shared untraced heap; bugfix: mem corruptions in message passing code

lib/system.nim

@@ -977,23 +977,56 @@ proc alloc*(size: int): pointer {.noconv, rtl.}
   ## block has to be freed with ``realloc(block, 0)`` or
   ## ``dealloc(block)``. The block is not initialized, so reading
   ## from it before writing to it is undefined behaviour!
+  ## The allocated memory belongs to its allocating thread!
+  ## Use `allocShared` to allocate from a shared heap.
 proc alloc0*(size: int): pointer {.noconv, rtl.}
   ## allocates a new memory block with at least ``size`` bytes. The
   ## block has to be freed with ``realloc(block, 0)`` or
   ## ``dealloc(block)``. The block is initialized with all bytes
   ## containing zero, so it is somewhat safer than ``alloc``.
+  ## The allocated memory belongs to its allocating thread!
+  ## Use `allocShared0` to allocate from a shared heap.
 proc realloc*(p: Pointer, newsize: int): pointer {.noconv, rtl.}
   ## grows or shrinks a given memory block. If p is **nil** then a new
   ## memory block is returned. In either way the block has at least
   ## ``newsize`` bytes. If ``newsize == 0`` and p is not **nil**
   ## ``realloc`` calls ``dealloc(p)``. In other cases the block has to
   ## be freed with ``dealloc``.
+  ## The allocated memory belongs to its allocating thread!
+  ## Use `reallocShared` to reallocate from a shared heap.
 proc dealloc*(p: Pointer) {.noconv, rtl.}
   ## frees the memory allocated with ``alloc``, ``alloc0`` or
   ## ``realloc``. This procedure is dangerous! If one forgets to
   ## free the memory a leak occurs; if one tries to access freed
   ## memory (or just freeing it twice!) a core dump may happen
   ## or other memory may be corrupted. 
+  ## The freed memory must belong to its allocating thread!
+  ## Use `deallocShared` to deallocate from a shared heap.
+
+proc allocShared*(size: int): pointer {.noconv, rtl.}
+  ## allocates a new memory block on the shared heap with at
+  ## least ``size`` bytes. The block has to be freed with
+  ## ``reallocShared(block, 0)`` or ``deallocShared(block)``. The block
+  ## is not initialized, so reading from it before writing to it is 
+  ## undefined behaviour!
+proc allocShared0*(size: int): pointer {.noconv, rtl.}
+  ## allocates a new memory block on the shared heap with at 
+  ## least ``size`` bytes. The block has to be freed with
+  ## ``reallocShared(block, 0)`` or ``deallocShared(block)``.
+  ## The block is initialized with all bytes
+  ## containing zero, so it is somewhat safer than ``allocShared``.
+proc reallocShared*(p: Pointer, newsize: int): pointer {.noconv, rtl.}
+  ## grows or shrinks a given memory block on the heap. If p is **nil**
+  ## then a new memory block is returned. In either way the block has at least
+  ## ``newsize`` bytes. If ``newsize == 0`` and p is not **nil**
+  ## ``reallocShared`` calls ``deallocShared(p)``. In other cases the
+  ## block has to be freed with ``deallocShared``.
+proc deallocShared*(p: Pointer) {.noconv, rtl.}
+  ## frees the memory allocated with ``allocShared``, ``allocShared0`` or
+  ## ``reallocShared``. This procedure is dangerous! If one forgets to
+  ## free the memory a leak occurs; if one tries to access freed
+  ## memory (or just freeing it twice!) a core dump may happen
+  ## or other memory may be corrupted.
 
 proc assert*(cond: bool) {.magic: "Assert", noSideEffect.}
   ## provides a means to implement `programming by contracts`:idx: in Nimrod.

lib/system/alloc.nim

@@ -550,6 +550,35 @@ proc isAllocatedPtr(a: TMemRegion, p: pointer): bool =
         var c = cast[PBigChunk](c)
         result = p == addr(c.data) and cast[ptr TFreeCell](p).zeroField >% 1
 
