Merge branch 'release' of git://git.kernel.org/pub/scm/linux/kernel/git/aegl/linux-2.6
[linux-2.6] / arch / ia64 / sn / kernel / sn2 / sn2_smp.c
1 /*
2  * SN2 Platform specific SMP Support
3  *
4  * This file is subject to the terms and conditions of the GNU General Public
5  * License.  See the file "COPYING" in the main directory of this archive
6  * for more details.
7  *
8  * Copyright (C) 2000-2006 Silicon Graphics, Inc. All rights reserved.
9  */
10
11 #include <linux/init.h>
12 #include <linux/kernel.h>
13 #include <linux/spinlock.h>
14 #include <linux/threads.h>
15 #include <linux/sched.h>
16 #include <linux/smp.h>
17 #include <linux/interrupt.h>
18 #include <linux/irq.h>
19 #include <linux/mmzone.h>
20 #include <linux/module.h>
21 #include <linux/bitops.h>
22 #include <linux/nodemask.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25
26 #include <asm/processor.h>
27 #include <asm/irq.h>
28 #include <asm/sal.h>
29 #include <asm/system.h>
30 #include <asm/delay.h>
31 #include <asm/io.h>
32 #include <asm/smp.h>
33 #include <asm/tlb.h>
34 #include <asm/numa.h>
35 #include <asm/hw_irq.h>
36 #include <asm/current.h>
37 #include <asm/sn/sn_cpuid.h>
38 #include <asm/sn/sn_sal.h>
39 #include <asm/sn/addrs.h>
40 #include <asm/sn/shub_mmr.h>
41 #include <asm/sn/nodepda.h>
42 #include <asm/sn/rw_mmr.h>
43
44 DEFINE_PER_CPU(struct ptc_stats, ptcstats);
45 DECLARE_PER_CPU(struct ptc_stats, ptcstats);
46
47 static  __cacheline_aligned DEFINE_SPINLOCK(sn2_global_ptc_lock);
48
49 /* 0 = old algorithm (no IPI flushes), 1 = ipi deadlock flush, 2 = ipi instead of SHUB ptc, >2 = always ipi */
50 static int sn2_flush_opt = 0;
51
52 extern unsigned long
53 sn2_ptc_deadlock_recovery_core(volatile unsigned long *, unsigned long,
54                                volatile unsigned long *, unsigned long,
55                                volatile unsigned long *, unsigned long);
56 void
57 sn2_ptc_deadlock_recovery(short *, short, short, int,
58                           volatile unsigned long *, unsigned long,
59                           volatile unsigned long *, unsigned long);
60
61 /*
62  * Note: some is the following is captured here to make degugging easier
63  * (the macros make more sense if you see the debug patch - not posted)
64  */
65 #define sn2_ptctest     0
66 #define local_node_uses_ptc_ga(sh1)     ((sh1) ? 1 : 0)
67 #define max_active_pio(sh1)             ((sh1) ? 32 : 7)
68 #define reset_max_active_on_deadlock()  1
69 #define PTC_LOCK(sh1)                   ((sh1) ? &sn2_global_ptc_lock : &sn_nodepda->ptc_lock)
70
71 struct ptc_stats {
72         unsigned long ptc_l;
73         unsigned long change_rid;
74         unsigned long shub_ptc_flushes;
75         unsigned long nodes_flushed;
76         unsigned long deadlocks;
77         unsigned long deadlocks2;
78         unsigned long lock_itc_clocks;
79         unsigned long shub_itc_clocks;
80         unsigned long shub_itc_clocks_max;
81         unsigned long shub_ptc_flushes_not_my_mm;
82         unsigned long shub_ipi_flushes;
83         unsigned long shub_ipi_flushes_itc_clocks;
84 };
85
86 #define sn2_ptctest     0
87
88 static inline unsigned long wait_piowc(void)
89 {
90         volatile unsigned long *piows;
91         unsigned long zeroval, ws;
92
93         piows = pda->pio_write_status_addr;
94         zeroval = pda->pio_write_status_val;
95         do {
96                 cpu_relax();
97         } while (((ws = *piows) & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK) != zeroval);
98         return (ws & SH_PIO_WRITE_STATUS_WRITE_DEADLOCK_MASK) != 0;
99 }
100
101 /**
102  * sn_migrate - SN-specific task migration actions
103  * @task: Task being migrated to new CPU
104  *
105  * SN2 PIO writes from separate CPUs are not guaranteed to arrive in order.
