xref: /titanic_41/usr/src/uts/sun4v/vm/mach_sfmmu.c (revision 9c9af2590af49bb395bc8d2eace0f2d4ea16d165)
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 2008 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 <sys/t_lock.h>
49 #include <sys/vm_machparam.h>
50 #include <sys/promif.h>
51 #include <sys/prom_isa.h>
52 #include <sys/prom_plat.h>
53 #include <sys/prom_debug.h>
54 #include <sys/privregs.h>
55 #include <sys/bootconf.h>
56 #include <sys/memlist.h>
57 #include <sys/memlist_plat.h>
58 #include <sys/cpu_module.h>
59 #include <sys/reboot.h>
60 #include <sys/kdi.h>
61 #include <sys/hypervisor_api.h>
62 #include <sys/hsvc.h>
63 
64 /*
65  * External routines and data structures
66  */
67 extern void	sfmmu_cache_flushcolor(int, pfn_t);
68 extern uint_t	mmu_page_sizes;
69 
70 /*
71  * Static routines
72  */
73 static void	sfmmu_set_tlb(void);
74 
75 /*
76  * Global Data:
77  */
78 caddr_t	textva, datava;
79 tte_t	ktext_tte, kdata_tte;		/* ttes for kernel text and data */
80 
81 int	enable_bigktsb = 1;
82 int	shtsb4m_first = 0;
83 
84 tte_t bigktsb_ttes[MAX_BIGKTSB_TTES];
85 int bigktsb_nttes = 0;
86 
87 /*
88  * Controls the logic which enables the use of the
89  * QUAD_LDD_PHYS ASI for TSB accesses.
90  */
91 int	ktsb_phys = 1;
92 
93 #ifdef SET_MMU_STATS
94 struct mmu_stat	mmu_stat_area[NCPU];
95 #endif /* SET_MMU_STATS */
96 
97 #ifdef DEBUG
98 /*
99  * The following two variables control if the hypervisor/hardware will
100  * be used to do the TSB table walk for kernel and user contexts.
101  */
102 int hv_use_0_tsb = 1;
103 int hv_use_non0_tsb = 1;
104 #endif /* DEBUG */
105 
106 static void
107 sfmmu_set_fault_status_area(void)
108 {
109 	caddr_t mmfsa_va;
110 	extern	caddr_t mmu_fault_status_area;
111 
112 	mmfsa_va =
113 	    mmu_fault_status_area + (MMFSA_SIZE  * getprocessorid());
114 	set_mmfsa_scratchpad(mmfsa_va);
115 	prom_set_mmfsa_traptable(&trap_table, va_to_pa(mmfsa_va));
116 }
117 
118 void
119 sfmmu_set_tsbs()
120 {
121 	uint64_t rv;
122 	struct hv_tsb_block *hvbp = &ksfmmup->sfmmu_hvblock;
123 
124 #ifdef DEBUG
125 	if (hv_use_0_tsb == 0)
126 		return;
127 #endif /* DEBUG */
128 
129 	rv = hv_set_ctx0(hvbp->hv_tsb_info_cnt,
130 	    hvbp->hv_tsb_info_pa);
131 	if (rv != H_EOK)
132 		prom_printf("cpu%d: hv_set_ctx0() returned %lx\n",
133 		    getprocessorid(), rv);
134 
135 #ifdef SET_MMU_STATS
136 	ASSERT(getprocessorid() < NCPU);
137 	rv = hv_mmu_set_stat_area(va_to_pa(&mmu_stat_area[getprocessorid()]),
138 	    sizeof (mmu_stat_area[0]));
139 	if (rv != H_EOK)
140 		prom_printf("cpu%d: hv_mmu_set_stat_area() returned %lx\n",
141 		    getprocessorid(), rv);
142 #endif /* SET_MMU_STATS */
143 }
144 
145 /*
146  * This routine remaps the kernel using large ttes
147  * All entries except locked ones will be removed from the tlb.
148  * It assumes that both the text and data segments reside in a separate
149  * 4mb virtual and physical contigous memory chunk.  This routine
150  * is only executed by the first cpu.  The remaining cpus execute
151  * sfmmu_mp_startup() instead.
152  * XXX It assumes that the start of the text segment is KERNELBASE.  It should
153  * actually be based on start.
