/ duct-tape / xnu / osfmk / kern / gzalloc.c
gzalloc.c
  1  /*
  2   * Copyright (c) 2000-2020 Apple Inc. All rights reserved.
  3   *
  4   * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
  5   *
  6   * This file contains Original Code and/or Modifications of Original Code
  7   * as defined in and that are subject to the Apple Public Source License
  8   * Version 2.0 (the 'License'). You may not use this file except in
  9   * compliance with the License. The rights granted to you under the License
 10   * may not be used to create, or enable the creation or redistribution of,
 11   * unlawful or unlicensed copies of an Apple operating system, or to
 12   * circumvent, violate, or enable the circumvention or violation of, any
 13   * terms of an Apple operating system software license agreement.
 14   *
 15   * Please obtain a copy of the License at
 16   * http://www.opensource.apple.com/apsl/ and read it before using this file.
 17   *
 18   * The Original Code and all software distributed under the License are
 19   * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 20   * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 21   * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 22   * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 23   * Please see the License for the specific language governing rights and
 24   * limitations under the License.
 25   *
 26   * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
 27   */
 28  /*
 29   *	File:	kern/gzalloc.c
 30   *	Author:	Derek Kumar
 31   *
 32   *	"Guard mode" zone allocator, used to trap use-after-free errors,
 33   *	overruns, underruns, mismatched allocations/frees, uninitialized
 34   *	zone element use, timing dependent races etc.
 35   *
 36   *	The allocator is configured by these boot-args:
 37   *	gzalloc_size=<size>: target all zones with elements of <size> bytes
 38   *	gzalloc_min=<size>: target zones with elements >= size
 39   *	gzalloc_max=<size>: target zones with elements <= size
 40   *      gzalloc_min/max can be specified in conjunction to target a range of
 41   *	sizes
 42   *	gzalloc_fc_size=<size>: number of zone elements (effectively page
 43   *	multiple sized) to retain in the free VA cache. This cache is evicted
 44   *	(backing pages and VA released) in a least-recently-freed fashion.
 45   *	Larger free VA caches allow for a longer window of opportunity to trap
 46   *	delayed use-after-free operations, but use more memory.
 47   *	-gzalloc_wp: Write protect, rather than unmap, freed allocations
 48   *	lingering in the free VA cache. Useful to disambiguate between
 49   *	read-after-frees/read overruns and writes. Also permits direct inspection
 50   *	of the freed element in the cache via the kernel debugger. As each
 51   *	element has a "header" (trailer in underflow detection mode), the zone
 52   *	of origin of the element can be easily determined in this mode.
 53   *	-gzalloc_uf_mode: Underflow detection mode, where the guard page
 54   *	adjoining each element is placed *before* the element page rather than
 55   *	after. The element is also located at the top of the page, rather than
 56   *	abutting the bottom as with the standard overflow detection mode.
 57   *	-gzalloc_noconsistency: disable consistency checks that flag mismatched
 58   *	frees, corruptions of the header/trailer signatures etc.
 59   *	-nogzalloc_mode: Disables the guard mode allocator. The DEBUG kernel
 60   *	enables the guard allocator for zones sized 1K (if present) by
 61   *	default, this option can disable that behaviour.
 62   *	gzname=<name> target a zone by name. Can be coupled with size-based
 63   *	targeting. Naming conventions match those of the zlog boot-arg, i.e.
