xref: /openbmc/linux/drivers/misc/sgi-gru/grumain.c (revision b9b77222)
1 /*
2  * SN Platform GRU Driver
3  *
4  *            DRIVER TABLE MANAGER + GRU CONTEXT LOAD/UNLOAD
5  *
6  *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
7  *
8  *  This program is free software; you can redistribute it and/or modify
9  *  it under the terms of the GNU General Public License as published by
10  *  the Free Software Foundation; either version 2 of the License, or
11  *  (at your option) any later version.
12  *
13  *  This program is distributed in the hope that it will be useful,
14  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
15  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  *  GNU General Public License for more details.
17  *
18  *  You should have received a copy of the GNU General Public License
19  *  along with this program; if not, write to the Free Software
20  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
21  */
22 
23 #include <linux/kernel.h>
24 #include <linux/slab.h>
25 #include <linux/mm.h>
26 #include <linux/spinlock.h>
27 #include <linux/sched.h>
28 #include <linux/device.h>
29 #include <linux/list.h>
30 #include <linux/err.h>
31 #include <linux/prefetch.h>
32 #include <asm/uv/uv_hub.h>
33 #include "gru.h"
34 #include "grutables.h"
35 #include "gruhandles.h"
36 
37 unsigned long gru_options __read_mostly;
38 
39 static struct device_driver gru_driver = {
40 	.name = "gru"
41 };
42 
43 static struct device gru_device = {
44 	.init_name = "",
45 	.driver = &gru_driver,
46 };
47 
48 struct device *grudev = &gru_device;
49 
50 /*
51  * Select a gru fault map to be used by the current cpu. Note that
52  * multiple cpus may be using the same map.
53  *	ZZZ should be inline but did not work on emulator
54  */
55 int gru_cpu_fault_map_id(void)
56 {
57 #ifdef CONFIG_IA64
58 	return uv_blade_processor_id() % GRU_NUM_TFM;
59 #else
60 	int cpu = smp_processor_id();
61 	int id, core;
62 
63 	core = uv_cpu_core_number(cpu);
64 	id = core + UV_MAX_INT_CORES * uv_cpu_socket_number(cpu);
65 	return id;
66 #endif
67 }
68 
69 /*--------- ASID Management -------------------------------------------
70  *
71  *  Initially, assign asids sequentially from MIN_ASID .. MAX_ASID.
72  *  Once MAX is reached, flush the TLB & start over. However,
73  *  some asids may still be in use. There won't be many (percentage wise) still
74  *  in use. Search active contexts & determine the value of the first
75  *  asid in use ("x"s below). Set "limit" to this value.
76  *  This defines a block of assignable asids.
77  *
78  *  When "limit" is reached, search forward from limit+1 and determine the
79  *  next block of assignable asids.
80  *
81  *  Repeat until MAX_ASID is reached, then start over again.
82  *
83  *  Each time MAX_ASID is reached, increment the asid generation. Since
84  *  the search for in-use asids only checks contexts with GRUs currently
85  *  assigned, asids in some contexts will be missed. Prior to loading
86  *  a context, the asid generation of the GTS asid is rechecked. If it
87  *  doesn't match the current generation, a new asid will be assigned.
88  *
89  *   	0---------------x------------x---------------------x----|
90  *	  ^-next	^-limit	   				^-MAX_ASID
91  *
92  * All asid manipulation & context loading/unloading is protected by the
93  * gs_lock.
94  */
95 
96 /* Hit the asid limit. Start over */
97 static int gru_wrap_asid(struct gru_state *gru)
98 {
99 	gru_dbg(grudev, "gid %d\n", gru->gs_gid);
100 	STAT(asid_wrap);
101 	gru->gs_asid_gen++;
102 	return MIN_ASID;
103 }
104 
105 /* Find the next chunk of unused asids */
106 static int gru_reset_asid_limit(struct gru_state *gru, int asid)
107 {
108 	int i, gid, inuse_asid, limit;
109 
110 	gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid);
111 	STAT(asid_next);
112 	limit = MAX_ASID;
113 	if (asid >= limit)
114 		asid = gru_wrap_asid(gru);
115 	gru_flush_all_tlb(gru);
116 	gid = gru->gs_gid;
117 again:
118 	for (i = 0; i < GRU_NUM_CCH; i++) {
119 		if (!gru->gs_gts[i] || is_kernel_context(gru->gs_gts[i]))
120 			continue;
121 		inuse_asid = gru->gs_gts[i]->ts_gms->ms_asids[gid].mt_asid;
122 		gru_dbg(grudev, "gid %d, gts %p, gms %p, inuse 0x%x, cxt %d\n",
123 			gru->gs_gid, gru->gs_gts[i], gru->gs_gts[i]->ts_gms,
124 			inuse_asid, i);
125 		if (inuse_asid == asid) {
126 			asid += ASID_INC;
127 			if (asid >= limit) {
128 				/*
129 				 * empty range: reset the range limit and
130 				 * start over
131 				 */
132 				limit = MAX_ASID;
133 				if (asid >= MAX_ASID)
134 					asid = gru_wrap_asid(gru);
135 				goto again;
136 			}
137 		}
138 
139 		if ((inuse_asid > asid) && (inuse_asid < limit))
140 			limit = inuse_asid;
141 	}
142 	gru->gs_asid_limit = limit;
143 	gru->gs_asid = asid;
144 	gru_dbg(grudev, "gid %d, new asid 0x%x, new_limit 0x%x\n", gru->gs_gid,
145 					asid, limit);
146 	return asid;
147 }
148 
149 /* Assign a new ASID to a thread context.  */
150 static int gru_assign_asid(struct gru_state *gru)
151 {
152 	int asid;
153 
154 	gru->gs_asid += ASID_INC;
155 	asid = gru->gs_asid;
156 	if (asid >= gru->gs_asid_limit)
157 		asid = gru_reset_asid_limit(gru, asid);
158 
159 	gru_dbg(grudev, "gid %d, asid 0x%x\n", gru->gs_gid, asid);
160 	return asid;
161 }
162 
163 /*
164  * Clear n bits in a word. Return a word indicating the bits that were cleared.
