Commit | Line | Data |
---|---|---|
9a0deecc JS |
1 | /* |
2 | * SN Platform GRU Driver | |
3 | * | |
4 | * DRIVER TABLE MANAGER + GRU CONTEXT LOAD/UNLOAD | |
5 | * | |
6 | * This file is subject to the terms and conditions of the GNU General Public | |
7 | * License. See the file "COPYING" in the main directory of this archive | |
8 | * for more details. | |
9 | * | |
10 | * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved. | |
11 | */ | |
12 | ||
13 | #include <linux/kernel.h> | |
14 | #include <linux/slab.h> | |
15 | #include <linux/mm.h> | |
16 | #include <linux/spinlock.h> | |
17 | #include <linux/sched.h> | |
18 | #include <linux/device.h> | |
19 | #include <linux/list.h> | |
20 | #include <asm/uv/uv_hub.h> | |
21 | #include "gru.h" | |
22 | #include "grutables.h" | |
23 | #include "gruhandles.h" | |
24 | ||
9ca8e40c | 25 | unsigned long gru_options __read_mostly; |
9a0deecc JS |
26 | |
27 | static struct device_driver gru_driver = { | |
28 | .name = "gru" | |
29 | }; | |
30 | ||
31 | static struct device gru_device = { | |
bb0dc43e | 32 | .init_name = "", |
9a0deecc JS |
33 | .driver = &gru_driver, |
34 | }; | |
35 | ||
36 | struct device *grudev = &gru_device; | |
37 | ||
38 | /* | |
39 | * Select a gru fault map to be used by the current cpu. Note that | |
40 | * multiple cpus may be using the same map. | |
41 | * ZZZ should "shift" be used?? Depends on HT cpu numbering | |
42 | * ZZZ should be inline but did not work on emulator | |
43 | */ | |
44 | int gru_cpu_fault_map_id(void) | |
45 | { | |
46 | return uv_blade_processor_id() % GRU_NUM_TFM; | |
47 | } | |
48 | ||
49 | /*--------- ASID Management ------------------------------------------- | |
50 | * | |
51 | * Initially, assign asids sequentially from MIN_ASID .. MAX_ASID. | |
52 | * Once MAX is reached, flush the TLB & start over. However, | |
53 | * some asids may still be in use. There won't be many (percentage wise) still | |
54 | * in use. Search active contexts & determine the value of the first | |
55 | * asid in use ("x"s below). Set "limit" to this value. | |
56 | * This defines a block of assignable asids. | |
57 | * | |
58 | * When "limit" is reached, search forward from limit+1 and determine the | |
59 | * next block of assignable asids. | |
60 | * | |
61 | * Repeat until MAX_ASID is reached, then start over again. | |
62 | * | |
63 | * Each time MAX_ASID is reached, increment the asid generation. Since | |
64 | * the search for in-use asids only checks contexts with GRUs currently | |
65 | * assigned, asids in some contexts will be missed. Prior to loading | |
66 | * a context, the asid generation of the GTS asid is rechecked. If it | |
67 | * doesn't match the current generation, a new asid will be assigned. | |
68 | * | |
69 | * 0---------------x------------x---------------------x----| | |
70 | * ^-next ^-limit ^-MAX_ASID | |
71 | * | |
72 | * All asid manipulation & context loading/unloading is protected by the | |
73 | * gs_lock. | |
74 | */ | |
75 | ||
76 | /* Hit the asid limit. Start over */ | |
77 | static int gru_wrap_asid(struct gru_state *gru) | |
78 | { | |
43884604 | 79 | gru_dbg(grudev, "gid %d\n", gru->gs_gid); |
9a0deecc JS |
80 | STAT(asid_wrap); |
81 | gru->gs_asid_gen++; | |
9a0deecc JS |
82 | return MIN_ASID; |
83 | } | |
84 | ||
85 | /* Find the next chunk of unused asids */ | |
86 | static int gru_reset_asid_limit(struct gru_state *gru, int asid) | |
87 | { | |
88 | int i, gid, inuse_asid, limit; | |
89 | ||
43884604 | 90 | gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid); |
9a0deecc JS |
91 | STAT(asid_next); |
92 | limit = MAX_ASID; | |
93 | if (asid >= limit) | |
94 | asid = gru_wrap_asid(gru); | |
87419412 | 95 | gru_flush_all_tlb(gru); |
9a0deecc JS |
96 | gid = gru->gs_gid; |
97 | again: | |
98 | for (i = 0; i < GRU_NUM_CCH; i++) { | |
836ce679 | 99 | if (!gru->gs_gts[i] || is_kernel_context(gru->gs_gts[i])) |
9a0deecc JS |
100 | continue; |
101 | inuse_asid = gru->gs_gts[i]->ts_gms->ms_asids[gid].mt_asid; | |
