xref: /illumos-gate/usr/src/uts/sun4v/vm/mach_sfmmu.c (revision 080adf91)
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 <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 
63 /*
64  * External routines and data structures
65  */
66 extern void	sfmmu_cache_flushcolor(int, pfn_t);
67 extern uint_t	mmu_page_sizes;
68 
69 /*
70  * Static routines
71  */
72 static void	sfmmu_set_tlb(void);
73 
74 /*
75  * Global Data:
76  */
77 caddr_t	textva, datava;
78 tte_t	ktext_tte, kdata_tte;		/* ttes for kernel text and data */
79 
80 int	enable_bigktsb = 1;
81 int	shtsb4m_first = 0;
82 
83 tte_t bigktsb_ttes[MAX_BIGKTSB_TTES];
84 int bigktsb_nttes = 0;
85 
86 /*
87  * Controls the logic which enables the use of the
88  * QUAD_LDD_PHYS ASI for TSB accesses.
89  */
90 int	ktsb_phys = 1;
91 
92 #ifdef SET_MMU_STATS
93 struct mmu_stat	mmu_stat_area[NCPU];
94 #endif /* SET_MMU_STATS */
95 
96 #ifdef DEBUG
97 /*
98  * The following two variables control if the hypervisor/hardware will
99  * be used to do the TSB table walk for kernel and user contexts.
100  */
101 int hv_use_0_tsb = 1;
102 int hv_use_non0_tsb = 1;
103 #endif /* DEBUG */
104 
105 static void
106 sfmmu_set_fault_status_area(void)
107 {
108 	caddr_t mmfsa_va;
109 	extern	caddr_t mmu_fault_status_area;
110 
111 	mmfsa_va =
112 	    mmu_fault_status_area + (MMFSA_SIZE  * getprocessorid());
113 	set_mmfsa_scratchpad(mmfsa_va);
114 	prom_set_mmfsa_traptable(&trap_table, va_to_pa(mmfsa_va));
115 }
116 
117 void
118 sfmmu_set_tsbs()
119 {
120 	uint64_t rv;
121 	struct hv_tsb_block *hvbp = &ksfmmup->sfmmu_hvblock;
122 
123 #ifdef DEBUG
124 	if (hv_use_0_tsb == 0)
125 		return;
126 #endif /* DEBUG */
127 
128 	rv = hv_set_ctx0(hvbp->hv_tsb_info_cnt,
129 	    hvbp->hv_tsb_info_pa);
130 	if (rv != H_EOK)
131 		prom_printf("cpu%d: hv_set_ctx0() returned %lx\n",
132 		    getprocessorid(), rv);
133 
134 #ifdef SET_MMU_STATS
135 	ASSERT(getprocessorid() < NCPU);
136 	rv = hv_mmu_set_stat_area(va_to_pa(&mmu_stat_area[getprocessorid()]),
137 	    sizeof (mmu_stat_area[0]));
138 	if (rv != H_EOK)
139 		prom_printf("cpu%d: hv_mmu_set_stat_area() returned %lx\n",
140 		    getprocessorid(), rv);
141 #endif /* SET_MMU_STATS */
142 }
143 
144 /*
145  * This routine remaps the kernel using large ttes
146  * All entries except locked ones will be removed from the tlb.
147  * It assumes that both the text and data segments reside in a separate
148  * 4mb virtual and physical contigous memory chunk.  This routine
149  * is only executed by the first cpu.  The remaining cpus execute
150  * sfmmu_mp_startup() instead.
151  * XXX It assumes that the start of the text segment is KERNELBASE.  It should
152  * actually be based on start.
153  */
154 void
155 sfmmu_remap_kernel(void)
156 {
157 	pfn_t	pfn;
158 	uint_t	attr;
159 	int	flags;
160 
161 	extern char end[];
162 	extern struct as kas;
163 
164 	textva = (caddr_t)(KERNELBASE & MMU_PAGEMASK4M);
165 	pfn = va_to_pfn(textva);
166 	if (pfn == PFN_INVALID)
167 		prom_panic("can't find kernel text pfn");
168 	pfn &= TTE_PFNMASK(TTE4M);
169 
170 	attr = PROC_TEXT | HAT_NOSYNC;
171 	flags = HAT_LOAD_LOCK | SFMMU_NO_TSBLOAD;
172 	sfmmu_memtte(&ktext_tte, pfn, attr, TTE4M);
173 	/*
174 	 * We set the lock bit in the tte to lock the translation in
175 	 * the tlb.
