xref: /illumos-gate/usr/src/uts/common/vm/seg_kmem.c (revision 902408c7)
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 <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 /*
52  * seg_kmem is the primary kernel memory segment driver.  It
53  * maps the kernel heap [kernelheap, ekernelheap), module text,
54  * and all memory which was allocated before the VM was initialized
55  * into kas.
56  *
57  * Pages which belong to seg_kmem are hashed into &kvp vnode at
58  * an offset equal to (u_offset_t)virt_addr, and have p_lckcnt >= 1.
59  * They must never be paged out since segkmem_fault() is a no-op to
60  * prevent recursive faults.
61  *
62  * Currently, seg_kmem pages are sharelocked (p_sharelock == 1) on
63  * __x86 and are unlocked (p_sharelock == 0) on __sparc.  Once __x86
64  * supports relocation the #ifdef kludges can be removed.
65  *
66  * seg_kmem pages may be subject to relocation by page_relocate(),
67  * provided that the HAT supports it; if this is so, segkmem_reloc
68  * will be set to a nonzero value. All boot time allocated memory as
69  * well as static memory is considered off limits to relocation.
70  * Pages are "relocatable" if p_state does not have P_NORELOC set, so
71  * we request P_NORELOC pages for memory that isn't safe to relocate.
72  *
73  * The kernel heap is logically divided up into four pieces:
74  *
75  *   heap32_arena is for allocations that require 32-bit absolute
76  *   virtual addresses (e.g. code that uses 32-bit pointers/offsets).
77  *
78  *   heap_core is for allocations that require 2GB *relative*
79  *   offsets; in other words all memory from heap_core is within
80  *   2GB of all other memory from the same arena. This is a requirement
81  *   of the addressing modes of some processors in supervisor code.
82  *
83  *   heap_arena is the general heap arena.
84  *
85  *   static_arena is the static memory arena.  Allocations from it
86  *   are not subject to relocation so it is safe to use the memory
87  *   physical address as well as the virtual address (e.g. the VA to
88  *   PA translations are static).  Caches may import from static_arena;
89  *   all other static memory allocations should use static_alloc_arena.
90  *
91  * On some platforms which have limited virtual address space, seg_kmem
92  * may share [kernelheap, ekernelheap) with seg_kp; if this is so,
93  * segkp_bitmap is non-NULL, and each bit represents a page of virtual
94  * address space which is actually seg_kp mapped.
95  */
96 
97 extern ulong_t *segkp_bitmap;   /* Is set if segkp is from the kernel heap */
98 
99 char *kernelheap;		/* start of primary kernel heap */
100 char *ekernelheap;		/* end of primary kernel heap */
101 struct seg kvseg;		/* primary kernel heap segment */
102 struct seg kvseg_core;		/* "core" kernel heap segment */
103 struct seg kzioseg;		/* Segment for zio mappings */
104 vmem_t *heap_arena;		/* primary kernel heap arena */
105 vmem_t *heap_core_arena;	/* core kernel heap arena */
106 char *heap_core_base;		/* start of core kernel heap arena */
107 char *heap_lp_base;		/* start of kernel large page heap arena */
108 char *heap_lp_end;		/* end of kernel large page heap arena */
109 vmem_t *hat_memload_arena;	/* HAT translation data */
110 struct seg kvseg32;		/* 32-bit kernel heap segment */
111 vmem_t *heap32_arena;		/* 32-bit kernel heap arena */
112 vmem_t *heaptext_arena;		/* heaptext arena */
113 struct as kas;			/* kernel address space */
114 int segkmem_reloc;		/* enable/disable relocatable segkmem pages */
115 vmem_t *static_arena;		/* arena for caches to import static memory */
116 vmem_t *static_alloc_arena;	/* arena for allocating static memory */
117 vmem_t *zio_arena = NULL;	/* arena for allocating zio memory */
118 vmem_t *zio_alloc_arena = NULL;	/* arena for allocating zio memory */
119 
120 /*
121  * seg_kmem driver can map part of the kernel heap with large pages.
122  * Currently this functionality is implemented for sparc platforms only.
123  *
124  * The large page size "segkmem_lpsize" for kernel heap is selected in the
125  * platform specific code. It can also be modified via /etc/system file.
126  * Setting segkmem_lpsize to PAGESIZE in /etc/system disables usage of large
127  * pages for kernel heap. "segkmem_lpshift" is adjusted appropriately to
128  * match segkmem_lpsize.
129  *
130  * At boot time we carve from kernel heap arena a range of virtual addresses
131  * that will be used for large page mappings. This range [heap_lp_base,
132  * heap_lp_end) is set up as a separate vmem arena - "heap_lp_arena". We also
133  * create "kmem_lp_arena" that caches memory already backed up by large
134  * pages. kmem_lp_arena imports virtual segments from heap_lp_arena.
135  */
136 
137 size_t	segkmem_lpsize;
138 static  uint_t	segkmem_lpshift = PAGESHIFT;
139 int	segkmem_lpszc = 0;
140 
141 size_t  segkmem_kmemlp_quantum = 0x400000;	/* 4MB */
142 size_t  segkmem_heaplp_quantum;
143 vmem_t *heap_lp_arena;
144 static  vmem_t *kmem_lp_arena;
145 static  vmem_t *segkmem_ppa_arena;
146 static	segkmem_lpcb_t segkmem_lpcb;
147 
148 /*
149  * We use "segkmem_kmemlp_max" to limit the total amount of physical memory
150  * consumed by the large page heap. By default this parameter is set to 1/8 of
151  * physmem but can be adjusted through /etc/system either directly or
152  * indirectly by setting "segkmem_kmemlp_pcnt" to the percent of physmem
153  * we allow for large page heap.
154  */
155 size_t  segkmem_kmemlp_max;
156 static  uint_t  segkmem_kmemlp_pcnt;
157 
158 /*
159  * Getting large pages for kernel heap could be problematic due to
160  * physical memory fragmentation. That's why we allow to preallocate
161  * "segkmem_kmemlp_min" bytes at boot time.
162  */
163 static  size_t	segkmem_kmemlp_min;
164 
165 /*
166  * Throttling is used to avoid expensive tries to allocate large pages
167  * for kernel heap when a lot of succesive attempts to do so fail.
168  */
169 static  ulong_t segkmem_lpthrottle_max = 0x400000;
170 static  ulong_t segkmem_lpthrottle_start = 0x40;
171 static  ulong_t segkmem_use_lpthrottle = 1;
172 
173 /*
174  * Freed pages accumulate on a garbage list until segkmem is ready,
175  * at which point we call segkmem_gc() to free it all.
176  */
177 typedef struct segkmem_gc_list {
178 	struct segkmem_gc_list	*gc_next;
179 	vmem_t			*gc_arena;
180 	size_t			gc_size;
181 } segkmem_gc_list_t;
182 
183 static segkmem_gc_list_t *segkmem_gc_list;
184 
185 /*
186  * Allocations from the hat_memload arena add VM_MEMLOAD to their
187  * vmflags so that segkmem_xalloc() can inform the hat layer that it needs
188  * to take steps to prevent infinite recursion.  HAT allocations also
189  * must be non-relocatable to prevent recursive page faults.
190  */
191 static void *
192 hat_memload_alloc(vmem_t *vmp, size_t size, int flags)
193 {
194 	flags |= (VM_MEMLOAD | VM_NORELOC);
195 	return (segkmem_alloc(vmp, size, flags));
196 }
197 
198 /*
199  * Allocations from static_arena arena (or any other arena that uses
200  * segkmem_alloc_permanent()) require non-relocatable (permanently
201  * wired) memory pages, since these pages are referenced by physical
202  * as well as virtual address.
203  */
204 void *
205 segkmem_alloc_permanent(vmem_t *vmp, size_t size, int flags)
206 {
207 	return (segkmem_alloc(vmp, size, flags | VM_NORELOC));
208 }
209 
210 /*
211  * Initialize kernel heap boundaries.
