• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * SN Platform GRU Driver
3  *
4  *              FAULT HANDLER FOR GRU DETECTED TLB MISSES
5  *
6  * This file contains code that handles TLB misses within the GRU.
7  * These misses are reported either via interrupts or user polling of
8  * the user CB.
9  *
10  *  Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
11  *
12  *  This program is free software; you can redistribute it and/or modify
13  *  it under the terms of the GNU General Public License as published by
14  *  the Free Software Foundation; either version 2 of the License, or
15  *  (at your option) any later version.
16  *
17  *  This program is distributed in the hope that it will be useful,
18  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
19  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
20  *  GNU General Public License for more details.
21  *
22  *  You should have received a copy of the GNU General Public License
23  *  along with this program; if not, write to the Free Software
24  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
25  */
26 
27 #include <linux/kernel.h>
28 #include <linux/errno.h>
29 #include <linux/spinlock.h>
30 #include <linux/mm.h>
31 #include <linux/hugetlb.h>
32 #include <linux/device.h>
33 #include <linux/io.h>
34 #include <linux/uaccess.h>
35 #include <linux/security.h>
36 #include <linux/prefetch.h>
37 #include <asm/pgtable.h>
38 #include "gru.h"
39 #include "grutables.h"
40 #include "grulib.h"
41 #include "gru_instructions.h"
42 #include <asm/uv/uv_hub.h>
43 
44 /* Return codes for vtop functions */
45 #define VTOP_SUCCESS               0
46 #define VTOP_INVALID               -1
47 #define VTOP_RETRY                 -2
48 
49 
50 /*
51  * Test if a physical address is a valid GRU GSEG address
52  */
is_gru_paddr(unsigned long paddr)53 static inline int is_gru_paddr(unsigned long paddr)
54 {
55 	return paddr >= gru_start_paddr && paddr < gru_end_paddr;
56 }
57 
58 /*
59  * Find the vma of a GRU segment. Caller must hold mmap_sem.
60  */
gru_find_vma(unsigned long vaddr)61 struct vm_area_struct *gru_find_vma(unsigned long vaddr)
62 {
63 	struct vm_area_struct *vma;
64 
65 	vma = find_vma(current->mm, vaddr);
66 	if (vma && vma->vm_start <= vaddr && vma->vm_ops == &gru_vm_ops)
67 		return vma;
68 	return NULL;
69 }
70 
71 /*
72  * Find and lock the gts that contains the specified user vaddr.
73  *
74  * Returns:
75  * 	- *gts with the mmap_sem locked for read and the GTS locked.
76  *	- NULL if vaddr invalid OR is not a valid GSEG vaddr.
77  */
78 
gru_find_lock_gts(unsigned long vaddr)79 static struct gru_thread_state *gru_find_lock_gts(unsigned long vaddr)
80 {
81 	struct mm_struct *mm = current->mm;
82 	struct vm_area_struct *vma;
83 	struct gru_thread_state *gts = NULL;
84 
85 	down_read(&mm->mmap_sem);
86 	vma = gru_find_vma(vaddr);
87 	if (vma)
88 		gts = gru_find_thread_state(vma, TSID(vaddr, vma));
89 	if (gts)
90 		mutex_lock(&gts->ts_ctxlock);
91 	else
92 		up_read(&mm->mmap_sem);
93 	return gts;
94 }
95 
gru_alloc_locked_gts(unsigned long vaddr)96 static struct gru_thread_state *gru_alloc_locked_gts(unsigned long vaddr)
97 {
98 	struct mm_struct *mm = current->mm;
99 	struct vm_area_struct *vma;
100 	struct gru_thread_state *gts = ERR_PTR(-EINVAL);
101 
102 	down_write(&mm->mmap_sem);
103 	vma = gru_find_vma(vaddr);
104 	if (!vma)
105 		goto err;
106 
107 	gts = gru_alloc_thread_state(vma, TSID(vaddr, vma));
108 	if (IS_ERR(gts))
109 		goto err;
110 	mutex_lock(&gts->ts_ctxlock);
111 	downgrade_write(&mm->mmap_sem);
112 	return gts;
113 
114 err:
115 	up_write(&mm->mmap_sem);
116 	return gts;
117 }
118 
119 /*
120  * Unlock a GTS that was previously locked with gru_find_lock_gts().
121  */
gru_unlock_gts(struct gru_thread_state * gts)122 static void gru_unlock_gts(struct gru_thread_state *gts)
123 {
124 	mutex_unlock(&gts->ts_ctxlock);
125 	up_read(&current->mm->mmap_sem);
126 }
127 
128 /*
129  * Set a CB.istatus to active using a user virtual address. This must be done
130  * just prior to a TFH RESTART. The new cb.istatus is an in-cache status ONLY.
