xref: /titanic_44/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 753d2d2e8e7fd0c9bcf736d9bf2f2faf4d6234cc)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
27 /*	All Rights Reserved   */
28 
29 /*
30  * Portions of this source code were derived from Berkeley 4.3 BSD
31  * under license from the Regents of the University of California.
32  */
33 
34 #pragma ident	"%Z%%M%	%I%	%E% SMI"
35 
36 /*
37  * UNIX machine dependent virtual memory support.
38  */
39 
40 #include <sys/vm.h>
41 #include <sys/exec.h>
42 #include <sys/cmn_err.h>
43 #include <sys/cpu_module.h>
44 #include <sys/cpu.h>
45 #include <sys/elf_SPARC.h>
46 #include <sys/archsystm.h>
47 #include <vm/hat_sfmmu.h>
48 #include <sys/memnode.h>
49 #include <sys/mem_cage.h>
50 #include <vm/vm_dep.h>
51 #include <sys/error.h>
52 #include <sys/machsystm.h>
53 #include <vm/seg_kmem.h>
54 
55 uint_t page_colors = 0;
56 uint_t page_colors_mask = 0;
57 uint_t page_coloring_shift = 0;
58 int consistent_coloring;
59 
60 uint_t mmu_page_sizes = MMU_PAGE_SIZES;
61 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
62 uint_t mmu_hashcnt = MAX_HASHCNT;
63 uint_t max_mmu_hashcnt = MAX_HASHCNT;
64 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
65 
66 /*
67  * A bitmask of the page sizes supported by hardware based upon szc.
68  * The base pagesize (p_szc == 0) must always be supported by the hardware.
69  */
70 int mmu_exported_pagesize_mask;
71 uint_t mmu_exported_page_sizes;
72 
73 uint_t szc_2_userszc[MMU_PAGE_SIZES];
74 uint_t userszc_2_szc[MMU_PAGE_SIZES];
75 
76 extern uint_t vac_colors_mask;
77 extern int vac_shift;
78 
79 hw_pagesize_t hw_page_array[] = {
80 	{MMU_PAGESIZE, MMU_PAGESHIFT, MMU_PAGESIZE >> MMU_PAGESHIFT},
81 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, MMU_PAGESIZE64K >> MMU_PAGESHIFT},
82 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K,
83 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
84 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
85 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, MMU_PAGESIZE32M >> MMU_PAGESHIFT},
86 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M,
87 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
88 	{0, 0, 0}
89 };
90 
91 /*
92  * Enable usage of 64k/4M pages for text and 64k pages for initdata for
93  * all sun4v platforms. These variables can be overwritten by the platmod
94  * or the CPU module. User can also change the setting via /etc/system.
95  */
96 
97 int	use_text_pgsz64k = 1;
98 int	use_text_pgsz4m = 1;
99 int	use_initdata_pgsz64k = 1;
100 
101 /*
102  * disable_text_largepages and disable_initdata_largepages bitmaks reflect
103  * both unconfigured and undesirable page sizes. Current implementation
104  * supports 64K and 4M page sizes for text and only 64K for data. Rest of
105  * the page sizes are not currently supported, hence disabled below. In
106  * future, when support is added for any other page size, it should be
107  * reflected below.
108  *
109  * Note that these bitmask can be set in platform or CPU specific code to
110  * disable page sizes that should not be used. These variables normally
111  * shouldn't be changed via /etc/system.
112  *
113  * These bitmasks are also updated within hat_init to reflect unsupported
114  * page sizes on a sun4v processor per mmu_exported_pagesize_mask global
115  * variable.
116  */
117 
118 int disable_text_largepages =
119 	(1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M) | (1 << TTE2G) |
120 	(1 << TTE16G);
121 int disable_initdata_largepages =
122 	(1 << TTE512K) | (1 << TTE4M) | (1 << TTE32M) | (1 << TTE256M) |
123 	(1 << TTE2G) | (1 << TTE16G);
124 
125 /*
126  * Minimum segment size tunables before 64K or 4M large pages
127  * should be used to map it.