 64   *	"a period in the logname will match a space in the zone name"
 65   *	-gzalloc_no_dfree_check Eliminate double free checks
 66   *	gzalloc_zscale=<value> specify size multiplier for the dedicated gzalloc submap
 67   */
 68  
 69  #include <mach/mach_types.h>
 70  #include <mach/vm_param.h>
 71  #include <mach/kern_return.h>
 72  #include <mach/machine/vm_types.h>
 73  #include <mach_debug/zone_info.h>
 74  #include <mach/vm_map.h>
 75  
 76  #include <kern/kern_types.h>
 77  #include <kern/assert.h>
 78  #include <kern/sched.h>
 79  #include <kern/locks.h>
 80  #include <kern/misc_protos.h>
 81  #include <kern/zalloc_internal.h>
 82  
 83  #include <vm/pmap.h>
 84  #include <vm/vm_map.h>
 85  #include <vm/vm_kern.h>
 86  #include <vm/vm_page.h>
 87  
 88  #include <pexpert/pexpert.h>
 89  
 90  #include <machine/machparam.h>
 91  
 92  #include <libkern/OSDebug.h>
 93  #include <libkern/OSAtomic.h>
 94  #include <sys/kdebug.h>
 95  
 96  boolean_t gzalloc_mode = FALSE;
 97  uint32_t pdzalloc_count, pdzfree_count;
 98  
 99  #define GZALLOC_MIN_DEFAULT (1024)
100  #define GZDEADZONE ((zone_t) 0xDEAD201E)
101  #define GZALLOC_SIGNATURE (0xABADCAFE)
102  #define GZALLOC_RESERVE_SIZE_DEFAULT (2 * 1024 * 1024)
103  #define GZFC_DEFAULT_SIZE (1536)
104  
105  char gzalloc_fill_pattern = 0x67; /* 'g' */
106  
107  uint32_t gzalloc_min = ~0U;
108  uint32_t gzalloc_max = 0;
109  uint32_t gzalloc_size = 0;
110  uint64_t gzalloc_allocated, gzalloc_freed, gzalloc_early_alloc, gzalloc_early_free, gzalloc_wasted;
111  boolean_t gzalloc_uf_mode = FALSE, gzalloc_consistency_checks = TRUE, gzalloc_dfree_check = TRUE;
112  vm_prot_t gzalloc_prot = VM_PROT_NONE;
113  uint32_t gzalloc_guard = KMA_GUARD_LAST;
114  uint32_t gzfc_size = GZFC_DEFAULT_SIZE;
115  uint32_t gzalloc_zonemap_scale = 6;
116  
117  vm_map_t gzalloc_map;
118  vm_offset_t gzalloc_map_min, gzalloc_map_max;
119  vm_offset_t gzalloc_reserve;
120  vm_size_t gzalloc_reserve_size;
121  
122  typedef struct gzalloc_header {
123  	zone_t gzone;
124  	uint32_t  gzsize;
125  	uint32_t  gzsig;
126  } gzhdr_t;
127  
128  #define GZHEADER_SIZE (sizeof(gzhdr_t))
129  
130  extern zone_t vm_page_zone;
131  
132  static zone_t gztrackzone = NULL;
133  static char gznamedzone[MAX_ZONE_NAME] = "";
134  
135  boolean_t
136  gzalloc_enabled(void)
137  {
138  	return gzalloc_mode;
139  }
140  
141  void
142  gzalloc_zone_init(zone_t z)
143  {
144  	if (gzalloc_mode == 0) {
145  		return;
146  	}
147  
148  	bzero(&z->gz, sizeof(z->gz));
149  
150  	if (track_this_zone(z->z_name, gznamedzone)) {
151  		gztrackzone = z;
152  	}
153  
154  	if (!z->gzalloc_exempt) {
155  		z->gzalloc_tracked = (z == gztrackzone) ||
156  		    ((zone_elem_size(z) >= gzalloc_min) && (zone_elem_size(z) <= gzalloc_max));
157  	}
158  
159  	if (gzfc_size && z->gzalloc_tracked) {
160  		vm_size_t gzfcsz = round_page(sizeof(*z->gz.gzfc) * gzfc_size);
161  		kern_return_t kr;
162  
163  		/* If the VM/kmem system aren't yet configured, carve
164  		 * out the free element cache structure directly from the
165  		 * gzalloc_reserve supplied by the pmap layer.
