Draining is currently done by traversing the SimObject graph and
calling drain()/drainResume() on the SimObjects. This is not ideal
when non-SimObjects (e.g., ports) need draining since this means that
SimObjects owning those objects need to be aware of this.
This changeset moves the responsibility for finding objects that need
draining from SimObjects and the Python-side of the simulator to the
DrainManager. The DrainManager now maintains a set of all objects that
need draining. To reduce the overhead in classes owning non-SimObjects
that need draining, objects inheriting from Drainable now
automatically register with the DrainManager. If such an object is
destroyed, it is automatically unregistered. This means that drain()
and drainResume() should never be called directly on a Drainable
object.
While implementing the new functionality, the DrainManager has now
been made thread safe. In practice, this means that it takes a lock
whenever it manipulates the set of Drainable objects since SimObjects
in different threads may create Drainable objects
dynamically. Similarly, the drain counter is now an atomic_uint, which
ensures that it is manipulated correctly when objects signal that they
are done draining.
A nice side effect of these changes is that it makes the drain state
changes stricter, which the simulation scripts can exploit to avoid
redundant drains.
The drain state enum is currently a part of the Drainable
interface. The same state machine will be used by the DrainManager to
identify the global state of the simulator. Make the drain state a
global typed enum to better cater for this usage scenario.
Objects that are can be serialized are supposed to inherit from the
Serializable class. This class is meant to provide a unified API for
such objects. However, so far it has mainly been used by SimObjects
due to some fundamental design limitations. This changeset redesigns
to the serialization interface to make it more generic and hide the
underlying checkpoint storage. Specifically:
* Add a set of APIs to serialize into a subsection of the current
object. Previously, objects that needed this functionality would
use ad-hoc solutions using nameOut() and section name
generation. In the new world, an object that implements the
interface has the methods serializeSection() and
unserializeSection() that serialize into a named /subsection/ of
the current object. Calling serialize() serializes an object into
the current section.
* Move the name() method from Serializable to SimObject as it is no
longer needed for serialization. The fully qualified section name
is generated by the main serialization code on the fly as objects
serialize sub-objects.
* Add a scoped ScopedCheckpointSection helper class. Some objects
need to serialize data structures, that are not deriving from
Serializable, into subsections. Previously, this was done using
nameOut() and manual section name generation. To simplify this,
this changeset introduces a ScopedCheckpointSection() helper
class. When this class is instantiated, it adds a new /subsection/
and subsequent serialization calls during the lifetime of this
helper class happen inside this section (or a subsection in case
of nested sections).
* The serialize() call is now const which prevents accidental state
manipulation during serialization. Objects that rely on modifying
state can use the serializeOld() call instead. The default
implementation simply calls serialize(). Note: The old-style calls
need to be explicitly called using the
serializeOld()/serializeSectionOld() style APIs. These are used by
default when serializing SimObjects.
* Both the input and output checkpoints now use their own named
types. This hides underlying checkpoint implementation from
objects that need checkpointing and makes it easier to change the
underlying checkpoint storage code.
This patch drops the NetworkMessage class. The relevant data members and functions
have been moved to the Message class, which was the parent of NetworkMessage.
The accessor function getDestination() for Destination variable in the
coherence message clashes with the getDestination() that is part of the Message
class. Hence the name change.
This structure's only purpose was to provide a comparison function for
ordering messages in the MessageBuffer. The comparison function is now
being moved to the Message class itself. So we no longer require this
structure.
This patch increases the default read/write buffer sizes for the DDR4
controller config to values that are more suitable for the high
bandwidth and high bank count.
This patch updates the command arbitration so that bank group timing
as well as rank-to-rank delays will be taken into account. The
resulting arbitration no longer selects commands (prepped or not) that
cannot issue seamlessly if there are commands that can issue
back-to-back, minimizing the effect of rank-to-rank (tCS) & same bank
group (tCCD_L) delays.
The arbitration selects a new command based on the following priority.