106  * Context switching user threads which have memory-mapped MMIO may cause
107  * PIOs to issue from separate CPUs, thus the PIO writes must be drained
108  * from the previous CPU's Shub before execution resumes on the new CPU.
109  */
110 void sn_migrate(struct task_struct *task)
111 {
112         pda_t *last_pda = pdacpu(task_thread_info(task)->last_cpu);
113         volatile unsigned long *adr = last_pda->pio_write_status_addr;
114         unsigned long val = last_pda->pio_write_status_val;
115
116         /* Drain PIO writes from old CPU's Shub */
117         while (unlikely((*adr & SH_PIO_WRITE_STATUS_PENDING_WRITE_COUNT_MASK)
118                         != val))
119                 cpu_relax();
120 }
121
122 void sn_tlb_migrate_finish(struct mm_struct *mm)
123 {
124         /* flush_tlb_mm is inefficient if more than 1 users of mm */
125         if (mm == current->mm && mm && atomic_read(&mm->mm_users) == 1)
126                 flush_tlb_mm(mm);
127 }
128
129 static void
130 sn2_ipi_flush_all_tlb(struct mm_struct *mm)
131 {
132         unsigned long itc;
133
134         itc = ia64_get_itc();
135         smp_flush_tlb_cpumask(mm->cpu_vm_mask);
136         itc = ia64_get_itc() - itc;
137         __get_cpu_var(ptcstats).shub_ipi_flushes_itc_clocks += itc;
138         __get_cpu_var(ptcstats).shub_ipi_flushes++;
139 }
140
141 /**
142  * sn2_global_tlb_purge - globally purge translation cache of virtual address range
143  * @mm: mm_struct containing virtual address range
144  * @start: start of virtual address range
145  * @end: end of virtual address range
146  * @nbits: specifies number of bytes to purge per instruction (num = 1<<(nbits & 0xfc))
147  *
148  * Purges the translation caches of all processors of the given virtual address
149  * range.
150  *
151  * Note:
152  *      - cpu_vm_mask is a bit mask that indicates which cpus have loaded the context.
153  *      - cpu_vm_mask is converted into a nodemask of the nodes containing the
154  *        cpus in cpu_vm_mask.
155  *      - if only one bit is set in cpu_vm_mask & it is the current cpu & the
156  *        process is purging its own virtual address range, then only the
157  *        local TLB needs to be flushed. This flushing can be done using
158  *        ptc.l. This is the common case & avoids the global spinlock.
159  *      - if multiple cpus have loaded the context, then flushing has to be
160  *        done with ptc.g/MMRs under protection of the global ptc_lock.
161  */
162
163 void
164 sn2_global_tlb_purge(struct mm_struct *mm, unsigned long start,
165                      unsigned long end, unsigned long nbits)
166 {
167         int i, ibegin, shub1, cnode, mynasid, cpu, lcpu = 0, nasid;
168         int mymm = (mm == current->active_mm && mm == current->mm);
169         int use_cpu_ptcga;
170         volatile unsigned long *ptc0, *ptc1;
171         unsigned long itc, itc2, flags, data0 = 0, data1 = 0, rr_value, old_rr = 0;
172         short nasids[MAX_NUMNODES], nix;
173         nodemask_t nodes_flushed;
174         int active, max_active, deadlock, flush_opt = sn2_flush_opt;
175
176         if (flush_opt > 2) {
177                 sn2_ipi_flush_all_tlb(mm);
178                 return;
179         }
180
181         nodes_clear(nodes_flushed);
182         i = 0;
183
184         for_each_cpu_mask(cpu, mm->cpu_vm_mask) {
185                 cnode = cpu_to_node(cpu);
186                 node_set(cnode, nodes_flushed);
187                 lcpu = cpu;
188                 i++;
189         }
190
191         if (i == 0)
192                 return;
193
194         preempt_disable();
195
196         if (likely(i == 1 && lcpu == smp_processor_id() && mymm)) {
197                 do {
198                         ia64_ptcl(start, nbits << 2);
199                         start += (1UL << nbits);
200                 } while (start < end);
201                 ia64_srlz_i();
202                 __get_cpu_var(ptcstats).ptc_l++;
203                 preempt_enable();
204                 return;
205         }
206
207         if (atomic_read(&mm->mm_users) == 1 && mymm) {
208                 flush_tlb_mm(mm);
209                 __get_cpu_var(ptcstats).change_rid++;
210                 preempt_enable();
211                 return;
212         }
213
214         if (flush_opt == 2) {
215                 sn2_ipi_flush_all_tlb(mm);
216                 preempt_enable();
217                 return;
218         }
219
220         itc = ia64_get_itc();
221         nix = 0;
222         for_each_node_mask(cnode, nodes_flushed)
