gru: fix bug in allocation of kernel contexts
[linux-2.6.git] / drivers / misc / sgi-gru / grufile.c
1 /*
2  * SN Platform GRU Driver
3  *
4  *              FILE OPERATIONS & DRIVER INITIALIZATION
5  *
6  * This file supports the user system call for file open, close, mmap, etc.
7  * This also incudes the driver initialization code.
8  *
9  *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
10  *
11  *  This program is free software; you can redistribute it and/or modify
12  *  it under the terms of the GNU General Public License as published by
13  *  the Free Software Foundation; either version 2 of the License, or
14  *  (at your option) any later version.
15  *
16  *  This program is distributed in the hope that it will be useful,
17  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
19  *  GNU General Public License for more details.
20  *
21  *  You should have received a copy of the GNU General Public License
22  *  along with this program; if not, write to the Free Software
23  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
24  */
25
26 #include <linux/module.h>
27 #include <linux/kernel.h>
28 #include <linux/errno.h>
29 #include <linux/slab.h>
30 #include <linux/mm.h>
31 #include <linux/io.h>
32 #include <linux/spinlock.h>
33 #include <linux/device.h>
34 #include <linux/miscdevice.h>
35 #include <linux/interrupt.h>
36 #include <linux/proc_fs.h>
37 #include <linux/uaccess.h>
38 #ifdef CONFIG_X86_64
39 #include <asm/uv/uv_irq.h>
40 #endif
41 #include <asm/uv/uv.h>
42 #include "gru.h"
43 #include "grulib.h"
44 #include "grutables.h"
45
46 #include <asm/uv/uv_hub.h>
47 #include <asm/uv/uv_mmrs.h>
48
49 struct gru_blade_state *gru_base[GRU_MAX_BLADES] __read_mostly;
50 unsigned long gru_start_paddr __read_mostly;
51 void *gru_start_vaddr __read_mostly;
52 unsigned long gru_end_paddr __read_mostly;
53 unsigned int gru_max_gids __read_mostly;
54 struct gru_stats_s gru_stats;
55
56 /* Guaranteed user available resources on each node */
57 static int max_user_cbrs, max_user_dsr_bytes;
58
59 static struct miscdevice gru_miscdev;
60
61
62 /*
63  * gru_vma_close
64  *
65  * Called when unmapping a device mapping. Frees all gru resources
66  * and tables belonging to the vma.
67  */
68 static void gru_vma_close(struct vm_area_struct *vma)
69 {
70         struct gru_vma_data *vdata;
71         struct gru_thread_state *gts;
72         struct list_head *entry, *next;
73
74         if (!vma->vm_private_data)
75                 return;
76
77         vdata = vma->vm_private_data;
78         vma->vm_private_data = NULL;
79         gru_dbg(grudev, "vma %p, file %p, vdata %p\n", vma, vma->vm_file,
80                                 vdata);
81         list_for_each_safe(entry, next, &vdata->vd_head) {
82                 gts =
83                     list_entry(entry, struct gru_thread_state, ts_next);
84                 list_del(&gts->ts_next);
85                 mutex_lock(&gts->ts_ctxlock);
86                 if (gts->ts_gru)
87                         gru_unload_context(gts, 0);
88                 mutex_unlock(&gts->ts_ctxlock);
89                 gts_drop(gts);
90         }
91         kfree(vdata);
92         STAT(vdata_free);
93 }
94
95 /*
96  * gru_file_mmap
97  *
98  * Called when mmapping the device.  Initializes the vma with a fault handler
99  * and private data structure necessary to allocate, track, and free the
100  * underlying pages.