43884604 JS |
102 | gru_dbg(grudev, "gid %d, gts %p, gms %p, inuse 0x%x, cxt %d\n", |
103 | gru->gs_gid, gru->gs_gts[i], gru->gs_gts[i]->ts_gms, | |
104 | inuse_asid, i); | |
9a0deecc JS |
105 | if (inuse_asid == asid) { |
106 | asid += ASID_INC; | |
107 | if (asid >= limit) { | |
108 | /* | |
109 | * empty range: reset the range limit and | |
110 | * start over | |
111 | */ | |
112 | limit = MAX_ASID; | |
113 | if (asid >= MAX_ASID) | |
114 | asid = gru_wrap_asid(gru); | |
115 | goto again; | |
116 | } | |
117 | } | |
118 | ||
119 | if ((inuse_asid > asid) && (inuse_asid < limit)) | |
120 | limit = inuse_asid; | |
121 | } | |
122 | gru->gs_asid_limit = limit; | |
123 | gru->gs_asid = asid; | |
43884604 JS |
124 | gru_dbg(grudev, "gid %d, new asid 0x%x, new_limit 0x%x\n", gru->gs_gid, |
125 | asid, limit); | |
9a0deecc JS |
126 | return asid; |
127 | } | |
128 | ||
129 | /* Assign a new ASID to a thread context. */ | |
130 | static int gru_assign_asid(struct gru_state *gru) | |
131 | { | |
132 | int asid; | |
133 | ||
9a0deecc JS |
134 | gru->gs_asid += ASID_INC; |
135 | asid = gru->gs_asid; | |
136 | if (asid >= gru->gs_asid_limit) | |
137 | asid = gru_reset_asid_limit(gru, asid); | |
9a0deecc | 138 | |
43884604 | 139 | gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid); |
9a0deecc JS |
140 | return asid; |
141 | } | |
142 | ||
143 | /* | |
144 | * Clear n bits in a word. Return a word indicating the bits that were cleared. | |
145 | * Optionally, build an array of chars that contain the bit numbers allocated. | |
146 | */ | |
147 | static unsigned long reserve_resources(unsigned long *p, int n, int mmax, | |
148 | char *idx) | |
149 | { | |
150 | unsigned long bits = 0; | |
151 | int i; | |
152 | ||
153 | do { | |
154 | i = find_first_bit(p, mmax); | |
155 | if (i == mmax) | |
156 | BUG(); | |
157 | __clear_bit(i, p); | |
158 | __set_bit(i, &bits); | |
159 | if (idx) | |
160 | *idx++ = i; | |
161 | } while (--n); | |
162 | return bits; | |
163 | } | |
164 | ||
9ca8e40c | 165 | unsigned long gru_reserve_cb_resources(struct gru_state *gru, int cbr_au_count, |
9a0deecc JS |
166 | char *cbmap) |
167 | { | |
168 | return reserve_resources(&gru->gs_cbr_map, cbr_au_count, GRU_CBR_AU, | |
169 | cbmap); | |
170 | } | |
171 | ||
9ca8e40c | 172 | unsigned long gru_reserve_ds_resources(struct gru_state *gru, int dsr_au_count, |
9a0deecc JS |
173 | char *dsmap) |
174 | { | |
175 | return reserve_resources(&gru->gs_dsr_map, dsr_au_count, GRU_DSR_AU, | |
176 | dsmap); | |
177 | } | |
178 | ||
179 | static void reserve_gru_resources(struct gru_state *gru, | |
180 | struct gru_thread_state *gts) | |
181 | { | |
182 | gru->gs_active_contexts++; | |
183 | gts->ts_cbr_map = | |
9ca8e40c | 184 | gru_reserve_cb_resources(gru, gts->ts_cbr_au_count, |
9a0deecc JS |
185 | gts->ts_cbr_idx); |
186 | gts->ts_dsr_map = | |
9ca8e40c | 187 | gru_reserve_ds_resources(gru, gts->ts_dsr_au_count, NULL); |
9a0deecc JS |
188 | } |
189 | ||
190 | static void free_gru_resources(struct gru_state *gru, | |
191 | struct gru_thread_state *gts) | |
192 | { | |
193 | gru->gs_active_contexts--; | |
194 | gru->gs_cbr_map |= gts->ts_cbr_map; | |
195 | gru->gs_dsr_map |= gts->ts_dsr_map; | |
196 | } | |
197 | ||
198 | /* | |
199 | * Check if a GRU has sufficient free resources to satisfy an allocation | |
200 | * request. Note: GRU locks may or may not be held when this is called. If | |
201 | * not held, recheck after acquiring the appropriate locks. | |
202 | * | |
203 | * Returns 1 if sufficient resources, 0 if not | |
204 | */ | |
205 | static int check_gru_resources(struct gru_state *gru, int cbr_au_count, | |
206 | int dsr_au_count, int max_active_contexts) | |
207 | { | |
208 | return hweight64(gru->gs_cbr_map) >= cbr_au_count | |
209 | && hweight64(gru->gs_dsr_map) >= dsr_au_count | |
210 | && gru->gs_active_contexts < max_active_contexts; | |
211 | } | |
212 | ||
213 | /* | |
