blob: 330b674ea73bbb585a15c454c1b3a6f2a66b26d4 [file] [log] [blame]
/* **********************************************************
* Copyright (c) 2010-2013 Google, Inc. All rights reserved.
* Copyright (c) 2001-2010 VMware, Inc. All rights reserved.
* **********************************************************/
/*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* * Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* * Redistributions in binary form must reproduce the above copyright notice,
* this list of conditions and the following disclaimer in the documentation
* and/or other materials provided with the distribution.
*
* * Neither the name of VMware, Inc. nor the names of its contributors may be
* used to endorse or promote products derived from this software without
* specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL VMWARE, INC. OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
* DAMAGE.
*/
/* Copyright (c) 2003-2007 Determina Corp. */
/* Copyright (c) 2001-2003 Massachusetts Institute of Technology */
/* Copyright (c) 2001 Hewlett-Packard Company */
/*
* heap.c - heap manager
*/
#include "globals.h"
#include <string.h> /* for memcpy */
#include <limits.h>
#include "fragment.h" /* for struct sizes */
#include "link.h" /* for struct sizes */
#include "instr.h" /* for struct sizes */
#include "fcache.h" /* fcache_low_on_memory */
#ifdef DEBUG
# include "hotpatch.h" /* To handle leak for case 9593. */
#endif
#ifdef HEAP_ACCOUNTING
# ifndef DEBUG
# error HEAP_ACCOUNTING requires DEBUG
# endif
#endif
#ifdef DEBUG_MEMORY
/* on by default but higher than general asserts */
# define CHKLVL_MEMFILL CHKLVL_DEFAULT
#endif
extern bool vm_areas_exited;
/***************************************************************************
* we cannot use malloc in the middle of interpreting the client program
* because we could be in the middle of interpreting malloc, which is not
* always reentrant
*
* We have a virtual memory manager which makes sure memory is
* reserved within the application address space so that we don't have
* to fight with the application. We call os_heap_reserve to allocate
* virtual space in a single consecutive region. We later use
* os_heap_commit to get committed memory in large chunks and manage
* the chunks using a simple scheme of free lists of different sizes.
* The virtual memory manager has to store out of band information about
* used and free blocks, since of course there is no real memory to use.
* The chunks (heap units) store in band extra information both for
* used and free. However, in the allocated blocks within a unit we
* don't need to store any information since heap_free passes in the
* size; we store the next pointers for the free lists at the start of
* the free blocks themselves. We have one large reservation for most of
* our allocations, and yet another for allocations that we do not
* plan on ever freeing up on detach - the only unavoidable tombstones
* are those for thread private code system calls that may be stuck on
* callbacks. In case we run out of reserved memory we do fall back
* on requests from the OS, but any of these may fail if we are
* competing with the application.
*
* looking at dynamo behavior as of Jan 2001, most heap_alloc requests are
* for < 128 bytes, very few for larger, so we have a bunch of fixed-size
* blocks of small sizes
*
* the UINT_MAX size is a variable-length block, we keep one byte to store
* the size (again storing the next pointer when free at the start of
* what we pass to the user)
*/
static const uint BLOCK_SIZES[] = {
8, /* for instr bits */
#ifndef X64
/* for x64 future_fragment_t is 24 bytes (could be 20 if we could put flags last) */
sizeof(future_fragment_t), /* 12 (24 x64) */
#endif
/* we have a lot of size 16 requests for IR but they are transient */
24, /* fcache empties and vm_area_t are now 20, vm area extras still 24 */
ALIGN_FORWARD(sizeof(fragment_t) + sizeof(indirect_linkstub_t), HEAP_ALIGNMENT), /* 40 dbg / 36 rel */
#if defined(X64) || defined(PROFILE_LINKCOUNT) || defined(CUSTOM_EXIT_STUBS)
sizeof(instr_t), /* 64 (104 x64) */
sizeof(fragment_t) + sizeof(direct_linkstub_t)
+ sizeof(cbr_fallthrough_linkstub_t), /* 68 dbg / 64 rel, 112 x64 */
/* all other bb/trace buckets are 8 larger but in same order */
#else
sizeof(fragment_t) + sizeof(direct_linkstub_t)
+ sizeof(cbr_fallthrough_linkstub_t), /* 60 dbg / 56 rel */
sizeof(instr_t), /* 64 */
#endif
/* we keep this bucket even though only 10% or so of normal bbs
* hit this.
* FIXME: release == instr_t here so a small waste when walking buckets
*/
ALIGN_FORWARD(sizeof(fragment_t) + 2*sizeof(direct_linkstub_t),
HEAP_ALIGNMENT), /* 68 dbg / 64 rel (128 x64) */
ALIGN_FORWARD(sizeof(trace_t) + 2*sizeof(direct_linkstub_t) + sizeof(uint),
HEAP_ALIGNMENT), /* 80 dbg / 76 rel (148 x64 => 152) */
/* FIXME: measure whether should put in indirect mixes as well */
ALIGN_FORWARD(sizeof(trace_t) + 3*sizeof(direct_linkstub_t) + sizeof(uint),
HEAP_ALIGNMENT), /* 96 dbg / 92 rel (180 x64 => 184) */
ALIGN_FORWARD(sizeof(trace_t) + 5*sizeof(direct_linkstub_t) + sizeof(uint),
HEAP_ALIGNMENT), /* 128 dbg / 124 rel (244 x64 => 248) */
256,
512,
UINT_MAX /* variable-length */
};
#define BLOCK_TYPES (sizeof(BLOCK_SIZES)/sizeof(uint))
#ifdef DEBUG
/* FIXME: would be nice to have these stats per HEAPACCT category */
/* These are ints only b/c we used to do non-atomic adds and wanted to
* gracefully handle underflow to negative values
*/
DECLARE_NEVERPROT_VAR(static int block_total_count[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static int block_count[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static int block_peak_count[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static int block_wasted[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static int block_peak_wasted[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static int block_align_pad[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static int block_peak_align_pad[BLOCK_TYPES], {0});
DECLARE_NEVERPROT_VAR(static bool out_of_vmheap_once, false);
#endif
/* variable-length: we steal one int for the size */
#define HEADER_SIZE (sizeof(size_t))
/* VARIABLE_SIZE is assignable */
#define VARIABLE_SIZE(p) (*(size_t *)((p)-HEADER_SIZE))
#define MEMSET_HEADER(p, value) VARIABLE_SIZE(p) = HEAP_TO_PTR_UINT(value)
#define GET_VARIABLE_ALLOCATION_SIZE(p) (VARIABLE_SIZE(p) + HEADER_SIZE)
/* heap is allocated in units
* we start out with a small unit, then each additional unit we
* need doubles in size, up to a maximum, we default to 32kb initial size
* (24kb useful with guard pages), max size defaults to 64kb (56kb useful with
* guard pages), we keep the max small to save memory, it doesn't seem to be
* perf hit! Though with guard pages we are wasting quite a bit of reserved
* (though not committed) space */
/* the only big things global heap is used for are pc sampling
* hash table and sideline sampling hash table -- if none of those
* are in use, 16KB should be plenty, we default to 32kb since guard
* pages are on by default (gives 24kb useful) max size is same as for
* normal heap units.
*/
/* the old defaults were 32kb (usable) for initial thread private and 16kb
* (usable) for initial global, changed to simplify the logic for allocating
* in multiples of the os allocation granularity. The new defaults prob.
* make more sense with the shared cache then the old did anyways.
*/
/* restrictions -
* any guard pages are included in the size, size must be > UNITOVERHEAD
* for best performance sizes should be of the form
* 2^n * page_size (where n is a positve integer) and max should be a multiple
* of the os allocation granularity so that with enough doublings we are
* reserving memory in multiples of the allocation granularity and not wasting
* any virtual address space (beyond our guard pages)
*/
#define HEAP_UNIT_MIN_SIZE INTERNAL_OPTION(initial_heap_unit_size)
#define HEAP_UNIT_MAX_SIZE INTERNAL_OPTION(max_heap_unit_size)
#define GLOBAL_UNIT_MIN_SIZE INTERNAL_OPTION(initial_global_heap_unit_size)
#define GUARD_PAGE_ADJUSTMENT (dynamo_options.guard_pages ? 2 * PAGE_SIZE : 0)
/* gets usable space in the unit */
#define UNITROOM(u) ((size_t) (u->end_pc - u->start_pc))
#define UNIT_RESERVED_ROOM(u) (u->reserved_end_pc - u->start_pc)
/* we keep the heap_unit_t header at top of the unit, this macro calculates
* the committed size of the unit by adding header size to available size
*/
#define UNIT_COMMIT_SIZE(u) (UNITROOM(u) + sizeof(heap_unit_t))
#define UNIT_RESERVED_SIZE(u) (UNIT_RESERVED_ROOM(u) + sizeof(heap_unit_t))
#define UNIT_ALLOC_START(u) (u->start_pc - sizeof(heap_unit_t))
#define UNIT_GET_START_PC(u) (byte*)(((ptr_uint_t)u) + sizeof(heap_unit_t))
#define UNIT_COMMIT_END(u) (u->end_pc)
#define UNIT_RESERVED_END(u) (u->reserved_end_pc)
/* gets the allocated size of the unit (reserved size + guard pages) */
#define UNITALLOC(u) (UNIT_RESERVED_SIZE(u) + GUARD_PAGE_ADJUSTMENT)
/* gets unit overhead, includes the reserved (guard pages) and committed
* (sizeof(heap_unit_t)) portions
*/
#define UNITOVERHEAD (sizeof(heap_unit_t) + GUARD_PAGE_ADJUSTMENT)
/* any alloc request larger than this needs a special unit */
#define MAXROOM (HEAP_UNIT_MAX_SIZE - UNITOVERHEAD)
/* maximum valid allocation (to guard against internal integer overflows) */
#define MAX_VALID_HEAP_ALLOCATION INT_MAX
/* thread-local heap structure
* this struct is kept at top of unit itself, not in separate allocation
*/
typedef struct _heap_unit_t {
heap_pc start_pc; /* start address of heap storage */
heap_pc end_pc; /* open-ended end address of heap storage */
heap_pc cur_pc; /* open-ended current end of allocated storage */
heap_pc reserved_end_pc; /* open-ended end of reserved (not nec committed) memory */
bool in_vmarea_list; /* perf opt for delayed batch vmarea updating */
#ifdef DEBUG
int id; /* # of this unit */
#endif
struct _heap_unit_t *next_local; /* used to link thread's units */
struct _heap_unit_t *next_global; /* used to link all units */
struct _heap_unit_t *prev_global; /* used to link all units */
} heap_unit_t;
#ifdef HEAP_ACCOUNTING
typedef struct _heap_acct_t {
size_t alloc_reuse[ACCT_LAST];
size_t alloc_new[ACCT_LAST];
size_t cur_usage[ACCT_LAST];
size_t max_usage[ACCT_LAST];
size_t max_single[ACCT_LAST];
uint num_alloc[ACCT_LAST];
} heap_acct_t;
#endif
/* FIXME (case 6336): rename to heap_t:
* a heap_t is a collection of units with the same properties
* to reflect that this is used for more than just thread-private memory.
* Also rename the "tu" vars to "h"
*/
typedef struct _thread_units_t {
heap_unit_t *top_unit; /* start of linked list of heap units */
heap_unit_t *cur_unit; /* current unit in heap list */
heap_pc free_list[BLOCK_TYPES];
#ifdef DEBUG
int num_units; /* total # of heap units */
#endif
dcontext_t *dcontext; /* back pointer to owner */
bool writable; /* remember state of heap protection */
#ifdef HEAP_ACCOUNTING
heap_acct_t acct;
#endif
} thread_units_t;
/* per-thread structure: */
typedef struct _thread_heap_t {
thread_units_t *local_heap;
thread_units_t *nonpersistent_heap;
} thread_heap_t;
/* global, unique thread-shared structure:
* FIXME: give this name to thread_units_t, and name this AllHeapUnits
*/
typedef struct _heap_t {
heap_unit_t *units; /* list of all allocated units */
heap_unit_t *dead; /* list of deleted units ready for re-allocation */
/* FIXME: num_dead duplicates stats->heap_num_free, but we want num_dead
* for release build too, so it's separate...can we do better?
*/
uint num_dead;
} heap_t;
/* no synch needed since only written once */
static bool heap_exiting = false;
#ifdef DEBUG
DECLARE_NEVERPROT_VAR(static bool ever_beyond_vmm, false);
#endif
/* Lock used only for managing heap units, not for normal thread-local alloc.
* Must be recursive due to circular dependencies between vmareas and global heap.
* Furthermore, always grab dynamo_vm_areas_lock() before grabbing this lock,
* to make DR areas update and heap alloc/free atomic!
*/
DECLARE_CXTSWPROT_VAR(static recursive_lock_t heap_unit_lock,
INIT_RECURSIVE_LOCK(heap_unit_lock));
/* N.B.: if these two locks are ever owned at the same time, the convention is
* that global_alloc_lock MUST be grabbed first, to avoid deadlocks
*/
/* separate lock for global heap access to avoid contention between local unit
* creation and global heap alloc
* must be recursive so that heap_vmareas_synch_units can hold it and heap_unit_lock
* up front to avoid deadlocks, and still allow vmareas to global_alloc --
* BUT we do NOT want global_alloc() to be able to recurse!
