Brian Demsky [Fri, 14 Sep 2012 08:31:09 +0000 (01:31 -0700)]
(1) add actions for the fence
(2) a little more support for cyclegraph -- show rmw edges
(3) add extra documentation for norris
Brian Demsky [Fri, 14 Sep 2012 05:04:45 +0000 (22:04 -0700)]
add support for dumping cyclegraphs as dot files... also eliminate redundant edges to make them easier to view
Brian Norris [Thu, 13 Sep 2012 22:42:21 +0000 (15:42 -0700)]
bugfix: straighten out STL vector allocation (snapshotted vs. persistent)
When using a STL data structure allocated on the stack, we must make sure its
elements are allocated with the same allocator as the structure. For instance,
in a model-checking context, the stack is persistent (non-snapshotting), so any
stack-allocated std::vector should use the non-snapshotting allocator (i.e.,
MyAlloc).
At the same time, clarify CycleGraph and CycleNode classes by labelling them as
SNAPSHOTALLOC.
Brian Norris [Thu, 13 Sep 2012 21:58:36 +0000 (14:58 -0700)]
model: bugfix - don't clobber 'updated'
The 'updated' variable should stay 'true' even if modification orders and
promises aren't updated by the "write" checks. This could be a RMW which made
updates via the "read" checks.
Brian Norris [Thu, 13 Sep 2012 21:56:16 +0000 (14:56 -0700)]
Revert "Makefile: don't always rebuild make.deps"
The Makefile wasn't using its dependency generation correctly. Revert this for
now.
This reverts commit
870ba814eceb1afee4eefb17dab3980549ca73e2.
Brian Norris [Thu, 13 Sep 2012 21:47:44 +0000 (14:47 -0700)]
snapshot: print stack trace on segfault
Brian Norris [Thu, 13 Sep 2012 21:47:23 +0000 (14:47 -0700)]
common: add print_trace() for backtracing
Brian Norris [Thu, 13 Sep 2012 18:45:42 +0000 (11:45 -0700)]
model: trivial fixups
Brian Norris [Thu, 13 Sep 2012 18:35:48 +0000 (11:35 -0700)]
Makefile: don't always rebuild make.deps
The dependencies don't need to be regenerated for "make tags", "make docs",
etc.
Brian Norris [Thu, 13 Sep 2012 17:16:37 +0000 (10:16 -0700)]
model: simple ASSERT() bug
Don't ASSERT(rf->is_write()) until we're sure that rf is non-NULL.
Brian Norris [Thu, 13 Sep 2012 17:04:28 +0000 (10:04 -0700)]
main: wrap help message
Brian Norris [Thu, 13 Sep 2012 16:53:59 +0000 (09:53 -0700)]
model: kill unneeded local variable
The "Thread *th" variable is used only on a few control paths, and it is never
reused. Might as well just calculate it on the fly. Also, it's recalculated and
masked by another local variable at a deeper level of nesting.
Brian Norris [Thu, 13 Sep 2012 16:42:54 +0000 (09:42 -0700)]
spacing
Brian Norris [Thu, 13 Sep 2012 16:21:08 +0000 (09:21 -0700)]
Merge remote-tracking branch 'origin/makefile'
Brian Demsky [Thu, 13 Sep 2012 08:33:24 +0000 (01:33 -0700)]
lots of debugging here... finally working with my rmw example...
Brian Norris [Thu, 13 Sep 2012 07:22:55 +0000 (00:22 -0700)]
Makefile: fixup Mac dependencies
Now, Mac OSX builds can just use "make", not "make mac". (This also allows
"make debug" for Mac OSX)
Brian Norris [Thu, 13 Sep 2012 07:01:01 +0000 (00:01 -0700)]
cleanup style
Brian Demsky [Thu, 13 Sep 2012 05:59:03 +0000 (22:59 -0700)]
commit new test case
Brian Demsky [Thu, 13 Sep 2012 05:57:00 +0000 (22:57 -0700)]
fix for horrible bug... turns out that we could generate an infinite set of bad executions due to future values...
fix--don't send future values backwards until all of your promises are resolved...
Brian Demsky [Thu, 13 Sep 2012 05:08:22 +0000 (22:08 -0700)]
missing changes
Brian Demsky [Thu, 13 Sep 2012 05:06:51 +0000 (22:06 -0700)]
right fix for avoid rmw cycles... bad assumption in the cyclegraph
Brian Demsky [Thu, 13 Sep 2012 04:03:37 +0000 (21:03 -0700)]
another bug fix...