212  */
213 void
214 kernelheap_init(
215 	void *heap_start,
216 	void *heap_end,
217 	char *first_avail,
218 	void *core_start,
219 	void *core_end)
220 {
221 	uintptr_t textbase;
222 	size_t core_size;
223 	size_t heap_size;
224 	vmem_t *heaptext_parent;
225 	size_t	heap_lp_size = 0;
226 #ifdef __sparc
227 	size_t kmem64_sz = kmem64_aligned_end - kmem64_base;
228 #endif	/* __sparc */
229 
230 	kernelheap = heap_start;
231 	ekernelheap = heap_end;
232 
233 #ifdef __sparc
234 	heap_lp_size = (((uintptr_t)heap_end - (uintptr_t)heap_start) / 4);
235 	/*
236 	 * Bias heap_lp start address by kmem64_sz to reduce collisions
237 	 * in 4M kernel TSB between kmem64 area and heap_lp
238 	 */
239 	kmem64_sz = P2ROUNDUP(kmem64_sz, MMU_PAGESIZE256M);
240 	if (kmem64_sz <= heap_lp_size / 2)
241 		heap_lp_size -= kmem64_sz;
242 	heap_lp_base = ekernelheap - heap_lp_size;
243 	heap_lp_end = heap_lp_base + heap_lp_size;
244 #endif	/* __sparc */
245 
246 	/*
247 	 * If this platform has a 'core' heap area, then the space for
248 	 * overflow module text should be carved out of the end of that
249 	 * heap.  Otherwise, it gets carved out of the general purpose
250 	 * heap.
251 	 */
252 	core_size = (uintptr_t)core_end - (uintptr_t)core_start;
253 	if (core_size > 0) {
254 		ASSERT(core_size >= HEAPTEXT_SIZE);
255 		textbase = (uintptr_t)core_end - HEAPTEXT_SIZE;
256 		core_size -= HEAPTEXT_SIZE;
257 	}
258 #ifndef __sparc
259 	else {
260 		ekernelheap -= HEAPTEXT_SIZE;
261 		textbase = (uintptr_t)ekernelheap;
262 	}
263 #endif
264 
265 	heap_size = (uintptr_t)ekernelheap - (uintptr_t)kernelheap;
266 	heap_arena = vmem_init("heap", kernelheap, heap_size, PAGESIZE,
267 	    segkmem_alloc, segkmem_free);
268 
269 	if (core_size > 0) {
270 		heap_core_arena = vmem_create("heap_core", core_start,
271 		    core_size, PAGESIZE, NULL, NULL, NULL, 0, VM_SLEEP);
272 		heap_core_base = core_start;
273 	} else {
274 		heap_core_arena = heap_arena;
275 		heap_core_base = kernelheap;
276 	}
277 
278 	/*
279 	 * reserve space for the large page heap. If large pages for kernel
280 	 * heap is enabled large page heap arean will be created later in the
281 	 * boot sequence in segkmem_heap_lp_init(). Otherwise the allocated
282 	 * range will be returned back to the heap_arena.
283 	 */
284 	if (heap_lp_size) {
285 		(void) vmem_xalloc(heap_arena, heap_lp_size, PAGESIZE, 0, 0,
286 		    heap_lp_base, heap_lp_end,
287 		    VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
288 	}
289 
290 	/*
291 	 * Remove the already-spoken-for memory range [kernelheap, first_avail).
292 	 */
293 	(void) vmem_xalloc(heap_arena, first_avail - kernelheap, PAGESIZE,
294 	    0, 0, kernelheap, first_avail, VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
295 
296 #ifdef __sparc
297 	heap32_arena = vmem_create("heap32", (void *)SYSBASE32,
298 	    SYSLIMIT32 - SYSBASE32 - HEAPTEXT_SIZE, PAGESIZE, NULL,
299 	    NULL, NULL, 0, VM_SLEEP);
300 
301 	textbase = SYSLIMIT32 - HEAPTEXT_SIZE;
302 	heaptext_parent = NULL;
303 #else	/* __sparc */
304 	heap32_arena = heap_core_arena;
305 	heaptext_parent = heap_core_arena;
306 #endif	/* __sparc */
307 
308 	heaptext_arena = vmem_create("heaptext", (void *)textbase,
309 	    HEAPTEXT_SIZE, PAGESIZE, NULL, NULL, heaptext_parent, 0, VM_SLEEP);
310 
311 	/*
312 	 * Create a set of arenas for memory with static translations
313 	 * (e.g. VA -> PA translations cannot change).  Since using
314 	 * kernel pages by physical address implies it isn't safe to
315 	 * walk across page boundaries, the static_arena quantum must
316 	 * be PAGESIZE.  Any kmem caches that require static memory
317 	 * should source from static_arena, while direct allocations
318 	 * should only use static_alloc_arena.
319 	 */
320 	static_arena = vmem_create("static", NULL, 0, PAGESIZE,
321 	    segkmem_alloc_permanent, segkmem_free, heap_arena, 0, VM_SLEEP);
322 	static_alloc_arena = vmem_create("static_alloc", NULL, 0,
323 	    sizeof (uint64_t), vmem_alloc, vmem_free, static_arena,
324 	    0, VM_SLEEP);
325 
326 	/*
327 	 * Create an arena for translation data (ptes, hmes, or hblks).
328 	 * We need an arena for this because hat_memload() is essential
329 	 * to vmem_populate() (see comments in common/os/vmem.c).
330 	 *
331 	 * Note: any kmem cache that allocates from hat_memload_arena
332 	 * must be created as a KMC_NOHASH cache (i.e. no external slab
333 	 * and bufctl structures to allocate) so that slab creation doesn't
334 	 * require anything more than a single vmem_alloc().
335 	 */
336 	hat_memload_arena = vmem_create("hat_memload", NULL, 0, PAGESIZE,
337 	    hat_memload_alloc, segkmem_free, heap_arena, 0,
338 	    VM_SLEEP | VMC_POPULATOR | VMC_DUMPSAFE);
339 }
340 
341 void
342 boot_mapin(caddr_t addr, size_t size)
343 {
344 	caddr_t	 eaddr;
345 	page_t	*pp;
346 	pfn_t	 pfnum;
347 
348 	if (page_resv(btop(size), KM_NOSLEEP) == 0)
349 		panic("boot_mapin: page_resv failed");
350 
351 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
352 		pfnum = va_to_pfn(addr);
353 		if (pfnum == PFN_INVALID)
354 			continue;
355 		if ((pp = page_numtopp_nolock(pfnum)) == NULL)
356 			panic("boot_mapin(): No pp for pfnum = %lx", pfnum);
357 
358 		/*
359 		 * must break up any large pages that may have constituent
360 		 * pages being utilized for BOP_ALLOC()'s before calling
361 		 * page_numtopp().The locking code (ie. page_reclaim())
362 		 * can't handle them
363 		 */
364 		if (pp->p_szc != 0)
365 			page_boot_demote(pp);
366 
367 		pp = page_numtopp(pfnum, SE_EXCL);
368 		if (pp == NULL || PP_ISFREE(pp))
369 			panic("boot_alloc: pp is NULL or free");
370 
371 		/*
372 		 * If the cage is on but doesn't yet contain this page,
373 		 * mark it as non-relocatable.
374 		 */
375 		if (kcage_on && !PP_ISNORELOC(pp)) {
376 			PP_SETNORELOC(pp);
377 			PLCNT_XFER_NORELOC(pp);
378 		}
379 
380 		(void) page_hashin(pp, &kvp, (u_offset_t)(uintptr_t)addr, NULL);
381 		pp->p_lckcnt = 1;
382 #if defined(__x86)
383 		page_downgrade(pp);
384 #else
385 		page_unlock(pp);
386 #endif
387 	}
388 }
389 
390 /*
391  * Get pages from boot and hash them into the kernel's vp.
392  * Used after page structs have been allocated, but before segkmem is ready.
