xref: /titanic_41/usr/src/uts/sun4v/vm/mach_vm_dep.c (revision 40e5e17b3361b3eea56a9723071c406894a20b78)
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 2007 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 #include <sys/stack.h>
55 #include <sys/atomic.h>
56 
57 uint_t page_colors = 0;
58 uint_t page_colors_mask = 0;
59 uint_t page_coloring_shift = 0;
60 int consistent_coloring;
61 
62 uint_t mmu_page_sizes = MMU_PAGE_SIZES;
63 uint_t max_mmu_page_sizes = MMU_PAGE_SIZES;
64 uint_t mmu_hashcnt = MAX_HASHCNT;
65 uint_t max_mmu_hashcnt = MAX_HASHCNT;
66 size_t mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
67 
68 /*
69  * A bitmask of the page sizes supported by hardware based upon szc.
70  * The base pagesize (p_szc == 0) must always be supported by the hardware.
71  */
72 int mmu_exported_pagesize_mask;
73 uint_t mmu_exported_page_sizes;
74 
75 uint_t szc_2_userszc[MMU_PAGE_SIZES];
76 uint_t userszc_2_szc[MMU_PAGE_SIZES];
77 
78 extern uint_t vac_colors_mask;
79 extern int vac_shift;
80 
81 hw_pagesize_t hw_page_array[] = {
82 	{MMU_PAGESIZE, MMU_PAGESHIFT, 0, MMU_PAGESIZE >> MMU_PAGESHIFT},
83 	{MMU_PAGESIZE64K, MMU_PAGESHIFT64K, 0,
84 	    MMU_PAGESIZE64K >> MMU_PAGESHIFT},
85 	{MMU_PAGESIZE512K, MMU_PAGESHIFT512K, 0,
86 	    MMU_PAGESIZE512K >> MMU_PAGESHIFT},
87 	{MMU_PAGESIZE4M, MMU_PAGESHIFT4M, 0, MMU_PAGESIZE4M >> MMU_PAGESHIFT},
88 	{MMU_PAGESIZE32M, MMU_PAGESHIFT32M, 0,
89 	    MMU_PAGESIZE32M >> MMU_PAGESHIFT},
90 	{MMU_PAGESIZE256M, MMU_PAGESHIFT256M, 0,
91 	    MMU_PAGESIZE256M >> MMU_PAGESHIFT},
92 	{0, 0, 0, 0}
93 };
94 
95 /*
96  * Maximum page size used to map 64-bit memory segment kmem64_base..kmem64_end
97  */
98 int	max_bootlp_tteszc = TTE256M;
99 
100 /*
101  * Maximum and default segment size tunables for user heap, stack, private
102  * and shared anonymous memory, and user text and initialized data.
103  */
104 size_t max_uheap_lpsize = MMU_PAGESIZE64K;
105 size_t default_uheap_lpsize = MMU_PAGESIZE64K;
106 size_t max_ustack_lpsize = MMU_PAGESIZE64K;
107 size_t default_ustack_lpsize = MMU_PAGESIZE64K;
108 size_t max_privmap_lpsize = MMU_PAGESIZE64K;
109 size_t max_uidata_lpsize = MMU_PAGESIZE64K;
110 size_t max_utext_lpsize = MMU_PAGESIZE4M;
111 size_t max_shm_lpsize = MMU_PAGESIZE4M;
112 
113 /*
114  * map_addr_proc() is the routine called when the system is to
115  * choose an address for the user.  We will pick an address
116  * range which is just below the current stack limit.  The
117  * algorithm used for cache consistency on machines with virtual
118  * address caches is such that offset 0 in the vnode is always
119  * on a shm_alignment'ed aligned address.  Unfortunately, this
120  * means that vnodes which are demand paged will not be mapped
121  * cache consistently with the executable images.  When the
122  * cache alignment for a given object is inconsistent, the
123  * lower level code must manage the translations so that this
124  * is not seen here (at the cost of efficiency, of course).
125  *
126  * addrp is a value/result parameter.
127  *	On input it is a hint from the user to be used in a completely
128  *	machine dependent fashion.  For MAP_ALIGN, addrp contains the
129  *	minimal alignment.
130  *
131  *	On output it is NULL if no address can be found in the current
132  *	processes address space or else an address that is currently
133  *	not mapped for len bytes with a page of red zone on either side.
134  *	If vacalign is true, then the selected address will obey the alignment
135  *	constraints of a vac machine based on the given off value.
