mirror of
https://git.checksum.fail/alec/erythros
synced 2025-12-12 05:59:53 +02:00
Everywhere: Live patch MAlloc/Free to use RMAlloc/RFree
To make MAlloc/Free operations consistent across multiple processors, we use a dedicated task on core 5 to service the requests.
This commit is contained in:
446
System/Setup/MAlloc.HC
Normal file
446
System/Setup/MAlloc.HC
Normal file
@@ -0,0 +1,446 @@
|
||||
#define RMALLOC_REQUEST_PTR 0x280000
|
||||
|
||||
#define RMALLOC_BUSY_WAIT 5
|
||||
#define RMALLOC_TASK_CPU 5
|
||||
#define RMALLOC_TIMEOUT 1024
|
||||
#define RMALLOC_TYPE_MALLOC 0
|
||||
#define RMALLOC_TYPE_CALLOC 1
|
||||
#define RMALLOC_TYPE_FREE 2
|
||||
|
||||
asm {
|
||||
//************************************
|
||||
// See ::/Doc/Credits.DD.
|
||||
_RMALLOC::
|
||||
// Throws 'OutMem'
|
||||
PUSH RBP
|
||||
MOV RBP,RSP
|
||||
PUSH RSI
|
||||
PUSH RDI
|
||||
|
||||
XOR RBX,RBX
|
||||
MOV RDX,U64 SF_ARG2[RBP]
|
||||
TEST RDX,RDX
|
||||
JNZ @@05
|
||||
MOV RDX,U64 FS:CTask.addr[RBX]
|
||||
@@05: CMP U32 CTask.task_signature[RDX],TASK_SIGNATURE_VAL
|
||||
|
||||
#assert CTask.task_signature == CHeapCtrl.hc_signature // location signature same
|
||||
|
||||
JNE @@10
|
||||
MOV RDX,U64 CTask.data_heap[RDX]
|
||||
@@10: CMP U32 CHeapCtrl.hc_signature[RDX],HEAP_CTRL_SIGNATURE_VAL
|
||||
JE @@15
|
||||
PUSH RDX
|
||||
CALL &SysBadMAlloc
|
||||
JMP I32 _SYS_HLT
|
||||
|
||||
@@15: MOV RAX,U64 SF_ARG1[RBP]
|
||||
PUSHFD
|
||||
ADD RAX,CMemUsed.start+7 // round-up to I64
|
||||
AND AL,0xF8
|
||||
#assert CMemUsed.start >= sizeof(CMemUnused)
|
||||
CMP RAX,CMemUsed.start
|
||||
JAE @@20
|
||||
MOV RAX,CMemUsed.start
|
||||
@@20:
|
||||
|
||||
CLI
|
||||
@@25: LOCK
|
||||
BTS U32 CHeapCtrl.locked_flags[RDX],HClf_LOCKED
|
||||
PAUSE // don't know if this inst helps
|
||||
JC @@25
|
||||
|
||||
CMP RAX,MEM_HEAP_HASH_SIZE
|
||||
JAE @@30
|
||||
MOV RSI,U64 CHeapCtrl.heap_hash[RAX+RDX]
|
||||
TEST RSI,RSI
|
||||
JZ @@35
|
||||
MOV RCX,U64 CMemUnused.next[RSI]
|
||||
MOV U64 CHeapCtrl.heap_hash[RAX+RDX],RCX
|
||||
JMP I32 RMALLOC_ALMOST_DONE
|
||||
|
||||
// Big allocation
|
||||
@@30: ADD RAX,sizeof(CMemBlk)+MEM_PAG_SIZE-1
|
||||
SHR RAX,MEM_PAG_BITS
|
||||
|
||||
PUSH RDX // preserve HeapCtrl
|
||||
PUSH RDX
|
||||
PUSH RAX
|
||||