165  * Optionally, build an array of chars that contain the bit numbers allocated.
166  */
167 static unsigned long reserve_resources(unsigned long *p, int n, int mmax,
168 				       char *idx)
169 {
170 	unsigned long bits = 0;
171 	int i;
172 
173 	while (n--) {
174 		i = find_first_bit(p, mmax);
175 		if (i == mmax)
176 			BUG();
177 		__clear_bit(i, p);
178 		__set_bit(i, &bits);
179 		if (idx)
180 			*idx++ = i;
181 	}
182 	return bits;
183 }
184 
185 unsigned long gru_reserve_cb_resources(struct gru_state *gru, int cbr_au_count,
186 				       char *cbmap)
187 {
188 	return reserve_resources(&gru->gs_cbr_map, cbr_au_count, GRU_CBR_AU,
189 				 cbmap);
190 }
191 
192 unsigned long gru_reserve_ds_resources(struct gru_state *gru, int dsr_au_count,
193 				       char *dsmap)
194 {
195 	return reserve_resources(&gru->gs_dsr_map, dsr_au_count, GRU_DSR_AU,
196 				 dsmap);
197 }
198 
199 static void reserve_gru_resources(struct gru_state *gru,
200 				  struct gru_thread_state *gts)
201 {
202 	gru->gs_active_contexts++;
203 	gts->ts_cbr_map =
204 	    gru_reserve_cb_resources(gru, gts->ts_cbr_au_count,
205 				     gts->ts_cbr_idx);
206 	gts->ts_dsr_map =
207 	    gru_reserve_ds_resources(gru, gts->ts_dsr_au_count, NULL);
208 }
209 
210 static void free_gru_resources(struct gru_state *gru,
211 			       struct gru_thread_state *gts)
212 {
213 	gru->gs_active_contexts--;
214 	gru->gs_cbr_map |= gts->ts_cbr_map;
215 	gru->gs_dsr_map |= gts->ts_dsr_map;
216 }
217 
218 /*
219  * Check if a GRU has sufficient free resources to satisfy an allocation
220  * request. Note: GRU locks may or may not be held when this is called. If
221  * not held, recheck after acquiring the appropriate locks.
222  *
223  * Returns 1 if sufficient resources, 0 if not
224  */
225 static int check_gru_resources(struct gru_state *gru, int cbr_au_count,
226 			       int dsr_au_count, int max_active_contexts)
227 {
228 	return hweight64(gru->gs_cbr_map) >= cbr_au_count
229 		&& hweight64(gru->gs_dsr_map) >= dsr_au_count
230 		&& gru->gs_active_contexts < max_active_contexts;
231 }
232 
233 /*
234  * TLB manangment requires tracking all GRU chiplets that have loaded a GSEG
235  * context.