166  		 */
167  		if (__improbable(startup_phase < STARTUP_SUB_KMEM)) {
168  			if (gzalloc_reserve_size < gzfcsz) {
169  				panic("gzalloc reserve exhausted");
170  			}
171  
172  			z->gz.gzfc = (vm_offset_t *)gzalloc_reserve;
173  			gzalloc_reserve += gzfcsz;
174  			gzalloc_reserve_size -= gzfcsz;
175  			bzero(z->gz.gzfc, gzfcsz);
176  		} else {
177  			kr = kernel_memory_allocate(kernel_map,
178  			    (vm_offset_t *)&z->gz.gzfc, gzfcsz, 0,
179  			    KMA_KOBJECT | KMA_ZERO, VM_KERN_MEMORY_OSFMK);
180  			if (kr != KERN_SUCCESS) {
181  				panic("%s: kernel_memory_allocate failed (%d) for 0x%lx bytes",
182  				    __func__, kr, (unsigned long)gzfcsz);
183  			}
184  		}
185  	}
186  }
187  
188  /* Called by zdestroy() to dump the free cache elements so the zone count can drop to zero. */
189  void
190  gzalloc_empty_free_cache(zone_t zone)
191  {
192  	kern_return_t kr;
193  	int freed_elements = 0;
194  	vm_offset_t free_addr = 0;
195  	vm_offset_t rounded_size = round_page(zone_elem_size(zone) + GZHEADER_SIZE);
196  	vm_offset_t gzfcsz = round_page(sizeof(*zone->gz.gzfc) * gzfc_size);
197  	vm_offset_t gzfc_copy;
198  
199  	assert(zone->gzalloc_tracked); // the caller is responsible for checking
200  
201  	kr = kmem_alloc(kernel_map, &gzfc_copy, gzfcsz, VM_KERN_MEMORY_OSFMK);
202  	if (kr != KERN_SUCCESS) {
203  		panic("gzalloc_empty_free_cache: kmem_alloc: 0x%x", kr);
204  	}
205  
206  	/* Reset gzalloc_data. */
207  	zone_lock(zone);
208  	memcpy((void *)gzfc_copy, (void *)zone->gz.gzfc, gzfcsz);
209  	bzero((void *)zone->gz.gzfc, gzfcsz);
210  	zone->gz.gzfc_index = 0;
211  	zone_unlock(zone);
212  
213  	/* Free up all the cached elements. */
214  	for (uint32_t index = 0; index < gzfc_size; index++) {
215  		free_addr = ((vm_offset_t *)gzfc_copy)[index];
216  		if (free_addr && free_addr >= gzalloc_map_min && free_addr < gzalloc_map_max) {
217  			kr = vm_map_remove(gzalloc_map, free_addr,
218  			    free_addr + rounded_size + (1 * PAGE_SIZE),
219  			    VM_MAP_REMOVE_KUNWIRE);
220  			if (kr != KERN_SUCCESS) {
221  				panic("gzalloc_empty_free_cache: vm_map_remove: %p, 0x%x", (void *)free_addr, kr);
222  			}
223  			OSAddAtomic64((SInt32)rounded_size, &gzalloc_freed);
224  			OSAddAtomic64(-((SInt32) (rounded_size - zone_elem_size(zone))), &gzalloc_wasted);
225  
226  			freed_elements++;
227  		}
228  	}
229  	/*
230  	 * TODO: Consider freeing up zone->gz.gzfc as well if it didn't come from the gzalloc_reserve pool.
231  	 * For now we're reusing this buffer across zdestroy's. We would have to allocate it again on a
232  	 * subsequent zinit() as well.
233  	 */
234  
235  	/* Decrement zone counters. */
236  	zone_lock(zone);
237  	zone->z_elems_free += freed_elements;
238  	zone->z_wired_cur -= freed_elements;
239  	zone_unlock(zone);
240  
241  	kmem_free(kernel_map, gzfc_copy, gzfcsz);
242  }
243  
244  __startup_func
245  static void
246  gzalloc_configure(void)
247  {
248  #if !KASAN_ZALLOC
249  	char temp_buf[16];
250  
251  	if (PE_parse_boot_argn("-gzalloc_mode", temp_buf, sizeof(temp_buf))) {
252  		gzalloc_mode = TRUE;
253  		gzalloc_min = GZALLOC_MIN_DEFAULT;
254  		gzalloc_max = ~0U;
255  	}
256  
257  	if (PE_parse_boot_argn("gzalloc_min", &gzalloc_min, sizeof(gzalloc_min))) {
258  		gzalloc_mode = TRUE;
259  		gzalloc_max = ~0U;
260  	}
261  
262  	if (PE_parse_boot_argn("gzalloc_max", &gzalloc_max, sizeof(gzalloc_max))) {
263  		gzalloc_mode = TRUE;
264  		if (gzalloc_min == ~0U) {
265  			gzalloc_min = 0;