131  * If the line is evicted, the status may be lost. The in-cache update
132  * is necessary to prevent the user from seeing a stale cb.istatus that will
133  * change as soon as the TFH restart is complete. Races may cause an
134  * occasional failure to clear the cb.istatus, but that is ok.
135  */
gru_cb_set_istatus_active(struct gru_instruction_bits * cbk)136 static void gru_cb_set_istatus_active(struct gru_instruction_bits *cbk)
137 {
138 	if (cbk) {
139 		cbk->istatus = CBS_ACTIVE;
140 	}
141 }
142 
143 /*
144  * Read & clear a TFM
145  *
146  * The GRU has an array of fault maps. A map is private to a cpu
147  * Only one cpu will be accessing a cpu's fault map.
148  *
149  * This function scans the cpu-private fault map & clears all bits that
150  * are set. The function returns a bitmap that indicates the bits that
151  * were cleared. Note that sense the maps may be updated asynchronously by
152  * the GRU, atomic operations must be used to clear bits.
153  */
get_clear_fault_map(struct gru_state * gru,struct gru_tlb_fault_map * imap,struct gru_tlb_fault_map * dmap)154 static void get_clear_fault_map(struct gru_state *gru,
155 				struct gru_tlb_fault_map *imap,
156 				struct gru_tlb_fault_map *dmap)
157 {
158 	unsigned long i, k;
159 	struct gru_tlb_fault_map *tfm;
160 
161 	tfm = get_tfm_for_cpu(gru, gru_cpu_fault_map_id());
162 	prefetchw(tfm);		/* Helps on hardware, required for emulator */
163 	for (i = 0; i < BITS_TO_LONGS(GRU_NUM_CBE); i++) {
164 		k = tfm->fault_bits[i];
165 		if (k)
166 			k = xchg(&tfm->fault_bits[i], 0UL);
167 		imap->fault_bits[i] = k;
168 		k = tfm->done_bits[i];
169 		if (k)
170 			k = xchg(&tfm->done_bits[i], 0UL);
171 		dmap->fault_bits[i] = k;
172 	}
173 
174 	/*
175 	 * Not functionally required but helps performance. (Required
176 	 * on emulator)
177 	 */
178 	gru_flush_cache(tfm);
179 }
180 
181 /*
182  * Atomic (interrupt context) & non-atomic (user context) functions to
183  * convert a vaddr into a physical address. The size of the page
184  * is returned in pageshift.
185  * 	returns:
186  * 		  0 - successful
187  * 		< 0 - error code
188  * 		  1 - (atomic only) try again in non-atomic context
189  */
non_atomic_pte_lookup(struct vm_area_struct * vma,unsigned long vaddr,int write,unsigned long * paddr,int * pageshift)190 static int non_atomic_pte_lookup(struct vm_area_struct *vma,
191 				 unsigned long vaddr, int write,
192 				 unsigned long *paddr, int *pageshift)
193 {
194 	struct page *page;
195 
196 #ifdef CONFIG_HUGETLB_PAGE
197 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
198 #else
199 	*pageshift = PAGE_SHIFT;
200 #endif
201 	if (get_user_pages(vaddr, 1, write ? FOLL_WRITE : 0, &page, NULL) <= 0)
202 		return -EFAULT;
203 	*paddr = page_to_phys(page);
204 	put_page(page);
205 	return 0;
206 }
207 
208 /*
209  * atomic_pte_lookup
210  *
211  * Convert a user virtual address to a physical address
212  * Only supports Intel large pages (2MB only) on x86_64.
213  *	ZZZ - hugepage support is incomplete
214  *
215  * NOTE: mmap_sem is already held on entry to this function. This
216  * guarantees existence of the page tables.
