xref: /illumos-gate/usr/src/uts/common/vm/seg_kmem.c (revision 1a2d662a91cee3bf82f41cd47c7ae6f3825d9db2)
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 (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2019 Joyent, Inc.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/t_lock.h>
28 #include <sys/param.h>
29 #include <sys/sysmacros.h>
30 #include <sys/tuneable.h>
31 #include <sys/systm.h>
32 #include <sys/vm.h>
33 #include <sys/kmem.h>
34 #include <sys/vmem.h>
35 #include <sys/mman.h>
36 #include <sys/cmn_err.h>
37 #include <sys/debug.h>
38 #include <sys/dumphdr.h>
39 #include <sys/bootconf.h>
40 #include <sys/lgrp.h>
41 #include <vm/seg_kmem.h>
42 #include <vm/hat.h>
43 #include <vm/page.h>
44 #include <vm/vm_dep.h>
45 #include <vm/faultcode.h>
46 #include <sys/promif.h>
47 #include <vm/seg_kp.h>
48 #include <sys/bitmap.h>
49 #include <sys/mem_cage.h>
50 
51 #ifdef __sparc
52 #include <sys/ivintr.h>
53 #include <sys/panic.h>
54 #endif
55 
56 /*
57  * seg_kmem is the primary kernel memory segment driver.  It
58  * maps the kernel heap [kernelheap, ekernelheap), module text,
59  * and all memory which was allocated before the VM was initialized
60  * into kas.
61  *
62  * Pages which belong to seg_kmem are hashed into &kvp vnode at
63  * an offset equal to (u_offset_t)virt_addr, and have p_lckcnt >= 1.
64  * They must never be paged out since segkmem_fault() is a no-op to
65  * prevent recursive faults.
66  *
67  * Currently, seg_kmem pages are sharelocked (p_sharelock == 1) on
68  * __x86 and are unlocked (p_sharelock == 0) on __sparc.  Once __x86
69  * supports relocation the #ifdef kludges can be removed.
70  *
71  * seg_kmem pages may be subject to relocation by page_relocate(),
72  * provided that the HAT supports it; if this is so, segkmem_reloc
73  * will be set to a nonzero value. All boot time allocated memory as
74  * well as static memory is considered off limits to relocation.
75  * Pages are "relocatable" if p_state does not have P_NORELOC set, so
76  * we request P_NORELOC pages for memory that isn't safe to relocate.
77  *
78  * The kernel heap is logically divided up into four pieces:
79  *
80  *   heap32_arena is for allocations that require 32-bit absolute
81  *   virtual addresses (e.g. code that uses 32-bit pointers/offsets).
82  *
83  *   heap_core is for allocations that require 2GB *relative*
84  *   offsets; in other words all memory from heap_core is within
85  *   2GB of all other memory from the same arena. This is a requirement
86  *   of the addressing modes of some processors in supervisor code.
87  *
88  *   heap_arena is the general heap arena.
89  *
90  *   static_arena is the static memory arena.  Allocations from it
91  *   are not subject to relocation so it is safe to use the memory
92  *   physical address as well as the virtual address (e.g. the VA to
93  *   PA translations are static).  Caches may import from static_arena;
94  *   all other static memory allocations should use static_alloc_arena.
95  *
96  * On some platforms which have limited virtual address space, seg_kmem
97  * may share [kernelheap, ekernelheap) with seg_kp; if this is so,
98  * segkp_bitmap is non-NULL, and each bit represents a page of virtual
99  * address space which is actually seg_kp mapped.
100  */
101 
102 extern ulong_t *segkp_bitmap;   /* Is set if segkp is from the kernel heap */
103 
104 char *kernelheap;		/* start of primary kernel heap */
105 char *ekernelheap;		/* end of primary kernel heap */
106 struct seg kvseg;		/* primary kernel heap segment */
107 struct seg kvseg_core;		/* "core" kernel heap segment */
108 struct seg kzioseg;		/* Segment for zio mappings */
109 vmem_t *heap_arena;		/* primary kernel heap arena */
110 vmem_t *heap_core_arena;	/* core kernel heap arena */
111 char *heap_core_base;		/* start of core kernel heap arena */
112 char *heap_lp_base;		/* start of kernel large page heap arena */
113 char *heap_lp_end;		/* end of kernel large page heap arena */
114 vmem_t *hat_memload_arena;	/* HAT translation data */
115 struct seg kvseg32;		/* 32-bit kernel heap segment */
116 vmem_t *heap32_arena;		/* 32-bit kernel heap arena */
117 vmem_t *heaptext_arena;		/* heaptext arena */
118 struct as kas;			/* kernel address space */
119 int segkmem_reloc;		/* enable/disable relocatable segkmem pages */
120 vmem_t *static_arena;		/* arena for caches to import static memory */
121 vmem_t *static_alloc_arena;	/* arena for allocating static memory */
122 vmem_t *zio_arena = NULL;	/* arena for allocating zio memory */
123 vmem_t *zio_alloc_arena = NULL;	/* arena for allocating zio memory */
124 
125 #if defined(__amd64)
126 vmem_t *kvmm_arena;		/* arena for vmm VA */
127 struct seg kvmmseg;		/* Segment for vmm memory */
128 #endif
129 
130 /*
131  * seg_kmem driver can map part of the kernel heap with large pages.
132  * Currently this functionality is implemented for sparc platforms only.
133  *
134  * The large page size "segkmem_lpsize" for kernel heap is selected in the
135  * platform specific code. It can also be modified via /etc/system file.
136  * Setting segkmem_lpsize to PAGESIZE in /etc/system disables usage of large
137  * pages for kernel heap. "segkmem_lpshift" is adjusted appropriately to
138  * match segkmem_lpsize.
139  *
140  * At boot time we carve from kernel heap arena a range of virtual addresses
141  * that will be used for large page mappings. This range [heap_lp_base,
142  * heap_lp_end) is set up as a separate vmem arena - "heap_lp_arena". We also
143  * create "kmem_lp_arena" that caches memory already backed up by large
144  * pages. kmem_lp_arena imports virtual segments from heap_lp_arena.
145  */
146 
147 size_t	segkmem_lpsize;
148 int	segkmem_lpszc = 0;
149 
150 size_t  segkmem_kmemlp_quantum = 0x400000;	/* 4MB */
151 size_t  segkmem_heaplp_quantum;
152 vmem_t *heap_lp_arena;
153 static  vmem_t *kmem_lp_arena;
154 static	segkmem_lpcb_t segkmem_lpcb;
155 
156 #ifdef __sparc
157 static  uint_t	segkmem_lpshift = PAGESHIFT;
158 static  vmem_t *segkmem_ppa_arena;
159 #endif
160 
161 /*
162  * We use "segkmem_kmemlp_max" to limit the total amount of physical memory
163  * consumed by the large page heap. By default this parameter is set to 1/8 of
164  * physmem but can be adjusted through /etc/system either directly or
165  * indirectly by setting "segkmem_kmemlp_pcnt" to the percent of physmem
166  * we allow for large page heap.
167  */
168 size_t  segkmem_kmemlp_max;
169 uint_t  segkmem_kmemlp_pcnt;
170 
171 /*
172  * Getting large pages for kernel heap could be problematic due to
173  * physical memory fragmentation. That's why we allow to preallocate
174  * "segkmem_kmemlp_min" bytes at boot time.
175  */
176 size_t	segkmem_kmemlp_min;
177 
178 /*
179  * Throttling is used to avoid expensive tries to allocate large pages
180  * for kernel heap when a lot of succesive attempts to do so fail.
181  */
182 static  ulong_t segkmem_lpthrottle_max = 0x400000;
183 static  ulong_t segkmem_lpthrottle_start = 0x40;
184 static  ulong_t segkmem_use_lpthrottle = 1;
185 
186 /*
187  * Freed pages accumulate on a garbage list until segkmem is ready,
188  * at which point we call segkmem_gc() to free it all.
189  */
190 typedef struct segkmem_gc_list {
191 	struct segkmem_gc_list	*gc_next;
192 	vmem_t			*gc_arena;
193 	size_t			gc_size;
194 } segkmem_gc_list_t;
195 
196 static segkmem_gc_list_t *segkmem_gc_list;
197 
198 /*
199  * Allocations from the hat_memload arena add VM_MEMLOAD to their
200  * vmflags so that segkmem_xalloc() can inform the hat layer that it needs
201  * to take steps to prevent infinite recursion.  HAT allocations also
202  * must be non-relocatable to prevent recursive page faults.
