Ravenbrook / Projects / Memory Pool System / Master Product Sources / Design Documents

Memory Pool System Project


                 DEBUGGING FEATURES FOR CLIENT OBJECTS
                        design.mps.object-debug
                             incomplete doc
                            pekka 1998-09-10

INTRODUCTION

.intro: This is the design for all the various debugging features that MPS 
clients (and sometimes MPS developers) can use to discover what is happening to 
their objects and the memory space.

.readership: MPS developers.


Document History

.hist.0: The first draft merely records all the various ideas about 
fenceposting that came up in discussions in June, July and September 1998.  
This includes the format wrapping idea from mail.ptw.1998-06-19.21-13(0).  
pekka 1998-09-10


OVERVIEW

.over.fenceposts: In its current state, this document mostly talks about 
fenceposts, straying a little into tagging where theses features have an effect 
on each other.  [There exist other documents that list other required features, 
and propose interfaces and implementations.  These will eventually be folded 
into this one.  pekka 1998-09-10]


REQUIREMENTS

.req.fencepost: Try to detect overwrites and underwrites of allocated blocks by 
adding fenceposts (source req.product.??? VC++, req.epcore.fun.debug.support).  
.req.fencepost.size: The fenceposts should be at least 4 bytes on either side 
or 8 bytes if on one side only, with an adjustable content (although VC++ only 
has 4 bytes with pattern 0xFDFDFDFD, having unwisely combined the 
implementation with other debug features).

.req.fencepost.check: There should be a function to check all the fenceposts 
(source req.epcore.fun.debug.support).

.req.free-block: Try to detect attempts to write and read free blocks.

.req.walk: There should be a way to map ("walk") a user function over all 
allocated objects (except PS VM objects), possibly only in a separate debugging 
variety/mode (source req.epcore.fun.debug.support).

.req.tag: There should be a way to store at least a word of user data (a "tag", 
borrowing the SW term) with every object in debugging mode, to be used in 
memory dumps (source req.product.??? VC++).  .req.tag.walk: The walking 
function (as required by .req.walk) should have access to this data (source 
req.epcore.fun.debug.support).

.req.dump.aver: It must be possible to perform a memory dump after an AVER has 
fired (naturally, if the information required for the dump has been corrupted, 
it will fail, as softly as possible).  (source @@@@)

[There are more, especially about memory dumps and allocation locations.  pekka 
1998-09-10]


SOLUTION IDEAS

.note.assumptions: I've tried not to assume anything about the coincidence of 
manual/automatic, formatted/unformatted, and ap/mps_alloc.  I think those 
questions deserve to be decided on their own merits. instead of being 
constrained by a debug feature.

.fence.content.repeat: The content of a fencepost could be specified as a 
byte/word which used repeatedly to fill the fencepost.  
.fence.content.template: The content could be given as a template which is of 
the right size and is simply copied onto the fencepost.

.fence.walk: .req.fencepost.check requires the ability to find all the 
allocated objects.  In formatted pools, this is not a problem.  In unformatted 
pools, we could use the walker.  It's a feasible strategy to bet that any pool 
that might have to support fenceposting will also have a walking requirement.  
.fence.tag: Fenceposting also needs to keep track which objects have 
fenceposts. unless we manage to do them all.  It would be easiest to put this 
in the tags.

.fence.check.object: A function to check the fenceposts on a given object would 
be nice.

.fence.ap: AP's could support fenceposting transparently by having a mode where 
mps_reserve always goes out-of-line and fills in the fenceposts (the pool's 
BufferFill method isn't involved).  This would leave the MPS with more freedom 
of implementation, especially when combined with some of the other ideas.  We 
think doing a function call for every allocation is not too bad for debugging.

.fence.outside-ap: We could also let the client insert their own fenceposts 
outside the MPS allocation mechanism.  Even if fenceposting were done like 
this, we'd still want it to be an MPS feature, so we'd offer sample C macros 
for adding the size of the fencepost and filling in the fencepost pattern.  
Possibly something like this (while we could still store the parameters in the 
pool or ap, there seems little point in doing so in this case, and having them 
as explicit parameters to the macros allows the client to specify constants to 
gain effiency):
  #define mps_add_fencepost(size, fp_size)
  #define mps_fill_fenceposts(obj, size, fp_size, fp_pattern)
The client would need to supply their own fencepost checking function, 
obviously, but again we could offer one that matches the sample macros.

.fence.tail-only: In automatic pools, the presence of a fencepost at the head 
of the allocated block results in the object reference being an internal 
pointer.  This means that the format or the pool would need to know about 
fenceposting and convert between references and pointers.  This would slow down 
the critical path when fenceposting is used.  This can be ameliorated by 
putting a fencepost at the tail of
the block only: this obviates the internal pointer problem and could provide 
almost the same degree of checking (provided the size was twice as large), 
especially in copying pools, where there are normally no gaps between allocated 
blocks.  In addition to the inescapable effects on allocation and freeing 
(including copying and reclaim thereunder), only scanning would have to know 
about fenceposts.  .fence.tail-only.under: Walking over all the objects in the 
pool would be necessary to detect underwrites, as one couldn't be sure that 
there is a fencepost before any given object (or where it's located exactly).  
If the pool were doing the checking, it could be sure: it would know about 
alignments and it could put fenceposts in padding objects (free blocks will 
have them because they were once allocated) so there'd be one on either side of 
any object (except at the head of a segment, which is not a major problem, and 
could be fixed by adding a padding object at the beginning of every segment).  
This requires some cleverness to avoid splinters smaller than the fencepost 
size, but it can be done.