236  */
237 static int gru_load_mm_tracker(struct gru_state *gru,
238 					struct gru_thread_state *gts)
239 {
240 	struct gru_mm_struct *gms = gts->ts_gms;
241 	struct gru_mm_tracker *asids = &gms->ms_asids[gru->gs_gid];
242 	unsigned short ctxbitmap = (1 << gts->ts_ctxnum);
243 	int asid;
244 
245 	spin_lock(&gms->ms_asid_lock);
246 	asid = asids->mt_asid;
247 
248 	spin_lock(&gru->gs_asid_lock);
249 	if (asid == 0 || (asids->mt_ctxbitmap == 0 && asids->mt_asid_gen !=
250 			  gru->gs_asid_gen)) {
251 		asid = gru_assign_asid(gru);
252 		asids->mt_asid = asid;
253 		asids->mt_asid_gen = gru->gs_asid_gen;
254 		STAT(asid_new);
255 	} else {
256 		STAT(asid_reuse);
257 	}
258 	spin_unlock(&gru->gs_asid_lock);
259 
260 	BUG_ON(asids->mt_ctxbitmap & ctxbitmap);
261 	asids->mt_ctxbitmap |= ctxbitmap;
262 	if (!test_bit(gru->gs_gid, gms->ms_asidmap))
263 		__set_bit(gru->gs_gid, gms->ms_asidmap);
264 	spin_unlock(&gms->ms_asid_lock);
265 
266 	gru_dbg(grudev,
267 		"gid %d, gts %p, gms %p, ctxnum %d, asid 0x%x, asidmap 0x%lx\n",
268 		gru->gs_gid, gts, gms, gts->ts_ctxnum, asid,
269 		gms->ms_asidmap[0]);
270 	return asid;
271 }
272 
273 static void gru_unload_mm_tracker(struct gru_state *gru,
274 					struct gru_thread_state *gts)
275 {
276 	struct gru_mm_struct *gms = gts->ts_gms;
277 	struct gru_mm_tracker *asids;
278 	unsigned short ctxbitmap;
279 
280 	asids = &gms->ms_asids[gru->gs_gid];
281 	ctxbitmap = (1 << gts->ts_ctxnum);
282 	spin_lock(&gms->ms_asid_lock);
283 	spin_lock(&gru->gs_asid_lock);
284 	BUG_ON((asids->mt_ctxbitmap & ctxbitmap) != ctxbitmap);
285 	asids->mt_ctxbitmap ^= ctxbitmap;
286 	gru_dbg(grudev, "gid %d, gts %p, gms %p, ctxnum %d, asidmap 0x%lx\n",
287 		gru->gs_gid, gts, gms, gts->ts_ctxnum, gms->ms_asidmap[0]);
288 	spin_unlock(&gru->gs_asid_lock);
289 	spin_unlock(&gms->ms_asid_lock);
290 }
291 
292 /*
293  * Decrement the reference count on a GTS structure. Free the structure
294  * if the reference count goes to zero.
295  */
296 void gts_drop(struct gru_thread_state *gts)
297 {
298 	if (gts && atomic_dec_return(&gts->ts_refcnt) == 0) {
299 		if (gts->ts_gms)
300 			gru_drop_mmu_notifier(gts->ts_gms);
301 		kfree(gts);
302 		STAT(gts_free);
303 	}
304 }
305 
306 /*
307  * Locate the GTS structure for the current thread.
308  */
309 static struct gru_thread_state *gru_find_current_gts_nolock(struct gru_vma_data
310 			    *vdata, int tsid)
311 {
312 	struct gru_thread_state *gts;
313 
314 	list_for_each_entry(gts, &vdata->vd_head, ts_next)
315 	    if (gts->ts_tsid == tsid)
316 		return gts;
317 	return NULL;
318 }
319 
320 /*
321  * Allocate a thread state structure.
322  */
323 struct gru_thread_state *gru_alloc_gts(struct vm_area_struct *vma,
324 		int cbr_au_count, int dsr_au_count,
325 		unsigned char tlb_preload_count, int options, int tsid)
326 {
327 	struct gru_thread_state *gts;
328 	struct gru_mm_struct *gms;
329 	int bytes;
330 
331 	bytes = DSR_BYTES(dsr_au_count) + CBR_BYTES(cbr_au_count);
332 	bytes += sizeof(struct gru_thread_state);
333 	gts = kmalloc(bytes, GFP_KERNEL);
334 	if (!gts)
335 		return ERR_PTR(-ENOMEM);
336 
337 	STAT(gts_alloc);
338 	memset(gts, 0, sizeof(struct gru_thread_state)); /* zero out header */
339 	atomic_set(&gts->ts_refcnt, 1);
340 	mutex_init(&gts->ts_ctxlock);
341 	gts->ts_cbr_au_count = cbr_au_count;
342 	gts->ts_dsr_au_count = dsr_au_count;
343 	gts->ts_tlb_preload_count = tlb_preload_count;
344 	gts->ts_user_options = options;
345 	gts->ts_user_blade_id = -1;
346 	gts->ts_user_chiplet_id = -1;
347 	gts->ts_tsid = tsid;
348 	gts->ts_ctxnum = NULLCTX;
349 	gts->ts_tlb_int_select = -1;
350 	gts->ts_cch_req_slice = -1;
351 	gts->ts_sizeavail = GRU_SIZEAVAIL(PAGE_SHIFT);
352 	if (vma) {
353 		gts->ts_mm = current->mm;
354 		gts->ts_vma = vma;
355 		gms = gru_register_mmu_notifier();
356 		if (IS_ERR(gms))
357 			goto err;
358 		gts->ts_gms = gms;
359 	}
360 
361 	gru_dbg(grudev, "alloc gts %p\n", gts);
362 	return gts;
363 
364 err:
365 	gts_drop(gts);
366 	return ERR_CAST(gms);
367 }
368 
369 /*
370  * Allocate a vma private data structure.