136  */
137 /*ARGSUSED3*/
138 void
139 map_addr_proc(caddr_t *addrp, size_t len, offset_t off, int vacalign,
140     caddr_t userlimit, struct proc *p, uint_t flags)
141 {
142 	struct as *as = p->p_as;
143 	caddr_t addr;
144 	caddr_t base;
145 	size_t slen;
146 	uintptr_t align_amount;
147 	int allow_largepage_alignment = 1;
148 
149 	base = p->p_brkbase;
150 	if (userlimit < as->a_userlimit) {
151 		/*
152 		 * This happens when a program wants to map something in
153 		 * a range that's accessible to a program in a smaller
154 		 * address space.  For example, a 64-bit program might
155 		 * be calling mmap32(2) to guarantee that the returned
156 		 * address is below 4Gbytes.
157 		 */
158 		ASSERT(userlimit > base);
159 		slen = userlimit - base;
160 	} else {
161 		slen = p->p_usrstack - base - (((size_t)rctl_enforced_value(
162 		    rctlproc_legacy[RLIMIT_STACK], p->p_rctls, p) + PAGEOFFSET)
163 		    & PAGEMASK);
164 	}
165 	len = (len + PAGEOFFSET) & PAGEMASK;
166 
167 	/*
168 	 * Redzone for each side of the request. This is done to leave
169 	 * one page unmapped between segments. This is not required, but
170 	 * it's useful for the user because if their program strays across
171 	 * a segment boundary, it will catch a fault immediately making
172 	 * debugging a little easier.
173 	 */
174 	len += (2 * PAGESIZE);
175 
176 	/*
177 	 *  If the request is larger than the size of a particular
178 	 *  mmu level, then we use that level to map the request.
179 	 *  But this requires that both the virtual and the physical
180 	 *  addresses be aligned with respect to that level, so we
181 	 *  do the virtual bit of nastiness here.
182 	 *
183 	 *  For 32-bit processes, only those which have specified
184 	 *  MAP_ALIGN or an addr will be aligned on a page size > 4MB. Otherwise
185 	 *  we can potentially waste up to 256MB of the 4G process address
186 	 *  space just for alignment.
187 	 *
188 	 * XXXQ Should iterate trough hw_page_array here to catch
189 	 * all supported pagesizes
190 	 */
191 	if (p->p_model == DATAMODEL_ILP32 && ((flags & MAP_ALIGN) == 0 ||
192 	    ((uintptr_t)*addrp) != 0)) {
193 		allow_largepage_alignment = 0;
194 	}
195 	if ((mmu_page_sizes == max_mmu_page_sizes) &&
196 	    allow_largepage_alignment &&
197 		(len >= MMU_PAGESIZE256M)) {	/* 256MB mappings */
198 		align_amount = MMU_PAGESIZE256M;
199 	} else if ((mmu_page_sizes == max_mmu_page_sizes) &&
200 	    allow_largepage_alignment &&
201 		(len >= MMU_PAGESIZE32M)) {	/* 32MB mappings */
202 		align_amount = MMU_PAGESIZE32M;
203 	} else if (len >= MMU_PAGESIZE4M) {  /* 4MB mappings */
204 		align_amount = MMU_PAGESIZE4M;
205 	} else if (len >= MMU_PAGESIZE512K) { /* 512KB mappings */
206 		align_amount = MMU_PAGESIZE512K;
207 	} else if (len >= MMU_PAGESIZE64K) { /* 64KB mappings */
208 		align_amount = MMU_PAGESIZE64K;
209 	} else  {
210 		/*
211 		 * Align virtual addresses on a 64K boundary to ensure
212 		 * that ELF shared libraries are mapped with the appropriate
213 		 * alignment constraints by the run-time linker.
214 		 */
215 		align_amount = ELF_SPARC_MAXPGSZ;
216 		if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp != 0) &&
217 			((uintptr_t)*addrp < align_amount))
218 			align_amount = (uintptr_t)*addrp;
219 	}
220 
221 	/*
222 	 * 64-bit processes require 1024K alignment of ELF shared libraries.
223 	 */
224 	if (p->p_model == DATAMODEL_LP64)
225 		align_amount = MAX(align_amount, ELF_SPARCV9_MAXPGSZ);
226 #ifdef VAC
227 	if (vac && vacalign && (align_amount < shm_alignment))
228 		align_amount = shm_alignment;
229 #endif
230 
231 	if ((flags & MAP_ALIGN) && ((uintptr_t)*addrp > align_amount)) {
232 		align_amount = (uintptr_t)*addrp;
233 	}
234 	len += align_amount;
235 
236 	/*
237 	 * Look for a large enough hole starting below the stack limit.
238 	 * After finding it, use the upper part.  Addition of PAGESIZE is
239 	 * for the redzone as described above.
240 	 */
241 	as_purge(as);
242 	if (as_gap(as, len, &base, &slen, AH_HI, NULL) == 0) {
243 		caddr_t as_addr;
244 
245 		addr = base + slen - len + PAGESIZE;
246 		as_addr = addr;
247 		/*
248 		 * Round address DOWN to the alignment amount,
249 		 * add the offset, and if this address is less
250 		 * than the original address, add alignment amount.