214 | * TLB manangment requires tracking all GRU chiplets that have loaded a GSEG | |
215 | * context. | |
216 | */ | |
43884604 JS |
217 | static int gru_load_mm_tracker(struct gru_state *gru, |
218 | struct gru_thread_state *gts) | |
9a0deecc | 219 | { |
43884604 | 220 | struct gru_mm_struct *gms = gts->ts_gms; |
9a0deecc | 221 | struct gru_mm_tracker *asids = &gms->ms_asids[gru->gs_gid]; |
43884604 | 222 | unsigned short ctxbitmap = (1 << gts->ts_ctxnum); |
9a0deecc JS |
223 | int asid; |
224 | ||
225 | spin_lock(&gms->ms_asid_lock); | |
226 | asid = asids->mt_asid; | |
227 | ||
87419412 JS |
228 | spin_lock(&gru->gs_asid_lock); |
229 | if (asid == 0 || (asids->mt_ctxbitmap == 0 && asids->mt_asid_gen != | |
230 | gru->gs_asid_gen)) { | |
9a0deecc JS |
231 | asid = gru_assign_asid(gru); |
232 | asids->mt_asid = asid; | |
233 | asids->mt_asid_gen = gru->gs_asid_gen; | |
234 | STAT(asid_new); | |
235 | } else { | |
236 | STAT(asid_reuse); | |
237 | } | |
87419412 | 238 | spin_unlock(&gru->gs_asid_lock); |
9a0deecc JS |
239 | |
240 | BUG_ON(asids->mt_ctxbitmap & ctxbitmap); | |
241 | asids->mt_ctxbitmap |= ctxbitmap; | |
242 | if (!test_bit(gru->gs_gid, gms->ms_asidmap)) | |
243 | __set_bit(gru->gs_gid, gms->ms_asidmap); | |
244 | spin_unlock(&gms->ms_asid_lock); | |
245 | ||
246 | gru_dbg(grudev, | |
43884604 JS |
247 | "gid %d, gts %p, gms %p, ctxnum %d, asid 0x%x, asidmap 0x%lx\n", |
248 | gru->gs_gid, gts, gms, gts->ts_ctxnum, asid, | |
249 | gms->ms_asidmap[0]); | |
9a0deecc JS |
250 | return asid; |
251 | } | |
252 | ||
253 | static void gru_unload_mm_tracker(struct gru_state *gru, | |
43884604 | 254 | struct gru_thread_state *gts) |
9a0deecc | 255 | { |
43884604 | 256 | struct gru_mm_struct *gms = gts->ts_gms; |
9a0deecc JS |
257 | struct gru_mm_tracker *asids; |
258 | unsigned short ctxbitmap; | |
259 | ||
260 | asids = &gms->ms_asids[gru->gs_gid]; | |
43884604 | 261 | ctxbitmap = (1 << gts->ts_ctxnum); |
9a0deecc | 262 | spin_lock(&gms->ms_asid_lock); |
87419412 | 263 | spin_lock(&gru->gs_asid_lock); |
9a0deecc JS |
264 | BUG_ON((asids->mt_ctxbitmap & ctxbitmap) != ctxbitmap); |
265 | asids->mt_ctxbitmap ^= ctxbitmap; | |
43884604 JS |
266 | gru_dbg(grudev, "gid %d, gts %p, gms %p, ctxnum 0x%d, asidmap 0x%lx\n", |
267 | gru->gs_gid, gts, gms, gts->ts_ctxnum, gms->ms_asidmap[0]); | |
87419412 | 268 | spin_unlock(&gru->gs_asid_lock); |
9a0deecc JS |
269 | spin_unlock(&gms->ms_asid_lock); |
270 | } | |
271 | ||
272 | /* | |
273 | * Decrement the reference count on a GTS structure. Free the structure | |
274 | * if the reference count goes to zero. | |
275 | */ | |
276 | void gts_drop(struct gru_thread_state *gts) | |
277 | { | |
278 | if (gts && atomic_dec_return(>s->ts_refcnt) == 0) { | |
279 | gru_drop_mmu_notifier(gts->ts_gms); | |
280 | kfree(gts); | |
281 | STAT(gts_free); | |
282 | } | |
283 | } | |
284 | ||
285 | /* | |
286 | * Locate the GTS structure for the current thread. | |
287 | */ | |
288 | static struct gru_thread_state *gru_find_current_gts_nolock(struct gru_vma_data | |
289 | *vdata, int tsid) | |
290 | { | |
291 | struct gru_thread_state *gts; | |
292 | ||
293 | list_for_each_entry(gts, &vdata->vd_head, ts_next) | |
294 | if (gts->ts_tsid == tsid) | |
295 | return gts; | |
296 | return NULL; | |
297 | } | |
298 | ||
299 | /* | |
300 | * Allocate a thread state structure. | |
301 | */ | |
364b76df JS |
302 | struct gru_thread_state *gru_alloc_gts(struct vm_area_struct *vma, |
303 | int cbr_au_count, int dsr_au_count, int options, int tsid) | |
9a0deecc JS |
304 | { |
305 | struct gru_thread_state *gts; | |
306 | int bytes; | |
307 | ||
364b76df | 308 | bytes = DSR_BYTES(dsr_au_count) + CBR_BYTES(cbr_au_count); |
9a0deecc JS |
309 | bytes += sizeof(struct gru_thread_state); |
310 | gts = kzalloc(bytes, GFP_KERNEL); | |
311 | if (!gts) | |
312 | return NULL; | |
313 | ||
314 | STAT(gts_alloc); | |
315 | atomic_set(>s->ts_refcnt, 1); | |
316 | mutex_init(>s->ts_ctxlock); | |
364b76df JS |