266  		}
267  	}
268  
269  	if (PE_parse_boot_argn("gzalloc_size", &gzalloc_size, sizeof(gzalloc_size))) {
270  		gzalloc_min = gzalloc_max = gzalloc_size;
271  		gzalloc_mode = TRUE;
272  	}
273  
274  	(void)PE_parse_boot_argn("gzalloc_fc_size", &gzfc_size, sizeof(gzfc_size));
275  
276  	if (PE_parse_boot_argn("-gzalloc_wp", temp_buf, sizeof(temp_buf))) {
277  		gzalloc_prot = VM_PROT_READ;
278  	}
279  
280  	if (PE_parse_boot_argn("-gzalloc_uf_mode", temp_buf, sizeof(temp_buf))) {
281  		gzalloc_uf_mode = TRUE;
282  		gzalloc_guard = KMA_GUARD_FIRST;
283  	}
284  
285  	if (PE_parse_boot_argn("-gzalloc_no_dfree_check", temp_buf, sizeof(temp_buf))) {
286  		gzalloc_dfree_check = FALSE;
287  	}
288  
289  	(void) PE_parse_boot_argn("gzalloc_zscale", &gzalloc_zonemap_scale, sizeof(gzalloc_zonemap_scale));
290  
291  	if (PE_parse_boot_argn("-gzalloc_noconsistency", temp_buf, sizeof(temp_buf))) {
292  		gzalloc_consistency_checks = FALSE;
293  	}
294  
295  	if (PE_parse_boot_argn("gzname", gznamedzone, sizeof(gznamedzone))) {
296  		gzalloc_mode = TRUE;
297  	}
298  #if DEBUG
299  	if (gzalloc_mode == FALSE) {
300  		gzalloc_min = 1024;
301  		gzalloc_max = 1024;
302  		strlcpy(gznamedzone, "pmap", sizeof(gznamedzone));
303  		gzalloc_prot = VM_PROT_READ;
304  		gzalloc_mode = TRUE;
305  	}
306  #endif
307  	if (PE_parse_boot_argn("-nogzalloc_mode", temp_buf, sizeof(temp_buf))) {
308  		gzalloc_mode = FALSE;
309  	}
310  
311  	if (gzalloc_mode) {
312  		gzalloc_reserve_size = GZALLOC_RESERVE_SIZE_DEFAULT;
313  		gzalloc_reserve = (vm_offset_t) pmap_steal_memory(gzalloc_reserve_size);
314  	}
315  #endif
316  }
317  STARTUP(PMAP_STEAL, STARTUP_RANK_FIRST, gzalloc_configure);
318  
319  void
320  gzalloc_init(vm_size_t max_zonemap_size)
321  {
322  	kern_return_t retval;
323  
324  	if (gzalloc_mode) {
325  		vm_map_kernel_flags_t vmk_flags;
326  
327  		vmk_flags = VM_MAP_KERNEL_FLAGS_NONE;
328  		vmk_flags.vmkf_permanent = TRUE;
329  		retval = kmem_suballoc(kernel_map, &gzalloc_map_min, (max_zonemap_size * gzalloc_zonemap_scale),
330  		    FALSE, VM_FLAGS_ANYWHERE, vmk_flags, VM_KERN_MEMORY_ZONE,
331  		    &gzalloc_map);
332  
333  		if (retval != KERN_SUCCESS) {
334  			panic("zone_init: kmem_suballoc(gzalloc_map, 0x%lx, %u) failed",
335  			    max_zonemap_size, gzalloc_zonemap_scale);
336  		}
337  		gzalloc_map_max = gzalloc_map_min + (max_zonemap_size * gzalloc_zonemap_scale);
338  	}
339  }
340  
341  vm_offset_t
342  gzalloc_alloc(zone_t zone, zone_stats_t zstats, zalloc_flags_t flags)
343  {
344  	vm_offset_t addr = 0;
345  
346  	assert(zone->gzalloc_tracked); // the caller is responsible for checking
347  
348  	if (get_preemption_level() != 0) {
349  		if (flags & Z_NOWAIT) {
350  			return 0;
351  		}
352  		pdzalloc_count++;
353  	}
354  
355  	bool kmem_ready = (startup_phase >= STARTUP_SUB_KMEM);
356  	vm_offset_t rounded_size = round_page(zone_elem_size(zone) + GZHEADER_SIZE);
357  	vm_offset_t residue = rounded_size - zone_elem_size(zone);
358  	vm_offset_t gzaddr = 0;
359  	gzhdr_t *gzh, *gzhcopy = NULL;
360  	bool new_va = false;
361  
362  	if (!kmem_ready || (vm_page_zone == ZONE_NULL)) {
363  		/* Early allocations are supplied directly from the
364  		 * reserve.
365  		 */
366  		if (gzalloc_reserve_size < (rounded_size + PAGE_SIZE)) {
367  			panic("gzalloc reserve exhausted");
368  		}
369  		gzaddr = gzalloc_reserve;
370  		/* No guard page for these early allocations, just
371  		 * waste an additional page.