* FIXME: either find a better solution to the heap_vmareas_synch_units deadlock
* that is as efficient, or find a way to assert that the only recursion is
* from heap_vmareas_synch_units to global_alloc
*/
DECLARE_CXTSWPROT_VAR(static recursive_lock_t global_alloc_lock,
INIT_RECURSIVE_LOCK(global_alloc_lock));
#if defined(DEBUG) && defined(HEAP_ACCOUNTING) && defined(HOT_PATCHING_INTERFACE)
static int get_special_heap_header_size(void);
#endif
vm_area_vector_t *landing_pad_areas; /* PR 250294 */
#ifdef WINDOWS
/* i#939: we steal space from ntdll's +rx segment */
static app_pc lpad_temp_writable_start;
static size_t lpad_temp_writable_size;
static void release_landing_pad_mem(void);
#endif
/* Indicates whether should back out of a global alloc/free and grab the
* DR areas lock first, to retry
*/
static bool
safe_to_allocate_or_free_heap_units()
{
return ((!self_owns_recursive_lock(&global_alloc_lock) &&
!self_owns_recursive_lock(&heap_unit_lock)) ||
self_owns_dynamo_vm_area_lock());
}
/* indicates a dynamo vm area remove was delayed
* protected by the heap_unit_lock
*/
DECLARE_FREQPROT_VAR(static bool dynamo_areas_pending_remove, false);
#ifdef HEAP_ACCOUNTING
const char * whichheap_name[] = {
/* max length for aligned output is length of "BB Fragments" */
"BB Fragments",
"Coarse Links",
"Future Frag",
"Frag Tables",
"IBL Tables",
"Traces",
"FC Empties",
"Vm Multis",
"IR",
"RCT Tables",
"VM Areas",
"Symbols",
# ifdef SIDELINE
"Sideline",
# endif
"TH Counter",
"Tombstone",
"Hot Patching",
"Thread Mgt",
"Memory Mgt",
"Stats",
"SpecialHeap",
# ifdef CLIENT_INTERFACE
"Client",
# endif
"Lib Dup",
"Clean Call",
/* NOTE: Add your heap name here */
"Other",
};
/* Since using a lock for these stats adds a lot of contention, we
* follow a two-pronged strategy:
* 1) For accurate stats we add a thread's final stats to the global only
* when it is cleaned up. But, this prevents global stats from being
* available in the middle of a run or if a run is not cleaned up nicely.
* 2) We have a set of heap_accounting stats for incremental global stats
* that are available at any time, yet racy and so may be off a little.
*/
/* all set to 0 is only initialization we need */
DECLARE_NEVERPROT_VAR(static thread_units_t global_racy_units, {0});
/* macro to get the type abstracted */
# define ACCOUNT_FOR_ALLOC_HELPER(type, tu, which, alloc_sz, ask_sz) do { \
(tu)->acct.type[which] += alloc_sz; \
(tu)->acct.num_alloc[which]++; \
(tu)->acct.cur_usage[which] += alloc_sz; \
if ((tu)->acct.cur_usage[which] > (tu)->acct.max_usage[which]) \
(tu)->acct.max_usage[which] = (tu)->acct.cur_usage[which]; \
if (ask_sz > (tu)->acct.max_single[which]) \
(tu)->acct.max_single[which] = ask_sz; \
} while (0)
# define ACCOUNT_FOR_ALLOC(type, tu, which, alloc_sz, ask_sz) do { \
STATS_ADD_PEAK(heap_claimed, alloc_sz); \
ACCOUNT_FOR_ALLOC_HELPER(type, tu, which, alloc_sz, ask_sz); \
ACCOUNT_FOR_ALLOC_HELPER(type, &global_racy_units, which, \
alloc_sz, ask_sz); \
} while (0)
# define ACCOUNT_FOR_FREE(tu, which, size) do { \
STATS_SUB(heap_claimed, (size)); \
(tu)->acct.cur_usage[which] -= size; \
global_racy_units.acct.cur_usage[which] -= size; \
} while (0)
#else
# define ACCOUNT_FOR_ALLOC(type, tu, which, alloc_sz, ask_sz)
# define ACCOUNT_FOR_FREE(tu, which, size)
#endif
typedef byte *vm_addr_t;
#ifdef X64
/* designates the closed interval within which we must allocate DR heap space */
static byte *heap_allowable_region_start = (byte *)PTR_UINT_0;
static byte *heap_allowable_region_end = (byte *)POINTER_MAX;
/* used only to protect read/write access to the must_reach_* static variables in
* request_region_be_heap_reachable() */
DECLARE_CXTSWPROT_VAR(static mutex_t request_region_be_heap_reachable_lock,
INIT_LOCK_FREE(request_region_be_heap_reachable_lock));
/* Request that the supplied region be 32bit offset reachable from the DR heap. Should
* be called before vmm_heap_init() so we can place the DR heap to meet these constraints.
* Can also be called post vmm_heap_init() but at that point acts as an assert that the
* supplied region is reachable since the heap is already reserved.
*
* Must be called at least once up front, for the -heap_in_lower_4GB code here
* to kick in!
*/
void
request_region_be_heap_reachable(byte *start, size_t size)
{
/* initialize so will be overridden on first call; protected by the
* request_region_be_heap_reachable_lock */
static byte *must_reach_region_start = (byte *)POINTER_MAX;
static byte *must_reach_region_end = (byte *)PTR_UINT_0; /* closed */
LOG(GLOBAL, LOG_HEAP, 2,
"Adding must-be-reachable-from-heap region "PFX"-"PFX"\n"
"Existing must-be-reachable region "PFX"-"PFX"\n"
"Existing allowed range "PFX"-"PFX"\n",
start, start+size, must_reach_region_start, must_reach_region_end,
heap_allowable_region_start, heap_allowable_region_end);
ASSERT(!POINTER_OVERFLOW_ON_ADD(start, size));
ASSERT(size > 0);
mutex_lock(&request_region_be_heap_reachable_lock);
if (start < must_reach_region_start) {
byte *allowable_end_tmp;
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
must_reach_region_start = start;
allowable_end_tmp = REACHABLE_32BIT_END(must_reach_region_start,
must_reach_region_end);
/* PR 215395 - add in absolute address reachability */
if (DYNAMO_OPTION(heap_in_lower_4GB) &&
allowable_end_tmp > ( byte *)POINTER_MAX_32BIT) {
allowable_end_tmp = (byte *)POINTER_MAX_32BIT;
}
/* Write assumed to be atomic so we don't have to hold a lock to use
* heap_allowable_region_end. */
heap_allowable_region_end = allowable_end_tmp;
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
}
if (start + size - 1 > must_reach_region_end) {
SELF_UNPROTECT_DATASEC(DATASEC_RARELY_PROT);
must_reach_region_end = start + size - 1; /* closed */
/* Write assumed to be atomic so we don't have to hold a lock to use
* heap_allowable_region_start. */
heap_allowable_region_start = REACHABLE_32BIT_START(must_reach_region_start,
must_reach_region_end);
SELF_PROTECT_DATASEC(DATASEC_RARELY_PROT);
}
ASSERT(must_reach_region_start <= must_reach_region_end); /* correctness check */
/* verify can be addressed absolutely (if required), correctness check */
ASSERT(!DYNAMO_OPTION(heap_in_lower_4GB) ||
heap_allowable_region_end <= (byte *)POINTER_MAX_32BIT);
mutex_unlock(&request_region_be_heap_reachable_lock);
LOG(GLOBAL, LOG_HEAP, 1,
"Added must-be-reachable-from-heap region "PFX"-"PFX"\n"
"New must-be-reachable region "PFX"-"PFX"\n"
"New allowed range "PFX"-"PFX"\n",
start, start+size, must_reach_region_start, must_reach_region_end,
heap_allowable_region_start, heap_allowable_region_end);
/* Reachability checks (xref PR 215395, note since we currently can't directly
* control where DR/client dlls are loaded these could fire if rebased). */
ASSERT(heap_allowable_region_start <= must_reach_region_start &&
"x64 reachability contraints not satisfiable");
ASSERT(must_reach_region_end <= heap_allowable_region_end &&
"x64 reachability contraints not satisfiable");
/* Handle release build failure. */
if (heap_allowable_region_start > must_reach_region_start ||
must_reach_region_end > heap_allowable_region_end) {
/* FIXME - in a released product we may want to detach or something else less
* drastic than triggering a FATAL_USAGE_ERROR. */
FATAL_USAGE_ERROR(HEAP_CONTRAINTS_UNSATISFIABLE, 2,
get_application_name(), get_application_pid());
}
}
void
vmcode_get_reachable_region(byte **region_start OUT, byte **region_end OUT)
{
/* We track sub-page for more accuracy on additional constraints, and
* align when asked about it.
*/
if (region_start != NULL)
*region_start = (byte *) ALIGN_FORWARD(heap_allowable_region_start, PAGE_SIZE);
if (region_end != NULL)
*region_end = (byte *) ALIGN_BACKWARD(heap_allowable_region_end, PAGE_SIZE);
}
#endif
/* forward declarations of static functions */
static void threadunits_init(dcontext_t *dcontext, thread_units_t *tu, size_t size);
/* dcontext only used for debugging */
static void threadunits_exit(thread_units_t *tu, dcontext_t *dcontext);
static void *common_heap_alloc(thread_units_t *tu, size_t size
HEAPACCT(which_heap_t which));
static bool common_heap_free(thread_units_t *tu, void *p, size_t size
HEAPACCT(which_heap_t which));
static void release_real_memory(void *p, size_t size, bool remove_vm);
static void release_guarded_real_memory(vm_addr_t p, size_t size, bool remove_vm,
bool guarded);
typedef enum {
/* I - Init, Interop - first allocation failed
* check for incompatible kernel drivers
*/
OOM_INIT = 0x1,
/* R - Reserve - out of virtual reservation *
* increase -vm_size to reserve more memory
*/
OOM_RESERVE = 0x2,
/* C - Commit - systemwide page file limit, or current process job limit hit
* Increase pagefile size, check for memory leak in any application.
*
* FIXME: possible automatic actions
* if systemwide failure we may want to wait if transient
* FIXME: if in a job latter we want to detect and just die
* (though after freeing as much memory as we can)
*/
OOM_COMMIT = 0x4,
/* E - Extending Commit - same reasons as Commit
* as a possible workaround increasing -heap_commit_increment
* may make expose us to commit-ing less frequently,
* On the other hand committing smaller chunks has a higher
* chance of getting through when there is very little memory.
*
* FIXME: not much more informative than OOM_COMMIT
*/
OOM_EXTEND = 0x8,
} oom_source_t;
static void report_low_on_memory(oom_source_t source,
heap_error_code_t os_error_code);
/* virtual memory manager */
enum {VMM_BLOCK_SIZE = IF_WINDOWS_ELSE(64,16)*1024}; /* 64KB or 16KB */
/* Our current allocation unit matches the allocation granularity on
* windows, to avoid worrying about external fragmentation
* Since most of our allocations fall within this range this makes the
* common operation be finding a single empty block.
*
* On Linux we save a lot of wasted alignment space by using a smaller
* granularity (PR 415959).
*
* FIXME: for Windows, if we reserve the whole region up front and
* just commit pieces, why do we need to match the Windows kernel
* alloc granularity?
*/
enum {
/* maximum 512MB virtual memory units */
MAX_VMM_HEAP_UNIT_SIZE = 512*1024*1024,
/* We should normally have only one large unit, so this is in fact
* the maximum we should count on in one process
*/
/* minimum will be used only if an invalid option is set */
MIN_VMM_HEAP_UNIT_SIZE = VMM_BLOCK_SIZE
};
typedef struct {
vm_addr_t start_addr; /* base virtual address */
vm_addr_t end_addr; /* noninclusive virtual memory range [start,end) */
vm_addr_t alloc_start; /* base allocation virtual address */
size_t alloc_size; /* allocation size */
/* for 64-bit do we want to shift to size_t to allow a larger region?
* if so must update the bitmap_t routines
*/
uint num_blocks; /* total number of blocks in virtual allocation */
mutex_t lock; /* write access to the rest of the fields is protected */
/* We make an assumption about the bitmap_t implementation being
static therefore we don't grab locks on read accesses. Anyways,
currently the bitmap_t is used with no write intent only for ASSERTs. */
uint num_free_blocks; /* currently free blocks */
/* Bitmap uses 1KB static data for granularity 64KB and static maximum 512MB */
/* Since we expect only two of these, for now it is ok for users
to have static max rather than dynamically allocating with
exact size - however this field is left last in the structure
in case we do want to save some memory
*/
bitmap_element_t blocks[BITMAP_INDEX(MAX_VMM_HEAP_UNIT_SIZE/VMM_BLOCK_SIZE)];
} vm_heap_t;
/* We keep our heap management structs on the heap for selfprot (case 8074).
* Note that we do have static structs for bootstrapping and we later move
* the data here.
*/
typedef struct _heap_management_t {
/* high-level management */
/* we reserve only a single vm_heap_t for guaranteed allocation,
* we fall back to OS when run out of reservation space */
vm_heap_t vmheap;
heap_t heap;
/* thread-shared heaps: */
thread_units_t global_units;
thread_units_t global_nonpersistent_units;
bool global_heap_writable;
thread_units_t global_unprotected_units;
} heap_management_t;
/* For bootstrapping until we can allocate our real heapmgt (case 8074).
* temp_heapmgt.lock is initialized in vmm_heap_unit_init().