203  */
204 static void *
205 hat_memload_alloc(vmem_t *vmp, size_t size, int flags)
206 {
207 	flags |= (VM_MEMLOAD | VM_NORELOC);
208 	return (segkmem_alloc(vmp, size, flags));
209 }
210 
211 /*
212  * Allocations from static_arena arena (or any other arena that uses
213  * segkmem_alloc_permanent()) require non-relocatable (permanently
214  * wired) memory pages, since these pages are referenced by physical
215  * as well as virtual address.
216  */
217 void *
218 segkmem_alloc_permanent(vmem_t *vmp, size_t size, int flags)
219 {
220 	return (segkmem_alloc(vmp, size, flags | VM_NORELOC));
221 }
222 
223 /*
224  * Initialize kernel heap boundaries.
225  */
226 void
227 kernelheap_init(
228 	void *heap_start,
229 	void *heap_end,
230 	char *first_avail,
231 	void *core_start,
232 	void *core_end)
233 {
234 	uintptr_t textbase;
235 	size_t core_size;
236 	size_t heap_size;
237 	vmem_t *heaptext_parent;
238 	size_t	heap_lp_size = 0;
239 #ifdef __sparc
240 	size_t kmem64_sz = kmem64_aligned_end - kmem64_base;
241 #endif	/* __sparc */
242 
243 	kernelheap = heap_start;
244 	ekernelheap = heap_end;
245 
246 #ifdef __sparc
247 	heap_lp_size = (((uintptr_t)heap_end - (uintptr_t)heap_start) / 4);
248 	/*
249 	 * Bias heap_lp start address by kmem64_sz to reduce collisions
250 	 * in 4M kernel TSB between kmem64 area and heap_lp
251 	 */
252 	kmem64_sz = P2ROUNDUP(kmem64_sz, MMU_PAGESIZE256M);
253 	if (kmem64_sz <= heap_lp_size / 2)
254 		heap_lp_size -= kmem64_sz;
255 	heap_lp_base = ekernelheap - heap_lp_size;
256 	heap_lp_end = heap_lp_base + heap_lp_size;
257 #endif	/* __sparc */
258 
259 	/*
260 	 * If this platform has a 'core' heap area, then the space for
261 	 * overflow module text should be carved out of the end of that
262 	 * heap.  Otherwise, it gets carved out of the general purpose
263 	 * heap.
264 	 */
265 	core_size = (uintptr_t)core_end - (uintptr_t)core_start;
266 	if (core_size > 0) {
267 		ASSERT(core_size >= HEAPTEXT_SIZE);
268 		textbase = (uintptr_t)core_end - HEAPTEXT_SIZE;
269 		core_size -= HEAPTEXT_SIZE;
270 	}
271 #ifndef __sparc
272 	else {
273 		ekernelheap -= HEAPTEXT_SIZE;
274 		textbase = (uintptr_t)ekernelheap;
275 	}
276 #endif
277 
278 	heap_size = (uintptr_t)ekernelheap - (uintptr_t)kernelheap;
279 	heap_arena = vmem_init("heap", kernelheap, heap_size, PAGESIZE,
280 	    segkmem_alloc, segkmem_free);
281 
282 	if (core_size > 0) {
283 		heap_core_arena = vmem_create("heap_core", core_start,
284 		    core_size, PAGESIZE, NULL, NULL, NULL, 0, VM_SLEEP);
285 		heap_core_base = core_start;
286 	} else {
287 		heap_core_arena = heap_arena;
288 		heap_core_base = kernelheap;
289 	}
290 
291 	/*
292 	 * reserve space for the large page heap. If large pages for kernel
293 	 * heap is enabled large page heap arean will be created later in the
294 	 * boot sequence in segkmem_heap_lp_init(). Otherwise the allocated
295 	 * range will be returned back to the heap_arena.
296 	 */
297 	if (heap_lp_size) {
298 		(void) vmem_xalloc(heap_arena, heap_lp_size, PAGESIZE, 0, 0,
299 		    heap_lp_base, heap_lp_end,
300 		    VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
301 	}
302 
303 	/*
304 	 * Remove the already-spoken-for memory range [kernelheap, first_avail).
305 	 */
306 	(void) vmem_xalloc(heap_arena, first_avail - kernelheap, PAGESIZE,
307 	    0, 0, kernelheap, first_avail, VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
308 
309 #ifdef __sparc
310 	heap32_arena = vmem_create("heap32", (void *)SYSBASE32,
311 	    SYSLIMIT32 - SYSBASE32 - HEAPTEXT_SIZE, PAGESIZE, NULL,
312 	    NULL, NULL, 0, VM_SLEEP);
313 	/*
314 	 * Prom claims the physical and virtual resources used by panicbuf
315 	 * and inter_vec_table. So reserve space for panicbuf, intr_vec_table,
316 	 * reserved interrupt vector data structures from 32-bit heap.
317 	 */
318 	(void) vmem_xalloc(heap32_arena, PANICBUFSIZE, PAGESIZE, 0, 0,
319 	    panicbuf, panicbuf + PANICBUFSIZE,
320 	    VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
321 
322 	(void) vmem_xalloc(heap32_arena, IVSIZE, PAGESIZE, 0, 0,
323 	    intr_vec_table, (caddr_t)intr_vec_table + IVSIZE,
324 	    VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
325 
326 	textbase = SYSLIMIT32 - HEAPTEXT_SIZE;
327 	heaptext_parent = NULL;
328 #else	/* __sparc */
329 	heap32_arena = heap_core_arena;
330 	heaptext_parent = heap_core_arena;
331 #endif	/* __sparc */
332 
333 	heaptext_arena = vmem_create("heaptext", (void *)textbase,
334 	    HEAPTEXT_SIZE, PAGESIZE, NULL, NULL, heaptext_parent, 0, VM_SLEEP);
335 
336 	/*
337 	 * Create a set of arenas for memory with static translations
338 	 * (e.g. VA -> PA translations cannot change).  Since using
339 	 * kernel pages by physical address implies it isn't safe to
340 	 * walk across page boundaries, the static_arena quantum must
341 	 * be PAGESIZE.  Any kmem caches that require static memory
342 	 * should source from static_arena, while direct allocations
343 	 * should only use static_alloc_arena.
344 	 */
345 	static_arena = vmem_create("static", NULL, 0, PAGESIZE,
346 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP);
347 	static_alloc_arena = vmem_create("static_alloc", NULL, 0,
348 	    sizeof (uint64_t), vmem_alloc, vmem_free, static_arena,
349 	    0, VM_SLEEP);
350 
351 	/*
352 	 * Create an arena for translation data (ptes, hmes, or hblks).
353 	 * We need an arena for this because hat_memload() is essential
354 	 * to vmem_populate() (see comments in common/os/vmem.c).
355 	 *
356 	 * Note: any kmem cache that allocates from hat_memload_arena
357 	 * must be created as a KMC_NOHASH cache (i.e. no external slab
358 	 * and bufctl structures to allocate) so that slab creation doesn't
359 	 * require anything more than a single vmem_alloc().
360 	 */
361 	hat_memload_arena = vmem_create("hat_memload", NULL, 0, PAGESIZE,
362 	    hat_memload_alloc, segkmem_free, heap_arena, 0,
363 	    VM_SLEEP | VMC_POPULATOR | VMC_DUMPSAFE);
364 }
365 
366 void
367 boot_mapin(caddr_t addr, size_t size)
368 {
369 	caddr_t	 eaddr;
370 	page_t	*pp;
371 	pfn_t	 pfnum;
372 
373 	if (page_resv(btop(size), KM_NOSLEEP) == 0)
374 		panic("boot_mapin: page_resv failed");
375 
376 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
377 		pfnum = va_to_pfn(addr);
378 		if (pfnum == PFN_INVALID)
379 			continue;
380 		if ((pp = page_numtopp_nolock(pfnum)) == NULL)
381 			panic("boot_mapin(): No pp for pfnum = %lx", pfnum);
382 
383 		/*
384 		 * must break up any large pages that may have constituent
385 		 * pages being utilized for BOP_ALLOC()'s before calling
386 		 * page_numtopp().The locking code (ie. page_reclaim())
387 		 * can't handle them
388 		 */
389 		if (pp->p_szc != 0)
390 			page_boot_demote(pp);
391 
392 		pp = page_numtopp(pfnum, SE_EXCL);
393 		if (pp == NULL || PP_ISFREE(pp))
394 			panic("boot_alloc: pp is NULL or free");
395 
396 		/*
397 		 * If the cage is on but doesn't yet contain this page,
398 		 * mark it as non-relocatable.