372  		 */
373  		gzalloc_reserve += rounded_size + PAGE_SIZE;
374  		gzalloc_reserve_size -= rounded_size + PAGE_SIZE;
375  		OSAddAtomic64((SInt32) (rounded_size), &gzalloc_early_alloc);
376  	} else {
377  		kern_return_t kr = kernel_memory_allocate(gzalloc_map,
378  		    &gzaddr, rounded_size + (1 * PAGE_SIZE),
379  		    0, KMA_KOBJECT | KMA_ATOMIC | gzalloc_guard,
380  		    VM_KERN_MEMORY_OSFMK);
381  		if (kr != KERN_SUCCESS) {
382  			panic("gzalloc: kernel_memory_allocate for size 0x%llx failed with %d",
383  			    (uint64_t)rounded_size, kr);
384  		}
385  		new_va = true;
386  	}
387  
388  	if (gzalloc_uf_mode) {
389  		gzaddr += PAGE_SIZE;
390  		/* The "header" becomes a "footer" in underflow
391  		 * mode.
392  		 */
393  		gzh = (gzhdr_t *) (gzaddr + zone_elem_size(zone));
394  		addr = gzaddr;
395  		gzhcopy = (gzhdr_t *) (gzaddr + rounded_size - sizeof(gzhdr_t));
396  	} else {
397  		gzh = (gzhdr_t *) (gzaddr + residue - GZHEADER_SIZE);
398  		addr = (gzaddr + residue);
399  	}
400  
401  	if (zone->z_free_zeroes) {
402  		bzero((void *)gzaddr, rounded_size);
403  	} else {
404  		/* Fill with a pattern on allocation to trap uninitialized
405  		 * data use. Since the element size may be "rounded up"
406  		 * by higher layers such as the kalloc layer, this may
407  		 * also identify overruns between the originally requested
408  		 * size and the rounded size via visual inspection.
409  		 * TBD: plumb through the originally requested size,
410  		 * prior to rounding by kalloc/IOMalloc etc.
411  		 * We also add a signature and the zone of origin in a header
412  		 * prefixed to the allocation.
413  		 */
414  		memset((void *)gzaddr, gzalloc_fill_pattern, rounded_size);
415  	}
416  
417  	gzh->gzone = (kmem_ready && vm_page_zone) ? zone : GZDEADZONE;
418  	gzh->gzsize = (uint32_t)zone_elem_size(zone);
419  	gzh->gzsig = GZALLOC_SIGNATURE;
420  
421  	/* In underflow detection mode, stash away a copy of the
422  	 * metadata at the edge of the allocated range, for
423  	 * retrieval by gzalloc_element_size()
424  	 */
425  	if (gzhcopy) {
426  		*gzhcopy = *gzh;
427  	}
428  
429  	zone_lock(zone);
430  	assert(zone->z_self == zone);
431  	zone->z_elems_free--;
432  	if (new_va) {
433  		zone->z_va_cur += 1;
434  	}
435  	zone->z_wired_cur += 1;
436  	zpercpu_get(zstats)->zs_mem_allocated += rounded_size;
437  	zone_unlock(zone);
438  
439  	OSAddAtomic64((SInt32) rounded_size, &gzalloc_allocated);
440  	OSAddAtomic64((SInt32) (rounded_size - zone_elem_size(zone)), &gzalloc_wasted);
441  
442  	return addr;
443  }
444  
445  void
446  gzalloc_free(zone_t zone, zone_stats_t zstats, void *addr)
447  {
448  	kern_return_t kr;
449  
450  	assert(zone->gzalloc_tracked); // the caller is responsible for checking
451  
452  	gzhdr_t *gzh;
453  	vm_offset_t rounded_size = round_page(zone_elem_size(zone) + GZHEADER_SIZE);
454  	vm_offset_t residue = rounded_size - zone_elem_size(zone);
455  	vm_offset_t saddr;
456  	vm_offset_t free_addr = 0;
457  
458  	if (gzalloc_uf_mode) {
459  		gzh = (gzhdr_t *)((vm_offset_t)addr + zone_elem_size(zone));
460  		saddr = (vm_offset_t) addr - PAGE_SIZE;
461  	} else {
462  		gzh = (gzhdr_t *)((vm_offset_t)addr - GZHEADER_SIZE);
463  		saddr = ((vm_offset_t)addr) - residue;
464  	}
465  
466  	if ((saddr & PAGE_MASK) != 0) {