*/
static heap_management_t temp_heapmgt;
static heap_management_t *heapmgt = &temp_heapmgt; /* initial value until alloced */
static bool vmm_heap_exited = false; /* FIXME: used only to thwart stack_free from trying,
should change the interface for the last stack
*/
static inline
uint
vmm_addr_to_block(vm_heap_t *vmh, vm_addr_t p)
{
ASSERT(CHECK_TRUNCATE_TYPE_uint((p - vmh->start_addr) / VMM_BLOCK_SIZE));
return (uint) ((p - vmh->start_addr) / VMM_BLOCK_SIZE);
}
static inline
vm_addr_t
vmm_block_to_addr(vm_heap_t *vmh, uint block)
{
ASSERT(block >=0 && block < vmh->num_blocks);
return (vm_addr_t)(vmh->start_addr + block*VMM_BLOCK_SIZE);
}
static bool
vmm_in_same_block(vm_addr_t p1, vm_addr_t p2)
{
return vmm_addr_to_block(&heapmgt->vmheap, p1) ==
vmm_addr_to_block(&heapmgt->vmheap, p2);
}
#if defined(DEBUG) && defined(INTERNAL)
static void
vmm_dump_map(vm_heap_t *vmh)
{
uint i;
bitmap_element_t *b = vmh->blocks;
uint bitmap_size = vmh->num_blocks;
uint last_i = 0;
bool is_used = bitmap_test(b, 0) == 0;
LOG(GLOBAL, LOG_HEAP, 3, "vmm_dump_map("PFX")\n", vmh);
/* raw dump first - if you really want binary dump use windbg's dyd */
DOLOG(3, LOG_HEAP, {
dump_buffer_as_bytes(GLOBAL, b,
BITMAP_INDEX(bitmap_size)*sizeof(bitmap_element_t),
DUMP_RAW|DUMP_ADDRESS);
});
LOG(GLOBAL, LOG_HEAP, 1, "\nvmm_dump_map("PFX") virtual regions\n", vmh);
#define VMM_DUMP_MAP_LOG(i, last_i) \
LOG(GLOBAL, LOG_HEAP, 1, PFX"-"PFX" size=%d %s\n", vmm_block_to_addr(vmh, last_i), \
vmm_block_to_addr(vmh, i-1) + VMM_BLOCK_SIZE - 1, \
(i-last_i)*VMM_BLOCK_SIZE, \
is_used ? "reserved" : "free");
for (i=0; i < bitmap_size; i++) {
/* start counting at free/used boundaries */
if (is_used != (bitmap_test(b, i) == 0)) {
VMM_DUMP_MAP_LOG(i, last_i);
is_used = (bitmap_test(b, i) == 0);
last_i = i;
}
}
VMM_DUMP_MAP_LOG(bitmap_size, last_i);
}
#endif /* DEBUG */
void
print_vmm_heap_data(file_t outf)
{
mutex_lock(&heapmgt->vmheap.lock);
print_file(outf, "VM heap: addr range "PFX"--"PFX", # free blocks %d\n",
heapmgt->vmheap.start_addr, heapmgt->vmheap.end_addr,
heapmgt->vmheap.num_free_blocks);
mutex_unlock(&heapmgt->vmheap.lock);
}
static inline
void
vmm_heap_initialize_unusable(vm_heap_t *vmh)
{
vmh->start_addr = vmh->end_addr = NULL;
vmh->num_free_blocks = vmh->num_blocks = 0;
}
static
void
vmm_heap_unit_init(vm_heap_t *vmh, size_t size)
{
ptr_uint_t preferred;
heap_error_code_t error_code;
ASSIGN_INIT_LOCK_FREE(vmh->lock, vmh_lock);
size = ALIGN_FORWARD(size, VMM_BLOCK_SIZE);
ASSERT(size <= MAX_VMM_HEAP_UNIT_SIZE);
vmh->alloc_size = size;
if (size == 0) {
vmm_heap_initialize_unusable(&heapmgt->vmheap);
return;
}
/* Out of 32 bits = 12 bits are page offset, windows wastes 4 more
* since its allocation base is 64KB, and if we want to stay
* safely in say 0x20000000-0x2fffffff we're left with only 12
* bits of randomness - which may be too little. On the other
* hand changing any of the lower 16 bits will make our bugs
* non-deterministic. */
/* Make sure we don't waste the lower bits from our random number */
preferred = (DYNAMO_OPTION(vm_base)
+ get_random_offset(DYNAMO_OPTION(vm_max_offset)/VMM_BLOCK_SIZE)
*VMM_BLOCK_SIZE);
preferred = ALIGN_FORWARD(preferred, VMM_BLOCK_SIZE);
/* overflow check: w/ vm_base shouldn't happen so debug-only check */
ASSERT(!POINTER_OVERFLOW_ON_ADD(preferred, size));
/* let's assume a single chunk is sufficient to reserve */
vmh->start_addr = NULL;
#ifdef X64
if ((byte *)preferred < heap_allowable_region_start ||
(byte *)preferred + size > heap_allowable_region_end) {
error_code = HEAP_ERROR_NOT_AT_PREFERRED;
} else {
#endif
vmh->start_addr = os_heap_reserve((void*)preferred, size, &error_code,
true/*+x*/);
LOG(GLOBAL, LOG_HEAP, 1,
"vmm_heap_unit_init preferred="PFX" got start_addr="PFX"\n",
preferred, vmh->start_addr);
#ifdef X64
}
#endif
while (vmh->start_addr == NULL && DYNAMO_OPTION(vm_allow_not_at_base)) {
/* Since we prioritize low-4GB or near-app over -vm_base, we do not
* syslog or assert here
*/
/* need extra size to ensure alignment */
vmh->alloc_size = size + VMM_BLOCK_SIZE;
#ifdef X64
/* PR 215395, make sure allocation satisfies heap reachability contraints */
vmh->alloc_start = os_heap_reserve_in_region
((void *)ALIGN_FORWARD(heap_allowable_region_start, PAGE_SIZE),
(void *)ALIGN_BACKWARD(heap_allowable_region_end, PAGE_SIZE),
size + VMM_BLOCK_SIZE, &error_code,
true/*+x*/);
#else
vmh->alloc_start = (heap_pc)
os_heap_reserve(NULL, size + VMM_BLOCK_SIZE, &error_code, true/*+x*/);
#endif
vmh->start_addr = (heap_pc) ALIGN_FORWARD(vmh->alloc_start, VMM_BLOCK_SIZE);
LOG(GLOBAL, LOG_HEAP, 1, "vmm_heap_unit_init unable to allocate at preferred="
PFX" letting OS place sz=%dM addr="PFX" \n",
preferred, size/(1024*1024), vmh->start_addr);
if (vmh->alloc_start == NULL && DYNAMO_OPTION(vm_allow_smaller)) {
/* Just a little smaller might fit */
size_t sub = (size_t) ALIGN_FORWARD(size/16, 1024*1024);
SYSLOG_INTERNAL_WARNING_ONCE("Full size vmm heap allocation failed");
if (size > sub)
size -= sub;
else
break;
} else
break;
}
#ifdef X64
/* ensure future out-of-block heap allocations are reachable from this allocation */
if (vmh->start_addr != NULL) {
ASSERT(vmh->start_addr >= heap_allowable_region_start &&
!POINTER_OVERFLOW_ON_ADD(vmh->start_addr, size) &&
vmh->start_addr + size <= heap_allowable_region_end);
request_region_be_heap_reachable(vmh->start_addr, size);
}
#endif
if (vmh->start_addr == 0) {
vmm_heap_initialize_unusable(vmh);
/* we couldn't even reserve initial virtual memory - we're out of luck */
/* XXX case 7373: make sure we tag as a potential
* interoperability issue, in staging mode we should probably
* get out from the process since we haven't really started yet
*/
report_low_on_memory(OOM_INIT, error_code);
ASSERT_NOT_REACHED();
}
vmh->end_addr = vmh->start_addr + size;
ASSERT_TRUNCATE(vmh->num_blocks, uint, size / VMM_BLOCK_SIZE);
vmh->num_blocks = (uint) (size / VMM_BLOCK_SIZE);
vmh->num_free_blocks = vmh->num_blocks;
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_unit_init ["PFX","PFX") total=%d free=%d\n",
vmh->start_addr, vmh->end_addr, vmh->num_blocks, vmh->num_free_blocks);
/* make sure static bitmap_t size is properly aligned on block boundaries */
ASSERT(ALIGNED(MAX_VMM_HEAP_UNIT_SIZE, VMM_BLOCK_SIZE));
bitmap_initialize_free(vmh->blocks, vmh->num_blocks);
DOLOG(1, LOG_HEAP, {
vmm_dump_map(vmh);
});
ASSERT(bitmap_check_consistency(vmh->blocks, vmh->num_blocks, vmh->num_free_blocks));
}
static
void
vmm_heap_unit_exit(vm_heap_t *vmh)
{
LOG(GLOBAL, LOG_HEAP, 1, "vmm_heap_unit_exit ["PFX","PFX") total=%d free=%d\n",
vmh->start_addr, vmh->end_addr, vmh->num_blocks, vmh->num_free_blocks);
/* we assume single thread in DR at this point */
DELETE_LOCK(vmh->lock);
if (vmh->start_addr == NULL)
return;
DOLOG(1, LOG_HEAP, { vmm_dump_map(vmh); });
ASSERT(bitmap_check_consistency(vmh->blocks,
vmh->num_blocks, vmh->num_free_blocks));
ASSERT(vmh->num_blocks * VMM_BLOCK_SIZE ==
(ptr_uint_t)(vmh->end_addr - vmh->start_addr));
/* In case there are no tombstones we can just free the unit and
* that is what we'll do, otherwise it will stay up forever.
*/
if (vmh->num_free_blocks == vmh->num_blocks) {
heap_error_code_t error_code;
os_heap_free(vmh->alloc_start, vmh->alloc_size, &error_code);
ASSERT(error_code == HEAP_ERROR_SUCCESS);
} else {
/* FIXME: doing nothing for now - we only care about this in
* detach scenarios where we should try to clean up from the
* virtual address space
*/
}
vmm_heap_initialize_unusable(vmh);
}
/* Returns whether within the region we reserved from the OS for doling
* out internally via our vm_heap_t; asserts that the address was also
* logically reserved within the vm_heap_t.
*/
static
bool
vmm_is_reserved_unit(vm_heap_t *vmh, vm_addr_t p, size_t size)
{
size = ALIGN_FORWARD(size, VMM_BLOCK_SIZE);
if (p < vmh->start_addr || vmh->end_addr < p/*overflow*/ ||
vmh->end_addr < (p + size))
return false;
ASSERT(CHECK_TRUNCATE_TYPE_uint(size/VMM_BLOCK_SIZE));
ASSERT(bitmap_are_reserved_blocks(vmh->blocks, vmh->num_blocks,
vmm_addr_to_block(vmh, p),
(uint)size/VMM_BLOCK_SIZE));
return true;
}
/* Returns whether entirely within the region we reserved from the OS for doling
* out internally via our vm_heap_t
*/
bool
is_vmm_reserved_address(byte *pc, size_t size)
{
ASSERT(heapmgt != NULL);
/* Case 10293: we don't call vmm_is_reserved_unit to avoid its
* assert, which we want to maintain for callers only dealing with
* DR-allocated addresses, while this routine is called w/ random
* addresses
*/
return (heapmgt != NULL && heapmgt->vmheap.start_addr != NULL &&
pc >= heapmgt->vmheap.start_addr &&
!POINTER_OVERFLOW_ON_ADD(pc, size) &&
(pc + size) <= heapmgt->vmheap.end_addr);
}
void
get_vmm_heap_bounds(byte **heap_start/*OUT*/, byte **heap_end/*OUT*/)
{
ASSERT(heapmgt != NULL);
ASSERT(heap_start != NULL && heap_end != NULL);
*heap_start = heapmgt->vmheap.start_addr;
*heap_end = heapmgt->vmheap.end_addr;
}
/* i#774: eventually we'll split vmheap from vmcode. For now, vmcode queries
* refer to the single vmheap reservation.
*/
byte *
vmcode_get_start(void)
{
byte *start, *end;
get_vmm_heap_bounds(&start, &end);
return start;
}
byte *
vmcode_get_end(void)
{
byte *start, *end;
get_vmm_heap_bounds(&start, &end);
return end;
}
byte *
vmcode_unreachable_pc(void)
{
ptr_uint_t start, end;
get_vmm_heap_bounds((byte **)&start, (byte **)&end);
if (start > INT_MAX)
return NULL;
else
return (byte *) PTR_UINT_MINUS_1;
}
bool
rel32_reachable_from_vmcode(byte *tgt)
{
byte *vmcode_start = vmcode_get_start();
byte *vmcode_end = vmcode_get_end();
ptr_int_t new_offs = (tgt > vmcode_start) ? (tgt - vmcode_start) : (vmcode_end - tgt);
/* Beyond-vmm-reservation allocs are handled b/c those are subject to the
* reachability constraints we set up on every new reservation, including
* the initial vm_reserve.
*/
return REL32_REACHABLE_OFFS(new_offs);
}
/* Reservations here are done with VMM_BLOCK_SIZE alignment
* (e.g. 64KB) but the caller is not forced to request at that
* alignment. We explicitly synchronize reservations and decommits
* within the vm_heap_t.
* Returns NULL if the VMMHeap is full or too fragmented to satisfy
* the request.
*/
static vm_addr_t
vmm_heap_reserve_blocks(vm_heap_t *vmh, size_t size_in)
{
vm_addr_t p;
uint request;
uint first_block;
size_t size;
size = ALIGN_FORWARD(size_in, VMM_BLOCK_SIZE);
ASSERT_TRUNCATE(request, uint, size/VMM_BLOCK_SIZE);
request = (uint) size/VMM_BLOCK_SIZE;
LOG(GLOBAL, LOG_HEAP, 2,
"vmm_heap_reserve_blocks: size=%d => %d in blocks=%d free_blocks~=%d\n",
size_in, size, request, vmh->num_free_blocks);
mutex_lock(&vmh->lock);
if (vmh->num_free_blocks < request) {
mutex_unlock(&vmh->lock);
return NULL;
}
first_block = bitmap_allocate_blocks(vmh->blocks, vmh->num_blocks, request);
if (first_block != BITMAP_NOT_FOUND) {
vmh->num_free_blocks -= request;
}
mutex_unlock(&vmh->lock);
if (first_block != BITMAP_NOT_FOUND) {
p = vmm_block_to_addr(vmh, first_block);
STATS_ADD_PEAK(vmm_vsize_used, size);
STATS_ADD_PEAK(vmm_vsize_blocks_used, request);
STATS_ADD_PEAK(vmm_vsize_wasted, size - size_in);
DOSTATS({
if (request > 1) {
STATS_INC(vmm_multi_block_allocs);
STATS_ADD(vmm_multi_blocks, request);
}
});
} else {
p = NULL;
}
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_reserve_blocks: size=%d blocks=%d p="PFX"\n",
size, request, p);
DOLOG(5, LOG_HEAP, { vmm_dump_map(vmh); });
return p;
}
/* We explicitly synchronize reservations and decommits within the vm_heap_t.
* Update bookkeeping information about the freed region.
