | |
6 | * This code is released under the GNU General Public License version 2 or | |
7 | * later. | |
8 | */ | |
aef8f5b8 | 9 | #include <linux/seq_file.h> |
1812924b CW |
10 | #include <linux/proc_fs.h> |
11 | #include <linux/kernel.h> | |
12 | ||
1812924b | 13 | #include <asm/mmu_context.h> |
bdbcdd48 | 14 | #include <asm/uv/uv.h> |
1812924b | 15 | #include <asm/uv/uv_mmrs.h> |
b4c286e6 | 16 | #include <asm/uv/uv_hub.h> |
1812924b | 17 | #include <asm/uv/uv_bau.h> |
7b6aa335 | 18 | #include <asm/apic.h> |
b4c286e6 | 19 | #include <asm/idle.h> |
b194b120 | 20 | #include <asm/tsc.h> |
99dd8713 | 21 | #include <asm/irq_vectors.h> |
1812924b | 22 | |
b4c286e6 IM |
23 | static struct bau_control **uv_bau_table_bases __read_mostly; |
24 | static int uv_bau_retry_limit __read_mostly; | |
25 | ||
26 | /* position of pnode (which is nasid>>1): */ | |
27 | static int uv_nshift __read_mostly; | |
28 | ||
29 | static unsigned long uv_mmask __read_mostly; | |
1812924b | 30 | |
dc163a41 IM |
31 | static DEFINE_PER_CPU(struct ptc_stats, ptcstats); |
32 | static DEFINE_PER_CPU(struct bau_control, bau_control); | |
1812924b | 33 | |
9674f35b CW |
34 | /* |
35 | * Determine the first node on a blade. | |
36 | */ | |
37 | static int __init blade_to_first_node(int blade) | |
38 | { | |
39 | int node, b; | |
40 | ||
41 | for_each_online_node(node) { | |
42 | b = uv_node_to_blade_id(node); | |
43 | if (blade == b) | |
44 | return node; | |
45 | } | |
46 | BUG(); | |
47 | } | |
48 | ||
49 | /* | |
50 | * Determine the apicid of the first cpu on a blade. | |
51 | */ | |
52 | static int __init blade_to_first_apicid(int blade) | |
53 | { | |
54 | int cpu; | |
55 | ||
56 | for_each_present_cpu(cpu) | |
57 | if (blade == uv_cpu_to_blade_id(cpu)) | |
58 | return per_cpu(x86_cpu_to_apicid, cpu); | |
59 | return -1; | |
60 | } | |
61 | ||
1812924b CW |
62 | /* |
63 | * Free a software acknowledge hardware resource by clearing its Pending | |
64 | * bit. This will return a reply to the sender. | |
65 | * If the message has timed out, a reply has already been sent by the | |
66 | * hardware but the resource has not been released. In that case our | |
67 | * clear of the Timeout bit (as well) will free the resource. No reply will | |
68 | * be sent (the hardware will only do one reply per message). | |
69 | */ | |
b194b120 | 70 | static void uv_reply_to_message(int resource, |
b4c286e6 IM |
71 | struct bau_payload_queue_entry *msg, |
72 | struct bau_msg_status *msp) | |
1812924b | 73 | { |
b194b120 | 74 | unsigned long dw; |
1812924b | 75 | |
b194b120 | 76 | dw = (1 << (resource + UV_SW_ACK_NPENDING)) | (1 << resource); |
1812924b CW |
77 | msg->replied_to = 1; |
78 | msg->sw_ack_vector = 0; | |
79 | if (msp) | |
80 | msp->seen_by.bits = 0; | |
b194b120 | 81 | uv_write_local_mmr(UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE_ALIAS, dw); |
1812924b CW |
82 | } |
83 | ||
84 | /* | |
85 | * Do all the things a cpu should do for a TLB shootdown message. | |
86 | * Other cpu's may come here at the same time for this message. | |
87 | */ | |
b194b120 | 88 | static void uv_bau_process_message(struct bau_payload_queue_entry *msg, |
b4c286e6 | 89 | int msg_slot, int sw_ack_slot) |
1812924b | 90 | { |
1812924b CW |
91 | unsigned long this_cpu_mask; |
92 | struct bau_msg_status *msp; | |
b4c286e6 | 93 | int cpu; |
1812924b CW |
94 | |
95 | msp = __get_cpu_var(bau_control).msg_statuses + msg_slot; | |
96 | cpu = uv_blade_processor_id(); | |
97 | msg->number_of_cpus = | |
9674f35b | 98 | uv_blade_nr_online_cpus(uv_node_to_blade_id(numa_node_id())); |
dc163a41 | 99 | this_cpu_mask = 1UL << cpu; |
1812924b CW |
100 | if (msp->seen_by.bits & this_cpu_mask) |
101 | return; | |
102 | atomic_or_long(&msp->seen_by.bits, this_cpu_mask); | |
103 | ||
104 | if (msg->replied_to == 1) | |
105 | return; | |
106 | ||
107 | if (msg->address == TLB_FLUSH_ALL) { | |
108 | local_flush_tlb(); | |
109 | __get_cpu_var(ptcstats).alltlb++; | |
110 | } else { | |
111 | __flush_tlb_one(msg->address); | |