Within each priority band, the arbitration will use FCFS to select the
appropriate command:
1) Bank is prepped and burst can issue seamlessly, without a bubble
2) Bank is not prepped, but can prep and issue seamlessly, without a
bubble
3) Bank is prepped but burst cannot issue seamlessly. In this case, a
bubble will occur on the bus
Thus, to enable more parallelism in subsequent selections, an
unprepped packet is given higher priority if the bank prep can be
hidden. If the bank prep cannot be hidden, the selection logic will
choose a prepped packet that cannot issue seamlessly if one exist.
Otherwise, the default selection will choose the packet with the
minimum bank prep delay.
This patch adds a simple lookup structure to avoid iterating over the
write queue to find read matches, and for the merging of write
bursts. Instead of relying on iteration we simply store a set of
currently-buffered write-burst addresses and compare against
these. For the reads we still perform the iteration if we have a
match. For the writes, we rely entirely on the set. Note that there
are corner-cases where sub-bursts would actually not be mergeable
without a read-modify-write. We ignore these cases and opt for speed.
This patch changes how the crossbar classes deal with
responses. Instead of forwarding responses directly and burdening the
neighbouring modules in paying for the latency (through the
pkt->headerDelay), we now queue them before sending them.
The coherency protocol is not affected as requests and any snoop
requests/responses are still passed on in zero time. Thus, the
responses end up paying for any header delay accumulated when passing
through the crossbar. Any latency incurred on the request path will be
paid for on the response side, if no other module has dealt with it.
As a result of this patch, responses are returned at a later
point. This affects the number of outstanding transactions, and quite
a few regressions see an impact in blocking due to no MSHRs, increased
cache-miss latencies, etc.
Going forward we should be able to use the same concept also for snoop
responses, and any request that is not an express snoop.
This patch takes the final step in removing the is_top_level parameter
from the cache. With the recent changes to read requests and write
invalidations, the parameter is no longer needed, and consequently
removed.
This also means that asymmetric cache hierarchies are now fully
supported (and we are actually using them already with L1 caches, but
no table-walker caches, connected to a shared L2).
WriteInvalidateReq ensures that a whole-line write does not incur the
cost of first doing a read exclusive, only to later overwrite the
data. This patch splits the existing WriteInvalidateReq into a
WriteLineReq, which is done locally, and an InvalidateReq that is sent
out throughout the memory system. The WriteLineReq re-uses the normal
WriteResp.
The change allows us to better express the difference between the
cache that is performing the write, and the ones that are merely
invalidating. As a consequence, we no longer have to rely on the
isTopLevel flag. Moreover, the actual memory in the system does not
see the intitial write, only the writeback. We were marking the
written line as dirty already, so there is really no need to also push
the write all the way to the memory.
The overall flow of the write-invalidate operation remains the same,
i.e. the operation is only carried out once the response for the
invalidate comes back. This patch adds the InvalidateResp for this
very reason.
This patch adds two new read requests packets:
ReadCleanReq - For a cache to explicitly request clean data. The
response is thus exclusive or shared, but not owned or modified. The
read-only caches (see previous patch) use this request type to ensure
they do not get dirty data.
ReadSharedReq - We add this to distinguish cache read requests from
those issued by other masters, such as devices and CPUs. Thus, devices
use ReadReq, and caches use ReadCleanReq, ReadExReq, or
ReadSharedReq. For the latter, the response can be any state, shared,
exclusive, owned or even modified.
Both ReadCleanReq and ReadSharedReq re-use the normal ReadResp. The
two transactions are aligned with the emerging cache-coherent TLM
standard and the AMBA nomenclature.
With this change, the normal ReadReq should never be used by a cache,
and is reserved for the actual (non-caching) masters in the system. We
thus have a way of identifying if a request came from a cache or
not. The introduction of ReadSharedReq thus removes the need for the
current isTopLevel hack, and also allows us to stop relying on
checking the packet size to determine if the source is a cache or
not. This is fixed in follow-on patches.
This patch adds a parameter to the BaseCache to enable a read-only
cache, for example for the instruction cache, or table-walker cache
(not for x86). A number of checks are put in place in the code to
ensure a read-only cache does not end up with dirty data.