101  */
102 static int gru_file_mmap(struct file *file, struct vm_area_struct *vma)
103 {
104         if ((vma->vm_flags & (VM_SHARED | VM_WRITE)) != (VM_SHARED | VM_WRITE))
105                 return -EPERM;
106
107         if (vma->vm_start & (GRU_GSEG_PAGESIZE - 1) ||
108                                 vma->vm_end & (GRU_GSEG_PAGESIZE - 1))
109                 return -EINVAL;
110
111         vma->vm_flags |=
112             (VM_IO | VM_DONTCOPY | VM_LOCKED | VM_DONTEXPAND | VM_PFNMAP |
113                         VM_RESERVED);
114         vma->vm_page_prot = PAGE_SHARED;
115         vma->vm_ops = &gru_vm_ops;
116
117         vma->vm_private_data = gru_alloc_vma_data(vma, 0);
118         if (!vma->vm_private_data)
119                 return -ENOMEM;
120
121         gru_dbg(grudev, "file %p, vaddr 0x%lx, vma %p, vdata %p\n",
122                 file, vma->vm_start, vma, vma->vm_private_data);
123         return 0;
124 }
125
126 /*
127  * Create a new GRU context
128  */
129 static int gru_create_new_context(unsigned long arg)
130 {
131         struct gru_create_context_req req;
132         struct vm_area_struct *vma;
133         struct gru_vma_data *vdata;
134         int ret = -EINVAL;
135
136         if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
137                 return -EFAULT;
138
139         if (req.data_segment_bytes > max_user_dsr_bytes)
140                 return -EINVAL;
141         if (req.control_blocks > max_user_cbrs || !req.maximum_thread_count)
142                 return -EINVAL;
143
144         if (!(req.options & GRU_OPT_MISS_MASK))
145                 req.options |= GRU_OPT_MISS_FMM_INTR;
146
147         down_write(&current->mm->mmap_sem);
148         vma = gru_find_vma(req.gseg);
149         if (vma) {
150                 vdata = vma->vm_private_data;
151                 vdata->vd_user_options = req.options;
152                 vdata->vd_dsr_au_count =
153                     GRU_DS_BYTES_TO_AU(req.data_segment_bytes);
154                 vdata->vd_cbr_au_count = GRU_CB_COUNT_TO_AU(req.control_blocks);
155                 vdata->vd_tlb_preload_count = req.tlb_preload_count;
156                 ret = 0;
157         }
158         up_write(&current->mm->mmap_sem);
159
160         return ret;
161 }
162
163 /*
164  * Get GRU configuration info (temp - for emulator testing)
165  */
166 static long gru_get_config_info(unsigned long arg)
167 {
168         struct gru_config_info info;
169         int nodesperblade;
170
171         if (num_online_nodes() > 1 &&
172                         (uv_node_to_blade_id(1) == uv_node_to_blade_id(0)))
173                 nodesperblade = 2;
174         else
175                 nodesperblade = 1;
176         info.cpus = num_online_cpus();
177         info.nodes = num_online_nodes();
178         info.blades = info.nodes / nodesperblade;
179         info.chiplets = GRU_CHIPLETS_PER_BLADE * info.blades;
180
181         if (copy_to_user((void __user *)arg, &info, sizeof(info)))
182                 return -EFAULT;
183         return 0;
184 }
185
186 /*
187  * gru_file_unlocked_ioctl
188  *
189  * Called to update file attributes via IOCTL calls.
190  */
191 static long gru_file_unlocked_ioctl(struct file *file, unsigned int req,
192                                     unsigned long arg)
193 {
194         int err = -EBADRQC;
195
196         gru_dbg(grudev, "file %p, req 0x%x, 0x%lx\n", file, req, arg);
197
198         switch (req) {
199         case GRU_CREATE_CONTEXT:
200                 err = gru_create_new_context(arg);
201                 break;
202         case GRU_SET_CONTEXT_OPTION:
203                 err = gru_set_context_option(arg);
204                 break;
205         case GRU_USER_GET_EXCEPTION_DETAIL:
206                 err = gru_get_exception_detail(arg);
207                 break;
208         case GRU_USER_UNLOAD_CONTEXT:
209                 err = gru_user_unload_context(arg);
210                 break;
211         case GRU_USER_FLUSH_TLB:
212                 err = gru_user_flush_tlb(arg);
213                 break;
214         case GRU_USER_CALL_OS:
215                 err = gru_handle_user_call_os(arg);
216                 break;
217         case GRU_GET_GSEG_STATISTICS:
218                 err = gru_get_gseg_statistics(arg);
219                 break;
220         case GRU_KTEST:
221                 err = gru_ktest(arg);
222                 break;
223         case GRU_GET_CONFIG_INFO:
224                 err = gru_get_config_info(arg);
225                 break;
226         case GRU_DUMP_CHIPLET_STATE:
227                 err = gru_dump_chiplet_request(arg);
228                 break;
229         }
230         return err;
231 }
232
233 /*
234  * Called at init time to build tables for all GRUs that are present in the
235  * system.
