This patch allows messages to be stalled in their input buffers and wait
until a corresponding address changes state. In order to make this work,
all in_ports must be ranked in order of dependence and those in_ports that
may unblock an address, must wake up the stalled messages. Alot of this
complexity is handled in slicc and the specification files simply
annotate the in_ports.
--HG--
rename : src/mem/slicc/ast/CheckAllocateStatementAST.py => src/mem/slicc/ast/StallAndWaitStatementAST.py
rename : src/mem/slicc/ast/CheckAllocateStatementAST.py => src/mem/slicc/ast/WakeUpDependentsStatementAST.py
Patch allows each individual message buffer to have different recycle latencies
and allows the overall recycle latency to be specified at the cmd line. The
patch also adds profiling info to make sure no one processor's requests are
recycled too much.
The main purpose for clearing stats in the unserialize process is so
that the profiler can correctly set its start time to the unserialized
value of curTick.
This patch allows one to disable migratory sharing for those cache blocks that
are accessed by atomic requests. While the implementations are different
between the token and hammer protocols, the motivation is the same. For
Alpha, LLSC semantics expect that normal loads do not unlock cache blocks that
have been locked by LL accesses. Therefore, locked blocks should not transfer
write permissions when responding to these load requests. Instead, only they
only transfer read permissions so that the subsequent SC access can possibly
succeed.
This patch fixes several bugs related to previous inconsistent assumptions on
how many tokens the Owner had. Mike Marty should have fixes these bugs years
ago. :)
Previously, the MOESI_hammer protocol calculated the same latency for L1 and
L2 hits. This was because the protocol was written using the old ruby
assumption that L1 hits used the sequencer fast path. Since ruby no longer
uses the fast-path, the protocol delays L2 hits by placing them on the
trigger queue.
The previous slower ruby latencies created a mismatch between the faster M5
cpu models and the much slower ruby memory system. Specifically smp
interrupts were much slower and infrequent, as well as cpus moving in and out
of spin locks. The result was many cpus were idle for large periods of time.
These changes fix the latency mismatch.
This patch adds back to ruby the capability to understand the response time
for messages that hit in different levels of the cache heirarchy.
Specifically add support for the MI_example, MOESI_hammer, and MOESI_CMP_token
protocols.
This patch adds DMA testing to the Memtester and is inherits many changes from
Polina's old tester_dma_extension patch. Since Ruby does not work in atomic
mode, the atomic mode options are removed.
Clean up some minor things left over from the default responder
change in rev 9af6fb59752f. Mostly renaming the 'responder_set'
param to 'use_default_range' to actually reflect what it does...
old name wasn't that descriptive in the first place, but now
it really doesn't make sense at all.
Also got rid of the bogus obsolete assignment to 'bus.responder'
which used to be a parameter but now is interpreted as an
implicit child assignment, and which was giving me problems in
the config restructuring to come. (A good argument for not
allowing implicit child assignments, IMO, but that's water under
the bridge, I'm afraid.)
Also moved the Bus constructor to the .cc file since that's
where it should have been all along.
Requires new "SCUpgradeReq" message that marks upgrades
for store conditionals, so downstream caches can fail
these when they run into invalidations.
See http://www.m5sim.org/flyspray/task/197
Only set the dirty bit when we actually write to a block
(not if we thought we might but didn't, as in a failed
SC or CAS). This requires makeing sure the dirty bit
stays set when we get an exclusive (writable) copy
in a cache-to-cache transfer from another owner, which
n turn requires copying the mem-inhibit flag from
timing-mode requests to their associated responses.
One big difference is that PrioHeap puts the smallest element at the
top of the heap, whereas stl puts the largest element on top, so I
changed all comparisons so they did the right thing.
Some usage of PrioHeap was simply changed to a std::vector, using sort
at the right time, other usage had me just use the various heap functions
in the stl.