467  		panic("%s: invalid address supplied: "
468  		    "%p (adjusted: 0x%lx) for zone with element sized 0x%lx\n",
469  		    __func__, addr, saddr, zone_elem_size(zone));
470  	}
471  
472  	if (gzfc_size && gzalloc_dfree_check) {
473  		zone_lock(zone);
474  		assert(zone->z_self == zone);
475  		for (uint32_t gd = 0; gd < gzfc_size; gd++) {
476  			if (zone->gz.gzfc[gd] != saddr) {
477  				continue;
478  			}
479  			panic("%s: double free detected, freed address: 0x%lx, "
480  			    "current free cache index: %d, freed index: %d",
481  			    __func__, saddr, zone->gz.gzfc_index, gd);
482  		}
483  		zone_unlock(zone);
484  	}
485  
486  	if (gzalloc_consistency_checks) {
487  		if (gzh->gzsig != GZALLOC_SIGNATURE) {
488  			panic("GZALLOC signature mismatch for element %p, "
489  			    "expected 0x%x, found 0x%x",
490  			    addr, GZALLOC_SIGNATURE, gzh->gzsig);
491  		}
492  
493  		if (gzh->gzone != zone && (gzh->gzone != GZDEADZONE)) {
494  			panic("%s: Mismatched zone or under/overflow, "
495  			    "current zone: %p, recorded zone: %p, address: %p",
496  			    __func__, zone, gzh->gzone, (void *)addr);
497  		}
498  		/* Partially redundant given the zone check, but may flag header corruption */
499  		if (gzh->gzsize != zone_elem_size(zone)) {
500  			panic("Mismatched zfree or under/overflow for zone %p, "
501  			    "recorded size: 0x%x, element size: 0x%x, address: %p",
502  			    zone, gzh->gzsize, (uint32_t)zone_elem_size(zone), (void *)addr);
503  		}
504  
505  		char *gzc, *checkstart, *checkend;
506  		if (gzalloc_uf_mode) {
507  			checkstart = (char *) ((uintptr_t) gzh + sizeof(gzh));
508  			checkend = (char *) ((((vm_offset_t)addr) & ~PAGE_MASK) + PAGE_SIZE);
509  		} else {
510  			checkstart = (char *) trunc_page_64(addr);
511  			checkend = (char *)gzh;
512  		}
513  
514  		for (gzc = checkstart; gzc < checkend; gzc++) {
515  			if (*gzc == gzalloc_fill_pattern) {
516  				continue;
517  			}
518  			panic("%s: detected over/underflow, byte at %p, element %p, "
519  			    "contents 0x%x from 0x%lx byte sized zone (%s%s) "
520  			    "doesn't match fill pattern (%c)",
521  			    __func__, gzc, addr, *gzc, zone_elem_size(zone),
522  			    zone_heap_name(zone), zone->z_name, gzalloc_fill_pattern);
523  		}
524  	}
525  
526  	if ((startup_phase < STARTUP_SUB_KMEM) || gzh->gzone == GZDEADZONE) {
527  		/* For now, just leak frees of early allocations
528  		 * performed before kmem is fully configured.
529  		 * They don't seem to get freed currently;
530  		 * consider ml_static_mfree in the future.
531  		 */
532  		OSAddAtomic64((SInt32) (rounded_size), &gzalloc_early_free);
533  		return;
534  	}
535  
536  	if (get_preemption_level() != 0) {
537  		pdzfree_count++;
538  	}
539  
540  	if (gzfc_size) {
541  		/* Either write protect or unmap the newly freed
542  		 * allocation
543  		 */
544  		kr = vm_map_protect(gzalloc_map, saddr,
545  		    saddr + rounded_size + (1 * PAGE_SIZE),
546  		    gzalloc_prot, FALSE);
547  		if (kr != KERN_SUCCESS) {
548  			panic("%s: vm_map_protect: %p, 0x%x", __func__, (void *)saddr, kr);
549  		}
550  	} else {
551  		free_addr = saddr;
552  	}
553  
554  	zone_lock(zone);
555  	assert(zone->z_self == zone);
556  
557  	/* Insert newly freed element into the protected free element
558  	 * cache, and rotate out the LRU element.