371  */
372 struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma, int tsid)
373 {
374 	struct gru_vma_data *vdata = NULL;
375 
376 	vdata = kmalloc(sizeof(*vdata), GFP_KERNEL);
377 	if (!vdata)
378 		return NULL;
379 
380 	STAT(vdata_alloc);
381 	INIT_LIST_HEAD(&vdata->vd_head);
382 	spin_lock_init(&vdata->vd_lock);
383 	gru_dbg(grudev, "alloc vdata %p\n", vdata);
384 	return vdata;
385 }
386 
387 /*
388  * Find the thread state structure for the current thread.
389  */
390 struct gru_thread_state *gru_find_thread_state(struct vm_area_struct *vma,
391 					int tsid)
392 {
393 	struct gru_vma_data *vdata = vma->vm_private_data;
394 	struct gru_thread_state *gts;
395 
396 	spin_lock(&vdata->vd_lock);
397 	gts = gru_find_current_gts_nolock(vdata, tsid);
398 	spin_unlock(&vdata->vd_lock);
399 	gru_dbg(grudev, "vma %p, gts %p\n", vma, gts);
400 	return gts;
401 }
402 
403 /*
404  * Allocate a new thread state for a GSEG. Note that races may allow
405  * another thread to race to create a gts.
406  */
407 struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct *vma,
408 					int tsid)
409 {
410 	struct gru_vma_data *vdata = vma->vm_private_data;
411 	struct gru_thread_state *gts, *ngts;
412 
413 	gts = gru_alloc_gts(vma, vdata->vd_cbr_au_count,
414 			    vdata->vd_dsr_au_count,
415 			    vdata->vd_tlb_preload_count,
416 			    vdata->vd_user_options, tsid);
417 	if (IS_ERR(gts))
418 		return gts;
419 
420 	spin_lock(&vdata->vd_lock);
421 	ngts = gru_find_current_gts_nolock(vdata, tsid);
422 	if (ngts) {
423 		gts_drop(gts);
424 		gts = ngts;
425 		STAT(gts_double_allocate);
426 	} else {
427 		list_add(&gts->ts_next, &vdata->vd_head);
428 	}
429 	spin_unlock(&vdata->vd_lock);
430 	gru_dbg(grudev, "vma %p, gts %p\n", vma, gts);
431 	return gts;
432 }
433 
434 /*
435  * Free the GRU context assigned to the thread state.
436  */
437 static void gru_free_gru_context(struct gru_thread_state *gts)
438 {
439 	struct gru_state *gru;
440 
441 	gru = gts->ts_gru;
442 	gru_dbg(grudev, "gts %p, gid %d\n", gts, gru->gs_gid);
443 
444 	spin_lock(&gru->gs_lock);
445 	gru->gs_gts[gts->ts_ctxnum] = NULL;
446 	free_gru_resources(gru, gts);
447 	BUG_ON(test_bit(gts->ts_ctxnum, &gru->gs_context_map) == 0);
448 	__clear_bit(gts->ts_ctxnum, &gru->gs_context_map);
449 	gts->ts_ctxnum = NULLCTX;
450 	gts->ts_gru = NULL;
451 	gts->ts_blade = -1;
452 	spin_unlock(&gru->gs_lock);
453 
454 	gts_drop(gts);
455 	STAT(free_context);
456 }
457 
458 /*
459  * Prefetching cachelines help hardware performance.
460  * (Strictly a performance enhancement. Not functionally required).