251 		 */
252 		addr = (caddr_t)((uintptr_t)addr & (~(align_amount - 1l)));
253 		addr += (long)(off & (align_amount - 1l));
254 		if (addr < as_addr) {
255 			addr += align_amount;
256 		}
257 
258 		ASSERT(addr <= (as_addr + align_amount));
259 		ASSERT(((uintptr_t)addr & (align_amount - 1l)) ==
260 		    ((uintptr_t)(off & (align_amount - 1l))));
261 		*addrp = addr;
262 
263 	} else {
264 		*addrp = NULL;	/* no more virtual space */
265 	}
266 }
267 
268 /*
269  * Platform-dependent page scrub call.
270  * We call hypervisor to scrub the page.
271  */
272 void
273 pagescrub(page_t *pp, uint_t off, uint_t len)
274 {
275 	uint64_t pa, length;
276 
277 	pa = (uint64_t)(pp->p_pagenum << MMU_PAGESHIFT + off);
278 	length = (uint64_t)len;
279 
280 	(void) mem_scrub(pa, length);
281 }
282 
283 void
284 sync_data_memory(caddr_t va, size_t len)
285 {
286 	/* Call memory sync function */
287 	(void) mem_sync(va, len);
288 }
289 
290 size_t
291 mmu_get_kernel_lpsize(size_t lpsize)
292 {
293 	extern int mmu_exported_pagesize_mask;
294 	uint_t tte;
295 
296 	if (lpsize == 0) {
297 		/* no setting for segkmem_lpsize in /etc/system: use default */
298 		if (mmu_exported_pagesize_mask & (1 << TTE256M)) {
299 			lpsize = MMU_PAGESIZE256M;
300 		} else if (mmu_exported_pagesize_mask & (1 << TTE4M)) {
301 			lpsize = MMU_PAGESIZE4M;
302 		} else if (mmu_exported_pagesize_mask & (1 << TTE64K)) {
303 			lpsize = MMU_PAGESIZE64K;
304 		} else {
305 			lpsize = MMU_PAGESIZE;
306 		}
307 
308 		return (lpsize);
309 	}
310 
311 	for (tte = TTE8K; tte <= TTE256M; tte++) {
312 
313 		if ((mmu_exported_pagesize_mask & (1 << tte)) == 0)
314 			continue;
315 
316 		if (lpsize == TTEBYTES(tte))
317 			return (lpsize);
318 	}
319 
320 	lpsize = TTEBYTES(TTE8K);
321 	return (lpsize);
322 }
323 
324 void
325 mmu_init_kcontext()
326 {
327 }
328 
329 /*ARGSUSED*/
330 void
331 mmu_init_kernel_pgsz(struct hat *hat)
332 {
333 }
334 
335 #define	QUANTUM_SIZE	64
336 
337 static	vmem_t	*contig_mem_slab_arena;
338 static	vmem_t	*contig_mem_arena;
339 
340 uint_t contig_mem_slab_size = MMU_PAGESIZE4M;
341 
342 static void *
343 contig_mem_span_alloc(vmem_t *vmp, size_t size, int vmflag)
344 {
345 	page_t *ppl;
346 	page_t *rootpp;
347 	caddr_t addr = NULL;
348 	pgcnt_t npages = btopr(size);
349 	page_t **ppa;
350 	int pgflags;
351 	int i = 0;
352 
353 
354 	/*
355 	 * The import request should be at least
356 	 * contig_mem_slab_size because that is the
357 	 * slab arena's quantum. The size can be
358 	 * further restricted since contiguous
359 	 * allocations larger than contig_mem_slab_size
360 	 * are not supported here.
361 	 */
362 	ASSERT(size == contig_mem_slab_size);
363 
364 	if ((addr = vmem_xalloc(vmp, size, size, 0, 0,
365 	    NULL, NULL, vmflag)) == NULL) {
366 		return (NULL);
367 	}
368 
369 	/* The address should be slab-size aligned. */
370 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
371 
372 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
373 		vmem_xfree(vmp, addr, size);
374 		return (NULL);
375 	}
376 
377 	pgflags = PG_EXCL;
378 	if ((vmflag & VM_NOSLEEP) == 0)
379 		pgflags |= PG_WAIT;
380 	if (vmflag & VM_PANIC)
381 		pgflags |= PG_PANIC;
382 	if (vmflag & VM_PUSHPAGE)
383 		pgflags |= PG_PUSHPAGE;
384 
385 	ppl = page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
386 	    pgflags, &kvseg, addr, NULL);
387 
388 	if (ppl == NULL) {
389 		vmem_xfree(vmp, addr, size);
390 		page_unresv(npages);
391 		return (NULL);
392 	}
393 
394 	rootpp = ppl;
395 	ppa = kmem_zalloc(npages * sizeof (page_t *), KM_SLEEP);
396 	while (ppl != NULL) {
397 		page_t *pp = ppl;
398 		ppa[i++] = pp;
399 		page_sub(&ppl, pp);
400 		ASSERT(page_iolock_assert(pp));
401 		page_io_unlock(pp);
402 	}
403 
404 	/*
405 	 * Load the locked entry.  It's OK to preload the entry into
406 	 * the TSB since we now support large mappings in the kernel TSB.