176 	 */
177 	TTE_SET_LOCKED(&ktext_tte);
178 	sfmmu_tteload(kas.a_hat, &ktext_tte, textva, NULL, flags);
179 
180 	datava = (caddr_t)((uintptr_t)end & MMU_PAGEMASK4M);
181 	pfn = va_to_pfn(datava);
182 	if (pfn == PFN_INVALID)
183 		prom_panic("can't find kernel data pfn");
184 	pfn &= TTE_PFNMASK(TTE4M);
185 
186 	attr = PROC_DATA | HAT_NOSYNC;
187 	sfmmu_memtte(&kdata_tte, pfn, attr, TTE4M);
188 	/*
189 	 * We set the lock bit in the tte to lock the translation in
190 	 * the tlb.  We also set the mod bit to avoid taking dirty bit
191 	 * traps on kernel data.
192 	 */
193 	TTE_SET_LOCKED(&kdata_tte);
194 	TTE_SET_LOFLAGS(&kdata_tte, 0, TTE_HWWR_INT);
195 	sfmmu_tteload(kas.a_hat, &kdata_tte, datava,
196 	    (struct page *)NULL, flags);
197 
198 	/*
199 	 * create bigktsb ttes if necessary.
200 	 */
201 	if (enable_bigktsb) {
202 		int i = 0;
203 		caddr_t va = ktsb_base;
204 		size_t tsbsz = ktsb_sz;
205 		tte_t tte;
206 
207 		ASSERT(va >= datava + MMU_PAGESIZE4M);
208 		ASSERT(tsbsz >= MMU_PAGESIZE4M);
209 		ASSERT(IS_P2ALIGNED(tsbsz, tsbsz));
210 		ASSERT(IS_P2ALIGNED(va, tsbsz));
211 		attr = PROC_DATA | HAT_NOSYNC;
212 		while (tsbsz != 0) {
213 			ASSERT(i < MAX_BIGKTSB_TTES);
214 			pfn = va_to_pfn(va);
215 			ASSERT(pfn != PFN_INVALID);
216 			ASSERT((pfn & ~TTE_PFNMASK(TTE4M)) == 0);
217 			sfmmu_memtte(&tte, pfn, attr, TTE4M);
218 			ASSERT(TTE_IS_MOD(&tte));
219 			/*
220 			 * No need to lock if we use physical addresses.
221 			 * Since we invalidate the kernel TSB using virtual
222 			 * addresses, it's an optimization to load them now
223 			 * so that we won't have to load them later.
224 			 */
225 			if (!ktsb_phys) {
226 				TTE_SET_LOCKED(&tte);
227 			}
228 			sfmmu_tteload(kas.a_hat, &tte, va, NULL, flags);
229 			bigktsb_ttes[i] = tte;
230 			va += MMU_PAGESIZE4M;
231 			tsbsz -= MMU_PAGESIZE4M;
232 			i++;
233 		}
234 		bigktsb_nttes = i;
235 	}
236 
237 	sfmmu_set_tlb();
238 }
239 
240 /*
241  * Setup the kernel's locked tte's
242  */
243 void
244 sfmmu_set_tlb(void)
245 {
246 	(void) hv_mmu_map_perm_addr(textva, KCONTEXT, *(uint64_t *)&ktext_tte,
247 	    MAP_ITLB | MAP_DTLB);
248 	(void) hv_mmu_map_perm_addr(datava, KCONTEXT, *(uint64_t *)&kdata_tte,
249 	    MAP_DTLB);
250 
251 	if (!ktsb_phys && enable_bigktsb) {
252 		int i;
253 		caddr_t va = ktsb_base;
254 		uint64_t tte;
255 
256 		ASSERT(bigktsb_nttes <= MAX_BIGKTSB_TTES);
257 		for (i = 0; i < bigktsb_nttes; i++) {
258 			tte = *(uint64_t *)&bigktsb_ttes[i];
259 			(void) hv_mmu_map_perm_addr(va, KCONTEXT, tte,
260 			    MAP_DTLB);
261 			va += MMU_PAGESIZE4M;
262 		}
263 	}
264 }
265 
266 /*
267  * This routine is executed by all other cpus except the first one
268  * at initialization time.  It is responsible for taking over the
269  * mmu from the prom.  We follow these steps.