399 		 */
400 		if (kcage_on && !PP_ISNORELOC(pp)) {
401 			PP_SETNORELOC(pp);
402 			PLCNT_XFER_NORELOC(pp);
403 		}
404 
405 		(void) page_hashin(pp, &kvp, (u_offset_t)(uintptr_t)addr, NULL);
406 		pp->p_lckcnt = 1;
407 #if defined(__x86)
408 		page_downgrade(pp);
409 #else
410 		page_unlock(pp);
411 #endif
412 	}
413 }
414 
415 /*
416  * Get pages from boot and hash them into the kernel's vp.
417  * Used after page structs have been allocated, but before segkmem is ready.
418  */
419 void *
420 boot_alloc(void *inaddr, size_t size, uint_t align)
421 {
422 	caddr_t addr = inaddr;
423 
424 	if (bootops == NULL)
425 		prom_panic("boot_alloc: attempt to allocate memory after "
426 		    "BOP_GONE");
427 
428 	size = ptob(btopr(size));
429 #ifdef __sparc
430 	if (bop_alloc_chunk(addr, size, align) != (caddr_t)addr)
431 		panic("boot_alloc: bop_alloc_chunk failed");
432 #else
433 	if (BOP_ALLOC(bootops, addr, size, align) != addr)
434 		panic("boot_alloc: BOP_ALLOC failed");
435 #endif
436 	boot_mapin((caddr_t)addr, size);
437 	return (addr);
438 }
439 
440 static void
441 segkmem_badop()
442 {
443 	panic("segkmem_badop");
444 }
445 
446 #define	SEGKMEM_BADOP(t)	(t(*)())(uintptr_t)segkmem_badop
447 
448 /*ARGSUSED*/
449 static faultcode_t
450 segkmem_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t size,
451     enum fault_type type, enum seg_rw rw)
452 {
453 	pgcnt_t npages;
454 	spgcnt_t pg;
455 	page_t *pp;
456 	struct vnode *vp = seg->s_data;
457 
458 	ASSERT(RW_READ_HELD(&seg->s_as->a_lock));
459 
460 	if (seg->s_as != &kas || size > seg->s_size ||
461 	    addr < seg->s_base || addr + size > seg->s_base + seg->s_size)
462 		panic("segkmem_fault: bad args");
463 
464 	/*
465 	 * If it is one of segkp pages, call segkp_fault.
466 	 */
467 	if (segkp_bitmap && seg == &kvseg &&
468 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
469 		return (SEGOP_FAULT(hat, segkp, addr, size, type, rw));
470 
471 	if (rw != S_READ && rw != S_WRITE && rw != S_OTHER)
472 		return (FC_NOSUPPORT);
473 
474 	npages = btopr(size);
475 
476 	switch (type) {
477 	case F_SOFTLOCK:	/* lock down already-loaded translations */
478 		for (pg = 0; pg < npages; pg++) {
479 			pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr,
480 			    SE_SHARED);
481 			if (pp == NULL) {
482 				/*
483 				 * Hmm, no page. Does a kernel mapping
484 				 * exist for it?
485 				 */
486 				if (!hat_probe(kas.a_hat, addr)) {
487 					addr -= PAGESIZE;
488 					while (--pg >= 0) {
489 						pp = page_find(vp, (u_offset_t)
490 						    (uintptr_t)addr);
491 						if (pp)
492 							page_unlock(pp);
493 						addr -= PAGESIZE;
494 					}
495 					return (FC_NOMAP);
496 				}
497 			}
498 			addr += PAGESIZE;
499 		}
500 		if (rw == S_OTHER)
501 			hat_reserve(seg->s_as, addr, size);
502 		return (0);
503 	case F_SOFTUNLOCK:
504 		while (npages--) {
505 			pp = page_find(vp, (u_offset_t)(uintptr_t)addr);
506 			if (pp)
507 				page_unlock(pp);
508 			addr += PAGESIZE;
509 		}
510 		return (0);
511 	default:
512 		return (FC_NOSUPPORT);
513 	}
514 	/*NOTREACHED*/
515 }
516 
517 static int
518 segkmem_setprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
519 {
520 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
521 
522 	if (seg->s_as != &kas || size > seg->s_size ||
523 	    addr < seg->s_base || addr + size > seg->s_base + seg->s_size)
524 		panic("segkmem_setprot: bad args");
525 
526 	/*
527 	 * If it is one of segkp pages, call segkp.
528 	 */
529 	if (segkp_bitmap && seg == &kvseg &&
530 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
531 		return (SEGOP_SETPROT(segkp, addr, size, prot));
532 
533 	if (prot == 0)
534 		hat_unload(kas.a_hat, addr, size, HAT_UNLOAD);
535 	else
536 		hat_chgprot(kas.a_hat, addr, size, prot);
537 	return (0);
538 }
539 
540 /*
541  * This is a dummy segkmem function overloaded to call segkp
542  * when segkp is under the heap.
543  */
544 /* ARGSUSED */
545 static int
546 segkmem_checkprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
547 {
548 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
549 
550 	if (seg->s_as != &kas)
551 		segkmem_badop();
552 
553 	/*
554 	 * If it is one of segkp pages, call into segkp.
555 	 */
556 	if (segkp_bitmap && seg == &kvseg &&
557 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
558 		return (SEGOP_CHECKPROT(segkp, addr, size, prot));
559 
560 	segkmem_badop();
561 	return (0);
562 }
563 
564 /*
565  * This is a dummy segkmem function overloaded to call segkp
566  * when segkp is under the heap.
567  */
568 /* ARGSUSED */
569 static int
570 segkmem_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
571 {
572 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
573 
574 	if (seg->s_as != &kas)
575 		segkmem_badop();
576 
577 	/*
578 	 * If it is one of segkp pages, call into segkp.
579 	 */
580 	if (segkp_bitmap && seg == &kvseg &&
581 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
582 		return (SEGOP_KLUSTER(segkp, addr, delta));
583 
584 	segkmem_badop();
585 	return (0);
586 }
587 
588 static void
589 segkmem_xdump_range(void *arg, void *start, size_t size)
590 {
591 	struct as *as = arg;
592 	caddr_t addr = start;
593 	caddr_t addr_end = addr + size;
594 
595 	while (addr < addr_end) {
596 		pfn_t pfn = hat_getpfnum(kas.a_hat, addr);
597 		if (pfn != PFN_INVALID && pfn <= physmax && pf_is_memory(pfn))
598 			dump_addpage(as, addr, pfn);
599 		addr += PAGESIZE;
600 		dump_timeleft = dump_timeout;
601 	}
602 }
603 
604 static void
605 segkmem_dump_range(void *arg, void *start, size_t size)
606 {
607 	caddr_t addr = start;
608 	caddr_t addr_end = addr + size;
609 
610 	/*
611 	 * If we are about to start dumping the range of addresses we
612 	 * carved out of the kernel heap for the large page heap walk
613 	 * heap_lp_arena to find what segments are actually populated
614 	 */
615 	if (SEGKMEM_USE_LARGEPAGES &&
616 	    addr == heap_lp_base && addr_end == heap_lp_end &&
617 	    vmem_size(heap_lp_arena, VMEM_ALLOC) < size) {
618 		vmem_walk(heap_lp_arena, VMEM_ALLOC | VMEM_REENTRANT,
619 		    segkmem_xdump_range, arg);
620 	} else {
621 		segkmem_xdump_range(arg, start, size);
622 	}
623 }
624 
625 static void
626 segkmem_dump(struct seg *seg)
627 {
628 	/*
629 	 * The kernel's heap_arena (represented by kvseg) is a very large
630 	 * VA space, most of which is typically unused.  To speed up dumping
631 	 * we use vmem_walk() to quickly find the pieces of heap_arena that
632 	 * are actually in use.  We do the same for heap32_arena and
633 	 * heap_core.