112 | __get_cpu_var(ptcstats).onetlb++; | |
113 | } | |
114 | ||
115 | __get_cpu_var(ptcstats).requestee++; | |
116 | ||
117 | atomic_inc_short(&msg->acknowledge_count); | |
118 | if (msg->number_of_cpus == msg->acknowledge_count) | |
119 | uv_reply_to_message(sw_ack_slot, msg, msp); | |
1812924b CW |
120 | } |
121 | ||
122 | /* | |
dc163a41 | 123 | * Examine the payload queue on one distribution node to see |
1812924b CW |
124 | * which messages have not been seen, and which cpu(s) have not seen them. |
125 | * | |
126 | * Returns the number of cpu's that have not responded. | |
127 | */ | |
dc163a41 | 128 | static int uv_examine_destination(struct bau_control *bau_tablesp, int sender) |
1812924b | 129 | { |
1812924b CW |
130 | struct bau_payload_queue_entry *msg; |
131 | struct bau_msg_status *msp; | |
b4c286e6 IM |
132 | int count = 0; |
133 | int i; | |
134 | int j; | |
1812924b | 135 | |
dc163a41 IM |
136 | for (msg = bau_tablesp->va_queue_first, i = 0; i < DEST_Q_SIZE; |
137 | msg++, i++) { | |
138 | if ((msg->sending_cpu == sender) && (!msg->replied_to)) { | |
139 | msp = bau_tablesp->msg_statuses + i; | |
140 | printk(KERN_DEBUG | |
141 | "blade %d: address:%#lx %d of %d, not cpu(s): ", | |
142 | i, msg->address, msg->acknowledge_count, | |
143 | msg->number_of_cpus); | |
144 | for (j = 0; j < msg->number_of_cpus; j++) { | |
b4c286e6 | 145 | if (!((1L << j) & msp->seen_by.bits)) { |
dc163a41 IM |
146 | count++; |
147 | printk("%d ", j); | |
148 | } | |
149 | } | |
150 | printk("\n"); | |
151 | } | |
152 | } | |
153 | return count; | |
154 | } | |
155 | ||
156 | /* | |
157 | * Examine the payload queue on all the distribution nodes to see | |
158 | * which messages have not been seen, and which cpu(s) have not seen them. | |
159 | * | |
160 | * Returns the number of cpu's that have not responded. | |
161 | */ | |
162 | static int uv_examine_destinations(struct bau_target_nodemask *distribution) | |
163 | { | |
164 | int sender; | |
165 | int i; | |
166 | int count = 0; | |
167 | ||
1812924b | 168 | sender = smp_processor_id(); |
b4c286e6 | 169 | for (i = 0; i < sizeof(struct bau_target_nodemask) * BITSPERBYTE; i++) { |
b194b120 CW |
170 | if (!bau_node_isset(i, distribution)) |
171 | continue; | |
dc163a41 | 172 | count += uv_examine_destination(uv_bau_table_bases[i], sender); |
1812924b CW |
173 | } |
174 | return count; | |
175 | } | |
176 | ||
b194b120 CW |
177 | /* |
178 | * wait for completion of a broadcast message | |
179 | * | |
180 | * return COMPLETE, RETRY or GIVEUP | |
181 | */ | |
dc163a41 | 182 | static int uv_wait_completion(struct bau_desc *bau_desc, |
b194b120 CW |
183 | unsigned long mmr_offset, int right_shift) |
184 | { | |
185 | int exams = 0; | |
186 | long destination_timeouts = 0; | |
187 | long source_timeouts = 0; | |
188 | unsigned long descriptor_status; | |
189 | ||
190 | while ((descriptor_status = (((unsigned long) | |
191 | uv_read_local_mmr(mmr_offset) >> | |
192 | right_shift) & UV_ACT_STATUS_MASK)) != | |
193 | DESC_STATUS_IDLE) { | |
194 | if (descriptor_status == DESC_STATUS_SOURCE_TIMEOUT) { | |
195 | source_timeouts++; | |
196 | if (source_timeouts > SOURCE_TIMEOUT_LIMIT) | |
197 | source_timeouts = 0; | |
198 | __get_cpu_var(ptcstats).s_retry++; | |
199 | return FLUSH_RETRY; | |
200 | } | |
201 | /* | |
202 | * spin here looking for progress at the destinations | |
203 | */ | |
204 | if (descriptor_status == DESC_STATUS_DESTINATION_TIMEOUT) { | |
205 | destination_timeouts++; | |
206 | if (destination_timeouts > DESTINATION_TIMEOUT_LIMIT) { | |
207 | /* | |
208 | * returns number of cpus not responding | |
209 | */ | |
210 | if (uv_examine_destinations | |
211 | (&bau_desc->distribution) == 0) { | |
212 | __get_cpu_var(ptcstats).d_retry++; | |
213 | return FLUSH_RETRY; | |
214 | } | |
215 | exams++; | |
216 | if (exams >= uv_bau_retry_limit) { | |
217 | printk(KERN_DEBUG | |
218 | "uv_flush_tlb_others"); | |
219 | printk("giving up on cpu %d\n", | |
220 | smp_processor_id()); | |
221 | return FLUSH_GIVEUP; | |
222 | } | |