128  */
129 size_t text_pgsz64k_minsize = MMU_PAGESIZE64K;
130 size_t text_pgsz4m_minsize = MMU_PAGESIZE4M;
131 size_t initdata_pgsz64k_minsize = MMU_PAGESIZE64K;
132 
133 size_t max_shm_lpsize = MMU_PAGESIZE4M;
134 
135 /* Auto large page tunables. */
136 int auto_lpg_tlb_threshold = 32;
137 int auto_lpg_minszc = TTE64K;
138 int auto_lpg_maxszc = TTE64K;
139 size_t auto_lpg_heap_default = MMU_PAGESIZE64K;
140 size_t auto_lpg_stack_default = MMU_PAGESIZE64K;
141 size_t auto_lpg_va_default = MMU_PAGESIZE64K;
142 size_t auto_lpg_remap_threshold = 0; /* always remap */
143 /*
144  * Number of pages in 1 GB.  Don't enable automatic large pages if we have
145  * fewer than this many pages.
146  */
147 pgcnt_t auto_lpg_min_physmem = 1 << (30 - MMU_PAGESHIFT);
148 
149 /*
150  * map_addr_proc() is the routine called when the system is to
151  * choose an address for the user.  We will pick an address
152  * range which is just below the current stack limit.  The
153  * algorithm used for cache consistency on machines with virtual
154  * address caches is such that offset 0 in the vnode is always
155  * on a shm_alignment'ed aligned address.  Unfortunately, this
156  * means that vnodes which are demand paged will not be mapped
157  * cache consistently with the executable images.  When the
158  * cache alignment for a given object is inconsistent, the
159  * lower level code must manage the translations so that this
160  * is not seen here (at the cost of efficiency, of course).
161  *
162  * addrp is a value/result parameter.
163  *	On input it is a hint from the user to be used in a completely
164  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
165  *	minimal alignment.
166  *
167  *	On output it is NULL if no address can be found in the current
168  *	processes address space or else an address that is currently
169  *	not mapped for len bytes with a page of red zone on either side.
170  *	If vacalign is true, then the selected address will obey the alignment
171  *	constraints of a vac machine based on the given off value.
172  */
173 /*ARGSUSED3*/
174 void
175 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
176     caddr_t userlimit, struct proc *p, uint_t flags)
177 {
178 	struct as *as = p->p_as;
179 	caddr_t addr;
180 	caddr_t base;
181 	size_t slen;
182 	uintptr_t align_amount;
183 	int allow_largepage_alignment = 1;
184 
185 	base = p->p_brkbase;
186 	if (userlimit < as->a_userlimit) {
187 		/*
188 		 * This happens when a program wants to map something in
189 		 * a range that's accessible to a program in a smaller
190 		 * address space.  For example, a 64-bit program might
191 		 * be calling mmap32(2) to guarantee that the returned
192 		 * address is below 4Gbytes.
193 		 */
194 		ASSERT(userlimit > base);
195 		slen = userlimit - base;
196 	} else {
197 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
198 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
199 		    & PAGEMASK);
200 	}
201 	len = (len + PAGEOFFSET) & PAGEMASK;
202 
203 	/*
204 	 * Redzone for each side of the request. This is done to leave
205 	 * one page unmapped between segments. This is not required, but
206 	 * it's useful for the user because if their program strays across
207 	 * a segment boundary, it will catch a fault immediately making
208 	 * debugging a little easier.
209 	 */
210 	len += (2 * PAGESIZE);
211 
212 	/*
213 	 *  If the request is larger than the size of a particular
214 	 *  mmu level, then we use that level to map the request.
215 	 *  But this requires that both the virtual and the physical
216 	 *  addresses be aligned with respect to that level, so we
217 	 *  do the virtual bit of nastiness here.
218 	 *
219 	 *  For 32-bit processes, only those which have specified
220 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
221 	 *  we can potentially waste up to 256MB of the 4G process address
222 	 *  space just for alignment.
223 	 *
224 	 * XXXQ Should iterate trough hw_page_array here to catch
225 	 * all supported pagesizes
226 	 */
227 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
228 	    ((uintptr_t)*addrp) != 0)) {
229 		allow_largepage_alignment = 0;
230 	}
231 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
232 	    allow_largepage_alignment &&
233 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
234 		align_amount = MMU_PAGESIZE256M;
235 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
236 	    allow_largepage_alignment &&
237 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
238 		align_amount = MMU_PAGESIZE32M;
239 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
240 		align_amount = MMU_PAGESIZE4M;
241 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
242 		align_amount = MMU_PAGESIZE512K;
243 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
244 		align_amount = MMU_PAGESIZE64K;
245 	} else  {
246 		/*
247 		 * Align virtual addresses on a 64K boundary to ensure
248 		 * that ELF shared libraries are mapped with the appropriate
249 		 * alignment constraints by the run-time linker.