461  */
462 static void prefetch_data(void *p, int num, int stride)
463 {
464 	while (num-- > 0) {
465 		prefetchw(p);
466 		p += stride;
467 	}
468 }
469 
470 static inline long gru_copy_handle(void *d, void *s)
471 {
472 	memcpy(d, s, GRU_HANDLE_BYTES);
473 	return GRU_HANDLE_BYTES;
474 }
475 
476 static void gru_prefetch_context(void *gseg, void *cb, void *cbe,
477 				unsigned long cbrmap, unsigned long length)
478 {
479 	int i, scr;
480 
481 	prefetch_data(gseg + GRU_DS_BASE, length / GRU_CACHE_LINE_BYTES,
482 		      GRU_CACHE_LINE_BYTES);
483 
484 	for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
485 		prefetch_data(cb, 1, GRU_CACHE_LINE_BYTES);
486 		prefetch_data(cbe + i * GRU_HANDLE_STRIDE, 1,
487 			      GRU_CACHE_LINE_BYTES);
488 		cb += GRU_HANDLE_STRIDE;
489 	}
490 }
491 
492 static void gru_load_context_data(void *save, void *grubase, int ctxnum,
493 				  unsigned long cbrmap, unsigned long dsrmap,
494 				  int data_valid)
495 {
496 	void *gseg, *cb, *cbe;
497 	unsigned long length;
498 	int i, scr;
499 
500 	gseg = grubase + ctxnum * GRU_GSEG_STRIDE;
501 	cb = gseg + GRU_CB_BASE;
502 	cbe = grubase + GRU_CBE_BASE;
503 	length = hweight64(dsrmap) * GRU_DSR_AU_BYTES;
504 	gru_prefetch_context(gseg, cb, cbe, cbrmap, length);
505 
506 	for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
507 		if (data_valid) {
508 			save += gru_copy_handle(cb, save);
509 			save += gru_copy_handle(cbe + i * GRU_HANDLE_STRIDE,
510 						save);
511 		} else {
512 			memset(cb, 0, GRU_CACHE_LINE_BYTES);
513 			memset(cbe + i * GRU_HANDLE_STRIDE, 0,
514 						GRU_CACHE_LINE_BYTES);
515 		}
516 		/* Flush CBE to hide race in context restart */
517 		mb();
518 		gru_flush_cache(cbe + i * GRU_HANDLE_STRIDE);
519 		cb += GRU_HANDLE_STRIDE;
520 	}
521 
522 	if (data_valid)
523 		memcpy(gseg + GRU_DS_BASE, save, length);
524 	else
525 		memset(gseg + GRU_DS_BASE, 0, length);
526 }
527 
528 static void gru_unload_context_data(void *save, void *grubase, int ctxnum,
529 				    unsigned long cbrmap, unsigned long dsrmap)
530 {
531 	void *gseg, *cb, *cbe;
532 	unsigned long length;
533 	int i, scr;
534 
535 	gseg = grubase + ctxnum * GRU_GSEG_STRIDE;
536 	cb = gseg + GRU_CB_BASE;
537 	cbe = grubase + GRU_CBE_BASE;
538 	length = hweight64(dsrmap) * GRU_DSR_AU_BYTES;
539 
540 	/* CBEs may not be coherent. Flush them from cache */
541 	for_each_cbr_in_allocation_map(i, &cbrmap, scr)
542 		gru_flush_cache(cbe + i * GRU_HANDLE_STRIDE);
543 	mb();		/* Let the CL flush complete */
544 
545 	gru_prefetch_context(gseg, cb, cbe, cbrmap, length);
546 
547 	for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
548 		save += gru_copy_handle(save, cb);
549 		save += gru_copy_handle(save, cbe + i * GRU_HANDLE_STRIDE);
550 		cb += GRU_HANDLE_STRIDE;
551 	}
552 	memcpy(save, gseg + GRU_DS_BASE, length);
553 }
554 
555 void gru_unload_context(struct gru_thread_state *gts, int savestate)
556 {
557 	struct gru_state *gru = gts->ts_gru;
558 	struct gru_context_configuration_handle *cch;
559 	int ctxnum = gts->ts_ctxnum;
560 
561 	if (!is_kernel_context(gts))
562 		zap_vma_ptes(gts->ts_vma, UGRUADDR(gts), GRU_GSEG_PAGESIZE);
563 	cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
564 
565 	gru_dbg(grudev, "gts %p, cbrmap 0x%lx, dsrmap 0x%lx\n",
566 		gts, gts->ts_cbr_map, gts->ts_dsr_map);
567 	lock_cch_handle(cch);
568 	if (cch_interrupt_sync(cch))
569 		BUG();
570 
571 	if (!is_kernel_context(gts))
572 		gru_unload_mm_tracker(gru, gts);
573 	if (savestate) {
574 		gru_unload_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr,
575 					ctxnum, gts->ts_cbr_map,
576 					gts->ts_dsr_map);
577 		gts->ts_data_valid = 1;
578 	}
579 
580 	if (cch_deallocate(cch))
581 		BUG();
582 	unlock_cch_handle(cch);
583 
584 	gru_free_gru_context(gts);
585 }
586 
587 /*
588  * Load a GRU context by copying it from the thread data structure in memory
589  * to the GRU.