223                 nasids[nix++] = cnodeid_to_nasid(cnode);
224
225         rr_value = (mm->context << 3) | REGION_NUMBER(start);
226
227         shub1 = is_shub1();
228         if (shub1) {
229                 data0 = (1UL << SH1_PTC_0_A_SHFT) |
230                         (nbits << SH1_PTC_0_PS_SHFT) |
231                         (rr_value << SH1_PTC_0_RID_SHFT) |
232                         (1UL << SH1_PTC_0_START_SHFT);
233                 ptc0 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH1_PTC_0);
234                 ptc1 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH1_PTC_1);
235         } else {
236                 data0 = (1UL << SH2_PTC_A_SHFT) |
237                         (nbits << SH2_PTC_PS_SHFT) |
238                         (1UL << SH2_PTC_START_SHFT);
239                 ptc0 = (long *)GLOBAL_MMR_PHYS_ADDR(0, SH2_PTC + 
240                         (rr_value << SH2_PTC_RID_SHFT));
241                 ptc1 = NULL;
242         }
243         
244
245         mynasid = get_nasid();
246         use_cpu_ptcga = local_node_uses_ptc_ga(shub1);
247         max_active = max_active_pio(shub1);
248
249         itc = ia64_get_itc();
250         spin_lock_irqsave(PTC_LOCK(shub1), flags);
251         itc2 = ia64_get_itc();
252
253         __get_cpu_var(ptcstats).lock_itc_clocks += itc2 - itc;
254         __get_cpu_var(ptcstats).shub_ptc_flushes++;
255         __get_cpu_var(ptcstats).nodes_flushed += nix;
256         if (!mymm)
257                  __get_cpu_var(ptcstats).shub_ptc_flushes_not_my_mm++;
258
259         if (use_cpu_ptcga && !mymm) {
260                 old_rr = ia64_get_rr(start);
261                 ia64_set_rr(start, (old_rr & 0xff) | (rr_value << 8));
262                 ia64_srlz_d();
263         }
264
265         wait_piowc();
266         do {
267                 if (shub1)
268                         data1 = start | (1UL << SH1_PTC_1_START_SHFT);
269                 else
270                         data0 = (data0 & ~SH2_PTC_ADDR_MASK) | (start & SH2_PTC_ADDR_MASK);
271                 deadlock = 0;
272                 active = 0;
273                 for (ibegin = 0, i = 0; i < nix; i++) {
274                         nasid = nasids[i];
275                         if (use_cpu_ptcga && unlikely(nasid == mynasid)) {
276                                 ia64_ptcga(start, nbits << 2);
277                                 ia64_srlz_i();
278                         } else {
279                                 ptc0 = CHANGE_NASID(nasid, ptc0);
280                                 if (ptc1)
281                                         ptc1 = CHANGE_NASID(nasid, ptc1);
282                                 pio_atomic_phys_write_mmrs(ptc0, data0, ptc1, data1);
283                                 active++;
284                         }
285                         if (active >= max_active || i == (nix - 1)) {
286                                 if ((deadlock = wait_piowc())) {
287                                         if (flush_opt == 1)
288                                                 goto done;
289                                         sn2_ptc_deadlock_recovery(nasids, ibegin, i, mynasid, ptc0, data0, ptc1, data1);
290                                         if (reset_max_active_on_deadlock())
291                                                 max_active = 1;
292                                 }
293                                 active = 0;
294                                 ibegin = i + 1;
295                         }
296                 }
297                 start += (1UL << nbits);
298         } while (start < end);
299
300 done:
301         itc2 = ia64_get_itc() - itc2;
302         __get_cpu_var(ptcstats).shub_itc_clocks += itc2;
303         if (itc2 > __get_cpu_var(ptcstats).shub_itc_clocks_max)
304                 __get_cpu_var(ptcstats).shub_itc_clocks_max = itc2;
305
306         if (old_rr) {
307                 ia64_set_rr(start, old_rr);
308                 ia64_srlz_d();
309         }
310
311         spin_unlock_irqrestore(PTC_LOCK(shub1), flags);
312
313         if (flush_opt == 1 && deadlock) {
314                 __get_cpu_var(ptcstats).deadlocks++;
315                 sn2_ipi_flush_all_tlb(mm);
316         }
317
318         preempt_enable();
319 }
320
321 /*
322  * sn2_ptc_deadlock_recovery
323  *
324  * Recover from PTC deadlocks conditions. Recovery requires stepping thru each 
325  * TLB flush transaction.  The recovery sequence is somewhat tricky & is
326  * coded in assembly language.
