xref: /titanic_41/usr/src/uts/sun4v/vm/mach_sfmmu.c (revision af33352a19f03e8fe054fe9b44b19c20aeed0da6)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/types.h>
27 #include <vm/hat.h>
28 #include <vm/hat_sfmmu.h>
29 #include <vm/page.h>
30 #include <sys/pte.h>
31 #include <sys/systm.h>
32 #include <sys/mman.h>
33 #include <sys/sysmacros.h>
34 #include <sys/machparam.h>
35 #include <sys/vtrace.h>
36 #include <sys/kmem.h>
37 #include <sys/mmu.h>
38 #include <sys/cmn_err.h>
39 #include <sys/cpu.h>
40 #include <sys/cpuvar.h>
41 #include <sys/debug.h>
42 #include <sys/lgrp.h>
43 #include <sys/archsystm.h>
44 #include <sys/machsystm.h>
45 #include <sys/vmsystm.h>
46 #include <sys/bitmap.h>
47 #include <vm/rm.h>
48 #include <vm/vm_dep.h>
49 #include <sys/t_lock.h>
50 #include <sys/vm_machparam.h>
51 #include <sys/promif.h>
52 #include <sys/prom_isa.h>
53 #include <sys/prom_plat.h>
54 #include <sys/prom_debug.h>
55 #include <sys/privregs.h>
56 #include <sys/bootconf.h>
57 #include <sys/memlist.h>
58 #include <sys/memlist_plat.h>
59 #include <sys/cpu_module.h>
60 #include <sys/reboot.h>
61 #include <sys/kdi.h>
62 #include <sys/hypervisor_api.h>
63 #include <sys/hsvc.h>
64 
65 /*
66  * External routines and data structures
67  */
68 extern void	sfmmu_cache_flushcolor(int, pfn_t);
69 extern uint_t	mmu_page_sizes;
70 
71 /*
72  * Static routines
73  */
74 static void	sfmmu_set_tlb(void);
75 
76 /*
77  * Global Data:
78  */
79 caddr_t	textva, datava;
80 tte_t	ktext_tte, kdata_tte;		/* ttes for kernel text and data */
81 
82 int	enable_bigktsb = 1;
83 int	shtsb4m_first = 0;
84 
85 tte_t bigktsb_ttes[MAX_BIGKTSB_TTES];
86 int bigktsb_nttes = 0;
87 
88 /*
89  * Controls the logic which enables the use of the
90  * QUAD_LDD_PHYS ASI for TSB accesses.
91  */
92 int	ktsb_phys = 1;
93 
94 #ifdef SET_MMU_STATS
95 struct mmu_stat	mmu_stat_area[NCPU];
96 #endif /* SET_MMU_STATS */
97 
98 #ifdef DEBUG
99 /*
100  * The following two variables control if the hypervisor/hardware will
101  * be used to do the TSB table walk for kernel and user contexts.
102  */
103 int hv_use_0_tsb = 1;
104 int hv_use_non0_tsb = 1;
105 #endif /* DEBUG */
106 
107 static void
108 sfmmu_set_fault_status_area(void)
109 {
110 	caddr_t mmfsa_va;
111 	extern	caddr_t mmu_fault_status_area;
112 
113 	mmfsa_va =
114 	    mmu_fault_status_area + (MMFSA_SIZE  * getprocessorid());
115 	set_mmfsa_scratchpad(mmfsa_va);
116 	prom_set_mmfsa_traptable(&trap_table, va_to_pa(mmfsa_va));
117 }
118 
119 void
120 sfmmu_set_tsbs()
121 {
122 	uint64_t rv;
123 	struct hv_tsb_block *hvbp = &ksfmmup->sfmmu_hvblock;
124 
125 #ifdef DEBUG
126 	if (hv_use_0_tsb == 0)
127 		return;
128 #endif /* DEBUG */
129 
130 	rv = hv_set_ctx0(hvbp->hv_tsb_info_cnt,
131 	    hvbp->hv_tsb_info_pa);
132 	if (rv != H_EOK)
133 		prom_printf("cpu%d: hv_set_ctx0() returned %lx\n",
134 		    getprocessorid(), rv);
135 
136 #ifdef SET_MMU_STATS
137 	ASSERT(getprocessorid() < NCPU);
138 	rv = hv_mmu_set_stat_area(va_to_pa(&mmu_stat_area[getprocessorid()]),
139 	    sizeof (mmu_stat_area[0]));
140 	if (rv != H_EOK)
141 		prom_printf("cpu%d: hv_mmu_set_stat_area() returned %lx\n",
142 		    getprocessorid(), rv);
143 #endif /* SET_MMU_STATS */
144 }
145 
146 /*
147  * This routine remaps the kernel using large ttes
148  * All entries except locked ones will be removed from the tlb.
