1 %************************************************************************
3 \section[SMinterface.lh]{Main storage manager interface}
5 %************************************************************************
7 %% I have changed most of the text here, in an attempt to understand
8 %% what's going on. Please let me know about any mistakes, so that
9 %% I can correct them! KH@15/10/92 (UK)
11 %% I have also split the original monster into SMinterface.lh,
12 %% SMClosures.lh and SMInfoTables.lh. The latter two are
15 This describes the interface used between the STG-machine
16 reducer and the storage manager. The overriding goal is to isolate
17 the implementation details of each from the other.
19 Multi-slurp protection:
26 {}\input{epsf} % Uses encapsulated PostScript diagrams
29 %************************************************************************
31 \subsection[SM-calling-interface]{Calling interface}
33 %************************************************************************
35 The @smInfo@ structure is used to pass all information back and forth
36 between the storage manager and the STG world.
38 WARNING: If you modify this structure, you {\em must} modify the
39 native-code generator as well, because the offsets for various fields
40 are hard-coded into the NCG. (In nativeGen/StixMacro.lhs).
44 P_ hp; /* last successfully allocated word */
45 P_ hplim; /* last allocatable word */
47 I_ rootno; /* No of heap roots stored in roots */
48 P_ *roots; /* Array of heap roots -- must be allocated (not static) */
49 P_ CAFlist; /* List of updated CAF's */
51 #if defined(GCap) || defined(GCgn)
52 P_ OldMutables; /* List of old generation mutable closures */
53 P_ OldLim; /* Ptr to end of the old generation */
57 P_ MallocPtrList; /* List of all Malloc Pointers (in new generation) */
59 #if defined(GCap) || defined(GCgn)
60 P_ OldMallocPtrList; /* List of all Malloc Pointers in old generation */
63 P_ StablePointerTable;
64 /* Heap allocated table used to store stable pointers in */
67 I_ hardHpOverflowSize; /* Some slop at the top of the heap which
68 (hopefully) provides enough space to let
69 us recover from heap overflow exceptions */
72 extern smInfo StorageMgrInfo;
76 Maximum number of roots storable in the heap roots array.
77 Question: Where are the stable pointer roots? (JSM)
78 Answer: They're on the heap in a "Stable Pointer Table". (ADR)
81 # define SM_MAXROOTS 8 /* 8 Vanilla Regs */
85 # define SM_MAXROOTS (10 + (MAX_PROC*4) + 2 + (MAX_PROC*2) + MAX_SPARKS)
86 /* unthreaded + spark/thread queues + Current/Main TSOs
89 # define SM_MAXROOTS 5 /* See c-as-asm/HpOverflow.lc */
92 # define SM_MAXROOTS 6 /* See c-as-asm/HpOverflow.lc */
97 The storage manager is accessed exclusively through these routines:
99 IF_RTS(void initSM (STG_NO_ARGS);)
100 IF_RTS(rtsBool exitSM PROTO((smInfo *sm));)
101 IF_RTS(rtsBool initStacks PROTO((smInfo *sm));)
102 IF_RTS(rtsBool initHeap PROTO((smInfo *sm));)
104 IF_RTS(rtsBool initThreadPools (STG_NO_ARGS);)
107 IF_RTS(void init_gr_profiling PROTO((int, char **, int, char **));)
110 I_ collectHeap PROTO((W_ reqsize, smInfo *sm, rtsBool do_full_collection));
112 IF_RTS(void unmapMiddleStackPage PROTO((char *, int));) /* char * == caddr_t ? */
114 #if defined(PROFILING) || defined(PAR)
115 IF_RTS(void handle_tick_serial(STG_NO_ARGS);)
116 IF_RTS(void handle_tick_noserial(STG_NO_ARGS);)
119 /* EXTFUN(_startMarkWorld); */
121 StgDouble usertime(STG_NO_ARGS);
122 StgDouble elapsedtime(STG_NO_ARGS);
123 void start_time(STG_NO_ARGS);
124 void end_init(STG_NO_ARGS);
127 void EvacuateLocalGAs PROTO((rtsBool full));
128 void RebuildGAtables PROTO((rtsBool full));
133 @initSM@ finalizes any runtime parameters of the storage manager.