393  */
394 void *
395 boot_alloc(void *inaddr, size_t size, uint_t align)
396 {
397 	caddr_t addr = inaddr;
398 
399 	if (bootops == NULL)
400 		prom_panic("boot_alloc: attempt to allocate memory after "
401 		    "BOP_GONE");
402 
403 	size = ptob(btopr(size));
404 #ifdef __sparc
405 	if (bop_alloc_chunk(addr, size, align) != (caddr_t)addr)
406 		panic("boot_alloc: bop_alloc_chunk failed");
407 #else
408 	if (BOP_ALLOC(bootops, addr, size, align) != addr)
409 		panic("boot_alloc: BOP_ALLOC failed");
410 #endif
411 	boot_mapin((caddr_t)addr, size);
412 	return (addr);
413 }
414 
415 static void
416 segkmem_badop()
417 {
418 	panic("segkmem_badop");
419 }
420 
421 #define	SEGKMEM_BADOP(t)	(t(*)())segkmem_badop
422 
423 /*ARGSUSED*/
424 static faultcode_t
425 segkmem_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t size,
426 	enum fault_type type, enum seg_rw rw)
427 {
428 	pgcnt_t npages;
429 	spgcnt_t pg;
430 	page_t *pp;
431 	struct vnode *vp = seg->s_data;
432 
433 	ASSERT(RW_READ_HELD(&seg->s_as->a_lock));
434 
435 	if (seg->s_as != &kas || size > seg->s_size ||
436 	    addr < seg->s_base || addr + size > seg->s_base + seg->s_size)
437 		panic("segkmem_fault: bad args");
438 
439 	/*
440 	 * If it is one of segkp pages, call segkp_fault.
441 	 */
442 	if (segkp_bitmap && seg == &kvseg &&
443 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
444 		return (SEGOP_FAULT(hat, segkp, addr, size, type, rw));
445 
446 	if (rw != S_READ && rw != S_WRITE && rw != S_OTHER)
447 		return (FC_NOSUPPORT);
448 
449 	npages = btopr(size);
450 
451 	switch (type) {
452 	case F_SOFTLOCK:	/* lock down already-loaded translations */
453 		for (pg = 0; pg < npages; pg++) {
454 			pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr,
455 			    SE_SHARED);
456 			if (pp == NULL) {
457 				/*
458 				 * Hmm, no page. Does a kernel mapping
459 				 * exist for it?
460 				 */
461 				if (!hat_probe(kas.a_hat, addr)) {
462 					addr -= PAGESIZE;
463 					while (--pg >= 0) {
464 						pp = page_find(vp, (u_offset_t)
465 						    (uintptr_t)addr);
466 						if (pp)
467 							page_unlock(pp);
468 						addr -= PAGESIZE;
469 					}
470 					return (FC_NOMAP);
471 				}
472 			}
473 			addr += PAGESIZE;
474 		}
475 		if (rw == S_OTHER)
476 			hat_reserve(seg->s_as, addr, size);
477 		return (0);
478 	case F_SOFTUNLOCK:
479 		while (npages--) {
480 			pp = page_find(vp, (u_offset_t)(uintptr_t)addr);
481 			if (pp)
482 				page_unlock(pp);
483 			addr += PAGESIZE;
484 		}
485 		return (0);
486 	default:
487 		return (FC_NOSUPPORT);
488 	}
489 	/*NOTREACHED*/
490 }
491 
492 static int
493 segkmem_setprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
494 {
495 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
496 
497 	if (seg->s_as != &kas || size > seg->s_size ||
498 	    addr < seg->s_base || addr + size > seg->s_base + seg->s_size)
499 		panic("segkmem_setprot: bad args");
500 
501 	/*
502 	 * If it is one of segkp pages, call segkp.
503 	 */
504 	if (segkp_bitmap && seg == &kvseg &&
505 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
506 		return (SEGOP_SETPROT(segkp, addr, size, prot));
507 
508 	if (prot == 0)
509 		hat_unload(kas.a_hat, addr, size, HAT_UNLOAD);
510 	else
511 		hat_chgprot(kas.a_hat, addr, size, prot);
512 	return (0);
513 }
514 
515 /*
516  * This is a dummy segkmem function overloaded to call segkp
517  * when segkp is under the heap.
518  */
519 /* ARGSUSED */
520 static int
521 segkmem_checkprot(struct seg *seg, caddr_t addr, size_t size, uint_t prot)
522 {
523 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
524 
525 	if (seg->s_as != &kas)
526 		segkmem_badop();
527 
528 	/*
529 	 * If it is one of segkp pages, call into segkp.
530 	 */
531 	if (segkp_bitmap && seg == &kvseg &&
532 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
533 		return (SEGOP_CHECKPROT(segkp, addr, size, prot));
534 
535 	segkmem_badop();
536 	return (0);
537 }
538 
539 /*
540  * This is a dummy segkmem function overloaded to call segkp
541  * when segkp is under the heap.
542  */
543 /* ARGSUSED */
544 static int
545 segkmem_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
546 {
547 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
548 
549 	if (seg->s_as != &kas)
550 		segkmem_badop();
551 
552 	/*
553 	 * If it is one of segkp pages, call into segkp.
554 	 */
555 	if (segkp_bitmap && seg == &kvseg &&
556 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
557 		return (SEGOP_KLUSTER(segkp, addr, delta));
558 
559 	segkmem_badop();
560 	return (0);
561 }
562 
563 static void
564 segkmem_xdump_range(void *arg, void *start, size_t size)
565 {
566 	struct as *as = arg;
567 	caddr_t addr = start;
568 	caddr_t addr_end = addr + size;
569 
570 	while (addr < addr_end) {
571 		pfn_t pfn = hat_getpfnum(kas.a_hat, addr);
572 		if (pfn != PFN_INVALID && pfn <= physmax && pf_is_memory(pfn))
573 			dump_addpage(as, addr, pfn);
574 		addr += PAGESIZE;
575 		dump_timeleft = dump_timeout;
576 	}
577 }
578 
579 static void
580 segkmem_dump_range(void *arg, void *start, size_t size)
581 {
582 	caddr_t addr = start;
583 	caddr_t addr_end = addr + size;
584 
585 	/*
586 	 * If we are about to start dumping the range of addresses we
587 	 * carved out of the kernel heap for the large page heap walk
588 	 * heap_lp_arena to find what segments are actually populated
589 	 */
590 	if (SEGKMEM_USE_LARGEPAGES &&
591 	    addr == heap_lp_base && addr_end == heap_lp_end &&
592 	    vmem_size(heap_lp_arena, VMEM_ALLOC) < size) {
593 		vmem_walk(heap_lp_arena, VMEM_ALLOC | VMEM_REENTRANT,
594 		    segkmem_xdump_range, arg);
595 	} else {
596 		segkmem_xdump_range(arg, start, size);
597 	}
598 }
599 
600 static void
601 segkmem_dump(struct seg *seg)
602 {
603 	/*
604 	 * The kernel's heap_arena (represented by kvseg) is a very large
605 	 * VA space, most of which is typically unused.  To speed up dumping
606 	 * we use vmem_walk() to quickly find the pieces of heap_arena that
607 	 * are actually in use.  We do the same for heap32_arena and
608 	 * heap_core.
609 	 *
610 	 * We specify VMEM_REENTRANT to vmem_walk() because dump_addpage()
611 	 * may ultimately need to allocate memory.  Reentrant walks are
612 	 * necessarily imperfect snapshots.  The kernel heap continues
613 	 * to change during a live crash dump, for example.  For a normal
614 	 * crash dump, however, we know that there won't be any other threads
615 	 * messing with the heap.  Therefore, at worst, we may fail to dump
616 	 * the pages that get allocated by the act of dumping; but we will
617 	 * always dump every page that was allocated when the walk began.
618 	 *
619 	 * The other segkmem segments are dense (fully populated), so there's
620 	 * no need to use this technique when dumping them.
621 	 *
622 	 * Note: when adding special dump handling for any new sparsely-
623 	 * populated segments, be sure to add similar handling to the ::kgrep
624 	 * code in mdb.