217  */
atomic_pte_lookup(struct vm_area_struct * vma,unsigned long vaddr,int write,unsigned long * paddr,int * pageshift)218 static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
219 	int write, unsigned long *paddr, int *pageshift)
220 {
221 	pgd_t *pgdp;
222 	pmd_t *pmdp;
223 	pud_t *pudp;
224 	pte_t pte;
225 
226 	pgdp = pgd_offset(vma->vm_mm, vaddr);
227 	if (unlikely(pgd_none(*pgdp)))
228 		goto err;
229 
230 	pudp = pud_offset(pgdp, vaddr);
231 	if (unlikely(pud_none(*pudp)))
232 		goto err;
233 
234 	pmdp = pmd_offset(pudp, vaddr);
235 	if (unlikely(pmd_none(*pmdp)))
236 		goto err;
237 #ifdef CONFIG_X86_64
238 	if (unlikely(pmd_large(*pmdp)))
239 		pte = *(pte_t *) pmdp;
240 	else
241 #endif
242 		pte = *pte_offset_kernel(pmdp, vaddr);
243 
244 	if (unlikely(!pte_present(pte) ||
245 		     (write && (!pte_write(pte) || !pte_dirty(pte)))))
246 		return 1;
247 
248 	*paddr = pte_pfn(pte) << PAGE_SHIFT;
249 #ifdef CONFIG_HUGETLB_PAGE
250 	*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
251 #else
252 	*pageshift = PAGE_SHIFT;
253 #endif
254 	return 0;
255 
256 err:
257 	return 1;
258 }
259 
gru_vtop(struct gru_thread_state * gts,unsigned long vaddr,int write,int atomic,unsigned long * gpa,int * pageshift)260 static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,
261 		    int write, int atomic, unsigned long *gpa, int *pageshift)
262 {
263 	struct mm_struct *mm = gts->ts_mm;
264 	struct vm_area_struct *vma;
265 	unsigned long paddr;
266 	int ret, ps;
267 
268 	vma = find_vma(mm, vaddr);
269 	if (!vma)
270 		goto inval;
271 
272 	/*
273 	 * Atomic lookup is faster & usually works even if called in non-atomic
274 	 * context.
275 	 */
276 	rmb();	/* Must/check ms_range_active before loading PTEs */
277 	ret = atomic_pte_lookup(vma, vaddr, write, &paddr, &ps);
278 	if (ret) {
279 		if (atomic)
280 			goto upm;
281 		if (non_atomic_pte_lookup(vma, vaddr, write, &paddr, &ps))
282 			goto inval;
283 	}
284 	if (is_gru_paddr(paddr))
285 		goto inval;
286 	paddr = paddr & ~((1UL << ps) - 1);
287 	*gpa = uv_soc_phys_ram_to_gpa(paddr);
288 	*pageshift = ps;
289 	return VTOP_SUCCESS;
290 
291 inval:
292 	return VTOP_INVALID;
293 upm:
294 	return VTOP_RETRY;
295 }
296 
297 
298 /*
299  * Flush a CBE from cache. The CBE is clean in the cache. Dirty the
300  * CBE cacheline so that the line will be written back to home agent.
301  * Otherwise the line may be silently dropped. This has no impact
302  * except on performance.
303  */
gru_flush_cache_cbe(struct gru_control_block_extended * cbe)304 static void gru_flush_cache_cbe(struct gru_control_block_extended *cbe)
305 {
306 	if (unlikely(cbe)) {
307 		cbe->cbrexecstatus = 0;         /* make CL dirty */
308 		gru_flush_cache(cbe);
309 	}
310 }
311 
312 /*
313  * Preload the TLB with entries that may be required. Currently, preloading
314  * is implemented only for BCOPY. Preload  <tlb_preload_count> pages OR to
315  * the end of the bcopy tranfer, whichever is smaller.
316  */
gru_preload_tlb(struct gru_state * gru,struct gru_thread_state * gts,int atomic,unsigned long fault_vaddr,int asid,int write,unsigned char tlb_preload_count,struct gru_tlb_fault_handle * tfh,struct gru_control_block_extended * cbe)317 static void gru_preload_tlb(struct gru_state *gru,
318 			struct gru_thread_state *gts, int atomic,
319 			unsigned long fault_vaddr, int asid, int write,
320 			unsigned char tlb_preload_count,
321 			struct gru_tlb_fault_handle *tfh,
322 			struct gru_control_block_extended *cbe)
323 {
324 	unsigned long vaddr = 0, gpa;
325 	int ret, pageshift;
326 
327 	if (cbe->opccpy != OP_BCOPY)
328 		return;
329 
330 	if (fault_vaddr == cbe->cbe_baddr0)
331 		vaddr = fault_vaddr + GRU_CACHE_LINE_BYTES * cbe->cbe_src_cl - 1;
332 	else if (fault_vaddr == cbe->cbe_baddr1)
333 		vaddr = fault_vaddr + (1 << cbe->xtypecpy) * cbe->cbe_nelemcur - 1;
334 
335 	fault_vaddr &= PAGE_MASK;
336 	vaddr &= PAGE_MASK;
337 	vaddr = min(vaddr, fault_vaddr + tlb_preload_count * PAGE_SIZE);
338 
339 	while (vaddr > fault_vaddr) {
340 		ret = gru_vtop(gts, vaddr, write, atomic, &gpa, &pageshift);
341 		if (ret || tfh_write_only(tfh, gpa, GAA_RAM, vaddr, asid, write,
342 					  GRU_PAGESIZE(pageshift)))
343 			return;
344 		gru_dbg(grudev,
345 			"%s: gid %d, gts 0x%p, tfh 0x%p, vaddr 0x%lx, asid 0x%x, rw %d, ps %d, gpa 0x%lx\n",
346 			atomic ? "atomic" : "non-atomic", gru->gs_gid, gts, tfh,
347 			vaddr, asid, write, pageshift, gpa);
348 		vaddr -= PAGE_SIZE;
349 		STAT(tlb_preload_page);
350 	}
351 }
352 
353 /*
354  * Drop a TLB entry into the GRU. The fault is described by info in an TFH.