407 	 */
408 	hat_memload_array(kas.a_hat, (caddr_t)rootpp->p_offset, size,
409 	    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC, HAT_LOAD_LOCK);
410 
411 	for (--i; i >= 0; --i) {
412 		(void) page_pp_lock(ppa[i], 0, 1);
413 		page_unlock(ppa[i]);
414 	}
415 
416 	kmem_free(ppa, npages * sizeof (page_t *));
417 	return (addr);
418 }
419 
420 void
421 contig_mem_span_free(vmem_t *vmp, void *inaddr, size_t size)
422 {
423 	page_t *pp;
424 	caddr_t addr = inaddr;
425 	caddr_t eaddr;
426 	pgcnt_t npages = btopr(size);
427 	pgcnt_t pgs_left = npages;
428 	page_t *rootpp = NULL;
429 
430 	ASSERT(((uintptr_t)addr & (contig_mem_slab_size - 1)) == 0);
431 
432 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
433 
434 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
435 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
436 		if (pp == NULL)
437 			panic("contig_mem_span_free: page not found");
438 
439 		ASSERT(PAGE_EXCL(pp));
440 		page_pp_unlock(pp, 0, 1);
441 
442 		if (rootpp == NULL)
443 			rootpp = pp;
444 		if (--pgs_left == 0) {
445 			/*
446 			 * similar logic to segspt_free_pages, but we know we
447 			 * have one large page.
448 			 */
449 			page_destroy_pages(rootpp);
450 		}
451 	}
452 	page_unresv(npages);
453 
454 	if (vmp != NULL)
455 		vmem_xfree(vmp, inaddr, size);
456 }
457 
458 static void *
459 contig_vmem_xalloc_aligned_wrapper(vmem_t *vmp, size_t size, int vmflag)
460 {
461 	return (vmem_xalloc(vmp, size, size, 0, 0, NULL, NULL, vmflag));
462 }
463 
464 /*
465  * conting_mem_alloc_align allocates real contiguous memory with the specified
466  * alignment upto contig_mem_slab_size. The alignment must be a power of 2.
467  */
468 void *
469 contig_mem_alloc_align(size_t size, size_t align)
470 {
471 	ASSERT(align <= contig_mem_slab_size);
472 
473 	if ((align & (align - 1)) != 0)
474 		return (NULL);
475 
476 	return (vmem_xalloc(contig_mem_arena, size, align, 0, 0,
477 	    NULL, NULL, VM_NOSLEEP));
478 }
479 
480 /*
481  * Allocates size aligned contiguous memory upto contig_mem_slab_size.
482  * Size must be a power of 2.
483  */
484 void *
485 contig_mem_alloc(size_t size)
486 {
487 	ASSERT((size & (size - 1)) == 0);
488 	return (contig_mem_alloc_align(size, size));
489 }
490 
491 void
492 contig_mem_free(void *vaddr, size_t size)
493 {
494 	vmem_xfree(contig_mem_arena, vaddr, size);
495 }
496 
497 /*
498  * We create a set of stacked vmem arenas to enable us to
499  * allocate large >PAGESIZE chucks of contiguous Real Address space
500  * This is  what the Dynamics TSB support does for TSBs.
501  * The contig_mem_arena import functions are exactly the same as the
502  * TSB kmem_default arena import functions.
503  */
504 void
505 contig_mem_init(void)
506 {
507 
508 	contig_mem_slab_arena = vmem_create("contig_mem_slab_arena", NULL, 0,
509 	    contig_mem_slab_size, contig_vmem_xalloc_aligned_wrapper,
510 	    vmem_xfree, heap_arena, 0, VM_SLEEP);
511 
512 	contig_mem_arena = vmem_create("contig_mem_arena", NULL, 0,
513 	    QUANTUM_SIZE, contig_mem_span_alloc, contig_mem_span_free,
514 	    contig_mem_slab_arena, 0, VM_SLEEP | VM_BESTFIT);
515 
516 }
517 
518 
519 static uint_t sp_color_stride = 16;
520 static uint_t sp_color_mask = 0x1f;
521 static uint_t sp_current_color = (uint_t)-1;
522 
523 size_t
524 exec_get_spslew(void)
525 {
526 	uint_t spcolor = atomic_inc_32_nv(&sp_current_color);
527 	return ((size_t)((spcolor & sp_color_mask) * SA(sp_color_stride)));
528 }
529