250 		 */
251 		align_amount = ELF_SPARC_MAXPGSZ;
252 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
253 			((uintptr_t)*addrp < align_amount))
254 			align_amount = (uintptr_t)*addrp;
255 	}
256 
257 	/*
258 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
259 	 */
260 	if (p->p_model == DATAMODEL_LP64)
261 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
262 #ifdef VAC
263 	if (vac && vacalign && (align_amount < shm_alignment))
264 		align_amount = shm_alignment;
265 #endif
266 
267 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
268 		align_amount = (uintptr_t)*addrp;
269 	}
270 	len += align_amount;
271 
272 	/*
273 	 * Look for a large enough hole starting below the stack limit.
274 	 * After finding it, use the upper part.  Addition of PAGESIZE is
275 	 * for the redzone as described above.
276 	 */
277 	as_purge(as);
278 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
279 		caddr_t as_addr;
280 
281 		addr = base + slen - len + PAGESIZE;
282 		as_addr = addr;
283 		/*
284 		 * Round address DOWN to the alignment amount,
285 		 * add the offset, and if this address is less
286 		 * than the original address, add alignment amount.
287 		 */
288 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
289 		addr += (long)(off & (align_amount - 1l));
290 		if (addr < as_addr) {
291 			addr += align_amount;
292 		}
293 
294 		ASSERT(addr <= (as_addr + align_amount));
295 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
296 		    ((uintptr_t)(off & (align_amount - 1l))));
297 		*addrp = addr;
298 
299 	} else {
300 		*addrp = NULL;	/* no more virtual space */
301 	}
302 }
303 
304 /*
305  * Platform-dependent page scrub call.
306  * We call hypervisor to scrub the page.
307  */
308 void
309 pagescrub(page_t *pp, uint_t off, uint_t len)
310 {
311 	uint64_t pa, length;
312 
313 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
314 	length = (uint64_t)len;
315 
316 	(void) mem_scrub(pa, length);
317 }
318 
319 void
320 sync_data_memory(caddr_t va, size_t len)
321 {
322 	/* Call memory sync function */
323 	mem_sync(va, len);
324 }
325 
326 size_t
327 mmu_get_kernel_lpsize(size_t lpsize)
328 {
329 	extern int mmu_exported_pagesize_mask;
330 	uint_t tte;
331 
332 	if (lpsize == 0) {
333 		/* no setting for segkmem_lpsize in /etc/system: use default */
334 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
335 			lpsize = MMU_PAGESIZE256M;
336 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
337 			lpsize = MMU_PAGESIZE4M;
338 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
339 			lpsize = MMU_PAGESIZE64K;
340 		} else {
341 			lpsize = MMU_PAGESIZE;
342 		}
343 
344 		return (lpsize);
345 	}
346 
347 	for (tte = TTE8K; tte <= TTE256M; tte++) {
348 
349 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
350 			continue;
351 
352 		if (lpsize == TTEBYTES(tte))
353 			return (lpsize);
354 	}
355 
356 	lpsize = TTEBYTES(TTE8K);
357 	return (lpsize);
358 }
359 
360 void
361 mmu_init_kcontext()
362 {
363 }
364 
365 /*ARGSUSED*/
366 void
367 mmu_init_kernel_pgsz(struct hat *hat)
368 {
369 }
370 
371 #define	QUANTUM_SIZE	64
372 
373 static	vmem_t	*contig_mem_slab_arena;
374 static	vmem_t	*contig_mem_arena;
375 
376 uint_t contig_mem_slab_size = MMU_PAGESIZE4M;
377 
378 static void *
379 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
380 {
381 	page_t *ppl;
382 	page_t *rootpp;
383 	caddr_t addr = NULL;
384 	pgcnt_t npages = btopr(size);
385 	page_t **ppa;
386 	int pgflags;
387 	int i = 0;
388 
389 
390 	/*
391 	 * The import request should be at least
392 	 * contig_mem_slab_size because that is the
393 	 * slab arena's quantum. The size can be
394 	 * further restricted since contiguous
395 	 * allocations larger than contig_mem_slab_size
396 	 * are not supported here.