This was somewhat tricky because the RefCnt API was somewhat odd. The
biggest confusion was that the the RefCnt object's constructor that
took a TYPE& cloned the object. I created an explicit virtual clone()
function for things that took advantage of this version of the
constructor. I was conservative and used clone() when I was in doubt
of whether or not it was necessary. I still think that there are
probably too many instances of clone(), but hopefully not too many.
I converted several instances of const MsgPtr & to a simple MsgPtr.
If the function wants to avoid the overhead of creating another
reference, then it should just use a regular pointer instead of a ref
counting ptr.
There were a couple of instances where refcounted objects were created
on the stack. This seems pretty dangerous since if you ever
accidentally make a reference to that object with a ref counting
pointer, bad things are bound to happen.
Further cleanup should probably be done to make this class be non-Ruby
specific and put it in src/base.
There are probably several cases where this class is used, std::bitset
could be used instead.
In addition to obvious changes, this required a slight change to the slicc
grammar to allow types with :: in them. Otherwise slicc barfs on std::string
which we need for the headers that slicc generates.
Previously, the set size was set to 4. This was mostly do to the fact that a
crazy graduate student use to create networks with 256 l2 cache banks. Now it
is far more likely that users will create systems with less than 64 of any
particular controller type. Therefore Ruby should be optimized for a set size
of 1.
On the config end, if a shared L2 is created for the system, it is
parameterized to have n sharers as defined by option.num_cpus. In addition to
making the cache sharing aware so that discriminating tag policies can make use
of context_ids to make decisions, I added an occupancy AverageStat and an occ %
stat to each cache so that you could know which contexts are occupying how much
cache on average, both in terms of blocks and percentage. Note that since
devices have context_id -1, having an array of occ stats that correspond to
each context_id will break here, so in FS mode I add an extra bucket for device
blocks. This bucket is explicitly not added in SE mode in order to not only
avoid ugliness in the stats.txt file, but to avoid broken stats (some formulas
break when a bucket is 0).
This patch includes the necessary regression updates to test the new ruby
configuration system. The patch includes support for multiple ruby protocols
and adds the ruby random tester. The patch removes atomic mode test for
ruby since ruby does not support atomic mode acceses. These tests can be
added back in when ruby supports atomic mode for real.
--HG--
rename : tests/quick/50.memtest/test.py => tests/quick/60.rubytest/test.py
Removed the dummy power function implementations so that Orion can implement
them correctly. Since Orion lacks modular design, this patch simply enables
scons to compile it. There are no python configuration changes in this patch.
Renamed the MESI directory file to be consistent with all other protocols.
--HG--
rename : src/mem/protocol/MESI_CMP_directory-mem.sm => src/mem/protocol/MESI_CMP_directory-dir.sm
Cleaned up the ruby profilers by moving the memory controller profiling code
out of the main profiler object and into a separate object similar to the
current CacheProfiler. Both the CacheProfiler and MemCntrlProfiler are
specific to a particular Ruby object, CacheMemory and MemoryControl
respectively. Therefore, these profilers should not be SimObjects and
created by the python configuration system, but instead private objects. This
simplifies the creation of these profilers.
Reorganized ruby python configuration so that protocol and ruby memory system
configuration code can be shared by multiple front-end configuration files
(i.e. memory tester, full system, and hopefully the regression tester). This
code works for memory tester, but have not tested fs mode.
Modified ruby's tracing support to no longer rely on the RubySystem map
to convert a sequencer string name to a sequencer pointer. As a
temporary solution, the code uses the sim_object find function.
Eventually, we should develop a better fix.
This patch includes a rather substantial change to the memory controller
profiler in order to work with the new configuration system. Most
noteably, the mem_cntrl_profiler no longer uses a string map, but instead
a vector. Eventually this support should be removed from the main
profiler and go into a separate object. Each memory controller should have
a pointer to that new mem_cntrl profile object.