135 @exitSM@ does any cleaning up required by the storage manager before
136 the program is executed. Its main purpose is to print any summary
139 @initStacks@ allocates the A and B stacks (sequential only). It
140 initialises the @spa@, @spb@, @sua@, and @sub@ fields of @sm@
141 appropriately for empty stacks. Successive calls to @initStacks@
142 re-initialise the stacks.
144 @initHeap@ allocates the heap. It initialises the @hp@ and @hplim@
145 fields of @sm@ to represent an empty heap for the compiled-in garbage
146 collector. It also allocates the @roots@ array for later use within
147 @collectHeap@, and initialises @CAFlist@ to be the empty list. The
148 @roots@ array must be large enough to hold at least @SM_MAXROOTS@
149 roots. If we are using Appel's collector it also initialises the
152 In the sequential system, it also initialises the stable pointer table
153 and the @MallocPtr@ (and @OldMallocPtrList@) fields.
155 @collectHeap@ invokes the garbage collector that was requested at
156 compile time. @reqsize@ is the size of the request (in words) that
157 resulted in the overflow. If the garbage collection succeeds, then at
158 least @reqsize@ words will be available. @collectHeap@ requires all
159 the fields of @sm@ to be initialised appropriately (from the
160 STG-machine registers). The following are identified as
163 \item The @roots@ array.
164 \item The updated CAFs recorded in @CAFlist@.
166 \item Update frames on the B Stack. These may be ``squeezed'' out
167 if they are the only reference to a closure --- thus avoiding the
169 \item The stable pointer table. (In sequential system.)
172 There are three possible results from a garbage collection:
174 \item[\tr{GC_HARD_LIMIT_EXCEEDED} (\tr{reqsize > hplim - hp})]
175 The heap size exceeds the hard heap limit: we report an error and
178 \item[\tr{GC_SOFT_LIMIT_EXCEEDED} (\tr{reqsize + hardHpOverflowSize > hplim - hp})]
179 The heap size exceeds the soft heap limit: set \tr{hardHpOverflowSize}
180 to \tr{0} so that we can use the overflow space, unwind the stack and
181 call an appropriate piece of Haskell to handle the error.
183 \item[\tr{GC_SUCCESS} (\tr{reqsize + hardHpOverflowSize <= hplim - hp})]
184 The heap size is less than the soft heap limit.
187 \item @hp@ and @hplim@ will indicate the new space available for
188 allocation. But we'll subtract \tr{hardHpOverflowSize} from
189 \tr{hplim} so that we'll GC when we hit the soft limit.
191 \item The elements of the @roots@ array will point to the new
192 locations of the closures.
194 \item @spb@ and @sub@ will be updated to reflect the new state of the
195 B stack arising from any update frame ``squeezing'' [sequential only].
197 \item The elements of @CAFlist@ and the stable pointers will be
198 updated to point to the new locations of the closures they reference.
200 \item Any members of @MallocPtrList@ which became garbage should have
201 been reported (by calling @FreeMallocPtr@; and the @(Old)MallocPtrList@
202 updated to contain only those Malloc Pointers which are still live.
208 #define GC_HARD_LIMIT_EXCEEDED 0
209 #define GC_SOFT_LIMIT_EXCEEDED 1
213 %************************************************************************
215 \subsection[SM-what-really-happens]{``What really happens in a garbage collection?''}
217 %************************************************************************
219 This is a brief tutorial on ``what really happens'' going to/from the
220 storage manager in a garbage collection.
223 %------------------------------------------------------------------------
224 \item[The heap check:]
228 If you gaze into the C output of GHC, you see many macros calls like:
230 HEAP_CHK_2PtrsLive((_FHS+2));
233 This expands into the C (roughly speaking...):
235 Hp = Hp + (_FHS+2); /* optimistically move heap pointer forward */
237 GC_WHILE_OR_IF (HEAP_OVERFLOW_OP(Hp, HpLim) OR_INTERVAL_EXPIRED) {
238 STGCALL2_GC(PerformGC, <liveness-bits>, (_FHS+2));
239 /* Heap full. Call "PerformGC" with 2 arguments, "<liveness>",
240 (info about what ptrs are live) and "_FHS+2" (words
241 requested), via the magical routine "callWrapper_GC",
242 which indicates ``I am calling a routine in which GC
243 may happen'' (a safe bet for `PerformGC').