625 	 */
626 	if (seg == &kvseg) {
627 		vmem_walk(heap_arena, VMEM_ALLOC | VMEM_REENTRANT,
628 		    segkmem_dump_range, seg->s_as);
629 #ifndef __sparc
630 		vmem_walk(heaptext_arena, VMEM_ALLOC | VMEM_REENTRANT,
631 		    segkmem_dump_range, seg->s_as);
632 #endif
633 	} else if (seg == &kvseg_core) {
634 		vmem_walk(heap_core_arena, VMEM_ALLOC | VMEM_REENTRANT,
635 		    segkmem_dump_range, seg->s_as);
636 	} else if (seg == &kvseg32) {
637 		vmem_walk(heap32_arena, VMEM_ALLOC | VMEM_REENTRANT,
638 		    segkmem_dump_range, seg->s_as);
639 		vmem_walk(heaptext_arena, VMEM_ALLOC | VMEM_REENTRANT,
640 		    segkmem_dump_range, seg->s_as);
641 	} else if (seg == &kzioseg) {
642 		/*
643 		 * We don't want to dump pages attached to kzioseg since they
644 		 * contain file data from ZFS.  If this page's segment is
645 		 * kzioseg return instead of writing it to the dump device.
646 		 */
647 		return;
648 	} else {
649 		segkmem_dump_range(seg->s_as, seg->s_base, seg->s_size);
650 	}
651 }
652 
653 /*
654  * lock/unlock kmem pages over a given range [addr, addr+len).
655  * Returns a shadow list of pages in ppp. If there are holes
656  * in the range (e.g. some of the kernel mappings do not have
657  * underlying page_ts) returns ENOTSUP so that as_pagelock()
658  * will handle the range via as_fault(F_SOFTLOCK).
659  */
660 /*ARGSUSED*/
661 static int
662 segkmem_pagelock(struct seg *seg, caddr_t addr, size_t len,
663 	page_t ***ppp, enum lock_type type, enum seg_rw rw)
664 {
665 	page_t **pplist, *pp;
666 	pgcnt_t npages;
667 	spgcnt_t pg;
668 	size_t nb;
669 	struct vnode *vp = seg->s_data;
670 
671 	ASSERT(ppp != NULL);
672 
673 	/*
674 	 * If it is one of segkp pages, call into segkp.
675 	 */
676 	if (segkp_bitmap && seg == &kvseg &&
677 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
678 		return (SEGOP_PAGELOCK(segkp, addr, len, ppp, type, rw));
679 
680 	npages = btopr(len);
681 	nb = sizeof (page_t *) * npages;
682 
683 	if (type == L_PAGEUNLOCK) {
684 		pplist = *ppp;
685 		ASSERT(pplist != NULL);
686 
687 		for (pg = 0; pg < npages; pg++) {
688 			pp = pplist[pg];
689 			page_unlock(pp);
690 		}
691 		kmem_free(pplist, nb);
692 		return (0);
693 	}
694 
695 	ASSERT(type == L_PAGELOCK);
696 
697 	pplist = kmem_alloc(nb, KM_NOSLEEP);
698 	if (pplist == NULL) {
699 		*ppp = NULL;
700 		return (ENOTSUP);	/* take the slow path */
701 	}
702 
703 	for (pg = 0; pg < npages; pg++) {
704 		pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, SE_SHARED);
705 		if (pp == NULL) {
706 			while (--pg >= 0)
707 				page_unlock(pplist[pg]);
708 			kmem_free(pplist, nb);
709 			*ppp = NULL;
710 			return (ENOTSUP);
711 		}
712 		pplist[pg] = pp;
713 		addr += PAGESIZE;
714 	}
715 
716 	*ppp = pplist;
717 	return (0);
718 }
719 
720 /*
721  * This is a dummy segkmem function overloaded to call segkp
722  * when segkp is under the heap.
723  */
724 /* ARGSUSED */
725 static int
726 segkmem_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
727 {
728 	ASSERT(RW_LOCK_HELD(&seg->s_as->a_lock));
729 
730 	if (seg->s_as != &kas)
731 		segkmem_badop();
732 
733 	/*
734 	 * If it is one of segkp pages, call into segkp.
735 	 */
736 	if (segkp_bitmap && seg == &kvseg &&
737 	    BT_TEST(segkp_bitmap, btop((uintptr_t)(addr - seg->s_base))))
738 		return (SEGOP_GETMEMID(segkp, addr, memidp));
739 
740 	segkmem_badop();
741 	return (0);
742 }
743 
744 /*ARGSUSED*/
745 static lgrp_mem_policy_info_t *
746 segkmem_getpolicy(struct seg *seg, caddr_t addr)
747 {
748 	return (NULL);
749 }
750 
751 /*ARGSUSED*/
752 static int
753 segkmem_capable(struct seg *seg, segcapability_t capability)
754 {
755 	if (capability == S_CAPABILITY_NOMINFLT)
756 		return (1);
757 	return (0);
758 }
759 
760 static struct seg_ops segkmem_ops = {
761 	SEGKMEM_BADOP(int),		/* dup */
762 	SEGKMEM_BADOP(int),		/* unmap */
763 	SEGKMEM_BADOP(void),		/* free */
764 	segkmem_fault,
765 	SEGKMEM_BADOP(faultcode_t),	/* faulta */
766 	segkmem_setprot,
767 	segkmem_checkprot,
768 	segkmem_kluster,
769 	SEGKMEM_BADOP(size_t),		/* swapout */
770 	SEGKMEM_BADOP(int),		/* sync */
771 	SEGKMEM_BADOP(size_t),		/* incore */
772 	SEGKMEM_BADOP(int),		/* lockop */
773 	SEGKMEM_BADOP(int),		/* getprot */
774 	SEGKMEM_BADOP(u_offset_t),	/* getoffset */
775 	SEGKMEM_BADOP(int),		/* gettype */
776 	SEGKMEM_BADOP(int),		/* getvp */
777 	SEGKMEM_BADOP(int),		/* advise */
778 	segkmem_dump,
779 	segkmem_pagelock,
780 	SEGKMEM_BADOP(int),		/* setpgsz */
781 	segkmem_getmemid,
782 	segkmem_getpolicy,		/* getpolicy */
783 	segkmem_capable,		/* capable */
784 };
785 
786 int
787 segkmem_zio_create(struct seg *seg)
788 {
789 	ASSERT(seg->s_as == &kas && RW_WRITE_HELD(&kas.a_lock));
790 	seg->s_ops = &segkmem_ops;
791 	seg->s_data = &zvp;
792 	kas.a_size += seg->s_size;
793 	return (0);
794 }
795 
796 int
797 segkmem_create(struct seg *seg)
798 {
799 	ASSERT(seg->s_as == &kas && RW_WRITE_HELD(&kas.a_lock));
800 	seg->s_ops = &segkmem_ops;
801 	seg->s_data = &kvp;
802 	kas.a_size += seg->s_size;
803 	return (0);
804 }
805 
806 /*ARGSUSED*/
807 page_t *
808 segkmem_page_create(void *addr, size_t size, int vmflag, void *arg)
809 {
810 	struct seg kseg;
811 	int pgflags;
812 	struct vnode *vp = arg;
813 
814 	if (vp == NULL)
815 		vp = &kvp;
816 
817 	kseg.s_as = &kas;
818 	pgflags = PG_EXCL;
819 
820 	if (segkmem_reloc == 0 || (vmflag & VM_NORELOC))
821 		pgflags |= PG_NORELOC;
822 	if ((vmflag & VM_NOSLEEP) == 0)
823 		pgflags |= PG_WAIT;
824 	if (vmflag & VM_PANIC)
825 		pgflags |= PG_PANIC;
826 	if (vmflag & VM_PUSHPAGE)
827 		pgflags |= PG_PUSHPAGE;
828 
829 	return (page_create_va(vp, (u_offset_t)(uintptr_t)addr, size,
830 	    pgflags, &kseg, addr));
831 }
832 
833 /*
834  * Allocate pages to back the virtual address range [addr, addr + size).
835  * If addr is NULL, allocate the virtual address space as well.