This patch includes the necessary changes to connect ruby objects using
the python configuration system. Mainly it consists of removing
unnecessary ruby object pointers and connecting the necessary object
pointers using the generated param objects. This patch includes the
slicc changes necessary to connect generated ruby objects together using
the python configuraiton system.
The necessary companion conversion of Ruby objects generated by SLICC
are converted to M5 SimObjects in the following patch, so this patch
alone does not compile.
Conversion of Garnet network models is also handled in a separate
patch; that code is temporarily disabled from compiling to allow
testing of interim code.
Though OutPort's message type is not used to generate code, this fix checks
that the programmer's intent is correct. Eventually, we may want to
remove the message type from the OutPort declaration statement.
1) Move alpha-specific code out of page_table.cc:serialize().
2) Begin serializing M5_pid and unserializing it, but adding an function to do optional paramIn so that old checkpoints don't need to be fixed up.
3) Fix up alpha startup code so that the unserialized M5_pid value is properly written to DTB_IPR_ASN.
4) Fix the memory unserialize that I forgot somehow in the last changeset.
5) Add in an agg_se.py to handle aggregated checkpoints. --bench foo-bar plus positional arguments foo bar are the only changes in usage from se.py.
Note this aggregation stuff has only been tested for Alpha and nothing else, though it should take a very minimal amount of work to get it to work with another ISA.
This patch changes the way that Ruby handles atomic RMW instructions. This implementation, unlike the prior one, is protocol independent. It works by locking an address from the sequencer immediately after the read portion of an RMW completes. When that address is locked, the coherence controller will only satisfy requests coming from one port (e.g., the mandatory queue) and will ignore all others. After the write portion completed, the line is unlocked. This should also work with multi-line atomics, as long as the blocks are always acquired in the same order.
Added error messages when:
- a state does not exist in a machine's list of known states.
- an event does not exist in a machine
- the actions of a certain machine have not been declared
Connects M5 cpu and dma ports directly to ruby sequencers and dma
sequencers. Rubymem also includes a pio port so that pio requests
and be forwarded to a special pio bus connecting to device pio
ports.
Right now .cc and .hh files are handled separately, but then
they're just munged together at the end by scons, so it
doesn't buy us anything. Might as well munge from the start
since we'll eventually be adding generated Python files
to the list too.
This mostly was a matter of changing the license owner to Princeton
which is as it should have been. The code was originally licensed
under the GPL but was relicensed as BSD by Li-Shiuan Peh on July 27,
2009. This relicensing was in an explicit e-mail to Nathan Binkert,
Brad Beckmann, Mark Hill, David Wood, and Steve Reinhardt.
This prevents redundant prefetches from being issued, solving the
occasional 'needsExclusive && !blk->isWritable()' assertion failure
in cache_impl.hh that several people have run into.
Eliminates "prefetch_cache_check_push" flag, neither setting of
which really solved the problem.
This is simply a translation of the C++ slicc into python with very minimal
reorganization of the code. The output can be verified as nearly identical
by doing a "diff -wBur".
Slicc can easily be run manually by using util/slicc
Get rid of misc.py and just stick misc things in __init__.py
Move utility functions out of SCons files and into m5.util
Move utility type stuff from m5/__init__.py to m5/util/__init__.py
Remove buildEnv from m5 and allow access only from m5.defines
Rename AddToPath to addToPath while we're moving it to m5.util
Rename read_command to readCommand while we're moving it
Rename compare_versions to compareVersions while we're moving it.
--HG--
rename : src/python/m5/convert.py => src/python/m5/util/convert.py
rename : src/python/m5/smartdict.py => src/python/m5/util/smartdict.py
This changeset contains a lot of different changes that are too
mingled to separate. They are:
1. Added MOESI_CMP_directory
I made the changes necessary to bring back MOESI_CMP_directory,
including adding a DMA controller. I got rid of MOESI_CMP_directory_m
and made MOESI_CMP_directory use a memory controller. Added a new
configuration for two level protocols in general, and
MOESI_CMP_directory in particular.