154  */
155 void
156 sfmmu_remap_kernel(void)
157 {
158 	pfn_t	pfn;
159 	uint_t	attr;
160 	int	flags;
161 
162 	extern char end[];
163 	extern struct as kas;
164 
165 	textva = (caddr_t)(KERNELBASE & MMU_PAGEMASK4M);
166 	pfn = va_to_pfn(textva);
167 	if (pfn == PFN_INVALID)
168 		prom_panic("can't find kernel text pfn");
169 	pfn &= TTE_PFNMASK(TTE4M);
170 
171 	attr = PROC_TEXT | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC;
172 	flags = HAT_LOAD_LOCK | SFMMU_NO_TSBLOAD;
173 	sfmmu_memtte(&ktext_tte, pfn, attr, TTE4M);
174 	/*
175 	 * We set the lock bit in the tte to lock the translation in
176 	 * the tlb.
177 	 */
178 	TTE_SET_LOCKED(&ktext_tte);
179 	sfmmu_tteload(kas.a_hat, &ktext_tte, textva, NULL, flags);
180 
181 	datava = (caddr_t)((uintptr_t)end & MMU_PAGEMASK4M);
182 	pfn = va_to_pfn(datava);
183 	if (pfn == PFN_INVALID)
184 		prom_panic("can't find kernel data pfn");
185 	pfn &= TTE_PFNMASK(TTE4M);
186 
187 	attr = PROC_DATA | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC;
188 	sfmmu_memtte(&kdata_tte, pfn, attr, TTE4M);
189 	/*
190 	 * We set the lock bit in the tte to lock the translation in
191 	 * the tlb.  We also set the mod bit to avoid taking dirty bit
192 	 * traps on kernel data.
193 	 */
194 	TTE_SET_LOCKED(&kdata_tte);
195 	TTE_SET_LOFLAGS(&kdata_tte, 0, TTE_HWWR_INT);
196 	sfmmu_tteload(kas.a_hat, &kdata_tte, datava,
197 	    (struct page *)NULL, flags);
198 
199 	/*
200 	 * create bigktsb ttes if necessary.
201 	 */
202 	if (enable_bigktsb) {
203 		int i = 0;
204 		caddr_t va = ktsb_base;
205 		size_t tsbsz = ktsb_sz;
206 		tte_t tte;
207 
208 		ASSERT(va >= datava + MMU_PAGESIZE4M);
209 		ASSERT(tsbsz >= MMU_PAGESIZE4M);
210 		ASSERT(IS_P2ALIGNED(tsbsz, tsbsz));
211 		ASSERT(IS_P2ALIGNED(va, tsbsz));
212 		attr = PROC_DATA | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC;
213 		while (tsbsz != 0) {
214 			ASSERT(i < MAX_BIGKTSB_TTES);
215 			pfn = va_to_pfn(va);
216 			ASSERT(pfn != PFN_INVALID);
217 			ASSERT((pfn & ~TTE_PFNMASK(TTE4M)) == 0);
218 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
219 			ASSERT(TTE_IS_MOD(&tte));
220 			/*
221 			 * No need to lock if we use physical addresses.
222 			 * Since we invalidate the kernel TSB using virtual
223 			 * addresses, it's an optimization to load them now
224 			 * so that we won't have to load them later.
225 			 */
226 			if (!ktsb_phys) {
227 				TTE_SET_LOCKED(&tte);
228 			}
229 			sfmmu_tteload(kas.a_hat, &tte, va, NULL, flags);
230 			bigktsb_ttes[i] = tte;
231 			va += MMU_PAGESIZE4M;
232 			tsbsz -= MMU_PAGESIZE4M;
233 			i++;
234 		}
235 		bigktsb_nttes = i;
236 	}
237 
238 	sfmmu_set_tlb();
239 }
240 
241 /*
242  * Setup the kernel's locked tte's
243  */
244 void
245 sfmmu_set_tlb(void)
246 {
247 	(void) hv_mmu_map_perm_addr(textva, KCONTEXT, *(uint64_t *)&ktext_tte,
248 	    MAP_ITLB | MAP_DTLB);
249 	(void) hv_mmu_map_perm_addr(datava, KCONTEXT, *(uint64_t *)&kdata_tte,
250 	    MAP_DTLB);
251 
252 	if (!ktsb_phys && enable_bigktsb) {
253 		int i;
254 		caddr_t va = ktsb_base;
255 		uint64_t tte;
256 
257 		ASSERT(bigktsb_nttes <= MAX_BIGKTSB_TTES);
258 		for (i = 0; i < bigktsb_nttes; i++) {
259 			tte = *(uint64_t *)&bigktsb_ttes[i];
260 			(void) hv_mmu_map_perm_addr(va, KCONTEXT, tte,
261 			    MAP_DTLB);
262 			va += MMU_PAGESIZE4M;
263 		}
264 	}
265 }
266 
267 /*
268  * This routine is executed by all other cpus except the first one
269  * at initialization time.  It is responsible for taking over the
270  * mmu from the prom.  We follow these steps.