634 	 *
635 	 * We specify VMEM_REENTRANT to vmem_walk() because dump_addpage()
636 	 * may ultimately need to allocate memory.  Reentrant walks are
637 	 * necessarily imperfect snapshots.  The kernel heap continues
638 	 * to change during a live crash dump, for example.  For a normal
639 	 * crash dump, however, we know that there won't be any other threads
640 	 * messing with the heap.  Therefore, at worst, we may fail to dump
641 	 * the pages that get allocated by the act of dumping; but we will
642 	 * always dump every page that was allocated when the walk began.
643 	 *
644 	 * The other segkmem segments are dense (fully populated), so there's
645 	 * no need to use this technique when dumping them.
646 	 *
647 	 * Note: when adding special dump handling for any new sparsely-
648 	 * populated segments, be sure to add similar handling to the ::kgrep
649 	 * code in mdb.
650 	 */
651 	if (seg == &kvseg) {
652 		vmem_walk(heap_arena, VMEM_ALLOC | VMEM_REENTRANT,
653 		    segkmem_dump_range, seg->s_as);
654 #ifndef __sparc
655 		vmem_walk(heaptext_arena, VMEM_ALLOC | VMEM_REENTRANT,
656 		    segkmem_dump_range, seg->s_as);
657 #endif
658 	} else if (seg == &kvseg_core) {
659 		vmem_walk(heap_core_arena, VMEM_ALLOC | VMEM_REENTRANT,
660 		    segkmem_dump_range, seg->s_as);
661 	} else if (seg == &kvseg32) {
662 		vmem_walk(heap32_arena, VMEM_ALLOC | VMEM_REENTRANT,
663 		    segkmem_dump_range, seg->s_as);
664 		vmem_walk(heaptext_arena, VMEM_ALLOC | VMEM_REENTRANT,
665 		    segkmem_dump_range, seg->s_as);
666 	/*
667 	 * We don't want to dump pages attached to kzioseg since they
668 	 * contain file data from ZFS.  If this page's segment is
669 	 * kzioseg return instead of writing it to the dump device.
670 	 *
671 	 * Same applies to VM memory allocations.
672 	 */
673 	} else if (seg == &kzioseg) {
674 		return;
675 #if defined(__amd64)
676 	} else if (seg == &kvmmseg) {
677 		return;
678 #endif
679 	} else {
680 		segkmem_dump_range(seg->s_as, seg->s_base, seg->s_size);
681 	}
682 }
683 
684 /*
685  * lock/unlock kmem pages over a given range [addr, addr+len).
686  * Returns a shadow list of pages in ppp. If there are holes
687  * in the range (e.g. some of the kernel mappings do not have
688  * underlying page_ts) returns ENOTSUP so that as_pagelock()
689  * will handle the range via as_fault(F_SOFTLOCK).
690  */
691 /*ARGSUSED*/
692 static int
693 segkmem_pagelock(struct seg *seg, caddr_t addr, size_t len,
694     page_t ***ppp, enum lock_type type, enum seg_rw rw)
695 {
696 	page_t **pplist, *pp;
697 	pgcnt_t npages;
698 	spgcnt_t pg;
699 	size_t nb;
700 	struct vnode *vp = seg->s_data;
701 
702 	ASSERT(ppp != NULL);
703 
704 	/*
705 	 * If it is one of segkp pages, call into segkp.
706 	 */
707 	if (segkp_bitmap && seg == &kvseg &&
708 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
709 		return (SEGOP_PAGELOCK(segkp, addr, len, ppp, type, rw));
710 
711 	npages = btopr(len);
712 	nb = sizeof (page_t *) * npages;
713 
714 	if (type == L_PAGEUNLOCK) {
715 		pplist = *ppp;
716 		ASSERT(pplist != NULL);
717 
718 		for (pg = 0; pg < npages; pg++) {
719 			pp = pplist[pg];
720 			page_unlock(pp);
721 		}
722 		kmem_free(pplist, nb);
723 		return (0);
724 	}
725 
726 	ASSERT(type == L_PAGELOCK);
727 
728 	pplist = kmem_alloc(nb, KM_NOSLEEP);
729 	if (pplist == NULL) {
730 		*ppp = NULL;
731 		return (ENOTSUP);	/* take the slow path */
732 	}
733 
734 	for (pg = 0; pg < npages; pg++) {
735 		pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, SE_SHARED);
736 		if (pp == NULL) {
737 			while (--pg >= 0)
738 				page_unlock(pplist[pg]);
739 			kmem_free(pplist, nb);
740 			*ppp = NULL;
741 			return (ENOTSUP);
742 		}
743 		pplist[pg] = pp;
744 		addr += PAGESIZE;
745 	}
746 
747 	*ppp = pplist;
748 	return (0);
749 }
750 
751 /*
752  * This is a dummy segkmem function overloaded to call segkp
753  * when segkp is under the heap.
754  */
755 /* ARGSUSED */
756 static int
757 segkmem_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
758 {
759 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
760 
761 	if (seg->s_as != &kas)
762 		segkmem_badop();
763 
764 	/*
765 	 * If it is one of segkp pages, call into segkp.
766 	 */
767 	if (segkp_bitmap && seg == &kvseg &&
768 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
769 		return (SEGOP_GETMEMID(segkp, addr, memidp));
770 
771 	segkmem_badop();
772 	return (0);
773 }
774 
775 /*ARGSUSED*/
776 static lgrp_mem_policy_info_t *
777 segkmem_getpolicy(struct seg *seg, caddr_t addr)
778 {
779 	return (NULL);
780 }
781 
782 /*ARGSUSED*/
783 static int
784 segkmem_capable(struct seg *seg, segcapability_t capability)
785 {
786 	if (capability == S_CAPABILITY_NOMINFLT)
787 		return (1);
788 	return (0);
789 }
790 
791 struct seg_ops segkmem_ops = {
792 	SEGKMEM_BADOP(int),		/* dup */
793 	SEGKMEM_BADOP(int),		/* unmap */
794 	SEGKMEM_BADOP(void),		/* free */
795 	segkmem_fault,
796 	SEGKMEM_BADOP(faultcode_t),	/* faulta */
797 	segkmem_setprot,
798 	segkmem_checkprot,
799 	segkmem_kluster,
800 	SEGKMEM_BADOP(size_t),		/* swapout */
801 	SEGKMEM_BADOP(int),		/* sync */
802 	SEGKMEM_BADOP(size_t),		/* incore */
803 	SEGKMEM_BADOP(int),		/* lockop */
804 	SEGKMEM_BADOP(int),		/* getprot */
805 	SEGKMEM_BADOP(u_offset_t),	/* getoffset */
806 	SEGKMEM_BADOP(int),		/* gettype */
807 	SEGKMEM_BADOP(int),		/* getvp */
808 	SEGKMEM_BADOP(int),		/* advise */
809 	segkmem_dump,
810 	segkmem_pagelock,
811 	SEGKMEM_BADOP(int),		/* setpgsz */
812 	segkmem_getmemid,
813 	segkmem_getpolicy,		/* getpolicy */
814 	segkmem_capable,		/* capable */
815 	seg_inherit_notsup		/* inherit */
816 };
817 
818 int
819 segkmem_create(struct seg *seg)
820 {
821 	ASSERT(seg->s_as == &kas && RW_WRITE_HELD(&kas.a_lock));
822 	seg->s_ops = &segkmem_ops;
823 	if (seg == &kzioseg)
824 		seg->s_data = &kvps[KV_ZVP];
825 #if defined(__amd64)
826 	else if (seg == &kvmmseg)
827 		seg->s_data = &kvps[KV_VVP];
828 #endif
829 	else
830 		seg->s_data = &kvps[KV_KVP];
831 	kas.a_size += seg->s_size;
832 	return (0);
833 }
834 
835 /*ARGSUSED*/
836 page_t *
837 segkmem_page_create(void *addr, size_t size, int vmflag, void *arg)
838 {
839 	struct seg kseg = { 0 };
840 	int pgflags = PG_EXCL;
841 	struct vnode *vp = arg;
842 
843 	if (vp == NULL)
844 		vp = &kvp;
845 
846 	kseg.s_as = &kas;
847 
848 	if (segkmem_reloc == 0 || (vmflag & VM_NORELOC))
849 		pgflags |= PG_NORELOC;
850 	if ((vmflag & VM_NOSLEEP) == 0)
851 		pgflags |= PG_WAIT;
852 	if (vmflag & VM_PANIC)
853 		pgflags |= PG_PANIC;
854 	if (vmflag & VM_PUSHPAGE)
855 		pgflags |= PG_PUSHPAGE;
856 	if (vmflag & VM_NORMALPRI) {
857 		ASSERT(vmflag & VM_NOSLEEP);
858 		pgflags |= PG_NORMALPRI;
859 	}
860 
861 	return (page_create_va(vp, (u_offset_t)(uintptr_t)addr, size,
862 	    pgflags, &kseg, addr));
863 }
864 
865 /*
866  * Allocate pages to back the virtual address range [addr, addr + size).