270  * Lock the kernel's ttes in the TLB
271  * Initialize the tsb hardware registers
272  * Take over the trap table
273  * Flush the prom's locked entries from the TLB
274  */
275 void
276 sfmmu_mp_startup(void)
277 {
278 	sfmmu_set_tlb();
279 	setwstate(WSTATE_KERN);
280 	/*
281 	 * sfmmu_set_fault_status_area() takes over trap_table
282 	 */
283 	sfmmu_set_fault_status_area();
284 	sfmmu_set_tsbs();
285 	install_va_to_tte();
286 }
287 
288 void
289 kdi_tlb_page_lock(caddr_t va, int do_dtlb)
290 {
291 	tte_t tte;
292 	pfn_t pfn = va_to_pfn(va);
293 	uint64_t ret;
294 
295 	sfmmu_memtte(&tte, pfn, (PROC_TEXT | HAT_NOSYNC), TTE8K);
296 	ret = hv_mmu_map_perm_addr(va, KCONTEXT, *(uint64_t *)&tte,
297 	    MAP_ITLB | (do_dtlb ? MAP_DTLB : 0));
298 
299 	if (ret != H_EOK) {
300 		cmn_err(CE_PANIC, "cpu%d: cannot set permanent mapping for "
301 		    "va=0x%p, hv error code 0x%lx",
302 		    getprocessorid(), (void *)va, ret);
303 	}
304 }
305 
306 void
307 kdi_tlb_page_unlock(caddr_t va, int do_dtlb)
308 {
309 	(void) hv_mmu_unmap_perm_addr(va, KCONTEXT,
310 	    MAP_ITLB | (do_dtlb ? MAP_DTLB : 0));
311 }
312 
313 /*
314  * Clear machine specific TSB information for a user process
315  */
316 void
317 sfmmu_clear_utsbinfo()
318 {
319 	(void) hv_set_ctxnon0(0, 0);
320 }
321 
322 /*
323  * The tsbord[] array is set up to translate from the order of tsbs in the sfmmu
324  * list to the order of tsbs in the tsb descriptor array passed to the hv, which
325  * is the search order used during Hardware Table Walk.
326  * So, the tsb with index i in the sfmmu list will have search order tsbord[i].
327  *
328  * The order of tsbs in the sfmmu list will be as follows:
329  *
330  *              0 8K - 512K private TSB
331  *              1 4M - 256M private TSB
332  *              2 8K - 512K shared TSB
333  *              3 4M - 256M shared TSB
334  *
335  * Shared TSBs are only used if a process is part of an SCD.
336  *
337  * So, e.g. tsbord[3] = 1;
338  *         corresponds to searching the shared 4M TSB second.
339  *
340  * The search order is selected so that the 8K-512K private TSB is always first.
341  * Currently shared context is not expected to map many 8K-512K pages that cause
342  * TLB misses so we order the shared TSB for 4M-256M pages in front of the
343  * shared TSB for 8K-512K pages. We also expect more TLB misses against private
344  * context mappings than shared context mappings and place private TSBs ahead of
345  * shared TSBs in descriptor order. The shtsb4m_first /etc/system tuneable can
346  * be used to change the default ordering of private and shared TSBs for
347  * 4M-256M pages.