355  *	Input:
356  *		cb    Address of user CBR. Null if not running in user context
357  * 	Return:
358  * 		  0 = dropin, exception, or switch to UPM successful
359  * 		  1 = range invalidate active
360  * 		< 0 = error code
361  *
362  */
gru_try_dropin(struct gru_state * gru,struct gru_thread_state * gts,struct gru_tlb_fault_handle * tfh,struct gru_instruction_bits * cbk)363 static int gru_try_dropin(struct gru_state *gru,
364 			  struct gru_thread_state *gts,
365 			  struct gru_tlb_fault_handle *tfh,
366 			  struct gru_instruction_bits *cbk)
367 {
368 	struct gru_control_block_extended *cbe = NULL;
369 	unsigned char tlb_preload_count = gts->ts_tlb_preload_count;
370 	int pageshift = 0, asid, write, ret, atomic = !cbk, indexway;
371 	unsigned long gpa = 0, vaddr = 0;
372 
373 	/*
374 	 * NOTE: The GRU contains magic hardware that eliminates races between
375 	 * TLB invalidates and TLB dropins. If an invalidate occurs
376 	 * in the window between reading the TFH and the subsequent TLB dropin,
377 	 * the dropin is ignored. This eliminates the need for additional locks.
378 	 */
379 
380 	/*
381 	 * Prefetch the CBE if doing TLB preloading
382 	 */
383 	if (unlikely(tlb_preload_count)) {
384 		cbe = gru_tfh_to_cbe(tfh);
385 		prefetchw(cbe);
386 	}
387 
388 	/*
389 	 * Error if TFH state is IDLE or FMM mode & the user issuing a UPM call.
390 	 * Might be a hardware race OR a stupid user. Ignore FMM because FMM
391 	 * is a transient state.
392 	 */
393 	if (tfh->status != TFHSTATUS_EXCEPTION) {
394 		gru_flush_cache(tfh);
395 		sync_core();
396 		if (tfh->status != TFHSTATUS_EXCEPTION)
397 			goto failnoexception;
398 		STAT(tfh_stale_on_fault);
399 	}
400 	if (tfh->state == TFHSTATE_IDLE)
401 		goto failidle;
402 	if (tfh->state == TFHSTATE_MISS_FMM && cbk)
403 		goto failfmm;
404 
405 	write = (tfh->cause & TFHCAUSE_TLB_MOD) != 0;
406 	vaddr = tfh->missvaddr;
407 	asid = tfh->missasid;
408 	indexway = tfh->indexway;
409 	if (asid == 0)
410 		goto failnoasid;
411 
412 	rmb();	/* TFH must be cache resident before reading ms_range_active */
413 
414 	/*
415 	 * TFH is cache resident - at least briefly. Fail the dropin
416 	 * if a range invalidate is active.
