diff options
author | Rickard Green <[email protected]> | 2013-08-23 17:29:58 +0200 |
---|---|---|
committer | Sverker Eriksson <[email protected]> | 2013-09-30 17:45:44 +0200 |
commit | ef3da907bd566b43a4022f1cbb1ae3d103b9ec3e (patch) | |
tree | 4268ac511df4754087bccb80367466fd2eea3492 /erts/emulator/sys | |
parent | c2dbcb69929ac18e7687f1df1de6613b34e2897b (diff) | |
download | otp-ef3da907bd566b43a4022f1cbb1ae3d103b9ec3e.tar.gz otp-ef3da907bd566b43a4022f1cbb1ae3d103b9ec3e.tar.bz2 otp-ef3da907bd566b43a4022f1cbb1ae3d103b9ec3e.zip |
erts: erts_mmap supercarrier management and erts_mseg usage
* Coalescing and trimming of free segments in supercarrier
* Management of super aligned and super unaligned areas in
supercarrier
* Management of reservation of physical memory
* erts_mseg usage of erts_mmap
Diffstat (limited to 'erts/emulator/sys')
-rw-r--r-- | erts/emulator/sys/common/erl_mmap.c | 1120 | ||||
-rw-r--r-- | erts/emulator/sys/common/erl_mmap.h | 89 | ||||
-rw-r--r-- | erts/emulator/sys/common/erl_mseg.c | 860 | ||||
-rw-r--r-- | erts/emulator/sys/common/erl_mseg.h | 34 |
4 files changed, 1354 insertions, 749 deletions
diff --git a/erts/emulator/sys/common/erl_mmap.c b/erts/emulator/sys/common/erl_mmap.c index a16ee7ae39..aac01bf93c 100644 --- a/erts/emulator/sys/common/erl_mmap.c +++ b/erts/emulator/sys/common/erl_mmap.c @@ -20,11 +20,57 @@ # include "config.h" #endif -#include <sys.h> -#include <erl_mmap.h> +#include "sys.h" #include <stddef.h> +#include "erl_smp.h" +#include "erl_mmap.h" -#ifdef DEBUG +#if defined(DEBUG) || 0 +# undef ERTS_MMAP_DEBUG +# define ERTS_MMAP_DEBUG +#endif + +/* #define ERTS_MMAP_DEBUG_FILL_AREAS */ + +#ifdef ERTS_MMAP_DEBUG +# define ERTS_MMAP_ASSERT(A) \ + ((void) (!(A) \ + ? erts_mmap_assert_failed(#A, __func__, __FILE__, __LINE__)\ + : 1)) +static int +erts_mmap_assert_failed(const char *a, const char *func, const char *file, int line) +{ + erts_fprintf(stderr, "%s:%d:%s() Assertion failed: %s\n", + (char *) file, line, (char *) func, (char *) a); + abort(); + return 0; +} +#else +# define ERTS_MMAP_ASSERT(A) ((void) 1) +#endif + +/* + * `mmap_state.sa.bot` and `mmap_state.sua.top` are read only after + * initialization, but the other pointers are not; i.e., only + * ERTS_MMAP_IN_SUPERCARRIER() is allowed without the mutex held. + */ +#define ERTS_MMAP_IN_SUPERCARRIER(PTR) \ + (((UWord) (PTR)) - ((UWord) mmap_state.sa.bot) \ + < ((UWord) mmap_state.sua.top) - ((UWord) mmap_state.sa.bot)) +#define ERTS_MMAP_IN_SUPERALIGNED_AREA(PTR) \ + (ERTS_SMP_LC_ASSERT(erts_lc_mtx_is_locked(&mmap_state.mtx)), \ + (((UWord) (PTR)) - ((UWord) mmap_state.sa.bot) \ + < ((UWord) mmap_state.sa.top) - ((UWord) mmap_state.sa.bot))) +#define ERTS_MMAP_IN_SUPERUNALIGNED_AREA(PTR) \ + (ERTS_SMP_LC_ASSERT(erts_lc_mtx_is_locked(&mmap_state.mtx)), \ + (((UWord) (PTR)) - ((UWord) mmap_state.sua.bot) \ + < ((UWord) mmap_state.sua.top) - ((UWord) mmap_state.sua.bot))) + +int erts_have_erts_mmap; +UWord erts_page_inv_mask; + +#if defined(DEBUG) || defined(ERTS_MMAP_DEBUG) +# undef RBT_DEBUG # define RBT_DEBUG #endif #ifdef RBT_DEBUG @@ -60,7 +106,7 @@ typedef struct { #define SET_RED(N) ((N)->flags |= RED_FLG) #define SET_BLACK(N) ((N)->flags &= ~RED_FLG) -#define HARD_DEBUG /*SVERK*/ +/* #define HARD_DEBUG */ #ifdef HARD_DEBUG # define HARD_CHECK_IS_MEMBER(ROOT,NODE) rbt_assert_is_member(ROOT,NODE) # define HARD_CHECK_TREE(TREE,SZ) check_tree(TREE, SZ) @@ -79,6 +125,134 @@ typedef struct { char* end; }ErtsFreeSegDesc; +typedef struct { + RBTree stree; + RBTree atree; +}ErtsFreeSegMap; + +static struct { + int (*reserve_physical)(char *, UWord); + void (*unreserve_physical)(char *, UWord); + int supercarrier; + int no_os_mmap; + /* + * Super unaligend area is located above super aligned + * area. That is, `sa.bot` is beginning of the super + * carrier, `sua.top` is the end of the super carrier, + * and sa.top and sua.bot moves towards eachother. + */ + struct { + char *top; + char *bot; + ErtsFreeSegMap map; + } sua; + struct { + char *top; + char *bot; + ErtsFreeSegMap map; + } sa; +#if HAVE_MMAP && (!defined(MAP_ANON) && !defined(MAP_ANONYMOUS)) + int mmap_fd; +#endif + erts_smp_mtx_t mtx; + char *desc_free_list; + struct { + struct { + UWord total; + struct { + UWord total; + UWord sa; + UWord sua; + } used; + } supercarrier; + struct { + UWord used; + } os; + } size; +} mmap_state; + +#define ERTS_MMAP_SIZE_SC_SA_INC(SZ) \ + do { \ + mmap_state.size.supercarrier.used.total += (SZ); \ + mmap_state.size.supercarrier.used.sa += (SZ); \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.total \ + <= mmap_state.size.supercarrier.total); \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.sa \ + <= mmap_state.size.supercarrier.used.total); \ + } while (0) +#define ERTS_MMAP_SIZE_SC_SA_DEC(SZ) \ + do { \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.total >= (SZ)); \ + mmap_state.size.supercarrier.used.total -= (SZ); \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.sa >= (SZ)); \ + mmap_state.size.supercarrier.used.sa -= (SZ); \ + } while (0) +#define ERTS_MMAP_SIZE_SC_SUA_INC(SZ) \ + do { \ + mmap_state.size.supercarrier.used.total += (SZ); \ + mmap_state.size.supercarrier.used.sua += (SZ); \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.total \ + <= mmap_state.size.supercarrier.total); \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.sua \ + <= mmap_state.size.supercarrier.used.total); \ + } while (0) +#define ERTS_MMAP_SIZE_SC_SUA_DEC(SZ) \ + do { \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.total >= (SZ)); \ + mmap_state.size.supercarrier.used.total -= (SZ); \ + ERTS_MMAP_ASSERT(mmap_state.size.supercarrier.used.sua >= (SZ)); \ + mmap_state.size.supercarrier.used.sua -= (SZ); \ + } while (0) +#define ERTS_MMAP_SIZE_OS_INC(SZ) \ + do { \ + ERTS_MMAP_ASSERT(mmap_state.size.os.used + (SZ) >= (SZ)); \ + mmap_state.size.os.used += (SZ); \ + } while (0) +#define ERTS_MMAP_SIZE_OS_DEC(SZ) \ + do { \ + ERTS_MMAP_ASSERT(mmap_state.size.os.used >= (SZ)); \ + mmap_state.size.os.used -= (SZ); \ + } while (0) + +static void +add_free_desc_area(char *start, char *end) +{ + if (end > start && sizeof(ErtsFreeSegDesc) <= end - start) { + ErtsFreeSegDesc *prev_desc, *desc; + char *desc_end; + + prev_desc = (ErtsFreeSegDesc *) start; + prev_desc->start = mmap_state.desc_free_list; + desc = (ErtsFreeSegDesc *) (start + sizeof(ErtsFreeSegDesc)); + desc_end = start + 2*sizeof(ErtsFreeSegDesc); + + while (desc_end <= end) { + desc->start = (char *) prev_desc; + prev_desc = desc; + desc = (ErtsFreeSegDesc *) desc_end; + desc_end += sizeof(ErtsFreeSegDesc); + } + mmap_state.desc_free_list = (char *) prev_desc; + } +} + +static ERTS_INLINE ErtsFreeSegDesc * +alloc_desc(void) +{ + ErtsFreeSegDesc *res; + res = (ErtsFreeSegDesc *) mmap_state.desc_free_list; + if (res) + mmap_state.desc_free_list = res->start; + return res; +} + +static ERTS_INLINE void +free_desc(ErtsFreeSegDesc *desc) +{ + desc->start = mmap_state.desc_free_list; + mmap_state.desc_free_list = (char *) desc; +} + static ERTS_INLINE ErtsFreeSegDesc* anode_to_desc(RBTNode* anode) { return (ErtsFreeSegDesc*) ((char*)anode - offsetof(ErtsFreeSegDesc, anode)); @@ -94,12 +268,6 @@ static ERTS_INLINE ErtsFreeSegDesc* node_to_desc(enum SortOrder order, RBTNode* return order==ADDR_ORDER ? anode_to_desc(node) : snode_to_desc(node); } -typedef struct { - RBTree stree; - RBTree atree; -}ErtsFreeSegMap; - - #ifdef HARD_DEBUG static ERTS_INLINE SWord cmp_blocks(enum SortOrder order, RBTNode* lhs, RBTNode* rhs) @@ -671,19 +839,927 @@ static ErtsFreeSegDesc* lookup_free_seg(ErtsFreeSegMap* map, SWord need_sz) return best_desc; } +#if ERTS_HAVE_OS_MMAP +/* Implementation of os_mmap()/os_munmap()/os_mremap()... */ + +#if HAVE_MMAP +# define ERTS_MMAP_PROT (PROT_READ|PROT_WRITE) +# if defined(MAP_ANONYMOUS) +# define ERTS_MMAP_FLAGS (MAP_ANON|MAP_PRIVATE) +# define ERTS_MMAP_FD (-1) +# elif defined(MAP_ANON) +# define ERTS_MMAP_FLAGS (MAP_ANON|MAP_PRIVATE) +# define ERTS_MMAP_FD (-1) +# else +# define ERTS_MMAP_FLAGS (MAP_PRIVATE) +# define ERTS_MMAP_FD mmap_state.mmap_fd +# endif +#endif -void erts_mmap_init(ErtsMMapInit* init) +static ERTS_INLINE void * +os_mmap(UWord size, int try_superalign) { -#ifdef HARD_DEBUG - erts_fprintf(stderr, "SVERK: scs = %bpu\n", init->scs); - erts_fprintf(stderr, "SVERK: sco = %i\n", init->sco); - erts_fprintf(stderr, "SVERK: scmgc = %i\n", init->scmgc); +#if HAVE_MMAP + void *res; +#ifdef MAP_ALIGN + if (try_superalign) + res = mmap((void *) ERTS_SUPERALIGNED_SIZE, size, ERTS_MMAP_PROT, + ERTS_MMAP_FLAGS|MAP_ALIGN, ERTS_MMAP_FD, 0); + else +#endif + res = mmap((void *) 0, size, ERTS_MMAP_PROT, + ERTS_MMAP_FLAGS, ERTS_MMAP_FD, 0); + if (res == MAP_FAILED) + return NULL; + return res; +#elif HAVE_VIRTUALALLOC + return (void *) VirtualAlloc(NULL, (SIZE_T) size, + MEM_COMMIT|MEM_RESERVE, PAGE_READWRITE); +#else +# error "missing mmap() or similar" +#endif +} + +static ERTS_INLINE void +os_munmap(void *ptr, UWord size) +{ +#if HAVE_MMAP +#ifdef ERTS_MMAP_DEBUG + int res = +#endif + munmap(ptr, size); + ERTS_MMAP_ASSERT(res == 0); +#elif HAVE_VIRTUALALLOC +#ifdef DEBUG + BOOL res = +#endif + VirtualFree((LPVOID) ptr, (SIZE_T) 0, MEM_RELEASE); + ERTS_MMAP_ASSERT(res != 0); +#else +# error "missing munmap() or similar" +#endif +} + +#ifdef ERTS_HAVE_OS_MREMAP +# if HAVE_MREMAP +# if defined(__NetBSD__) +# define ERTS_MREMAP_FLAGS (0) +# else +# define ERTS_MREMAP_FLAGS (MREMAP_MAYMOVE) +# endif +# endif +static ERTS_INLINE void * +os_mremap(void *ptr, UWord old_size, UWord new_size, int try_superalign) +{ + void *new_seg; +#if HAVE_MREMAP + new_seg = mremap(ptr, (size_t) old_size, +# if defined(__NetBSD__) + NULL, +# endif + (size_t) new_size, ERTS_MREMAP_FLAGS); + if (new_seg == (void *) MAP_FAILED) + return NULL; + return new_seg; +#else +# error "missing mremap() or similar" +#endif +} +#endif + +#ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION +#if HAVE_MMAP + +#define ERTS_MMAP_RESERVE_PROT (ERTS_MMAP_PROT) +#define ERTS_MMAP_RESERVE_FLAGS (ERTS_MMAP_FLAGS|MAP_FIXED) +#define ERTS_MMAP_UNRESERVE_PROT (PROT_NONE) +#define ERTS_MMAP_UNRESERVE_FLAGS (ERTS_MMAP_FLAGS|MAP_NORESERVE|MAP_FIXED) +#define ERTS_MMAP_VIRTUAL_PROT (PROT_NONE) +#define ERTS_MMAP_VIRTUAL_FLAGS (ERTS_MMAP_FLAGS|MAP_NORESERVE) + +static int +os_reserve_physical(char *ptr, UWord size) +{ + void *res = mmap((void *) ptr, (size_t) size, ERTS_MMAP_RESERVE_PROT, + ERTS_MMAP_RESERVE_FLAGS, ERTS_MMAP_FD, 0); + if (res == (void *) MAP_FAILED) + return 0; + return 1; +} + +static void +os_unreserve_physical(char *ptr, UWord size) +{ + void *res = mmap((void *) ptr, (size_t) size, ERTS_MMAP_UNRESERVE_PROT, + ERTS_MMAP_UNRESERVE_FLAGS, ERTS_MMAP_FD, 0); + if (res == (void *) MAP_FAILED) + erl_exit(ERTS_ABORT_EXIT, "Failed to unreserve memory"); +} + +static void * +os_mmap_virtual(char *ptr, UWord size) +{ + void *res = mmap((void *) ptr, (size_t) size, ERTS_MMAP_VIRTUAL_PROT, + ERTS_MMAP_VIRTUAL_FLAGS, ERTS_MMAP_FD, 0); + if (res == (void *) MAP_FAILED) + return NULL; + return res; +} + +#else +#error "Missing reserve/unreserve physical memory implementation" +#endif +#endif /* ERTS_HAVE_OS_RESERVE_PHYSICAL_MEMORY */ + +#endif /* ERTS_HAVE_OS_MMAP */ + +static int reserve_noop(char *ptr, UWord size) +{ +#ifdef ERTS_MMAP_DEBUG_FILL_AREAS + Uint32 *uip, *end = (Uint32 *) (ptr + size); + + for (uip = (Uint32 *) ptr; uip < end; uip++) + ERTS_MMAP_ASSERT(*uip == (Uint32) 0xdeadbeef); + for (uip = (Uint32 *) ptr; uip < end; uip++) + *uip = (Uint32) 0xfeedfeed; +#endif + return 1; +} + +static void unreserve_noop(char *ptr, UWord size) +{ +#ifdef ERTS_MMAP_DEBUG_FILL_AREAS + Uint32 *uip, *end = (Uint32 *) (ptr + size); + + for (uip = (Uint32 *) ptr; uip < end; uip++) + *uip = (Uint32) 0xdeadbeef; +#endif +} + +void * +erts_mmap(Uint32 flags, UWord *sizep) +{ + char *seg; + UWord asize = ERTS_PAGEALIGNED_CEILING(*sizep); + + /* Map in premapped supercarrier */ + if (mmap_state.supercarrier && !(ERTS_MMAPFLG_OS_ONLY & flags)) { + char *end; + ErtsFreeSegDesc *desc; + Uint32 superaligned = (ERTS_MMAPFLG_SUPERALIGNED & flags); + + erts_smp_mtx_lock(&mmap_state.mtx); + + if (!superaligned) { + desc = lookup_free_seg(&mmap_state.sua.map, asize); + if (desc) { + seg = desc->start; + end = seg+asize; + if (!mmap_state.reserve_physical(seg, asize)) + goto supercarrier_reserve_failure; + if (desc->end == end) { + delete_free_seg(&mmap_state.sua.map, desc); + free_desc(desc); + } + else { + ERTS_MMAP_ASSERT(end < desc->end); + resize_free_seg(&mmap_state.sua.map, desc, end, desc->end); + } + ERTS_MMAP_SIZE_SC_SUA_INC(asize); + goto supercarrier_success; + } + + if (asize <= mmap_state.sua.bot - mmap_state.sa.top) { + if (!mmap_state.reserve_physical(mmap_state.sua.bot - asize, + asize)) + goto supercarrier_reserve_failure; + mmap_state.sua.bot -= asize; + seg = mmap_state.sua.bot; + ERTS_MMAP_SIZE_SC_SUA_INC(asize); + goto supercarrier_success; + } + } + + asize = ERTS_SUPERALIGNED_CEILING(asize); + + desc = lookup_free_seg(&mmap_state.sa.map, asize); + if (desc) { + seg = desc->start; + end = seg+asize; + if (!mmap_state.reserve_physical(seg, asize)) + goto supercarrier_reserve_failure; + if (desc->end == end) { + delete_free_seg(&mmap_state.sa.map, desc); + free_desc(desc); + } + else { + ERTS_MMAP_ASSERT(end < desc->end); + resize_free_seg(&mmap_state.sa.map, desc, end, desc->end); + } + ERTS_MMAP_SIZE_SC_SA_INC(asize); + goto supercarrier_success; + } + + if (superaligned) { + + if (asize <= mmap_state.sua.bot - mmap_state.sa.top) { + seg = (void *) mmap_state.sa.top; + if (!mmap_state.reserve_physical(seg, asize)) + goto supercarrier_reserve_failure; + mmap_state.sa.top += asize; + ERTS_MMAP_SIZE_SC_SA_INC(asize); + goto supercarrier_success; + } + + desc = lookup_free_seg(&mmap_state.sua.map, asize + ERTS_SUPERALIGNED_SIZE); + if (desc) { + char *org_start = desc->start; + char *org_end = desc->end; + + seg = (char *) ERTS_SUPERALIGNED_CEILING(org_start); + end = seg + asize; + if (!mmap_state.reserve_physical(seg, asize)) + goto supercarrier_reserve_failure; + if (org_start != seg) { + ERTS_MMAP_ASSERT(org_start < seg); + resize_free_seg(&mmap_state.sua.map, desc, org_start, seg); + desc = NULL; + } + if (end != org_end) { + ERTS_MMAP_ASSERT(end < org_end); + if (desc) + resize_free_seg(&mmap_state.sua.map, desc, end, org_end); + else { + desc = alloc_desc(); + if (!desc) + add_free_desc_area(end, org_end); + else + insert_free_seg(&mmap_state.sua.map, desc, end, org_end); + } + } + ERTS_MMAP_SIZE_SC_SA_INC(asize); + goto supercarrier_success; + } + } + + erts_smp_mtx_unlock(&mmap_state.mtx); + } + +#if ERTS_HAVE_OS_MMAP + /* Map using OS primitives */ + if (!(ERTS_MMAPFLG_SUPERCARRIER_ONLY & flags) && !mmap_state.no_os_mmap) { + if (!(ERTS_MMAPFLG_SUPERALIGNED & flags)) { + seg = os_mmap(asize, 0); + if (!seg) + return NULL; + } + else { + asize = ERTS_SUPERALIGNED_CEILING(*sizep); + seg = os_mmap(asize, 1); + if (!seg) + return NULL; + + if (!ERTS_IS_SUPERALIGNED(seg)) { + char *ptr; + UWord sz; + + os_munmap(seg, asize); + + ptr = os_mmap(asize + ERTS_SUPERALIGNED_SIZE, 1); + if (!ptr) + return NULL; + + seg = (char *) ERTS_SUPERALIGNED_CEILING(ptr); + sz = (UWord) (seg - ptr); + ERTS_MMAP_ASSERT(sz <= ERTS_SUPERALIGNED_SIZE); + if (sz) + os_munmap(ptr, sz); + sz = ERTS_SUPERALIGNED_SIZE - sz; + if (sz) + os_munmap(seg+asize, sz); + } + } + + ERTS_MMAP_SIZE_OS_INC(asize); + *sizep = asize; + return (void *) seg; + } +#endif + *sizep = 0; + return NULL; + +supercarrier_success: + +#ifdef ERTS_MMAP_DEBUG + if ((ERTS_MMAPFLG_SUPERALIGNED & flags) + || ERTS_MMAP_IN_SUPERALIGNED_AREA(seg)) { + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(seg)); + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(asize)); + } + else { + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(seg)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(asize)); + } +#endif + + erts_smp_mtx_unlock(&mmap_state.mtx); + + *sizep = asize; + return (void *) seg; + +supercarrier_reserve_failure: + erts_smp_mtx_unlock(&mmap_state.mtx); + + *sizep = 0; + return NULL; + +} + +void +erts_munmap(Uint32 flags, void **ptrp, UWord *sizep) +{ + void *ptr = *ptrp; + UWord size = *sizep; + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(size)); + if (!ERTS_MMAP_IN_SUPERCARRIER(ptr)) { + ERTS_MMAP_ASSERT(!mmap_state.no_os_mmap); +#if ERTS_HAVE_OS_MMAP + ERTS_MMAP_SIZE_OS_DEC(size); + os_munmap(ptr, size); +#endif + } + else { + char *start, *end; + ErtsFreeSegMap *map; + ErtsFreeSegDesc *prev, *next, *desc; + + ERTS_MMAP_ASSERT(mmap_state.supercarrier); + + start = (char *) ptr; + end = start + size; + + erts_smp_mtx_lock(&mmap_state.mtx); + + if (ERTS_MMAP_IN_SUPERALIGNED_AREA(ptr)) { + + start = (char *) ERTS_SUPERALIGNED_CEILING(start); + end = (char *) ERTS_SUPERALIGNED_FLOOR(end); + + size = (UWord) (end - start); + *ptrp = start; + *sizep = size; + + map = &mmap_state.sa.map; + adjacent_free_seg(map, start, end, &prev, &next); + + ERTS_MMAP_SIZE_SC_SA_DEC(size); + if (end == mmap_state.sa.top) { + ERTS_MMAP_ASSERT(!next); + if (prev) { + start = prev->start; + delete_free_seg(map, prev); + free_desc(prev); + } + mmap_state.sa.top = start; + goto supercarrier_success; + } + } + else { + ERTS_MMAP_ASSERT(ERTS_MMAP_IN_SUPERUNALIGNED_AREA(ptr)); + + map = &mmap_state.sua.map; + adjacent_free_seg(map, start, end, &prev, &next); + + ERTS_MMAP_SIZE_SC_SUA_DEC(size); + if (start == mmap_state.sua.bot) { + ERTS_MMAP_ASSERT(!prev); + if (next) { + end = next->end; + delete_free_seg(map, next); + free_desc(next); + } + mmap_state.sua.bot = end; + goto supercarrier_success; + } + } + + desc = NULL; + + if (next) { + ERTS_MMAP_ASSERT(end < next->end); + end = next->end; + if (prev) { + delete_free_seg(map, next); + free_desc(next); + goto save_prev; + } + desc = next; + } else if (prev) { + save_prev: + ERTS_MMAP_ASSERT(prev->start < start); + start = prev->start; + desc = prev; + } + + if (desc) + resize_free_seg(map, desc, start, end); + else { + desc = alloc_desc(); + if (desc) + insert_free_seg(map, desc, start, end); + else { + if (map == &mmap_state.sa.map) + ERTS_MMAP_SIZE_SC_SA_INC(size); + else + ERTS_MMAP_SIZE_SC_SUA_INC(size); + add_free_desc_area(start, end); + } + } + + supercarrier_success: + erts_smp_mtx_unlock(&mmap_state.mtx); + + mmap_state.unreserve_physical((char *) ptr, size); + } +} + +static void * +remap_move(Uint32 flags, void *ptr, UWord old_size, UWord *sizep) +{ + UWord size = *sizep; + UWord um_size = old_size; + void *um_ptr = ptr; + void *new_ptr = erts_mmap(flags, &size); + if (!new_ptr) + return NULL; + *sizep = size; + if (old_size < size) + size = old_size; + sys_memcpy(new_ptr, ptr, (size_t) size); + erts_munmap(flags, &um_ptr, &um_size); + ERTS_MMAP_ASSERT(um_ptr == ptr); + ERTS_MMAP_ASSERT(um_size == old_size); + return new_ptr; +} + +void * +erts_mremap(Uint32 flags, void *ptr, UWord old_size, UWord *sizep) +{ + void *new_ptr; + Uint32 superaligned; + UWord asize; + + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(old_size)); + ERTS_MMAP_ASSERT(sizep && ERTS_IS_PAGEALIGNED(*sizep)); + + if (!ERTS_MMAP_IN_SUPERCARRIER(ptr)) { + + ERTS_MMAP_ASSERT(!mmap_state.no_os_mmap); + + if (!(ERTS_MMAPFLG_OS_ONLY & flags) && mmap_state.supercarrier) { + new_ptr = remap_move(ERTS_MMAPFLG_SUPERCARRIER_ONLY|flags, ptr, + old_size, sizep); + if (new_ptr) + return new_ptr; + } + + if (ERTS_MMAPFLG_SUPERCARRIER_ONLY & flags) + return NULL; + +#if ERTS_HAVE_OS_MREMAP || ERTS_HAVE_GENUINE_OS_MMAP + superaligned = (ERTS_MMAPFLG_SUPERALIGNED & flags); + + if (superaligned) { + asize = ERTS_SUPERALIGNED_CEILING(*sizep); + if (asize == old_size && ERTS_IS_SUPERALIGNED(ptr)) { + *sizep = asize; + return ptr; + } + } + else { + asize = ERTS_PAGEALIGNED_CEILING(*sizep); + if (asize == old_size) { + *sizep = asize; + return ptr; + } + } + +#if ERTS_HAVE_GENUINE_OS_MMAP + if (asize < old_size + && (!superaligned + || ERTS_IS_SUPERALIGNED(ptr))) { + UWord um_sz; + new_ptr = ((char *) ptr) + asize; + ERTS_MMAP_ASSERT((((char *)ptr) + old_size) > (char *) new_ptr); + um_sz = (UWord) ((((char *) ptr) + old_size) - (char *) new_ptr); + ERTS_MMAP_SIZE_OS_DEC(um_sz); + os_munmap(new_ptr, um_sz); + *sizep = asize; + return ptr; + } +#endif +#if ERTS_HAVE_OS_MREMAP + if (superaligned) + return remap_move(flags, new_ptr, old_size, sizep); + else { + new_ptr = os_mremap(ptr, old_size, asize, 0); + if (!new_ptr) + return NULL; + if (asize > old_size) + ERTS_MMAP_SIZE_OS_INC(asize - old_size); + else + ERTS_MMAP_SIZE_OS_DEC(old_size - asize); + *sizep = asize; + return new_ptr; + } +#endif +#endif + } + else { /* In super carrier */ + char *start, *end, *new_end; + ErtsFreeSegMap *map; + ErtsFreeSegDesc *prev, *next, *desc; + + ERTS_MMAP_ASSERT(mmap_state.supercarrier); + + if (ERTS_MMAPFLG_OS_ONLY & flags) + return remap_move(flags, ptr, old_size, sizep); + + superaligned = (ERTS_MMAPFLG_SUPERALIGNED & flags); + + asize = (superaligned + ? ERTS_SUPERALIGNED_CEILING(*sizep) + : ERTS_PAGEALIGNED_CEILING(*sizep)); + + erts_smp_mtx_lock(&mmap_state.mtx); + + if (ERTS_MMAP_IN_SUPERALIGNED_AREA(ptr) + ? (!superaligned && lookup_free_seg(&mmap_state.sua.map, asize)) + : (superaligned && lookup_free_seg(&mmap_state.sa.map, asize))) { + erts_smp_mtx_unlock(&mmap_state.mtx); + /* + * Segment currently in wrong area (due to a previous memory + * shortage), move it to the right area. + * (remap_move() will succeed) + */ + return remap_move(ERTS_MMAPFLG_SUPERCARRIER_ONLY|flags, ptr, + old_size, sizep); + } + + if (asize == old_size) { + new_ptr = ptr; + goto supercarrier_resize_success; + } + + start = (char *) ptr; + end = start + old_size; + new_end = start+asize; + + if (asize < old_size) { + new_ptr = ptr; + if (!ERTS_MMAP_IN_SUPERALIGNED_AREA(ptr)) { + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(old_size)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(asize)); + map = &mmap_state.sua.map; + ERTS_MMAP_SIZE_SC_SUA_DEC(old_size - asize); + } + else { + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(old_size)); + if (!superaligned) { + /* must be a superaligned size in this area */ + asize = ERTS_SUPERALIGNED_CEILING(asize); + ERTS_MMAP_ASSERT(asize <= old_size); + if (asize == old_size) + goto supercarrier_resize_success; + new_end = start+asize; + } + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(asize)); + if (end == mmap_state.sa.top) { + mmap_state.sa.top = new_end; + mmap_state.unreserve_physical(((char *) ptr) + asize, + old_size - asize); + goto supercarrier_resize_success; + } + ERTS_MMAP_SIZE_SC_SA_DEC(old_size - asize); + map = &mmap_state.sa.map; + } + + adjacent_free_seg(map, start, end, &prev, &next); + + if (next) + resize_free_seg(map, next, new_end, next->end); + else { + desc = alloc_desc(); + if (desc) + insert_free_seg(map, desc, new_end, end); + else { + if (map == &mmap_state.sa.map) + ERTS_MMAP_SIZE_SC_SA_INC(old_size - asize); + else + ERTS_MMAP_SIZE_SC_SUA_INC(old_size - asize); + add_free_desc_area(new_end, end); + goto supercarrier_resize_success; + } + } + mmap_state.unreserve_physical(((char *) ptr) + asize, + old_size - asize); + goto supercarrier_resize_success; + } + + if (!ERTS_MMAP_IN_SUPERALIGNED_AREA(ptr)) { + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(old_size)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(asize)); + + adjacent_free_seg(&mmap_state.sua.map, start, end, &prev, &next); + + if (next && new_end <= next->end) { + if (!mmap_state.reserve_physical(((char *) ptr) + old_size, + asize - old_size)) + goto supercarrier_reserve_failure; + if (new_end < next->end) + resize_free_seg(&mmap_state.sua.map, next, new_end, next->end); + else { + delete_free_seg(&mmap_state.sua.map, next); + free_desc(next); + } + new_ptr = ptr; + ERTS_MMAP_SIZE_SC_SUA_INC(asize - old_size); + goto supercarrier_resize_success; + } + } + else { /* Superaligned area */ + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(old_size)); + + if (!superaligned) { + /* must be a superaligned size in this area */ + asize = ERTS_PAGEALIGNED_CEILING(asize); + new_end = start+asize; + } + + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(asize)); + + if (end == mmap_state.sa.top) { + if (new_end <= mmap_state.sua.bot) { + if (!mmap_state.reserve_physical(((char *) ptr) + old_size, + asize - old_size)) + goto supercarrier_reserve_failure; + mmap_state.sa.top = new_end; + new_ptr = ptr; + ERTS_MMAP_SIZE_SC_SA_INC(asize - old_size); + goto supercarrier_resize_success; + } + } + else { + adjacent_free_seg(&mmap_state.sa.map, start, end, &prev, &next); + if (next && new_end <= next->end) { + if (!mmap_state.reserve_physical(((char *) ptr) + old_size, + asize - old_size)) + goto supercarrier_reserve_failure; + if (new_end < next->end) + resize_free_seg(&mmap_state.sa.map, next, new_end, next->end); + else { + delete_free_seg(&mmap_state.sa.map, next); + free_desc(next); + } + new_ptr = ptr; + ERTS_MMAP_SIZE_SC_SA_INC(asize - old_size); + goto supercarrier_resize_success; + } + } + } + erts_smp_mtx_unlock(&mmap_state.mtx); + + /* Failed to resize... */ + } + + return remap_move(flags, ptr, old_size, sizep); + +supercarrier_resize_success: + +#ifdef ERTS_MMAP_DEBUG + if ((ERTS_MMAPFLG_SUPERALIGNED & flags) + || ERTS_MMAP_IN_SUPERALIGNED_AREA(new_ptr)) { + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(new_ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_SUPERALIGNED(asize)); + } + else { + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(new_ptr)); + ERTS_MMAP_ASSERT(ERTS_IS_PAGEALIGNED(asize)); + } +#endif + + erts_smp_mtx_unlock(&mmap_state.mtx); + + *sizep = asize; + return new_ptr; + +supercarrier_reserve_failure: + + erts_smp_mtx_unlock(&mmap_state.mtx); + *sizep = 0; + return NULL; + +} + +int erts_mmap_in_supercarrier(void *ptr) +{ + return ERTS_MMAP_IN_SUPERCARRIER(ptr); +} + +void +erts_mmap_init(ErtsMMapInit *init) +{ + int virtual_map = 0; + char *start = NULL, *end = NULL; + UWord pagesize; +#if defined(__WIN32__) + SYSTEM_INFO sysinfo; + GetSystemInfo(&sysinfo); + pagesize = (UWord) sysinfo.dwPageSize; +#elif defined(_SC_PAGESIZE) + pagesize = (UWord) sysconf(_SC_PAGESIZE); +#elif defined(HAVE_GETPAGESIZE) + pagesize = (UWord) getpagesize(); +#else +# error "Do not know how to get page size" +#endif +#if defined(HARD_DEBUG) || 0 + erts_fprintf(stderr, "erts_mmap: scs = %bpu\n", init->scs); + erts_fprintf(stderr, "erts_mmap: sco = %i\n", init->sco); + erts_fprintf(stderr, "erts_mmap: scmgc = %i\n", init->scmgc); +#endif + erts_page_inv_mask = pagesize - 1; + if (pagesize & erts_page_inv_mask) + erl_exit(-1, "erts_mmap: Invalid pagesize: %bpu\n", + pagesize); + + erts_have_erts_mmap = 0; + + mmap_state.reserve_physical = reserve_noop; + mmap_state.unreserve_physical = unreserve_noop; + +#if HAVE_MMAP && !defined(MAP_ANON) + mmap_state.mmap_fd = open("/dev/zero", O_RDWR); + if (mmap_state.mmap_fd < 0) + erl_exit(-1, "erts_mmap: Failed to open /dev/zero\n"); +#endif + + erts_smp_mtx_init(&mmap_state.mtx, "erts_mmap"); + +#ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION + if (init->virtual_range.start) { + char *ptr; + UWord sz; + ptr = (char *) ERTS_PAGEALIGNED_CEILING(init->virtual_range.start); + end = (char *) ERTS_PAGEALIGNED_FLOOR(init->virtual_range.end); + sz = end - ptr; + start = os_mmap_virtual(ptr, sz); + if (!start || start > ptr || start >= end) + erl_exit(-1, + "erts_mmap: Failed to create virtual range for super carrier\n"); + sz = start - ptr; + if (sz) + os_munmap(end, sz); + mmap_state.reserve_physical = os_reserve_physical; + mmap_state.unreserve_physical = os_unreserve_physical; + virtual_map = 1; + } + else +#endif + if (init->predefined_area.start) { + start = init->predefined_area.start; + end = init->predefined_area.end; + if (end != (void *) 0 && end < start) + end = start; + } +#if ERTS_HAVE_OS_MMAP + else if (init->scs) { + UWord sz; + sz = ERTS_PAGEALIGNED_CEILING(init->scs); +#ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION + if (!init->scrpm) { + start = os_mmap_virtual(NULL, sz); + mmap_state.reserve_physical = os_reserve_physical; + mmap_state.unreserve_physical = os_unreserve_physical; + virtual_map = 1; + } + else +#endif + { + /* + * The whole supercarrier will by physically + * reserved all the time. + */ + start = os_mmap(sz, 1); + } + if (!start) + erl_exit(-1, + "erts_mmap: Failed to create super carrier of size %bpu MB\n", + init->scs/1024/1024); + end = start + sz; +#ifdef ERTS_MMAP_DEBUG_FILL_AREAS + if (!virtual_map) { + Uint32 *uip; + + for (uip = (Uint32 *) start; uip < (Uint32 *) end; uip++) + *uip = (Uint32) 0xdeadbeef; + } +#endif + } + if (!mmap_state.no_os_mmap) + erts_have_erts_mmap |= ERTS_HAVE_ERTS_OS_MMAP; +#endif + + mmap_state.size.supercarrier.total = 0; + mmap_state.size.supercarrier.used.total = 0; + mmap_state.size.supercarrier.used.sa = 0; + mmap_state.size.supercarrier.used.sua = 0; + mmap_state.size.os.used = 0; + + if (!start) { + mmap_state.sa.bot = NULL; + mmap_state.sua.top = NULL; + mmap_state.sa.bot = NULL; + mmap_state.sua.top = NULL; + mmap_state.no_os_mmap = 0; + } + else { + size_t desc_size; + + mmap_state.no_os_mmap = init->sco; + + desc_size = init->scmgc; + if (desc_size < 100) + desc_size = 100; + desc_size *= sizeof(ErtsFreeSegDesc); + if ((desc_size + + ERTS_SUPERALIGNED_SIZE + + ERTS_PAGEALIGNED_SIZE) > end - start) + erl_exit(-1, "erts_mmap: No space for segments in super carrier\n"); + + mmap_state.sa.bot = start; + mmap_state.sa.bot += desc_size; + mmap_state.sa.bot = (char *) ERTS_SUPERALIGNED_CEILING(mmap_state.sa.bot); + mmap_state.sa.top = mmap_state.sa.bot; + mmap_state.sua.top = (char *) ERTS_SUPERALIGNED_FLOOR(end); + mmap_state.sua.bot = mmap_state.sua.top; + + mmap_state.size.os.used += (UWord) (mmap_state.sa.bot - start); + + if (end == (void *) 0) { + /* + * Very unlikely, but we need a guarantee + * that `mmap_state.sua.top` always will + * compare as larger than all segment pointers + * into the super carrier... + */ + mmap_state.sua.top -= ERTS_PAGEALIGNED_SIZE; + mmap_state.size.os.used += ERTS_PAGEALIGNED_SIZE; + } + + mmap_state.size.supercarrier.total = (UWord) (mmap_state.sua.top - mmap_state.sa.bot); + + /* + * Area before (and after) super carrier + * will be used for free segment descritors. + */ + mmap_state.desc_free_list = NULL; +#ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION + if (virtual_map && mmap_state.sa.bot - start > 0) + os_reserve_physical(start, mmap_state.sa.bot - start); +#endif + add_free_desc_area(start, mmap_state.sa.bot); +#ifdef ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION + if (virtual_map && end - mmap_state.sua.top > 0) + os_reserve_physical(mmap_state.sua.top, end - mmap_state.sua.top); +#endif + add_free_desc_area(mmap_state.sua.top, end); + + init_free_seg_map(&mmap_state.sa.map, 0); + init_free_seg_map(&mmap_state.sua.map, 1); + + mmap_state.supercarrier = 1; + erts_have_erts_mmap |= ERTS_HAVE_ERTS_SUPERCARRIER_MMAP; + +#ifdef HARD_DEBUG { void test_it(void); test_it(); } #endif + + } +#if !ERTS_HAVE_OS_MMAP + mmap_state.no_os_mmap = 1; +#endif + } @@ -897,12 +1973,6 @@ print_tree(enum SortOrder order, RBTNode* root) #endif /* PRINT_TREE */ - -static ErtsFreeSegDesc* new_desc(void) -{ - return (ErtsFreeSegDesc*) malloc(sizeof(ErtsFreeSegDesc)); -} - void test_it(void) { ErtsFreeSegMap map; @@ -912,13 +1982,13 @@ void test_it(void) for (i=0; i<2; i++) { init_free_seg_map(&map, i); - insert_free_seg(&map, new_desc(), (char*)0x11000, (char*)0x12000); + insert_free_seg(&map, alloc_desc(), (char*)0x11000, (char*)0x12000); check_tree(&map.atree, 0); check_tree(&map.stree, 0); - insert_free_seg(&map, new_desc(), (char*)0x13000, (char*)0x14000); + insert_free_seg(&map, alloc_desc(), (char*)0x13000, (char*)0x14000); check_tree(&map.atree, 0); check_tree(&map.stree, 0); - insert_free_seg(&map, new_desc(), (char*)0x15000, (char*)0x17000); + insert_free_seg(&map, alloc_desc(), (char*)0x15000, (char*)0x17000); check_tree(&map.atree, 0); check_tree(&map.stree, 0); - insert_free_seg(&map, new_desc(), (char*)0x8000, (char*)0x10000); + insert_free_seg(&map, alloc_desc(), (char*)0x8000, (char*)0x10000); check_tree(&map.atree, 0); check_tree(&map.stree, 0); desc = lookup_free_seg(&map, 0x500); diff --git a/erts/emulator/sys/common/erl_mmap.h b/erts/emulator/sys/common/erl_mmap.h index 64baa6c493..143f1aff3e 100644 --- a/erts/emulator/sys/common/erl_mmap.h +++ b/erts/emulator/sys/common/erl_mmap.h @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2002-2013. All Rights Reserved. + * Copyright Ericsson AB 2013. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -17,10 +17,95 @@ * %CopyrightEnd% */ +#ifndef ERL_MMAP_H__ +#define ERL_MMAP_H__ + +#include "sys.h" + +#define ERTS_MMAP_SUPERALIGNED_BITS (18) +/* Affects hard limits for sbct and lmbcs documented in erts_alloc.xml */ + +#define ERTS_MMAPFLG_OS_ONLY (((Uint32) 1) << 0) +#define ERTS_MMAPFLG_SUPERCARRIER_ONLY (((Uint32) 1) << 1) +#define ERTS_MMAPFLG_SUPERALIGNED (((Uint32) 1) << 2) + +#define ERTS_HAVE_ERTS_OS_MMAP (1 << 0) +#define ERTS_HAVE_ERTS_SUPERCARRIER_MMAP (1 << 1) +extern int erts_have_erts_mmap; +extern UWord erts_page_inv_mask; + typedef struct { - SWord scs; /* super carrier size */ + struct { + char *start; + char *end; + } virtual_range; + struct { + char *start; + char *end; + } predefined_area; + UWord scs; /* super carrier size */ int sco; /* super carrier only? */ Uint scmgc; /* super carrier: max guaranteed (number of) carriers */ + int scrpm; }ErtsMMapInit; +#define ERTS_MMAP_INIT_DEFAULT_INITER \ + {{NULL, NULL}, {NULL, NULL}, 0, 1, (1 << 16), 1} + +void *erts_mmap(Uint32 flags, UWord *sizep); +void erts_munmap(Uint32 flags, void **ptrp, UWord *sizep); +void *erts_mremap(Uint32 flags, void *ptr, UWord old_size, UWord *sizep); +int erts_mmap_in_supercarrier(void *ptr); void erts_mmap_init(ErtsMMapInit*); + +#define ERTS_SUPERALIGNED_SIZE \ + (1 << ERTS_MMAP_SUPERALIGNED_BITS) +#define ERTS_INV_SUPERALIGNED_MASK \ + ((UWord) (ERTS_SUPERALIGNED_SIZE - 1)) +#define ERTS_SUPERALIGNED_MASK \ + (~ERTS_INV_SUPERALIGNED_MASK) +#define ERTS_SUPERALIGNED_FLOOR(X) \ + (((UWord) (X)) & ERTS_SUPERALIGNED_MASK) +#define ERTS_SUPERALIGNED_CEILING(X) \ + ERTS_SUPERALIGNED_FLOOR((X) + ERTS_INV_SUPERALIGNED_MASK) +#define ERTS_IS_SUPERALIGNED(X) \ + (((UWord) (X) & ERTS_INV_SUPERALIGNED_MASK) == 0) + +#define ERTS_INV_PAGEALIGNED_MASK \ + (erts_page_inv_mask) +#define ERTS_PAGEALIGNED_MASK \ + (~ERTS_INV_PAGEALIGNED_MASK) +#define ERTS_PAGEALIGNED_FLOOR(X) \ + (((UWord) (X)) & ERTS_PAGEALIGNED_MASK) +#define ERTS_PAGEALIGNED_CEILING(X) \ + ERTS_PAGEALIGNED_FLOOR((X) + ERTS_INV_PAGEALIGNED_MASK) +#define ERTS_IS_PAGEALIGNED(X) \ + (((UWord) (X) & ERTS_INV_PAGEALIGNED_MASK) == 0) +#define ERTS_PAGEALIGNED_SIZE \ + (ERTS_INV_PAGEALIGNED_MASK + 1) + +#ifndef HAVE_MMAP +# define HAVE_MMAP 0 +#endif +#ifndef HAVE_MREMAP +# define HAVE_MREMAP 0 +#endif +#if HAVE_MMAP +# define ERTS_HAVE_OS_MMAP 1 +# define ERTS_HAVE_GENUINE_OS_MMAP 1 +# if HAVE_MREMAP +# define ERTS_HAVE_OS_MREMAP 1 +# endif +# if defined(MAP_FIXED) && defined(MAP_NORESERVE) +# define ERTS_HAVE_OS_PHYSICAL_MEMORY_RESERVATION 1 +# endif +#endif + +#ifndef HAVE_VIRTUALALLOC +# define HAVE_VIRTUALALLOC 0 +#endif +#if HAVE_VIRTUALALLOC +# define ERTS_HAVE_OS_MMAP 1 +#endif + +#endif /* ERL_MMAP_H__ */ diff --git a/erts/emulator/sys/common/erl_mseg.c b/erts/emulator/sys/common/erl_mseg.c index 64fcb6bb40..b21d6ca393 100644 --- a/erts/emulator/sys/common/erl_mseg.c +++ b/erts/emulator/sys/common/erl_mseg.c @@ -100,45 +100,6 @@ static int atoms_initialized; typedef struct mem_kind_t MemKind; -#if HALFWORD_HEAP -static int initialize_pmmap(void); -static void *pmmap(size_t size); -static int pmunmap(void *p, size_t size); -static void *pmremap(void *old_address, size_t old_size, - size_t new_size); -#endif - -#if HAVE_MMAP -/* Mmap ... */ - -#define MMAP_PROT (PROT_READ|PROT_WRITE) - - -#ifdef MAP_ANON -# define MMAP_FLAGS (MAP_ANON|MAP_PRIVATE) -# define MMAP_FD (-1) -#else -# define MMAP_FLAGS (MAP_PRIVATE) -# define MMAP_FD mmap_fd -static int mmap_fd; -#endif - -#if HAVE_MREMAP -# define HAVE_MSEG_RECREATE 1 -#else -# define HAVE_MSEG_RECREATE 0 -#endif - -#if HALFWORD_HEAP -#define CAN_PARTLY_DESTROY 0 -#else -#define CAN_PARTLY_DESTROY 1 -#endif -#else /* #if HAVE_MMAP */ -#define CAN_PARTLY_DESTROY 0 -#error "Not supported" -#endif /* #if HAVE_MMAP */ - const ErtsMsegOpt_t erts_mseg_default_opt = { 1, /* Use cache */ 1, /* Preserv data */ @@ -163,9 +124,7 @@ typedef struct { CallCounter create; CallCounter create_resize; CallCounter destroy; -#if HAVE_MSEG_RECREATE CallCounter recreate; -#endif CallCounter clear_cache; CallCounter check_cache; } ErtsMsegCalls; @@ -236,11 +195,6 @@ struct ErtsMsegAllctr_t_ { Uint rel_max_cache_bad_fit; ErtsMsegCalls calls; - -#if CAN_PARTLY_DESTROY - Uint min_seg_size; -#endif - }; typedef union { @@ -344,69 +298,31 @@ schedule_cache_check(ErtsMsegAllctr_t *ma) { } } -/* remove ErtsMsegAllctr_t from arguments? - * only used for statistics - */ -static ERTS_INLINE void * -mmap_align(ErtsMsegAllctr_t *ma, void *addr, size_t length, int prot, int flags, int fd, off_t offset) { - - char *p, *q; - UWord d; - - p = mmap(addr, length, prot, flags, fd, offset); - - if (MAP_IS_ALIGNED(p) || p == MAP_FAILED) - return p; - - if (ma) - INC_CC(ma, create_resize); - - munmap(p, length); - - if ((p = mmap(addr, length + MSEG_ALIGNED_SIZE, prot, flags, fd, offset)) == MAP_FAILED) - return MAP_FAILED; - - q = (void *)ALIGNED_CEILING((char *)p); - d = (UWord)(q - p); - - if (d > 0) - munmap(p, d); - - if (MSEG_ALIGNED_SIZE - d > 0) - munmap((void *)(q + length), MSEG_ALIGNED_SIZE - d); - - return q; -} +/* #define ERTS_PRINT_ERTS_MMAP */ static ERTS_INLINE void * -mseg_create(ErtsMsegAllctr_t *ma, MemKind* mk, Uint size) +mseg_create(ErtsMsegAllctr_t *ma, Uint flags, MemKind* mk, UWord *sizep) { +#ifdef ERTS_PRINT_ERTS_MMAP + UWord req_size = *sizep; +#endif void *seg; - ASSERT(size % MSEG_ALIGNED_SIZE == 0); - + Uint32 mmap_flags = 0; #if HALFWORD_HEAP - if (mk == &ma->low_mem) { - seg = pmmap(size); - if ((unsigned long) seg & CHECK_POINTER_MASK) { - erts_fprintf(stderr,"Pointer mask failure (0x%08lx)\n",(unsigned long) seg); - return NULL; - } - } else + mmap_flags |= ((mk == &ma->low_mem) + ? ERTS_MMAPFLG_SUPERCARRIER_ONLY + : ERTS_MMAPFLG_OS_ONLY); #endif - { -#if HAVE_MMAP - { - seg = (void *) mmap_align(ma, (void *) 0, (size_t) size, - MMAP_PROT, MMAP_FLAGS, MMAP_FD, 0); - if (seg == (void *) MAP_FAILED) - seg = NULL; - - ASSERT(MAP_IS_ALIGNED(seg) || !seg); - } -#else -# error "Missing mseg_create() implementation" + if (MSEG_FLG_IS_2POW(flags)) + mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED; + + seg = erts_mmap(mmap_flags, sizep); + +#ifdef ERTS_PRINT_ERTS_MMAP + erts_fprintf(stderr, "%p = erts_mmap(%s, {%bpu, %bpu});\n", seg, + (mmap_flags & ERTS_MMAPFLG_SUPERALIGNED) ? "sa" : "sua", + req_size, *sizep); #endif - } INC_CC(ma, create); @@ -414,91 +330,55 @@ mseg_create(ErtsMsegAllctr_t *ma, MemKind* mk, Uint size) } static ERTS_INLINE void -mseg_destroy(ErtsMsegAllctr_t *ma, MemKind* mk, void *seg, Uint size) { - ERTS_DECLARE_DUMMY(int res); - +mseg_destroy(ErtsMsegAllctr_t *ma, Uint flags, MemKind* mk, void **seg_pp, UWord *size_p) { + + Uint32 mmap_flags = 0; #if HALFWORD_HEAP - if (mk == &ma->low_mem) { - res = pmunmap((void *) seg, size); - } - else + mmap_flags |= ((mk == &ma->low_mem) + ? ERTS_MMAPFLG_SUPERCARRIER_ONLY + : ERTS_MMAPFLG_OS_ONLY); #endif - { -#ifdef HAVE_MMAP - res = munmap((void *) seg, size); -#else -# error "Missing mseg_destroy() implementation" + if (MSEG_FLG_IS_2POW(flags)) + mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED; + + erts_munmap(mmap_flags, seg_pp, size_p); +#ifdef ERTS_PRINT_ERTS_MMAP + erts_fprintf(stderr, "erts_munmap(%s, %p, %bpu);\n", + (mmap_flags & ERTS_MMAPFLG_SUPERALIGNED) ? "sa" : "sua", + *seg_pp, *size_p); #endif - } - - ASSERT(size % MSEG_ALIGNED_SIZE == 0); - ASSERT(res == 0); - INC_CC(ma, destroy); } -#if HAVE_MSEG_RECREATE -#if defined(__NetBSD__) -#define MREMAP_FLAGS (0) -#else -#define MREMAP_FLAGS (MREMAP_MAYMOVE) -#endif - - -/* mseg_recreate - * May return *unaligned* segments as in address not aligned to MSEG_ALIGNMENT - * it is still page aligned - * - * This is fine for single block carriers as long as we don't cache misaligned - * segments (since multiblock carriers may use them) - * - * For multiblock carriers we *need* MSEG_ALIGNMENT but mbc's will never be - * reallocated. - * - * This should probably be fixed the following way: - * 1) Use an option to segment allocation - NEED_ALIGNMENT - * 2) Add mremap_align which takes care of aligning a new a mremaped area - * 3) Fix the cache to handle of aligned and unaligned segments - */ - static ERTS_INLINE void * -mseg_recreate(ErtsMsegAllctr_t *ma, MemKind* mk, void *old_seg, Uint old_size, Uint new_size) +mseg_recreate(ErtsMsegAllctr_t *ma, Uint flags, MemKind* mk, void *old_seg, UWord old_size, UWord *sizep) { +#ifdef ERTS_PRINT_ERTS_MMAP + UWord req_size = *sizep; +#endif void *new_seg; - - ASSERT(old_size % MSEG_ALIGNED_SIZE == 0); - ASSERT(new_size % MSEG_ALIGNED_SIZE == 0); - + Uint32 mmap_flags = 0; #if HALFWORD_HEAP - if (mk == &ma->low_mem) { - new_seg = (void *) pmremap((void *) old_seg, - (size_t) old_size, - (size_t) new_size); - } - else -#endif - { -#if HAVE_MREMAP -#if defined(__NetBSD__) - new_seg = mremap(old_seg, (size_t)old_size, NULL, new_size, MREMAP_FLAGS); -#else - new_seg = mremap(old_seg, (size_t)old_size, (size_t)new_size, MREMAP_FLAGS); + mmap_flags |= ((mk == &ma->low_mem) + ? ERTS_MMAPFLG_SUPERCARRIER_ONLY + : ERTS_MMAPFLG_OS_ONLY); #endif - if (new_seg == (void *) MAP_FAILED) - new_seg = NULL; -#else -#error "Missing mseg_recreate() implementation" -#endif - } + if (MSEG_FLG_IS_2POW(flags)) + mmap_flags |= ERTS_MMAPFLG_SUPERALIGNED; + + new_seg = erts_mremap(mmap_flags, old_seg, old_size, sizep); +#ifdef ERTS_PRINT_ERTS_MMAP + erts_fprintf(stderr, "%p = erts_mremap(%s, %p, %bpu, {%bpu, %bpu});\n", + new_seg, (mmap_flags & ERTS_MMAPFLG_SUPERALIGNED) ? "sa" : "sua", + old_seg, old_size, req_size, *sizep); +#endif INC_CC(ma, recreate); return new_seg; } -#endif /* #if HAVE_MSEG_RECREATE */ - #ifdef DEBUG #define ERTS_DBG_MA_CHK_THR_ACCESS(MA) \ do { \ @@ -566,14 +446,19 @@ static ERTS_INLINE int cache_bless_segment(MemKind *mk, void *seg, Uint size, Ui return 1; } else if (!MSEG_FLG_IS_2POW(flags) && !erts_circleq_is_empty(&(mk->cache_unpowered_node))) { - + void *destr_seg; + UWord destr_size; /* No free slots. * Evict oldest slot from unpowered cache so we can cache an unpowered (sbc) segment */ c = erts_circleq_tail(&(mk->cache_unpowered_node)); erts_circleq_remove(c); - mseg_destroy(mk->ma, mk, c->seg, c->size); + destr_seg = c->seg; + destr_size = c->size; + mseg_destroy(mk->ma, ERTS_MSEG_FLG_NONE, mk, &destr_seg, &destr_size); + ASSERT(destr_seg == c->seg); + ASSERT(destr_size == c->size); mseg_cache_clear_node(c); c->seg = seg; @@ -593,13 +478,20 @@ static ERTS_INLINE int cache_bless_segment(MemKind *mk, void *seg, Uint size, Ui int i; for( i = 0; i < CACHE_AREAS; i++) { + void *destr_seg; + UWord destr_size; if (erts_circleq_is_empty(&(mk->cache_powered_node[i]))) continue; c = erts_circleq_tail(&(mk->cache_powered_node[i])); erts_circleq_remove(c); - mseg_destroy(mk->ma, mk, c->seg, c->size); + destr_seg = seg; + destr_size = c->size; + mseg_destroy(mk->ma, ERTS_MSEG_FLG_2POW, mk, &destr_seg, &destr_size); + ASSERT(destr_seg == c->seg); + ASSERT(destr_size == c->size); + mseg_cache_clear_node(c); c->seg = seg; @@ -614,9 +506,9 @@ static ERTS_INLINE int cache_bless_segment(MemKind *mk, void *seg, Uint size, Ui return 0; } -static ERTS_INLINE void *cache_get_segment(MemKind *mk, Uint *size_p, Uint flags) { +static ERTS_INLINE void *cache_get_segment(MemKind *mk, UWord *size_p, Uint flags) { - Uint size = *size_p; + Uint size = (UWord) *size_p; ERTS_DBG_MK_CHK_THR_ACCESS(mk); @@ -653,7 +545,11 @@ static ERTS_INLINE void *cache_get_segment(MemKind *mk, Uint *size_p, Uint flags ASSERT(!