149  * It assumes that both the text and data segments reside in a separate
150  * 4mb virtual and physical contigous memory chunk.  This routine
151  * is only executed by the first cpu.  The remaining cpus execute
152  * sfmmu_mp_startup() instead.
153  * XXX It assumes that the start of the text segment is KERNELBASE.  It should
154  * actually be based on start.
155  */
156 void
157 sfmmu_remap_kernel(void)
158 {
159 	pfn_t	pfn;
160 	uint_t	attr;
161 	int	flags;
162 
163 	extern char end[];
164 	extern struct as kas;
165 
166 	textva = (caddr_t)(KERNELBASE & MMU_PAGEMASK4M);
167 	pfn = va_to_pfn(textva);
168 	if (pfn == PFN_INVALID)
169 		prom_panic("can't find kernel text pfn");
170 	pfn &= TTE_PFNMASK(TTE4M);
171 
172 	attr = PROC_TEXT | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC;
173 	flags = HAT_LOAD_LOCK | SFMMU_NO_TSBLOAD;
174 	sfmmu_memtte(&ktext_tte, pfn, attr, TTE4M);
175 	/*
176 	 * We set the lock bit in the tte to lock the translation in
177 	 * the tlb.
178 	 */
179 	TTE_SET_LOCKED(&ktext_tte);
180 	sfmmu_tteload(kas.a_hat, &ktext_tte, textva, NULL, flags);
181 
182 	datava = (caddr_t)((uintptr_t)end & MMU_PAGEMASK4M);
183 	pfn = va_to_pfn(datava);
184 	if (pfn == PFN_INVALID)
185 		prom_panic("can't find kernel data pfn");
186 	pfn &= TTE_PFNMASK(TTE4M);
187 
188 	attr = PROC_DATA | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC;
189 	sfmmu_memtte(&kdata_tte, pfn, attr, TTE4M);
190 	/*
191 	 * We set the lock bit in the tte to lock the translation in
192 	 * the tlb.  We also set the mod bit to avoid taking dirty bit
193 	 * traps on kernel data.
194 	 */
195 	TTE_SET_LOCKED(&kdata_tte);
196 	TTE_SET_LOFLAGS(&kdata_tte, 0, TTE_HWWR_INT);
197 	sfmmu_tteload(kas.a_hat, &kdata_tte, datava,
198 	    (struct page *)NULL, flags);
199 
200 	/*
201 	 * create bigktsb ttes if necessary.
202 	 */
203 	if (enable_bigktsb) {
204 		int i = 0;
205 		caddr_t va = ktsb_base;
206 		size_t tsbsz = ktsb_sz;
207 		tte_t tte;
208 
209 		ASSERT(va >= datava + MMU_PAGESIZE4M);
210 		ASSERT(tsbsz >= MMU_PAGESIZE4M);
211 		ASSERT(IS_P2ALIGNED(tsbsz, tsbsz));
212 		ASSERT(IS_P2ALIGNED(va, tsbsz));
213 		attr = PROC_DATA | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC;
214 		while (tsbsz != 0) {
215 			ASSERT(i < MAX_BIGKTSB_TTES);
216 			pfn = va_to_pfn(va);
217 			ASSERT(pfn != PFN_INVALID);
218 			ASSERT((pfn & ~TTE_PFNMASK(TTE4M)) == 0);
219 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
220 			ASSERT(TTE_IS_MOD(&tte));
221 			/*
222 			 * No need to lock if we use physical addresses.