A follow-on patch adds suitable read requests to allow a read-only
cache to explicitly ask for clean data.
This patch adds eviction notices to the caches, to provide accurate
tracking of cache blocks in snoop filters. We add the CleanEvict
message to the memory heirarchy and use both CleanEvicts and
Writebacks with BLOCK_CACHED flags to propagate notice of clean and
dirty evictions respectively, down the memory hierarchy. Note that the
BLOCK_CACHED flag indicates whether there exist any copies of the
evicted block in the caches above the evicting cache.
The purpose of the CleanEvict message is to notify snoop filters of
silent evictions in the relevant caches. The CleanEvict message
behaves much like a Writeback. CleanEvict is a write and a request but
unlike a Writeback, CleanEvict does not have data and does not need
exclusive access to the block. The cache generates the CleanEvict
message on a fill resulting in eviction of a clean block. Before
travelling downwards CleanEvict requests generate zero-time snoop
requests to check if the same block is cached in upper levels of the
memory heirarchy. If the block exists, the cache discards the
CleanEvict message. The snoops check the tags, writeback queue and the
MSHRs of upper level caches in a manner similar to snoops generated
from HardPFReqs. Currently CleanEvicts keep travelling towards main
memory unless they encounter the block corresponding to their address
or reach main memory (since we have no well defined point of
serialisation). Main memory simply discards CleanEvict messages.
We have modified the behavior of Writebacks, such that they generate
snoops to check for the presence of blocks in upper level caches. It
is possible in our current implmentation for a lower level cache to be
writing back a block while a shared copy of the same block exists in
the upper level cache. If the snoops find the same block in upper
level caches, we set the BLOCK_CACHED flag in the Writeback message.
We have also added logic to account for interaction of other message
types with CleanEvicts waiting in the writeback queue. A simple
example is of a response arriving at a cache removing any CleanEvicts
to the same address from the cache's writeback queue.
This patch fixes an issue which is very wide spread in the codebase,
causing sporadic linking failures. The issue is that we declare static
const class variables in the header, without any definition (as part
of a source file). In most cases the compiler propagates the value and
we have no issues. However, especially for less optimising builds such
as debug, we get sporadic linking failures due to undefined
references.
This patch fixes the Request class, by turning the static const flags
and master IDs into C++11 typed enums.
Remove the assert when adding a port to the RubyPort retry list.
Instead of asserting, just ignore the added port, since it's
already on the list.
Without this patch, Ruby+detailed fails for even the simplest tests
Snoop packets share the request pointer with the originating
packets. We need to ensure that the snoop packet destruction does not
delete the request. Snoops are used for reads, invalidations,
HardPFReqs, Writebacks and CleansEvicts. Reads, invalidations, and
HardPFReqs need a response so their snoops do not delete the
request. For Writebacks and CleanEvicts we need to check explicitly
for whethere the current packet is an express snoop, in whcih case do
not delete the request.
Fixes missed forward eviction to CPU. With the O3CPU this can lead to load-load
reordering, as the LQ is never notified of the invalidate.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
A single HMC-2500 x32 model based on:
[1] DRAMSpec: a high-level DRAM bank modelling tool developed at the University
of Kaiserslautern. This high level tool uses RC (resistance-capacitance) and CV
(capacitance-voltage) models to estimate the DRAM bank latency and power
numbers.
[2] A Logic-base Interconnect for Supporting Near Memory Computation in the
Hybrid Memory Cube (E. Azarkhish et. al) Assumed for the HMC model is a 30 nm
technology node. The modelled HMC consists of a 4 Gbit part with 4 layers
connected with TSVs. Each layer has 16 vaults and each vault consists of 2
banks per layer. In order to be able to use the same controller used for 2D
DRAM generations for HMC, the following analogy is done: Channel (DDR) => Vault
(HMC) device_size (DDR) => size of a single layer in a vault ranks per channel
(DDR) => number of layers banks per rank (DDR) => banks per layer devices per
rank (DDR) => devices per layer ( 1 for HMC). The parameters for which no
input is available are inherited from the DDR3 configuration.