2. DMA Sequencer uses a generic SequencerMsg
I will eventually make the cache Sequencer use this type as well. It
doesn't contain an offset field, just a physical address and a length.
MI_example has been updated to deal with this.
3. Parameterized Controllers
SLICC controllers can now take custom parameters to use for mapping,
latencies, etc. Currently, only int parameters are supported.
The inconsistency was causing a subtle bug with some of the
constructors where the params had the same name as the fields.
This is also a first step to switching the accessors over to
our new "standard", e.g., getVaddr() -> vaddr().
Caches are now responsible for their own statistic gathering. This
requires a direct callback from the protocol on misses, and so all
future protocols need to take this into account.
The DMASequencer was still using a parameter from the old RubyConfig,
causing an offset error when the requested data wasn't block aligned.
This changeset also includes a fix to MI_example for a similar bug.
2. Reintroduced RMW_Read and RMW_Write
3. Defined -2 in the Sequencer as well as made a note about mandatory queue
Did not address the issues in the slicc because remaking the atomics altogether to allow
multiple processors to issue atomic requests at once
This also includes a change to the default Ruby random seed, which was
previously set using the wall clock. It is now set to 1234 so that
the stat files don't change for the regression tester.
This was done with an automated process, so there could be things that were
done in this tree in the past that didn't make it. One known regression
is that atomic memory operations do not seem to work properly anymore.
This changeset also includes a lot of work from Derek Hower <drh5@cs.wisc.edu>
RubyMemory is now both a driver for Ruby and a port for M5. Changed
makeRequest/hitCallback interface. Brought packets (superficially)
into the sequencer. Modified tester infrastructure to be packet based.
and Ruby can be used together through the example ruby_se.py
script. SPARC parallel applications work, and the timing *seems* right
from combined M5/Ruby debug traces. To run,
% build/ALPHA_SE/m5.debug configs/example/ruby_se.py -c
tests/test-progs/hello/bin/alpha/linux/hello -n 4 -t
1. removed checks from tester files
2. removed else clause in Sequencer and DirectoryMemory else clause is
needed by the tester, it is up to Derek to revive it elsewhere when he
gets to it
Also:
1. Changed m_entries in DirectoryMemory to a map
2. And replaced SIMICS_read_physical_memory with a call to now-dummy
Derek's-to-be readPhysMem function
Add the PROTOCOL sticky option sets the coherence protocol that slicc
will parse and therefore ruby will use. This whole process was made
difficult by the fact that the set of files that are output by slicc
are not easily known ahead of time. The easiest thing wound up being
to write a parser for slicc that would tell me. Incidentally this
means we now have a slicc grammar written in python.
This basically means changing all #include statements and changing
autogenerated code so that it generates the correct paths. Because
slicc generates #includes, I had to hard code the include paths to
mem/protocol.
1) Removing files from the ruby build left some unresovled
symbols. Those have been fixed.
2) Most of the dependencies on Simics data types and the simics
interface files have been removed.
3) Almost all mention of opal is gone.
4) Huge chunks of LogTM are now gone.
5) Handling 1-4 left ~hundreds of unresolved references, which were
fixed, yielding a snowball effect (and the massive size of this
delta).
I did the macro cleanup because I was worried that the SCons scanner
would get confused. This code will hopefully go away soon anyway.
--HG--
rename : src/mem/ruby/config/config.include => src/mem/ruby/config/config.hh
Previously there was one per bus, which caused some coherence problems
when more than one decided to respond. Now there is just one on
the main memory bus. The default bus responder on all other buses
is now the downstream cache's cpu_side port. Caches no longer need
to do address range filtering; instead, we just have a simple flag
to prevent snoops from propagating to the I/O bus.
This frees up needed space for more public flags. Also:
- remove unused Request accessor methods
- make Packet use public Request accessors, so it need not be a friend
Apparently we broke it with the cache rewrite and never noticed.