867  * If addr is NULL, allocate the virtual address space as well.
868  */
869 void *
870 segkmem_xalloc(vmem_t *vmp, void *inaddr, size_t size, int vmflag, uint_t attr,
871     page_t *(*page_create_func)(void *, size_t, int, void *), void *pcarg)
872 {
873 	page_t *ppl;
874 	caddr_t addr = inaddr;
875 	pgcnt_t npages = btopr(size);
876 	int allocflag;
877 
878 	if (inaddr == NULL && (addr = vmem_alloc(vmp, size, vmflag)) == NULL)
879 		return (NULL);
880 
881 	ASSERT(((uintptr_t)addr & PAGEOFFSET) == 0);
882 
883 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
884 		if (inaddr == NULL)
885 			vmem_free(vmp, addr, size);
886 		return (NULL);
887 	}
888 
889 	ppl = page_create_func(addr, size, vmflag, pcarg);
890 	if (ppl == NULL) {
891 		if (inaddr == NULL)
892 			vmem_free(vmp, addr, size);
893 		page_unresv(npages);
894 		return (NULL);
895 	}
896 
897 	/*
898 	 * Under certain conditions, we need to let the HAT layer know
899 	 * that it cannot safely allocate memory.  Allocations from
900 	 * the hat_memload vmem arena always need this, to prevent
901 	 * infinite recursion.
902 	 *
903 	 * In addition, the x86 hat cannot safely do memory
904 	 * allocations while in vmem_populate(), because there
905 	 * is no simple bound on its usage.
906 	 */
907 	if (vmflag & VM_MEMLOAD)
908 		allocflag = HAT_NO_KALLOC;
909 #if defined(__x86)
910 	else if (vmem_is_populator())
911 		allocflag = HAT_NO_KALLOC;
912 #endif
913 	else
914 		allocflag = 0;
915 
916 	while (ppl != NULL) {
917 		page_t *pp = ppl;
918 		page_sub(&ppl, pp);
919 		ASSERT(page_iolock_assert(pp));
920 		ASSERT(PAGE_EXCL(pp));
921 		page_io_unlock(pp);
922 		hat_memload(kas.a_hat, (caddr_t)(uintptr_t)pp->p_offset, pp,
923 		    (PROT_ALL & ~PROT_USER) | HAT_NOSYNC | attr,
924 		    HAT_LOAD_LOCK | allocflag);
925 		pp->p_lckcnt = 1;
926 #if defined(__x86)
927 		page_downgrade(pp);
928 #else
929 		if (vmflag & SEGKMEM_SHARELOCKED)
930 			page_downgrade(pp);
931 		else
932 			page_unlock(pp);
933 #endif
934 	}
935 
936 	return (addr);
937 }
938 
939 static void *
940 segkmem_alloc_vn(vmem_t *vmp, size_t size, int vmflag, struct vnode *vp)
941 {
942 	void *addr;
943 	segkmem_gc_list_t *gcp, **prev_gcpp;
944 
945 	ASSERT(vp != NULL);
946 
947 	if (kvseg.s_base == NULL) {
948 #ifndef __sparc
949 		if (bootops->bsys_alloc == NULL)
950 			halt("Memory allocation between bop_alloc() and "
951 			    "kmem_alloc().\n");
952 #endif
953 
954 		/*
955 		 * There's not a lot of memory to go around during boot,
956 		 * so recycle it if we can.
957 		 */
958 		for (prev_gcpp = &segkmem_gc_list; (gcp = *prev_gcpp) != NULL;
959 		    prev_gcpp = &gcp->gc_next) {
960 			if (gcp->gc_arena == vmp && gcp->gc_size == size) {
961 				*prev_gcpp = gcp->gc_next;
962 				return (gcp);
963 			}
964 		}
965 
966 		addr = vmem_alloc(vmp, size, vmflag | VM_PANIC);
967 		if (boot_alloc(addr, size, BO_NO_ALIGN) != addr)
968 			panic("segkmem_alloc: boot_alloc failed");
969 		return (addr);
970 	}
971 	return (segkmem_xalloc(vmp, NULL, size, vmflag, 0,
972 	    segkmem_page_create, vp));
973 }
974 
975 void *
976 segkmem_alloc(vmem_t *vmp, size_t size, int vmflag)
977 {
978 	return (segkmem_alloc_vn(vmp, size, vmflag, &kvp));
979 }
980 
981 static void *
982 segkmem_zio_alloc(vmem_t *vmp, size_t size, int vmflag)
983 {
984 	return (segkmem_alloc_vn(vmp, size, vmflag, &kvps[KV_ZVP]));
985 }
986 
987 /*
988  * Any changes to this routine must also be carried over to
989  * devmap_free_pages() in the seg_dev driver. This is because
990  * we currently don't have a special kernel segment for non-paged
991  * kernel memory that is exported by drivers to user space.
992  */
993 void
994 segkmem_xfree(vmem_t *vmp, void *inaddr, size_t size, struct vnode *vp,
995     void (*func)(page_t *))
996 {
997 	page_t *pp;
998 	caddr_t addr = inaddr;
999 	caddr_t eaddr;
1000 	pgcnt_t npages = btopr(size);
1001 
1002 	ASSERT(((uintptr_t)addr & PAGEOFFSET) == 0);
1003 	ASSERT(vp != NULL);
1004 
1005 	if (kvseg.s_base == NULL) {
1006 		segkmem_gc_list_t *gc = inaddr;
1007 		gc->gc_arena = vmp;
1008 		gc->gc_size = size;
1009 		gc->gc_next = segkmem_gc_list;
1010 		segkmem_gc_list = gc;
1011 		return;
1012 	}
1013 
1014 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1015 
1016 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
1017 #if defined(__x86)
1018 		pp = page_find(vp, (u_offset_t)(uintptr_t)addr);
1019 		if (pp == NULL)
1020 			panic("segkmem_free: page not found");
1021 		if (!page_tryupgrade(pp)) {
1022 			/*
1023 			 * Some other thread has a sharelock. Wait for
1024 			 * it to drop the lock so we can free this page.
1025 			 */
1026 			page_unlock(pp);
1027 			pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr,
1028 			    SE_EXCL);
1029 		}
1030 #else
1031 		pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
1032 #endif
1033 		if (pp == NULL)
1034 			panic("segkmem_free: page not found");
1035 		/* Clear p_lckcnt so page_destroy() doesn't update availrmem */
1036 		pp->p_lckcnt = 0;
1037 		if (func)
1038 			func(pp);
1039 		else
1040 			page_destroy(pp, 0);
1041 	}
1042 	if (func == NULL)
1043 		page_unresv(npages);
1044 
1045 	if (vmp != NULL)
1046 		vmem_free(vmp, inaddr, size);
1047 
1048 }
1049 
1050 void
1051 segkmem_free(vmem_t *vmp, void *inaddr, size_t size)
1052 {
1053 	segkmem_xfree(vmp, inaddr, size, &kvp, NULL);
1054 }
1055 
1056 static void
1057 segkmem_zio_free(vmem_t *vmp, void *inaddr, size_t size)
1058 {
1059 	segkmem_xfree(vmp, inaddr, size, &kvps[KV_ZVP], NULL);
1060 }
1061 
1062 void
1063 segkmem_gc(void)
1064 {
1065 	ASSERT(kvseg.s_base != NULL);
1066 	while (segkmem_gc_list != NULL) {
1067 		segkmem_gc_list_t *gc = segkmem_gc_list;
1068 		segkmem_gc_list = gc->gc_next;
1069 		segkmem_free(gc->gc_arena, gc, gc->gc_size);
1070 	}
1071 }
1072 
1073 /*
1074  * Legacy entry points from here to end of file.