590  */
591 void gru_load_context(struct gru_thread_state *gts)
592 {
593 	struct gru_state *gru = gts->ts_gru;
594 	struct gru_context_configuration_handle *cch;
595 	int i, err, asid, ctxnum = gts->ts_ctxnum;
596 
597 	cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
598 	lock_cch_handle(cch);
599 	cch->tfm_fault_bit_enable =
600 	    (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL
601 	     || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
602 	cch->tlb_int_enable = (gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
603 	if (cch->tlb_int_enable) {
604 		gts->ts_tlb_int_select = gru_cpu_fault_map_id();
605 		cch->tlb_int_select = gts->ts_tlb_int_select;
606 	}
607 	if (gts->ts_cch_req_slice >= 0) {
608 		cch->req_slice_set_enable = 1;
609 		cch->req_slice = gts->ts_cch_req_slice;
610 	} else {
611 		cch->req_slice_set_enable =0;
612 	}
613 	cch->tfm_done_bit_enable = 0;
614 	cch->dsr_allocation_map = gts->ts_dsr_map;
615 	cch->cbr_allocation_map = gts->ts_cbr_map;
616 
617 	if (is_kernel_context(gts)) {
618 		cch->unmap_enable = 1;
619 		cch->tfm_done_bit_enable = 1;
620 		cch->cb_int_enable = 1;
621 		cch->tlb_int_select = 0;	/* For now, ints go to cpu 0 */
622 	} else {
623 		cch->unmap_enable = 0;
624 		cch->tfm_done_bit_enable = 0;
625 		cch->cb_int_enable = 0;
626 		asid = gru_load_mm_tracker(gru, gts);
627 		for (i = 0; i < 8; i++) {
628 			cch->asid[i] = asid + i;
629 			cch->sizeavail[i] = gts->ts_sizeavail;
630 		}
631 	}
632 
633 	err = cch_allocate(cch);
634 	if (err) {
635 		gru_dbg(grudev,
636 			"err %d: cch %p, gts %p, cbr 0x%lx, dsr 0x%lx\n",
637 			err, cch, gts, gts->ts_cbr_map, gts->ts_dsr_map);
638 		BUG();
639 	}
640 
641 	gru_load_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, ctxnum,
642 			gts->ts_cbr_map, gts->ts_dsr_map, gts->ts_data_valid);
643 
644 	if (cch_start(cch))
645 		BUG();
646 	unlock_cch_handle(cch);
647 
648 	gru_dbg(grudev, "gid %d, gts %p, cbrmap 0x%lx, dsrmap 0x%lx, tie %d, tis %d\n",
649 		gts->ts_gru->gs_gid, gts, gts->ts_cbr_map, gts->ts_dsr_map,
650 		(gts->ts_user_options == GRU_OPT_MISS_FMM_INTR), gts->ts_tlb_int_select);
651 }
652 
653 /*
654  * Update fields in an active CCH:
655  * 	- retarget interrupts on local blade
656  * 	- update sizeavail mask
657  */
658 int gru_update_cch(struct gru_thread_state *gts)
659 {
660 	struct gru_context_configuration_handle *cch;
661 	struct gru_state *gru = gts->ts_gru;
662 	int i, ctxnum = gts->ts_ctxnum, ret = 0;
663 
664 	cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
665 
666 	lock_cch_handle(cch);
667 	if (cch->state == CCHSTATE_ACTIVE) {
668 		if (gru->gs_gts[gts->ts_ctxnum] != gts)
669 			goto exit;
670 		if (cch_interrupt(cch))
671 			BUG();
672 		for (i = 0; i < 8; i++)
673 			cch->sizeavail[i] = gts->ts_sizeavail;
674 		gts->ts_tlb_int_select = gru_cpu_fault_map_id();
675 		cch->tlb_int_select = gru_cpu_fault_map_id();
676 		cch->tfm_fault_bit_enable =
677 		  (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL
678 		    || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
679 		if (cch_start(cch))
680 			BUG();
681 		ret = 1;
682 	}
683 exit:
684 	unlock_cch_handle(cch);
685 	return ret;
686 }
687 
688 /*
689  * Update CCH tlb interrupt select. Required when all the following is true:
690  * 	- task's GRU context is loaded into a GRU
691  * 	- task is using interrupt notification for TLB faults
692  * 	- task has migrated to a different cpu on the same blade where
693  * 	  it was previously running.
694  */
695 static int gru_retarget_intr(struct gru_thread_state *gts)
696 {
697 	if (gts->ts_tlb_int_select < 0
698 	    || gts->ts_tlb_int_select == gru_cpu_fault_map_id())
699 		return 0;
700 
701 	gru_dbg(grudev, "retarget from %d to %d\n", gts->ts_tlb_int_select,
702 		gru_cpu_fault_map_id());
703 	return gru_update_cch(gts);
704 }
705 
706 /*
707  * Check if a GRU context is allowed to use a specific chiplet. By default
708  * a context is assigned to any blade-local chiplet. However, users can
709  * override this.
710  * 	Returns 1 if assignment allowed, 0 otherwise
711  */
712 static int gru_check_chiplet_assignment(struct gru_state *gru,
713 					struct gru_thread_state *gts)
714 {
715 	int blade_id;
716 	int chiplet_id;
717 
718 	blade_id = gts->ts_user_blade_id;
719 	if (blade_id < 0)
720 		blade_id = uv_numa_blade_id();
721 
722 	chiplet_id = gts->ts_user_chiplet_id;
723 	return gru->gs_blade_id == blade_id &&
724 		(chiplet_id < 0 || chiplet_id == gru->gs_chiplet_id);
725 }
726 
727 /*
728  * Unload the gru context if it is not assigned to the correct blade or
729  * chiplet. Misassignment can occur if the process migrates to a different
730  * blade or if the user changes the selected blade/chiplet.
731  */
732 void gru_check_context_placement(struct gru_thread_state *gts)
733 {
734 	struct gru_state *gru;
735 
736 	/*
737 	 * If the current task is the context owner, verify that the
738 	 * context is correctly placed. This test is skipped for non-owner
739 	 * references. Pthread apps use non-owner references to the CBRs.