248 In the parallel world, where we will need to re-try the heap check,
249 @GC_WHILE_OR_IF@ will be a ``while''; in the sequential world, it will
252 The ``heap lookahead'' checks, which are similar and used for
253 multi-precision @Integer@ ops, have some further complications. See
254 the commentary there (\tr{StgMacros.lh}).
256 %------------------------------------------------------------------------
257 \item[Into @callWrapper_GC@...:]
259 When we failed the heap check (above), we were inside the
260 GCC-registerised ``threaded world.'' @callWrapper_GC@ is all about
261 getting in and out of the threaded world. On SPARCs, with register
262 windows, the name of the game is not shifting windows until we have
263 what we want out of the old one. In tricky cases like this, it's best
264 written in assembly language.
266 Though the principle of ``save everything away'' is the same in both
267 the sequential and parallel worlds, the details are different.
269 For the sequential world:
272 @callWrapper_GC@ saves the return address.
274 It saves the arguments passed to it (so it doesn't get lost).
276 Save the machine registers used in the STG threaded world in their
277 \tr{*_SAVE} global-variable backup locations. E.g., register \tr{Hp}
278 is saved into \tr{Hp_SAVE}.
280 Call the routine it was asked to call; in this example, call
281 @PerformGC@ with arguments \tr{<liveness>}, and @_FHS+2@ (some constant)...
284 For the parallel world, a GC means giving up the thread of control.
285 So we must fill in the thread-state-object (TSO) [and its associated
286 stk object] with enough information for later resumption:
289 Save the return address in the TSO's PC field.
291 Save the machine registers used in the STG threaded world in their
292 corresponding TSO fields. We also save the pointer-liveness
293 information in the TSO.
295 The registers that are not thread-specific, notably \tr{Hp} and
296 \tr{HpLim}, are saved in the @StorageMgrInfo@ structure.
298 Call the routine it was asked to call; in this example, call
299 @PerformGC@ with arguments \tr{<liveness>} and @_FHS+2@ (some constant)...
301 (In the parallel world, we don't expect it to return...)
304 %------------------------------------------------------------------------
305 \item[Into the heap overflow wrapper, @PerformGC@ [sequential]:]
307 The first argument (\tr{<liveness>}, in our example) say what registers
308 are live, i.e., are ``roots'' the storage manager needs to know.
310 StorageMgrInfo.rootno = 2;
311 StorageMgrInfo.roots[0] = (P_) Ret1_SAVE;
312 StorageMgrInfo.roots[1] = (P_) Ret2_SAVE;
315 We further: (a)~move the heap-pointer back [we had optimistically
316 advanced it, in the initial heap check], (b)~load up the @smInfo@ data
317 from the STG registers' \tr{*_SAVE} locations, and (c)~FINALLY: call
320 IT IS AT THIS POINT THAT THE WORLD IS COMPLETELY TIDY.
322 %------------------------------------------------------------------------
323 \item[Into the heap overflow wrapper, @PerformGC@ [parallel]:]
325 Parallel execution is only slightly different. Most information has
326 already been saved in the TSO.
330 We still need to set up the storage manager's @roots@ array.
332 We mark on the scheduler's big ``blackboard'' that a GC is
335 We reschedule, i.e., this thread gives up control. (The scheduler
336 will presumably initiate a garbage-collection, but it may have to do
337 any number of other things---flushing, for example---before ``normal
338 execution'' resumes; and it most certainly may not be this thread that
339 resumes at that point!)
342 %------------------------------------------------------------------------
343 \item[Into/out of @collectHeap@ [sequential only]:]
345 @collectHeap@ does the business and reports back whether it freed up
348 %------------------------------------------------------------------------
349 \item[Out of the heap overflow wrapper, @PerformGC@ [sequential only]:]
351 We begin our return back to doing useful work by: (a)~reloading the
352 appropriate STG-register \tr{*_SAVE} locations from (presumably
353 changed) @smInfo@; (b) re-advance the heap-pointer---which we've been
354 trying to do for a week or two---now that there is enough space.