*/
static void
vmm_heap_free_blocks(vm_heap_t *vmh, vm_addr_t p, size_t size_in)
{
uint first_block = vmm_addr_to_block(vmh, p);
uint request;
size_t size;
size = ALIGN_FORWARD(size_in, VMM_BLOCK_SIZE);
ASSERT_TRUNCATE(request, uint, size/VMM_BLOCK_SIZE);
request = (uint) size/VMM_BLOCK_SIZE;
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_free_blocks: size=%d blocks=%d p="PFX"\n",
size, request, p);
mutex_lock(&vmh->lock);
bitmap_free_blocks(vmh->blocks, vmh->num_blocks, first_block, request);
vmh->num_free_blocks += request;
mutex_unlock(&vmh->lock);
ASSERT(vmh->num_free_blocks <= vmh->num_blocks);
STATS_SUB(vmm_vsize_used, size);
STATS_SUB(vmm_vsize_blocks_used, request);
STATS_SUB(vmm_vsize_wasted, size - size_in);
}
/* This is the proper interface for the rest of heap.c to the os_heap_* functions */
/* place all the local-scope static vars (from DO_THRESHOLD) into .fspdata to avoid
* protection changes */
START_DATA_SECTION(FREQ_PROTECTED_SECTION, "w");
static bool
at_reset_at_vmm_limit()
{
return
(DYNAMO_OPTION(reset_at_vmm_percent_free_limit) != 0 &&
100 * heapmgt->vmheap.num_free_blocks <
DYNAMO_OPTION(reset_at_vmm_percent_free_limit) * heapmgt->vmheap.num_blocks) ||
(DYNAMO_OPTION(reset_at_vmm_free_limit) != 0 &&
heapmgt->vmheap.num_free_blocks * VMM_BLOCK_SIZE <
DYNAMO_OPTION(reset_at_vmm_free_limit));
}
/* Reserve virtual address space without committing swap space for it */
static vm_addr_t
vmm_heap_reserve(size_t size, heap_error_code_t *error_code, bool executable)
{
vm_addr_t p;
/* should only be used on sizable aligned pieces */
ASSERT(size > 0 && ALIGNED(size, PAGE_SIZE));
ASSERT(!OWN_MUTEX(&reset_pending_lock));
if (DYNAMO_OPTION(vm_reserve)) {
/* FIXME: should we make this an external option? */
if (INTERNAL_OPTION(vm_use_last) ||
(DYNAMO_OPTION(switch_to_os_at_vmm_reset_limit) && at_reset_at_vmm_limit())) {
DO_ONCE({
if (DYNAMO_OPTION(reset_at_switch_to_os_at_vmm_limit))
schedule_reset(RESET_ALL);
DOCHECK(1, {
if (!INTERNAL_OPTION(vm_use_last)) {
ASSERT_CURIOSITY(false && "running low on vm reserve");
}
});
/* FIXME - for our testing would be nice to have some release build
* notification of this ... */
});
DODEBUG(ever_beyond_vmm = true;);
#ifdef X64
/* PR 215395, make sure allocation satisfies heap reachability contraints */
p = os_heap_reserve_in_region
((void *)ALIGN_FORWARD(heap_allowable_region_start, PAGE_SIZE),
(void *)ALIGN_BACKWARD(heap_allowable_region_end, PAGE_SIZE),
size, error_code, executable);
/* ensure future heap allocations are reachable from this allocation */
if (p != NULL)
request_region_be_heap_reachable(p, size);
#else
p = os_heap_reserve(NULL, size, error_code, executable);
#endif
if (p != NULL)
return p;
LOG(GLOBAL, LOG_HEAP, 1, "vmm_heap_reserve: failed "PFX"\n",
*error_code);
}
if (at_reset_at_vmm_limit()) {
/* We're running low on our reservation, trigger a reset */
if (schedule_reset(RESET_ALL)) {
STATS_INC(reset_low_vmm_count);
DO_THRESHOLD_SAFE(DYNAMO_OPTION(report_reset_vmm_threshold),
FREQ_PROTECTED_SECTION,
{/* < max - nothing */},
{/* >= max */
/* FIXME - do we want to report more then once to give some idea of
* how much thrashing there is? */
DO_ONCE({
SYSLOG_CUSTOM_NOTIFY(SYSLOG_WARNING, MSG_LOW_ON_VMM_MEMORY, 2,
"Potentially thrashing on low virtual "
"memory resetting.", get_application_name(),
get_application_pid());
/* want QA to notice */
ASSERT_CURIOSITY(false && "vmm heap limit reset thrashing");
});
});
}
}
p = vmm_heap_reserve_blocks(&heapmgt->vmheap, size);
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_reserve: size=%d p="PFX"\n",
size, p);
if (p)
return p;
DO_ONCE({
DODEBUG({ out_of_vmheap_once = true; });
if (!INTERNAL_OPTION(skip_out_of_vm_reserve_curiosity)) {
/* this maybe unsafe for early services w.r.t. case 666 */
SYSLOG_INTERNAL_WARNING("Out of vmheap reservation - reserving %dKB."
"Falling back onto OS allocation", size/1024);
ASSERT_CURIOSITY(false && "Out of vmheap reservation");
}
/* This actually-out trigger is only trying to help issues like a
* thread-private configuration being a memory hog (and thus we use up
* our reserve). Reset needs memory, and this is asynchronous, so no
* guarantees here anyway (the app may have already reserved all memory
* beyond our reservation, see sqlsrvr.exe and cisvc.exe for ex.) which is
* why we have -reset_at_vmm_threshold to make reset success more likely. */
if (DYNAMO_OPTION(reset_at_vmm_full)) {
schedule_reset(RESET_ALL);
}
});
}
/* if we fail to allocate from our reservation we fall back to the OS */
DODEBUG(ever_beyond_vmm = true;);
#ifdef X64
/* PR 215395, make sure allocation satisfies heap reachability contraints */
p = os_heap_reserve_in_region
((void *)ALIGN_FORWARD(heap_allowable_region_start, PAGE_SIZE),
(void *)ALIGN_BACKWARD(heap_allowable_region_end, PAGE_SIZE),
size, error_code, executable);
/* ensure future heap allocations are reachable from this allocation */
if (p != NULL)
request_region_be_heap_reachable(p, size);
#else
p = os_heap_reserve(NULL, size, error_code, executable);
#endif
return p;
}
/* Commit previously reserved pages, returns false when out of memory
* This is here just to complement the vmm interface, in fact it is
* almost an alias for os_heap_commit. (If we had strict types then
* here we'd convert a vm_addr_t into a heap_pc.)
*/
static inline bool
vmm_heap_commit(vm_addr_t p, size_t size, uint prot, heap_error_code_t *error_code)
{
bool res = os_heap_commit(p, size, prot, error_code);
size_t commit_used, commit_limit;
ASSERT(!OWN_MUTEX(&reset_pending_lock));
if ((DYNAMO_OPTION(reset_at_commit_percent_free_limit) != 0 ||
DYNAMO_OPTION(reset_at_commit_free_limit) != 0) &&
os_heap_get_commit_limit(&commit_used, &commit_limit)) {
size_t commit_left = commit_limit - commit_used;
ASSERT(commit_used <= commit_limit);
/* FIXME - worry about overflow in the multiplies below? With 4kb pages isn't
* an issue till 160GB of committable memory. */
if ((DYNAMO_OPTION(reset_at_commit_free_limit) != 0 &&
commit_left < DYNAMO_OPTION(reset_at_commit_free_limit) / PAGE_SIZE) ||
(DYNAMO_OPTION(reset_at_commit_percent_free_limit) != 0 &&
100 * commit_left <
DYNAMO_OPTION(reset_at_commit_percent_free_limit) * commit_limit)) {
/* Machine is getting low on memory, trigger a reset */
/* FIXME - if we aren't the ones hogging committed memory (rougue app) then
* do we want a version of reset that doesn't de-commit our already grabbed
* memory to avoid someone else stealing it (or perhaps keep just a minimal
* level to ensure we make some progress)? */
/* FIXME - the commit limit is for the whole system; we have no good way of
* telling if we're running in a job and if so what the commit limit for the
* job is. */
/* FIXME - if a new process is started under dr while the machine is already
* past the threshold we will just spin resetting here and not make any
* progress, may be better to only reset when we have a reasonable amount of
* non-persistent memory to free (so that we can at least make some progress
* before resetting again). */
/* FIXME - the threshold is calculated at the current page file size, but
* it's possible that the pagefile is expandable (dependent on disk space of
* course) and thus we're preventing a potentially beneficial (to us)
* upsizing of the pagefile here. See "HKLM\SYSTEM\CCS\ControlSession /
* Manager\Memory Management" for the initial/max size of the various page
* files (query SystemPafefileInformation only gets you the current size). */
/* xref case 345 on fixmes (and link to wiki discussion) */
if (schedule_reset(RESET_ALL)) {
STATS_INC(reset_low_commit_count);
DO_THRESHOLD_SAFE(DYNAMO_OPTION(report_reset_commit_threshold),
FREQ_PROTECTED_SECTION,
{/* < max - nothing */},
{/* >= max */
/* FIXME - do we want to report more then once to give some idea of
* how much thrashing there is? */
DO_ONCE({
SYSLOG_CUSTOM_NOTIFY(SYSLOG_WARNING,
MSG_LOW_ON_COMMITTABLE_MEMORY, 2,
"Potentially thrashing on low committable "
"memory resetting.", get_application_name(),
get_application_pid());
/* want QA to notice */
ASSERT_CURIOSITY(false && "commit limit reset thrashing");
});
});
}
}
}
if (!res &&
DYNAMO_OPTION(oom_timeout) != 0) {
DEBUG_DECLARE(heap_error_code_t old_error_code = *error_code;)
ASSERT(old_error_code != HEAP_ERROR_SUCCESS);
/* check whether worth retrying */
if (!os_heap_systemwide_overcommit(*error_code)) {
/* FIXME: we should check whether current process is the hog */
/* unless we have used the memory, there is still a
* miniscule chance another thread will free up some or
* will attempt suicide, so could retry even if current
* process has a leak */
ASSERT_NOT_IMPLEMENTED(false);
/* retry */
}
SYSLOG_INTERNAL_WARNING("vmm_heap_commit oom: timeout and retry");
/* let's hope a memory hog dies in the mean time */
os_timeout(DYNAMO_OPTION(oom_timeout));
res = os_heap_commit(p, size, prot, error_code);
DODEBUG({
if (res) {
SYSLOG_INTERNAL_WARNING("vmm_heap_commit retried, got away! old="PFX" new="PFX"\n",
old_error_code, *error_code);
} else {
SYSLOG_INTERNAL_WARNING("vmm_heap_commit retrying, no luck. old="PFX" new="PFX"\n",
old_error_code, *error_code);
}
});
}
return res;
}
/* back to normal section */
END_DATA_SECTION()
/* Free previously reserved and possibly committed memory. Check if
* it is within the memory managed by the virtual memory manager we
* only decommit back to the OS, and we remove the vmm reservation.
* Keep in mind that this can be called on units that are not fully
* committed, e.g. guard pages are added to this - as long as the
* os_heap_decommit interface can handle this we're OK
*/
static void
vmm_heap_free(vm_addr_t p, size_t size, heap_error_code_t *error_code)
{
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_free: size=%d p="PFX" is_reserved=%d\n",
size, p, vmm_is_reserved_unit(&heapmgt->vmheap, p, size));
/* the memory doesn't have to be within our VM reserve if it
was allocated as an extra OS call when if we ran out */
if (DYNAMO_OPTION(vm_reserve)) {
if (vmm_is_reserved_unit(&heapmgt->vmheap, p, size)) {
os_heap_decommit(p, size, error_code);
vmm_heap_free_blocks(&heapmgt->vmheap, p, size);
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_free: freed size=%d p="PFX"\n",
size, p);
return;
} else {
/* FIXME: check if this is stack_free getting in the way, then ignore it */
/* FIXME: could do this by overriding the meaning of the vmheap fields
after cleanup to a different combination that start_pc = end_pc = NULL
*/
/* FIXME: see vmm_heap_unit_exit for the current stack_free problem */
if (vmm_heap_exited) {
*error_code = HEAP_ERROR_SUCCESS;
return;
}
}
}
os_heap_free(p, size, error_code);
}
static void
vmm_heap_decommit(vm_addr_t p, size_t size, heap_error_code_t *error_code)
{
LOG(GLOBAL, LOG_HEAP, 2, "vmm_heap_decommit: size=%d p="PFX" is_reserved=%d\n",
size, p, vmm_is_reserved_unit(&heapmgt->vmheap, p, size));
os_heap_decommit(p, size, error_code);
/* nothing to be done to vmm blocks */
}
/* Caller is required to handle thread synchronization and to update dynamo vm areas.
* size must be PAGE_SIZE-aligned.
* Returns NULL if fails to allocate memory!
*/
static void *
vmm_heap_alloc(size_t size, uint prot, heap_error_code_t *error_code)
{
vm_addr_t p = vmm_heap_reserve(size, error_code, TEST(MEMPROT_EXEC, prot));
if (!p)
return NULL; /* out of reserved memory */
if (!vmm_heap_commit(p, size, prot, error_code))
return NULL; /* out of committed memory */
return p;
}
/* set reachability constraints before loading any client libs */
void
vmm_heap_init_constraints()
{
#ifdef X64
/* add reachable regions before we allocate the heap, xref PR 215395 */
/* i#774, i#901: we no longer need the DR library nor ntdll.dll to be
* reachable by the vmheap reservation. But, for -heap_in_lower_4GB,
* we must call request_region_be_heap_reachable() up front.
*/
if (DYNAMO_OPTION(heap_in_lower_4GB))
request_region_be_heap_reachable((byte *)(ptr_uint_t)0x80000000/*middle*/, 1);
else if (DYNAMO_OPTION(vm_base_near_app)) {
/* Required for STATIC_LIBRARY: must be near app b/c clients are there.
* Non-static: still a good idea for fewer rip-rel manglings.
* Asking for app base means we'll prefer before the app, which
* has less of an impact on its heap.
*/
app_pc base = get_application_base();
request_region_be_heap_reachable(base, get_application_end() - base);
} else {
/* It seems silly to let the 1st client lib set the region, so we give
* -vm_base priority.
*/
request_region_be_heap_reachable
((byte *)DYNAMO_OPTION(vm_base), DYNAMO_OPTION(vm_size));
}
/* XXX: really we should iterate and try other options: right now we'll
* just fail if we run out of space. E.g., if the app is quite large, we might
* fit the client near, and then our vm reservation could fail and at that
* point we'd just abort. We need to restructure the code to allow
* iterating over the client lib loads and vm reservation at once.