A step towards removing RubyMemoryControl and shift users to
DRAMCtrl. The latter is faster, more representative, very versatile,
and is integrated with power models.
The processes of warming up and cooling down Ruby caches are simulation-wide
processes, not just RubySystem instance-specific processes. Thus, the warm-up
and cool-down variables should be globally visible to any Ruby components
participating in either process. Make these variables static members and track
the warm-up and cool-down processes as appropriate.
This patch also has two side benefits:
1) It removes references to the RubySystem g_system_ptr, which are problematic
for allowing multiple RubySystem instances in a single simulation. Warmup and
cooldown variables being static (global) reduces the need for instance-specific
dereferences through the RubySystem.
2) From the AbstractController, it removes local RubySystem pointers, which are
used inconsistently with other uses of the RubySystem: 11 other uses reference
the RubySystem with the g_system_ptr. Only sequencers have local pointers.
Sometimes, we need to defer an express snoop in an MSHR, but the original
request might complete and deallocate the original pkt->req. In those cases,
create a copy of the request so that someone who is inspecting the delayed
snoop can also inspect the request still. All of this is rather hacky, but the
allocation / linking and general life-time management of Packet and Request is
rather tricky. Deleting the copy is another tricky area, testing so far has
shown that the right copy is deleted at the right time.
The Request::UNCACHEABLE flag currently has two different
functions. The first, and obvious, function is to prevent the memory
system from caching data in the request. The second function is to
prevent reordering and speculation in CPU models.
This changeset gives the order/speculation requirement a separate flag
(Request::STRICT_ORDER). This flag prevents CPU models from doing the
following optimizations:
* Speculation: CPU models are not allowed to issue speculative
loads.
* Write combining: CPU models and caches are not allowed to merge
writes to the same cache line.
Note: The memory system may still reorder accesses unless the
UNCACHEABLE flag is set. It is therefore expected that the
STRICT_ORDER flag is combined with the UNCACHEABLE flag to prevent
this behavior.
This patch takes a last step in fixing issues related to uncacheable
accesses. We do not separate uncacheable memory from uncacheable
devices, and in cases where it is really memory, there are valid
scenarios where we need to snoop since we do not support cache
maintenance instructions (yet). On snooping an uncacheable access we
thus provide data if possible. In essence this makes uncacheable
accesses IO coherent.
The snoop filter is also queried to steer the snoops, but not updated
since the uncacheable accesses do not allocate a block.
This patch ensures that we pass on information about a packet being
shared (rather than exclusive), when forwarding a packet downstream.
Without this patch there is a risk that a downstream cache considers
the line exclusive when it really isn't.
This patch adds a missing counter update for the uncacheable
accesses. By updating this counter we also get a meaningful average
latency for uncacheable accesses (previously inf).
This patch changes the cache implementation to rely on virtual methods
rather than using the replacement policy as a template argument.
There is no impact on the simulation performance, and overall the
changes make it easier to modify (and subclass) the cache and/or
replacement policy.
Both open_adaptive and close_adaptive page polices keep the page
open if a row hit is found. If a row hit is not found, close_adaptive
page policy precharges the row, and open_adaptive policy precharges
the row only if there is a bank conflict request waiting in the queue.
This patch makes the checks for above conditions simpler.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
Restoring from a checkpoint with ruby + the DRAMCtrl memory model was not
working, because ruby and DRAMCtrl disagreed on the current tick during warmup.
Since there is no reason to do timing requests during warmup, use functional
requests instead.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
The stride prefetcher had a hardcoded number of contexts (i.e. master-IDs)
that it could handle. Since master IDs need to be unique per system, and
every core, cache etc. requires a separate master port, a static limit on
these does not make much sense.
Instead, this patch adds a small hash map that will map all master IDs to
the right prefetch state and dynamically allocates new state for new master
IDs.