CALL &MemPagTaskAlloc
|
||||
POP RDX
|
||||
TEST RAX,RAX
|
||||
JZ @@45 // Out of memory
|
||||
MOV RSI,RAX
|
||||
MOV EAX,U32 CMemBlk.pags[RSI]
|
||||
|
||||
SHL RAX,MEM_PAG_BITS
|
||||
SUB RAX,sizeof(CMemBlk)
|
||||
ADD RSI,sizeof(CMemBlk)
|
||||
JMP I32 RMALLOC_ALMOST_DONE
|
||||
|
||||
// Little allocation, chunk-off piece from free lst chunks
|
||||
@@35: LEA RSI,U64 CHeapCtrl.malloc_free_lst-CMemUnused.next[RDX]
|
||||
|
||||
@@40: MOV RBX,RSI
|
||||
MOV RSI,U64 CMemUnused.next[RBX]
|
||||
TEST RSI,RSI
|
||||
JNZ I32 @@60
|
||||
PUSH RAX //-**** save byte size
|
||||
ADD RAX,16*MEM_PAG_SIZE-1
|
||||
SHR RAX,MEM_PAG_BITS
|
||||
|
||||
PUSH RDX // preserve HeapCtrl
|
||||
PUSH RDX
|
||||
PUSH RAX
|
||||
CALL &MemPagTaskAlloc
|
||||
POP RDX
|
||||
TEST RAX,RAX
|
||||
JNZ @@50
|
||||
|
||||
// Out of memory
|
||||
@@45: LOCK
|
||||
BTR U32 CHeapCtrl.locked_flags[RDX],HClf_LOCKED
|
||||
POPFD
|
||||
PUSH TRUE
|
||||
MOV RAX,'OutMem'
|
||||
PUSH RAX
|
||||
CALL I32 &throw
|
||||
JMP I32 RMALLOC_FINAL_EXIT // Never gets here, hopefully.
|
||||
|
||||
@@50: MOV RSI,RAX
|
||||
MOV EAX,U32 CMemBlk.pags[RSI]
|
||||
SHL RAX,MEM_PAG_BITS
|
||||
|
||||
// Can it be combined with last chunk? (Never Free these chunks.)
|
||||
MOV RDI,U64 CHeapCtrl.last_mergable[RDX]
|
||||
LEA RBX,U64 [RSI+RAX]
|
||||
CMP RDI,RBX
|
||||
JNE @@55
|
||||
|
||||
PUSH RAX
|
||||
MOV EAX,U32 CMemBlk.pags[RDI]
|
||||
ADD U32 CMemBlk.pags[RSI],EAX
|
||||
// QueRem
|
||||
MOV RAX,U64 CMemBlk.next[RDI]
|
||||
MOV RBX,U64 CMemBlk.last[RDI]
|
||||
MOV U64 CMemBlk.last[RAX],RBX
|
||||
MOV U64 CMemBlk.next[RBX],RAX
|
||||
POP RAX
|
||||
|
||||
@@55: MOV U64 CHeapCtrl.last_mergable[RDX],RSI
|
||||
LEA RSI,U64 sizeof(CMemBlk)[RSI]
|
||||
SUB RAX,sizeof(CMemBlk)
|
||||
LEA RBX,U64 CHeapCtrl.malloc_free_lst-CMemUnused.next[RDX]
|
||||
MOV RDI,U64 CMemUnused.next[RBX]
|
||||
MOV U64 CMemUnused.next[RSI],RDI
|
||||
MOV U64 CMemUnused.size[RSI],RAX
|
||||
MOV U64 CMemUnused.next[RBX],RSI
|
||||
POP RAX //+****
|
||||
JMP @@70
|
||||
@@60: CMP U64 CMemUnused.size[RSI],RAX
|
||||
JB I32 @@40
|
||||
JNE @@70
|
||||
|
||||
@@65: MOV RDI,U64 CMemUnused.next[RSI]
|
||||
MOV U64 CMemUnused.next[RBX],RDI
|
||||
JMP RMALLOC_ALMOST_DONE
|
||||
|
||||
@@70: SUB U64 CMemUnused.size[RSI],RAX // UPDATE FREE ENTRY