317 | gts->ts_cbr_au_count = cbr_au_count; |
318 | gts->ts_dsr_au_count = dsr_au_count; | |
319 | gts->ts_user_options = options; | |
9a0deecc | 320 | gts->ts_tsid = tsid; |
9a0deecc | 321 | gts->ts_ctxnum = NULLCTX; |
9a0deecc | 322 | gts->ts_tlb_int_select = -1; |
7b8274e9 | 323 | gts->ts_sizeavail = GRU_SIZEAVAIL(PAGE_SHIFT); |
364b76df JS |
324 | if (vma) { |
325 | gts->ts_mm = current->mm; | |
326 | gts->ts_vma = vma; | |
327 | gts->ts_gms = gru_register_mmu_notifier(); | |
328 | if (!gts->ts_gms) | |
329 | goto err; | |
330 | } | |
9a0deecc | 331 | |
364b76df | 332 | gru_dbg(grudev, "alloc gts %p\n", gts); |
9a0deecc JS |
333 | return gts; |
334 | ||
335 | err: | |
336 | gts_drop(gts); | |
337 | return NULL; | |
338 | } | |
339 | ||
340 | /* | |
341 | * Allocate a vma private data structure. | |
342 | */ | |
343 | struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma, int tsid) | |
344 | { | |
345 | struct gru_vma_data *vdata = NULL; | |
346 | ||
347 | vdata = kmalloc(sizeof(*vdata), GFP_KERNEL); | |
348 | if (!vdata) | |
349 | return NULL; | |
350 | ||
351 | INIT_LIST_HEAD(&vdata->vd_head); | |
352 | spin_lock_init(&vdata->vd_lock); | |
353 | gru_dbg(grudev, "alloc vdata %p\n", vdata); | |
354 | return vdata; | |
355 | } | |
356 | ||
357 | /* | |
358 | * Find the thread state structure for the current thread. | |
359 | */ | |
360 | struct gru_thread_state *gru_find_thread_state(struct vm_area_struct *vma, | |
361 | int tsid) | |
362 | { | |
363 | struct gru_vma_data *vdata = vma->vm_private_data; | |
364 | struct gru_thread_state *gts; | |
365 | ||
366 | spin_lock(&vdata->vd_lock); | |
367 | gts = gru_find_current_gts_nolock(vdata, tsid); | |
368 | spin_unlock(&vdata->vd_lock); | |
369 | gru_dbg(grudev, "vma %p, gts %p\n", vma, gts); | |
370 | return gts; | |
371 | } | |
372 | ||
373 | /* | |
374 | * Allocate a new thread state for a GSEG. Note that races may allow | |
375 | * another thread to race to create a gts. | |
376 | */ | |
377 | struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct *vma, | |
378 | int tsid) | |
379 | { | |
380 | struct gru_vma_data *vdata = vma->vm_private_data; | |
381 | struct gru_thread_state *gts, *ngts; | |
382 | ||
364b76df JS |
383 | gts = gru_alloc_gts(vma, vdata->vd_cbr_au_count, vdata->vd_dsr_au_count, |
384 | vdata->vd_user_options, tsid); | |
9a0deecc JS |
385 | if (!gts) |
386 | return NULL; | |
387 | ||
388 | spin_lock(&vdata->vd_lock); | |
389 | ngts = gru_find_current_gts_nolock(vdata, tsid); | |
390 | if (ngts) { | |
391 | gts_drop(gts); | |
392 | gts = ngts; | |
393 | STAT(gts_double_allocate); | |
394 | } else { | |
395 | list_add(>s->ts_next, &vdata->vd_head); | |
396 | } | |
397 | spin_unlock(&vdata->vd_lock); | |
398 | gru_dbg(grudev, "vma %p, gts %p\n", vma, gts); | |
399 | return gts; | |
400 | } | |
401 | ||
402 | /* | |
403 | * Free the GRU context assigned to the thread state. | |
404 | */ | |
405 | static void gru_free_gru_context(struct gru_thread_state *gts) | |
406 | { | |
407 | struct gru_state *gru; | |
408 | ||
409 | gru = gts->ts_gru; | |
43884604 | 410 | gru_dbg(grudev, "gts %p, gid %d\n", gts, gru->gs_gid); |
9a0deecc JS |
411 | |
412 | spin_lock(&gru->gs_lock); | |
413 | gru->gs_gts[gts->ts_ctxnum] = NULL; | |
414 | free_gru_resources(gru, gts); | |
415 | BUG_ON(test_bit(gts->ts_ctxnum, &gru->gs_context_map) == 0); | |
416 | __clear_bit(gts->ts_ctxnum, &gru->gs_context_map); | |
417 | gts->ts_ctxnum = NULLCTX; | |
418 | gts->ts_gru = NULL; | |
87419412 | 419 | gts->ts_blade = -1; |
9a0deecc JS |
420 | spin_unlock(&gru->gs_lock); |
421 | ||
422 | gts_drop(gts); | |
423 | STAT(free_context); | |
424 | } | |
425 | ||
426 | /* | |
427 | * Prefetching cachelines help hardware performance. | |
9ca8e40c | 428 | * (Strictly a performance enhancement. Not functionally required). |
9a0deecc JS |
429 | */ |
430 | static void prefetch_data(void *p, int num, int stride) | |
431 | { | |
432 | while (num-- > 0) { | |
433 | prefetchw(p); | |