Thanks to Bao Yungang <baoyungang@gmail.com> for a significant part
of these changes (and for inspiring me to work on the rest).
Some other overdue cleanup on the prefetch code too.
Bogus calls to ChunkGenerator with negative size were triggering
a new assertion that was added there.
Also did a little renaming and cleanup in the process.
I think readData() and writeData() were used for Erik's compression
work, but that code is gone, these aren't called anymore, and they
don't even really do what their names imply.
I did some of the flags and assertions wrong. Thanks to Brad Beckmann
for pointing this out. I should have run the opt regressions instead
of the fast. I also screwed up some of the logical functions in the Flags
class.
the primary identifier for a hardware context should be contextId(). The
concept of threads within a CPU remains, in the form of threadId() because
sometimes you need to know which context within a cpu to manipulate.
Since the early days of M5, an event needed to know which event queue
it was on, and that data was required at the time of construction of
the event object. In the future parallelized M5, this sort of
requirement does not work well since the proper event queue will not
always be known at the time of construction of an event. Now, events
are created, and the EventQueue itself has the schedule function,
e.g. eventq->schedule(event, when). To simplify the syntax, I created
a class called EventManager which holds a pointer to an EventQueue and
provides the schedule interface that is a proxy for the EventQueue.
The intent is that objects that frequently schedule events can be
derived from EventManager and then they have the schedule interface.
SimObject and Port are examples of objects that will become
EventManagers. The end result is that any SimObject can just call
schedule(event, when) and it will just call that SimObject's
eventq->schedule function. Of course, some objects may have more than
one EventQueue, so this interface might not be perfect for those, but
they should be relatively few.
This appears to work, but I don't want to commit it until it gets tested a lot more.
I haven't deleted the functionality in this patch that will come later, but one question
is how to enforce encourage objects that call getVirtPort() to not cache the virtual port
since if the CPU changes out from under them it will be worse than useless. Perhaps a null
function like delVirtPort() is still useful in that case.
It runs out that if a MemObject turns around and does a send in its
receive callback, and there are other sends already scheduled, then
it could observe a state where it's not at the head of the list but
the bus's sendEvent is not scheduled (because we're still in the
middle of processing the prior sendEvent).
I was asserting that the only reason you would defer targets is if
a write came in while you had an outstanding read miss, but there's
another case where you could get a read access after you've snooped
an invalidation and buffered it because it applies to a prior
outstanding miss.
Make OutputDirectory::resolve() private and change the functions using
resolve() to instead use create().
--HG--
extra : convert_revision : 36d4be629764d0c4c708cec8aa712cd15f966453
if a prior write miss arrived while an even earlier
read miss was still outstanding.
--HG--
extra : convert_revision : 4924e145829b2ecf4610b88d33f4773510c6801a
where we defer a response to a read from a far-away cache A, then later
defer a ReadExcl from a cache B on the same bus as us. We'll assert
MemInhibit in both cases, but in the latter case MemInhibit will keep
the invalidation from reaching cache A. This special response tells
cache A that it gets the block to satisfy its read, but must immediately
invalidate it.
--HG--
extra : convert_revision : f85c8b47bb30232da37ac861b50a6539dc81161b
Don't mark upstream MSHR as pending if downstream MSHR is already in service.
--HG--
extra : convert_revision : e1c135ff00217291db58ce8a06ccde34c403d37f
Not so much noise on failed sends, and more complete
info when grepping a trace using an address.
--HG--
extra : convert_revision : 05a8261c9452072ca08b906200c6322b33e2b9f1
SimObjects not yet updated:
- Process and subclasses
- BaseCPU and subclasses
The SimObject(const std::string &name) constructor was removed. Subclasses
that still rely on that behavior must call the parent initializer as
: SimObject(makeParams(name))
--HG--
extra : convert_revision : d6faddde76e7c3361ebdbd0a7b372a40941c12ed
The page table now stores actual page table entries. It is still a templated
class here, but this will be corrected in the near future.