Brian Demsky [Thu, 13 Sep 2012 03:35:47 +0000 (20:35 -0700)]
separate out rmw actions
make sure we don't insert promises twice for a node
Brian Demsky [Wed, 12 Sep 2012 23:24:07 +0000 (16:24 -0700)]
(1) structure code a little better
(2) add startChanges to cyclegraph... This is just to check some assertions that rollback is being used correctly and doesn't rollback unintended changes
Brian Demsky [Wed, 12 Sep 2012 22:12:31 +0000 (15:12 -0700)]
deal with looping due to bogus future value via promise expiration
fix promise bug
Brian Demsky [Wed, 12 Sep 2012 09:52:34 +0000 (02:52 -0700)]
fun issues...
(1) need to compute promises for rmw nodes...they could satisfy them
(2) don't call rmwedge if you don't know where you are reading from yet
(3) miscopied commit from big merge (already_added -> !already_added)
(4) avoid adding rmw edges back to itself
(5) probably worth adding a check that the execution isfeasible before sending future values
(6) resolve promises should add rmwedges when appropriate
(7) fix seq_cst to mask earlier writes..
Brian Norris [Wed, 12 Sep 2012 02:48:38 +0000 (19:48 -0700)]
threads, model, schedule: refactor thread joining
The Thread can hold a list of ModelAction (instead of Threads) that are waiting
for its completion. This will give the exiting Thread a better ability to
handle the event.
This also deletes Scheduler::wait, since it is no superceded by Scheduler::sleep.
Brian Norris [Wed, 12 Sep 2012 02:45:57 +0000 (19:45 -0700)]
model: add ModelChecker::get_thread(ModelAction *act)
Many times, I need to convert ModelAction to a Thread. Overload get_thread() to
make it easier to use.
Brian Norris [Wed, 12 Sep 2012 02:38:50 +0000 (19:38 -0700)]
schedule: add sleep() function
The wait() and wake() functions are not very complementary. I will be removing
wait() in favor of a simpler sleep() function, which is introduced here.
Brian Norris [Wed, 12 Sep 2012 00:54:29 +0000 (17:54 -0700)]
cyclegraph: add non-NULL assertions
Brian Norris [Wed, 12 Sep 2012 00:52:02 +0000 (17:52 -0700)]
Merge branch 'demsky'
Branch cleaned up by Brian Norris
Brian Demsky [Tue, 11 Sep 2012 03:17:10 +0000 (20:17 -0700)]
model: cleaning up some code
Brian Norris [Tue, 11 Sep 2012 18:43:18 +0000 (11:43 -0700)]
model: handle RMW, unresolved reads in w_modification_order checks
Brian Demsky [Mon, 10 Sep 2012 00:43:20 +0000 (17:43 -0700)]
model: fix the maxreads support
Brian Demsky [Sun, 9 Sep 2012 09:25:47 +0000 (02:25 -0700)]
model: some bug fixes to the model checker
Brian Demsky [Sun, 9 Sep 2012 09:25:47 +0000 (02:25 -0700)]
linuxrwlocks: two bug fixes; guess the model checker helps find bugs
Brian Demsky [Sun, 9 Sep 2012 02:37:11 +0000 (19:37 -0700)]
model: bug fixes to new code
Brian Demsky [Sun, 9 Sep 2012 02:37:11 +0000 (19:37 -0700)]
cyclegraph: support rolling back changes
Brian Norris [Tue, 11 Sep 2012 19:24:15 +0000 (12:24 -0700)]
model: replace action conditions with switch block
Brian Norris [Tue, 11 Sep 2012 17:20:06 +0000 (10:20 -0700)]
model: add check_recency()
Note the weaknesses of the code, as documented in @todo.
Brian Norris [Tue, 11 Sep 2012 18:19:50 +0000 (11:19 -0700)]
model: refactor "get next thread" code
get_next_replay_thread() might as well handle all the Thread decisions for the
model-checker. Refactor, and rename to get_next_thread().
Brian Norris [Tue, 11 Sep 2012 17:19:46 +0000 (10:19 -0700)]
model: add too_many_reads flag
Brian Norris [Tue, 11 Sep 2012 17:19:10 +0000 (10:19 -0700)]
nodestack: add get_read_from_size()
Brian Norris [Tue, 11 Sep 2012 17:10:39 +0000 (10:10 -0700)]
main: add "maxreads" parameter
Brian Norris [Sat, 8 Sep 2012 07:38:03 +0000 (00:38 -0700)]
main: add parameter parsing
Only has a help message for now.