836  */
837 void *
838 segkmem_xalloc(vmem_t *vmp, void *inaddr, size_t size, int vmflag, uint_t attr,
839 	page_t *(*page_create_func)(void *, size_t, int, void *), void *pcarg)
840 {
841 	page_t *ppl;
842 	caddr_t addr = inaddr;
843 	pgcnt_t npages = btopr(size);
844 	int allocflag;
845 
846 	if (inaddr == NULL && (addr = vmem_alloc(vmp, size, vmflag)) == NULL)
847 		return (NULL);
848 
849 	ASSERT(((uintptr_t)addr & PAGEOFFSET) == 0);
850 
851 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
852 		if (inaddr == NULL)
853 			vmem_free(vmp, addr, size);
854 		return (NULL);
855 	}
856 
857 	ppl = page_create_func(addr, size, vmflag, pcarg);
858 	if (ppl == NULL) {
859 		if (inaddr == NULL)
860 			vmem_free(vmp, addr, size);
861 		page_unresv(npages);
862 		return (NULL);
863 	}
864 
865 	/*
866 	 * Under certain conditions, we need to let the HAT layer know
867 	 * that it cannot safely allocate memory.  Allocations from
868 	 * the hat_memload vmem arena always need this, to prevent
869 	 * infinite recursion.
870 	 *
871 	 * In addition, the x86 hat cannot safely do memory
872 	 * allocations while in vmem_populate(), because there
873 	 * is no simple bound on its usage.
874 	 */
875 	if (vmflag & VM_MEMLOAD)
876 		allocflag = HAT_NO_KALLOC;
877 #if defined(__x86)
878 	else if (vmem_is_populator())
879 		allocflag = HAT_NO_KALLOC;
880 #endif
881 	else
882 		allocflag = 0;
883 
884 	while (ppl != NULL) {
885 		page_t *pp = ppl;
886 		page_sub(&ppl, pp);
887 		ASSERT(page_iolock_assert(pp));
888 		ASSERT(PAGE_EXCL(pp));
889 		page_io_unlock(pp);
890 		hat_memload(kas.a_hat, (caddr_t)(uintptr_t)pp->p_offset, pp,
891 		    (PROT_ALL & ~PROT_USER) | HAT_NOSYNC | attr,
892 		    HAT_LOAD_LOCK | allocflag);
893 		pp->p_lckcnt = 1;
894 #if defined(__x86)
895 		page_downgrade(pp);
896 #else
897 		if (vmflag & SEGKMEM_SHARELOCKED)
898 			page_downgrade(pp);
899 		else
900 			page_unlock(pp);
901 #endif
902 	}
903 
904 	return (addr);
905 }
906 
907 static void *
908 segkmem_alloc_vn(vmem_t *vmp, size_t size, int vmflag, struct vnode *vp)
909 {
910 	void *addr;
911 	segkmem_gc_list_t *gcp, **prev_gcpp;
912 
913 	ASSERT(vp != NULL);
914 
915 	if (kvseg.s_base == NULL) {
916 #ifndef __sparc
917 		if (bootops->bsys_alloc == NULL)
918 			halt("Memory allocation between bop_alloc() and "
919 			    "kmem_alloc().\n");
920 #endif
921 
922 		/*
923 		 * There's not a lot of memory to go around during boot,
924 		 * so recycle it if we can.
925 		 */
926 		for (prev_gcpp = &segkmem_gc_list; (gcp = *prev_gcpp) != NULL;
927 		    prev_gcpp = &gcp->gc_next) {
928 			if (gcp->gc_arena == vmp && gcp->gc_size == size) {
929 				*prev_gcpp = gcp->gc_next;
930 				return (gcp);
931 			}
932 		}
933 
934 		addr = vmem_alloc(vmp, size, vmflag | VM_PANIC);
935 		if (boot_alloc(addr, size, BO_NO_ALIGN) != addr)
936 			panic("segkmem_alloc: boot_alloc failed");
937 		return (addr);
938 	}
939 	return (segkmem_xalloc(vmp, NULL, size, vmflag, 0,
940 	    segkmem_page_create, vp));
941 }
942 
943 void *
944 segkmem_alloc(vmem_t *vmp, size_t size, int vmflag)
945 {
946 	return (segkmem_alloc_vn(vmp, size, vmflag, &kvp));
947 }
948 
949 void *
950 segkmem_zio_alloc(vmem_t *vmp, size_t size, int vmflag)
951 {
952 	return (segkmem_alloc_vn(vmp, size, vmflag, &zvp));
953 }
954 
955 /*
956  * Any changes to this routine must also be carried over to
957  * devmap_free_pages() in the seg_dev driver. This is because
958  * we currently don't have a special kernel segment for non-paged
959  * kernel memory that is exported by drivers to user space.
960  */
961 static void
962 segkmem_free_vn(vmem_t *vmp, void *inaddr, size_t size, struct vnode *vp,
963     void (*func)(page_t *))
964 {
965 	page_t *pp;
966 	caddr_t addr = inaddr;
967 	caddr_t eaddr;
968 	pgcnt_t npages = btopr(size);
969 
970 	ASSERT(((uintptr_t)addr & PAGEOFFSET) == 0);
971 	ASSERT(vp != NULL);
972 
973 	if (kvseg.s_base == NULL) {
974 		segkmem_gc_list_t *gc = inaddr;
975 		gc->gc_arena = vmp;
976 		gc->gc_size = size;
977 		gc->gc_next = segkmem_gc_list;
978 		segkmem_gc_list = gc;
979 		return;
980 	}
981 
982 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
983 
984 	for (eaddr = addr + size; addr < eaddr; addr += PAGESIZE) {
985 #if defined(__x86)
986 		pp = page_find(vp, (u_offset_t)(uintptr_t)addr);
987 		if (pp == NULL)
988 			panic("segkmem_free: page not found");
989 		if (!page_tryupgrade(pp)) {
990 			/*
991 			 * Some other thread has a sharelock. Wait for
992 			 * it to drop the lock so we can free this page.
993 			 */
994 			page_unlock(pp);
995 			pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr,
996 			    SE_EXCL);
997 		}
998 #else
999 		pp = page_lookup(vp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
1000 #endif
1001 		if (pp == NULL)
1002 			panic("segkmem_free: page not found");
1003 		/* Clear p_lckcnt so page_destroy() doesn't update availrmem */
1004 		pp->p_lckcnt = 0;
1005 		if (func)
1006 			func(pp);
1007 		else
1008 			page_destroy(pp, 0);
1009 	}
1010 	if (func == NULL)
1011 		page_unresv(npages);
1012 
1013 	if (vmp != NULL)
1014 		vmem_free(vmp, inaddr, size);
1015 
1016 }
1017 
1018 void
1019 segkmem_xfree(vmem_t *vmp, void *inaddr, size_t size, void (*func)(page_t *))
1020 {
1021 	segkmem_free_vn(vmp, inaddr, size, &kvp, func);
1022 }
1023 
1024 void
1025 segkmem_free(vmem_t *vmp, void *inaddr, size_t size)
1026 {
1027 	segkmem_free_vn(vmp, inaddr, size, &kvp, NULL);
1028 }
1029 
1030 void
1031 segkmem_zio_free(vmem_t *vmp, void *inaddr, size_t size)
1032 {
1033 	segkmem_free_vn(vmp, inaddr, size, &zvp, NULL);
1034 }
1035 
1036 void
1037 segkmem_gc(void)
1038 {
1039 	ASSERT(kvseg.s_base != NULL);
1040 	while (segkmem_gc_list != NULL) {
1041 		segkmem_gc_list_t *gc = segkmem_gc_list;
1042 		segkmem_gc_list = gc->gc_next;
1043 		segkmem_free(gc->gc_arena, gc, gc->gc_size);
1044 	}
1045 }
1046 
1047 /*
1048  * Legacy entry points from here to end of file.