|
||||
CMP U64 CMemUnused.size[RSI],sizeof(CMemUnused)
|
||||
JAE @@75 // take from top of block
|
||||
ADD U64 CMemUnused.size[RSI],RAX // doesn't fit, undo
|
||||
JMP I32 @@40
|
||||
|
||||
@@75: ADD RSI,U64 CMemUnused.size[RSI]
|
||||
|
||||
RMALLOC_ALMOST_DONE:
|
||||
// RSI=res-CMemUsed.size
|
||||
// RAX=size+CMemUsed.size
|
||||
// RDX=HeapCtrl
|
||||
ADD U64 CHeapCtrl.used_u8s[RDX],RAX
|
||||
|
||||
#if _CFG_HEAP_DBG
|
||||
// QueIns
|
||||
MOV RDI,U64 CHeapCtrl.last_um[RDX]
|
||||
MOV U64 CMemUsed.next[RDI],RSI
|
||||
MOV U64 CHeapCtrl.last_um[RDX],RSI
|
||||
MOV U64 CMemUsed.last[RSI],RDI
|
||||
LEA RDI,U64 CHeapCtrl.next_um-CMemUsed.next[RDX]
|
||||
MOV U64 CMemUsed.next[RSI],RDI
|
||||
|
||||
// Caller1/Caller2
|
||||
PUSH RDX
|
||||
MOV RDX,U64 [MEM_HEAP_LIMIT]
|
||||
MOV RDI,U64 SF_RIP[RBP]
|
||||
CMP RDI,RDX
|
||||
JB @@80
|
||||
XOR RDI,RDI
|
||||
MOV U64 CMemUsed.caller1[RSI],RDI
|
||||
JMP @@90
|
||||
@@80: MOV U64 CMemUsed.caller1[RSI],RDI
|
||||
MOV RDI,U64 SF_RBP[RBP]
|
||||
CMP RDI,RDX
|
||||
JB @@85
|
||||
XOR RDI,RDI
|
||||
JMP @@90
|
||||
@@85: MOV RDI,U64 SF_RIP[RDI]
|
||||
CMP RDI,RDX
|
||||
JB @@90
|
||||
XOR RDI,RDI
|
||||
@@90: MOV U64 CMemUsed.caller2[RSI],RDI
|
||||
POP RDX
|
||||
|
||||
#endif
|
||||
LOCK
|
||||
BTR U32 CHeapCtrl.locked_flags[RDX],HClf_LOCKED
|
||||
POPFD
|
||||
|
||||
MOV U64 CMemUsed.size[RSI],RAX
|
||||
MOV U64 CMemUsed.hc[RSI],RDX
|
||||
LEA RAX,U64 CMemUsed.start[RSI]
|
||||
|
||||
TEST U8 [SYS_SEMAS+SEMA_HEAPLOG_ACTIVE*DFT_CACHE_LINE_WIDTH],1
|
||||
JZ @@105
|
||||
PUSH RAX
|
||||
PUSH RAX
|
||||
MOV RAX,U64 [SYS_EXTERN_TABLE]
|
||||
MOV RAX,U64 EXT_HEAPLOG_MALLOC*8[RAX]
|
||||
TEST RAX,RAX
|
||||
JZ @@95
|
||||
CALL RAX
|
||||
JMP @@100
|
||||
@@95: ADD RSP,8
|
||||
@@100: POP RAX
|
||||
|
||||
@@105: TEST U8 [SYS_HEAP_INIT_FLAG],1
|
||||
JZ RMALLOC_FINAL_EXIT
|
||||
|
||||
PUSH RAX
|
||||
MOV RCX,U64 CMemUsed.size-CMemUsed.start[RAX]
|
||||
SUB RCX,CMemUsed.start
|
||||
MOV RDI,RAX
|
||||
MOV AL,U8 [SYS_HEAP_INIT_VAL]
|
||||
REP_STOSB
|
||||
POP RAX
|
||||
|
||||
RMALLOC_FINAL_EXIT:
|
||||
POP RDI
|
||||
POP RSI
|
||||
POP RBP
|
||||
RET1 16
|
||||
//************************************
|
||||
_RFREE::
|
||||
// Be aware of heap_hash in MemPagTaskAlloc().