Brian Norris [Sat, 8 Sep 2012 06:36:00 +0000 (23:36 -0700)]
model: fix release sequence with RMW, acq vs. acq/rel
Check in a bugfix from Brian D
Brian Norris [Fri, 7 Sep 2012 20:12:46 +0000 (13:12 -0700)]
model: maintain a count of the pending lazy synchronizations
lazy_sync_size should be the sum of the size of all the lists in
lazy_sync_with_release.
Brian Norris [Fri, 7 Sep 2012 19:58:06 +0000 (12:58 -0700)]
model: group snapshottable ModelChecker members in struct
These items shouldn't be reset individually, but rather should be grouped
together and allocated on the snapshotting heap. This puts us a step closer to
utilizing partial rollback within executions.
Brian Norris [Fri, 7 Sep 2012 18:25:25 +0000 (11:25 -0700)]
model: force THREAD_START to immediately follow THREAD_CREATE
Note that on the current userprog.c test case, this change causes more
infeasible executions to appear. This is somewhat unexpected but probably just
a result of sub-optimal scheduling.
Brian Norris [Fri, 7 Sep 2012 17:38:58 +0000 (10:38 -0700)]
model: fix release_seq for open-ended synchronization
This fix handles the case where a thread may not yet have synchronized in a way
that ensures no future writes may break a release sequence. If we encounter
such a thread, we can return false (uncertain) and re-check this potential
release sequence later in the execution.
Brian Norris [Fri, 7 Sep 2012 17:31:08 +0000 (10:31 -0700)]
model: use std::find for release sequence search
Brian Norris [Fri, 7 Sep 2012 17:16:41 +0000 (10:16 -0700)]
model: don't use global 'model' unnecessarily
Brian Demsky [Fri, 7 Sep 2012 06:48:25 +0000 (23:48 -0700)]
changes
Brian Demsky [Fri, 7 Sep 2012 01:56:39 +0000 (18:56 -0700)]
Merge branch 'master' of ssh://demsky.eecs.uci.edu/home/git/model-checker
check in my stuff...
Conflicts:
model.cc
threads.h
Brian Demsky [Fri, 7 Sep 2012 01:32:56 +0000 (18:32 -0700)]
changes
Brian Norris [Thu, 6 Sep 2012 23:16:51 +0000 (16:16 -0700)]
Merge branch 'norris'
Fixes up some THREAD_JOIN behavior, cleaning up a little bit of general model-checker structure.
Brian Norris [Thu, 6 Sep 2012 21:07:44 +0000 (14:07 -0700)]
model: complete the Thread teardown during THREAD_FINISH
Brian Norris [Thu, 6 Sep 2012 21:04:09 +0000 (14:04 -0700)]
model: enforce rule: current_action != NULL
Now that I've fixed THREAD_JOIN, modify the comments and safety checks.
Brian Norris [Thu, 6 Sep 2012 20:55:05 +0000 (13:55 -0700)]
libthreads: don't spin on THREAD_JOIN
THREAD_JOIN can now be implemented properly here. We shouldn't have to spin,
waiting for another thread to complete. This may help improve the ModelChecker
architecture, since we will never have NULL actions passed to the ModelChecker.
Brian Norris [Thu, 6 Sep 2012 20:54:06 +0000 (13:54 -0700)]
model: hook up THREAD_JOIN and THREAD_FINISH actions
Brian Norris [Thu, 6 Sep 2012 20:37:35 +0000 (13:37 -0700)]
schedule: add wait and wake functions
These functions will be used to prevent execution of threads that should be
waiting for another event (e.g., a THREAD_JOIN should cause a thread to wait
for another thread's exit).
Brian Norris [Thu, 6 Sep 2012 20:47:12 +0000 (13:47 -0700)]
threads: fixup THREAD_READY comment
This state will be used for more than just a context switch, so reword the
comment.
Brian Norris [Thu, 6 Sep 2012 20:41:12 +0000 (13:41 -0700)]
threads: add THREAD_BLOCKED state
To signal that a Thread cannot yet be added back to the Scheduler's ready list.
Brian Norris [Thu, 6 Sep 2012 20:17:06 +0000 (13:17 -0700)]
threads: add a wait_list
This list will contain all Threads that are waiting on the current Thread. Will
be used for Thread joins, in particular.
Brian Norris [Thu, 6 Sep 2012 20:10:11 +0000 (13:10 -0700)]
threads: add is_complete() helper function
Brian Norris [Thu, 6 Sep 2012 20:07:26 +0000 (13:07 -0700)]
action: add THREAD_FINISH action
For cleanly finalizing a thread.