559  	 */
560  	if (gzfc_size) {
561  		if (zone->gz.gzfc_index >= gzfc_size) {
562  			zone->gz.gzfc_index = 0;
563  		}
564  		free_addr = zone->gz.gzfc[zone->gz.gzfc_index];
565  		zone->gz.gzfc[zone->gz.gzfc_index++] = saddr;
566  	}
567  
568  	if (free_addr) {
569  		zone->z_elems_free++;
570  		zone->z_wired_cur -= 1;
571  	}
572  
573  	zpercpu_get(zstats)->zs_mem_freed += rounded_size;
574  	zone_unlock(zone);
575  
576  	if (free_addr) {
577  		// TODO: consider using physical reads to check for
578  		// corruption while on the protected freelist
579  		// (i.e. physical corruption)
580  		kr = vm_map_remove(gzalloc_map, free_addr,
581  		    free_addr + rounded_size + (1 * PAGE_SIZE),
582  		    VM_MAP_REMOVE_KUNWIRE);
583  		if (kr != KERN_SUCCESS) {
584  			panic("gzfree: vm_map_remove: %p, 0x%x", (void *)free_addr, kr);
585  		}
586  		// TODO: sysctl-ize for quick reference
587  		OSAddAtomic64((SInt32)rounded_size, &gzalloc_freed);
588  		OSAddAtomic64(-((SInt32) (rounded_size - zone_elem_size(zone))),
589  		    &gzalloc_wasted);
590  	}
591  }
592  
593  boolean_t
594  gzalloc_element_size(void *gzaddr, zone_t *z, vm_size_t *gzsz)
595  {
596  	uintptr_t a = (uintptr_t)gzaddr;
597  	if (__improbable(gzalloc_mode && (a >= gzalloc_map_min) && (a < gzalloc_map_max))) {
598  		gzhdr_t *gzh;
599  		boolean_t       vmef;
600  		vm_map_entry_t  gzvme = NULL;
601  		vm_map_lock_read(gzalloc_map);
602  		vmef = vm_map_lookup_entry(gzalloc_map, (vm_map_offset_t)a, &gzvme);
603  		vm_map_unlock(gzalloc_map);
604  		if (vmef == FALSE) {
605  			panic("GZALLOC: unable to locate map entry for %p\n", (void *)a);
606  		}
607  		assertf(gzvme->vme_atomic != 0, "GZALLOC: VM map entry inconsistency, "
608  		    "vme: %p, start: %llu end: %llu", gzvme, gzvme->vme_start, gzvme->vme_end);
609  
610  		/* Locate the gzalloc metadata adjoining the element */
611  		if (gzalloc_uf_mode == TRUE) {
612  			/* In underflow detection mode, locate the map entry describing
613  			 * the element, and then locate the copy of the gzalloc
614  			 * header at the trailing edge of the range.
615  			 */
616  			gzh = (gzhdr_t *)(gzvme->vme_end - GZHEADER_SIZE);
617  		} else {
618  			/* In overflow detection mode, scan forward from
619  			 * the base of the map entry to locate the
620  			 * gzalloc header.
621  			 */
622  			uint32_t *p = (uint32_t*) gzvme->vme_start;
623  			while (p < (uint32_t *) gzvme->vme_end) {
624  				if (*p == GZALLOC_SIGNATURE) {
625  					break;
626  				} else {
627  					p++;
628  				}
629  			}
630  			if (p >= (uint32_t *) gzvme->vme_end) {
631  				panic("GZALLOC signature missing addr %p, zone %p", gzaddr, z);
632  			}
633  			p++;
634  			uintptr_t q = (uintptr_t) p;
635  			gzh = (gzhdr_t *) (q - sizeof(gzhdr_t));
636  		}
637  
638  		if (gzh->gzsig != GZALLOC_SIGNATURE) {
639  			panic("GZALLOC signature mismatch for element %p, expected 0x%x, found 0x%x",
640  			    (void *)a, GZALLOC_SIGNATURE, gzh->gzsig);
641  		}
642  
643  		*gzsz = zone_elem_size(gzh->gzone);
644  		if (__improbable(!gzh->gzone->gzalloc_tracked)) {
645  			panic("GZALLOC: zone mismatch (%p)\n", gzh->gzone);
646  		}
647  
648  		if (z) {
649  			*z = gzh->gzone;
650  		}
651  		return TRUE;
652  	} else {
653  		return FALSE;
654  	}
655  }