434 | p += stride; | |
435 | } | |
436 | } | |
437 | ||
438 | static inline long gru_copy_handle(void *d, void *s) | |
439 | { | |
440 | memcpy(d, s, GRU_HANDLE_BYTES); | |
441 | return GRU_HANDLE_BYTES; | |
442 | } | |
443 | ||
fe5bb6b0 JS |
444 | static void gru_prefetch_context(void *gseg, void *cb, void *cbe, |
445 | unsigned long cbrmap, unsigned long length) | |
9a0deecc | 446 | { |
9a0deecc JS |
447 | int i, scr; |
448 | ||
9a0deecc JS |
449 | prefetch_data(gseg + GRU_DS_BASE, length / GRU_CACHE_LINE_BYTES, |
450 | GRU_CACHE_LINE_BYTES); | |
451 | ||
9a0deecc JS |
452 | for_each_cbr_in_allocation_map(i, &cbrmap, scr) { |
453 | prefetch_data(cb, 1, GRU_CACHE_LINE_BYTES); | |
454 | prefetch_data(cbe + i * GRU_HANDLE_STRIDE, 1, | |
455 | GRU_CACHE_LINE_BYTES); | |
456 | cb += GRU_HANDLE_STRIDE; | |
457 | } | |
923f7f69 JS |
458 | } |
459 | ||
460 | static void gru_load_context_data(void *save, void *grubase, int ctxnum, | |
461 | unsigned long cbrmap, unsigned long dsrmap) | |
462 | { | |
463 | void *gseg, *cb, *cbe; | |
464 | unsigned long length; | |
465 | int i, scr; | |
9a0deecc | 466 | |
923f7f69 | 467 | gseg = grubase + ctxnum * GRU_GSEG_STRIDE; |
9a0deecc | 468 | cb = gseg + GRU_CB_BASE; |
923f7f69 JS |
469 | cbe = grubase + GRU_CBE_BASE; |
470 | length = hweight64(dsrmap) * GRU_DSR_AU_BYTES; | |
471 | gru_prefetch_context(gseg, cb, cbe, cbrmap, length); | |
472 | ||
9a0deecc JS |
473 | for_each_cbr_in_allocation_map(i, &cbrmap, scr) { |
474 | save += gru_copy_handle(cb, save); | |
475 | save += gru_copy_handle(cbe + i * GRU_HANDLE_STRIDE, save); | |
476 | cb += GRU_HANDLE_STRIDE; | |
477 | } | |
478 | ||
479 | memcpy(gseg + GRU_DS_BASE, save, length); | |
480 | } | |
481 | ||
482 | static void gru_unload_context_data(void *save, void *grubase, int ctxnum, | |
483 | unsigned long cbrmap, unsigned long dsrmap) | |
484 | { | |
485 | void *gseg, *cb, *cbe; | |
486 | unsigned long length; | |
487 | int i, scr; | |
488 | ||
489 | gseg = grubase + ctxnum * GRU_GSEG_STRIDE; | |
9a0deecc JS |
490 | cb = gseg + GRU_CB_BASE; |
491 | cbe = grubase + GRU_CBE_BASE; | |
923f7f69 JS |
492 | length = hweight64(dsrmap) * GRU_DSR_AU_BYTES; |
493 | gru_prefetch_context(gseg, cb, cbe, cbrmap, length); | |
494 | ||
9a0deecc JS |
495 | for_each_cbr_in_allocation_map(i, &cbrmap, scr) { |
496 | save += gru_copy_handle(save, cb); | |
497 | save += gru_copy_handle(save, cbe + i * GRU_HANDLE_STRIDE); | |
498 | cb += GRU_HANDLE_STRIDE; | |
499 | } | |
9a0deecc JS |
500 | memcpy(save, gseg + GRU_DS_BASE, length); |
501 | } | |
502 | ||
503 | void gru_unload_context(struct gru_thread_state *gts, int savestate) | |
504 | { | |
505 | struct gru_state *gru = gts->ts_gru; | |
506 | struct gru_context_configuration_handle *cch; | |
507 | int ctxnum = gts->ts_ctxnum; | |
508 | ||
836ce679 JS |
509 | if (!is_kernel_context(gts)) |
510 | zap_vma_ptes(gts->ts_vma, UGRUADDR(gts), GRU_GSEG_PAGESIZE); | |
9a0deecc JS |
511 | cch = get_cch(gru->gs_gru_base_vaddr, ctxnum); |
512 | ||
43884604 | 513 | gru_dbg(grudev, "gts %p\n", gts); |
9a0deecc JS |
514 | lock_cch_handle(cch); |
515 | if (cch_interrupt_sync(cch)) | |
516 | BUG(); | |
9a0deecc | 517 | |
836ce679 JS |
518 | if (!is_kernel_context(gts)) |
519 | gru_unload_mm_tracker(gru, gts); | |
9a0deecc JS |
520 | if (savestate) |
521 | gru_unload_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, | |
522 | ctxnum, gts->ts_cbr_map, | |
523 | gts->ts_dsr_map); | |
524 | ||
525 | if (cch_deallocate(cch)) | |
526 | BUG(); | |
527 | gts->ts_force_unload = 0; /* ts_force_unload locked by CCH lock */ | |
528 | unlock_cch_handle(cch); | |
529 | ||
530 | gru_free_gru_context(gts); | |
9a0deecc JS |
531 | } |
532 | ||
533 | /* | |
534 | * Load a GRU context by copying it from the thread data structure in memory | |
535 | * to the GRU. | |
536 | */ | |
d57c82b1 | 537 | void gru_load_context(struct gru_thread_state *gts) |
9a0deecc JS |
538 | { |
539 | struct gru_state *gru = gts->ts_gru; | |