Brian Norris [Thu, 6 Sep 2012 18:43:06 +0000 (11:43 -0700)]
snapshot-interface: debug prints: lengthen the "regionname" field
Brian Demsky [Thu, 6 Sep 2012 06:04:30 +0000 (23:04 -0700)]
working towards making datarace detection work
Brian Demsky [Wed, 5 Sep 2012 07:27:23 +0000 (00:27 -0700)]
bug in race detector
Brian Demsky [Wed, 5 Sep 2012 06:28:00 +0000 (23:28 -0700)]
another example
Brian Norris [Tue, 4 Sep 2012 21:13:28 +0000 (14:13 -0700)]
model: document ModelChecker::check_current_action
Brian Norris [Tue, 4 Sep 2012 20:08:46 +0000 (13:08 -0700)]
model: check_current_action returns its 'nextThread'
To begin some code structure rearrangements, I make check_current_action()
return the 'nextThread' value as a true return value. Eventually, the
nextThread field might not be needed, and the model-checker behavior might make
more sense...
Brian Norris [Tue, 4 Sep 2012 19:45:49 +0000 (12:45 -0700)]
model/schedule: revise 'nextThread' data flow
The ModelChecker::nextThread field was being abused a little in my aging
design. It really should be either a private field (not accessed even via
accessors) or else just a return value / function parameter.
This commit makes a change so that nextThread is a Thread pointer and is
directly supplied to the Scheduler. If it is NULL, then the Scheduler is
allowed to pick its own Thread to run.
Brian Norris [Tue, 4 Sep 2012 17:59:47 +0000 (10:59 -0700)]
action: print '?' for unknown read-from-future
To provide more informative debugging traces, use a placeholder '?' as the "Rf"
value when printing a 'read' ModelAction that has no reads-from value. This
helps, for instance, when manually identifying problem points in an infeasible
execution.
Brian Norris [Fri, 24 Aug 2012 00:54:39 +0000 (17:54 -0700)]
model: re-check release sequences lazily
For now, I write this "lazy check" as follows:
Whenever one of the following occurs:
* a Promise is fulfilled
* a mo_graph edge is added
Then I recheck all the actions (for the relevant object location) that are
waiting in the lazy release head queue.
Brian Norris [Sat, 25 Aug 2012 01:37:15 +0000 (18:37 -0700)]
model: add resolve_release_sequences() function
This function can check for release sequence resolutions then propagate
synchronization and remove from the "lazy release" list. It does not add any
new mo_graph edges yet.
Brian Norris [Thu, 23 Aug 2012 02:16:58 +0000 (19:16 -0700)]
model: stash actions for lazy release-seq checking
Build up lists of actions to lazily check for new release sequence
developments.
Brian Norris [Fri, 24 Aug 2012 00:47:31 +0000 (17:47 -0700)]
model: report 'updates' when adding mo_graph edges
The rest of the model checker would like to know if any edges were added. Add a
boolean return value to report this.
Brian Norris [Fri, 24 Aug 2012 00:50:21 +0000 (17:50 -0700)]
model: report status of resolved promises
The rest of the model checker needs to know when promises are added, so add a
return status as a boolean.
Brian Norris [Thu, 23 Aug 2012 19:31:08 +0000 (12:31 -0700)]
action, clockvector: add 'has_synchronized_with()' functions
These functions will check whether a ClockVector (or corresponding ModelAction)
is *completely* synchronized with another already. This is different from
simply "happens before," because I may need to update and propagate a clock
vector after initial execution as more information becomes available, and so
this function helps determine whether a particular pair of vectors is worth
merging (and then - expensively - propagating) before actually performing the
synchronization.
[Not documented properly yet...]
Brian Norris [Thu, 23 Aug 2012 19:30:27 +0000 (12:30 -0700)]
action: make synchronize_with() public
Brian Norris [Wed, 15 Aug 2012 00:37:32 +0000 (17:37 -0700)]
action: utilize release sequence(s) for synchronization
Instead of checking only the trivial release sequence (i.e., a read-acquire
reads directly from a write-release) for establishing synchronization, make use
of the ModelChecker's more complete 'get_release_seq_head()' functionality,
then loop through all release heads and synchronize with each. This is
necessary because a read-acquire may synchronize with more than one
store-release.
Note that this step only implements support based on present knowledge. The
incomplete knowledge of the modification order, as given in mo_graph, as well
as "reading from the future" may require lazy checking.
Brian Norris [Wed, 15 Aug 2012 00:21:29 +0000 (17:21 -0700)]
model: add release sequence support
The ModelChecker now can find the head(s) of the release sequence(s) with which
a particular ModelAction (read-acquire) will synchronize.