223 | /* | |
224 | * delays can hang the simulator | |
225 | udelay(1000); | |
226 | */ | |
227 | destination_timeouts = 0; | |
228 | } | |
229 | } | |
18c07cf5 | 230 | cpu_relax(); |
b194b120 CW |
231 | } |
232 | return FLUSH_COMPLETE; | |
233 | } | |
234 | ||
235 | /** | |
236 | * uv_flush_send_and_wait | |
237 | * | |
238 | * Send a broadcast and wait for a broadcast message to complete. | |
239 | * | |
bdbcdd48 | 240 | * The flush_mask contains the cpus the broadcast was sent to. |
b194b120 | 241 | * |
bdbcdd48 TH |
242 | * Returns NULL if all remote flushing was done. The mask is zeroed. |
243 | * Returns @flush_mask if some remote flushing remains to be done. The | |
244 | * mask will have some bits still set. | |
b194b120 | 245 | */ |
9674f35b | 246 | const struct cpumask *uv_flush_send_and_wait(int cpu, int this_pnode, |
bdbcdd48 TH |
247 | struct bau_desc *bau_desc, |
248 | struct cpumask *flush_mask) | |
b194b120 CW |
249 | { |
250 | int completion_status = 0; | |
251 | int right_shift; | |
b194b120 | 252 | int tries = 0; |
9674f35b | 253 | int pnode; |
b4c286e6 | 254 | int bit; |
b194b120 | 255 | unsigned long mmr_offset; |
b4c286e6 | 256 | unsigned long index; |
b194b120 CW |
257 | cycles_t time1; |
258 | cycles_t time2; | |
259 | ||
260 | if (cpu < UV_CPUS_PER_ACT_STATUS) { | |
261 | mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_0; | |
262 | right_shift = cpu * UV_ACT_STATUS_SIZE; | |
263 | } else { | |
264 | mmr_offset = UVH_LB_BAU_SB_ACTIVATION_STATUS_1; | |
265 | right_shift = | |
266 | ((cpu - UV_CPUS_PER_ACT_STATUS) * UV_ACT_STATUS_SIZE); | |
267 | } | |
268 | time1 = get_cycles(); | |
269 | do { | |
270 | tries++; | |
dc163a41 IM |
271 | index = (1UL << UVH_LB_BAU_SB_ACTIVATION_CONTROL_PUSH_SHFT) | |
272 | cpu; | |
b194b120 CW |
273 | uv_write_local_mmr(UVH_LB_BAU_SB_ACTIVATION_CONTROL, index); |
274 | completion_status = uv_wait_completion(bau_desc, mmr_offset, | |
275 | right_shift); | |
276 | } while (completion_status == FLUSH_RETRY); | |
277 | time2 = get_cycles(); | |
278 | __get_cpu_var(ptcstats).sflush += (time2 - time1); | |
279 | if (tries > 1) | |
280 | __get_cpu_var(ptcstats).retriesok++; | |
281 | ||
282 | if (completion_status == FLUSH_GIVEUP) { | |
283 | /* | |
284 | * Cause the caller to do an IPI-style TLB shootdown on | |
285 | * the cpu's, all of which are still in the mask. | |
286 | */ | |
287 | __get_cpu_var(ptcstats).ptc_i++; | |
2749ebe3 | 288 | return flush_mask; |
b194b120 CW |
289 | } |
290 | ||
291 | /* | |
292 | * Success, so clear the remote cpu's from the mask so we don't | |
293 | * use the IPI method of shootdown on them. | |
294 | */ | |
bdbcdd48 | 295 | for_each_cpu(bit, flush_mask) { |
9674f35b CW |
296 | pnode = uv_cpu_to_pnode(bit); |
297 | if (pnode == this_pnode) | |
b194b120 | 298 | continue; |
bdbcdd48 | 299 | cpumask_clear_cpu(bit, flush_mask); |
b194b120 | 300 | } |
bdbcdd48 TH |
301 | if (!cpumask_empty(flush_mask)) |
302 | return flush_mask; | |
303 | return NULL; | |
b194b120 CW |
304 | } |
305 | ||
1812924b CW |
306 | /** |
307 | * uv_flush_tlb_others - globally purge translation cache of a virtual | |
308 | * address or all TLB's | |
bdbcdd48 | 309 | * @cpumask: mask of all cpu's in which the address is to be removed |
1812924b CW |
310 | * @mm: mm_struct containing virtual address range |
311 | * @va: virtual address to be removed (or TLB_FLUSH_ALL for all TLB's on cpu) | |
bdbcdd48 | 312 | * @cpu: the current cpu |
1812924b CW |
313 | * |
314 | * This is the entry point for initiating any UV global TLB shootdown. | |
315 | * | |
316 | * Purges the translation caches of all specified processors of the given | |
317 | * virtual address, or purges all TLB's on specified processors. | |
318 | * | |
bdbcdd48 TH |
319 | * The caller has derived the cpumask from the mm_struct. This function |
320 | * is called only if there are bits set in the mask. (e.g. flush_tlb_page()) | |
1812924b | 321 | * |
bdbcdd48 | 322 | * The cpumask is converted into a nodemask of the nodes containing |