271  * Lock the kernel's ttes in the TLB
272  * Initialize the tsb hardware registers
273  * Take over the trap table
274  * Flush the prom's locked entries from the TLB
275  */
276 void
277 sfmmu_mp_startup(void)
278 {
279 	sfmmu_set_tlb();
280 	setwstate(WSTATE_KERN);
281 	/*
282 	 * sfmmu_set_fault_status_area() takes over trap_table
283 	 */
284 	sfmmu_set_fault_status_area();
285 	sfmmu_set_tsbs();
286 	install_va_to_tte();
287 }
288 
289 void
290 kdi_tlb_page_lock(caddr_t va, int do_dtlb)
291 {
292 	tte_t tte;
293 	pfn_t pfn = va_to_pfn(va);
294 	uint64_t ret;
295 
296 	sfmmu_memtte(&tte, pfn, PROC_TEXT | HAT_NOSYNC | HAT_ATTR_NOSOFTEXEC,
297 	    TTE8K);
298 	ret = hv_mmu_map_perm_addr(va, KCONTEXT, *(uint64_t *)&tte,
299 	    MAP_ITLB | (do_dtlb ? MAP_DTLB : 0));
300 
301 	if (ret != H_EOK) {
302 		cmn_err(CE_PANIC, "cpu%d: cannot set permanent mapping for "
303 		    "va=0x%p, hv error code 0x%lx",
304 		    getprocessorid(), (void *)va, ret);
305 	}
306 }
307 
308 void
309 kdi_tlb_page_unlock(caddr_t va, int do_dtlb)
310 {
311 	(void) hv_mmu_unmap_perm_addr(va, KCONTEXT,
312 	    MAP_ITLB | (do_dtlb ? MAP_DTLB : 0));
313 }
314 
315 /*
316  * Clear machine specific TSB information for a user process
317  */
318 void
319 sfmmu_clear_utsbinfo()
320 {
321 	(void) hv_set_ctxnon0(0, NULL);
322 }
323 
324 /*
325  * The tsbord[] array is set up to translate from the order of tsbs in the sfmmu
326  * list to the order of tsbs in the tsb descriptor array passed to the hv, which
327  * is the search order used during Hardware Table Walk.
328  * So, the tsb with index i in the sfmmu list will have search order tsbord[i].
329  *
330  * The order of tsbs in the sfmmu list will be as follows:
331  *
332  *              0 8K - 512K private TSB
333  *              1 4M - 256M private TSB
334  *              2 8K - 512K shared TSB
335  *              3 4M - 256M shared TSB
336  *
337  * Shared TSBs are only used if a process is part of an SCD.
338  *
339  * So, e.g. tsbord[3] = 1;
340  *         corresponds to searching the shared 4M TSB second.
341  *
342  * The search order is selected so that the 8K-512K private TSB is always first.
343  * Currently shared context is not expected to map many 8K-512K pages that cause
344  * TLB misses so we order the shared TSB for 4M-256M pages in front of the
345  * shared TSB for 8K-512K pages. We also expect more TLB misses against private
346  * context mappings than shared context mappings and place private TSBs ahead of
347  * shared TSBs in descriptor order. The shtsb4m_first /etc/system tuneable can
348  * be used to change the default ordering of private and shared TSBs for
349  * 4M-256M pages.