The ModelChecker::release_seq_head function can locate a release sequence head
for a given ModelAction, based on information at a given moment. That is, it
knows happens-before and modification information for the present, but some
decisions may need to be made in the future as reads-from promises are
fulfilled or modification ordering is observed by future reads and writes.
Lazy checking for the latter cases has yet to be implemented.
Brian Norris [Fri, 24 Aug 2012 00:52:46 +0000 (17:52 -0700)]
model: change pointer spacing style
Brian Norris [Thu, 23 Aug 2012 02:15:04 +0000 (19:15 -0700)]
action: (assertion) disallow out-of-order synchronization
We build our ModelChecker around the assumption that synchronization never
occurs counter to the execution order. This assertion lets us know if that
fails.
Brian Norris [Thu, 23 Aug 2012 02:08:50 +0000 (19:08 -0700)]
model: improve some promise-related comments
Brian Norris [Mon, 20 Aug 2012 22:59:25 +0000 (15:59 -0700)]
action: update 'reads_from' field before synchronization
As the ModelChecker's release/acquire functionality gets more complex,
it makes more sense to assign the ModelAction::reads_from field before
checking synchronization. Currently, this makes no change in behavior.
Brian Norris [Wed, 22 Aug 2012 23:45:55 +0000 (16:45 -0700)]
nodestack: turn magic promise numbers into enum + typedef
Document the enum properly, since it's quite unclear what these flags really
mean.
Brian Norris [Wed, 22 Aug 2012 23:02:48 +0000 (16:02 -0700)]
nodestack, model: use uniform spacing, style
Brian Norris [Wed, 22 Aug 2012 22:56:13 +0000 (15:56 -0700)]
nodestack: don't use uint32_t
uint32_t (and similar) should only be used when an exact field-size is needed
across platforms with different integer sizes. In this case, we only need an
unsigned type, so just use unsigned int.
Brian Norris [Wed, 22 Aug 2012 22:51:43 +0000 (15:51 -0700)]
nodestack: remove unnecessary typedefs
These typedefs are used once or twice and don't add much meaning.
Brian Norris [Fri, 17 Aug 2012 00:06:31 +0000 (17:06 -0700)]
cyclegraph: add public CycleGraph::checkReachable()
The private CycleGraph::checkReachable() function can be useful externally,
when provided with two ModelActions. This implements a small wrapper for public
usage.
Brian Norris [Tue, 21 Aug 2012 06:58:38 +0000 (23:58 -0700)]
model: correct mo_graph comment
My understanding of the mo_graph was incorrect, due to some confusion over the
CycleGraph interfaces. Correct this.
Brian Norris [Tue, 21 Aug 2012 06:56:10 +0000 (23:56 -0700)]
cyclegraph: straighten out header vs. implementation vs. usage
The CycleGraph::addEdge and CycleGraph::addRMWEdge functions were a little
confusing to use, since their implementation and header prototypes had
different parameter naming. This swapped the 'to' and 'from' naming, such that
it appeared as if the addEdge() users were adding edges in the reverse
direction. The functionality was not actually incorrect, but my understanding
was...
This corrects the naming and switches the order of the arguments.
Brian Norris [Mon, 20 Aug 2012 19:11:42 +0000 (12:11 -0700)]
promise: rewrite into a simpler header file
Brian Norris [Fri, 10 Aug 2012 18:52:36 +0000 (11:52 -0700)]
Makefile: rewrite header dependencies
We shouldn't need to add new files in 3 different places. Automatically
generate most of this using GCC's '-MM' option.
Brian Norris [Mon, 20 Aug 2012 18:22:32 +0000 (11:22 -0700)]
model: document ModelChecker::mo_graph
The modification order graph is a complex data structure, and there are a few
pieces of high-level information that should be noted. For example, the edges
are actually directed from most recent to oldest, which is somewhat in reverse
of the usage in the literature, where
a --mo--> b
means that a comes *before* b in the modification order. This convention can be
changed in the future, but it should be documented here.
Brian Norris [Mon, 20 Aug 2012 18:03:36 +0000 (11:03 -0700)]
model: rename 'cyclegraph' to 'mo_graph'
This is not just an arbitrary graph with cycle-detection; it is specifically a
representation of the modification order of various atomic objects.
Brian Norris [Mon, 20 Aug 2012 19:53:28 +0000 (12:53 -0700)]
model: don't "leak" promises
Brian Norris [Fri, 17 Aug 2012 00:40:40 +0000 (17:40 -0700)]
test/Makefile: remove pointless variable