(mk->cache_size < 0)); if (csize != size) { - mseg_destroy(mk->ma, mk, (char *)seg + size, csize - size); + void *destr_seg = ((char *) seg) + size; + UWord destr_size = csize - size; + mseg_destroy(mk->ma, ERTS_MSEG_FLG_2POW, mk, &destr_seg, &destr_size); + *size_p = (UWord) (destr_seg - seg); + ASSERT(c->seg + c->size == destr_seg + destr_size); } return seg; @@ -683,7 +579,7 @@ static ERTS_INLINE void *cache_get_segment(MemKind *mk, Uint *size_p, Uint flags mseg_cache_clear_node(c); erts_circleq_push_head(&(mk->cache_free), c); - *size_p = csize; + *size_p = (UWord) csize; return seg; @@ -714,7 +610,7 @@ static ERTS_INLINE void *cache_get_segment(MemKind *mk, Uint *size_p, Uint flags ASSERT((size % GET_PAGE_SIZE) == 0); ASSERT((best->size % GET_PAGE_SIZE) == 0); - *size_p = size; + *size_p = (UWord) size; return seg; @@ -728,7 +624,9 @@ static ERTS_INLINE void *cache_get_segment(MemKind *mk, Uint *size_p, Uint flags * using callbacks from aux-work in the scheduler. */ -static ERTS_INLINE Uint mseg_drop_one_memkind_cache_size(MemKind *mk, cache_t *head) { +static ERTS_INLINE Uint mseg_drop_one_memkind_cache_size(MemKind *mk, Uint flags, cache_t *head) { + void *destr_seg; + UWord destr_size; cache_t *c = NULL; c = erts_circleq_tail(head); @@ -737,7 +635,11 @@ static ERTS_INLINE Uint mseg_drop_one_memkind_cache_size(MemKind *mk, cache_t *h if (erts_mtrace_enabled) erts_mtrace_crr_free(SEGTYPE, SEGTYPE, c->seg); - mseg_destroy(mk->ma, mk, c->seg, c->size); + destr_seg = c->seg; + destr_size = c->size; + mseg_destroy(mk->ma, flags, mk, &destr_seg, &destr_size); + ASSERT(destr_seg == c->seg); + ASSERT(destr_size == c->size); mseg_cache_clear_node(c); erts_circleq_push_head(&(mk->cache_free), c); @@ -749,10 +651,12 @@ static ERTS_INLINE Uint mseg_drop_one_memkind_cache_size(MemKind *mk, cache_t *h return mk->cache_size; } -static ERTS_INLINE Uint mseg_drop_memkind_cache_size(MemKind *mk, cache_t *head) { +static ERTS_INLINE Uint mseg_drop_memkind_cache_size(MemKind *mk, Uint flags, cache_t *head) { cache_t *c = NULL; while (!erts_circleq_is_empty(head)) { + void *destr_seg; + UWord destr_size; c = erts_circleq_tail(head); erts_circleq_remove(c); @@ -760,7 +664,11 @@ static ERTS_INLINE Uint mseg_drop_memkind_cache_size(MemKind *mk, cache_t *head) if (erts_mtrace_enabled) erts_mtrace_crr_free(SEGTYPE, SEGTYPE, c->seg); - mseg_destroy(mk->ma, mk, c->seg, c->size); + destr_seg = c->seg; + destr_size = c->size; + mseg_destroy(mk->ma, flags, mk, &destr_seg, &destr_size); + ASSERT(destr_seg == c->seg); + ASSERT(destr_size == c->size); mseg_cache_clear_node(c); erts_circleq_push_head(&(mk->cache_free), c); @@ -788,11 +696,11 @@ static Uint mseg_check_memkind_cache(MemKind *mk) { for (i = 0; i < CACHE_AREAS; i++) { if (!erts_circleq_is_empty(&(mk->cache_powered_node[i]))) - return mseg_drop_one_memkind_cache_size(mk, &(mk->cache_powered_node[i])); + return mseg_drop_one_memkind_cache_size(mk, ERTS_MSEG_FLG_2POW, &(mk->cache_powered_node[i])); } if (!erts_circleq_is_empty(&(mk->cache_unpowered_node))) - return mseg_drop_one_memkind_cache_size(mk, &(mk->cache_unpowered_node)); + return mseg_drop_one_memkind_cache_size(mk, ERTS_MSEG_FLG_NONE, &(mk->cache_unpowered_node)); return 0; } @@ -851,12 +759,12 @@ static void mseg_clear_memkind_cache(MemKind *mk) { if (erts_circleq_is_empty(&(mk->cache_powered_node[i]))) continue; - mseg_drop_memkind_cache_size(mk, &(mk->cache_powered_node[i])); + mseg_drop_memkind_cache_size(mk, ERTS_MSEG_FLG_2POW, &(mk->cache_powered_node[i])); ASSERT(erts_circleq_is_empty(&(mk->cache_powered_node[i]))); } /* drop varied caches */ if (!erts_circleq_is_empty(&(mk->cache_unpowered_node))) - mseg_drop_memkind_cache_size(mk, &(mk->cache_unpowered_node)); + mseg_drop_memkind_cache_size(mk, ERTS_MSEG_FLG_NONE, &(mk->cache_unpowered_node)); ASSERT(erts_circleq_is_empty(&(mk->cache_unpowered_node))); ASSERT(mk->cache_size == 0); @@ -896,36 +804,33 @@ static ERTS_INLINE MemKind* memkind(ErtsMsegAllctr_t *ma, } static void * -mseg_alloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, Uint *size_p, +mseg_alloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, UWord *size_p, Uint flags, const ErtsMsegOpt_t *opt) { - Uint size; + UWord size; void *seg; MemKind* mk = memkind(ma, opt); INC_CC(ma, alloc); - /* Carrier align */ - size = ALIGNED_CEILING(*size_p); - - /* Cache optim (if applicable) */ - if (MSEG_FLG_IS_2POW(flags) && !IS_2POW(size)) + if (!MSEG_FLG_IS_2POW(flags) && !IS_2POW(size)) + size = ERTS_PAGEALIGNED_CEILING(*size_p); + else { + size = ALIGNED_CEILING(*size_p); + /* Cache optim (if applicable) */ size = ceil_2pow(size); + } -#if CAN_PARTLY_DESTROY - if (size < ma->min_seg_size) - ma->min_seg_size = size; -#endif - if (opt->cache && mk->cache_size > 0 && (seg = cache_get_segment(mk, &size, flags)) != NULL) goto done; - if ((seg = mseg_create(ma, mk, size)) == NULL) - size = 0; + seg = mseg_create(ma, flags, mk, &size); + if (!seg) + *size_p = 0; + else { done: - *size_p = size; - if (seg) { + *size_p = size; if (erts_mtrace_enabled) erts_mtrace_crr_alloc(seg, atype, ERTS_MTRACE_SEGMENT_ID, size); @@ -937,14 +842,16 @@ done: static void -mseg_dealloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg, Uint size, +mseg_dealloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg, UWord size, Uint flags, const ErtsMsegOpt_t *opt) { + void *destr_seg; + UWord destr_size; MemKind* mk = memkind(ma, opt); ERTS_MSEG_DEALLOC_STAT(mk,size); - if (opt->cache && cache_bless_segment(mk, seg, size, flags)) { + if (opt->cache && cache_bless_segment(mk, seg, (Uint) size, flags)) { schedule_cache_check(ma); goto done; } @@ -952,7 +859,11 @@ mseg_dealloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg, Uint size, if (erts_mtrace_enabled) erts_mtrace_crr_free(atype, SEGTYPE, seg); - mseg_destroy(ma, mk, seg, size); + destr_seg = seg; + destr_size = size; + mseg_destroy(ma, flags, mk, &destr_seg, &destr_size); + ASSERT(destr_seg == seg); + ASSERT(destr_size == size); done: @@ -961,11 +872,11 @@ done: static void * mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg, - Uint old_size, Uint *new_size_p, Uint flags, const ErtsMsegOpt_t *opt) + UWord old_size, UWord *new_size_p, Uint flags, const ErtsMsegOpt_t *opt) { MemKind* mk; void *new_seg; - Uint new_size; + UWord new_size; /* Just allocate a new segment if we didn't have one before */ if (!seg || !old_size) { @@ -985,90 +896,46 @@ mseg_realloc(ErtsMsegAllctr_t *ma, ErtsAlcType_t atype, void *seg, mk = memkind(ma, opt); new_seg = seg; - /* Carrier align */ - new_size = ALIGNED_CEILING(*new_size_p); - /* Cache optim (if applicable) */ - if (MSEG_FLG_IS_2POW(flags) && !IS_2POW(new_size)) + if (!MSEG_FLG_IS_2POW(flags)) + new_size = ERTS_PAGEALIGNED_CEILING(*new_size_p); + else { + new_size = ALIGNED_CEILING(*new_size_p); + /* Cache optim (if applicable) */ new_size = ceil_2pow(new_size); + } - if (new_size == old_size) - ; - else if (new_size < old_size) { - Uint shrink_sz = old_size - new_size; - -#if CAN_PARTLY_DESTROY - if (new_size < ma->min_seg_size) - ma->min_seg_size = new_size; -#endif - /* +M<S>rsbcst <ratio> */ - if (shrink_sz < opt->abs_shrink_th - && 100*shrink_sz < opt->rel_shrink_th*old_size) { - new_size = old_size; + if (new_size > old_size) { + if (opt->preserv) { + new_seg = mseg_recreate(ma, flags, mk, (void *) seg, old_size, &new_size); + if (!new_seg) + new_size = old_size; } else { - -#if CAN_PARTLY_DESTROY - - if (erts_mtrace_enabled) - erts_mtrace_crr_realloc(new_seg, atype, SEGTYPE, seg, new_size); - - mseg_destroy(ma, mk, ((char *) seg) + new_size, shrink_sz); - -#elif HAVE_MSEG_RECREATE - goto do_recreate; -#else + mseg_dealloc(ma, atype, seg, old_size, flags, opt); new_seg = mseg_alloc(ma, atype, &new_size, flags, opt); - - ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg); - if (!new_seg) - new_size = old_size; - else { - sys_memcpy(((char *) new_seg), - ((char *) seg), - MIN(new_size, old_size)); - mseg_dealloc(ma, atype, seg, old_size, flags, opt); - } -#endif + new_size = 0; } } - else { + else if (new_size < old_size) { + UWord shrink_sz = old_size - new_size; - if (!opt->preserv) { - mseg_dealloc(ma, atype, seg, old_size, flags, opt); - new_seg = mseg_alloc(ma, atype, &new_size, flags, opt); - ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg); + /* +M<S>rsbcst <ratio> */ + if (shrink_sz < opt->abs_shrink_th + && 100*shrink_sz < opt->rel_shrink_th*old_size) { + new_size = old_size; } else { -#if HAVE_MSEG_RECREATE -#if !CAN_PARTLY_DESTROY - do_recreate: -#endif - new_seg = mseg_recreate(ma, mk, (void *) seg, old_size, new_size); - /* ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg); - * will not always be aligned and it ok for now - */ - - if (erts_mtrace_enabled) - erts_mtrace_crr_realloc(new_seg, atype, SEGTYPE, seg, new_size); + new_seg = mseg_recreate(ma, flags, mk, (void *) seg, old_size, &new_size); if (!new_seg) new_size = old_size; -#else - new_seg = mseg_alloc(ma, atype, &new_size, flags, opt); - - ASSERT(MAP_IS_ALIGNED(new_seg) || !new_seg); - - if (!new_seg) - new_size = old_size; - else { - sys_memcpy(((char *) new_seg), ((char *) seg), MIN(new_size, old_size)); - mseg_dealloc(ma, atype, seg, old_size, flags, opt); - } -#endif } } + if (erts_mtrace_enabled) + erts_mtrace_crr_realloc(new_seg, atype, SEGTYPE, seg, new_size); + INC_CC(ma, realloc); ASSERT(!MSEG_FLG_IS_2POW(flags) || IS_2POW(new_size)); @@ -1106,9 +973,7 @@ static struct { Eterm mseg_create; Eterm mseg_create_resize; Eterm mseg_destroy; -#if HAVE_MSEG_RECREATE Eterm mseg_recreate; -#endif Eterm mseg_clear_cache; Eterm mseg_check_cache; @@ -1163,9 +1028,7 @@ init_atoms(ErtsMsegAllctr_t *ma) AM_INIT(mseg_create); AM_INIT(mseg_create_resize); AM_INIT(mseg_destroy); -#if HAVE_MSEG_RECREATE AM_INIT(mseg_recreate); -#endif AM_INIT(mseg_clear_cache); AM_INIT(mseg_check_cache); @@ -1293,9 +1156,7 @@ info_calls(ErtsMsegAllctr_t *ma, int *print_to_p, void *print_to_arg, Uint **hpp PRINT_CC(to, arg, create); PRINT_CC(to, arg, create_resize); PRINT_CC(to, arg, destroy); -#if HAVE_MSEG_RECREATE PRINT_CC(to, arg, recreate); -#endif PRINT_CC(to, arg, clear_cache); PRINT_CC(to, arg, check_cache); @@ -1316,12 +1177,10 @@ info_calls(ErtsMsegAllctr_t *ma, int *print_to_p, void *print_to_arg, Uint **hpp bld_unstable_uint(hpp, szp, ma->calls.clear_cache.giga_no), bld_unstable_uint(hpp, szp, ma->calls.clear_cache.no)); -#if HAVE_MSEG_RECREATE add_3tup(hpp, szp, &res, am.mseg_recreate, bld_unstable_uint(hpp, szp, ma->calls.recreate.giga_no), bld_unstable_uint(hpp, szp, ma->calls.recreate.no)); -#endif add_3tup(hpp, szp, &res, am.mseg_destroy, bld_unstable_uint(hpp, szp, ma->calls.destroy.giga_no), @@ -1521,7 +1380,7 @@ erts_mseg_info(int ix, } void * -erts_mseg_alloc_opt(ErtsAlcType_t atype, Uint *size_p, Uint flags, const ErtsMsegOpt_t *opt) +erts_mseg_alloc_opt(ErtsAlcType_t atype, UWord *size_p, Uint flags, const ErtsMsegOpt_t *opt) { ErtsMsegAllctr_t *ma = ERTS_MSEG_ALLCTR_OPT(opt); void *seg; @@ -1533,14 +1392,14 @@ erts_mseg_alloc_opt(ErtsAlcType_t atype, Uint *size_p, Uint flags, const ErtsMse } void * -erts_mseg_alloc(ErtsAlcType_t atype, Uint *size_p, Uint flags) +erts_mseg_alloc(ErtsAlcType_t atype, UWord *size_p, Uint flags) { return erts_mseg_alloc_opt(atype, size_p, flags, &erts_mseg_default_opt); } void erts_mseg_dealloc_opt(ErtsAlcType_t atype, void *seg, - Uint size, Uint flags, const ErtsMsegOpt_t *opt) + UWord size, Uint flags, const ErtsMsegOpt_t *opt) { ErtsMsegAllctr_t *ma = ERTS_MSEG_ALLCTR_OPT(opt); ERTS_MSEG_LOCK(ma); @@ -1550,14 +1409,14 @@ erts_mseg_dealloc_opt(ErtsAlcType_t atype, void *seg, } void -erts_mseg_dealloc(ErtsAlcType_t atype, void *seg, Uint size, Uint flags) +erts_mseg_dealloc(ErtsAlcType_t atype, void *seg, UWord size, Uint flags) { erts_mseg_dealloc_opt(atype, seg, size, flags, &erts_mseg_default_opt); } void * erts_mseg_realloc_opt(ErtsAlcType_t atype, void *seg, - Uint old_size, Uint *new_size_p, + UWord old_size, UWord *new_size_p, Uint flags, const ErtsMsegOpt_t *opt) { @@ -1572,7 +1431,7 @@ erts_mseg_realloc_opt(ErtsAlcType_t atype, void *seg, void * erts_mseg_realloc(ErtsAlcType_t atype, void *seg, - Uint old_size, Uint *new_size_p, Uint flags) + UWord old_size, UWord *new_size_p, Uint flags) { return erts_mseg_realloc_opt(atype, seg, old_size, new_size_p, flags, &erts_mseg_default_opt); @@ -1662,19 +1521,16 @@ erts_mseg_init(ErtsMsegInit_t *init) erts_mtx_init(&init_atoms_mutex, "mseg_init_atoms"); -#if HAVE_MMAP && !defined(MAP_ANON) - mmap_fd = open("/dev/zero", O_RDWR); - if (mmap_fd < 0) - erl_exit(ERTS_ABORT_EXIT, "erts_mseg: unable to open /dev/zero\n"); +#if HALFWORD_HEAP + if (sizeof(void *) != 8) + erl_exit(-1,"Halfword emulator cannot be run in 32bit mode"); + + init->mmap.virtual_range.start = (char *) sbrk(0); + init->mmap.virtual_range.end = (char *) 0x100000000UL; + init->mmap.sco = 0; #endif -#if HAVE_MMAP -# if HALFWORD_HEAP - initialize_pmmap(); -# else erts_mmap_init(&init->mmap); -# endif -#endif if (!IS_2POW(GET_PAGE_SIZE)) erl_exit(ERTS_ABORT_EXIT, "erts_mseg: Unexpected page_size %beu\n", GET_PAGE_SIZE); @@ -1716,10 +1572,6 @@ erts_mseg_init(ErtsMsegInit_t *init) #endif sys_memzero((void *) &ma->calls, sizeof(ErtsMsegCalls)); - -#if CAN_PARTLY_DESTROY - ma->min_seg_size = ~((Uint) 0); -#endif } } @@ -1761,7 +1613,7 @@ erts_mseg_test(UWord op, UWord a1, UWord a2, UWord a3) case 0x400: /* Have erts_mseg */ return (UWord) 1; case 0x401: - return (UWord) erts_mseg_alloc(ERTS_ALC_A_INVALID, (Uint *) a1, (Uint) 0); + return (UWord) erts_mseg_alloc(ERTS_ALC_A_INVALID, (UWord *) a1, (Uint) 0); case 0x402: erts_mseg_dealloc(ERTS_ALC_A_INVALID, (void *) a1, (Uint) a2, (Uint) 0); return (UWord) 0; @@ -1769,7 +1621,7 @@ erts_mseg_test(UWord op, UWord a1, UWord a2, UWord a3) return (UWord) erts_mseg_realloc(ERTS_ALC_A_INVALID, (void *) a1, (Uint) a2, - (Uint *) a3, + (UWord *) a3, (Uint) 0); case 0x404: erts_mseg_clear_cache(); @@ -1792,405 +1644,3 @@ erts_mseg_test(UWord op, UWord a1, UWord a2, UWord a3) } } - - -#if HALFWORD_HEAP -/* - * Very simple page oriented mmap replacer. Works in the lower - * 32 bit address range of a 64bit program. - * Implements anonymous mmap mremap and munmap with address order first fit. - * The free list is expected to be very short... - * To be used for compressed pointers in Erlang halfword emulator - * implementation. The MacOS X version is more of a toy, it's not really - * for production as the halfword erlang VM relies on Linux specific memory - * mapping tricks. - */ - -/* #define HARDDEBUG 1 */ - -#ifdef HARDDEBUG -static void dump_freelist(void) -{ - FreeBlock *p = first; - - while (p) { - fprintf(stderr, "p = %p\r\np->num = %ld\r\np->next = %p\r\n\r\n", - (void *) p, (unsigned long) p->num, (void *) p->next); - p = p->next; - } -} - -#define HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(PTR, SZ) \ - fprintf(stderr,"Mapping of address %p with size %ld " \ - "does not map complete pages (%s:%d)\r\n", \ - (void *) (PTR), (unsigned long) (SZ),__FILE__, __LINE__) - -#define HARDDEBUG_HW_UNALIGNED_ALIGNMENT(PTR, SZ) \ - fprintf(stderr,"Mapping of address %p with size %ld " \ - "is not page aligned (%s:%d)\r\n", \ - (void *) (PTR), (unsigned long) (SZ),__FILE__, __LINE__) - -#define HARDDEBUG_MAP_FAILED(PTR, SZ) \ - fprintf(stderr, "Could not actually map memory " \ - "at address %p with size %ld (%s:%d) ..\r\n", \ - (void *) (PTR), (unsigned long) (SZ),__FILE__, __LINE__) -#else -#define HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(PTR, SZ) do{}while(0) -#define HARDDEBUG_HW_UNALIGNED_ALIGNMENT(PTR, SZ) do{}while(0) -#define HARDDEBUG_MAP_FAILED(PTR, SZ) do{}while(0) -#endif - - -#ifdef __APPLE__ -#define MAP_ANONYMOUS MAP_ANON -#endif - -#define INIT_LOCK() do {erts_mtx_init(&pmmap_mutex, "pmmap");} while(0) - -#define TAKE_LOCK() do {erts_mtx_lock(&pmmap_mutex);} while(0) - -#define RELEASE_LOCK() do {erts_mtx_unlock(&pmmap_mutex);} while(0) - -static erts_mtx_t pmmap_mutex; /* Also needed when !USE_THREADS */ - -typedef struct _free_block { - unsigned long num; /*pages*/ - struct _free_block *next; -} FreeBlock; - -/* Protect with lock */ -static FreeBlock *first; - -static void *do_map(void *ptr, size_t sz) -{ - void *res; - - if (ALIGNED_CEILING(sz) != sz) { - HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(ptr, sz); - return NULL; - } - - if (((unsigned long) ptr) % MSEG_ALIGNED_SIZE) { - HARDDEBUG_HW_UNALIGNED_ALIGNMENT(ptr, sz); - return NULL; - } - -#if HAVE_MMAP - res = mmap(ptr, sz, - PROT_READ | PROT_WRITE, MAP_PRIVATE | - MAP_ANONYMOUS | MAP_FIXED, - -1 , 0); -#else -# error "Missing mmap support" -#endif - - if (res == MAP_FAILED) { - HARDDEBUG_MAP_FAILED(ptr, sz); - return NULL; - } - - return res; -} - -static int do_unmap(void *ptr, size_t sz) -{ - void *res; - - if (ALIGNED_CEILING(sz) != sz) { - HARDDEBUG_HW_INCOMPLETE_ALIGNMENT(ptr, sz); - return 1; - } - - if (((unsigned long) ptr) % MSEG_ALIGNED_SIZE) { - HARDDEBUG_HW_UNALIGNED_ALIGNMENT(ptr, sz); - return 1; - } - - res = mmap(ptr, sz, - PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE | MAP_FIXED, - -1 , 0); - - if (res == MAP_FAILED) { - HARDDEBUG_MAP_FAILED(ptr, sz); - return 1; - } - - return 0; -} - -#ifdef __APPLE__ -/* - * The first 4 gig's are protected on Macos X for 64bit processes :( - * The range 0x1000000000 - 0x10FFFFFFFF is selected as an arbitrary - * value of a normally unused range... Real MMAP's will avoid - * it and all 32bit compressed pointers can be in that range... - * More expensive than on Linux where expansion of compressed - * poiters involves no masking (as they are in the first 4 gig's). - * It's also very uncertain if the MAP_NORESERVE flag really has - * any effect in MacOS X. Swap space may always be allocated... - */ -#define SET_RANGE_MIN() /* nothing */ -#define RANGE_MIN 0x1000000000UL -#define RANGE_MAX 0x1100000000UL -#define RANGE_MASK (RANGE_MIN) -#define EXTRA_MAP_FLAGS (MAP_FIXED) -#else -static size_t range_min; -#define SET_RANGE_MIN() do { range_min = (size_t) sbrk(0); } while (0) -#define RANGE_MIN range_min -#define RANGE_MAX 0x100000000UL -#define RANGE_MASK 0UL -#define EXTRA_MAP_FLAGS (0) -#endif - -static int initialize_pmmap(void) -{ - char *p,*q,*rptr; - size_t rsz; - FreeBlock *initial; - - SET_RANGE_MIN(); - if (sizeof(void *) != 8) { - erl_exit(1,"Halfword emulator cannot be run in 32bit mode"); - } - - p = (char *) RANGE_MIN; - q = (char *) RANGE_MAX; - - rsz = ALIGNED_FLOOR(q - p); - - rptr = mmap_align(NULL, (void *) p, rsz, - PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS | - MAP_NORESERVE | EXTRA_MAP_FLAGS, - -1 , 0); -#ifdef HARDDEBUG - printf("p=%p, rsz = %ld, pages = %ld, got range = %p -> %p\r\n", - p, (unsigned long) rsz, (unsigned long) (rsz / MSEG_ALIGNED_SIZE), - (void *) rptr, (void*)(rptr + rsz)); -#endif - if ((UWord)(rptr + rsz) > RANGE_MAX) { - size_t rsz_trunc = RANGE_MAX - (UWord)rptr; -#ifdef HARDDEBUG - printf("Reducing mmap'ed memory from %lu to %lu Mb, reduced range = %p -> %p\r\n", - rsz/(1024*1024), rsz_trunc/(1024*1024), rptr, rptr+rsz_trunc); -#endif - munmap((void*)RANGE_MAX, rsz - rsz_trunc); - rsz = rsz_trunc; - } - if (!do_map(rptr, MSEG_ALIGNED_SIZE)) { - erl_exit(1,"Could not actually mmap first page for halfword emulator...