1812924b | 323 | * the cpus. |
b194b120 | 324 | * |
bdbcdd48 TH |
325 | * Note that this function should be called with preemption disabled. |
326 | * | |
327 | * Returns NULL if all remote flushing was done. | |
328 | * Returns pointer to cpumask if some remote flushing remains to be | |
329 | * done. The returned pointer is valid till preemption is re-enabled. | |
1812924b | 330 | */ |
bdbcdd48 TH |
331 | const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask, |
332 | struct mm_struct *mm, | |
333 | unsigned long va, unsigned int cpu) | |
1812924b | 334 | { |
bdbcdd48 TH |
335 | static DEFINE_PER_CPU(cpumask_t, flush_tlb_mask); |
336 | struct cpumask *flush_mask = &__get_cpu_var(flush_tlb_mask); | |
1812924b | 337 | int i; |
b194b120 | 338 | int bit; |
9674f35b | 339 | int pnode; |
bdbcdd48 | 340 | int uv_cpu; |
9674f35b | 341 | int this_pnode; |
b194b120 | 342 | int locals = 0; |
dc163a41 | 343 | struct bau_desc *bau_desc; |
bdbcdd48 TH |
344 | |
345 | cpumask_andnot(flush_mask, cpumask, cpumask_of(cpu)); | |
346 | ||
347 | uv_cpu = uv_blade_processor_id(); | |
9674f35b | 348 | this_pnode = uv_hub_info->pnode; |
1812924b | 349 | bau_desc = __get_cpu_var(bau_control).descriptor_base; |
bdbcdd48 | 350 | bau_desc += UV_ITEMS_PER_DESCRIPTOR * uv_cpu; |
1812924b CW |
351 | |
352 | bau_nodes_clear(&bau_desc->distribution, UV_DISTRIBUTION_SIZE); | |
353 | ||
354 | i = 0; | |
bdbcdd48 | 355 | for_each_cpu(bit, flush_mask) { |
9674f35b CW |
356 | pnode = uv_cpu_to_pnode(bit); |
357 | BUG_ON(pnode > (UV_DISTRIBUTION_SIZE - 1)); | |
358 | if (pnode == this_pnode) { | |
b194b120 | 359 | locals++; |
1812924b | 360 | continue; |
b194b120 | 361 | } |
9674f35b | 362 | bau_node_set(pnode, &bau_desc->distribution); |
1812924b CW |
363 | i++; |
364 | } | |
b194b120 CW |
365 | if (i == 0) { |
366 | /* | |
367 | * no off_node flushing; return status for local node | |
368 | */ | |
369 | if (locals) | |
bdbcdd48 | 370 | return flush_mask; |
b194b120 | 371 | else |
bdbcdd48 | 372 | return NULL; |
b194b120 | 373 | } |
1812924b CW |
374 | __get_cpu_var(ptcstats).requestor++; |
375 | __get_cpu_var(ptcstats).ntargeted += i; | |
376 | ||
377 | bau_desc->payload.address = va; | |
bdbcdd48 | 378 | bau_desc->payload.sending_cpu = cpu; |
1812924b | 379 | |
9674f35b | 380 | return uv_flush_send_and_wait(uv_cpu, this_pnode, bau_desc, flush_mask); |
1812924b CW |
381 | } |
382 | ||
383 | /* | |
384 | * The BAU message interrupt comes here. (registered by set_intr_gate) | |
385 | * See entry_64.S | |
386 | * | |
387 | * We received a broadcast assist message. | |
388 | * | |
389 | * Interrupts may have been disabled; this interrupt could represent | |
390 | * the receipt of several messages. | |
391 | * | |
392 | * All cores/threads on this node get this interrupt. | |
393 | * The last one to see it does the s/w ack. | |
394 | * (the resource will not be freed until noninterruptable cpus see this | |
395 | * interrupt; hardware will timeout the s/w ack and reply ERROR) | |
396 | */ | |
b194b120 | 397 | void uv_bau_message_interrupt(struct pt_regs *regs) |
1812924b | 398 | { |
dc163a41 IM |
399 | struct bau_payload_queue_entry *va_queue_first; |
400 | struct bau_payload_queue_entry *va_queue_last; | |
b4c286e6 | 401 | struct bau_payload_queue_entry *msg; |
1812924b | 402 | struct pt_regs *old_regs = set_irq_regs(regs); |
b4c286e6 IM |
403 | cycles_t time1; |
404 | cycles_t time2; | |
1812924b CW |
405 | int msg_slot; |
406 | int sw_ack_slot; | |
407 | int fw; | |
408 | int count = 0; | |
409 | unsigned long local_pnode; | |
410 | ||
411 | ack_APIC_irq(); | |
412 | exit_idle(); | |
413 | irq_enter(); | |
414 | ||
b194b120 | 415 | time1 = get_cycles(); |
1812924b CW |
416 | |
417 | local_pnode = uv_blade_to_pnode(uv_numa_blade_id()); | |
418 | ||
b4c286e6 | 419 | va_queue_first = __get_cpu_var(bau_control).va_queue_first; |
dc163a41 | 420 | va_queue_last = __get_cpu_var(bau_control).va_queue_last; |
b4c286e6 | 421 | |
1812924b CW |
422 | msg = __get_cpu_var(bau_control).bau_msg_head; |