356 We must further restore appropriate @Ret?@ registers from the storage
357 manager's roots array; in this example:
360 Ret1_SAVE = (W_) StorageMgrInfo.roots[0];
361 Ret2_SAVE = (W_) StorageMgrInfo.roots[1];
364 %------------------------------------------------------------------------
365 \item[Out of @callWrapper_GC@ [sequential]:]
367 We pop out of heap-overflow code and are ready to resume STG
368 ``threaded world'' stuff.
370 The main thing is to re-load up the GCC-ised machine registers from
371 the relevant \tr{*_SAVE} locations; e.g., \tr{SpA} from \tr{SpA_SAVE}.
373 To conclude, @callWrapper_GC@ merely {\em jumps} back to the return
374 address which it was given originally.
376 WE'RE BACK IN (SEQUENTIAL) BUSINESS.
378 %------------------------------------------------------------------------
379 \item[Out of @callWrapper_GC@ [parallel]:]
381 When this thread is finally resumed after GC (and who knows what
382 else), it will restart by the normal enter-TSO/enter-stack-object
383 sequence, which has the effect of re-loading the registers, etc.,
384 (i.e., restoring the state).
386 Because the address we saved in the TSO's PC field was that at the end
387 of the heap check, and because the check is a while-loop in the
388 parallel system, we will now loop back around, and make sure there is
389 enough space before continuing.
392 %************************************************************************
394 \subsection[SM-stack-info]{Stacks}
396 %************************************************************************
398 There are two stacks, as in the STG paper \cite{new-stg-paper}.
401 The A stack contains only closure pointers.
403 The B stack contains, basic values, return addresses, and update
406 The A stack and B stack grow towards each other, so they overflow when
407 they collide. Currently the A stack grows downward (towards lower
408 addresses); the B stack grows upward. (We localise the stuff which
409 uses this information within macros defined in @StgDirections.h@)
411 During reduction, SpA and SpB point to the topmost allocated word of
412 the corresponding stack (though they may not be up to date in the
413 middle of a basic block).
415 Each stack also has a {\em stack update pointer}, SuA and SuB, which
416 point to the topmost word of the most recent update frame in the
417 corresponding stack. (Colloquially, SuA and Sub point to the first
418 items on their respective stacks ``that you cannot have.'')
420 A standard update frame (on the B stack) looks like this
421 (stack grows downward in this picture):
423 \mbox{\epsffile{update-frame.ps}}
425 The SuB therefore points to the Update return vector component of
426 the topmost update frame.
429 A {\em constructor} update frame, which is pushed only by closures
430 which know they will evaluate to a data object, looks just the
431 same, but without the saved SuA pointer.
433 We store the following information concerning the stacks in a global
434 structure. (sequential only).
438 PP_ botA; /* Points to bottom-most word of A stack */
439 P_ botB; /* Points to bottom-most word of B stack */
442 extern stackData stackInfo;
445 %************************************************************************
447 \subsection[SM-choose-flavour]{Deciding which GC flavour is in force...}
449 %************************************************************************
451 Each garbage collector requires different garbage collection entries
456 #define _INFO_COPYING
460 #define _INFO_COMPACTING
461 #define _INFO_MARKING
464 #if defined(GCdu) || defined (GCap) || defined (GCgn)
465 #define _INFO_COPYING
466 #define _INFO_COMPACTING
467 #define _INFO_MARKING
470 /* NO_INFO_SPECIFIED (ToDo: an #error ???) */
476 %************************************************************************
478 %\subsection[Info.lh]{Info Pointer Definitions}
480 %************************************************************************
486 %************************************************************************
488 %\subsection[Parallel.lh]{Parallel Machine Definitions}
490 %************************************************************************
497 %************************************************************************
499 %\subsection[CostCentre.lh]{Profiling Definitions}
501 %************************************************************************
504 \input{CostCentre.lh}
508 %************************************************************************
510 %\subsection[SM-closures]{Closure-Related Definitions}
512 %************************************************************************
515 \input{SMClosures.lh}
520 %************************************************************************
522 %\subsection[SM-info-tables]{Info-table Related Definitions}
524 %************************************************************************
527 \input{SMInfoTables.lh}
531 End multi-slurp protection:
533 #endif /* SMinterface_H */