417 	 */
418 	if (atomic_read(&gts->ts_gms->ms_range_active))
419 		goto failactive;
420 
421 	ret = gru_vtop(gts, vaddr, write, atomic, &gpa, &pageshift);
422 	if (ret == VTOP_INVALID)
423 		goto failinval;
424 	if (ret == VTOP_RETRY)
425 		goto failupm;
426 
427 	if (!(gts->ts_sizeavail & GRU_SIZEAVAIL(pageshift))) {
428 		gts->ts_sizeavail |= GRU_SIZEAVAIL(pageshift);
429 		if (atomic || !gru_update_cch(gts)) {
430 			gts->ts_force_cch_reload = 1;
431 			goto failupm;
432 		}
433 	}
434 
435 	if (unlikely(cbe) && pageshift == PAGE_SHIFT) {
436 		gru_preload_tlb(gru, gts, atomic, vaddr, asid, write, tlb_preload_count, tfh, cbe);
437 		gru_flush_cache_cbe(cbe);
438 	}
439 
440 	gru_cb_set_istatus_active(cbk);
441 	gts->ustats.tlbdropin++;
442 	tfh_write_restart(tfh, gpa, GAA_RAM, vaddr, asid, write,
443 			  GRU_PAGESIZE(pageshift));
444 	gru_dbg(grudev,
445 		"%s: gid %d, gts 0x%p, tfh 0x%p, vaddr 0x%lx, asid 0x%x, indexway 0x%x,"
446 		" rw %d, ps %d, gpa 0x%lx\n",
447 		atomic ? "atomic" : "non-atomic", gru->gs_gid, gts, tfh, vaddr, asid,
448 		indexway, write, pageshift, gpa);
449 	STAT(tlb_dropin);
450 	return 0;
451 
452 failnoasid:
453 	/* No asid (delayed unload). */
454 	STAT(tlb_dropin_fail_no_asid);
455 	gru_dbg(grudev, "FAILED no_asid tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
456 	if (!cbk)
457 		tfh_user_polling_mode(tfh);
458 	else
459 		gru_flush_cache(tfh);
460 	gru_flush_cache_cbe(cbe);
461 	return -EAGAIN;
462 
463 failupm:
464 	/* Atomic failure switch CBR to UPM */
465 	tfh_user_polling_mode(tfh);
466 	gru_flush_cache_cbe(cbe);
467 	STAT(tlb_dropin_fail_upm);
468 	gru_dbg(grudev, "FAILED upm tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
469 	return 1;
470 
471 failfmm:
472 	/* FMM state on UPM call */
473 	gru_flush_cache(tfh);
474 	gru_flush_cache_cbe(cbe);
475 	STAT(tlb_dropin_fail_fmm);
476 	gru_dbg(grudev, "FAILED fmm tfh: 0x%p, state %d\n", tfh, tfh->state);
477 	return 0;
478 
479 failnoexception:
480 	/* TFH status did not show exception pending */
481 	gru_flush_cache(tfh);
482 	gru_flush_cache_cbe(cbe);
483 	if (cbk)
484 		gru_flush_cache(cbk);
485 	STAT(tlb_dropin_fail_no_exception);
486 	gru_dbg(grudev, "FAILED non-exception tfh: 0x%p, status %d, state %d\n",
487 		tfh, tfh->status, tfh->state);
488 	return 0;
489 
490 failidle:
491 	/* TFH state was idle  - no miss pending */
492 	gru_flush_cache(tfh);
493 	gru_flush_cache_cbe(cbe);
494 	if (cbk)
495 		gru_flush_cache(cbk);
496 	STAT(tlb_dropin_fail_idle);
497 	gru_dbg(grudev, "FAILED idle tfh: 0x%p, state %d\n", tfh, tfh->state);
498 	return 0;
499 
500 failinval:
501 	/* All errors (atomic & non-atomic) switch CBR to EXCEPTION state */
502 	tfh_exception(tfh);
503 	gru_flush_cache_cbe(cbe);
504 	STAT(tlb_dropin_fail_invalid);
505 	gru_dbg(grudev, "FAILED inval tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
506 	return -EFAULT;
507 
508 failactive:
509 	/* Range invalidate active. Switch to UPM iff atomic */
510 	if (!cbk)
511 		tfh_user_polling_mode(tfh);
512 	else
513 		gru_flush_cache(tfh);
514 	gru_flush_cache_cbe(cbe);
515 	STAT(tlb_dropin_fail_range_active);
516 	gru_dbg(grudev, "FAILED range active: tfh 0x%p, vaddr 0x%lx\n",
517 		tfh, vaddr);
518 	return 1;
519 }
520 
521 /*
522  * Process an external interrupt from the GRU. This interrupt is
523  * caused by a TLB miss.
524  * Note that this is the interrupt handler that is registered with linux
525  * interrupt handlers.
526  */
gru_intr(int chiplet,int blade)527 static irqreturn_t gru_intr(int chiplet, int blade)
528 {
529 	struct gru_state *gru;
530 	struct gru_tlb_fault_map imap, dmap;
531 	struct gru_thread_state *gts;
532 	struct gru_tlb_fault_handle *tfh = NULL;
533 	struct completion *cmp;
534 	int cbrnum, ctxnum;
535 
536 	STAT(intr);
537 
538 	gru = &gru_base[blade]->bs_grus[chiplet];
539 	if (!gru) {
540 		dev_err(grudev, "GRU: invalid interrupt: cpu %d, chiplet %d\n",
541 			raw_smp_processor_id(), chiplet);
542 		return IRQ_NONE;
543 	}
544 	get_clear_fault_map(gru, &imap, &dmap);
545 	gru_dbg(grudev,
546 		"cpu %d, chiplet %d, gid %d, imap %016lx %016lx, dmap %016lx %016lx\n",
547 		smp_processor_id(), chiplet, gru->gs_gid,
548 		imap.fault_bits[0], imap.fault_bits[1],
549 		dmap.fault_bits[0], dmap.fault_bits[1]);
550 
551 	for_each_cbr_in_tfm(cbrnum, dmap.fault_bits) {
552 		STAT(intr_cbr);
553 		cmp = gru->gs_blade->bs_async_wq;
554 		if (cmp)
555 			complete(cmp);
556 		gru_dbg(grudev, "gid %d, cbr_done %d, done %d\n",
557 			gru->gs_gid, cbrnum, cmp ? cmp->done : -1);
558 	}
559 
560 	for_each_cbr_in_tfm(cbrnum, imap.fault_bits) {
561 		STAT(intr_tfh);
562 		tfh = get_tfh_by_index(gru, cbrnum);
563 		prefetchw(tfh);	/* Helps on hdw, required for emulator */
564 
565 		/*
566 		 * When hardware sets a bit in the faultmap, it implicitly
567 		 * locks the GRU context so that it cannot be unloaded.