1075  */
1076 void
1077 segkmem_mapin(struct seg *seg, void *addr, size_t size, uint_t vprot,
1078     pfn_t pfn, uint_t flags)
1079 {
1080 	hat_unload(seg->s_as->a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1081 	hat_devload(seg->s_as->a_hat, addr, size, pfn, vprot,
1082 	    flags | HAT_LOAD_LOCK);
1083 }
1084 
1085 void
1086 segkmem_mapout(struct seg *seg, void *addr, size_t size)
1087 {
1088 	hat_unload(seg->s_as->a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1089 }
1090 
1091 void *
1092 kmem_getpages(pgcnt_t npages, int kmflag)
1093 {
1094 	return (kmem_alloc(ptob(npages), kmflag));
1095 }
1096 
1097 void
1098 kmem_freepages(void *addr, pgcnt_t npages)
1099 {
1100 	kmem_free(addr, ptob(npages));
1101 }
1102 
1103 #ifdef __sparc
1104 /*
1105  * segkmem_page_create_large() allocates a large page to be used for the kmem
1106  * caches. If kpr is enabled we ask for a relocatable page unless requested
1107  * otherwise. If kpr is disabled we have to ask for a non-reloc page
1108  */
1109 static page_t *
1110 segkmem_page_create_large(void *addr, size_t size, int vmflag, void *arg)
1111 {
1112 	int pgflags;
1113 
1114 	pgflags = PG_EXCL;
1115 
1116 	if (segkmem_reloc == 0 || (vmflag & VM_NORELOC))
1117 		pgflags |= PG_NORELOC;
1118 	if (!(vmflag & VM_NOSLEEP))
1119 		pgflags |= PG_WAIT;
1120 	if (vmflag & VM_PUSHPAGE)
1121 		pgflags |= PG_PUSHPAGE;
1122 	if (vmflag & VM_NORMALPRI)
1123 		pgflags |= PG_NORMALPRI;
1124 
1125 	return (page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
1126 	    pgflags, &kvseg, addr, arg));
1127 }
1128 
1129 /*
1130  * Allocate a large page to back the virtual address range
1131  * [addr, addr + size).  If addr is NULL, allocate the virtual address
1132  * space as well.
1133  */
1134 static void *
1135 segkmem_xalloc_lp(vmem_t *vmp, void *inaddr, size_t size, int vmflag,
1136     uint_t attr, page_t *(*page_create_func)(void *, size_t, int, void *),
1137     void *pcarg)
1138 {
1139 	caddr_t addr = inaddr, pa;
1140 	size_t  lpsize = segkmem_lpsize;
1141 	pgcnt_t npages = btopr(size);
1142 	pgcnt_t nbpages = btop(lpsize);
1143 	pgcnt_t nlpages = size >> segkmem_lpshift;
1144 	size_t  ppasize = nbpages * sizeof (page_t *);
1145 	page_t *pp, *rootpp, **ppa, *pplist = NULL;
1146 	int i;
1147 
1148 	vmflag |= VM_NOSLEEP;
1149 
1150 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
1151 		return (NULL);
1152 	}
1153 
1154 	/*
1155 	 * allocate an array we need for hat_memload_array.
1156 	 * we use a separate arena to avoid recursion.
1157 	 * we will not need this array when hat_memload_array learns pp++
1158 	 */
1159 	if ((ppa = vmem_alloc(segkmem_ppa_arena, ppasize, vmflag)) == NULL) {
1160 		goto fail_array_alloc;
1161 	}
1162 
1163 	if (inaddr == NULL && (addr = vmem_alloc(vmp, size, vmflag)) == NULL)
1164 		goto fail_vmem_alloc;
1165 
1166 	ASSERT(((uintptr_t)addr & (lpsize - 1)) == 0);
1167 
1168 	/* create all the pages */
1169 	for (pa = addr, i = 0; i < nlpages; i++, pa += lpsize) {
1170 		if ((pp = page_create_func(pa, lpsize, vmflag, pcarg)) == NULL)
1171 			goto fail_page_create;
1172 		page_list_concat(&pplist, &pp);
1173 	}
1174 
1175 	/* at this point we have all the resource to complete the request */
1176 	while ((rootpp = pplist) != NULL) {
1177 		for (i = 0; i < nbpages; i++) {
1178 			ASSERT(pplist != NULL);
1179 			pp = pplist;
1180 			page_sub(&pplist, pp);
1181 			ASSERT(page_iolock_assert(pp));
1182 			page_io_unlock(pp);
1183 			ppa[i] = pp;
1184 		}
1185 		/*
1186 		 * Load the locked entry. It's OK to preload the entry into the
1187 		 * TSB since we now support large mappings in the kernel TSB.
1188 		 */
1189 		hat_memload_array(kas.a_hat,
1190 		    (caddr_t)(uintptr_t)rootpp->p_offset, lpsize,
1191 		    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC | attr,
1192 		    HAT_LOAD_LOCK);
1193 
1194 		for (--i; i >= 0; --i) {
1195 			ppa[i]->p_lckcnt = 1;
1196 			page_unlock(ppa[i]);
1197 		}
1198 	}
1199 
1200 	vmem_free(segkmem_ppa_arena, ppa, ppasize);
1201 	return (addr);
1202 
1203 fail_page_create:
1204 	while ((rootpp = pplist) != NULL) {
1205 		for (i = 0, pp = pplist; i < nbpages; i++, pp = pplist) {
1206 			ASSERT(pp != NULL);
1207 			page_sub(&pplist, pp);
1208 			ASSERT(page_iolock_assert(pp));
1209 			page_io_unlock(pp);
1210 		}
1211 		page_destroy_pages(rootpp);
1212 	}
1213 
1214 	if (inaddr == NULL)
1215 		vmem_free(vmp, addr, size);
1216 
1217 fail_vmem_alloc:
1218 	vmem_free(segkmem_ppa_arena, ppa, ppasize);
1219 
1220 fail_array_alloc:
1221 	page_unresv(npages);
1222 
1223 	return (NULL);
1224 }
1225 
1226 static void
1227 segkmem_free_one_lp(caddr_t addr, size_t size)
1228 {
1229 	page_t		*pp, *rootpp = NULL;
1230 	pgcnt_t		pgs_left = btopr(size);
1231 
1232 	ASSERT(size == segkmem_lpsize);
1233 
1234 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1235 
1236 	for (; pgs_left > 0; addr += PAGESIZE, pgs_left--) {
1237 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
1238 		if (pp == NULL)
1239 			panic("segkmem_free_one_lp: page not found");
1240 		ASSERT(PAGE_EXCL(pp));
1241 		pp->p_lckcnt = 0;
1242 		if (rootpp == NULL)
1243 			rootpp = pp;
1244 	}
1245 	ASSERT(rootpp != NULL);
1246 	page_destroy_pages(rootpp);
1247 
1248 	/* page_unresv() is done by the caller */
1249 }
1250 #endif /* __sparc */
1251 
1252 /*
1253  * This function is called to import new spans into the vmem arenas like
1254  * kmem_default_arena and kmem_oversize_arena. It first tries to import
1255  * spans from large page arena - kmem_lp_arena. In order to do this it might
1256  * have to "upgrade the requested size" to kmem_lp_arena quantum. If
1257  * it was not able to satisfy the upgraded request it then calls regular
1258  * segkmem_alloc() that satisfies the request by importing from "*vmp" arena
1259  */
1260 /*ARGSUSED*/
1261 void *
1262 segkmem_alloc_lp(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