This patch changes the order of writeback allocation such that any
writebacks resulting from a tag lookup (e.g. for an uncacheable
access), are added to the writebuffer before any new MSHR entries are
allocated. This ensures that the writebacks logically precedes the new
allocations.
The patch also changes the uncacheable flush to use proper timed (or
atomic) writebacks, as opposed to functional writes.
This patch simplifies the code dealing with uncacheable timing
accesses, aiming to align it with the existing miss handling. Similar
to what we do in atomic, a timing request now goes through
Cache::access (where the block is also flushed), and then proceeds to
ignore any existing MSHR for the block in question. This unifies the
flow for cacheable and uncacheable accesses, and for atomic and timing.
This patch changes how we search for matching MSHRs, ignoring any MSHR
that is allocated for an uncacheable access. By doing so, this patch
fixes a corner case in the MSHRs where incorrect data ended up being
copied into a (cacheable) read packet due to a first uncacheable MSHR
target of size 4, followed by a cacheable target to the same MSHR of
size 64. The latter target was filled with nonsense data.
This patch removes the no-longer-needed
allocateUncachedReadBuffer. Besides the checks it is exactly the same
as allocateMissBuffer and thus provides no value.
This patch aligns all MSHR queue entries to block boundaries to
simplify checks for matches. Previously there were corner cases that
could lead to existing entries not being identified as matches.
There are, rather alarmingly, a few regressions that change with this
patch.
This patch subsumes the PREFETCH_SNOOP_SQUASH flag with the more
generic BLOCK_CACHED flag. Future patches implementing cache eviction
messages can use the BLOCK_CACHED flag in almost the same manner as
hardware prefetches use the PREFETCH_SNOOP_SQUASH flag. The
PREFTECH_SNOOP_FLAG is set if the prefetch target is found in the tags
or the MSHRs in any state, so we are simply replacing calls to
setPrefetchSquashed() with setBlockCached(). The case of where the
prefetch target is found in the writeback MSHRs of upper level caches
continues to be covered by the MEM_INHIBIT flag.
The CommMonitor by default only allows memory traces to be gathered in
timing mode. This patch allows memory traces to be gathered in atomic
mode if all one needs is a functional trace of memory addresses used
and timing information is of a secondary concern.
For some reason we were checking mshr->hasTargets() even though
we had already called mshr->getTarget() unconditionally earlier
in the same function (which asserts if there are no targets).
Get rid of this useless check, and while we're at it get rid
of the redundant call to mshr->getTarget(), since we still have
the value saved in a local var.
Refactor the way that specific MemCmd values are generated for packets.
The new approach is a little more elegant in that we assign the right
value up front, and it's also more amenable to non-heap-allocated
Packet objects.
Also replaced the code in the Minor model that was still doing it the
ad-hoc way.
This is basically a refinement of http://repo.gem5.org/gem5/rev/711eb0e64249.
The 'if (writebacks.size)' check was redundant, because
writeBuffer.findMatches() would return false if the
writebacks list was empty.
Also renamed 'mshr' to 'wb_entry' in this context since
we are pointing at a writebuffer entry and not an MSHR
(even though it's the same C++ class).
This patch changes all the DPRINTF messages in the cache to use
'%#llx' every time a packet address is printed. The inclusion of '#'
ensures '0x' is prepended, and since the address type is a uint64_t %x
really should be %llx.
This patch fixes a rather subtle issue in the sending of MSHR requests
in the cache, where the logic previously did not check for conflicts
between the MSRH queue and the write queue when requests were not
ready. The correct thing to do is to always check, since not having a
ready MSHR does not guarantee that there is no conflict.
The underlying problem seems to have slipped past due to the symmetric
timings used for the write queue and MSHR queue. However, with the
recent timing changes the bug caused regressions to fail.
This patch changes the valid-bytes start/end to a proper byte
mask. With the changes in timing introduced in previous patches there
are more packets waiting in queues, and there are regressions using
the checker CPU failing due to non-contigous read data being found in
the various cache queues.
This patch also adds some more comments explaining what is going on,
and adds the fourth and missing case to Packet::checkFunctional.