740 	 */
741 	gru = gts->ts_gru;
742 	if (!gru || gts->ts_tgid_owner != current->tgid)
743 		return;
744 
745 	if (!gru_check_chiplet_assignment(gru, gts)) {
746 		STAT(check_context_unload);
747 		gru_unload_context(gts, 1);
748 	} else if (gru_retarget_intr(gts)) {
749 		STAT(check_context_retarget_intr);
750 	}
751 }
752 
753 
754 /*
755  * Insufficient GRU resources available on the local blade. Steal a context from
756  * a process. This is a hack until a _real_ resource scheduler is written....
757  */
758 #define next_ctxnum(n)	((n) <  GRU_NUM_CCH - 2 ? (n) + 1 : 0)
759 #define next_gru(b, g)	(((g) < &(b)->bs_grus[GRU_CHIPLETS_PER_BLADE - 1]) ?  \
760 				 ((g)+1) : &(b)->bs_grus[0])
761 
762 static int is_gts_stealable(struct gru_thread_state *gts,
763 		struct gru_blade_state *bs)
764 {
765 	if (is_kernel_context(gts))
766 		return down_write_trylock(&bs->bs_kgts_sema);
767 	else
768 		return mutex_trylock(&gts->ts_ctxlock);
769 }
770 
771 static void gts_stolen(struct gru_thread_state *gts,
772 		struct gru_blade_state *bs)
773 {
774 	if (is_kernel_context(gts)) {
775 		up_write(&bs->bs_kgts_sema);
776 		STAT(steal_kernel_context);
777 	} else {
778 		mutex_unlock(&gts->ts_ctxlock);
779 		STAT(steal_user_context);
780 	}
781 }
782 
783 void gru_steal_context(struct gru_thread_state *gts)
784 {
785 	struct gru_blade_state *blade;
786 	struct gru_state *gru, *gru0;
787 	struct gru_thread_state *ngts = NULL;
788 	int ctxnum, ctxnum0, flag = 0, cbr, dsr;
789 	int blade_id;
790 
791 	blade_id = gts->ts_user_blade_id;
792 	if (blade_id < 0)
793 		blade_id = uv_numa_blade_id();
794 	cbr = gts->ts_cbr_au_count;
795 	dsr = gts->ts_dsr_au_count;
796 
797 	blade = gru_base[blade_id];
798 	spin_lock(&blade->bs_lock);
799 
800 	ctxnum = next_ctxnum(blade->bs_lru_ctxnum);
801 	gru = blade->bs_lru_gru;
802 	if (ctxnum == 0)
803 		gru = next_gru(blade, gru);
804 	blade->bs_lru_gru = gru;
805 	blade->bs_lru_ctxnum = ctxnum;
806 	ctxnum0 = ctxnum;
807 	gru0 = gru;
808 	while (1) {
809 		if (gru_check_chiplet_assignment(gru, gts)) {
810 			if (check_gru_resources(gru, cbr, dsr, GRU_NUM_CCH))
811 				break;
812 			spin_lock(&gru->gs_lock);
813 			for (; ctxnum < GRU_NUM_CCH; ctxnum++) {
814 				if (flag && gru == gru0 && ctxnum == ctxnum0)
815 					break;
816 				ngts = gru->gs_gts[ctxnum];
817 				/*
818 			 	* We are grabbing locks out of order, so trylock is
819 			 	* needed. GTSs are usually not locked, so the odds of
820 			 	* success are high. If trylock fails, try to steal a
821 			 	* different GSEG.
822 			 	*/
823 				if (ngts && is_gts_stealable(ngts, blade))
824 					break;
825 				ngts = NULL;
826 			}
827 			spin_unlock(&gru->gs_lock);
828 			if (ngts || (flag && gru == gru0 && ctxnum == ctxnum0))
829 				break;
830 		}
831 		if (flag && gru == gru0)
832 			break;
833 		flag = 1;
834 		ctxnum = 0;
835 		gru = next_gru(blade, gru);
836 	}
837 	spin_unlock(&blade->bs_lock);
838 
839 	if (ngts) {
840 		gts->ustats.context_stolen++;
841 		ngts->ts_steal_jiffies = jiffies;
842 		gru_unload_context(ngts, is_kernel_context(ngts) ? 0 : 1);
843 		gts_stolen(ngts, blade);
844 	} else {
845 		STAT(steal_context_failed);
846 	}
847 	gru_dbg(grudev,
848 		"stole gid %d, ctxnum %d from gts %p. Need cb %d, ds %d;"
849 		" avail cb %ld, ds %ld\n",
850 		gru->gs_gid, ctxnum, ngts, cbr, dsr, hweight64(gru->gs_cbr_map),
851 		hweight64(gru->gs_dsr_map));
852 }
853 
854 /*
855  * Assign a gru context.