+proc ptrSize(p: pointer): int =
+  var x = cast[pointer](cast[TAddress](p) -% sizeof(TFreeCell))
+  result = pageAddr(x).size - sizeof(TFreeCell)
+
+proc alloc(allocator: var TMemRegion, size: int): pointer =
+  result = rawAlloc(allocator, size+sizeof(TFreeCell))
+  cast[ptr TFreeCell](result).zeroField = 1 # mark it as used
+  sysAssert(not isAllocatedPtr(allocator, result))
+  result = cast[pointer](cast[TAddress](result) +% sizeof(TFreeCell))
+
+proc alloc0(allocator: var TMemRegion, size: int): pointer =
+  result = alloc(size)
+  zeroMem(result, size)
+
+proc dealloc(allocator: var TMemRegion, p: pointer) =
+  var x = cast[pointer](cast[TAddress](p) -% sizeof(TFreeCell))
+  sysAssert(cast[ptr TFreeCell](x).zeroField == 1)
+  rawDealloc(allocator, x)
+  sysAssert(not isAllocatedPtr(allocator, x))
+
+proc realloc(allocator: var TMemRegion, p: pointer, newsize: int): pointer =
+  if newsize > 0:
+    result = alloc(allocator, newsize)
+    if p != nil:
+      copyMem(result, p, ptrSize(p))
+      dealloc(allocator, p)
+  elif p != nil:
+    dealloc(allocator, p)
+
 proc deallocOsPages(a: var TMemRegion) =
   # we free every 'ordinarily' allocated page by iterating over the page bits:
   for p in elements(a.chunkStarts):
@@ -569,48 +598,17 @@ proc getOccupiedMem(a: TMemRegion): int {.inline.} =
 template InstantiateForRegion(allocator: expr) =
   proc deallocOsPages = deallocOsPages(allocator)
 
-  proc unlockedAlloc(size: int): pointer =
-    result = rawAlloc(allocator, size+sizeof(TFreeCell))
-    cast[ptr TFreeCell](result).zeroField = 1 # mark it as used
-    sysAssert(not isAllocatedPtr(allocator, result))
-    result = cast[pointer](cast[TAddress](result) +% sizeof(TFreeCell))
-
-  proc unlockedAlloc0(size: int): pointer =
-    result = unlockedAlloc(size)
-    zeroMem(result, size)
-
-  proc unlockedDealloc(p: pointer) =
-    var x = cast[pointer](cast[TAddress](p) -% sizeof(TFreeCell))
-    sysAssert(cast[ptr TFreeCell](x).zeroField == 1)
-    rawDealloc(allocator, x)
-    sysAssert(not isAllocatedPtr(allocator, x))
-
   proc alloc(size: int): pointer =
-    when hasThreadSupport and hasSharedHeap: AcquireSys(HeapLock)
-    result = unlockedAlloc(size)
-    when hasThreadSupport and hasSharedHeap: ReleaseSys(HeapLock)
+    result = alloc(allocator, size)
 
   proc alloc0(size: int): pointer =
-    result = alloc(size)
-    zeroMem(result, size)
+    result = alloc0(allocator, size)
 
   proc dealloc(p: pointer) =
-    when hasThreadSupport and hasSharedHeap: AcquireSys(HeapLock)
-    unlockedDealloc(p)
-    when hasThreadSupport and hasSharedHeap: ReleaseSys(HeapLock)
-
-  proc ptrSize(p: pointer): int =
-    var x = cast[pointer](cast[TAddress](p) -% sizeof(TFreeCell))
-    result = pageAddr(x).size - sizeof(TFreeCell)
+    dealloc(allocator, p)
 
   proc realloc(p: pointer, newsize: int): pointer =
-    if newsize > 0:
-      result = alloc(newsize)
-      if p != nil:
-        copyMem(result, p, ptrSize(p))
-        dealloc(p)
-    elif p != nil:
-      dealloc(p)
+    result = realloc(allocator, p, newsize)
 
   when false:
     proc countFreeMem(): int =
@@ -627,3 +625,37 @@ template InstantiateForRegion(allocator: expr) =
   proc getTotalMem(): int = return allocator.currMem
   proc getOccupiedMem(): int = return getTotalMem() - getFreeMem()
 
+  # -------------------- shared heap region ----------------------------------
+  when hasThreadSupport:
+    var sharedHeap: TMemRegion
+    var heapLock: TSysLock
+    InitSysLock(HeapLock)
+
+  proc allocShared(size: int): pointer =
+    when hasThreadSupport:
+      AcquireSys(HeapLock)
+      result = alloc(sharedHeap, size)
+      ReleaseSys(HeapLock)
+    else:
+      result = alloc(size)
+
+  proc allocShared0(size: int): pointer =
+    result = allocShared(size)
+    zeroMem(result, size)
+
+  proc deallocShared(p: pointer) =
+    when hasThreadSupport: 
+      AcquireSys(HeapLock)
+      dealloc(sharedHeap, p)
+      ReleaseSys(HeapLock)
+    else:
+      dealloc(p)
+
+  proc reallocShared(p: pointer, newsize: int): pointer =
+    when hasThreadSupport: 
+      AcquireSys(HeapLock)
+      result = realloc(sharedHeap, p, newsize)
+      ReleaseSys(HeapLock)
+    else:
+      result = realloc(p, newsize)
+