By default, the packet queue is ordered by the ticks of the to-be-sent
packages. With the recent modifications of packages sinking their header time
when their resposne leaves the caches, there could be cases of MSHR targets
being allocated and ordered A, B, but their responses being sent out in the
order B,A. This led to inconsistencies in bus traffic, in particular the snoop
filter observing first a ReadExResp and later a ReadRespWithInv. Logically,
these were ordered the other way around behind the MSHR, but due to the timing
adjustments when inserting into the PacketQueue, they were sent out in the
wrong order on the bus, confusing the snoop filter.
This patch adds a flag (off by default) such that these special cases can
request in-order insertion into the packet queue, which might offset timing
slighty. This is expected to occur rarely and not affect timing results.
This patch makes the caches and memory controllers consume the delay
that is annotated to a packet by the crossbar. Previously many
components simply threw these delays away. Note that the devices still
do not pay for these delays.
This patch introduces a few subclasses to the CoherentXBar and
NoncoherentXBar to distinguish the different uses in the system. We
use the crossbar in a wide range of places: interfacing cores to the
L2, as a system interconnect, connecting I/O and peripherals,
etc. Needless to say, these crossbars have very different performance,
and the clock frequency alone is not enough to distinguish these
scenarios.
Instead of trying to capture every possible case, this patch
introduces dedicated subclasses for the three primary use-cases:
L2XBar, SystemXBar and IOXbar. More can be added if needed, and the
defaults can be overridden.
This patch introduces latencies in crossbar that were neglected
before. In particular, it adds three parameters in crossbar model:
front_end_latency, forward_latency, and response_latency. Along with
these parameters, three corresponding members are added:
frontEndLatency, forwardLatency, and responseLatency. The coherent
crossbar has an additional snoop_response_latency.
The latency of the request path through the xbar is set as
--> frontEndLatency + forwardLatency
In case the snoop filter is enabled, the request path latency is charged
also by look-up latency of the snoop filter.
--> frontEndLatency + SF(lookupLatency) + forwardLatency.
The latency of the response path through the xbar is set instead as
--> responseLatency.
In case of snoop response, if the response is treated as a normal response
the latency associated is again
--> responseLatency;
If instead it is forwarded as snoop response we add an additional variable
+ snoopResponseLatency
and the latency associated is
--> snoopResponseLatency;
Furthermore, this patch lets the crossbar progress on the next clock
edge after an unused retry, changing the time the crossbar considers
itself busy after sending a retry that was not acted upon.
This patch fixes a long-standing isue with the port flow
control. Before this patch the retry mechanism was shared between all
different packet classes. As a result, a snoop response could get
stuck behind a request waiting for a retry, even if the send/recv
functions were split. This caused message-dependent deadlocks in
stress-test scenarios.
The patch splits the retry into one per packet (message) class. Thus,
sendTimingReq has a corresponding recvReqRetry, sendTimingResp has
recvRespRetry etc. Most of the changes to the code involve simply
clarifying what type of request a specific object was accepting.
The biggest change in functionality is in the cache downstream packet
queue, facing the memory. This queue was shared by requests and snoop
responses, and it is now split into two queues, each with their own
flow control, but the same physical MasterPort. These changes fixes
the previously seen deadlocks.
This patch resolves a bug with hardware prefetches. Before a hardware prefetch
is sent towards the memory, the system generates a snoop request to check all
caches above the prefetch generating cache for the presence of the prefetth
target. If the prefetch target is found in the tags or the MSHRs of the upper
caches, the cache sets the prefetchSquashed flag in the snoop packet. When the
snoop packet returns with the prefetchSquashed flag set, the prefetch
generating cache deallocates the MSHR reserved for the prefetch. If the
prefetch target is found in the writeback buffer of the upper cache, the cache
sets the memInhibit flag, which signals the prefetch generating cache to
expect the data from the writeback. When the snoop packet returns with the
memInhibitAsserted flag set, it marks the allocated MSHR as inService and
waits for the data from the writeback.