1263 {
1264 	size_t size;
1265 	kthread_t *t = curthread;
1266 	segkmem_lpcb_t *lpcb = &segkmem_lpcb;
1267 
1268 	ASSERT(sizep != NULL);
1269 
1270 	size = *sizep;
1271 
1272 	if (lpcb->lp_uselp && !(t->t_flag & T_PANIC) &&
1273 	    !(vmflag & SEGKMEM_SHARELOCKED)) {
1274 
1275 		size_t kmemlp_qnt = segkmem_kmemlp_quantum;
1276 		size_t asize = P2ROUNDUP(size, kmemlp_qnt);
1277 		void  *addr = NULL;
1278 		ulong_t *lpthrtp = &lpcb->lp_throttle;
1279 		ulong_t lpthrt = *lpthrtp;
1280 		int	dowakeup = 0;
1281 		int	doalloc = 1;
1282 
1283 		ASSERT(kmem_lp_arena != NULL);
1284 		ASSERT(asize >= size);
1285 
1286 		if (lpthrt != 0) {
1287 			/* try to update the throttle value */
1288 			lpthrt = atomic_inc_ulong_nv(lpthrtp);
1289 			if (lpthrt >= segkmem_lpthrottle_max) {
1290 				lpthrt = atomic_cas_ulong(lpthrtp, lpthrt,
1291 				    segkmem_lpthrottle_max / 4);
1292 			}
1293 
1294 			/*
1295 			 * when we get above throttle start do an exponential
1296 			 * backoff at trying large pages and reaping
1297 			 */
1298 			if (lpthrt > segkmem_lpthrottle_start &&
1299 			    !ISP2(lpthrt)) {
1300 				lpcb->allocs_throttled++;
1301 				lpthrt--;
1302 				if (ISP2(lpthrt))
1303 					kmem_reap();
1304 				return (segkmem_alloc(vmp, size, vmflag));
1305 			}
1306 		}
1307 
1308 		if (!(vmflag & VM_NOSLEEP) &&
1309 		    segkmem_heaplp_quantum >= (8 * kmemlp_qnt) &&
1310 		    vmem_size(kmem_lp_arena, VMEM_FREE) <= kmemlp_qnt &&
1311 		    asize < (segkmem_heaplp_quantum - kmemlp_qnt)) {
1312 
1313 			/*
1314 			 * we are low on free memory in kmem_lp_arena
1315 			 * we let only one guy to allocate heap_lp
1316 			 * quantum size chunk that everybody is going to
1317 			 * share
1318 			 */
1319 			mutex_enter(&lpcb->lp_lock);
1320 
1321 			if (lpcb->lp_wait) {
1322 
1323 				/* we are not the first one - wait */
1324 				cv_wait(&lpcb->lp_cv, &lpcb->lp_lock);
1325 				if (vmem_size(kmem_lp_arena, VMEM_FREE) <
1326 				    kmemlp_qnt)  {
1327 					doalloc = 0;
1328 				}
1329 			} else if (vmem_size(kmem_lp_arena, VMEM_FREE) <=
1330 			    kmemlp_qnt) {
1331 
1332 				/*
1333 				 * we are the first one, make sure we import
1334 				 * a large page
1335 				 */
1336 				if (asize == kmemlp_qnt)
1337 					asize += kmemlp_qnt;
1338 				dowakeup = 1;
1339 				lpcb->lp_wait = 1;
1340 			}
1341 
1342 			mutex_exit(&lpcb->lp_lock);
1343 		}
1344 
1345 		/*
1346 		 * VM_ABORT flag prevents sleeps in vmem_xalloc when
1347 		 * large pages are not available. In that case this allocation
1348 		 * attempt will fail and we will retry allocation with small
1349 		 * pages. We also do not want to panic if this allocation fails
1350 		 * because we are going to retry.
1351 		 */
1352 		if (doalloc) {
1353 			addr = vmem_alloc(kmem_lp_arena, asize,
1354 			    (vmflag | VM_ABORT) & ~VM_PANIC);
1355 
1356 			if (dowakeup) {
1357 				mutex_enter(&lpcb->lp_lock);
1358 				ASSERT(lpcb->lp_wait != 0);
1359 				lpcb->lp_wait = 0;
1360 				cv_broadcast(&lpcb->lp_cv);
1361 				mutex_exit(&lpcb->lp_lock);
1362 			}
1363 		}
1364 
1365 		if (addr != NULL) {
1366 			*sizep = asize;
1367 			*lpthrtp = 0;
1368 			return (addr);
1369 		}
1370 
1371 		if (vmflag & VM_NOSLEEP)
1372 			lpcb->nosleep_allocs_failed++;
1373 		else
1374 			lpcb->sleep_allocs_failed++;
1375 		lpcb->alloc_bytes_failed += size;
1376 
1377 		/* if large page throttling is not started yet do it */
1378 		if (segkmem_use_lpthrottle && lpthrt == 0) {
1379 			lpthrt = atomic_cas_ulong(lpthrtp, lpthrt, 1);
1380 		}
1381 	}
1382 	return (segkmem_alloc(vmp, size, vmflag));
1383 }
1384 
1385 void
1386 segkmem_free_lp(vmem_t *vmp, void *inaddr, size_t size)
1387 {
1388 	if (kmem_lp_arena == NULL || !IS_KMEM_VA_LARGEPAGE((caddr_t)inaddr)) {
1389 		segkmem_free(vmp, inaddr, size);
1390 	} else {
1391 		vmem_free(kmem_lp_arena, inaddr, size);
1392 	}
1393 }
1394 
1395 #ifdef __sparc
1396 /*
1397  * segkmem_alloc_lpi() imports virtual memory from large page heap arena
1398  * into kmem_lp arena. In the process it maps the imported segment with
1399  * large pages
1400  */
1401 static void *
1402 segkmem_alloc_lpi(vmem_t *vmp, size_t size, int vmflag)
1403 {
1404 	segkmem_lpcb_t *lpcb = &segkmem_lpcb;
1405 	void  *addr;
1406 
1407 	ASSERT(size != 0);
1408 	ASSERT(vmp == heap_lp_arena);
1409 
1410 	/* do not allow large page heap grow beyound limits */
1411 	if (vmem_size(vmp, VMEM_ALLOC) >= segkmem_kmemlp_max) {
1412 		lpcb->allocs_limited++;
1413 		return (NULL);
1414 	}
1415 
1416 	addr = segkmem_xalloc_lp(vmp, NULL, size, vmflag, 0,
1417 	    segkmem_page_create_large, NULL);
1418 	return (addr);
1419 }
1420 
1421 /*
1422  * segkmem_free_lpi() returns virtual memory back into large page heap arena
1423  * from kmem_lp arena. Beore doing this it unmaps the segment and frees
1424  * large pages used to map it.
1425  */
1426 static void
1427 segkmem_free_lpi(vmem_t *vmp, void *inaddr, size_t size)
1428 {
1429 	pgcnt_t		nlpages = size >> segkmem_lpshift;
1430 	size_t		lpsize = segkmem_lpsize;
1431 	caddr_t		addr = inaddr;
1432 	pgcnt_t		npages = btopr(size);
1433 	int		i;
1434 
1435 	ASSERT(vmp == heap_lp_arena);
1436 	ASSERT(IS_KMEM_VA_LARGEPAGE(addr));
1437 	ASSERT(((uintptr_t)inaddr & (lpsize - 1)) == 0);
1438 
1439 	for (i = 0; i < nlpages; i++) {
1440 		segkmem_free_one_lp(addr, lpsize);
1441 		addr += lpsize;
1442 	}
1443 
1444 	page_unresv(npages);
1445 
1446 	vmem_free(vmp, inaddr, size);
1447 }
1448 #endif /* __sparc */
1449 
1450 /*
1451  * This function is called at system boot time by kmem_init right after
1452  * /etc/system file has been read. It checks based on hardware configuration
1453  * and /etc/system settings if system is going to use large pages. The
1454  * initialiazation necessary to actually start using large pages
1455  * happens later in the process after segkmem_heap_lp_init() is called.