236  */
237 static void gru_init_chiplet(struct gru_state *gru, unsigned long paddr,
238                              void *vaddr, int blade_id, int chiplet_id)
239 {
240         spin_lock_init(&gru->gs_lock);
241         spin_lock_init(&gru->gs_asid_lock);
242         gru->gs_gru_base_paddr = paddr;
243         gru->gs_gru_base_vaddr = vaddr;
244         gru->gs_gid = blade_id * GRU_CHIPLETS_PER_BLADE + chiplet_id;
245         gru->gs_blade = gru_base[blade_id];
246         gru->gs_blade_id = blade_id;
247         gru->gs_chiplet_id = chiplet_id;
248         gru->gs_cbr_map = (GRU_CBR_AU == 64) ? ~0 : (1UL << GRU_CBR_AU) - 1;
249         gru->gs_dsr_map = (1UL << GRU_DSR_AU) - 1;
250         gru->gs_asid_limit = MAX_ASID;
251         gru_tgh_flush_init(gru);
252         if (gru->gs_gid >= gru_max_gids)
253                 gru_max_gids = gru->gs_gid + 1;
254         gru_dbg(grudev, "bid %d, gid %d, vaddr %p (0x%lx)\n",
255                 blade_id, gru->gs_gid, gru->gs_gru_base_vaddr,
256                 gru->gs_gru_base_paddr);
257 }
258
259 static int gru_init_tables(unsigned long gru_base_paddr, void *gru_base_vaddr)
260 {
261         int pnode, nid, bid, chip;
262         int cbrs, dsrbytes, n;
263         int order = get_order(sizeof(struct gru_blade_state));
264         struct page *page;
265         struct gru_state *gru;
266         unsigned long paddr;
267         void *vaddr;
268
269         max_user_cbrs = GRU_NUM_CB;
270         max_user_dsr_bytes = GRU_NUM_DSR_BYTES;
271         for_each_possible_blade(bid) {
272                 pnode = uv_blade_to_pnode(bid);
273                 nid = uv_blade_to_memory_nid(bid);/* -1 if no memory on blade */
274                 page = alloc_pages_node(nid, GFP_KERNEL, order);
275                 if (!page)
276                         goto fail;
277                 gru_base[bid] = page_address(page);
278                 memset(gru_base[bid], 0, sizeof(struct gru_blade_state));
279                 gru_base[bid]->bs_lru_gru = &gru_base[bid]->bs_grus[0];
280                 spin_lock_init(&gru_base[bid]->bs_lock);
281                 init_rwsem(&gru_base[bid]->bs_kgts_sema);
282
283                 dsrbytes = 0;
284                 cbrs = 0;
285                 for (gru = gru_base[bid]->bs_grus, chip = 0;
286                                 chip < GRU_CHIPLETS_PER_BLADE;
287                                 chip++, gru++) {
288                         paddr = gru_chiplet_paddr(gru_base_paddr, pnode, chip);
289                         vaddr = gru_chiplet_vaddr(gru_base_vaddr, pnode, chip);
290                         gru_init_chiplet(gru, paddr, vaddr, bid, chip);
291                         n = hweight64(gru->gs_cbr_map) * GRU_CBR_AU_SIZE;
292                         cbrs = max(cbrs, n);
293                         n = hweight64(gru->gs_dsr_map) * GRU_DSR_AU_BYTES;
294                         dsrbytes = max(dsrbytes, n);
295                 }
296                 max_user_cbrs = min(max_user_cbrs, cbrs);
297                 max_user_dsr_bytes = min(max_user_dsr_bytes, dsrbytes);
298         }
299
300         return 0;
301
302 fail:
303         for (bid--; bid >= 0; bid--)
304                 free_pages((unsigned long)gru_base[bid], order);
305         return -ENOMEM;
306 }
307
308 static void gru_free_tables(void)
309 {
310         int bid;
311         int order = get_order(sizeof(struct gru_state) *