223 			 * Since we invalidate the kernel TSB using virtual
224 			 * addresses, it's an optimization to load them now
225 			 * so that we won't have to load them later.
226 			 */
227 			if (!ktsb_phys) {
228 				TTE_SET_LOCKED(&tte);
229 			}
230 			sfmmu_tteload(kas.a_hat, &tte, va, NULL, flags);
231 			bigktsb_ttes[i] = tte;
232 			va += MMU_PAGESIZE4M;
233 			tsbsz -= MMU_PAGESIZE4M;
234 			i++;
235 		}
236 		bigktsb_nttes = i;
237 	}
238 
239 	sfmmu_set_tlb();
240 }
241 
242 /*
243  * Setup the kernel's locked tte's
244  */
245 void
246 sfmmu_set_tlb(void)
247 {
248 	(void) hv_mmu_map_perm_addr(textva, KCONTEXT, *(uint64_t *)&ktext_tte,
249 	    MAP_ITLB | MAP_DTLB);
250 	(void) hv_mmu_map_perm_addr(datava, KCONTEXT, *(uint64_t *)&kdata_tte,
251 	    MAP_DTLB);
252 
253 	if (!ktsb_phys && enable_bigktsb) {
254 		int i;
255 		caddr_t va = ktsb_base;
256 		uint64_t tte;
257 
258 		ASSERT(bigktsb_nttes <= MAX_BIGKTSB_TTES);
259 		for (i = 0; i < bigktsb_nttes; i++) {
260 			tte = *(uint64_t *)&bigktsb_ttes[i];
261 			(void) hv_mmu_map_perm_addr(va, KCONTEXT, tte,
262 			    MAP_DTLB);
263 			va += MMU_PAGESIZE4M;
264 		}
265 	}
266 }
267 
268 /*
269  * This routine is executed by all other cpus except the first one
270  * at initialization time.  It is responsible for taking over the
271  * mmu from the prom.  We follow these steps.
272  * Lock the kernel's ttes in the TLB
273  * Initialize the tsb hardware registers
274  * Take over the trap table
275  * Flush the prom's locked entries from the TLB
276  */
277 void
278 sfmmu_mp_startup(void)
279 {
280 	sfmmu_set_tlb();
281 	setwstate(WSTATE_KERN);
282 	/*
283 	 * sfmmu_set_fault_status_area() takes over trap_table
284 	 */
285 	sfmmu_set_fault_status_area();
286 	sfmmu_set_tsbs();
287 	install_va_to_tte();
288 }
289 
290 void
291 kdi_tlb_page_lock(caddr_t va, int do_dtlb)
292 {
293 	tte_t tte;
294 	pfn_t pfn = va_to_pfn(va);
295 	uint64_t ret;
296 
297 	sfmmu_memtte(&tte, pfn, PROC_TEXT | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC,
298 	    TTE8K);
299 	ret = hv_mmu_map_perm_addr(va, KCONTEXT, *(uint64_t *)&tte,
300 	    MAP_ITLB | (do_dtlb ? MAP_DTLB : 0));
301 
302 	if (ret != H_EOK) {
303 		cmn_err(CE_PANIC, "cpu%d: cannot set permanent mapping for "
304 		    "va=0x%p, hv error code 0x%lx",
305 		    getprocessorid(), (void *)va, ret);
306 	}
307 }
308 
309 void
310 kdi_tlb_page_unlock(caddr_t va, int do_dtlb)
311 {
312 	(void) hv_mmu_unmap_perm_addr(va, KCONTEXT,
313 	    MAP_ITLB | (do_dtlb ? MAP_DTLB : 0));
314 }
315 
316 /*
317  * Clear machine specific TSB information for a user process
318  */
319 void
320 sfmmu_clear_utsbinfo()
321 {
322 	(void) hv_set_ctxnon0(0, NULL);
323 }
324 
325 /*
326  * The tsbord[] array is set up to translate from the order of tsbs in the sfmmu
327  * list to the order of tsbs in the tsb descriptor array passed to the hv, which
328  * is the search order used during Hardware Table Walk.