568 		 * The gts cannot change until a TFH start/writestart command
569 		 * is issued.
570 		 */
571 		ctxnum = tfh->ctxnum;
572 		gts = gru->gs_gts[ctxnum];
573 
574 		/* Spurious interrupts can cause this. Ignore. */
575 		if (!gts) {
576 			STAT(intr_spurious);
577 			continue;
578 		}
579 
580 		/*
581 		 * This is running in interrupt context. Trylock the mmap_sem.
582 		 * If it fails, retry the fault in user context.
583 		 */
584 		gts->ustats.fmm_tlbmiss++;
585 		if (!gts->ts_force_cch_reload &&
586 					down_read_trylock(&gts->ts_mm->mmap_sem)) {
587 			gru_try_dropin(gru, gts, tfh, NULL);
588 			up_read(&gts->ts_mm->mmap_sem);
589 		} else {
590 			tfh_user_polling_mode(tfh);
591 			STAT(intr_mm_lock_failed);
592 		}
593 	}
594 	return IRQ_HANDLED;
595 }
596 
gru0_intr(int irq,void * dev_id)597 irqreturn_t gru0_intr(int irq, void *dev_id)
598 {
599 	return gru_intr(0, uv_numa_blade_id());
600 }
601 
gru1_intr(int irq,void * dev_id)602 irqreturn_t gru1_intr(int irq, void *dev_id)
603 {
604 	return gru_intr(1, uv_numa_blade_id());
605 }
606 
gru_intr_mblade(int irq,void * dev_id)607 irqreturn_t gru_intr_mblade(int irq, void *dev_id)
608 {
609 	int blade;
610 
611 	for_each_possible_blade(blade) {
612 		if (uv_blade_nr_possible_cpus(blade))
613 			continue;
614 		 gru_intr(0, blade);
615 		 gru_intr(1, blade);
616 	}
617 	return IRQ_HANDLED;
618 }
619 
620 
gru_user_dropin(struct gru_thread_state * gts,struct gru_tlb_fault_handle * tfh,void * cb)621 static int gru_user_dropin(struct gru_thread_state *gts,
622 			   struct gru_tlb_fault_handle *tfh,
623 			   void *cb)
624 {
625 	struct gru_mm_struct *gms = gts->ts_gms;
626 	int ret;
627 
628 	gts->ustats.upm_tlbmiss++;
629 	while (1) {
630 		wait_event(gms->ms_wait_queue,
631 			   atomic_read(&gms->ms_range_active) == 0);
632 		prefetchw(tfh);	/* Helps on hdw, required for emulator */
633 		ret = gru_try_dropin(gts->ts_gru, gts, tfh, cb);
634 		if (ret <= 0)
635 			return ret;
636 		STAT(call_os_wait_queue);
637 	}
638 }
639 
640 /*
641  * This interface is called as a result of a user detecting a "call OS" bit
642  * in a user CB. Normally means that a TLB fault has occurred.
643  * 	cb - user virtual address of the CB
644  */
gru_handle_user_call_os(unsigned long cb)645 int gru_handle_user_call_os(unsigned long cb)
646 {
647 	struct gru_tlb_fault_handle *tfh;
648 	struct gru_thread_state *gts;
649 	void *cbk;
650 	int ucbnum, cbrnum, ret = -EINVAL;
651 
652 	STAT(call_os);
653 
654 	/* sanity check the cb pointer */
655 	ucbnum = get_cb_number((void *)cb);
656 	if ((cb & (GRU_HANDLE_STRIDE - 1)) || ucbnum >= GRU_NUM_CB)
657 		return -EINVAL;
658 
659 	gts = gru_find_lock_gts(cb);
660 	if (!gts)
661 		return -EINVAL;
662 	gru_dbg(grudev, "address 0x%lx, gid %d, gts 0x%p\n", cb, gts->ts_gru ? gts->ts_gru->gs_gid : -1, gts);
663 
664 	if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE)
665 		goto exit;
666 
667 	gru_check_context_placement(gts);
668 
669 	/*
670 	 * CCH may contain stale data if ts_force_cch_reload is set.