*/
#endif /* X64 */
}
/* virtual memory manager initialization */
void
vmm_heap_init()
{
IF_WINDOWS(ASSERT(VMM_BLOCK_SIZE == OS_ALLOC_GRANULARITY));
if (DYNAMO_OPTION(vm_reserve)) {
vmm_heap_unit_init(&heapmgt->vmheap, DYNAMO_OPTION(vm_size));
}
}
void
vmm_heap_exit()
{
/* virtual memory manager exit */
if (DYNAMO_OPTION(vm_reserve)) {
/* FIXME: we have three regions that are not explicitly
* deallocated current stack, init stack, global_do_syscall
*/
DOCHECK(1, {
uint perstack =
ALIGN_FORWARD_UINT(dynamo_options.stack_size +
(dynamo_options.guard_pages ? (2*PAGE_SIZE) : 0),
VMM_BLOCK_SIZE) /
VMM_BLOCK_SIZE;
uint unfreed_blocks = perstack * 1 /* initstack */ +
/* current stack */
perstack * ((IF_WINDOWS_ELSE(doing_detach, false)
IF_APP_EXPORTS(|| dr_api_exit)) ? 0 : 1);
/* FIXME: on detach arch_thread_exit should explicitly mark as
left behind all TPCs needed so then we can assert even for
detach
*/
ASSERT(IF_WINDOWS(doing_detach || ) /* not deterministic when detaching */
heapmgt->vmheap.num_free_blocks == heapmgt->vmheap.num_blocks
- unfreed_blocks ||
/* >=, not ==, b/c if we hit the vmm limit the cur dstack
* could be outside of vmm (i#1164)
*/
(ever_beyond_vmm && heapmgt->vmheap.num_free_blocks >=
heapmgt->vmheap.num_blocks - unfreed_blocks));
});
/* FIXME: On process exit we are currently executing off a
* stack in this region so we cannot free the whole allocation.
* FIXME: Any tombstone allocations will have to use a
* different interface than the generic heap_mmap() which is
* sometimes used to leave things behind. FIXME: Currently
* we'll leave behind the whole vm unit if any tombstones are
* left - which in fact is always the case, no matter whether
* thread private code needs to be left or not.
* global_do_syscall 32 byte allocation should be part of our
* dll and won't have to be left.
* The current stack is the main problem because it is later
* cleaned up in cleanup_and_terminate by calling stack_free which
* in turn gets all the way to vmm_heap_free. Therefore we add an
* explicit test for vmm_heap_exited, so that we can otherwise free
* bookkeeping information and delete the lock now.
* Potential solution to most of these problems is to have
* cleanup_and_terminate call vmm_heap_exit when cleaning up
* the process, or to just leave the vm mapping behind and
* simply pass a different argument to stack_free.
*/
vmm_heap_unit_exit(&heapmgt->vmheap);
vmm_heap_exited = true;
}
}
/* checks for compatibility among heap options, returns true if
* modified the value of any options to make them compatible
*/
bool
heap_check_option_compatibility()
{
bool ret = false;
ret = check_param_bounds(&dynamo_options.vm_size, MIN_VMM_HEAP_UNIT_SIZE,
MAX_VMM_HEAP_UNIT_SIZE, "vm_size")
|| ret;
#ifdef INTERNAL
/* if max_heap_unit_size is too small you may get a funny message
* "initial_heap_unit_size must be >= 8229 and <= 4096" but in
* release build we will take the min and then complain about
* max_heap_unit_size and set it to the min also, so it all works
* out w/o needing an extra check() call.
*/
/* case 7626: don't short-circuit checks, as later ones may be needed */
ret = check_param_bounds(&dynamo_options.initial_heap_unit_size,
/* if have units smaller than a page we end up
* allocating 64KB chunks for "oversized" units
* for just about every alloc! so round up to
* at least a page.
*/
ALIGN_FORWARD(UNITOVERHEAD + 1, PAGE_SIZE),
HEAP_UNIT_MAX_SIZE, "initial_heap_unit_size")
|| ret;
ret = check_param_bounds(&dynamo_options.initial_global_heap_unit_size,
ALIGN_FORWARD(UNITOVERHEAD + 1, PAGE_SIZE),
HEAP_UNIT_MAX_SIZE, "initial_global_heap_unit_size")
|| ret;
ret = check_param_bounds(&dynamo_options.max_heap_unit_size,
MAX(HEAP_UNIT_MIN_SIZE, GLOBAL_UNIT_MIN_SIZE),
INT_MAX, "max_heap_unit_size")
|| ret;
#endif
return ret;
}
/* thread-shared initialization that should be repeated after a reset */
void
heap_reset_init()
{
if (DYNAMO_OPTION(enable_reset)) {
threadunits_init(GLOBAL_DCONTEXT, &heapmgt->global_nonpersistent_units,
GLOBAL_UNIT_MIN_SIZE);
}
}
/* initialization */
void
heap_init()
{
int i;
uint prev_sz = 0;
LOG(GLOBAL, LOG_TOP|LOG_HEAP, 2, "Heap bucket sizes are:\n");
/* make sure we'll preserve alignment */
ASSERT(ALIGNED(HEADER_SIZE, HEAP_ALIGNMENT));
/* make sure free list pointers will fit */
ASSERT(BLOCK_SIZES[0] >= sizeof(heap_pc*));
/* since sizes depend on size of structs, make sure they're in order */
for (i = 0; i < BLOCK_TYPES; i++) {
ASSERT(BLOCK_SIZES[i] > prev_sz);
/* we assume all of our heap allocs are aligned */
ASSERT(i == BLOCK_TYPES-1 || ALIGNED(BLOCK_SIZES[i], HEAP_ALIGNMENT));
prev_sz = BLOCK_SIZES[i];
LOG(GLOBAL, LOG_TOP|LOG_HEAP, 2, "\t%d bytes\n", BLOCK_SIZES[i]);
}
/* we assume writes to some static vars are atomic,
* i.e., the vars don't cross cache lines. they shouldn't since
* they should all be 4-byte-aligned in the data segment.
* FIXME: ensure that release build aligns ok?
* I would be quite surprised if static vars were not 4-byte-aligned!
*/
ASSERT(ALIGN_BACKWARD(&heap_exiting, CACHE_LINE_SIZE()) ==
ALIGN_BACKWARD(&heap_exiting + 1, CACHE_LINE_SIZE()));
ASSERT(ALIGN_BACKWARD(&heap_unit_lock.owner, CACHE_LINE_SIZE()) ==
ALIGN_BACKWARD(&heap_unit_lock.owner + 1, CACHE_LINE_SIZE()));
/* For simplicity we go through our normal heap mechanism to allocate
* our post-init heapmgt struct
*/
ASSERT(heapmgt == &temp_heapmgt);
heapmgt->global_heap_writable = true;
threadunits_init(GLOBAL_DCONTEXT, &heapmgt->global_units, GLOBAL_UNIT_MIN_SIZE);
heapmgt = HEAP_TYPE_ALLOC(GLOBAL_DCONTEXT, heap_management_t, ACCT_MEM_MGT, PROTECTED);
memset(heapmgt, 0, sizeof(*heapmgt));
ASSERT(sizeof(temp_heapmgt) == sizeof(*heapmgt));
memcpy(heapmgt, &temp_heapmgt, sizeof(temp_heapmgt));
threadunits_init(GLOBAL_DCONTEXT, &heapmgt->global_unprotected_units,
GLOBAL_UNIT_MIN_SIZE);
heap_reset_init();
#ifdef WINDOWS
/* PR 250294: As part of 64-bit hook work, hook reachability was addressed
* using landing pads (see win32/callback.c for more explanation). Landing
* pad areas are a type of special heap, so they should be initialized
* during heap init.
* Each landing pad area has its own allocation pointer, so they shouldn't
* be merged automatically.
*/
VMVECTOR_ALLOC_VECTOR(landing_pad_areas, GLOBAL_DCONTEXT,
VECTOR_SHARED | VECTOR_NEVER_MERGE,
landing_pad_areas_lock);
#endif
}
/* need to not remove from vmareas on process exit -- vmareas has already exited! */
static void
really_free_unit(heap_unit_t *u)
{
STATS_SUB(heap_capacity, UNIT_COMMIT_SIZE(u));
STATS_ADD(heap_reserved_only, (stats_int_t)(UNIT_COMMIT_SIZE(u) - UNIT_RESERVED_SIZE(u)));
/* remember that u itself is inside unit, not separately allocated */
release_guarded_real_memory((vm_addr_t)u, UNIT_RESERVED_SIZE(u),
false/*do not update DR areas now*/, true);
}
/* Free all thread-shared state not critical to forward progress;
* heap_reset_init() will be called before continuing.
*/
void
heap_reset_free()
{
heap_unit_t *u, *next_u;
/* FIXME: share some code w/ heap_exit -- currently only called by reset */
ASSERT(DYNAMO_OPTION(enable_reset));
/* we must grab this lock before heap_unit_lock to avoid rank
* order violations when freeing
*/
dynamo_vm_areas_lock();
/* for combining stats into global_units we need this lock
* FIXME: remove if we go to separate stats sum location
*/
DODEBUG({ acquire_recursive_lock(&global_alloc_lock); });
acquire_recursive_lock(&heap_unit_lock);
LOG(GLOBAL, LOG_HEAP, 1, "Pre-reset, global heap unit stats:\n");
/* FIXME: free directly rather than putting on dead list first */
threadunits_exit(&heapmgt->global_nonpersistent_units, GLOBAL_DCONTEXT);
/* free all dead units */
u = heapmgt->heap.dead;
while (u != NULL) {
next_u = u->next_global;
LOG(GLOBAL, LOG_HEAP, 1, "\tfreeing dead unit "PFX"-"PFX" [-"PFX"]\n",
u, UNIT_COMMIT_END(u), UNIT_RESERVED_END(u));
RSTATS_DEC(heap_num_free);
really_free_unit(u);
u = next_u;
}
heapmgt->heap.dead = NULL;
heapmgt->heap.num_dead = 0;
release_recursive_lock(&heap_unit_lock);
DODEBUG({ release_recursive_lock(&global_alloc_lock); });
dynamo_vm_areas_unlock();
}
/* atexit cleanup */
void
heap_exit()
{
heap_unit_t *u, *next_u;
heap_management_t *temp;
heap_exiting = true;
/* FIXME: we shouldn't need either lock if executed last */
dynamo_vm_areas_lock();
acquire_recursive_lock(&heap_unit_lock);
#ifdef WINDOWS
release_landing_pad_mem(); /* PR 250294 */
#endif
LOG(GLOBAL, LOG_HEAP, 1, "Global unprotected heap unit stats:\n");
threadunits_exit(&heapmgt->global_unprotected_units, GLOBAL_DCONTEXT);
if (DYNAMO_OPTION(enable_reset)) {
LOG(GLOBAL, LOG_HEAP, 1, "Global nonpersistent heap unit stats:\n");
threadunits_exit(&heapmgt->global_nonpersistent_units, GLOBAL_DCONTEXT);
}
/* Now we need to go back to the static struct to clean up */
ASSERT(heapmgt != &temp_heapmgt);
memcpy(&temp_heapmgt, heapmgt, sizeof(temp_heapmgt));
temp = heapmgt;
heapmgt = &temp_heapmgt;
HEAP_TYPE_FREE(GLOBAL_DCONTEXT, temp, heap_management_t, ACCT_MEM_MGT, PROTECTED);
LOG(GLOBAL, LOG_HEAP, 1, "Global heap unit stats:\n");
threadunits_exit(&heapmgt->global_units, GLOBAL_DCONTEXT);
/* free heap for all unfreed units */
LOG(GLOBAL, LOG_HEAP, 1, "Unfreed units:\n");
u = heapmgt->heap.units;
while (u != NULL) {
next_u = u->next_global;
LOG(GLOBAL, LOG_HEAP, 1, "\tfreeing live unit "PFX"-"PFX" [-"PFX"]\n",
u, UNIT_COMMIT_END(u), UNIT_RESERVED_END(u));
RSTATS_DEC(heap_num_live);
really_free_unit(u);
u = next_u;
}
heapmgt->heap.units = NULL;
u = heapmgt->heap.dead;
while (u != NULL) {
next_u = u->next_global;
LOG(GLOBAL, LOG_HEAP, 1, "\tfreeing dead unit "PFX"-"PFX" [-"PFX"]\n",
u, UNIT_COMMIT_END(u), UNIT_RESERVED_END(u));
RSTATS_DEC(heap_num_free);
really_free_unit(u);
u = next_u;
}
heapmgt->heap.dead = NULL;
release_recursive_lock(&heap_unit_lock);
dynamo_vm_areas_unlock();
DELETE_RECURSIVE_LOCK(heap_unit_lock);
DELETE_RECURSIVE_LOCK(global_alloc_lock);
#ifdef X64
DELETE_LOCK(request_region_be_heap_reachable_lock);
#endif
}
/* FIXME:
* detect if the app is who we're fighting for memory, if so, don't
* free memory, else the app will just keep grabbing more.
* need a test for hitting 2GB (or 3GB!) user mode limit.
*/
static void
heap_low_on_memory()
{
/* free some memory! */
heap_unit_t *u, *next_u;
size_t freed = 0;
LOG(GLOBAL, LOG_CACHE|LOG_STATS, 1,
"heap_low_on_memory: about to free dead list units\n");
/* WARNING: this routine is called at arbitrary allocation failure points,
* so we have to be careful what locks we grab
* However, no allocation site can hold a lock weaker in rank than
* heap_unit_lock, b/c it could deadlock on the allocation itself!
* So we're safe.
*/
/* must grab this lock prior to heap_unit_lock if affecting DR vm areas
* this is recursive so ok if we ran out of memory while holding DR vm area lock
*/
ASSERT(safe_to_allocate_or_free_heap_units());
dynamo_vm_areas_lock();
acquire_recursive_lock(&heap_unit_lock);
u = heapmgt->heap.dead;
while (u != NULL) {
next_u = u->next_global;
freed += UNIT_COMMIT_SIZE(u);
/* FIXME: if out of committed pages only, could keep our reservations */
LOG(GLOBAL, LOG_HEAP, 1, "\tfreeing dead unit "PFX"-"PFX" [-"PFX"]\n",
u, UNIT_COMMIT_END(u), UNIT_RESERVED_END(u));
RSTATS_DEC(heap_num_free);
really_free_unit(u);
u = next_u;
heapmgt->heap.num_dead--;
}
heapmgt->heap.dead = NULL;
release_recursive_lock(&heap_unit_lock);
dynamo_vm_areas_unlock();
LOG(GLOBAL, LOG_CACHE|LOG_STATS, 1,
"heap_low_on_memory: freed %d KB\n", freed/1024);
/* FIXME: we don't keep a list of guard pages, which we may decide to throw
* out or compact at this time.