329  * So, the tsb with index i in the sfmmu list will have search order tsbord[i].
330  *
331  * The order of tsbs in the sfmmu list will be as follows:
332  *
333  *              0 8K - 512K private TSB
334  *              1 4M - 256M private TSB
335  *              2 8K - 512K shared TSB
336  *              3 4M - 256M shared TSB
337  *
338  * Shared TSBs are only used if a process is part of an SCD.
339  *
340  * So, e.g. tsbord[3] = 1;
341  *         corresponds to searching the shared 4M TSB second.
342  *
343  * The search order is selected so that the 8K-512K private TSB is always first.
344  * Currently shared context is not expected to map many 8K-512K pages that cause
345  * TLB misses so we order the shared TSB for 4M-256M pages in front of the
346  * shared TSB for 8K-512K pages. We also expect more TLB misses against private
347  * context mappings than shared context mappings and place private TSBs ahead of
348  * shared TSBs in descriptor order. The shtsb4m_first /etc/system tuneable can
349  * be used to change the default ordering of private and shared TSBs for
350  * 4M-256M pages.
351  */
352 void
353 sfmmu_setup_tsbinfo(sfmmu_t *sfmmup)
354 {
355 	struct tsb_info		*tsbinfop;
356 	hv_tsb_info_t		*tdp;
357 	int			i;
358 	int			j;
359 	int			scd = 0;
360 	int			tsbord[NHV_TSB_INFO];
361 
362 #ifdef DEBUG
363 	ASSERT(max_mmu_ctxdoms > 0);
364 	if (sfmmup != ksfmmup) {
365 		/* Process should have INVALID_CONTEXT on all MMUs. */
366 		for (i = 0; i < max_mmu_ctxdoms; i++) {
367 			ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
368 		}
369 	}
370 #endif
371 
372 	tsbinfop = sfmmup->sfmmu_tsb;
373 	if (tsbinfop == NULL) {
374 		sfmmup->sfmmu_hvblock.hv_tsb_info_pa = (uint64_t)-1;
375 		sfmmup->sfmmu_hvblock.hv_tsb_info_cnt = 0;
376 		return;
377 	}
378 
379 	ASSERT(sfmmup != ksfmmup || sfmmup->sfmmu_scdp == NULL);
380 	ASSERT(sfmmup->sfmmu_scdp == NULL ||
381 	    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb != NULL);
382 
383 	tsbord[0] = 0;
384 	if (sfmmup->sfmmu_scdp == NULL) {
385 		tsbord[1] = 1;
386 	} else {
387 		struct tsb_info *scd8ktsbp =
388 		    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb;
389 		ulong_t shared_4mttecnt = 0;
390 		ulong_t priv_4mttecnt = 0;
391 		int scd4mtsb = (scd8ktsbp->tsb_next != NULL);
392 
393 		for (i = TTE4M; i < MMU_PAGE_SIZES; i++) {
394 			if (scd4mtsb) {
395 				shared_4mttecnt +=
396 				    sfmmup->sfmmu_scdismttecnt[i] +
397 				    sfmmup->sfmmu_scdrttecnt[i];
398 			}
399 			if (tsbinfop->tsb_next != NULL) {
400 				priv_4mttecnt += sfmmup->sfmmu_ttecnt[i] +