540 | struct gru_context_configuration_handle *cch; | |
6e910074 | 541 | int i, err, asid, ctxnum = gts->ts_ctxnum; |
9a0deecc JS |
542 | |
543 | gru_dbg(grudev, "gts %p\n", gts); | |
544 | cch = get_cch(gru->gs_gru_base_vaddr, ctxnum); | |
545 | ||
546 | lock_cch_handle(cch); | |
9a0deecc JS |
547 | cch->tfm_fault_bit_enable = |
548 | (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL | |
549 | || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR); | |
550 | cch->tlb_int_enable = (gts->ts_user_options == GRU_OPT_MISS_FMM_INTR); | |
551 | if (cch->tlb_int_enable) { | |
552 | gts->ts_tlb_int_select = gru_cpu_fault_map_id(); | |
553 | cch->tlb_int_select = gts->ts_tlb_int_select; | |
554 | } | |
555 | cch->tfm_done_bit_enable = 0; | |
6e910074 JS |
556 | cch->dsr_allocation_map = gts->ts_dsr_map; |
557 | cch->cbr_allocation_map = gts->ts_cbr_map; | |
836ce679 JS |
558 | |
559 | if (is_kernel_context(gts)) { | |
560 | cch->unmap_enable = 1; | |
561 | } else { | |
562 | cch->unmap_enable = 0; | |
563 | asid = gru_load_mm_tracker(gru, gts); | |
564 | for (i = 0; i < 8; i++) { | |
565 | cch->asid[i] = asid + i; | |
566 | cch->sizeavail[i] = gts->ts_sizeavail; | |
567 | } | |
6e910074 JS |
568 | } |
569 | ||
570 | err = cch_allocate(cch); | |
9a0deecc JS |
571 | if (err) { |
572 | gru_dbg(grudev, | |
573 | "err %d: cch %p, gts %p, cbr 0x%lx, dsr 0x%lx\n", | |
574 | err, cch, gts, gts->ts_cbr_map, gts->ts_dsr_map); | |
575 | BUG(); | |
576 | } | |
577 | ||
578 | gru_load_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, ctxnum, | |
579 | gts->ts_cbr_map, gts->ts_dsr_map); | |
580 | ||
581 | if (cch_start(cch)) | |
582 | BUG(); | |
583 | unlock_cch_handle(cch); | |
9a0deecc JS |
584 | } |
585 | ||
586 | /* | |
587 | * Update fields in an active CCH: | |
588 | * - retarget interrupts on local blade | |
7b8274e9 | 589 | * - update sizeavail mask |
9a0deecc JS |
590 | * - force a delayed context unload by clearing the CCH asids. This |
591 | * forces TLB misses for new GRU instructions. The context is unloaded | |
592 | * when the next TLB miss occurs. | |
593 | */ | |
7b8274e9 | 594 | int gru_update_cch(struct gru_thread_state *gts, int force_unload) |
9a0deecc JS |
595 | { |
596 | struct gru_context_configuration_handle *cch; | |
597 | struct gru_state *gru = gts->ts_gru; | |
598 | int i, ctxnum = gts->ts_ctxnum, ret = 0; | |
599 | ||
600 | cch = get_cch(gru->gs_gru_base_vaddr, ctxnum); | |
601 | ||
602 | lock_cch_handle(cch); | |
603 | if (cch->state == CCHSTATE_ACTIVE) { | |
604 | if (gru->gs_gts[gts->ts_ctxnum] != gts) | |
605 | goto exit; | |
606 | if (cch_interrupt(cch)) | |
607 | BUG(); | |
7b8274e9 JS |
608 | if (!force_unload) { |
609 | for (i = 0; i < 8; i++) | |
610 | cch->sizeavail[i] = gts->ts_sizeavail; | |
611 | gts->ts_tlb_int_select = gru_cpu_fault_map_id(); | |
612 | cch->tlb_int_select = gru_cpu_fault_map_id(); | |
cd1334f0 | 613 | cch->tfm_fault_bit_enable = |
d57c82b1 JS |
614 | (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL |
615 | || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR); | |
9a0deecc JS |
616 | } else { |
617 | for (i = 0; i < 8; i++) | |
618 | cch->asid[i] = 0; | |
619 | cch->tfm_fault_bit_enable = 0; | |
620 | cch->tlb_int_enable = 0; | |
621 | gts->ts_force_unload = 1; | |
622 | } | |
623 | if (cch_start(cch)) | |
624 | BUG(); | |
625 | ret = 1; | |
626 | } | |
627 | exit: | |
628 | unlock_cch_handle(cch); | |
629 | return ret; | |
630 | } | |
631 | ||
632 | /* | |
633 | * Update CCH tlb interrupt select. Required when all the following is true: | |
634 | * - task's GRU context is loaded into a GRU | |
635 | * - task is using interrupt notification for TLB faults | |
636 | * - task has migrated to a different cpu on the same blade where | |
637 | * it was previously running. | |
638 | */ | |
639 | static int gru_retarget_intr(struct gru_thread_state *gts) | |
640 | { | |
641 | if (gts->ts_tlb_int_select < 0 | |
642 | || gts->ts_tlb_int_select == gru_cpu_fault_map_id()) | |