--HG--
extra : convert_revision : 804dcc6320414c2b3ab76a74a15295bd24e1d13d
Make sure not to keep processing functional accesses
after they've been responded to.
Also use checkFunctional() return value instead of checking
packet command field where possible, mostly just for consistency.
--HG--
extra : convert_revision : 29fc76bc18731bd93a4ed05a281297827028ef75
creation and initialization now happens in python. Parameter objects
are generated and initialized by python. The .ini file is now solely for
debugging purposes and is not used in construction of the objects in any
way.
--HG--
extra : convert_revision : 7e722873e417cb3d696f2e34c35ff488b7bff4ed
Turns out DeferredSnoop isn't quite the right bit of info
we needed... see new comment in cache_impl.hh.
--HG--
extra : convert_revision : a38de8c1677a37acafb743b7074ef88b21d3b7be
If the invalidation beats the upgrade at a lower level
then the upgrade must be converted to a read exclusive
"in the field".
Restructure target list & deferred target list to
factor out some common code.
--HG--
extra : convert_revision : 7bab4482dd6c48efdb619610f0d3778c60ff777a
- Add "deferred snoop" flag to Packet so upper-level caches
can distinguish whether lower-level cache request was
in-service or not at the time of the original snoop.
- Revamp response handling to properly handle deferred snoops
on non-cache-fill requests (i.e. upgrades).
- Make sure forwarded writebacks are kept in write buffer at
lower-level caches so they get snooped properly.
--HG--
extra : convert_revision : 17f8a3772a1ae31a16991a53f8225ddf54d31fc9
Move check for loops outside, since half the call sites
end up working around it anyway. Return integer port ID
instead of port object pointer.
--HG--
extra : convert_revision : 4c31fe9930f4d1aa4919e764efb7c50d43792ea3
Note that we should *not* print pointer values in DPRINTFs as
these needlessly clutter tracediff output.
--HG--
extra : convert_revision : 25a448f1b3ac8d453a717a104ad6dc0112fb30bb
src/cpu/simple/timing.cc:
Fix another SC problem.
src/mem/cache/cache_impl.hh:
Forgot to call makeTimingResponse() on uncached timing responses.
--HG--
extra : convert_revision : 5a5a58ca2053e4e8de2133205bfd37de15eb4209
Stats pretty much line up with old code, except:
- bug in old code included L1 latency in L2 miss time, making it too high
- UniCoherence did cache-to-cache transfers even from non-owner caches,
so occasionally the icache would get a block from the dcache not the L2
- L2 can now receive ReadExReq from L1 since L1s have coherence
--HG--
extra : convert_revision : 5052c1a1767b5a662f30a88f16012165a73b791c
Change target overflow from assertion to warning.
src/mem/cache/cache_impl.hh:
Change target overflow from assertion to warning.
--HG--
extra : convert_revision : ceca990ed916bbf96dedd4836c40df522803f173
src/mem/cache/cache_impl.hh:
Handle grants with no packet.
src/mem/cache/miss/mshr.cc:
Fix MSHR snoop hit handling.
--HG--
extra : convert_revision : f365283afddaa07cb9e050b2981ad6a898c14451
sure we don't re-request bus prematurely. Use callback to
avoid calling sendRetry() recursively within recvTiming.
--HG--
extra : convert_revision : a907a2781b4b00aa8eb1ea7147afc81d6b424140
src/cpu/memtest/memtest.cc:
Need to set packet source field so that response from cache
doesn't run into assertion failure when copying source to dest.
src/mem/packet.hh:
Copy source field when copying packets.
Assert that source is valid before copying it to dest
when turning packets around.
--HG--
extra : convert_revision : 09e3cfda424aa89fe170e21e955b295746832bf8
supposed to and make sure parameters have the right type.
Also make sure that any object that should be an intermediate
type has the right options set.
--HG--
extra : convert_revision : d56910628d9a067699827adbc0a26ab629d11e93