1049  */
1050 void
1051 segkmem_mapin(struct seg *seg, void *addr, size_t size, uint_t vprot,
1052     pfn_t pfn, uint_t flags)
1053 {
1054 	hat_unload(seg->s_as->a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1055 	hat_devload(seg->s_as->a_hat, addr, size, pfn, vprot,
1056 	    flags | HAT_LOAD_LOCK);
1057 }
1058 
1059 void
1060 segkmem_mapout(struct seg *seg, void *addr, size_t size)
1061 {
1062 	hat_unload(seg->s_as->a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1063 }
1064 
1065 void *
1066 kmem_getpages(pgcnt_t npages, int kmflag)
1067 {
1068 	return (kmem_alloc(ptob(npages), kmflag));
1069 }
1070 
1071 void
1072 kmem_freepages(void *addr, pgcnt_t npages)
1073 {
1074 	kmem_free(addr, ptob(npages));
1075 }
1076 
1077 /*
1078  * segkmem_page_create_large() allocates a large page to be used for the kmem
1079  * caches. If kpr is enabled we ask for a relocatable page unless requested
1080  * otherwise. If kpr is disabled we have to ask for a non-reloc page
1081  */
1082 static page_t *
1083 segkmem_page_create_large(void *addr, size_t size, int vmflag, void *arg)
1084 {
1085 	int pgflags;
1086 
1087 	pgflags = PG_EXCL;
1088 
1089 	if (segkmem_reloc == 0 || (vmflag & VM_NORELOC))
1090 		pgflags |= PG_NORELOC;
1091 	if (!(vmflag & VM_NOSLEEP))
1092 		pgflags |= PG_WAIT;
1093 	if (vmflag & VM_PUSHPAGE)
1094 		pgflags |= PG_PUSHPAGE;
1095 
1096 	return (page_create_va_large(&kvp, (u_offset_t)(uintptr_t)addr, size,
1097 	    pgflags, &kvseg, addr, arg));
1098 }
1099 
1100 /*
1101  * Allocate a large page to back the virtual address range
1102  * [addr, addr + size).  If addr is NULL, allocate the virtual address
1103  * space as well.
1104  */
1105 static void *
1106 segkmem_xalloc_lp(vmem_t *vmp, void *inaddr, size_t size, int vmflag,
1107     uint_t attr, page_t *(*page_create_func)(void *, size_t, int, void *),
1108     void *pcarg)
1109 {
1110 	caddr_t addr = inaddr, pa;
1111 	size_t  lpsize = segkmem_lpsize;
1112 	pgcnt_t npages = btopr(size);
1113 	pgcnt_t nbpages = btop(lpsize);
1114 	pgcnt_t nlpages = size >> segkmem_lpshift;
1115 	size_t  ppasize = nbpages * sizeof (page_t *);
1116 	page_t *pp, *rootpp, **ppa, *pplist = NULL;
1117 	int i;
1118 
1119 	vmflag |= VM_NOSLEEP;
1120 
1121 	if (page_resv(npages, vmflag & VM_KMFLAGS) == 0) {
1122 		return (NULL);
1123 	}
1124 
1125 	/*
1126 	 * allocate an array we need for hat_memload_array.
1127 	 * we use a separate arena to avoid recursion.
1128 	 * we will not need this array when hat_memload_array learns pp++
1129 	 */
1130 	if ((ppa = vmem_alloc(segkmem_ppa_arena, ppasize, vmflag)) == NULL) {
1131 		goto fail_array_alloc;
1132 	}
1133 
1134 	if (inaddr == NULL && (addr = vmem_alloc(vmp, size, vmflag)) == NULL)
1135 		goto fail_vmem_alloc;
1136 
1137 	ASSERT(((uintptr_t)addr & (lpsize - 1)) == 0);
1138 
1139 	/* create all the pages */
1140 	for (pa = addr, i = 0; i < nlpages; i++, pa += lpsize) {
1141 		if ((pp = page_create_func(pa, lpsize, vmflag, pcarg)) == NULL)
1142 			goto fail_page_create;
1143 		page_list_concat(&pplist, &pp);
1144 	}
1145 
1146 	/* at this point we have all the resource to complete the request */
1147 	while ((rootpp = pplist) != NULL) {
1148 		for (i = 0; i < nbpages; i++) {
1149 			ASSERT(pplist != NULL);
1150 			pp = pplist;
1151 			page_sub(&pplist, pp);
1152 			ASSERT(page_iolock_assert(pp));
1153 			page_io_unlock(pp);
1154 			ppa[i] = pp;
1155 		}
1156 		/*
1157 		 * Load the locked entry. It's OK to preload the entry into the
1158 		 * TSB since we now support large mappings in the kernel TSB.
1159 		 */
1160 		hat_memload_array(kas.a_hat,
1161 		    (caddr_t)(uintptr_t)rootpp->p_offset, lpsize,
1162 		    ppa, (PROT_ALL & ~PROT_USER) | HAT_NOSYNC | attr,
1163 		    HAT_LOAD_LOCK);
1164 
1165 		for (--i; i >= 0; --i) {
1166 			ppa[i]->p_lckcnt = 1;
1167 			page_unlock(ppa[i]);
1168 		}
1169 	}
1170 
1171 	vmem_free(segkmem_ppa_arena, ppa, ppasize);
1172 	return (addr);
1173 
1174 fail_page_create:
1175 	while ((rootpp = pplist) != NULL) {
1176 		for (i = 0, pp = pplist; i < nbpages; i++, pp = pplist) {
1177 			ASSERT(pp != NULL);
1178 			page_sub(&pplist, pp);
1179 			ASSERT(page_iolock_assert(pp));
1180 			page_io_unlock(pp);
1181 		}
1182 		page_destroy_pages(rootpp);
1183 	}
1184 
1185 	if (inaddr == NULL)
1186 		vmem_free(vmp, addr, size);
1187 
1188 fail_vmem_alloc:
1189 	vmem_free(segkmem_ppa_arena, ppa, ppasize);
1190 
1191 fail_array_alloc:
1192 	page_unresv(npages);
1193 
1194 	return (NULL);
1195 }
1196 
1197 static void
1198 segkmem_free_one_lp(caddr_t addr, size_t size)
1199 {
1200 	page_t		*pp, *rootpp = NULL;
1201 	pgcnt_t 	pgs_left = btopr(size);
1202 
1203 	ASSERT(size == segkmem_lpsize);
1204 
1205 	hat_unload(kas.a_hat, addr, size, HAT_UNLOAD_UNLOCK);
1206 
1207 	for (; pgs_left > 0; addr += PAGESIZE, pgs_left--) {
1208 		pp = page_lookup(&kvp, (u_offset_t)(uintptr_t)addr, SE_EXCL);
1209 		if (pp == NULL)
1210 			panic("segkmem_free_one_lp: page not found");
1211 		ASSERT(PAGE_EXCL(pp));
1212 		pp->p_lckcnt = 0;
1213 		if (rootpp == NULL)
1214 			rootpp = pp;
1215 	}
1216 	ASSERT(rootpp != NULL);
1217 	page_destroy_pages(rootpp);
1218 
1219 	/* page_unresv() is done by the caller */
1220 }
1221 
1222 /*
1223  * This function is called to import new spans into the vmem arenas like
1224  * kmem_default_arena and kmem_oversize_arena. It first tries to import
1225  * spans from large page arena - kmem_lp_arena. In order to do this it might
1226  * have to "upgrade the requested size" to kmem_lp_arena quantum. If
1227  * it was not able to satisfy the upgraded request it then calls regular
1228  * segkmem_alloc() that satisfies the request by importing from "*vmp" arena
1229  */
1230 /*ARGSUSED*/
1231 void *
1232 segkmem_alloc_lp(vmem_t *vmp, size_t *sizep, size_t align, int vmflag)
1233 {