|
||||
PUSH RBP
|
||||
MOV RBP,RSP
|
||||
PUSH RSI
|
||||
PUSH RDI
|
||||
|
||||
TEST U8 [SYS_SEMAS+SEMA_HEAPLOG_ACTIVE*DFT_CACHE_LINE_WIDTH],1
|
||||
JZ @@15
|
||||
MOV RBX,U64 SF_ARG1[RBP]
|
||||
TEST RBX,RBX
|
||||
JZ @@05
|
||||
MOV RAX,U64 CMemUsed.size-CMemUsed.start[RBX]
|
||||
TEST RAX,RAX
|
||||
JGE @@05 // Aligned alloced chunks have neg size
|
||||
ADD RBX,RAX
|
||||
@@05: PUSH RBX
|
||||
MOV RAX,U64 [SYS_EXTERN_TABLE]
|
||||
MOV RAX,U64 EXT_HEAPLOG_FREE*8[RAX]
|
||||
TEST RAX,RAX
|
||||
JZ @@10
|
||||
CALL RAX
|
||||
JMP @@15
|
||||
@@10: ADD RSP,8
|
||||
|
||||
@@15: MOV RSI,U64 SF_ARG1[RBP]
|
||||
TEST RSI,RSI
|
||||
|
||||
#if _CFG_HEAP_DBG
|
||||
JZ I32 RFREE_DONE
|
||||
#else
|
||||
JZ RFREE_DONE
|
||||
#endif
|
||||
|
||||
MOV RAX,U64 CMemUsed.size-CMemUsed.start[RSI]
|
||||
TEST RAX,RAX
|
||||
JGE @@20 // Aligned alloced chunks have neg size.
|
||||
// The neg size is offset to start of CMemUsed struct.
|
||||
ADD RSI,RAX
|
||||
|
||||
@@20: PUSHFD
|
||||
SUB RSI,CMemUsed.start
|
||||
MOV RDX,U64 CMemUsed.hc[RSI]
|
||||
CMP U32 CHeapCtrl.hc_signature[RDX],HEAP_CTRL_SIGNATURE_VAL
|
||||
JE @@25
|
||||
ADD RSI,CMemUsed.start
|
||||
PUSH RSI
|
||||
CALL &SysBadFree
|
||||
JMP I32 _SYS_HLT
|
||||
|
||||
@@25: MOV RAX,U64 CMemUsed.size[RSI]
|
||||
SUB U64 CHeapCtrl.used_u8s[RDX],RAX
|
||||
CLI
|
||||
@@30: LOCK
|
||||
BTS U32 CHeapCtrl.locked_flags[RDX],HClf_LOCKED
|
||||
PAUSE
|
||||
JC @@30
|
||||
#if _CFG_HEAP_DBG
|
||||
// QueRem
|
||||
MOV RDX,U64 CMemUsed.next[RSI]
|
||||
MOV RDI,U64 CMemUsed.last[RSI]
|
||||
MOV U64 CMemUsed.last[RDX],RDI
|
||||
MOV U64 CMemUsed.next[RDI],RDX
|
||||
|
||||
// Caller1/Caller2
|
||||
MOV RDX,U64 [MEM_HEAP_LIMIT]
|
||||
MOV RDI,U64 SF_RIP[RBP]
|
||||
CMP RDI,RDX
|
||||
JB @@35
|
||||
XOR RDI,RDI
|
||||
MOV U64 CMemUnused.caller1[RSI],RDI
|
||||
JMP @@45
|
||||
@@35: MOV U64 CMemUnused.caller1[RSI],RDI
|
||||
MOV RDI,U64 SF_RBP[RBP]
|
||||
CMP RDI,RDX
|
||||
JB @@40
|
||||
XOR RDI,RDI
|
||||
JMP @@45
|
||||
@@40: MOV RDI,U64 SF_RIP[RDI]
|
||||
CMP RDI,RDX
|
||||
JB @@45
|
||||
XOR RDI,RDI
|
||||
@@45: MOV U64 CMemUnused.caller2[RSI],RDI
|
||||
|
||||
MOV RDX,U64 CMemUsed.hc[RSI]
|
||||
#endif
|
||||
CMP RAX,MEM_HEAP_HASH_SIZE
|
||||
JAE @@50
|
||||
|
||||
#assert CMemUnused.size == CMemUsed.size
|
||||
// MOV U64 CMemUnused.size[RSI],RAX
|
||||
|
||||
MOV RBX,U64 CHeapCtrl.heap_hash[RAX+RDX]
|
||||
MOV U64 CMemUnused.next[RSI],RBX
|
||||
MOV U64 CHeapCtrl.heap_hash[RAX+RDX],RSI
|
||||
JMP @@55
|
||||
|
||||
@@50: SUB RSI,sizeof(CMemBlk)
|
||||
PUSH RDX
|
||||
PUSH RDX
|
||||
PUSH RSI
|
||||
CALL &MemPagTaskFree
|
||||
POP RDX
|
||||
|
||||
@@55: LOCK
|
||||