423 | while (msg->sw_ack_vector) { | |
424 | count++; | |
425 | fw = msg->sw_ack_vector; | |
b4c286e6 | 426 | msg_slot = msg - va_queue_first; |
1812924b CW |
427 | sw_ack_slot = ffs(fw) - 1; |
428 | ||
429 | uv_bau_process_message(msg, msg_slot, sw_ack_slot); | |
430 | ||
431 | msg++; | |
dc163a41 IM |
432 | if (msg > va_queue_last) |
433 | msg = va_queue_first; | |
1812924b CW |
434 | __get_cpu_var(bau_control).bau_msg_head = msg; |
435 | } | |
436 | if (!count) | |
437 | __get_cpu_var(ptcstats).nomsg++; | |
438 | else if (count > 1) | |
439 | __get_cpu_var(ptcstats).multmsg++; | |
440 | ||
b194b120 CW |
441 | time2 = get_cycles(); |
442 | __get_cpu_var(ptcstats).dflush += (time2 - time1); | |
1812924b CW |
443 | |
444 | irq_exit(); | |
445 | set_irq_regs(old_regs); | |
1812924b CW |
446 | } |
447 | ||
c4c4688f CW |
448 | /* |
449 | * uv_enable_timeouts | |
450 | * | |
451 | * Each target blade (i.e. blades that have cpu's) needs to have | |
452 | * shootdown message timeouts enabled. The timeout does not cause | |
453 | * an interrupt, but causes an error message to be returned to | |
454 | * the sender. | |
455 | */ | |
b194b120 | 456 | static void uv_enable_timeouts(void) |
1812924b | 457 | { |
1812924b | 458 | int blade; |
c4c4688f | 459 | int nblades; |
1812924b | 460 | int pnode; |
c4c4688f CW |
461 | unsigned long mmr_image; |
462 | ||
463 | nblades = uv_num_possible_blades(); | |
1812924b | 464 | |
c4c4688f CW |
465 | for (blade = 0; blade < nblades; blade++) { |
466 | if (!uv_blade_nr_possible_cpus(blade)) | |
1812924b | 467 | continue; |
c4c4688f | 468 | |
1812924b | 469 | pnode = uv_blade_to_pnode(blade); |
c4c4688f CW |
470 | mmr_image = |
471 | uv_read_global_mmr64(pnode, UVH_LB_BAU_MISC_CONTROL); | |
472 | /* | |
473 | * Set the timeout period and then lock it in, in three | |
474 | * steps; captures and locks in the period. | |
475 | * | |
476 | * To program the period, the SOFT_ACK_MODE must be off. | |
477 | */ | |
478 | mmr_image &= ~((unsigned long)1 << | |
479 | UV_ENABLE_INTD_SOFT_ACK_MODE_SHIFT); | |
480 | uv_write_global_mmr64 | |
481 | (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image); | |
482 | /* | |
483 | * Set the 4-bit period. | |
484 | */ | |
485 | mmr_image &= ~((unsigned long)0xf << | |
486 | UV_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHIFT); | |
487 | mmr_image |= (UV_INTD_SOFT_ACK_TIMEOUT_PERIOD << | |
488 | UV_INTD_SOFT_ACK_TIMEOUT_PERIOD_SHIFT); | |
489 | uv_write_global_mmr64 | |
490 | (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image); | |
491 | /* | |
492 | * Subsequent reversals of the timebase bit (3) cause an | |
493 | * immediate timeout of one or all INTD resources as | |
494 | * indicated in bits 2:0 (7 causes all of them to timeout). | |
495 | */ | |
496 | mmr_image |= ((unsigned long)1 << | |
497 | UV_ENABLE_INTD_SOFT_ACK_MODE_SHIFT); | |
498 | uv_write_global_mmr64 | |
499 | (pnode, UVH_LB_BAU_MISC_CONTROL, mmr_image); | |
1812924b | 500 | } |
1812924b CW |
501 | } |
502 | ||
b194b120 | 503 | static void *uv_ptc_seq_start(struct seq_file *file, loff_t *offset) |
1812924b CW |
504 | { |
505 | if (*offset < num_possible_cpus()) | |
506 | return offset; | |
507 | return NULL; | |
508 | } | |
509 | ||
b194b120 | 510 | static void *uv_ptc_seq_next(struct seq_file *file, void *data, loff_t *offset) |
1812924b CW |
511 | { |
512 | (*offset)++; | |
513 | if (*offset < num_possible_cpus()) | |
514 | return offset; | |
515 | return NULL; | |
516 | } | |
517 | ||
b194b120 | 518 | static void uv_ptc_seq_stop(struct seq_file *file, void *data) |
1812924b CW |
519 | { |
520 | } | |
521 | ||
522 | /* | |
523 | * Display the statistics thru /proc | |
524 | * data points to the cpu number | |
525 | */ | |
b194b120 | 526 | static int uv_ptc_seq_show(struct seq_file *file, void *data) |
1812924b CW |
527 | { |
528 | struct ptc_stats *stat; | |
529 | int cpu; | |
530 | ||
531 | cpu = *(loff_t *)data; | |
532 | ||
533 | if (!cpu) { | |
534 | seq_printf(file, | |
535 | "# cpu requestor requestee one all sretry dretry ptc_i "); | |