*/
/* FIXME: should also fix up the allocator to look in other free lists
* of sizes larger than asked for, we may have plenty of memory available
* in other lists! see comments in common_heap_alloc
*/
}
static const char*
get_oom_source_name(oom_source_t source)
{
/* currently only single character codenames,
* (still as a string though)
*/
const char *code_name = "?";
switch (source) {
case OOM_INIT : code_name = "I"; break;
case OOM_RESERVE : code_name = "R"; break;
case OOM_COMMIT : code_name = "C"; break;
case OOM_EXTEND : code_name = "E"; break;
default:
ASSERT_NOT_REACHED();
}
return code_name;
}
static bool
silent_oom_for_process(oom_source_t source)
{
if (TESTANY(OOM_COMMIT|OOM_EXTEND, source) &&
!IS_STRING_OPTION_EMPTY(silent_commit_oom_list)) {
bool onlist;
const char *process_name = get_short_name(get_application_name());
string_option_read_lock();
onlist = check_filter_with_wildcards(DYNAMO_OPTION(silent_commit_oom_list),
process_name);
string_option_read_unlock();
if (onlist) {
SYSLOG_INTERNAL_WARNING("not reporting last words of executable %s",
process_name);
return true;
}
}
return false;
}
/* oom_source_t identifies the action we were taking, os_error_code is
* the returned value from the last system call - opaque at this OS
* independent layer.
*/
static void
report_low_on_memory(oom_source_t source, heap_error_code_t os_error_code)
{
if (TESTANY(DYNAMO_OPTION(silent_oom_mask), source)
|| silent_oom_for_process(source)) {
SYSLOG_INTERNAL_WARNING("Mostly silent OOM: %s "PFX".\n",
get_oom_source_name(source), os_error_code);
/* still produce an ldmp for internal use */
if (TEST(DUMPCORE_OUT_OF_MEM_SILENT, DYNAMO_OPTION(dumpcore_mask)))
os_dump_core("Out of memory, silently aborting program.");
} else {
const char *oom_source_code = get_oom_source_name(source);
char status_hex[19]; /* FIXME: for 64bit hex need 16+NULL */
/* note 0x prefix added by the syslog */
snprintf(status_hex,
BUFFER_SIZE_ELEMENTS(status_hex), PFX, /* FIXME: 32bit */
os_error_code);
NULL_TERMINATE_BUFFER(status_hex);
/* SYSLOG first */
SYSLOG_CUSTOM_NOTIFY(SYSLOG_CRITICAL, MSG_OUT_OF_MEMORY, 4,
"Out of memory. Program aborted.",
get_application_name(), get_application_pid(),
oom_source_code, status_hex
);
/* FIXME: case 7306 can't specify arguments in SYSLOG_CUSTOM_NOTIFY */
SYSLOG_INTERNAL_WARNING("OOM Status: %s %s", oom_source_code, status_hex);
/* FIXME: case 7296 - ldmp even if we have decided not to produce an event above */
if (TEST(DUMPCORE_OUT_OF_MEM, DYNAMO_OPTION(dumpcore_mask)))
os_dump_core("Out of memory, aborting program.");
/* passing only status code to XML where we should have a stack dump and callstack */
report_diagnostics("Out of memory", status_hex, NO_VIOLATION_BAD_INTERNAL_STATE);
}
os_terminate(NULL, TERMINATE_PROCESS);
ASSERT_NOT_REACHED();
}
/* update statistics for committed memory, and add to vm_areas */
static inline void
account_for_memory(void *p, size_t size, uint prot, bool add_vm, bool image
_IF_DEBUG(const char *comment))
{
STATS_ADD_PEAK(memory_capacity, size);
/* case 3045: areas inside the vmheap reservation are not added to the list
* for clients that use DR-allocated memory, we have get_memory_info()
* query from the OS to see inside
*/
if (vmm_is_reserved_unit(&heapmgt->vmheap, p, size)) {
return;
}
if (add_vm) {
add_dynamo_vm_area(p, ((app_pc)p) + size, prot, image _IF_DEBUG(comment));
} else {
/* due to circular dependencies bet vmareas and global heap we do not call
* add_dynamo_vm_area here, instead we indicate that something has changed
*/
mark_dynamo_vm_areas_stale();
/* NOTE: 'prot' info is lost about this region, but is needed in
* heap_vmareas_synch_units to update all_memory_areas. Currently
* heap_create_unit is the only place that passes 'false' with prot rw-.
*/
ASSERT(TESTALL(MEMPROT_READ | MEMPROT_WRITE, prot));
}
}
/* remove_vm MUST be false iff this is heap memory, which is updated separately */
static void
update_dynamo_areas_on_release(app_pc start, app_pc end, bool remove_vm)
{
if (!vm_areas_exited && !heap_exiting) { /* avoid problems when exiting */
/* case 3045: areas inside the vmheap reservation are not added to the list
* for clients that use DR-allocated memory, we have get_memory_info()
* query from the OS to see inside
*/
if (vmm_is_reserved_unit(&heapmgt->vmheap, start, end - start)) {
return;
}
if (remove_vm) {
remove_dynamo_vm_area(start, end);
} else {
/* due to cyclic dependencies bet heap and vmareas we cannot remove incrementally.
* the pending set is protected by the same lock needed to synch the vm areas,
* so we will never mis-identify free memory as DR memory.
*/
mark_dynamo_vm_areas_stale();
dynamo_areas_pending_remove = true;
}
}
}
bool
lockwise_safe_to_allocate_memory()
{
/* check whether it's safe to hold a lock that normally can be held
* for memory allocation -- i.e., check whether we hold the
* global_alloc_lock
*/
return !self_owns_recursive_lock(&global_alloc_lock);
}
/* we indirect all os memory requests through here so we have a central place
* to handle the out-of-memory condition.
* add_vm MUST be false iff this is heap memory, which is updated separately.
*/
static void *
get_real_memory(size_t size, uint prot, bool add_vm _IF_DEBUG(const char *comment))
{
void *p;
heap_error_code_t error_code;
/* must round up to page sizes, else vmm_heap_alloc assert triggers */
size = ALIGN_FORWARD(size, PAGE_SIZE);
/* memory alloc/dealloc and updating DR list must be atomic */
dynamo_vm_areas_lock(); /* if already hold lock this is a nop */
p = vmm_heap_alloc(size, prot, &error_code);
if (p == NULL) {
SYSLOG_INTERNAL_WARNING_ONCE("Out of memory -- cannot reserve or "
"commit %dKB. Trying to recover.", size/1024);
/* we should be ok here, shouldn't come in here holding global_alloc_lock
* or heap_unit_lock w/o first having grabbed DR areas lock
*/
ASSERT(safe_to_allocate_or_free_heap_units());
heap_low_on_memory();
fcache_low_on_memory();
/* try again
* FIXME: have more sophisticated strategy of freeing a little, then getting
* more drastic with each subsequent failure
* FIXME: can only free live fcache units for current thread w/ current
* impl...should we wait a while and try again if out of memory, hoping
* other threads have freed some?!?!
*/
p = vmm_heap_alloc(size, prot, &error_code);
if (p == NULL) {
report_low_on_memory(OOM_RESERVE, error_code);
}
SYSLOG_INTERNAL_WARNING_ONCE("Out of memory -- but still alive after "
"emergency free.");
}
account_for_memory(p, size, prot, add_vm, false _IF_DEBUG(comment));
dynamo_vm_areas_unlock();
return p;
}
static void
release_memory_and_update_areas(app_pc p, size_t size, bool decommit, bool remove_vm)
{
heap_error_code_t error_code;
/* these two operations need to be atomic wrt DR area updates */
dynamo_vm_areas_lock(); /* if already hold lock this is a nop */
/* ref case 3035, we must remove from dynamo_areas before we free in case
* we end up allocating memory in the process of removing the area
* (we don't want to end up getting the memory we just freed since that
* would lead to errors in the list when we finally did remove it)
*/
update_dynamo_areas_on_release(p, p + size, remove_vm);
if (decommit)
vmm_heap_decommit(p, size, &error_code);
else
vmm_heap_free(p, size, &error_code);
ASSERT(error_code == HEAP_ERROR_SUCCESS);
dynamo_vm_areas_unlock();
}
/* remove_vm MUST be false iff this is heap memory, which is updated separately */
static void
release_real_memory(void *p, size_t size, bool remove_vm)
{
/* must round up to page sizes for vmm_heap_free */
size = ALIGN_FORWARD(size, PAGE_SIZE);
release_memory_and_update_areas((app_pc)p, size, false/*free*/, remove_vm);
/* avoid problem w/ being called by cleanup_and_terminate after dynamo_process_exit */
DOSTATS({
if (!dynamo_exited_log_and_stats)
STATS_SUB(memory_capacity, size);
});
}
static void
extend_commitment(vm_addr_t p, size_t size, uint prot,
bool initial_commit)
{
heap_error_code_t error_code;
ASSERT(ALIGNED(p, PAGE_SIZE));
size = ALIGN_FORWARD(size, PAGE_SIZE);
if (!vmm_heap_commit(p, size, prot, &error_code)) {
SYSLOG_INTERNAL_WARNING_ONCE("Out of memory - cannot extend commit "
"%dKB. Trying to recover.", size/1024);
heap_low_on_memory();
fcache_low_on_memory();
/* see low-memory ideas in get_real_memory */
if (!vmm_heap_commit(p, size, prot, &error_code)) {
report_low_on_memory(initial_commit ? OOM_COMMIT : OOM_EXTEND,
error_code);
}
SYSLOG_INTERNAL_WARNING_ONCE("Out of memory in extend - still alive "
"after emergency free.");
}
}
/* a wrapper around get_real_memory that adds a guard page on each side of the requested unit.
* These should consume only uncommitted virtual address and should not use any physical memory.
* add_vm MUST be false iff this is heap memory, which is updated separately.
*/
static vm_addr_t
get_guarded_real_memory(size_t reserve_size, size_t commit_size, uint prot,
bool add_vm, bool guarded _IF_DEBUG(const char *comment))
{
vm_addr_t p;
uint guard_size = PAGE_SIZE;
heap_error_code_t error_code;
ASSERT(reserve_size >= commit_size);
if (!guarded || !dynamo_options.guard_pages) {
if (reserve_size == commit_size)
return get_real_memory(reserve_size, prot, add_vm _IF_DEBUG(comment));
guard_size = 0;
}
reserve_size = ALIGN_FORWARD(reserve_size, PAGE_SIZE);
commit_size = ALIGN_FORWARD(commit_size, PAGE_SIZE);
reserve_size += 2* guard_size; /* add top and bottom guards */
/* memory alloc/dealloc and updating DR list must be atomic */
dynamo_vm_areas_lock(); /* if already hold lock this is a nop */
p = vmm_heap_reserve(reserve_size, &error_code, TEST(MEMPROT_EXEC, prot));
if (p == NULL) {
/* This is very unlikely to happen - we have to reach at least 2GB reserved memory. */
SYSLOG_INTERNAL_WARNING_ONCE("Out of memory - cannot reserve %dKB. "
"Trying to recover.", reserve_size/1024);
heap_low_on_memory();
fcache_low_on_memory();
p = vmm_heap_reserve(reserve_size, &error_code, TEST(MEMPROT_EXEC, prot));
if (p == NULL) {
report_low_on_memory(OOM_RESERVE, error_code);
}
SYSLOG_INTERNAL_WARNING_ONCE("Out of memory on reserve - but still "
"alive after emergency free.");
}
/* includes guard pages if add_vm -- else, heap_vmareas_synch_units() will
* add guard pages in by assuming one page on each side of every heap unit
* if dynamo_options.guard_pages
*/
account_for_memory((void *)p, reserve_size, prot, add_vm, false _IF_DEBUG(comment));
dynamo_vm_areas_unlock();
STATS_ADD_PEAK(reserved_memory_capacity, reserve_size);
STATS_ADD_PEAK(guard_pages, 2);
p += guard_size;
extend_commitment(p, commit_size, prot, true /* initial commit */);
return p;
}
/* a wrapper around get_release_memory that also frees the guard pages on each side of the requested unit.
* remove_vm MUST be false iff this is heap memory, which is updated separately
*/
static void
release_guarded_real_memory(vm_addr_t p, size_t size, bool remove_vm, bool guarded)
{
if (!guarded || !dynamo_options.guard_pages) {
release_real_memory(p, size, remove_vm);
return;
}
size = ALIGN_FORWARD(size, PAGE_SIZE);
size += PAGE_SIZE * 2; /* add top and bottom guards */
p -= PAGE_SIZE;
release_memory_and_update_areas((app_pc)p, size, false/*free*/, remove_vm);
/* avoid problem w/ being called by cleanup_and_terminate after dynamo_process_exit */
DOSTATS({
if (!dynamo_exited_log_and_stats) {
STATS_SUB(memory_capacity, size);
STATS_SUB(reserved_memory_capacity, size);
STATS_ADD(guard_pages, -2);
}
});
}
/* use heap_mmap to allocate large chunks of executable memory
* it's mainly used to allocate our fcache units
*/
void *
heap_mmap_ex(size_t reserve_size, size_t commit_size, uint prot, bool guarded)
{
/* XXX i#774: when we split vmheap and vmcode, if MEMPROT_EXEC is requested
* here (or this is a call from a client, for reachability
* compatibility), put it in vmcode; else in vmheap.
*/
void *p = get_guarded_real_memory(reserve_size, commit_size, prot, true, guarded
_IF_DEBUG("heap_mmap"));
#ifdef DEBUG_MEMORY
if (TEST(MEMPROT_WRITE, prot))
memset(p, HEAP_ALLOCATED_BYTE, commit_size);
#endif
/* We rely on this for freeing _post_stack in absence of dcontext */
ASSERT(!DYNAMO_OPTION(vm_reserve) ||
!DYNAMO_OPTION(stack_shares_gencode) ||
(ptr_uint_t)p - (guarded ? (GUARD_PAGE_ADJUSTMENT/2) : 0) ==
ALIGN_BACKWARD(p, VMM_BLOCK_SIZE) ||
at_reset_at_vmm_limit());
LOG(GLOBAL, LOG_HEAP, 2, "heap_mmap: %d bytes [/ %d] @ "PFX"\n",
commit_size, reserve_size, p);
STATS_ADD_PEAK(mmap_capacity, commit_size);
STATS_ADD_PEAK(mmap_reserved_only, (reserve_size - commit_size));
return p;
}
/* use heap_mmap to allocate large chunks of executable memory
* it's mainly used to allocate our fcache units
*/
void *
heap_mmap_reserve(size_t reserve_size, size_t commit_size)
{
/* heap_mmap always marks as executable */
return heap_mmap_ex(reserve_size, commit_size,
MEMPROT_EXEC|MEMPROT_READ|MEMPROT_WRITE, true);
}
/* It is up to the caller to ensure commit_size is a page size multiple,
* and that it does not extend beyond the initial reservation.