BTR U32 CHeapCtrl.locked_flags[RDX],HClf_LOCKED
|
||||
POPFD
|
||||
RFREE_DONE:
|
||||
POP RDI
|
||||
POP RSI
|
||||
POP RBP
|
||||
RET1 8
|
||||
//************************************
|
||||
}
|
||||
|
||||
_extern _RMALLOC U8* RMAlloc(I64 size, CTask* mem_task = NULL);
|
||||
_extern _RFREE U0 RFree(U8* addr);
|
||||
|
||||
U8* RCAlloc(I64 size, CTask* mem_task = NULL)
|
||||
{
|
||||
U8* res = RMAlloc(size, mem_task);
|
||||
if (res) {
|
||||
MemSet(res, NULL, size);
|
||||
}
|
||||
return res;
|
||||
}
|
||||
|
||||
class @rmalloc_request
|
||||
{
|
||||
U64 res_addr;
|
||||
CTask* task;
|
||||
U64 bytes;
|
||||
U64 ptr;
|
||||
Bool type;
|
||||
Bool kick;
|
||||
};
|
||||
|
||||
U0 @rmalloc_task()
|
||||
{
|
||||
U64* ptr = RMALLOC_REQUEST_PTR;
|
||||
@rmalloc_request* r = NULL;
|
||||
while (1) {
|
||||
r = *ptr;
|
||||
if (r) {
|
||||
switch (r->type) {
|
||||
case RMALLOC_TYPE_MALLOC:
|
||||
LXchgI64(r->res_addr, RMAlloc(r->bytes, r->task));
|
||||
break;
|
||||
case RMALLOC_TYPE_CALLOC:
|
||||
LXchgI64(r->res_addr, RCAlloc(r->bytes, r->task));
|
||||
break;
|
||||
case RMALLOC_TYPE_FREE:
|
||||
RFree(r->ptr);
|
||||
break;
|
||||
default:
|
||||
break;
|
||||
}
|
||||
LXchgU8(&r->kick, 1);
|
||||
*ptr = NULL;
|
||||
}
|
||||
Yield;
|
||||
}
|
||||
}
|
||||
|
||||
U64 RMAllocOper(I64 size, CTask* task = NULL, Bool type = RMALLOC_TYPE_MALLOC, U64 addr = NULL)
|
||||
{
|
||||
I64 count = 0;
|
||||
U64 res = NULL;
|
||||
U64* ptr = RMALLOC_REQUEST_PTR;
|
||||
@rmalloc_request r;
|
||||
|
||||
r.res_addr = &res;
|
||||
r.bytes = size;
|
||||
r.task = task;
|
||||
r.ptr = addr;
|
||||
r.type = type;
|
||||
r.kick = FALSE;
|
||||
|
||||
retry_rmalloc:
|
||||
count = 0;
|
||||
while (*ptr)
|
||||
Busy(RMALLOC_BUSY_WAIT);
|
||||
LXchgI64(ptr, &r);
|
||||
while (!r.kick) {
|
||||
++count;
|
||||
if (count > RMALLOC_TIMEOUT)
|
||||
goto retry_rmalloc;
|
||||
Busy(RMALLOC_BUSY_WAIT);
|
||||
}
|
||||
|
||||
return res;
|
||||
}
|
||||
|
||||
U64 MAlloc2(I64 size, CTask* task = NULL)
|
||||
{
|
||||
return RMAllocOper(size, task);
|
||||
}
|
||||
|
||||
U64 CAlloc2(I64 size, CTask* task = NULL)
|
||||
{
|
||||
return RMAllocOper(size, task, RMALLOC_TYPE_CALLOC);
|
||||
}
|
||||
|
||||
U0 Free2(U8* addr)
|
||||
{
|
||||
RMAllocOper(NULL, NULL, RMALLOC_TYPE_FREE, addr);
|
||||
}
|
||||
|
||||
CTask* sys_malloc_task = Spawn(&@rmalloc_task, , "RMAlloc Task", RMALLOC_TASK_CPU);
|
||||
|
||||
Suspend(sys_winmgr_task);
|
||||
@patch_jmp_rel32(&MAlloc, &MAlloc2);
|
||||
@patch_jmp_rel32(&CAlloc, &CAlloc2);
|
||||
@patch_jmp_rel32(&Free, &Free2);
|
||||
Suspend(sys_winmgr_task, FALSE);
|
||||
Reference in New Issue
Block a user