1234 	size_t size;
1235 	kthread_t *t = curthread;
1236 	segkmem_lpcb_t *lpcb = &segkmem_lpcb;
1237 
1238 	ASSERT(sizep != NULL);
1239 
1240 	size = *sizep;
1241 
1242 	if (lpcb->lp_uselp && !(t->t_flag & T_PANIC) &&
1243 	    !(vmflag & SEGKMEM_SHARELOCKED)) {
1244 
1245 		size_t kmemlp_qnt = segkmem_kmemlp_quantum;
1246 		size_t asize = P2ROUNDUP(size, kmemlp_qnt);
1247 		void  *addr = NULL;
1248 		ulong_t *lpthrtp = &lpcb->lp_throttle;
1249 		ulong_t lpthrt = *lpthrtp;
1250 		int	dowakeup = 0;
1251 		int	doalloc = 1;
1252 
1253 		ASSERT(kmem_lp_arena != NULL);
1254 		ASSERT(asize >= size);
1255 
1256 		if (lpthrt != 0) {
1257 			/* try to update the throttle value */
1258 			lpthrt = atomic_add_long_nv(lpthrtp, 1);
1259 			if (lpthrt >= segkmem_lpthrottle_max) {
1260 				lpthrt = atomic_cas_ulong(lpthrtp, lpthrt,
1261 				    segkmem_lpthrottle_max / 4);
1262 			}
1263 
1264 			/*
1265 			 * when we get above throttle start do an exponential
1266 			 * backoff at trying large pages and reaping
1267 			 */
1268 			if (lpthrt > segkmem_lpthrottle_start &&
1269 			    (lpthrt & (lpthrt - 1))) {
1270 				lpcb->allocs_throttled++;
1271 				lpthrt--;
1272 				if ((lpthrt & (lpthrt - 1)) == 0)
1273 					kmem_reap();
1274 				return (segkmem_alloc(vmp, size, vmflag));
1275 			}
1276 		}
1277 
1278 		if (!(vmflag & VM_NOSLEEP) &&
1279 		    segkmem_heaplp_quantum >= (8 * kmemlp_qnt) &&
1280 		    vmem_size(kmem_lp_arena, VMEM_FREE) <= kmemlp_qnt &&
1281 		    asize < (segkmem_heaplp_quantum - kmemlp_qnt)) {
1282 
1283 			/*
1284 			 * we are low on free memory in kmem_lp_arena
1285 			 * we let only one guy to allocate heap_lp
1286 			 * quantum size chunk that everybody is going to
1287 			 * share
1288 			 */
1289 			mutex_enter(&lpcb->lp_lock);
1290 
1291 			if (lpcb->lp_wait) {
1292 
1293 				/* we are not the first one - wait */
1294 				cv_wait(&lpcb->lp_cv, &lpcb->lp_lock);
1295 				if (vmem_size(kmem_lp_arena, VMEM_FREE) <
1296 				    kmemlp_qnt)  {
1297 					doalloc = 0;
1298 				}
1299 			} else if (vmem_size(kmem_lp_arena, VMEM_FREE) <=
1300 			    kmemlp_qnt) {
1301 
1302 				/*
1303 				 * we are the first one, make sure we import
1304 				 * a large page
1305 				 */
1306 				if (asize == kmemlp_qnt)
1307 					asize += kmemlp_qnt;
1308 				dowakeup = 1;
1309 				lpcb->lp_wait = 1;
1310 			}
1311 
1312 			mutex_exit(&lpcb->lp_lock);
1313 		}
1314 
1315 		/*
1316 		 * VM_ABORT flag prevents sleeps in vmem_xalloc when
1317 		 * large pages are not available. In that case this allocation
1318 		 * attempt will fail and we will retry allocation with small
1319 		 * pages. We also do not want to panic if this allocation fails
1320 		 * because we are going to retry.
1321 		 */
1322 		if (doalloc) {
1323 			addr = vmem_alloc(kmem_lp_arena, asize,
1324 			    (vmflag | VM_ABORT) & ~VM_PANIC);
1325 
1326 			if (dowakeup) {
1327 				mutex_enter(&lpcb->lp_lock);
1328 				ASSERT(lpcb->lp_wait != 0);
1329 				lpcb->lp_wait = 0;
1330 				cv_broadcast(&lpcb->lp_cv);
1331 				mutex_exit(&lpcb->lp_lock);
1332 			}
1333 		}
1334 
1335 		if (addr != NULL) {
1336 			*sizep = asize;
1337 			*lpthrtp = 0;
1338 			return (addr);
1339 		}
1340 
1341 		if (vmflag & VM_NOSLEEP)
1342 			lpcb->nosleep_allocs_failed++;
1343 		else
1344 			lpcb->sleep_allocs_failed++;
1345 		lpcb->alloc_bytes_failed += size;
1346 
1347 		/* if large page throttling is not started yet do it */
1348 		if (segkmem_use_lpthrottle && lpthrt == 0) {
1349 			lpthrt = atomic_cas_ulong(lpthrtp, lpthrt, 1);
1350 		}
1351 	}
1352 	return (segkmem_alloc(vmp, size, vmflag));
1353 }
1354 
1355 void
1356 segkmem_free_lp(vmem_t *vmp, void *inaddr, size_t size)
1357 {
1358 	if (kmem_lp_arena == NULL || !IS_KMEM_VA_LARGEPAGE((caddr_t)inaddr)) {
1359 		segkmem_free(vmp, inaddr, size);
1360 	} else {
1361 		vmem_free(kmem_lp_arena, inaddr, size);
1362 	}
1363 }
1364 
1365 /*
1366  * segkmem_alloc_lpi() imports virtual memory from large page heap arena
1367  * into kmem_lp arena. In the process it maps the imported segment with
1368  * large pages
1369  */
1370 static void *
1371 segkmem_alloc_lpi(vmem_t *vmp, size_t size, int vmflag)
1372 {
1373 	segkmem_lpcb_t *lpcb = &segkmem_lpcb;
1374 	void  *addr;
1375 
1376 	ASSERT(size != 0);
1377 	ASSERT(vmp == heap_lp_arena);
1378 
1379 	/* do not allow large page heap grow beyound limits */
1380 	if (vmem_size(vmp, VMEM_ALLOC) >= segkmem_kmemlp_max) {
1381 		lpcb->allocs_limited++;
1382 		return (NULL);
1383 	}
1384 
1385 	addr = segkmem_xalloc_lp(vmp, NULL, size, vmflag, 0,
1386 	    segkmem_page_create_large, NULL);
1387 	return (addr);
1388 }
1389 
1390 /*
1391  * segkmem_free_lpi() returns virtual memory back into large page heap arena
1392  * from kmem_lp arena. Beore doing this it unmaps the segment and frees
1393  * large pages used to map it.
1394  */
1395 static void
1396 segkmem_free_lpi(vmem_t *vmp, void *inaddr, size_t size)
1397 {
1398 	pgcnt_t		nlpages = size >> segkmem_lpshift;
1399 	size_t		lpsize = segkmem_lpsize;
1400 	caddr_t		addr = inaddr;
1401 	pgcnt_t 	npages = btopr(size);
1402 	int		i;
1403 
1404 	ASSERT(vmp == heap_lp_arena);
1405 	ASSERT(IS_KMEM_VA_LARGEPAGE(addr));
1406 	ASSERT(((uintptr_t)inaddr & (lpsize - 1)) == 0);
1407 
1408 	for (i = 0; i < nlpages; i++) {
1409 		segkmem_free_one_lp(addr, lpsize);
1410 		addr += lpsize;
1411 	}
1412 
1413 	page_unresv(npages);
1414 
1415 	vmem_free(vmp, inaddr, size);
1416 }
1417 
1418 /*
1419  * This function is called at system boot time by kmem_init right after
1420  * /etc/system file has been read. It checks based on hardware configuration
1421  * and /etc/system settings if system is going to use large pages. The
1422  * initialiazation necessary to actually start using large pages
1423  * happens later in the process after segkmem_heap_lp_init() is called.