643 | return 0; | |
644 | ||
645 | gru_dbg(grudev, "retarget from %d to %d\n", gts->ts_tlb_int_select, | |
646 | gru_cpu_fault_map_id()); | |
7b8274e9 | 647 | return gru_update_cch(gts, 0); |
9a0deecc JS |
648 | } |
649 | ||
650 | ||
651 | /* | |
652 | * Insufficient GRU resources available on the local blade. Steal a context from | |
653 | * a process. This is a hack until a _real_ resource scheduler is written.... | |
654 | */ | |
655 | #define next_ctxnum(n) ((n) < GRU_NUM_CCH - 2 ? (n) + 1 : 0) | |
656 | #define next_gru(b, g) (((g) < &(b)->bs_grus[GRU_CHIPLETS_PER_BLADE - 1]) ? \ | |
657 | ((g)+1) : &(b)->bs_grus[0]) | |
658 | ||
836ce679 JS |
659 | static int is_gts_stealable(struct gru_thread_state *gts, |
660 | struct gru_blade_state *bs) | |
661 | { | |
662 | if (is_kernel_context(gts)) | |
663 | return down_write_trylock(&bs->bs_kgts_sema); | |
664 | else | |
665 | return mutex_trylock(>s->ts_ctxlock); | |
666 | } | |
667 | ||
668 | static void gts_stolen(struct gru_thread_state *gts, | |
669 | struct gru_blade_state *bs) | |
670 | { | |
671 | if (is_kernel_context(gts)) { | |
672 | up_write(&bs->bs_kgts_sema); | |
673 | STAT(steal_kernel_context); | |
674 | } else { | |
675 | mutex_unlock(>s->ts_ctxlock); | |
676 | STAT(steal_user_context); | |
677 | } | |
678 | } | |
679 | ||
d57c82b1 | 680 | void gru_steal_context(struct gru_thread_state *gts, int blade_id) |
9a0deecc JS |
681 | { |
682 | struct gru_blade_state *blade; | |
683 | struct gru_state *gru, *gru0; | |
684 | struct gru_thread_state *ngts = NULL; | |
685 | int ctxnum, ctxnum0, flag = 0, cbr, dsr; | |
686 | ||
687 | cbr = gts->ts_cbr_au_count; | |
688 | dsr = gts->ts_dsr_au_count; | |
689 | ||
364b76df | 690 | blade = gru_base[blade_id]; |
9a0deecc JS |
691 | spin_lock(&blade->bs_lock); |
692 | ||
693 | ctxnum = next_ctxnum(blade->bs_lru_ctxnum); | |
694 | gru = blade->bs_lru_gru; | |
695 | if (ctxnum == 0) | |
696 | gru = next_gru(blade, gru); | |
697 | ctxnum0 = ctxnum; | |
698 | gru0 = gru; | |
699 | while (1) { | |
700 | if (check_gru_resources(gru, cbr, dsr, GRU_NUM_CCH)) | |
701 | break; | |
702 | spin_lock(&gru->gs_lock); | |
703 | for (; ctxnum < GRU_NUM_CCH; ctxnum++) { | |
704 | if (flag && gru == gru0 && ctxnum == ctxnum0) | |
705 | break; | |
706 | ngts = gru->gs_gts[ctxnum]; | |
707 | /* | |
708 | * We are grabbing locks out of order, so trylock is | |
709 | * needed. GTSs are usually not locked, so the odds of | |
710 | * success are high. If trylock fails, try to steal a | |
711 | * different GSEG. | |
712 | */ | |
836ce679 | 713 | if (ngts && is_gts_stealable(ngts, blade)) |
9a0deecc JS |
714 | break; |
715 | ngts = NULL; | |
716 | flag = 1; | |
717 | } | |
718 | spin_unlock(&gru->gs_lock); | |
719 | if (ngts || (flag && gru == gru0 && ctxnum == ctxnum0)) | |
720 | break; | |
721 | ctxnum = 0; | |
722 | gru = next_gru(blade, gru); | |
723 | } | |
724 | blade->bs_lru_gru = gru; | |
725 | blade->bs_lru_ctxnum = ctxnum; | |
726 | spin_unlock(&blade->bs_lock); | |
9a0deecc JS |
727 | |
728 | if (ngts) { | |
9a0deecc | 729 | ngts->ts_steal_jiffies = jiffies; |
836ce679 JS |
730 | gru_unload_context(ngts, is_kernel_context(ngts) ? 0 : 1); |
731 | gts_stolen(ngts, blade); | |
9a0deecc JS |
732 | } else { |
733 | STAT(steal_context_failed); | |
734 | } | |
735 | gru_dbg(grudev, | |
43884604 | 736 | "stole gid %d, ctxnum %d from gts %p. Need cb %d, ds %d;" |
9a0deecc JS |
737 | " avail cb %ld, ds %ld\n", |
738 | gru->gs_gid, ctxnum, ngts, cbr, dsr, hweight64(gru->gs_cbr_map), | |
739 | hweight64(gru->gs_dsr_map)); | |
740 | } | |
741 | ||
742 | /* | |
743 | * Scan the GRUs on the local blade & assign a GRU context. | |
744 | */ | |
d57c82b1 | 745 | struct gru_state *gru_assign_gru_context(struct gru_thread_state *gts, |
364b76df | 746 | int blade) |
9a0deecc JS |
747 | { |
748 | struct gru_state *gru, *grux; | |
749 | int i, max_active_contexts; | |
750 | ||
9a0deecc JS |
751 | |
752 | again: | |