536 | seq_printf(file, | |
b194b120 | 537 | "sw_ack sflush dflush sok dnomsg dmult starget\n"); |
1812924b CW |
538 | } |
539 | if (cpu < num_possible_cpus() && cpu_online(cpu)) { | |
540 | stat = &per_cpu(ptcstats, cpu); | |
541 | seq_printf(file, "cpu %d %ld %ld %ld %ld %ld %ld %ld ", | |
542 | cpu, stat->requestor, | |
543 | stat->requestee, stat->onetlb, stat->alltlb, | |
544 | stat->s_retry, stat->d_retry, stat->ptc_i); | |
545 | seq_printf(file, "%lx %ld %ld %ld %ld %ld %ld\n", | |
9674f35b | 546 | uv_read_global_mmr64(uv_cpu_to_pnode(cpu), |
1812924b | 547 | UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE), |
b194b120 | 548 | stat->sflush, stat->dflush, |
1812924b CW |
549 | stat->retriesok, stat->nomsg, |
550 | stat->multmsg, stat->ntargeted); | |
551 | } | |
552 | ||
553 | return 0; | |
554 | } | |
555 | ||
556 | /* | |
557 | * 0: display meaning of the statistics | |
558 | * >0: retry limit | |
559 | */ | |
b194b120 | 560 | static ssize_t uv_ptc_proc_write(struct file *file, const char __user *user, |
b4c286e6 | 561 | size_t count, loff_t *data) |
1812924b CW |
562 | { |
563 | long newmode; | |
564 | char optstr[64]; | |
565 | ||
e7eb8726 | 566 | if (count == 0 || count > sizeof(optstr)) |
cef53278 | 567 | return -EINVAL; |
1812924b CW |
568 | if (copy_from_user(optstr, user, count)) |
569 | return -EFAULT; | |
570 | optstr[count - 1] = '\0'; | |
571 | if (strict_strtoul(optstr, 10, &newmode) < 0) { | |
572 | printk(KERN_DEBUG "%s is invalid\n", optstr); | |
573 | return -EINVAL; | |
574 | } | |
575 | ||
576 | if (newmode == 0) { | |
577 | printk(KERN_DEBUG "# cpu: cpu number\n"); | |
578 | printk(KERN_DEBUG | |
579 | "requestor: times this cpu was the flush requestor\n"); | |
580 | printk(KERN_DEBUG | |
581 | "requestee: times this cpu was requested to flush its TLBs\n"); | |
582 | printk(KERN_DEBUG | |
583 | "one: times requested to flush a single address\n"); | |
584 | printk(KERN_DEBUG | |
585 | "all: times requested to flush all TLB's\n"); | |
586 | printk(KERN_DEBUG | |
587 | "sretry: number of retries of source-side timeouts\n"); | |
588 | printk(KERN_DEBUG | |
589 | "dretry: number of retries of destination-side timeouts\n"); | |
590 | printk(KERN_DEBUG | |
591 | "ptc_i: times UV fell through to IPI-style flushes\n"); | |
592 | printk(KERN_DEBUG | |
593 | "sw_ack: image of UVH_LB_BAU_INTD_SOFTWARE_ACKNOWLEDGE\n"); | |
594 | printk(KERN_DEBUG | |
b194b120 | 595 | "sflush_us: cycles spent in uv_flush_tlb_others()\n"); |
1812924b | 596 | printk(KERN_DEBUG |
b194b120 | 597 | "dflush_us: cycles spent in handling flush requests\n"); |
1812924b CW |
598 | printk(KERN_DEBUG "sok: successes on retry\n"); |
599 | printk(KERN_DEBUG "dnomsg: interrupts with no message\n"); | |
600 | printk(KERN_DEBUG | |
601 | "dmult: interrupts with multiple messages\n"); | |
602 | printk(KERN_DEBUG "starget: nodes targeted\n"); | |
603 | } else { | |
604 | uv_bau_retry_limit = newmode; | |
605 | printk(KERN_DEBUG "timeout retry limit:%d\n", | |
606 | uv_bau_retry_limit); | |
607 | } | |
608 | ||
609 | return count; | |
610 | } | |
611 | ||
612 | static const struct seq_operations uv_ptc_seq_ops = { | |
dc163a41 IM |
613 | .start = uv_ptc_seq_start, |
614 | .next = uv_ptc_seq_next, | |
615 | .stop = uv_ptc_seq_stop, | |
616 | .show = uv_ptc_seq_show | |
1812924b CW |
617 | }; |
618 | ||
b194b120 | 619 | static int uv_ptc_proc_open(struct inode *inode, struct file *file) |
1812924b CW |
620 | { |
621 | return seq_open(file, &uv_ptc_seq_ops); | |
622 | } | |
623 | ||
624 | static const struct file_operations proc_uv_ptc_operations = { | |
b194b120 CW |
625 | .open = uv_ptc_proc_open, |
626 | .read = seq_read, | |
627 | .write = uv_ptc_proc_write, | |
628 | .llseek = seq_lseek, | |
629 | .release = seq_release, | |
1812924b CW |
630 | }; |
631 | ||
b194b120 | 632 | static int __init uv_ptc_init(void) |
1812924b | 633 | { |
b194b120 | 634 | struct proc_dir_entry *proc_uv_ptc; |
1812924b CW |
635 | |
636 | if (!is_uv_system()) | |
637 | return 0; | |