671 	 */
672 	if (gts->ts_gru && gts->ts_force_cch_reload) {
673 		gts->ts_force_cch_reload = 0;
674 		gru_update_cch(gts);
675 	}
676 
677 	ret = -EAGAIN;
678 	cbrnum = thread_cbr_number(gts, ucbnum);
679 	if (gts->ts_gru) {
680 		tfh = get_tfh_by_index(gts->ts_gru, cbrnum);
681 		cbk = get_gseg_base_address_cb(gts->ts_gru->gs_gru_base_vaddr,
682 				gts->ts_ctxnum, ucbnum);
683 		ret = gru_user_dropin(gts, tfh, cbk);
684 	}
685 exit:
686 	gru_unlock_gts(gts);
687 	return ret;
688 }
689 
690 /*
691  * Fetch the exception detail information for a CB that terminated with
692  * an exception.
693  */
gru_get_exception_detail(unsigned long arg)694 int gru_get_exception_detail(unsigned long arg)
695 {
696 	struct control_block_extended_exc_detail excdet;
697 	struct gru_control_block_extended *cbe;
698 	struct gru_thread_state *gts;
699 	int ucbnum, cbrnum, ret;
700 
701 	STAT(user_exception);
702 	if (copy_from_user(&excdet, (void __user *)arg, sizeof(excdet)))
703 		return -EFAULT;
704 
705 	gts = gru_find_lock_gts(excdet.cb);
706 	if (!gts)
707 		return -EINVAL;
708 
709 	gru_dbg(grudev, "address 0x%lx, gid %d, gts 0x%p\n", excdet.cb, gts->ts_gru ? gts->ts_gru->gs_gid : -1, gts);
710 	ucbnum = get_cb_number((void *)excdet.cb);
711 	if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE) {
712 		ret = -EINVAL;
713 	} else if (gts->ts_gru) {
714 		cbrnum = thread_cbr_number(gts, ucbnum);
715 		cbe = get_cbe_by_index(gts->ts_gru, cbrnum);
716 		gru_flush_cache(cbe);	/* CBE not coherent */
717 		sync_core();		/* make sure we are have current data */
718 		excdet.opc = cbe->opccpy;
719 		excdet.exopc = cbe->exopccpy;
720 		excdet.ecause = cbe->ecause;
721 		excdet.exceptdet0 = cbe->idef1upd;
722 		excdet.exceptdet1 = cbe->idef3upd;
723 		excdet.cbrstate = cbe->cbrstate;
724 		excdet.cbrexecstatus = cbe->cbrexecstatus;
725 		gru_flush_cache_cbe(cbe);
726 		ret = 0;
727 	} else {
728 		ret = -EAGAIN;
729 	}
730 	gru_unlock_gts(gts);
731 
732 	gru_dbg(grudev,
733 		"cb 0x%lx, op %d, exopc %d, cbrstate %d, cbrexecstatus 0x%x, ecause 0x%x, "
734 		"exdet0 0x%lx, exdet1 0x%x\n",
735 		excdet.cb, excdet.opc, excdet.exopc, excdet.cbrstate, excdet.cbrexecstatus,
736 		excdet.ecause, excdet.exceptdet0, excdet.exceptdet1);
737 	if (!ret && copy_to_user((void __user *)arg, &excdet, sizeof(excdet)))
738 		ret = -EFAULT;
739 	return ret;
740 }
741 
742 /*
743  * User request to unload a context. Content is saved for possible reload.
744  */
gru_unload_all_contexts(void)745 static int gru_unload_all_contexts(void)
746 {
747 	struct gru_thread_state *gts;
748 	struct gru_state *gru;
749 	int gid, ctxnum;
750 
751 	if (!capable(CAP_SYS_ADMIN))
752 		return -EPERM;
753 	foreach_gid(gid) {
754 		gru = GID_TO_GRU(gid);
755 		spin_lock(&gru->gs_lock);
756 		for (ctxnum = 0; ctxnum < GRU_NUM_CCH; ctxnum++) {
757 			gts = gru->gs_gts[ctxnum];
758 			if (gts && mutex_trylock(&gts->ts_ctxlock)) {
759 				spin_unlock(&gru->gs_lock);
760 				gru_unload_context(gts, 1);
761 				mutex_unlock(&gts->ts_ctxlock);
762 				spin_lock(&gru->gs_lock);
763 			}
764 		}
765 		spin_unlock(&gru->gs_lock);
766 	}
767 	return 0;
768 }
769 
gru_user_unload_context(unsigned long arg)770 int gru_user_unload_context(unsigned long arg)
771 {
772 	struct gru_thread_state *gts;
773 	struct gru_unload_context_req req;
774 
775 	STAT(user_unload_context);
776 	if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
777 		return -EFAULT;
778 
779 	gru_dbg(grudev, "gseg 0x%lx\n", req.gseg);
780 
781 	if (!req.gseg)
782 		return gru_unload_all_contexts();
783 
784 	gts = gru_find_lock_gts(req.gseg);
785 	if (!gts)
786 		return -EINVAL;
787 
788 	if (gts->ts_gru)
789 		gru_unload_context(gts, 1);
790 	gru_unlock_gts(gts);
791 
792 	return 0;
793 }
794 
795 /*
796  * User request to flush a range of virtual addresses from the GRU TLB
797  * (Mainly for testing).