753 | gru = NULL; | |
754 | max_active_contexts = GRU_NUM_CCH; | |
364b76df | 755 | for_each_gru_on_blade(grux, blade, i) { |
9a0deecc JS |
756 | if (check_gru_resources(grux, gts->ts_cbr_au_count, |
757 | gts->ts_dsr_au_count, | |
758 | max_active_contexts)) { | |
759 | gru = grux; | |
760 | max_active_contexts = grux->gs_active_contexts; | |
761 | if (max_active_contexts == 0) | |
762 | break; | |
763 | } | |
764 | } | |
765 | ||
766 | if (gru) { | |
767 | spin_lock(&gru->gs_lock); | |
768 | if (!check_gru_resources(gru, gts->ts_cbr_au_count, | |
769 | gts->ts_dsr_au_count, GRU_NUM_CCH)) { | |
770 | spin_unlock(&gru->gs_lock); | |
771 | goto again; | |
772 | } | |
773 | reserve_gru_resources(gru, gts); | |
774 | gts->ts_gru = gru; | |
87419412 | 775 | gts->ts_blade = gru->gs_blade_id; |
9a0deecc JS |
776 | gts->ts_ctxnum = |
777 | find_first_zero_bit(&gru->gs_context_map, GRU_NUM_CCH); | |
778 | BUG_ON(gts->ts_ctxnum == GRU_NUM_CCH); | |
779 | atomic_inc(>s->ts_refcnt); | |
780 | gru->gs_gts[gts->ts_ctxnum] = gts; | |
781 | __set_bit(gts->ts_ctxnum, &gru->gs_context_map); | |
782 | spin_unlock(&gru->gs_lock); | |
783 | ||
784 | STAT(assign_context); | |
785 | gru_dbg(grudev, | |
43884604 | 786 | "gseg %p, gts %p, gid %d, ctx %d, cbr %d, dsr %d\n", |
9a0deecc JS |
787 | gseg_virtual_address(gts->ts_gru, gts->ts_ctxnum), gts, |
788 | gts->ts_gru->gs_gid, gts->ts_ctxnum, | |
789 | gts->ts_cbr_au_count, gts->ts_dsr_au_count); | |
790 | } else { | |
791 | gru_dbg(grudev, "failed to allocate a GTS %s\n", ""); | |
792 | STAT(assign_context_failed); | |
793 | } | |
794 | ||
9a0deecc JS |
795 | return gru; |
796 | } | |
797 | ||
798 | /* | |
799 | * gru_nopage | |
800 | * | |
801 | * Map the user's GRU segment | |
9ca8e40c JS |
802 | * |
803 | * Note: gru segments alway mmaped on GRU_GSEG_PAGESIZE boundaries. | |
9a0deecc JS |
804 | */ |
805 | int gru_fault(struct vm_area_struct *vma, struct vm_fault *vmf) | |
806 | { | |
807 | struct gru_thread_state *gts; | |
808 | unsigned long paddr, vaddr; | |
364b76df | 809 | int blade_id; |
9a0deecc JS |
810 | |
811 | vaddr = (unsigned long)vmf->virtual_address; | |
812 | gru_dbg(grudev, "vma %p, vaddr 0x%lx (0x%lx)\n", | |
813 | vma, vaddr, GSEG_BASE(vaddr)); | |
814 | STAT(nopfn); | |
815 | ||
9ca8e40c | 816 | /* The following check ensures vaddr is a valid address in the VMA */ |
9a0deecc JS |
817 | gts = gru_find_thread_state(vma, TSID(vaddr, vma)); |
818 | if (!gts) | |
819 | return VM_FAULT_SIGBUS; | |
820 | ||
821 | again: | |
9a0deecc | 822 | mutex_lock(>s->ts_ctxlock); |
fe5bb6b0 | 823 | preempt_disable(); |
364b76df JS |
824 | blade_id = uv_numa_blade_id(); |
825 | ||
9a0deecc | 826 | if (gts->ts_gru) { |
364b76df | 827 | if (gts->ts_gru->gs_blade_id != blade_id) { |
9a0deecc JS |
828 | STAT(migrated_nopfn_unload); |
829 | gru_unload_context(gts, 1); | |
830 | } else { | |
831 | if (gru_retarget_intr(gts)) | |
832 | STAT(migrated_nopfn_retarget); | |
833 | } | |
834 | } | |
835 | ||
836 | if (!gts->ts_gru) { | |
836ce679 | 837 | STAT(load_user_context); |
364b76df | 838 | if (!gru_assign_gru_context(gts, blade_id)) { |
9a0deecc | 839 | preempt_enable(); |
364b76df JS |
840 | mutex_unlock(>s->ts_ctxlock); |
841 | set_current_state(TASK_INTERRUPTIBLE); | |
9a0deecc | 842 | schedule_timeout(GRU_ASSIGN_DELAY); /* true hack ZZZ */ |
364b76df | 843 | blade_id = uv_numa_blade_id(); |
9a0deecc | 844 | if (gts->ts_steal_jiffies + GRU_STEAL_DELAY < jiffies) |
364b76df | 845 | gru_steal_context(gts, blade_id); |
9a0deecc JS |
846 | goto again; |
847 | } | |
848 | gru_load_context(gts); | |
849 | paddr = gseg_physical_address(gts->ts_gru, gts->ts_ctxnum); | |
850 | remap_pfn_range(vma, vaddr & ~(GRU_GSEG_PAGESIZE - 1), | |
851 | paddr >> PAGE_SHIFT, GRU_GSEG_PAGESIZE, | |
852 | vma->vm_page_prot); | |
853 | } | |
854 | ||
9a0deecc | 855 | preempt_enable(); |
364b76df | 856 | mutex_unlock(>s->ts_ctxlock); |
9a0deecc JS |
857 | |
858 | return VM_FAULT_NOPAGE; | |
859 | } | |
860 |