\n"); - } - initial = (FreeBlock *) rptr; - initial->num = (rsz / MSEG_ALIGNED_SIZE); - initial->next = NULL; - first = initial; - INIT_LOCK(); - return 0; -} - -static void *pmmap(size_t size) -{ - size_t real_size = ALIGNED_CEILING(size); - size_t num_pages = real_size / MSEG_ALIGNED_SIZE; - FreeBlock **block; - FreeBlock *tail; - FreeBlock *res; - - TAKE_LOCK(); - - for (block = &first; - *block != NULL && (*block)->num < num_pages; - block = &((*block)->next)) - ; - if (!(*block)) { - RELEASE_LOCK(); - return NULL; - } - if ((*block)->num == num_pages) { - /* nice, perfect fit */ - res = *block; - *block = (*block)->next; - } else { - tail = (FreeBlock *) (((char *) ((void *) (*block))) + real_size); - if (!do_map(tail, MSEG_ALIGNED_SIZE)) { - HARDDEBUG_MAP_FAILED(tail, MSEG_ALIGNED_SIZE); - RELEASE_LOCK(); - return NULL; - } - tail->num = (*block)->num - num_pages; - tail->next = (*block)->next; - res = *block; - *block = tail; - } - - RELEASE_LOCK(); - - if (!do_map(res, real_size)) { - HARDDEBUG_MAP_FAILED(res, real_size); - return NULL; - } - - return (void *) res; -} - -static int pmunmap(void *p, size_t size) -{ - size_t real_size = ALIGNED_CEILING(size); - size_t num_pages = real_size / MSEG_ALIGNED_SIZE; - - FreeBlock *block; - FreeBlock *last; - FreeBlock *nb = (FreeBlock *) p; - - ASSERT(((unsigned long)p & CHECK_POINTER_MASK)==0); - - if (real_size > MSEG_ALIGNED_SIZE) { - if (do_unmap(((char *) p) + MSEG_ALIGNED_SIZE, real_size - MSEG_ALIGNED_SIZE)) { - return 1; - } - } - - TAKE_LOCK(); - - last = NULL; - block = first; - while(block != NULL && ((void *) block) < p) { - last = block; - block = block->next; - } - - if (block != NULL && - ((void *) block) == ((void *) (((char *) p) + real_size))) { - /* Merge new free block with following */ - nb->num = block->num + num_pages; - nb->next = block->next; - if (do_unmap(block, MSEG_ALIGNED_SIZE)) { - RELEASE_LOCK(); - return 1; - } - } else { - /* just link in */ - nb->num = num_pages; - nb->next = block; - } - if (last != NULL) { - if (p == ((void *) (((char *) last) + (last->num * MSEG_ALIGNED_SIZE)))) { - /* Merge with previous */ - last->num += nb->num; - last->next = nb->next; - if (do_unmap(nb, MSEG_ALIGNED_SIZE)) { - RELEASE_LOCK(); - return 1; - } - } else { - last->next = nb; - } - } else { - first = nb; - } - RELEASE_LOCK(); - return 0; -} - -static void *pmremap(void *old_address, size_t old_size, - size_t new_size) -{ - size_t new_real_size = ALIGNED_CEILING(new_size); - size_t new_num_pages = new_real_size / MSEG_ALIGNED_SIZE; - size_t old_real_size = ALIGNED_CEILING(old_size); - size_t old_num_pages = old_real_size / MSEG_ALIGNED_SIZE; - if (new_num_pages == old_num_pages) { - return old_address; - } else if (new_num_pages < old_num_pages) { /* Shrink */ - size_t nfb_pages = old_num_pages - new_num_pages; - size_t nfb_real_size = old_real_size - new_real_size; - void *vnfb = (void *) (((char *)old_address) + new_real_size); - FreeBlock *nfb = (FreeBlock *) vnfb; - FreeBlock **block; - TAKE_LOCK(); - for (block = &first; - *block != NULL && (*block) < nfb; - block = &((*block)->next)) - ; - if (!(*block) || - (*block) > ((FreeBlock *)(((char *) vnfb) + nfb_real_size))) { - /* Normal link in */ - if (nfb_pages > 1) { - if (do_unmap((void *)(((char *) vnfb) + MSEG_ALIGNED_SIZE), - (nfb_pages - 1)*MSEG_ALIGNED_SIZE)) { - return NULL; - } - } - nfb->next = (*block); - nfb->num = nfb_pages; - (*block) = nfb; - } else { /* block merge */ - nfb->next = (*block)->next; - nfb->num = nfb_pages + (*block)->num; - /* unmap also the first page of the next freeblock */ - (*block) = nfb; - if (do_unmap((void *)(((char *) vnfb) + MSEG_ALIGNED_SIZE), - nfb_pages*MSEG_ALIGNED_SIZE)) { - return NULL; - } - } - RELEASE_LOCK(); - return old_address; - } else { /* Enlarge */ - FreeBlock **block; - void *old_end = (void *) (((char *)old_address) + old_real_size); - TAKE_LOCK(); - for (block = &first; - *block != NULL && (*block) < (FreeBlock *) old_address; - block = &((*block)->next)) - ; - if ((*block) == NULL || old_end > ((void *) RANGE_MAX) || - (*block) != old_end || - (*block)->num < (new_num_pages - old_num_pages)) { - /* cannot extend */ - void *result; - RELEASE_LOCK(); - result = pmmap(new_size); - if (result == NULL) { - return NULL; - } - memcpy(result,old_address,old_size); - if (pmunmap(old_address,old_size)) { - /* Oups... */ - pmunmap(result,new_size); - return NULL; - } - return result; - } else { /* extend */ - size_t remaining_pages = (*block)->num - - (new_num_pages - old_num_pages); - if (!remaining_pages) { - void *p = (void *) (((char *) (*block)) + MSEG_ALIGNED_SIZE); - void *n = (*block)->next; - size_t x = ((*block)->num - 1) * MSEG_ALIGNED_SIZE; - if (x > 0) { - if (do_map(p,x) == NULL) { - RELEASE_LOCK(); - return NULL; - } - } - (*block) = n; - } else { - FreeBlock *nfb = (FreeBlock *) ((void *) - (((char *) old_address) + - new_real_size)); - void *p = (void *) (((char *) (*block)) + MSEG_ALIGNED_SIZE); - if (do_map(p,new_real_size - old_real_size) == NULL) { - RELEASE_LOCK(); - return NULL; - } - nfb->num = remaining_pages; - nfb->next = (*block)->next; - (*block) = nfb; - } - RELEASE_LOCK(); - return old_address; - } - } -} -#endif /* HALFWORD_HEAP */ diff --git a/erts/emulator/sys/common/erl_mseg.h b/erts/emulator/sys/common/erl_mseg.h index 7454e5c473..2284b3f8f1 100644 --- a/erts/emulator/sys/common/erl_mseg.h +++ b/erts/emulator/sys/common/erl_mseg.h @@ -24,14 +24,14 @@ #include "erl_alloc_types.h" #include "erl_mmap.h" -#ifndef HAVE_MMAP -# define HAVE_MMAP 0 -#endif -#ifndef HAVE_MREMAP -# define HAVE_MREMAP 0 -#endif - -#if HAVE_MMAP +/* + * We currently only enable mseg_alloc if we got + * a genuine mmap()/munmap() primitive. It is possible + * to utilize erts_mmap() withiout a mmap support but + * alloc_util needs to be prepared before we can do + * that. + */ +#ifdef ERTS_HAVE_GENUINE_OS_MMAP # define HAVE_ERTS_MSEG 1 # define ERTS_HAVE_MSEG_SUPER_ALIGNED 1 #else @@ -40,8 +40,7 @@ #endif #if ERTS_HAVE_MSEG_SUPER_ALIGNED -# define MSEG_ALIGN_BITS (18) - /* Affects hard limits for sbct and lmbcs documented in erts_alloc.xml */ +# define MSEG_ALIGN_BITS ERTS_MMAP_SUPERALIGNED_BITS #else /* If we don't use super aligned multiblock carriers * we will mmap with page size alignment (and thus use corresponding @@ -77,7 +76,8 @@ typedef struct { 4*1024*1024, /* amcbf: Absolute max cache bad fit */ \ 20, /* rmcbf: Relative max cache bad fit */ \ 10, /* mcs: Max cache size */ \ - 1000 /* cci: Cache check interval */ \ + 1000, /* cci: Cache check interval */ \ + ERTS_MMAP_INIT_DEFAULT_INITER \ } typedef struct { @@ -93,12 +93,12 @@ typedef struct { extern const ErtsMsegOpt_t erts_mseg_default_opt; -void *erts_mseg_alloc(ErtsAlcType_t, Uint *, Uint); -void *erts_mseg_alloc_opt(ErtsAlcType_t, Uint *, Uint, const ErtsMsegOpt_t *); -void erts_mseg_dealloc(ErtsAlcType_t, void *, Uint, Uint); -void erts_mseg_dealloc_opt(ErtsAlcType_t, void *, Uint, Uint, const ErtsMsegOpt_t *); -void *erts_mseg_realloc(ErtsAlcType_t, void *, Uint, Uint *, Uint); -void *erts_mseg_realloc_opt(ErtsAlcType_t, void *, Uint, Uint *, Uint, const ErtsMsegOpt_t *); +void *erts_mseg_alloc(ErtsAlcType_t, UWord *, Uint); +void *erts_mseg_alloc_opt(ErtsAlcType_t, UWord *, Uint, const ErtsMsegOpt_t *); +void erts_mseg_dealloc(ErtsAlcType_t, void *, UWord, Uint); +void erts_mseg_dealloc_opt(ErtsAlcType_t, void *, UWord, Uint, const ErtsMsegOpt_t *); +void *erts_mseg_realloc(ErtsAlcType_t, void *, UWord, UWord *, Uint); +void *erts_mseg_realloc_opt(ErtsAlcType_t, void *, UWord, UWord *, Uint, const ErtsMsegOpt_t *); void erts_mseg_clear_cache(void); void erts_mseg_cache_check(void); Uint erts_mseg_no( const ErtsMsegOpt_t *); |