312                               GRU_CHIPLETS_PER_BLADE);
313
314         for (bid = 0; bid < GRU_MAX_BLADES; bid++)
315                 free_pages((unsigned long)gru_base[bid], order);
316 }
317
318 static unsigned long gru_chiplet_cpu_to_mmr(int chiplet, int cpu, int *corep)
319 {
320         unsigned long mmr = 0;
321         int core;
322
323         /*
324          * We target the cores of a blade and not the hyperthreads themselves.
325          * There is a max of 8 cores per socket and 2 sockets per blade,
326          * making for a max total of 16 cores (i.e., 16 CPUs without
327          * hyperthreading and 32 CPUs with hyperthreading).
328          */
329         core = uv_cpu_core_number(cpu) + UV_MAX_INT_CORES * uv_cpu_socket_number(cpu);
330         if (core >= GRU_NUM_TFM || uv_cpu_ht_number(cpu))
331                 return 0;
332
333         if (chiplet == 0) {
334                 mmr = UVH_GR0_TLB_INT0_CONFIG +
335                     core * (UVH_GR0_TLB_INT1_CONFIG - UVH_GR0_TLB_INT0_CONFIG);
336         } else if (chiplet == 1) {
337                 mmr = UVH_GR1_TLB_INT0_CONFIG +
338                     core * (UVH_GR1_TLB_INT1_CONFIG - UVH_GR1_TLB_INT0_CONFIG);
339         } else {
340                 BUG();
341         }
342
343         *corep = core;
344         return mmr;
345 }
346
347 #ifdef CONFIG_IA64
348
349 static int gru_irq_count[GRU_CHIPLETS_PER_BLADE];
350
351 static void gru_noop(unsigned int irq)
352 {
353 }
354
355 static struct irq_chip gru_chip[GRU_CHIPLETS_PER_BLADE] = {
356         [0 ... GRU_CHIPLETS_PER_BLADE - 1] {
357                 .mask           = gru_noop,
358                 .unmask         = gru_noop,
359                 .ack            = gru_noop
360         }
361 };
362
363 static int gru_chiplet_setup_tlb_irq(int chiplet, char *irq_name,
364                         irq_handler_t irq_handler, int cpu, int blade)
365 {
366         unsigned long mmr;
367         int irq = IRQ_GRU + chiplet;
368         int ret, core;
369
370         mmr = gru_chiplet_cpu_to_mmr(chiplet, cpu, &core);
371         if (mmr == 0)
372                 return 0;
373
374         if (gru_irq_count[chiplet] == 0) {
375                 gru_chip[chiplet].name = irq_name;
376                 ret = set_irq_chip(irq, &gru_chip[chiplet]);
377                 if (ret) {
378                         printk(KERN_ERR "%s: set_irq_chip failed, errno=%d\n",
379                                GRU_DRIVER_ID_STR, -ret);
380                         return ret;
381                 }
382
383                 ret = request_irq(irq, irq_handler, 0, irq_name, NULL);
384                 if (ret) {
385                         printk(KERN_ERR "%s: request_irq failed, errno=%d\n",
386                                GRU_DRIVER_ID_STR, -ret);
387                         return ret;
388                 }
389         }
390         gru_irq_count[chiplet]++;
391
392         return 0;
393 }
394
395 static void gru_chiplet_teardown_tlb_irq(int chiplet, int cpu, int blade)
396 {
397         unsigned long mmr;
398         int core, irq = IRQ_GRU + chiplet;
399
400         if (gru_irq_count[chiplet] == 0)
401                 return;
402
403         mmr = gru_chiplet_cpu_to_mmr(chiplet, cpu, &core);
404         if (mmr == 0)
405                 return;
406
407         if (--gru_irq_count[chiplet] == 0)
408                 free_irq(irq, NULL);
409 }
410
411 #elif defined CONFIG_X86_64
412