*/
void
heap_mmap_extend_commitment(void *p, size_t commit_size)
{
extend_commitment(p, commit_size, MEMPROT_EXEC|MEMPROT_READ|MEMPROT_WRITE,
false /*not initial commit*/);
STATS_SUB(mmap_reserved_only, commit_size);
STATS_ADD_PEAK(mmap_capacity, commit_size);
#ifdef DEBUG_MEMORY
memset(p, HEAP_ALLOCATED_BYTE, commit_size);
#endif
}
/* De-commits from a committed region. */
void
heap_mmap_retract_commitment(void *retract_start, size_t decommit_size)
{
heap_error_code_t error_code;
ASSERT(ALIGNED(decommit_size, PAGE_SIZE));
vmm_heap_decommit(retract_start, decommit_size, &error_code);
STATS_ADD(mmap_reserved_only, decommit_size);
STATS_ADD_PEAK(mmap_capacity, -(stats_int_t)decommit_size);
}
/* Allocates executable memory in the same allocation region as this thread's
* stack, to save address space (case 9474).
*/
void *
heap_mmap_reserve_post_stack(dcontext_t *dcontext,
size_t reserve_size, size_t commit_size)
{
void *p;
byte *stack_reserve_end = NULL;
heap_error_code_t error_code;
size_t available = 0;
uint prot;
bool known_stack = false;
ASSERT(reserve_size > 0 && commit_size < reserve_size);
/* 1.5 * guard page adjustment since we'll share the middle one */
if (DYNAMO_OPTION(stack_size) + reserve_size +
GUARD_PAGE_ADJUSTMENT + GUARD_PAGE_ADJUSTMENT / 2 > VMM_BLOCK_SIZE) {
/* there's not enough room to share the allocation block, stack is too big */
LOG(GLOBAL, LOG_HEAP, 1, "Not enough room to allocate 0x%08x bytes post stack "
"of size 0x%08x\n", reserve_size, DYNAMO_OPTION(stack_size));
return heap_mmap_reserve(reserve_size, commit_size);
}
if (DYNAMO_OPTION(stack_shares_gencode) &&
/* FIXME: we could support this w/o vm_reserve, or when beyond
* the reservation, but we don't bother */
DYNAMO_OPTION(vm_reserve) &&
dcontext != GLOBAL_DCONTEXT && dcontext != NULL) {
stack_reserve_end = dcontext->dstack + GUARD_PAGE_ADJUSTMENT/2;
#if defined(UNIX) && !defined(HAVE_MEMINFO)
prot = 0; /* avoid compiler warning: should only need inside if */
if (!dynamo_initialized) {
/* memory info is not yet set up. since so early we only support
* post-stack if inside vmm (won't be true only for pathological
* tiny vmm sizes)
*/
if (vmm_is_reserved_unit(&heapmgt->vmheap, stack_reserve_end, reserve_size)) {
known_stack = true;
available = reserve_size;
} else
known_stack = false;
} else
#elif defined(UNIX)
/* the all_memory_areas list doesn't keep details inside vmheap */
known_stack = get_memory_info_from_os(stack_reserve_end, NULL,
&available, &prot);
#else
known_stack = get_memory_info(stack_reserve_end, NULL, &available, &prot);
#endif
/* If ever out of vmheap, then may have free space beyond stack,
* which we could support but don't (see FIXME above) */
ASSERT(out_of_vmheap_once ||
(known_stack && available >= reserve_size && prot == 0));
}
if (!known_stack ||
/* if -no_vm_reserve will short-circuit so no vmh deref danger */
!vmm_in_same_block(dcontext->dstack,
/* we do want a guard page at the end */
stack_reserve_end + reserve_size) ||
available < reserve_size) {
ASSERT(!DYNAMO_OPTION(stack_shares_gencode) ||
!DYNAMO_OPTION(vm_reserve) || out_of_vmheap_once);
DOLOG(1, LOG_HEAP, {
if (known_stack && available < reserve_size) {
LOG(GLOBAL, LOG_HEAP, 1,
"heap_mmap_reserve_post_stack: avail %d < needed %d\n",
available, reserve_size);
}
});
STATS_INC(mmap_no_share_stack_region);
return heap_mmap_reserve(reserve_size, commit_size);
}
ASSERT(DYNAMO_OPTION(vm_reserve));
ASSERT(stack_reserve_end != NULL);
prot = MEMPROT_EXEC|MEMPROT_READ|MEMPROT_WRITE;
/* memory alloc/dealloc and updating DR list must be atomic */
dynamo_vm_areas_lock(); /* if already hold lock this is a nop */
/* We share the stack's end guard page as our start guard page */
if (vmm_is_reserved_unit(&heapmgt->vmheap, stack_reserve_end, reserve_size)) {
/* Memory is already reserved with OS */
p = stack_reserve_end;
} else {
p = os_heap_reserve(stack_reserve_end, reserve_size, &error_code, true/*+x*/);
#ifdef X64
/* ensure future heap allocations are reachable from this allocation
* (this will also verify that this region meets reachability requirements) */
if (p != NULL)
request_region_be_heap_reachable(p, reserve_size);
#endif
if (p == NULL) {
ASSERT_NOT_REACHED();
LOG(GLOBAL, LOG_HEAP, 1,
"heap_mmap_reserve_post_stack: reserve failed "PFX"\n", error_code);
dynamo_vm_areas_unlock();
STATS_INC(mmap_no_share_stack_region);
return heap_mmap_reserve(reserve_size, commit_size);
}
ASSERT(error_code == HEAP_ERROR_SUCCESS);
}
if (!vmm_heap_commit(p, commit_size, prot, &error_code)) {
ASSERT_NOT_REACHED();
LOG(GLOBAL, LOG_HEAP, 1, "heap_mmap_reserve_post_stack: commit failed "PFX"\n",
error_code);
if (!vmm_is_reserved_unit(&heapmgt->vmheap, stack_reserve_end, reserve_size)) {
os_heap_free(p, reserve_size, &error_code);
ASSERT(error_code == HEAP_ERROR_SUCCESS);
}
dynamo_vm_areas_unlock();
STATS_INC(mmap_no_share_stack_region);
return heap_mmap_reserve(reserve_size, commit_size);
}
account_for_memory(p, reserve_size, prot, true/*add now*/, false
_IF_DEBUG("heap_mmap_reserve_post_stack"));
dynamo_vm_areas_unlock();
/* We rely on this for freeing in absence of dcontext */
ASSERT((ptr_uint_t)p - GUARD_PAGE_ADJUSTMENT/2 !=
ALIGN_BACKWARD(p, VMM_BLOCK_SIZE));
#ifdef DEBUG_MEMORY
memset(p, HEAP_ALLOCATED_BYTE, commit_size);
#endif
LOG(GLOBAL, LOG_HEAP, 2, "heap_mmap w/ stack: %d bytes [/ %d] @ "PFX"\n",
commit_size, reserve_size, p);
STATS_ADD_PEAK(mmap_capacity, commit_size);
STATS_ADD_PEAK(mmap_reserved_only, (reserve_size - commit_size));
STATS_INC(mmap_share_stack_region);
return p;
}
/* De-commits memory that was allocated in the same allocation region as this
* thread's stack (case 9474).
*/
void
heap_munmap_post_stack(dcontext_t *dcontext, void *p, size_t reserve_size)
{
/* We would require a valid dcontext and compare to the stack reserve end,
* but on detach we have no dcontext, so we instead use block alignment.
*/
DOCHECK(1, {
if (dcontext != NULL && dcontext != GLOBAL_DCONTEXT &&
DYNAMO_OPTION(vm_reserve) && DYNAMO_OPTION(stack_shares_gencode)) {
bool at_stack_end = (p == dcontext->dstack + GUARD_PAGE_ADJUSTMENT/2);
bool at_block_start = ((ptr_uint_t)p - GUARD_PAGE_ADJUSTMENT/2 ==
ALIGN_BACKWARD(p, VMM_BLOCK_SIZE));
ASSERT((at_stack_end && !at_block_start) ||
(!at_stack_end && at_block_start));
}
});
if (!DYNAMO_OPTION(vm_reserve) ||
!DYNAMO_OPTION(stack_shares_gencode) ||
(ptr_uint_t)p - GUARD_PAGE_ADJUSTMENT/2 == ALIGN_BACKWARD(p, VMM_BLOCK_SIZE)) {
heap_munmap(p, reserve_size);
} else {
/* Detach makes it a pain to pass in the commit size so
* we use the reserve size, which works fine.
*/
release_memory_and_update_areas((app_pc)p, reserve_size, true/*decommit*/,
true/*update now*/);
LOG(GLOBAL, LOG_HEAP, 2, "heap_munmap_post_stack: %d bytes @ "PFX"\n",
reserve_size, p);
STATS_SUB(mmap_capacity, reserve_size);
STATS_SUB(mmap_reserved_only, reserve_size);
}
}
/* use heap_mmap to allocate large chunks of executable memory
* it's mainly used to allocate our fcache units
*/
void *
heap_mmap(size_t size)
{
return heap_mmap_reserve(size, size);
}
/* free memory-mapped storage */
void
heap_munmap_ex(void *p, size_t size, bool guarded)
{
#ifdef DEBUG_MEMORY
/* can't set to HEAP_UNALLOCATED_BYTE since really not in our address
* space anymore */
#endif
release_guarded_real_memory((vm_addr_t)p, size, true/*update DR areas immediately*/,
guarded);
DOSTATS({
/* avoid problem w/ being called by cleanup_and_terminate after dynamo_process_exit */
if (!dynamo_exited_log_and_stats) {
LOG(GLOBAL, LOG_HEAP, 2, "heap_munmap: %d bytes @ "PFX"\n", size, p);
STATS_SUB(mmap_capacity, size);
STATS_SUB(mmap_reserved_only, size);
}
});
}
/* free memory-mapped storage */
void
heap_munmap(void *p, size_t size)
{
heap_munmap_ex(p, size, true/*guarded*/);
}
#ifdef STACK_GUARD_PAGE
# define STACK_GUARD_PAGES 1
#endif
/* use stack_alloc to build a stack -- it returns TOS
* For STACK_GUARD_PAGE, it also marks the bottom STACK_GUARD_PAGES==1
* to detect overflows when used.
*/
void *
stack_alloc(size_t size)
{
void *p;
/* we reserve and commit at once for now
* FIXME case 2330: commit-on-demand could allow larger max sizes w/o
* hurting us in the common case
*/
p = get_guarded_real_memory(size, size, MEMPROT_READ|MEMPROT_WRITE, true, true
_IF_DEBUG("stack_alloc"));
#ifdef DEBUG_MEMORY
memset(p, HEAP_ALLOCATED_BYTE, size);
#endif
#ifdef STACK_GUARD_PAGE
/* mark the bottom page non-accessible to trap stack overflow */
/* NOTE: the guard page should be included in the total memory requested */
# ifdef WINDOWS
mark_page_as_guard((byte *)p + ((STACK_GUARD_PAGES - 1) * PAGE_SIZE));
# else
/* FIXME: make no access, not just no write -- and update signal.c to
* look at reads and not just writes -- though unwritable is nearly as good
*/
# if defined(CLIENT_INTERFACE) || defined(STANDALONE_UNIT_TEST)
if (!standalone_library)
# endif
make_unwritable(p, STACK_GUARD_PAGES * PAGE_SIZE);
# endif
#endif
STATS_ADD(stack_capacity, size);
STATS_MAX(peak_stack_capacity, stack_capacity);
/* stack grows from high to low */
return (void *) ((ptr_uint_t)p + size);
}
/* free stack storage */
void
stack_free(void *p, size_t size)
{
if (size == 0)
size = DYNAMORIO_STACK_SIZE;
p = (void *) ((vm_addr_t)p - size);
release_guarded_real_memory((vm_addr_t)p, size, true/*update DR areas immediately*/,
true);
DOSTATS({
if (!dynamo_exited_log_and_stats)
STATS_SUB(stack_capacity, size);
});
}
#ifdef STACK_GUARD_PAGE
/* only checks initstack and current dcontext
* does not check any dstacks on the callback stack (win32) */
bool
is_stack_overflow(dcontext_t *dcontext, byte *sp)
{
/* ASSUMPTION: size of stack is DYNAMORIO_STACK_SIZE = dynamo_options.stack_size
* currently sideline violates that for a thread stack
* but all dstacks and initstack should be this size
*/
byte *bottom = dcontext->dstack - DYNAMORIO_STACK_SIZE;
/* see if in bottom guard page of dstack */
if (sp >= bottom && sp < bottom + (STACK_GUARD_PAGES * PAGE_SIZE))
return true;
/* now check the initstack */
bottom = initstack - DYNAMORIO_STACK_SIZE;
if (sp >= bottom && sp < bottom + (STACK_GUARD_PAGES * PAGE_SIZE))
return true;
return false;
}
#endif
byte *
map_file(file_t f, size_t *size INOUT, uint64 offs, app_pc addr, uint prot,
map_flags_t map_flags)
{
byte *view;
/* memory alloc/dealloc and updating DR list must be atomic */
dynamo_vm_areas_lock(); /* if already hold lock this is a nop */
view = os_map_file(f, size, offs, addr, prot, map_flags);
if (view != NULL) {
STATS_ADD_PEAK(file_map_capacity, *size);
account_for_memory((void *)view, *size, prot, true/*add now*/, true/*image*/
_IF_DEBUG("map_file"));
}
dynamo_vm_areas_unlock();
return view;
}
bool
unmap_file(byte *map, size_t size)
{
bool success;
ASSERT(map != NULL && ALIGNED(map, PAGE_SIZE));
size = ALIGN_FORWARD(size, PAGE_SIZE);
/* memory alloc/dealloc and updating DR list must be atomic */
dynamo_vm_areas_lock(); /* if already hold lock this is a nop */
success = os_unmap_file(map, size);
if (success) {
/* Only update the all_memory_areas on success.