327  */
328
329 void
330 sn2_ptc_deadlock_recovery(short *nasids, short ib, short ie, int mynasid,
331                           volatile unsigned long *ptc0, unsigned long data0,
332                           volatile unsigned long *ptc1, unsigned long data1)
333 {
334         short nasid, i;
335         unsigned long *piows, zeroval, n;
336
337         __get_cpu_var(ptcstats).deadlocks++;
338
339         piows = (unsigned long *) pda->pio_write_status_addr;
340         zeroval = pda->pio_write_status_val;
341
342
343         for (i=ib; i <= ie; i++) {
344                 nasid = nasids[i];
345                 if (local_node_uses_ptc_ga(is_shub1()) && nasid == mynasid)
346                         continue;
347                 ptc0 = CHANGE_NASID(nasid, ptc0);
348                 if (ptc1)
349                         ptc1 = CHANGE_NASID(nasid, ptc1);
350
351                 n = sn2_ptc_deadlock_recovery_core(ptc0, data0, ptc1, data1, piows, zeroval);
352                 __get_cpu_var(ptcstats).deadlocks2 += n;
353         }
354
355 }
356
357 /**
358  * sn_send_IPI_phys - send an IPI to a Nasid and slice
359  * @nasid: nasid to receive the interrupt (may be outside partition)
360  * @physid: physical cpuid to receive the interrupt.
361  * @vector: command to send
362  * @delivery_mode: delivery mechanism
363  *
364  * Sends an IPI (interprocessor interrupt) to the processor specified by
365  * @physid
366  *
367  * @delivery_mode can be one of the following
368  *
369  * %IA64_IPI_DM_INT - pend an interrupt
370  * %IA64_IPI_DM_PMI - pend a PMI
371  * %IA64_IPI_DM_NMI - pend an NMI
372  * %IA64_IPI_DM_INIT - pend an INIT interrupt
373  */
374 void sn_send_IPI_phys(int nasid, long physid, int vector, int delivery_mode)
375 {
376         long val;
377         unsigned long flags = 0;
378         volatile long *p;
379
380         p = (long *)GLOBAL_MMR_PHYS_ADDR(nasid, SH_IPI_INT);
381         val = (1UL << SH_IPI_INT_SEND_SHFT) |
382             (physid << SH_IPI_INT_PID_SHFT) |
383             ((long)delivery_mode << SH_IPI_INT_TYPE_SHFT) |
384             ((long)vector << SH_IPI_INT_IDX_SHFT) |
385             (0x000feeUL << SH_IPI_INT_BASE_SHFT);
386
387         mb();
388         if (enable_shub_wars_1_1()) {
389                 spin_lock_irqsave(&sn2_global_ptc_lock, flags);
390         }
391         pio_phys_write_mmr(p, val);
392         if (enable_shub_wars_1_1()) {
393                 wait_piowc();
394                 spin_unlock_irqrestore(&sn2_global_ptc_lock, flags);
395         }
396
397 }
398
399 EXPORT_SYMBOL(sn_send_IPI_phys);
400
401 /**
402  * sn2_send_IPI - send an IPI to a processor
403  * @cpuid: target of the IPI
404  * @vector: command to send
405  * @delivery_mode: delivery mechanism
406  * @redirect: redirect the IPI?