638 | ||
1812924b CW |
639 | proc_uv_ptc = create_proc_entry(UV_PTC_BASENAME, 0444, NULL); |
640 | if (!proc_uv_ptc) { | |
641 | printk(KERN_ERR "unable to create %s proc entry\n", | |
642 | UV_PTC_BASENAME); | |
643 | return -EINVAL; | |
644 | } | |
645 | proc_uv_ptc->proc_fops = &proc_uv_ptc_operations; | |
646 | return 0; | |
647 | } | |
648 | ||
b194b120 CW |
649 | /* |
650 | * begin the initialization of the per-blade control structures | |
651 | */ | |
652 | static struct bau_control * __init uv_table_bases_init(int blade, int node) | |
1812924b | 653 | { |
b194b120 | 654 | int i; |
b194b120 | 655 | struct bau_msg_status *msp; |
dc163a41 | 656 | struct bau_control *bau_tabp; |
b194b120 | 657 | |
dc163a41 | 658 | bau_tabp = |
b194b120 | 659 | kmalloc_node(sizeof(struct bau_control), GFP_KERNEL, node); |
dc163a41 | 660 | BUG_ON(!bau_tabp); |
b4c286e6 | 661 | |
dc163a41 | 662 | bau_tabp->msg_statuses = |
b194b120 | 663 | kmalloc_node(sizeof(struct bau_msg_status) * |
dc163a41 IM |
664 | DEST_Q_SIZE, GFP_KERNEL, node); |
665 | BUG_ON(!bau_tabp->msg_statuses); | |
b4c286e6 | 666 | |
dc163a41 | 667 | for (i = 0, msp = bau_tabp->msg_statuses; i < DEST_Q_SIZE; i++, msp++) |
b194b120 CW |
668 | bau_cpubits_clear(&msp->seen_by, (int) |
669 | uv_blade_nr_possible_cpus(blade)); | |
b4c286e6 | 670 | |
dc163a41 | 671 | uv_bau_table_bases[blade] = bau_tabp; |
b4c286e6 | 672 | |
d400524a | 673 | return bau_tabp; |
1812924b CW |
674 | } |
675 | ||
b194b120 CW |
676 | /* |
677 | * finish the initialization of the per-blade control structures | |
678 | */ | |
b4c286e6 | 679 | static void __init |
9674f35b | 680 | uv_table_bases_finish(int blade, |
b4c286e6 IM |
681 | struct bau_control *bau_tablesp, |
682 | struct bau_desc *adp) | |
b194b120 | 683 | { |
b194b120 | 684 | struct bau_control *bcp; |
9674f35b | 685 | int cpu; |
b194b120 | 686 | |
9674f35b CW |
687 | for_each_present_cpu(cpu) { |
688 | if (blade != uv_cpu_to_blade_id(cpu)) | |
689 | continue; | |
b4c286e6 | 690 | |
9674f35b | 691 | bcp = (struct bau_control *)&per_cpu(bau_control, cpu); |
b4c286e6 IM |
692 | bcp->bau_msg_head = bau_tablesp->va_queue_first; |
693 | bcp->va_queue_first = bau_tablesp->va_queue_first; | |
694 | bcp->va_queue_last = bau_tablesp->va_queue_last; | |
b4c286e6 IM |
695 | bcp->msg_statuses = bau_tablesp->msg_statuses; |
696 | bcp->descriptor_base = adp; | |
b194b120 CW |
697 | } |
698 | } | |
1812924b CW |
699 | |
700 | /* | |
b194b120 | 701 | * initialize the sending side's sending buffers |
1812924b | 702 | */ |
dc163a41 | 703 | static struct bau_desc * __init |
b194b120 | 704 | uv_activation_descriptor_init(int node, int pnode) |
1812924b CW |
705 | { |
706 | int i; | |
1812924b | 707 | unsigned long pa; |
1812924b | 708 | unsigned long m; |
b194b120 | 709 | unsigned long n; |
1812924b | 710 | unsigned long mmr_image; |
dc163a41 IM |
711 | struct bau_desc *adp; |
712 | struct bau_desc *ad2; | |
b194b120 | 713 | |
9674f35b | 714 | adp = (struct bau_desc *)kmalloc_node(16384, GFP_KERNEL, node); |
dc163a41 | 715 | BUG_ON(!adp); |
b4c286e6 | 716 | |
b194b120 CW |
717 | pa = __pa((unsigned long)adp); |
718 | n = pa >> uv_nshift; | |
719 | m = pa & uv_mmask; | |
b4c286e6 | 720 | |
b194b120 | 721 | mmr_image = uv_read_global_mmr64(pnode, UVH_LB_BAU_SB_DESCRIPTOR_BASE); |
b4c286e6 | 722 | if (mmr_image) { |
b194b120 CW |
723 | uv_write_global_mmr64(pnode, (unsigned long) |
724 | UVH_LB_BAU_SB_DESCRIPTOR_BASE, | |
725 | (n << UV_DESC_BASE_PNODE_SHIFT | m)); | |
b4c286e6 IM |
726 | } |
727 | ||
b194b120 | 728 | for (i = 0, ad2 = adp; i < UV_ACTIVATION_DESCRIPTOR_SIZE; i++, ad2++) { |
dc163a41 | 729 | memset(ad2, 0, sizeof(struct bau_desc)); |
b194b120 | 730 | ad2->header.sw_ack_flag = 1; |
9674f35b | 731 | ad2->header.base_dest_nodeid = uv_cpu_to_pnode(0); |
b194b120 CW |
732 | ad2->header.command = UV_NET_ENDPOINT_INTD; |
733 | ad2->header.int_both = 1; | |
734 | /* | |
735 | * all others need to be set to zero: | |
736 | * fairness chaining multilevel count replied_to | |
737 | */ | |