* It should still be atomic to the outside observers.
*/
update_dynamo_areas_on_release(map, map+size, true/*remove now*/);
STATS_SUB(file_map_capacity, size);
}
dynamo_vm_areas_unlock();
return success;
}
/* We cannot incrementally keep dynamo vm area list up to date due to
* circular dependencies bet vmareas and global heap (trust me, I've tried
* to support it with reentrant routines and recursive locks, the hard part
* is getting add_vm_area to be reentrant or to queue up adding areas,
* I think this solution is much more elegant, plus it avoids race conditions
* between DR memory allocation and the vmareas list by ensuring the list
* is up to date at the exact time of each query).
* Instead we on-demand walk the units.
* Freed units can usually be removed incrementally, except when we
* hold the heap_unit_lock when we run out of memory -- when we set
* a flag telling the caller of this routine to remove all heap areas
* from the vm list prior to calling us to add the real ones back in.
* Re-adding everyone is the simplest policy, so we don't have to keep
* track of who's been added.
* The caller is assumed to hold the dynamo vm areas write lock.
*/
void
heap_vmareas_synch_units()
{
heap_unit_t *u, *next;
/* make sure to add guard page on each side, as well */
uint offs = (dynamo_options.guard_pages) ? PAGE_SIZE : 0;
/* we again have circular dependence w/ vmareas if it happens to need a
* new unit in the course of adding these areas, so we use a recursive lock!
* furthermore, we need to own the global lock now, to avoid deadlock with
* another thread who does global_alloc and then needs a new unit!
* which means that the global_alloc lock must be recursive since vmareas
* may need to global_alloc...
*/
/* if chance could own both locks, must grab both now
* always grab global_alloc first, then we won't have deadlocks
*/
acquire_recursive_lock(&global_alloc_lock);
acquire_recursive_lock(&heap_unit_lock);
if (dynamo_areas_pending_remove) {
dynamo_areas_pending_remove = false;
remove_dynamo_heap_areas();
/* When heap units are removed from the dynamo_area, they should be
* marked so. See case 4196.
*/
for (u = heapmgt->heap.units; u != NULL; u = u->next_global)
u->in_vmarea_list = false;
for (u = heapmgt->heap.dead; u != NULL; u = u->next_global)
u->in_vmarea_list = false;
}
for (u = heapmgt->heap.units; u != NULL; u = next) {
app_pc start = (app_pc)u - offs;
/* support un-aligned heap reservation end: PR 415269 (though as
* part of that PR we shouldn't have un-aligned anymore)
*/
app_pc end_align = (app_pc) ALIGN_FORWARD(UNIT_RESERVED_END(u), PAGE_SIZE);
app_pc end = end_align + offs;
/* u can be moved to dead list, so cache the next link; case 4196. */
next = u->next_global;
/* case 3045: areas inside the vmheap reservation are not added to the list */
if (!u->in_vmarea_list && !vmm_is_reserved_unit(&heapmgt->vmheap,
start, end - start)) {
/* case 4196 if next is used by dynamo_vmareas then next
* may become dead if vector is resized, then u should be
* alive and u->next_global should be reset AFTER add */
bool next_may_die =
is_dynamo_area_buffer(UNIT_GET_START_PC(next)); /* keep breaking abstractions */
/* dynamo_areas.buf vector may get resized and u can either
* go to the dead unit list, or it can be released back to
* the OS. We'll mark it as being in vmarea list to avoid
* re-adding when going through dead one's, and we'll mark
* _before_ the potential free. If dynamo_areas.buf is
* freed back to the OS we'll have another iteration in
* update_dynamo_vm_areas() until we get fully
* synchronized, so we don't need to worry about the
* inconsistency.
*/
u->in_vmarea_list = true;
add_dynamo_heap_vm_area(start, end, true, false _IF_DEBUG("heap unit"));
/* NOTE: Since we could mark_dynamo_vm_areas_stale instead of adding to
* it, we may lose prot info about this unit.
* FIXME: Currently, this is done only at one place, which allocates unit
* as MEMPROT_READ | MEMPROT_WRITE. If other places are added, then this
* needs to change.
*/
update_all_memory_areas((app_pc)u, end_align,
MEMPROT_READ | MEMPROT_WRITE,
DR_MEMTYPE_DATA); /* unit */
if (offs != 0) {
/* guard pages */
update_all_memory_areas((app_pc)u - offs, (app_pc)u, MEMPROT_NONE,
DR_MEMTYPE_DATA);
update_all_memory_areas(end_align, end, MEMPROT_NONE,
DR_MEMTYPE_DATA);
}
if (next_may_die) {
STATS_INC(num_vmareas_resize_synch);
/* if next was potentially on dead row, then current
* should still be live and point to the next live
*/
next = u->next_global;
}
}
}
for (u = heapmgt->heap.dead; u != NULL; u = next) {
app_pc start = (app_pc)u - offs;
/* support un-aligned heap reservation end: PR 415269 (though as
* part of that PR we shouldn't have un-aligned anymore)
*/
app_pc end_align = (app_pc) ALIGN_FORWARD(UNIT_RESERVED_END(u), PAGE_SIZE);
app_pc end = end_align + offs;
/* u can be moved to live list, so cache the next link; case 4196. */
next = u->next_global;
/* case 3045: areas inside the vmheap reservation are not added to the list */
if (!u->in_vmarea_list && !vmm_is_reserved_unit(&heapmgt->vmheap,
start, end - start)) {
u->in_vmarea_list = true;
add_dynamo_heap_vm_area(start, end, true, false _IF_DEBUG("dead heap unit"));
update_all_memory_areas((app_pc)u, end_align,
MEMPROT_READ | MEMPROT_WRITE,
DR_MEMTYPE_DATA); /* unit */
if (offs != 0) {
/* guard pages */
update_all_memory_areas(start, (app_pc)u, MEMPROT_NONE,
DR_MEMTYPE_DATA);
update_all_memory_areas(end_align, end, MEMPROT_NONE,
DR_MEMTYPE_DATA);
}
/* case 4196 if next was put back on live list for
* dynamo_areas.buf vector, then next will no longer be a
* valid iterator over dead list
*/
if (is_dynamo_area_buffer(UNIT_GET_START_PC(next))) { /* keep breaking abstractions */
STATS_INC(num_vmareas_resize_synch);
ASSERT_NOT_TESTED();
next = u->next_global;
}
}
}
release_recursive_lock(&heap_unit_lock);
release_recursive_lock(&global_alloc_lock);
}
/* shared between global and global_unprotected */
static void *
common_global_heap_alloc(thread_units_t *tu, size_t size HEAPACCT(which_heap_t which))
{
void *p;
acquire_recursive_lock(&global_alloc_lock);
p = common_heap_alloc(tu, size HEAPACCT(which));
release_recursive_lock(&global_alloc_lock);
if (p == NULL) {
/* circular dependence solution: we need to hold DR lock before
* global alloc lock -- so we back out, grab it, and retry
*/
dynamo_vm_areas_lock();
acquire_recursive_lock(&global_alloc_lock);
p = common_heap_alloc(tu, size HEAPACCT(which));
release_recursive_lock(&global_alloc_lock);
dynamo_vm_areas_unlock();
}
ASSERT(p != NULL);
return p;
}
/* shared between global and global_unprotected */
static void
common_global_heap_free(thread_units_t *tu, void *p, size_t size HEAPACCT(which_heap_t which))
{
bool ok;
if (p == NULL) {
ASSERT(false && "attempt to free NULL");
return;
}
acquire_recursive_lock(&global_alloc_lock);
ok = common_heap_free(tu, p, size HEAPACCT(which));
release_recursive_lock(&global_alloc_lock);
if (!ok) {
/* circular dependence solution: we need to hold DR lock before
* global alloc lock -- so we back out, grab it, and retry
*/
dynamo_vm_areas_lock();
acquire_recursive_lock(&global_alloc_lock);
ok = common_heap_free(tu, p, size HEAPACCT(which));
release_recursive_lock(&global_alloc_lock);
dynamo_vm_areas_unlock();
}
ASSERT(ok);
}
/* these functions use the global heap instead of a thread's heap: */
void *
global_heap_alloc(size_t size HEAPACCT(which_heap_t which))
{
void *p = common_global_heap_alloc(&heapmgt->global_units, size HEAPACCT(which));
ASSERT(p != NULL);
LOG(GLOBAL, LOG_HEAP, 6, "\nglobal alloc: "PFX" (%d bytes)\n", p, size);
return p;
}
void
global_heap_free(void *p, size_t size HEAPACCT(which_heap_t which))
{
common_global_heap_free(&heapmgt->global_units, p, size HEAPACCT(which));
LOG(GLOBAL, LOG_HEAP, 6, "\nglobal free: "PFX" (%d bytes)\n", p, size);
}
/* reallocate area
allocates new_num elements of element_size
if ptr is NULL acts like global_heap_alloc,
copies an old_num elements of given size in the new area */
/* FIXME: do a heap_realloc and a special_heap_realloc too */
void *
global_heap_realloc(void *ptr, size_t old_num, size_t new_num, size_t element_size
HEAPACCT(which_heap_t which))
{
void *new_area = global_heap_alloc(new_num * element_size HEAPACCT(which));
if (ptr) {
memcpy(new_area, ptr, (old_num < new_num ? old_num : new_num) * element_size);
global_heap_free(ptr, old_num * element_size HEAPACCT(which));
}
return new_area;
}
/* size does not include guard pages (if any) and is reserved, but only
* DYNAMO_OPTION(heap_commit_increment) is committed up front
*/
static heap_unit_t *
heap_create_unit(thread_units_t *tu, size_t size, bool must_be_new)
{
heap_unit_t *u = NULL, *dead = NULL, *prev_dead = NULL;
bool new_unit = false;
/* we do not restrict size to unit max as we have to make larger-than-max
* units for oversized requests
*/
/* modifying heap list and DR areas must be atomic, and must grab
* DR area lock before heap_unit_lock
*/
ASSERT(safe_to_allocate_or_free_heap_units());
dynamo_vm_areas_lock();
/* take from dead list if possible */
acquire_recursive_lock(&heap_unit_lock);
/* FIXME: need to unprotect units that we're going to perform
* {next,prev}_global assignments too -- but need to know whether
* to re-protect -- do all at once, or each we need? add a writable
* flag to heap_unit_t?
*/
if (!must_be_new) {
for (dead = heapmgt->heap.dead;
dead != NULL && UNIT_RESERVED_SIZE(dead) < size;
prev_dead = dead, dead = dead->next_global)
;
}
if (dead != NULL) {
if (prev_dead == NULL)
heapmgt->heap.dead = dead->next_global;
else
prev_dead->next_global = dead->next_global;
u = dead;
heapmgt->heap.num_dead--;
RSTATS_DEC(heap_num_free);
release_recursive_lock(&heap_unit_lock);
LOG(GLOBAL, LOG_HEAP, 2,
"Re-using dead heap unit: "PFX"-"PFX" %d KB (need %d KB)\n",
u, ((byte*)u)+size, UNIT_RESERVED_SIZE(u)/1024, size/1024);
} else {
size_t commit_size = DYNAMO_OPTION(heap_commit_increment);
release_recursive_lock(&heap_unit_lock); /* do not hold while asking for memory */
/* create new unit */
ASSERT(commit_size <= size);
u = (heap_unit_t *)
get_guarded_real_memory(size, commit_size, MEMPROT_READ|MEMPROT_WRITE,
false, true _IF_DEBUG(""));
new_unit = true;
/* FIXME: handle low memory conditions by freeing units, + fcache units? */
ASSERT(u);
LOG(GLOBAL, LOG_HEAP, 2, "New heap unit: "PFX"-"PFX"\n", u, ((byte*)u)+size);
/* u is kept at top of unit itself, so displace start pc */
u->start_pc = (heap_pc) (((ptr_uint_t)u) + sizeof(heap_unit_t));
u->end_pc = ((heap_pc)u) + commit_size;
u->reserved_end_pc = ((heap_pc)u) + size;
u->in_vmarea_list = false;
STATS_ADD(heap_capacity, commit_size);
STATS_MAX(peak_heap_capacity, heap_capacity);
/* FIXME: heap sizes are not always page-aligned so stats will be off */
STATS_ADD_PEAK(heap_reserved_only, (u->reserved_end_pc - u->end_pc));
}
RSTATS_ADD_PEAK(heap_num_live, 1);
u->cur_pc = u->start_pc;
u->next_local = NULL;
DODEBUG({
u->id = tu->num_units;
tu->num_units++;
});
acquire_recursive_lock(&heap_unit_lock);
u->next_global = heapmgt->heap.units;
if (heapmgt->heap.units != NULL)
heapmgt->heap.units->prev_global = u;
u->prev_global = NULL;
heapmgt->heap.units = u;
release_recursive_lock(&heap_unit_lock);
dynamo_vm_areas_unlock();
#ifdef DEBUG_MEMORY
DOCHECK(CHKLVL_MEMFILL,
memset(u->start_pc, HEAP_UNALLOCATED_BYTE, u->end_pc - u->start_pc););
#endif
return u;
}
/* dcontext only used to determine whether a global unit or not */
static void
heap_free_unit(heap_unit_t *unit, dcontext_t *dcontext)
{
heap_unit_t *u, *prev_u;
#ifdef DEBUG_MEMORY
/* Unit should already be set to all HEAP_UNALLOCATED by the individual
* frees and the free list cleanup, verify. */
/* NOTE - this assert fires if any memory in the unit wasn't freed. This
* would include memory allocated ACCT_TOMBSTONE (which we don't currently
* use). Using ACCT_TOMBSTONE is dangerous since we will still free the
* unit here (say at proc or thread exit) even if there are ACCT_TOMBSTONE
* allocations in it. */
/* Note, this memset check is done only on the special heap unit header,
* not on the unit itself - FIXME: case 10434. Maybe we should embed the
* special heap unit header in the first special heap unit itself. */
/* The hotp_only leak relaxation below is for case 9588 &a