407  *
408  * Sends an IPI (InterProcessor Interrupt) to the processor specified by
409  * @cpuid.  @vector specifies the command to send, while @delivery_mode can 
410  * be one of the following
411  *
412  * %IA64_IPI_DM_INT - pend an interrupt
413  * %IA64_IPI_DM_PMI - pend a PMI
414  * %IA64_IPI_DM_NMI - pend an NMI
415  * %IA64_IPI_DM_INIT - pend an INIT interrupt
416  */
417 void sn2_send_IPI(int cpuid, int vector, int delivery_mode, int redirect)
418 {
419         long physid;
420         int nasid;
421
422         physid = cpu_physical_id(cpuid);
423         nasid = cpuid_to_nasid(cpuid);
424
425         /* the following is used only when starting cpus at boot time */
426         if (unlikely(nasid == -1))
427                 ia64_sn_get_sapic_info(physid, &nasid, NULL, NULL);
428
429         sn_send_IPI_phys(nasid, physid, vector, delivery_mode);
430 }
431
432 #ifdef CONFIG_PROC_FS
433
434 #define PTC_BASENAME    "sgi_sn/ptc_statistics"
435
436 static void *sn2_ptc_seq_start(struct seq_file *file, loff_t * offset)
437 {
438         if (*offset < NR_CPUS)
439                 return offset;
440         return NULL;
441 }
442
443 static void *sn2_ptc_seq_next(struct seq_file *file, void *data, loff_t * offset)
444 {
445         (*offset)++;
446         if (*offset < NR_CPUS)
447                 return offset;
448         return NULL;
449 }
450
451 static void sn2_ptc_seq_stop(struct seq_file *file, void *data)
452 {
453 }
454
455 static int sn2_ptc_seq_show(struct seq_file *file, void *data)
456 {
457         struct ptc_stats *stat;
458         int cpu;
459
460         cpu = *(loff_t *) data;
461
462         if (!cpu) {
463                 seq_printf(file,
464                            "# cpu ptc_l newrid ptc_flushes nodes_flushed deadlocks lock_nsec shub_nsec shub_nsec_max not_my_mm deadlock2 ipi_fluches ipi_nsec\n");
465                 seq_printf(file, "# ptctest %d, flushopt %d\n", sn2_ptctest, sn2_flush_opt);
466         }
467
468         if (cpu < NR_CPUS && cpu_online(cpu)) {
469                 stat = &per_cpu(ptcstats, cpu);
470                 seq_printf(file, "cpu %d %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld %ld\n", cpu, stat->ptc_l,
471                                 stat->change_rid, stat->shub_ptc_flushes, stat->nodes_flushed,
472                                 stat->deadlocks,
473                                 1000 * stat->lock_itc_clocks / per_cpu(cpu_info, cpu).cyc_per_usec,
474                                 1000 * stat->shub_itc_clocks / per_cpu(cpu_info, cpu).cyc_per_usec,
475                                 1000 * stat->shub_itc_clocks_max / per_cpu(cpu_info, cpu).cyc_per_usec,
476                                 stat->shub_ptc_flushes_not_my_mm,
477                                 stat->deadlocks2,
478                                 stat->shub_ipi_flushes,
479                                 1000 * stat->shub_ipi_flushes_itc_clocks / per_cpu(cpu_info, cpu).cyc_per_usec);
480         }
481         return 0;
482 }
483
484 static ssize_t sn2_ptc_proc_write(struct file *file, const char __user *user, size_t count, loff_t *data)
485 {
486         int cpu;
487         char optstr[64];
488
489         if (copy_from_user(optstr, user, count))
490                 return -EFAULT;
491         optstr[count - 1] = '\0';
492         sn2_flush_opt = simple_strtoul(optstr, NULL, 0);
493
494         for_each_online_cpu(cpu)
495                 memset(&per_cpu(ptcstats, cpu), 0, sizeof(struct ptc_stats));
496
497         return count;
498 }
499
500 static struct seq_operations sn2_ptc_seq_ops = {
501         .start = sn2_ptc_seq_start,
502         .next = sn2_ptc_seq_next,
503         .stop = sn2_ptc_seq_stop,
504         .show = sn2_ptc_seq_show
505 };
506
507 static int sn2_ptc_proc_open(struct inode *inode, struct file *file)
508 {
509         return seq_open(file, &sn2_ptc_seq_ops);
510 }
511
512 static const struct file_operations proc_sn2_ptc_operations = {
513         .open = sn2_ptc_proc_open,
514         .read = seq_read,
515         .write = sn2_ptc_proc_write,
516         .llseek = seq_lseek,
517         .release = seq_release,
518 };
519
520 static struct proc_dir_entry *proc_sn2_ptc;
521
522 static int __init sn2_ptc_init(void)
523 {
524         if (!ia64_platform_is("sn2"))
525                 return 0;
526
527         if (!(proc_sn2_ptc = create_proc_entry(PTC_BASENAME, 0444, NULL))) {
528                 printk(KERN_ERR "unable to create %s proc entry", PTC_BASENAME);
529                 return -EINVAL;
530         }
531         proc_sn2_ptc->proc_fops = &proc_sn2_ptc_operations;
532         spin_lock_init(&sn2_global_ptc_lock);
533         return 0;
534 }
535
536 static void __exit sn2_ptc_exit(void)
537 {
538         remove_proc_entry(PTC_BASENAME, NULL);
539 }
540
541 module_init(sn2_ptc_init);
542 module_exit(sn2_ptc_exit);
543 #endif /* CONFIG_PROC_FS */
544