lib/system/cellsets.nim

@@ -47,9 +47,9 @@ proc contains(s: TCellSeq, c: PCell): bool {.inline.} =
 proc add(s: var TCellSeq, c: PCell) {.inline.} =
   if s.len >= s.cap:
     s.cap = s.cap * 3 div 2
-    var d = cast[PCellArray](unlockedAlloc(s.cap * sizeof(PCell)))
+    var d = cast[PCellArray](Alloc(s.cap * sizeof(PCell)))
     copyMem(d, s.d, s.len * sizeof(PCell))
-    unlockedDealloc(s.d)
+    Dealloc(s.d)
     s.d = d
     # XXX: realloc?
   s.d[s.len] = c
@@ -58,10 +58,10 @@ proc add(s: var TCellSeq, c: PCell) {.inline.} =
 proc init(s: var TCellSeq, cap: int = 1024) =
   s.len = 0
   s.cap = cap
-  s.d = cast[PCellArray](unlockedAlloc0(cap * sizeof(PCell)))
+  s.d = cast[PCellArray](Alloc0(cap * sizeof(PCell)))
 
 proc deinit(s: var TCellSeq) = 
-  unlockedDealloc(s.d)
+  Dealloc(s.d)
   s.d = nil
   s.len = 0
   s.cap = 0
@@ -70,7 +70,7 @@ const
   InitCellSetSize = 1024 # must be a power of two!
 
 proc Init(s: var TCellSet) =
-  s.data = cast[PPageDescArray](unlockedAlloc0(InitCellSetSize * sizeof(PPageDesc)))
+  s.data = cast[PPageDescArray](Alloc0(InitCellSetSize * sizeof(PPageDesc)))
   s.max = InitCellSetSize-1
   s.counter = 0
   s.head = nil
@@ -79,10 +79,10 @@ proc Deinit(s: var TCellSet) =
   var it = s.head
   while it != nil:
     var n = it.next
-    unlockedDealloc(it)
+    Dealloc(it)
     it = n
   s.head = nil # play it safe here
-  unlockedDealloc(s.data)
+  Dealloc(s.data)
   s.data = nil
   s.counter = 0
 
@@ -110,11 +110,11 @@ proc CellSetRawInsert(t: TCellSet, data: PPageDescArray, desc: PPageDesc) =
 proc CellSetEnlarge(t: var TCellSet) =
   var oldMax = t.max
   t.max = ((t.max+1)*2)-1
-  var n = cast[PPageDescArray](unlockedAlloc0((t.max + 1) * sizeof(PPageDesc)))
+  var n = cast[PPageDescArray](Alloc0((t.max + 1) * sizeof(PPageDesc)))
   for i in 0 .. oldmax:
     if t.data[i] != nil:
       CellSetRawInsert(t, n, t.data[i])
-  unlockedDealloc(t.data)
+  Dealloc(t.data)
   t.data = n
 
 proc CellSetPut(t: var TCellSet, key: TAddress): PPageDesc =
@@ -132,7 +132,7 @@ proc CellSetPut(t: var TCellSet, key: TAddress): PPageDesc =
   while t.data[h] != nil: h = nextTry(h, t.max)
   sysAssert(t.data[h] == nil)
   # the new page descriptor goes into result
-  result = cast[PPageDesc](unlockedAlloc0(sizeof(TPageDesc)))
+  result = cast[PPageDesc](Alloc0(sizeof(TPageDesc)))
   result.next = t.head
   result.key = key
   t.head = result