.fence.wrapper: On formatted pools, fenceposting could be implemented by 
"wrapping" the client-supplied format at creation time.  The wrapper can handle 
the conversion from the fenceposted object and back.  This will be invisible to 
the client and gives the added benefit that the wrapper can validate fenceposts 
on every format operation, should it desire.  That is, the pool would see the 
fenceposts as part of the client object, but the client would only see its 
object; the format wrapper would translate between the two.  Note that hiding 
the fenceposts from scan methods, which are required to take a contiguous range 
of objects, is a bit complicated.  .fence.client-format: The MPS would supply 
such a wrapper, but clients could also be allowed to write their own 
fenceposted formats (provided they coordinate with allocation, see below).  
This would make scanning fenceposted segments more efficient.

.fence.wrapper.variable: Furthermore, you could create different classes of 
fencepost within a pool, because the fencepost itself could have a variable 
format.  For instance, you might choose to have the fencepost be minimal (1-2 
words) for small objects, and more detailed/complex for large objects 
(imagining that large objects are likely vector-ish and subject to overruns).  
You could get really
fancy and have the fencepost class keyed to the object class (e.g., different 
AP's create different classes of fenceposting).

.fence.wrapper.alloc: Even with a wrapped format, allocation and freeing would 
still have know about the fenceposts.  If allocation points are used, either 
MPS-side (.fence.ap) or client-side (.fence.outside-ap) fenceposting could be 
used, with the obvious modifications.  .fence.wrapper.alloc.format: We could 
add three format methods, to adjust the pointer and the size for alloc and 
free, to put down the fenceposts during alloc, and to check them; to avoid 
slowing down all allocation, this would require some MOPping to make the format 
class affect the choice of the alloc and free methods (see 
mail.pekka.1998-06-11.18-18).  .fence.wrapper.alloc.size: We could just 
communicate the size of the fenceposts between the format and the allocation 
routines, but then you couldn't use variable fenceposts 
(.fence.wrapper.variable).  [All this applies to copying and reclaim
in a straight-forward manner, I think.]

.fence.pool.wrapper: Pools can be wrapped as well.  This could be a natural way 
to represent/implement the fenceposting changes to the Alloc and Free methods.  
[@@@@alignment]

.fence.pool.new-class: We could simply offer a debugging version of each pool 
class (e.g., mps_pool_class_mv_debug).  As we have seen, debugging features 
have synergies which make it advantageous to have a coordinated implementation, 
so splitting them up would not just complicate the client interface, it would 
also be an implementation problem; we can turn features on or off with pool 
init parameters.

.fence.pool.abstract: We could simply use pool init parameters only to control 
all debugging features (optargs would be useful here).  While there migh be 
subclasses and wrappers internally, the client would only see a single pool 
class; in the internal view, this would be an abstract class, and the 
parameters would determine which concrete class actually gets instantiated.

.tag.out-of-line: It would be nice if tags were stored out-of-line, so they can 
be used to study allocation patterns and fragmentation behaviours.  Such an  
implementation of tagging could also easily be shared among several pools.


ARCHITECTURE

.pool: The implementation is at the pool level, because pools manage allocated 
objects.  A lot of the code will be generic, naturally, but the data structures 
and the control interfaces attach to pools.  In particular, clients will be 
able to use tagging and fenceposting separately on each pool.

.fence.size: Having fenceposts of adjustable size and pattern is quite useful.  
We feel that restricting the size to an integral multiple of the [pool or 
format?] alignment is harmless and simplifies the implementation enormously.

.fence.template: We use templates (.fence.content.template) to fill in the 
fenceposts, but we do not give any guarantees about the location of the 
fenceposts, only that they're properly aligned.  This leaves us the opportunity 
to do tail-only fenceposting, if we choose.

.fence.slop: [see impl.c.dbgpool.FenceAlloc @@@@]

.fence.check.free: We check the fenceposts when freeing an object.

.unified-walk: Combine the walking and tagging requirements (.req.tag.walk and 
@@@@) into a generic facility for walking and tagging objects with just one 
interface and one name: tagging.  Also combine the existing formatted object 
walker into this metaphor, but allowing the format and tag parameters of the 
step function be optional [this part has not been implemented yet  pekka 
1998-09-10].

.init: It simplifies the implementation of both tagging and fenceposting if 
they are always on, so that we don't have to keep track of which objects have 
been fenceposted and which have not, and don't have to have three kinds of 
tags: for user data, for fenceposting, and for both.  So we determine this at 
pool init time (and let fenceposting turn on tagging, if necessary).