856  */
857 static int gru_assign_context_number(struct gru_state *gru)
858 {
859 	int ctxnum;
860 
861 	ctxnum = find_first_zero_bit(&gru->gs_context_map, GRU_NUM_CCH);
862 	__set_bit(ctxnum, &gru->gs_context_map);
863 	return ctxnum;
864 }
865 
866 /*
867  * Scan the GRUs on the local blade & assign a GRU context.
868  */
869 struct gru_state *gru_assign_gru_context(struct gru_thread_state *gts)
870 {
871 	struct gru_state *gru, *grux;
872 	int i, max_active_contexts;
873 	int blade_id = gts->ts_user_blade_id;
874 
875 	if (blade_id < 0)
876 		blade_id = uv_numa_blade_id();
877 again:
878 	gru = NULL;
879 	max_active_contexts = GRU_NUM_CCH;
880 	for_each_gru_on_blade(grux, blade_id, i) {
881 		if (!gru_check_chiplet_assignment(grux, gts))
882 			continue;
883 		if (check_gru_resources(grux, gts->ts_cbr_au_count,
884 					gts->ts_dsr_au_count,
885 					max_active_contexts)) {
886 			gru = grux;
887 			max_active_contexts = grux->gs_active_contexts;
888 			if (max_active_contexts == 0)
889 				break;
890 		}
891 	}
892 
893 	if (gru) {
894 		spin_lock(&gru->gs_lock);
895 		if (!check_gru_resources(gru, gts->ts_cbr_au_count,
896 					 gts->ts_dsr_au_count, GRU_NUM_CCH)) {
897 			spin_unlock(&gru->gs_lock);
898 			goto again;
899 		}
900 		reserve_gru_resources(gru, gts);
901 		gts->ts_gru = gru;
902 		gts->ts_blade = gru->gs_blade_id;
903 		gts->ts_ctxnum = gru_assign_context_number(gru);
904 		atomic_inc(&gts->ts_refcnt);
905 		gru->gs_gts[gts->ts_ctxnum] = gts;
906 		spin_unlock(&gru->gs_lock);
907 
908 		STAT(assign_context);
909 		gru_dbg(grudev,
910 			"gseg %p, gts %p, gid %d, ctx %d, cbr %d, dsr %d\n",
911 			gseg_virtual_address(gts->ts_gru, gts->ts_ctxnum), gts,
912 			gts->ts_gru->gs_gid, gts->ts_ctxnum,
913 			gts->ts_cbr_au_count, gts->ts_dsr_au_count);
914 	} else {
915 		gru_dbg(grudev, "failed to allocate a GTS %s\n", "");
916 		STAT(assign_context_failed);
917 	}
918 
919 	return gru;
920 }
921 
922 /*
923  * gru_nopage
924  *
925  * Map the user's GRU segment
926  *
927  * 	Note: gru segments alway mmaped on GRU_GSEG_PAGESIZE boundaries.
928  */
929 vm_fault_t gru_fault(struct vm_fault *vmf)
930 {
931 	struct vm_area_struct *vma = vmf->vma;
932 	struct gru_thread_state *gts;
933 	unsigned long paddr, vaddr;
934 	unsigned long expires;
935 
936 	vaddr = vmf->address;
937 	gru_dbg(grudev, "vma %p, vaddr 0x%lx (0x%lx)\n",
938 		vma, vaddr, GSEG_BASE(vaddr));
939 	STAT(nopfn);
940 
941 	/* The following check ensures vaddr is a valid address in the VMA */
942 	gts = gru_find_thread_state(vma, TSID(vaddr, vma));
943 	if (!gts)
944 		return VM_FAULT_SIGBUS;
945 
946 again:
947 	mutex_lock(&gts->ts_ctxlock);
948 	preempt_disable();
949 
950 	gru_check_context_placement(gts);
951 
952 	if (!gts->ts_gru) {
953 		STAT(load_user_context);
954 		if (!gru_assign_gru_context(gts)) {
955 			preempt_enable();
956 			mutex_unlock(&gts->ts_ctxlock);
957 			set_current_state(TASK_INTERRUPTIBLE);
958 			schedule_timeout(GRU_ASSIGN_DELAY);  /* true hack ZZZ */
959 			expires = gts->ts_steal_jiffies + GRU_STEAL_DELAY;
960 			if (time_before(expires, jiffies))
961 				gru_steal_context(gts);
962 			goto again;
963 		}
964 		gru_load_context(gts);
965 		paddr = gseg_physical_address(gts->ts_gru, gts->ts_ctxnum);
966 		remap_pfn_range(vma, vaddr & ~(GRU_GSEG_PAGESIZE - 1),
967 				paddr >> PAGE_SHIFT, GRU_GSEG_PAGESIZE,
968 				vma->vm_page_prot);
969 	}
970 
971 	preempt_enable();
972 	mutex_unlock(&gts->ts_ctxlock);
973 
974 	return VM_FAULT_NOPAGE;
975 }
976 
977