lib/system/inboxes.nim

@@ -77,7 +77,7 @@ proc storeAux(dest, src: Pointer, mt: PNimType, t: PInbox,
         x[] = nil
       else:
         var ss = cast[NimString](s2)
-        var ns = cast[NimString](rawAlloc(t.region, ss.len+1 + GenericSeqSize))
+        var ns = cast[NimString](Alloc(t.region, ss.len+1 + GenericSeqSize))
         copyMem(ns, ss, ss.len+1 + GenericSeqSize)
         x[] = ns
     else:
@@ -87,7 +87,7 @@ proc storeAux(dest, src: Pointer, mt: PNimType, t: PInbox,
         unsureAsgnRef(x, s2)
       else:
         unsureAsgnRef(x, copyString(cast[NimString](s2)))
-        rawDealloc(t.region, s2)
+        Dealloc(t.region, s2)
   of tySequence:
     var s2 = cast[ppointer](src)[]
     var seq = cast[PGenericSeq](s2)
@@ -100,7 +100,7 @@ proc storeAux(dest, src: Pointer, mt: PNimType, t: PInbox,
     else:
       sysAssert(dest != nil)
       if mode == mStore:
-        x[] = rawAlloc(t.region, seq.len *% mt.base.size +% GenericSeqSize)
+        x[] = Alloc(t.region, seq.len *% mt.base.size +% GenericSeqSize)
       else:
         unsureAsgnRef(x, newObj(mt, seq.len * mt.base.size + GenericSeqSize))
       var dst = cast[taddress](cast[ppointer](dest)[])
@@ -113,7 +113,7 @@ proc storeAux(dest, src: Pointer, mt: PNimType, t: PInbox,
       var dstseq = cast[PGenericSeq](dst)
       dstseq.len = seq.len
       dstseq.space = seq.len
-      if mode != mStore: rawDealloc(t.region, s2)
+      if mode != mStore: Dealloc(t.region, s2)
   of tyObject:
     # copy type field:
     var pint = cast[ptr PNimType](dest)
@@ -136,7 +136,7 @@ proc storeAux(dest, src: Pointer, mt: PNimType, t: PInbox,
         unsureAsgnRef(x, nil)
     else:
       if mode == mStore:
-        x[] = rawAlloc(t.region, mt.base.size)
+        x[] = Alloc(t.region, mt.base.size)
       else:
         # XXX we should use the dynamic type here too, but that is not stored in
         # the inbox at all --> use source[]'s object type? but how? we need a
@@ -144,7 +144,7 @@ proc storeAux(dest, src: Pointer, mt: PNimType, t: PInbox,
         var obj = newObj(mt.base, mt.base.size)
         unsureAsgnRef(x, obj)
       storeAux(x[], s, mt.base, t, mode)
-      if mode != mStore: rawDealloc(t.region, s)
+      if mode != mStore: Dealloc(t.region, s)
   else:
     copyMem(dest, src, mt.size) # copy raw bits
 
@@ -154,7 +154,7 @@ proc rawSend(q: PInbox, data: pointer, typ: PNimType) =
   if q.count >= cap:
     # start with capicity for 2 entries in the queue:
     if cap == 0: cap = 1
-    var n = cast[pbytes](rawAlloc0(q.region, cap*2*typ.size))
+    var n = cast[pbytes](Alloc0(q.region, cap*2*typ.size))
     var z = 0
     var i = q.rd
     var c = q.count
@@ -163,7 +163,7 @@ proc rawSend(q: PInbox, data: pointer, typ: PNimType) =
       copyMem(addr(n[z*typ.size]), addr(q.data[i*typ.size]), typ.size)
       i = (i + 1) and q.mask
       inc z
-    if q.data != nil: rawDealloc(q.region, q.data)
+    if q.data != nil: Dealloc(q.region, q.data)
     q.data = n
     q.mask = cap*2 - 1
     q.wr = q.count

lib/system/mmdisp.nim

@@ -88,12 +88,18 @@ when defined(boehmgc):
   proc realloc(p: Pointer, newsize: int): pointer =
     result = boehmRealloc(p, newsize)
     if result == nil: raiseOutOfMem()
-  proc dealloc(p: Pointer) =
-    boehmDealloc(p)
+  proc dealloc(p: Pointer) = boehmDealloc(p)
 
-  proc unlockedAlloc(size: int): pointer {.inline.} = result = alloc(size)
-  proc unlockedAlloc0(size: int): pointer {.inline.} = result = alloc0(size)
-  proc unlockedDealloc(p: pointer) {.inline.} = dealloc(p)
+  proc allocShared(size: int): pointer =
+    result = boehmAlloc(size)
+    if result == nil: raiseOutOfMem()
+  proc allocShared0(size: int): pointer =
+    result = alloc(size)
+    zeroMem(result, size)
+  proc reallocShared(p: Pointer, newsize: int): pointer =
+    result = boehmRealloc(p, newsize)
+    if result == nil: raiseOutOfMem()
+  proc deallocShared(p: Pointer) = boehmDealloc(p)
 
   proc initGC() = 
     when defined(macosx): boehmGCinit()
@@ -136,20 +142,13 @@ when defined(boehmgc):
   type
     TMemRegion = object {.final, pure.}
   