.pool-parameters: Fencepost templates and tag formats are passed in as pool 
parameters.

.modularity: While a combined generic implementation of tags and fenceposts is 
provided, it is structured so that each part of it could be implemented by a 
pool-specific mechanism with a minimum of new protocol.  [This will be 
improved, when we figure out formatted pools -- they don't need tags for 
fenceposting.]

.out-of-space: If there's no room for tags, we will not dip into the reservoir, 
just fail to allocate the tag.  If the alloc call had a reservoir permit, we 
let it succeed even without a tag, and just make sure the free method will not 
complain if it can't find a tag.  If the call didn't have a reservoir permit, 
we free the block allocated for the object and fail the allocation, so that the 
client gets a chance
to do whatever low-memory actions they might want to do.  [Should this depend 
on whether there is anything in the reservoir?]  This breaks the 1-to-1 
relationship between tags and objects, so some checks cannot be made, but we do 
count the "lost" tags.

[need to hash out how to do fenceposting in formatted pools]


Client interface

.interface.fenceposting:

.interface.fenceposting.check: A function to check all fenceposts in a pool 
(AVERs if a problem is found)
  void mps_pool_check_fenceposts(mps_pool_t pool);

[from here on, these are tentative and incomplete]

.interface.fenceposting.format: A function to wrap a format (class) to provide 
fenceposting

  mps_res_t mps_fmt_fencepost_wrap(mps_fmt_t *format_return,
                                   mps_arena_t arena,
                                   mps_fmt_t format,
                                   [fp parameters])

.interface.fenceposting.add: A format method to adjust size of a block about to 
be allocted to allow for fenceposts

  void (*mps_fmt_adjust_fencepost_t)(size_t *size_io)

.interface.fenceposting.add: A format method to add a fencepost around a block 
about to be allocated [the NULL method adds a tail fencepost]

  void (*mps_fmt_put_fencepost_t)(mps_addr_t * addr_io, size_t size)

.interface.fenceposting.add: A format method to check the fenceposts around an 
object [the NULL method checks tails]

  mps_bool_t (*mps_fmt_check_fenceposts_t)(mps_addr_t)

.interface.fenceposting.pool: A function to wrap a pool class to provide 
fenceposting (note absence of arena parameter)

  mps_class_t mps_debug_class(mps_class_t class)

.interface.tags.alloc: Three functions to replace existing mps_alloc 
(request.???.??? proposes to remove the varargs)

  mps_res_t mps_alloc(mps_addr_t *, mps_pool_t, size_t);
  mps_res_t mps_alloc_dbg(mps_addr_t *, mps_pool_t, size_t, ...);
  mps_res_t mps_alloc_dbg_v(mps_addr_t *, mps_pool_t, size_t,
                            va_list);

.interface.tags.walker: Functions to walk all the allocated objects in a pool 
or an arena (only client pools in this case), format and tag_data can be NULL 
(tag_data really wants to be void *, not mps_addr_t, because it's stored 
together with the internal tag data in an MPS internal pool)

  void (*mps_objects_step_t)(mps_addr_t addr, size_t size, mps_fmt_t format,
                             mps_pool_t pool, void *tag_data, void *p)

  void mps_pool_walk(mps_arena_t arena, mps_pool_t pool,
                     mps_objects_step_t step, void *p)

  void mps_arena_walk(mps_arena_t arena, mps_objects_step_t step, void *p)


Examples

.example.debug-alloc:

#define MPS_ALLOC_DBG(res_io, addr_io, pool, size)
  MPS_BEGIN
    static mps_tag_A_s _ts = { __FILE__, __LINE__ };

    *res_io = mps_alloc(addr_io, pool, size, _ts_)
  MPS_END


IMPLEMENTATION

.new-pool: The client interface to control fenceposting consists of the new 
classes mps_pool_class_mv_debug, mps_pool_class_epdl_debug, and 
mps_pool_class_epdr_debug, and their new init parameter of type 
mps_pool_debug_option_s.  [This is a temporary solution, to get it out without 
writing lots of new interface.  pekka 1998-09-10]

.new-pool.impl: The debug pools are implemented using the "class wrapper" 
EnsureDebugClass, which produces a subclass with modified init, finish, alloc, 
and free methods.  These methods are implemented in the generic debug class 
code (impl.c.dbgpool), and are basically wrappers around the superclass methods 
(invoked through the pool->class->super field).  To find the data stored in the 
class for the debugging features, they use the debugMixin method provided by 
the subclass.  So to make a debug subclass, three things should be provided: a 
structure definition of the instance containing a PoolDebugMixinStruct, a pool 
class function that uses
EnsureDebugClass, and a debugMixin method that locates the PoolDebugMixinStruct 
within an instance.

.tags.splay: The tags are stored in a splay tree of tags allocated from a 
subsidiary MFS pool.  The client needs to specify the (maximum) size of the 
client data in a tag, so that the pool can be created.

[Lots more should be said, eventually.  pekka 1998-09-10]

A. References

B. Document History

2002-06-07 RB Converted from MMInfo database design document.