1456  */
1457 int
1458 segkmem_lpsetup()
1459 {
1460 	int use_large_pages = 0;
1461 
1462 #ifdef __sparc
1463 
1464 	size_t memtotal = physmem * PAGESIZE;
1465 
1466 	if (heap_lp_base == NULL) {
1467 		segkmem_lpsize = PAGESIZE;
1468 		return (0);
1469 	}
1470 
1471 	/* get a platform dependent value of large page size for kernel heap */
1472 	segkmem_lpsize = get_segkmem_lpsize(segkmem_lpsize);
1473 
1474 	if (segkmem_lpsize <= PAGESIZE) {
1475 		/*
1476 		 * put virtual space reserved for the large page kernel
1477 		 * back to the regular heap
1478 		 */
1479 		vmem_xfree(heap_arena, heap_lp_base,
1480 		    heap_lp_end - heap_lp_base);
1481 		heap_lp_base = NULL;
1482 		heap_lp_end = NULL;
1483 		segkmem_lpsize = PAGESIZE;
1484 		return (0);
1485 	}
1486 
1487 	/* set heap_lp quantum if necessary */
1488 	if (segkmem_heaplp_quantum == 0 || !ISP2(segkmem_heaplp_quantum) ||
1489 	    P2PHASE(segkmem_heaplp_quantum, segkmem_lpsize)) {
1490 		segkmem_heaplp_quantum = segkmem_lpsize;
1491 	}
1492 
1493 	/* set kmem_lp quantum if necessary */
1494 	if (segkmem_kmemlp_quantum == 0 || !ISP2(segkmem_kmemlp_quantum) ||
1495 	    segkmem_kmemlp_quantum > segkmem_heaplp_quantum) {
1496 		segkmem_kmemlp_quantum = segkmem_heaplp_quantum;
1497 	}
1498 
1499 	/* set total amount of memory allowed for large page kernel heap */
1500 	if (segkmem_kmemlp_max == 0) {
1501 		if (segkmem_kmemlp_pcnt == 0 || segkmem_kmemlp_pcnt > 100)
1502 			segkmem_kmemlp_pcnt = 12;
1503 		segkmem_kmemlp_max = (memtotal * segkmem_kmemlp_pcnt) / 100;
1504 	}
1505 	segkmem_kmemlp_max = P2ROUNDUP(segkmem_kmemlp_max,
1506 	    segkmem_heaplp_quantum);
1507 
1508 	/* fix lp kmem preallocation request if necesssary */
1509 	if (segkmem_kmemlp_min) {
1510 		segkmem_kmemlp_min = P2ROUNDUP(segkmem_kmemlp_min,
1511 		    segkmem_heaplp_quantum);
1512 		if (segkmem_kmemlp_min > segkmem_kmemlp_max)
1513 			segkmem_kmemlp_min = segkmem_kmemlp_max;
1514 	}
1515 
1516 	use_large_pages = 1;
1517 	segkmem_lpszc = page_szc(segkmem_lpsize);
1518 	segkmem_lpshift = page_get_shift(segkmem_lpszc);
1519 
1520 #endif
1521 	return (use_large_pages);
1522 }
1523 
1524 void
1525 segkmem_zio_init(void *zio_mem_base, size_t zio_mem_size)
1526 {
1527 	ASSERT(zio_mem_base != NULL);
1528 	ASSERT(zio_mem_size != 0);
1529 
1530 	/*
1531 	 * To reduce VA space fragmentation, we set up quantum caches for the
1532 	 * smaller sizes;  we chose 32k because that translates to 128k VA
1533 	 * slabs, which matches nicely with the common 128k zio_data bufs.
1534 	 */
1535 	zio_arena = vmem_create("zfs_file_data", zio_mem_base, zio_mem_size,
1536 	    PAGESIZE, NULL, NULL, NULL, 32 * 1024, VM_SLEEP);
1537 
1538 	zio_alloc_arena = vmem_create("zfs_file_data_buf", NULL, 0, PAGESIZE,
1539 	    segkmem_zio_alloc, segkmem_zio_free, zio_arena, 0, VM_SLEEP);
1540 
1541 	ASSERT(zio_arena != NULL);
1542 	ASSERT(zio_alloc_arena != NULL);
1543 }
1544 
1545 #if defined(__amd64)
1546 
1547 void
1548 segkmem_kvmm_init(void *base, size_t size)
1549 {
1550 	ASSERT(base != NULL);
1551 	ASSERT(size != 0);
1552 
1553 	kvmm_arena = vmem_create("kvmm_arena", base, size, 1024 * 1024,
1554 	    NULL, NULL, NULL, 0, VM_SLEEP);
1555 
1556 	ASSERT(kvmm_arena != NULL);
1557 }
1558 
1559 #elif defined(__sparc)
1560 
1561 static void *
1562 segkmem_alloc_ppa(vmem_t *vmp, size_t size, int vmflag)
1563 {
1564 	size_t ppaquantum = btopr(segkmem_lpsize) * sizeof (page_t *);
1565 	void   *addr;
1566 
1567 	if (ppaquantum <= PAGESIZE)
1568 		return (segkmem_alloc(vmp, size, vmflag));
1569 
1570 	ASSERT((size & (ppaquantum - 1)) == 0);
1571 
1572 	addr = vmem_xalloc(vmp, size, ppaquantum, 0, 0, NULL, NULL, vmflag);
1573 	if (addr != NULL && segkmem_xalloc(vmp, addr, size, vmflag, 0,
1574 	    segkmem_page_create, NULL) == NULL) {
1575 		vmem_xfree(vmp, addr, size);
1576 		addr = NULL;
1577 	}
1578 
1579 	return (addr);
1580 }
1581 
1582 static void
1583 segkmem_free_ppa(vmem_t *vmp, void *addr, size_t size)
1584 {
1585 	size_t ppaquantum = btopr(segkmem_lpsize) * sizeof (page_t *);
1586 
1587 	ASSERT(addr != NULL);
1588 
1589 	if (ppaquantum <= PAGESIZE) {
1590 		segkmem_free(vmp, addr, size);
1591 	} else {
1592 		segkmem_free(NULL, addr, size);
1593 		vmem_xfree(vmp, addr, size);
1594 	}
1595 }
1596 
1597 void
1598 segkmem_heap_lp_init()
1599 {
1600 	segkmem_lpcb_t *lpcb = &segkmem_lpcb;
1601 	size_t heap_lp_size = heap_lp_end - heap_lp_base;
1602 	size_t lpsize = segkmem_lpsize;
1603 	size_t ppaquantum;
1604 	void   *addr;
1605 
1606 	if (segkmem_lpsize <= PAGESIZE) {
1607 		ASSERT(heap_lp_base == NULL);
1608 		ASSERT(heap_lp_end == NULL);
1609 		return;
1610 	}
1611 
1612 	ASSERT(segkmem_heaplp_quantum >= lpsize);
1613 	ASSERT((segkmem_heaplp_quantum & (lpsize - 1)) == 0);
1614 	ASSERT(lpcb->lp_uselp == 0);
1615 	ASSERT(heap_lp_base != NULL);
1616 	ASSERT(heap_lp_end != NULL);
1617 	ASSERT(heap_lp_base < heap_lp_end);
1618 	ASSERT(heap_lp_arena == NULL);
1619 	ASSERT(((uintptr_t)heap_lp_base & (lpsize - 1)) == 0);
1620 	ASSERT(((uintptr_t)heap_lp_end & (lpsize - 1)) == 0);
1621 
1622 	/* create large page heap arena */
1623 	heap_lp_arena = vmem_create("heap_lp", heap_lp_base, heap_lp_size,
1624 	    segkmem_heaplp_quantum, NULL, NULL, NULL, 0, VM_SLEEP);
1625 
1626 	ASSERT(heap_lp_arena != NULL);
1627 
1628 	/* This arena caches memory already mapped by large pages */
1629 	kmem_lp_arena = vmem_create("kmem_lp", NULL, 0, segkmem_kmemlp_quantum,
1630 	    segkmem_alloc_lpi, segkmem_free_lpi, heap_lp_arena, 0, VM_SLEEP);
1631 
1632 	ASSERT(kmem_lp_arena != NULL);
1633 
1634 	mutex_init(&lpcb->lp_lock, NULL, MUTEX_DEFAULT, NULL);
1635 	cv_init(&lpcb->lp_cv, NULL, CV_DEFAULT, NULL);
1636 
1637 	/*
1638 	 * this arena is used for the array of page_t pointers necessary
1639 	 * to call hat_mem_load_array
1640 	 */
1641 	ppaquantum = btopr(lpsize) * sizeof (page_t *);
1642 	segkmem_ppa_arena = vmem_create("segkmem_ppa", NULL, 0, ppaquantum,
1643 	    segkmem_alloc_ppa, segkmem_free_ppa, heap_arena, ppaquantum,
1644 	    VM_SLEEP);
1645 
1646 	ASSERT(segkmem_ppa_arena != NULL);
1647 
1648 	/* prealloacate some memory for the lp kernel heap */
1649 	if (segkmem_kmemlp_min) {
1650 
1651 		ASSERT(P2PHASE(segkmem_kmemlp_min,
1652 		    segkmem_heaplp_quantum) == 0);
1653 
1654 		if ((addr = segkmem_alloc_lpi(heap_lp_arena,
1655 		    segkmem_kmemlp_min, VM_SLEEP)) != NULL) {
1656 
1657 			addr = vmem_add(kmem_lp_arena, addr,
1658 			    segkmem_kmemlp_min, VM_SLEEP);
1659 			ASSERT(addr != NULL);
1660 		}
1661 	}
1662 
1663 	lpcb->lp_uselp = 1;
1664 }
1665 
1666 #endif
1667