-  var
-    dummy {.rtlThreadVar.}: int
-  
-  proc rawAlloc(r: var TMemRegion, size: int): pointer =
+  proc Alloc(r: var TMemRegion, size: int): pointer =
     result = boehmAlloc(size)
     if result == nil: raiseOutOfMem()
-  proc rawAlloc0(r: var TMemRegion, size: int): pointer =
+  proc Alloc0(r: var TMemRegion, size: int): pointer =
     result = alloc(size)
     zeroMem(result, size)
-  proc realloc(r: var TMemRegion, p: Pointer, newsize: int): pointer =
-    result = boehmRealloc(p, newsize)
-    if result == nil: raiseOutOfMem()
-  proc rawDealloc(r: var TMemRegion, p: Pointer) = boehmDealloc(p)
-
+  proc Dealloc(r: var TMemRegion, p: Pointer) = boehmDealloc(p)  
   proc deallocOsPages(r: var TMemRegion) {.inline.} = nil
   proc deallocOsPages() {.inline.} = nil
 
@@ -204,10 +203,6 @@ elif defined(nogc):
 else:
   include "system/alloc"
 
-  proc unlockedAlloc(size: int): pointer {.inline.} 
-  proc unlockedAlloc0(size: int): pointer {.inline.} 
-  proc unlockedDealloc(p: pointer) {.inline.} 
-  
   include "system/cellsets"
   sysAssert(sizeof(TCell) == sizeof(TFreeCell))
   include "system/gc"

lib/system/threads.nim

@@ -210,10 +210,7 @@ when not defined(useNimRtl):
       echo "too large thread local storage size requested"
       quit 1
   
-  when hasSharedHeap:
-    var heapLock: TSysLock
-    InitSysLock(HeapLock)
-
+  when hasSharedHeap and not defined(boehmgc) and not defined(nogc):
     var
       threadList: PGcThread
       
@@ -275,19 +272,17 @@ template ThreadProcWrapperBody(closure: expr) =
   var t = cast[ptr TThread[TMsg]](closure)
   when useStackMaskHack:
     var tls: TThreadLocalStorage
-  when not defined(boehmgc) and not hasSharedHeap:
+  when not defined(boehmgc) and not defined(nogc) and not hasSharedHeap:
     # init the GC for this thread:
     setStackBottom(addr(t))
     initGC()
-  when hasSharedHeap:
+  when defined(registerThread):
     t.stackBottom = addr(t)
     registerThread(t)
   if t.emptyFn == nil: t.dataFn(t.data)
   else: t.emptyFn()
-  #finally:
-  # XXX shut-down is not executed when the thread is forced down!
   freeInbox(addr(t.inbox))
-  when hasSharedHeap: unregisterThread(t)
+  when defined(registerThread): unregisterThread(t)
   when defined(deallocOsPages): deallocOsPages()
   # Since an unhandled exception terminates the whole process (!), there is
   # no need for a ``try finally`` here, nor would it be correct: The current

todo.txt

@@ -1,21 +1,28 @@
-version 0.9.0
-=============
+Version 0.8.14
+==============
 
 - ``var T`` as a return type; easy to prove that location does not escape its
   stack frame
+- strutils.unindent
 - document Nimrod's two phase symbol lookup for generics
+- make ^ available as operator
+- make threadvar efficient again on linux after testing
+- test the sort implementation again
+- export re-entrant and non-reentrant locks and condition vars
+
+
+version 0.9.0
+=============
+
 - add --deadlock_prevention:on|off switch? timeout for locks?
 - bug: tfFinal not passed to generic
 - bug: forward proc for generic seems broken
-- test the sort implementation again
 - warning for implicit openArray -> varargs convention
 - implement explicit varargs
 - tests: run modules that contain "#RUN_ME", compile the other
   modules; run the GC tests
-- fix overloading resolution
-- make ^ available as operator
+- change overloading resolution
 - implement closures; implement proper coroutines
-- make threadvar efficient again on linux after testing
 
 Bugs
 ----

web/community.txt

@@ -6,3 +6,5 @@ Wiki: http://github.com/Araq/Nimrod/wiki.
 
 Bug reports: http://github.com/Araq/Nimrod/issues.
 
+For quickest feedback, join our IRC channel: irc://freenode/nimrod
+