350  */
351 void
352 sfmmu_setup_tsbinfo(sfmmu_t *sfmmup)
353 {
354 	struct tsb_info		*tsbinfop;
355 	hv_tsb_info_t		*tdp;
356 	int			i;
357 	int			j;
358 	int			scd = 0;
359 	int			tsbord[NHV_TSB_INFO];
360 
361 #ifdef DEBUG
362 	ASSERT(max_mmu_ctxdoms > 0);
363 	if (sfmmup != ksfmmup) {
364 		/* Process should have INVALID_CONTEXT on all MMUs. */
365 		for (i = 0; i < max_mmu_ctxdoms; i++) {
366 			ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
367 		}
368 	}
369 #endif
370 
371 	tsbinfop = sfmmup->sfmmu_tsb;
372 	if (tsbinfop == NULL) {
373 		sfmmup->sfmmu_hvblock.hv_tsb_info_pa = (uint64_t)-1;
374 		sfmmup->sfmmu_hvblock.hv_tsb_info_cnt = 0;
375 		return;
376 	}
377 
378 	ASSERT(sfmmup != ksfmmup || sfmmup->sfmmu_scdp == NULL);
379 	ASSERT(sfmmup->sfmmu_scdp == NULL ||
380 	    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb != NULL);
381 
382 	tsbord[0] = 0;
383 	if (sfmmup->sfmmu_scdp == NULL) {
384 		tsbord[1] = 1;
385 	} else {
386 		struct tsb_info *scd8ktsbp =
387 		    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb;
388 		ulong_t shared_4mttecnt = 0;
389 		ulong_t priv_4mttecnt = 0;
390 		int scd4mtsb = (scd8ktsbp->tsb_next != NULL);
391 
392 		for (i = TTE4M; i < MMU_PAGE_SIZES; i++) {
393 			if (scd4mtsb) {
394 				shared_4mttecnt +=
395 				    sfmmup->sfmmu_scdismttecnt[i] +
396 				    sfmmup->sfmmu_scdrttecnt[i];
397 			}
398 			if (tsbinfop->tsb_next != NULL) {
399 				priv_4mttecnt += sfmmup->sfmmu_ttecnt[i] +
400 				    sfmmup->sfmmu_ismttecnt[i];
401 			}
402 		}
403 		if (tsbinfop->tsb_next == NULL) {
404 			if (shared_4mttecnt) {
405 				tsbord[1] = 2;
406 				tsbord[2] = 1;
407 			} else {
408 				tsbord[1] = 1;
409 				tsbord[2] = 2;
410 			}
411 		} else if (priv_4mttecnt) {
412 			if (shared_4mttecnt) {
413 				tsbord[1] = shtsb4m_first ? 2 : 1;
414 				tsbord[2] = 3;
415 				tsbord[3] = shtsb4m_first ? 1 : 2;
416 			} else {
417 				tsbord[1] = 1;
418 				tsbord[2] = 2;
419 				tsbord[3] = 3;
420 			}
421 		} else if (shared_4mttecnt) {
422 			tsbord[1] = 3;
423 			tsbord[2] = 2;
424 			tsbord[3] = 1;
425 		} else {
426 			tsbord[1] = 2;
427 			tsbord[2] = 1;
428 			tsbord[3] = 3;
429 		}
430 	}
431 
432 	ASSERT(tsbinfop != NULL);
433 	for (i = 0; tsbinfop != NULL && i < NHV_TSB_INFO; i++) {
434 		if (i == 0) {
435 			tdp = &sfmmup->sfmmu_hvblock.hv_tsb_info[i];
436 			sfmmup->sfmmu_hvblock.hv_tsb_info_pa = va_to_pa(tdp);
437 		}
438 
439 
440 		j = tsbord[i];
441 
442 		tdp = &sfmmup->sfmmu_hvblock.hv_tsb_info[j];
443 
444 		ASSERT(tsbinfop->tsb_ttesz_mask != 0);
445 		tdp->hvtsb_idxpgsz = lowbit(tsbinfop->tsb_ttesz_mask) - 1;
446 		tdp->hvtsb_assoc = 1;
447 		tdp->hvtsb_ntte = TSB_ENTRIES(tsbinfop->tsb_szc);
448 		tdp->hvtsb_ctx_index = scd;
449 		tdp->hvtsb_pgszs = tsbinfop->tsb_ttesz_mask;
450 		tdp->hvtsb_rsvd = 0;
451 		tdp->hvtsb_pa = tsbinfop->tsb_pa;
452 
453 		tsbinfop = tsbinfop->tsb_next;
454 		if (tsbinfop == NULL && !scd && sfmmup->sfmmu_scdp != NULL) {
455 			tsbinfop =
456 			    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb;
457 			scd = 1;
458 		}
459 	}
460 	sfmmup->sfmmu_hvblock.hv_tsb_info_cnt = i;
461 	ASSERT(tsbinfop == NULL);
462 }
463 
464 /*
465  * Invalidate a TSB via processor specific TSB invalidation routine
466  */
467 void
468 sfmmu_inv_tsb(caddr_t tsb_base, uint_t tsb_bytes)
469 {
470 	extern void cpu_inv_tsb(caddr_t, uint_t);
471 
472 	cpu_inv_tsb(tsb_base, tsb_bytes);
473 }
474 
475 /*
476  * Completely flush the D-cache on all cpus.
477  * Not applicable to sun4v.
478  */
479 void
480 sfmmu_cache_flushall()
481 {
482 }
483