348  */
349 void
350 sfmmu_setup_tsbinfo(sfmmu_t *sfmmup)
351 {
352 	struct tsb_info		*tsbinfop;
353 	hv_tsb_info_t		*tdp;
354 	int			i;
355 	int			j;
356 	int			scd = 0;
357 	int			tsbord[NHV_TSB_INFO];
358 
359 #ifdef DEBUG
360 	ASSERT(max_mmu_ctxdoms > 0);
361 	if (sfmmup != ksfmmup) {
362 		/* Process should have INVALID_CONTEXT on all MMUs. */
363 		for (i = 0; i < max_mmu_ctxdoms; i++) {
364 			ASSERT(sfmmup->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
365 		}
366 	}
367 #endif
368 
369 	tsbinfop = sfmmup->sfmmu_tsb;
370 	if (tsbinfop == NULL) {
371 		sfmmup->sfmmu_hvblock.hv_tsb_info_pa = (uint64_t)-1;
372 		sfmmup->sfmmu_hvblock.hv_tsb_info_cnt = 0;
373 		return;
374 	}
375 
376 	ASSERT(sfmmup != ksfmmup || sfmmup->sfmmu_scdp == NULL);
377 	ASSERT(sfmmup->sfmmu_scdp == NULL ||
378 	    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb != NULL);
379 
380 	tsbord[0] = 0;
381 	if (sfmmup->sfmmu_scdp == NULL) {
382 		tsbord[1] = 1;
383 	} else {
384 		struct tsb_info *scd8ktsbp =
385 		    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb;
386 		ulong_t shared_4mttecnt = 0;
387 		ulong_t priv_4mttecnt = 0;
388 		int scd4mtsb = (scd8ktsbp->tsb_next != NULL);
389 
390 		for (i = TTE4M; i < MMU_PAGE_SIZES; i++) {
391 			if (scd4mtsb) {
392 				shared_4mttecnt +=
393 				    sfmmup->sfmmu_scdismttecnt[i] +
394 				    sfmmup->sfmmu_scdrttecnt[i];
395 			}
396 			if (tsbinfop->tsb_next != NULL) {
397 				priv_4mttecnt += sfmmup->sfmmu_ttecnt[i] +
398 				    sfmmup->sfmmu_ismttecnt[i];
399 			}
400 		}
401 		if (tsbinfop->tsb_next == NULL) {
402 			if (shared_4mttecnt) {
403 				tsbord[1] = 2;
404 				tsbord[2] = 1;
405 			} else {
406 				tsbord[1] = 1;
407 				tsbord[2] = 2;
408 			}
409 		} else if (priv_4mttecnt) {
410 			if (shared_4mttecnt) {
411 				tsbord[1] = shtsb4m_first ? 2 : 1;
412 				tsbord[2] = 3;
413 				tsbord[3] = shtsb4m_first ? 1 : 2;
414 			} else {
415 				tsbord[1] = 1;
416 				tsbord[2] = 2;
417 				tsbord[3] = 3;
418 			}
419 		} else if (shared_4mttecnt) {
420 			tsbord[1] = 3;
421 			tsbord[2] = 2;
422 			tsbord[3] = 1;
423 		} else {
424 			tsbord[1] = 2;
425 			tsbord[2] = 1;
426 			tsbord[3] = 3;
427 		}
428 	}
429 
430 	ASSERT(tsbinfop != NULL);
431 	for (i = 0; tsbinfop != NULL && i < NHV_TSB_INFO; i++) {
432 		if (i == 0) {
433 			tdp = &sfmmup->sfmmu_hvblock.hv_tsb_info[i];
434 			sfmmup->sfmmu_hvblock.hv_tsb_info_pa = va_to_pa(tdp);
435 		}
436 
437 
438 		j = tsbord[i];
439 
440 		tdp = &sfmmup->sfmmu_hvblock.hv_tsb_info[j];
441 
442 		ASSERT(tsbinfop->tsb_ttesz_mask != 0);
443 		tdp->hvtsb_idxpgsz = lowbit(tsbinfop->tsb_ttesz_mask) - 1;
444 		tdp->hvtsb_assoc = 1;
445 		tdp->hvtsb_ntte = TSB_ENTRIES(tsbinfop->tsb_szc);
446 		tdp->hvtsb_ctx_index = scd;
447 		tdp->hvtsb_pgszs = tsbinfop->tsb_ttesz_mask;
448 		tdp->hvtsb_rsvd = 0;
449 		tdp->hvtsb_pa = tsbinfop->tsb_pa;
450 
451 		tsbinfop = tsbinfop->tsb_next;
452 		if (tsbinfop == NULL && !scd && sfmmup->sfmmu_scdp != NULL) {
453 			tsbinfop =
454 			    sfmmup->sfmmu_scdp->scd_sfmmup->sfmmu_tsb;
455 			scd = 1;
456 		}
457 	}
458 	sfmmup->sfmmu_hvblock.hv_tsb_info_cnt = i;
459 	ASSERT(tsbinfop == NULL);
460 }
461 
462 /*
463  * Invalidate a TSB via processor specific TSB invalidation routine
464  */
465 void
466 sfmmu_inv_tsb(caddr_t tsb_base, uint_t tsb_bytes)
467 {
468 	extern void cpu_inv_tsb(caddr_t, uint_t);
469 
470 	cpu_inv_tsb(tsb_base, tsb_bytes);
471 }
472 
473 /*
474  * Completely flush the D-cache on all cpus.
475  * Not applicable to sun4v.
476  */
477 void
478 sfmmu_cache_flushall()
479 {
480 }
481