413 static int gru_chiplet_setup_tlb_irq(int chiplet, char *irq_name,
414                         irq_handler_t irq_handler, int cpu, int blade)
415 {
416         unsigned long mmr;
417         int irq, core;
418         int ret;
419
420         mmr = gru_chiplet_cpu_to_mmr(chiplet, cpu, &core);
421         if (mmr == 0)
422                 return 0;
423
424         irq = uv_setup_irq(irq_name, cpu, blade, mmr, UV_AFFINITY_CPU);
425         if (irq < 0) {
426                 printk(KERN_ERR "%s: uv_setup_irq failed, errno=%d\n",
427                        GRU_DRIVER_ID_STR, -irq);
428                 return irq;
429         }
430
431         ret = request_irq(irq, irq_handler, 0, irq_name, NULL);
432         if (ret) {
433                 uv_teardown_irq(irq);
434                 printk(KERN_ERR "%s: request_irq failed, errno=%d\n",
435                        GRU_DRIVER_ID_STR, -ret);
436                 return ret;
437         }
438         gru_base[blade]->bs_grus[chiplet].gs_irq[core] = irq;
439         return 0;
440 }
441
442 static void gru_chiplet_teardown_tlb_irq(int chiplet, int cpu, int blade)
443 {
444         int irq, core;
445         unsigned long mmr;
446
447         mmr = gru_chiplet_cpu_to_mmr(chiplet, cpu, &core);
448         if (mmr) {
449                 irq = gru_base[blade]->bs_grus[chiplet].gs_irq[core];
450                 if (irq) {
451                         free_irq(irq, NULL);
452                         uv_teardown_irq(irq);
453                 }
454         }
455 }
456
457 #endif
458
459 static void gru_teardown_tlb_irqs(void)
460 {
461         int blade;
462         int cpu;
463
464         for_each_online_cpu(cpu) {
465                 blade = uv_cpu_to_blade_id(cpu);
466                 gru_chiplet_teardown_tlb_irq(0, cpu, blade);
467                 gru_chiplet_teardown_tlb_irq(1, cpu, blade);
468         }
469         for_each_possible_blade(blade) {
470                 if (uv_blade_nr_possible_cpus(blade))
471                         continue;
472                 gru_chiplet_teardown_tlb_irq(0, 0, blade);
473                 gru_chiplet_teardown_tlb_irq(1, 0, blade);
474         }
475 }
476
477 static int gru_setup_tlb_irqs(void)
478 {
479         int blade;
480         int cpu;
481         int ret;
482
483         for_each_online_cpu(cpu) {
484                 blade = uv_cpu_to_blade_id(cpu);
485                 ret = gru_chiplet_setup_tlb_irq(0, "GRU0_TLB", gru0_intr, cpu, blade);
486                 if (ret != 0)
487                         goto exit1;
488
489                 ret = gru_chiplet_setup_tlb_irq(1, "GRU1_TLB", gru1_intr, cpu, blade);
490                 if (ret != 0)
491                         goto exit1;
492         }
493         for_each_possible_blade(blade) {
494                 if (uv_blade_nr_possible_cpus(blade))
495                         continue;
496                 ret = gru_chiplet_setup_tlb_irq(0, "GRU0_TLB", gru_intr_mblade, 0, blade);
497                 if (ret != 0)
498                         goto exit1;
499
500                 ret = gru_chiplet_setup_tlb_irq(1, "GRU1_TLB", gru_intr_mblade, 0, blade);
501                 if (ret != 0)
502                         goto exit1;
503         }
504
505         return 0;
506
507 exit1:
508         gru_teardown_tlb_irqs();
509         return ret;
510 }
511
512 /*
513  * gru_init
514  *
515  * Called at boot or module load time to initialize the GRUs.