1424  */
1425 int
1426 segkmem_lpsetup()
1427 {
1428 	int use_large_pages = 0;
1429 
1430 #ifdef __sparc
1431 
1432 	size_t memtotal = physmem * PAGESIZE;
1433 
1434 	if (heap_lp_base == NULL) {
1435 		segkmem_lpsize = PAGESIZE;
1436 		return (0);
1437 	}
1438 
1439 	/* get a platform dependent value of large page size for kernel heap */
1440 	segkmem_lpsize = get_segkmem_lpsize(segkmem_lpsize);
1441 
1442 	if (segkmem_lpsize <= PAGESIZE) {
1443 		/*
1444 		 * put virtual space reserved for the large page kernel
1445 		 * back to the regular heap
1446 		 */
1447 		vmem_xfree(heap_arena, heap_lp_base,
1448 		    heap_lp_end - heap_lp_base);
1449 		heap_lp_base = NULL;
1450 		heap_lp_end = NULL;
1451 		segkmem_lpsize = PAGESIZE;
1452 		return (0);
1453 	}
1454 
1455 	/* set heap_lp quantum if necessary */
1456 	if (segkmem_heaplp_quantum == 0 ||
1457 	    (segkmem_heaplp_quantum & (segkmem_heaplp_quantum - 1)) ||
1458 	    P2PHASE(segkmem_heaplp_quantum, segkmem_lpsize)) {
1459 		segkmem_heaplp_quantum = segkmem_lpsize;
1460 	}
1461 
1462 	/* set kmem_lp quantum if necessary */
1463 	if (segkmem_kmemlp_quantum == 0 ||
1464 	    (segkmem_kmemlp_quantum & (segkmem_kmemlp_quantum - 1)) ||
1465 	    segkmem_kmemlp_quantum > segkmem_heaplp_quantum) {
1466 		segkmem_kmemlp_quantum = segkmem_heaplp_quantum;
1467 	}
1468 
1469 	/* set total amount of memory allowed for large page kernel heap */
1470 	if (segkmem_kmemlp_max == 0) {
1471 		if (segkmem_kmemlp_pcnt == 0 || segkmem_kmemlp_pcnt > 100)
1472 			segkmem_kmemlp_pcnt = 12;
1473 		segkmem_kmemlp_max = (memtotal * segkmem_kmemlp_pcnt) / 100;
1474 	}
1475 	segkmem_kmemlp_max = P2ROUNDUP(segkmem_kmemlp_max,
1476 	    segkmem_heaplp_quantum);
1477 
1478 	/* fix lp kmem preallocation request if necesssary */
1479 	if (segkmem_kmemlp_min) {
1480 		segkmem_kmemlp_min = P2ROUNDUP(segkmem_kmemlp_min,
1481 		    segkmem_heaplp_quantum);
1482 		if (segkmem_kmemlp_min > segkmem_kmemlp_max)
1483 			segkmem_kmemlp_min = segkmem_kmemlp_max;
1484 	}
1485 
1486 	use_large_pages = 1;
1487 	segkmem_lpszc = page_szc(segkmem_lpsize);
1488 	segkmem_lpshift = page_get_shift(segkmem_lpszc);
1489 
1490 #endif
1491 	return (use_large_pages);
1492 }
1493 
1494 void
1495 segkmem_zio_init(void *zio_mem_base, size_t zio_mem_size)
1496 {
1497 	ASSERT(zio_mem_base != NULL);
1498 	ASSERT(zio_mem_size != 0);
1499 
1500 	zio_arena = vmem_create("zfs_file_data", zio_mem_base, zio_mem_size,
1501 	    PAGESIZE, NULL, NULL, NULL, 0, VM_SLEEP);
1502 
1503 	zio_alloc_arena = vmem_create("zfs_file_data_buf", NULL, 0, PAGESIZE,
1504 	    segkmem_zio_alloc, segkmem_zio_free, zio_arena, 0, VM_SLEEP);
1505 
1506 	ASSERT(zio_arena != NULL);
1507 	ASSERT(zio_alloc_arena != NULL);
1508 }
1509 
1510 #ifdef __sparc
1511 
1512 
1513 static void *
1514 segkmem_alloc_ppa(vmem_t *vmp, size_t size, int vmflag)
1515 {
1516 	size_t ppaquantum = btopr(segkmem_lpsize) * sizeof (page_t *);
1517 	void   *addr;
1518 
1519 	if (ppaquantum <= PAGESIZE)
1520 		return (segkmem_alloc(vmp, size, vmflag));
1521 
1522 	ASSERT((size & (ppaquantum - 1)) == 0);
1523 
1524 	addr = vmem_xalloc(vmp, size, ppaquantum, 0, 0, NULL, NULL, vmflag);
1525 	if (addr != NULL && segkmem_xalloc(vmp, addr, size, vmflag, 0,
1526 	    segkmem_page_create, NULL) == NULL) {
1527 		vmem_xfree(vmp, addr, size);
1528 		addr = NULL;
1529 	}
1530 
1531 	return (addr);
1532 }
1533 
1534 static void
1535 segkmem_free_ppa(vmem_t *vmp, void *addr, size_t size)
1536 {
1537 	size_t ppaquantum = btopr(segkmem_lpsize) * sizeof (page_t *);
1538 
1539 	ASSERT(addr != NULL);
1540 
1541 	if (ppaquantum <= PAGESIZE) {
1542 		segkmem_free(vmp, addr, size);
1543 	} else {
1544 		segkmem_free(NULL, addr, size);
1545 		vmem_xfree(vmp, addr, size);
1546 	}
1547 }
1548 
1549 void
1550 segkmem_heap_lp_init()
1551 {
1552 	segkmem_lpcb_t *lpcb = &segkmem_lpcb;
1553 	size_t heap_lp_size = heap_lp_end - heap_lp_base;
1554 	size_t lpsize = segkmem_lpsize;
1555 	size_t ppaquantum;
1556 	void   *addr;
1557 
1558 	if (segkmem_lpsize <= PAGESIZE) {
1559 		ASSERT(heap_lp_base == NULL);
1560 		ASSERT(heap_lp_end == NULL);
1561 		return;
1562 	}
1563 
1564 	ASSERT(segkmem_heaplp_quantum >= lpsize);
1565 	ASSERT((segkmem_heaplp_quantum & (lpsize - 1)) == 0);
1566 	ASSERT(lpcb->lp_uselp == 0);
1567 	ASSERT(heap_lp_base != NULL);
1568 	ASSERT(heap_lp_end != NULL);
1569 	ASSERT(heap_lp_base < heap_lp_end);
1570 	ASSERT(heap_lp_arena == NULL);
1571 	ASSERT(((uintptr_t)heap_lp_base & (lpsize - 1)) == 0);
1572 	ASSERT(((uintptr_t)heap_lp_end & (lpsize - 1)) == 0);
1573 
1574 	/* create large page heap arena */
1575 	heap_lp_arena = vmem_create("heap_lp", heap_lp_base, heap_lp_size,
1576 	    segkmem_heaplp_quantum, NULL, NULL, NULL, 0, VM_SLEEP);
1577 
1578 	ASSERT(heap_lp_arena != NULL);
1579 
1580 	/* This arena caches memory already mapped by large pages */
1581 	kmem_lp_arena = vmem_create("kmem_lp", NULL, 0, segkmem_kmemlp_quantum,
1582 	    segkmem_alloc_lpi, segkmem_free_lpi, heap_lp_arena, 0, VM_SLEEP);
1583 
1584 	ASSERT(kmem_lp_arena != NULL);
1585 
1586 	mutex_init(&lpcb->lp_lock, NULL, MUTEX_DEFAULT, NULL);
1587 	cv_init(&lpcb->lp_cv, NULL, CV_DEFAULT, NULL);
1588 
1589 	/*
1590 	 * this arena is used for the array of page_t pointers necessary
1591 	 * to call hat_mem_load_array
1592 	 */
1593 	ppaquantum = btopr(lpsize) * sizeof (page_t *);
1594 	segkmem_ppa_arena = vmem_create("segkmem_ppa", NULL, 0, ppaquantum,
1595 	    segkmem_alloc_ppa, segkmem_free_ppa, heap_arena, ppaquantum,
1596 	    VM_SLEEP);
1597 
1598 	ASSERT(segkmem_ppa_arena != NULL);
1599 
1600 	/* prealloacate some memory for the lp kernel heap */
1601 	if (segkmem_kmemlp_min) {
1602 
1603 		ASSERT(P2PHASE(segkmem_kmemlp_min,
1604 		    segkmem_heaplp_quantum) == 0);
1605 
1606 		if ((addr = segkmem_alloc_lpi(heap_lp_arena,
1607 		    segkmem_kmemlp_min, VM_SLEEP)) != NULL) {
1608 
1609 			addr = vmem_add(kmem_lp_arena, addr,
1610 			    segkmem_kmemlp_min, VM_SLEEP);
1611 			ASSERT(addr != NULL);
1612 		}
1613 	}
1614 
1615 	lpcb->lp_uselp = 1;
1616 }
1617 
1618 #endif
1619