397 	 */
398 	ASSERT(size == contig_mem_slab_size);
399 
400 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
401 	    NULL, NULL, vmflag)) == NULL) {
402 		return (NULL);
403 	}
404 
405 	/* The address should be slab-size aligned. */
406 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
407 
408 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
409 		vmem_xfree(vmp, addr, size);
410 		return (NULL);
411 	}
412 
413 	pgflags = PG_EXCL;
414 	if ((vmflag & VM_NOSLEEP) == 0)
415 		pgflags |= PG_WAIT;
416 	if (vmflag & VM_PANIC)
417 		pgflags |= PG_PANIC;
418 	if (vmflag & VM_PUSHPAGE)
419 		pgflags |= PG_PUSHPAGE;
420 
421 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
422 	    pgflags, &kvseg, addr, NULL);
423 
424 	if (ppl == NULL) {
425 		vmem_xfree(vmp, addr, size);
426 		page_unresv(npages);
427 		return (NULL);
428 	}
429 
430 	rootpp = ppl;
431 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
432 	while (ppl != NULL) {
433 		page_t *pp = ppl;
434 		ppa[i++] = pp;
435 		page_sub(&ppl, pp);
436 		ASSERT(page_iolock_assert(pp));
437 		page_io_unlock(pp);
438 	}
439 
440 	/*
441 	 * Load the locked entry.  It's OK to preload the entry into
442 	 * the TSB since we now support large mappings in the kernel TSB.
443 	 */
444 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
445 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
446 
447 	for (--i; i >= 0; --i) {
448 		(void) page_pp_lock(ppa[i], 0, 1);
449 		page_unlock(ppa[i]);
450 	}
451 
452 	kmem_free(ppa, npages * sizeof (page_t *));
453 	return (addr);
454 }
455 
456 void
457 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
458 {
459 	page_t *pp;
460 	caddr_t addr = inaddr;
461 	caddr_t eaddr;
462 	pgcnt_t npages = btopr(size);
463 	pgcnt_t pgs_left = npages;
464 	page_t *rootpp = NULL;
465 
466 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
467 
468 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
469 
470 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
471 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
472 		if (pp == NULL)
473 			panic("contig_mem_span_free: page not found");
474 
475 		ASSERT(PAGE_EXCL(pp));
476 		page_pp_unlock(pp, 0, 1);
477 
478 		if (rootpp == NULL)
479 			rootpp = pp;
480 		if (--pgs_left == 0) {
481 			/*
482 			 * similar logic to segspt_free_pages, but we know we
483 			 * have one large page.
484 			 */
485 			page_destroy_pages(rootpp);
486 		}
487 	}
488 	page_unresv(npages);
489 
490 	if (vmp != NULL)
491 		vmem_xfree(vmp, inaddr, size);
492 }
493 
494 static void *
495 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
496 {
497 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
498 }
499 
500 /*
501  * conting_mem_alloc_align allocates real contiguous memory with the specified
502  * alignment upto contig_mem_slab_size. The alignment must be a power of 2.
503  */
504 void *
505 contig_mem_alloc_align(size_t size, size_t align)
506 {
507 	ASSERT(align <= contig_mem_slab_size);
508 
509 	if ((align & (align - 1)) != 0)
510 		return (NULL);
511 
512 	return (vmem_xalloc(contig_mem_arena, size, align, 0, 0,
513 	    NULL, NULL, VM_NOSLEEP));
514 }
515 
516 /*
517  * Allocates size aligned contiguous memory upto contig_mem_slab_size.
518  * Size must be a power of 2.
519  */
520 void *
521 contig_mem_alloc(size_t size)
522 {
523 	ASSERT((size & (size - 1)) == 0);
524 	return (contig_mem_alloc_align(size, size));
525 }
526 
527 void
528 contig_mem_free(void *vaddr, size_t size)
529 {
530 	vmem_xfree(contig_mem_arena, vaddr, size);
531 }
532 
533 /*
534  * We create a set of stacked vmem arenas to enable us to
535  * allocate large >PAGESIZE chucks of contiguous Real Address space
536  * This is  what the Dynamics TSB support does for TSBs.
537  * The contig_mem_arena import functions are exactly the same as the
538  * TSB kmem_default arena import functions.
539  */
540 void
541 contig_mem_init(void)
542 {
543 
544 	contig_mem_slab_arena = vmem_create("contig_mem_slab_arena", NULL, 0,
545 	    contig_mem_slab_size, contig_vmem_xalloc_aligned_wrapper,
546 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
547 
548 	contig_mem_arena = vmem_create("contig_mem_arena", NULL, 0,
549 	    QUANTUM_SIZE, contig_mem_span_alloc, contig_mem_span_free,
550 	    contig_mem_slab_arena, 0, VM_SLEEP | VM_BESTFIT);
551 
552 }
553