798  */
gru_user_flush_tlb(unsigned long arg)799 int gru_user_flush_tlb(unsigned long arg)
800 {
801 	struct gru_thread_state *gts;
802 	struct gru_flush_tlb_req req;
803 	struct gru_mm_struct *gms;
804 
805 	STAT(user_flush_tlb);
806 	if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
807 		return -EFAULT;
808 
809 	gru_dbg(grudev, "gseg 0x%lx, vaddr 0x%lx, len 0x%lx\n", req.gseg,
810 		req.vaddr, req.len);
811 
812 	gts = gru_find_lock_gts(req.gseg);
813 	if (!gts)
814 		return -EINVAL;
815 
816 	gms = gts->ts_gms;
817 	gru_unlock_gts(gts);
818 	gru_flush_tlb_range(gms, req.vaddr, req.len);
819 
820 	return 0;
821 }
822 
823 /*
824  * Fetch GSEG statisticss
825  */
gru_get_gseg_statistics(unsigned long arg)826 long gru_get_gseg_statistics(unsigned long arg)
827 {
828 	struct gru_thread_state *gts;
829 	struct gru_get_gseg_statistics_req req;
830 
831 	if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
832 		return -EFAULT;
833 
834 	/*
835 	 * The library creates arrays of contexts for threaded programs.
836 	 * If no gts exists in the array, the context has never been used & all
837 	 * statistics are implicitly 0.
838 	 */
839 	gts = gru_find_lock_gts(req.gseg);
840 	if (gts) {
841 		memcpy(&req.stats, &gts->ustats, sizeof(gts->ustats));
842 		gru_unlock_gts(gts);
843 	} else {
844 		memset(&req.stats, 0, sizeof(gts->ustats));
845 	}
846 
847 	if (copy_to_user((void __user *)arg, &req, sizeof(req)))
848 		return -EFAULT;
849 
850 	return 0;
851 }
852 
853 /*
854  * Register the current task as the user of the GSEG slice.
855  * Needed for TLB fault interrupt targeting.
856  */
gru_set_context_option(unsigned long arg)857 int gru_set_context_option(unsigned long arg)
858 {
859 	struct gru_thread_state *gts;
860 	struct gru_set_context_option_req req;
861 	int ret = 0;
862 
863 	STAT(set_context_option);
864 	if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
865 		return -EFAULT;
866 	gru_dbg(grudev, "op %d, gseg 0x%lx, value1 0x%lx\n", req.op, req.gseg, req.val1);
867 
868 	gts = gru_find_lock_gts(req.gseg);
869 	if (!gts) {
870 		gts = gru_alloc_locked_gts(req.gseg);
871 		if (IS_ERR(gts))
872 			return PTR_ERR(gts);
873 	}
874 
875 	switch (req.op) {
876 	case sco_blade_chiplet:
877 		/* Select blade/chiplet for GRU context */
878 		if (req.val0 < -1 || req.val0 >= GRU_CHIPLETS_PER_HUB ||
879 		    req.val1 < -1 || req.val1 >= GRU_MAX_BLADES ||
880 		    (req.val1 >= 0 && !gru_base[req.val1])) {
881 			ret = -EINVAL;
882 		} else {
883 			gts->ts_user_blade_id = req.val1;
884 			gts->ts_user_chiplet_id = req.val0;
885 			gru_check_context_placement(gts);
886 		}
887 		break;
888 	case sco_gseg_owner:
889  		/* Register the current task as the GSEG owner */
890 		gts->ts_tgid_owner = current->tgid;
891 		break;
892 	case sco_cch_req_slice:
893  		/* Set the CCH slice option */
894 		gts->ts_cch_req_slice = req.val1 & 3;
895 		break;
896 	default:
897 		ret = -EINVAL;
898 	}
899 	gru_unlock_gts(gts);
900 
901 	return ret;
902 }
903