401 				    sfmmup->sfmmu_ismttecnt[i];
402 			}
403 		}
404 		if (tsbinfop->tsb_next == NULL) {
405 			if (shared_4mttecnt) {
406 				tsbord[1] = 2;
407 				tsbord[2] = 1;
408 			} else {
409 				tsbord[1] = 1;
410 				tsbord[2] = 2;
411 			}
412 		} else if (priv_4mttecnt) {
413 			if (shared_4mttecnt) {
414 				tsbord[1] = shtsb4m_first ? 2 : 1;
415 				tsbord[2] = 3;
416 				tsbord[3] = shtsb4m_first ? 1 : 2;
417 			} else {
418 				tsbord[1] = 1;
419 				tsbord[2] = 2;
420 				tsbord[3] = 3;
421 			}
422 		} else if (shared_4mttecnt) {
423 			tsbord[1] = 3;
424 			tsbord[2] = 2;
425 			tsbord[3] = 1;
426 		} else {
427 			tsbord[1] = 2;
428 			tsbord[2] = 1;
429 			tsbord[3] = 3;
430 		}
431 	}
432 
433 	ASSERT(tsbinfop != NULL);
434 	for (i = 0; tsbinfop != NULL && i < NHV_TSB_INFO; i++) {
435 		if (i == 0) {
436 			tdp = &sfmmup->sfmmu_hvblock.hv_tsb_info[i];
437 			sfmmup->sfmmu_hvblock.hv_tsb_info_pa = va_to_pa(tdp);
438 		}
439 
440 
441 		j = tsbord[i];
442 
443 		tdp = &sfmmup->sfmmu_hvblock.hv_tsb_info[j];
444 
445 		ASSERT(tsbinfop->tsb_ttesz_mask != 0);
446 		tdp->hvtsb_idxpgsz = lowbit(tsbinfop->tsb_ttesz_mask) - 1;
447 		tdp->hvtsb_assoc = 1;
448 		tdp->hvtsb_ntte = TSB_ENTRIES(tsbinfop->tsb_szc);
449 		tdp->hvtsb_ctx_index = scd;
450 		tdp->hvtsb_pgszs = tsbinfop->tsb_ttesz_mask;
451 		tdp->hvtsb_rsvd = 0;
452 		tdp->hvtsb_pa = tsbinfop->tsb_pa;
453 
454 		tsbinfop = tsbinfop->tsb_next;
455 		if (tsbinfop == NULL && !scd && sfmmup->sfmmu_scdp != NULL) {
456 			tsbinfop =
457 			    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb;
458 			scd = 1;
459 		}
460 	}
461 	sfmmup->sfmmu_hvblock.hv_tsb_info_cnt = i;
462 	ASSERT(tsbinfop == NULL);
463 }
464 
465 /*
466  * Invalidate a TSB via processor specific TSB invalidation routine
467  */
468 void
469 sfmmu_inv_tsb(caddr_t tsb_base, uint_t tsb_bytes)
470 {
471 	extern void cpu_inv_tsb(caddr_t, uint_t);
472 
473 	cpu_inv_tsb(tsb_base, tsb_bytes);
474 }
475 
476 /*
477  * Completely flush the D-cache on all cpus.
478  * Not applicable to sun4v.
479  */
480 void
481 sfmmu_cache_flushall()
482 {
483 }
484 
485 /*
486  * Initialise the real address field in sfmmu_pgsz_order.
487  */
488 void
489 sfmmu_init_pgsz_hv(sfmmu_t *sfmmup)
490 {
491 	int i;
492 
493 	/*
494 	 * Initialize mmu counts for pagesize register programming.
495 	 */
496 	for (i = 0; i < max_mmu_page_sizes; i++) {
497 		sfmmup->sfmmu_mmuttecnt[i] = 0;
498 	}
499 
500 	sfmmup->sfmmu_pgsz_order.hv_pgsz_order_pa =
501 	    va_to_pa(&sfmmup->sfmmu_pgsz_order.hv_pgsz_order);
502 }
503