738 | } | |
739 | return adp; | |
740 | } | |
741 | ||
742 | /* | |
743 | * initialize the destination side's receiving buffers | |
744 | */ | |
b4c286e6 IM |
745 | static struct bau_payload_queue_entry * __init |
746 | uv_payload_queue_init(int node, int pnode, struct bau_control *bau_tablesp) | |
b194b120 | 747 | { |
1812924b | 748 | struct bau_payload_queue_entry *pqp; |
b4c286e6 | 749 | char *cp; |
1812924b | 750 | |
dc163a41 IM |
751 | pqp = (struct bau_payload_queue_entry *) kmalloc_node( |
752 | (DEST_Q_SIZE + 1) * sizeof(struct bau_payload_queue_entry), | |
753 | GFP_KERNEL, node); | |
754 | BUG_ON(!pqp); | |
b4c286e6 | 755 | |
b194b120 CW |
756 | cp = (char *)pqp + 31; |
757 | pqp = (struct bau_payload_queue_entry *)(((unsigned long)cp >> 5) << 5); | |
758 | bau_tablesp->va_queue_first = pqp; | |
759 | uv_write_global_mmr64(pnode, | |
760 | UVH_LB_BAU_INTD_PAYLOAD_QUEUE_FIRST, | |
761 | ((unsigned long)pnode << | |
762 | UV_PAYLOADQ_PNODE_SHIFT) | | |
763 | uv_physnodeaddr(pqp)); | |
764 | uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_TAIL, | |
765 | uv_physnodeaddr(pqp)); | |
dc163a41 | 766 | bau_tablesp->va_queue_last = pqp + (DEST_Q_SIZE - 1); |
b194b120 CW |
767 | uv_write_global_mmr64(pnode, UVH_LB_BAU_INTD_PAYLOAD_QUEUE_LAST, |
768 | (unsigned long) | |
769 | uv_physnodeaddr(bau_tablesp->va_queue_last)); | |
dc163a41 | 770 | memset(pqp, 0, sizeof(struct bau_payload_queue_entry) * DEST_Q_SIZE); |
b4c286e6 | 771 | |
b194b120 CW |
772 | return pqp; |
773 | } | |
1812924b | 774 | |
b194b120 CW |
775 | /* |
776 | * Initialization of each UV blade's structures | |
777 | */ | |
9674f35b | 778 | static int __init uv_init_blade(int blade) |
b194b120 | 779 | { |
9674f35b | 780 | int node; |
b194b120 CW |
781 | int pnode; |
782 | unsigned long pa; | |
783 | unsigned long apicid; | |
dc163a41 | 784 | struct bau_desc *adp; |
b194b120 CW |
785 | struct bau_payload_queue_entry *pqp; |
786 | struct bau_control *bau_tablesp; | |
1812924b | 787 | |
9674f35b | 788 | node = blade_to_first_node(blade); |
b194b120 CW |
789 | bau_tablesp = uv_table_bases_init(blade, node); |
790 | pnode = uv_blade_to_pnode(blade); | |
791 | adp = uv_activation_descriptor_init(node, pnode); | |
792 | pqp = uv_payload_queue_init(node, pnode, bau_tablesp); | |
9674f35b | 793 | uv_table_bases_finish(blade, bau_tablesp, adp); |
b194b120 CW |
794 | /* |
795 | * the below initialization can't be in firmware because the | |
796 | * messaging IRQ will be determined by the OS | |
797 | */ | |
9674f35b | 798 | apicid = blade_to_first_apicid(blade); |
b194b120 CW |
799 | pa = uv_read_global_mmr64(pnode, UVH_BAU_DATA_CONFIG); |
800 | if ((pa & 0xff) != UV_BAU_MESSAGE) { | |
801 | uv_write_global_mmr64(pnode, UVH_BAU_DATA_CONFIG, | |
802 | ((apicid << 32) | UV_BAU_MESSAGE)); | |
1812924b | 803 | } |
b194b120 CW |
804 | return 0; |
805 | } | |
806 | ||
807 | /* | |
808 | * Initialization of BAU-related structures | |
809 | */ | |
810 | static int __init uv_bau_init(void) | |
811 | { | |
812 | int blade; | |
b194b120 | 813 | int nblades; |
2c74d666 | 814 | int cur_cpu; |
b194b120 CW |
815 | |
816 | if (!is_uv_system()) | |
817 | return 0; | |
1812924b | 818 | |
b194b120 | 819 | uv_bau_retry_limit = 1; |
1812924b | 820 | uv_nshift = uv_hub_info->n_val; |
dc163a41 | 821 | uv_mmask = (1UL << uv_hub_info->n_val) - 1; |
9674f35b CW |
822 | nblades = uv_num_possible_blades(); |
823 | ||
1812924b CW |
824 | uv_bau_table_bases = (struct bau_control **) |
825 | kmalloc(nblades * sizeof(struct bau_control *), GFP_KERNEL); | |
dc163a41 | 826 | BUG_ON(!uv_bau_table_bases); |
b4c286e6 | 827 | |
9674f35b CW |
828 | for (blade = 0; blade < nblades; blade++) |
829 | if (uv_blade_nr_possible_cpus(blade)) | |
830 | uv_init_blade(blade); | |
831 | ||
99dd8713 | 832 | alloc_intr_gate(UV_BAU_MESSAGE, uv_bau_message_intr1); |
1812924b | 833 | uv_enable_timeouts(); |
b4c286e6 | 834 | |
1812924b CW |
835 | return 0; |
836 | } | |
1812924b | 837 | __initcall(uv_bau_init); |
b194b120 | 838 | __initcall(uv_ptc_init); |