516  */
517 static int __init gru_init(void)
518 {
519         int ret;
520
521         if (!is_uv_system())
522                 return 0;
523
524 #if defined CONFIG_IA64
525         gru_start_paddr = 0xd000000000UL; /* ZZZZZZZZZZZZZZZZZZZ fixme */
526 #else
527         gru_start_paddr = uv_read_local_mmr(UVH_RH_GAM_GRU_OVERLAY_CONFIG_MMR) &
528                                 0x7fffffffffffUL;
529 #endif
530         gru_start_vaddr = __va(gru_start_paddr);
531         gru_end_paddr = gru_start_paddr + GRU_MAX_BLADES * GRU_SIZE;
532         printk(KERN_INFO "GRU space: 0x%lx - 0x%lx\n",
533                gru_start_paddr, gru_end_paddr);
534         ret = misc_register(&gru_miscdev);
535         if (ret) {
536                 printk(KERN_ERR "%s: misc_register failed\n",
537                        GRU_DRIVER_ID_STR);
538                 goto exit0;
539         }
540
541         ret = gru_proc_init();
542         if (ret) {
543                 printk(KERN_ERR "%s: proc init failed\n", GRU_DRIVER_ID_STR);
544                 goto exit1;
545         }
546
547         ret = gru_init_tables(gru_start_paddr, gru_start_vaddr);
548         if (ret) {
549                 printk(KERN_ERR "%s: init tables failed\n", GRU_DRIVER_ID_STR);
550                 goto exit2;
551         }
552
553         ret = gru_setup_tlb_irqs();
554         if (ret != 0)
555                 goto exit3;
556
557         gru_kservices_init();
558
559         printk(KERN_INFO "%s: v%s\n", GRU_DRIVER_ID_STR,
560                GRU_DRIVER_VERSION_STR);
561         return 0;
562
563 exit3:
564         gru_free_tables();
565 exit2:
566         gru_proc_exit();
567 exit1:
568         misc_deregister(&gru_miscdev);
569 exit0:
570         return ret;
571
572 }
573
574 static void __exit gru_exit(void)
575 {
576         if (!is_uv_system())
577                 return;
578
579         gru_teardown_tlb_irqs();
580         gru_kservices_exit();
581         gru_free_tables();
582         misc_deregister(&gru_miscdev);
583         gru_proc_exit();
584 }
585
586 static const struct file_operations gru_fops = {
587         .owner          = THIS_MODULE,
588         .unlocked_ioctl = gru_file_unlocked_ioctl,
589         .mmap           = gru_file_mmap,
590 };
591
592 static struct miscdevice gru_miscdev = {
593         .minor          = MISC_DYNAMIC_MINOR,
594         .name           = "gru",
595         .fops           = &gru_fops,
596 };
597
598 const struct vm_operations_struct gru_vm_ops = {
599         .close          = gru_vma_close,
600         .fault          = gru_fault,
601 };
602
603 #ifndef MODULE
604 fs_initcall(gru_init);
605 #else
606 module_init(gru_init);
607 #endif
608 module_exit(gru_exit);
609
610 module_param(gru_options, ulong, 0644);
611 MODULE_PARM_DESC(gru_options, "Various debug options");
612
613 MODULE_AUTHOR("Silicon Graphics, Inc.");
614 MODULE_LICENSE("GPL");
615 MODULE_DESCRIPTION(GRU_DRIVER_ID_STR GRU_DRIVER_VERSION_STR);
616 MODULE_VERSION(GRU_DRIVER_VERSION_STR);
617