If the prefetch target is found in multiple upper level caches, specifically
in the tags or MSHRs of one upper level cache and the writeback buffer of
another, the snoop packet will return with both prefetchSquashed and
memInhibitAsserted set, while the current code is not written to handle such
an outcome. Current code checks for the prefetchSquashed flag first, if it
finds the flag, it deallocates the reserved MSHR. This leads to assert failure
when the data from the writeback appears at cache. In this fix, we simply
switch the order of checks. We first check for memInhibitAsserted and then for
prefetch squashed.
Previously, the user would have to manually set access_backing_store=True
on all RubyPorts (Sequencers) in the config files.
Now, instead there is one global option that each RubyPort checks on
initialization.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
In highly loaded cases, reads might actually overlap with writes to the
initial memory state. The mem checker needs to detect such cases and
permit the read reading either from the writes (what it is doing now) or
read from the initial, unknown value.
This patch adds this logic.
This patch ensures we can run simulations with very large simulated
memories (at least 64 TB based on some quick runs on a Linux
workstation). In essence this allows us to efficiently deal with
sparse address maps without having to implement a redirection layer in
the backing store.
This opens up for run-time errors if we eventually exhausts the hosts
memory and swap space, but this should hopefully never happen.
This patch changes the range cache used in the global physical memory
to be an iterator so that we can use it not only as part of isMemAddr,
but also access and functionalAccess. This matches use-cases where a
core is using the atomic non-caching memory mode, and repeatedly calls
isMemAddr and access.
Linux boot on aarch32, with a single atomic CPU, is now more than 30%
faster when using "--fastmem" compared to not using the direct memory
access.
This patch clarifies the packet timings annotated
when going through a crossbar.
The old 'firstWordDelay' is replaced by 'headerDelay' that represents
the delay associated to the delivery of the header of the packet.
The old 'lastWordDelay' is replaced by 'payloadDelay' that represents
the delay needed to processing the payload of the packet.
For now the uses and values remain identical. However, going forward
the payloadDelay will be additive, and not include the
headerDelay. Follow-on patches will make the headerDelay capture the
pipeline latency incurred in the crossbar, whereas the payloadDelay
will capture the additional serialisation delay.
This patch adds some much-needed clarity in the specification of the
cache timing. For now, hit_latency and response_latency are kept as
top-level parameters, but the cache itself has a number of local
variables to better map the individual timing variables to different
behaviours (and sub-components).
The introduced variables are:
- lookupLatency: latency of tag lookup, occuring on any access
- forwardLatency: latency that occurs in case of outbound miss
- fillLatency: latency to fill a cache block
We keep the existing responseLatency
The forwardLatency is used by allocateInternalBuffer() for:
- MSHR allocateWriteBuffer (unchached write forwarded to WriteBuffer);
- MSHR allocateMissBuffer (cacheable miss in MSHR queue);
- MSHR allocateUncachedReadBuffer (unchached read allocated in MSHR
queue)
It is our assumption that the time for the above three buffers is the
same. Similarly, for snoop responses passing through the cache we use
forwardLatency.
This patch adds a bit of clarification around the assumptions made in
the cache when packets are sent out, and dirty responses are
pending. As part of the change, the marking of an MSHR as in service
is simplified slightly, and comments are added to explain what
assumptions are made.
This patch changes the DRAM channel interleaving default behaviour to
be more representative. The default address mapping (RoRaBaCoCh) moves
the channel bits towards the least significant bits, and uses 128 byte
as the default channel interleaving granularity.
These defaults can be overridden if desired, but should serve as a
sensible starting point for most use-cases.
This patch takes the final step in removing the src and dest fields in
the packet. These fields were rather confusing in that they only
remember a single multiplexing component, and pushed the
responsibility to the bridge and caches to store the fields in a
senderstate, thus effectively creating a stack. With the recent
changes to the crossbar response routing the crossbar is now
responsible without relying on the packet fields. Thus, these
variables are now unused and can be removed.
This patch removes the source field from the ForwardResponseRecord,
but keeps the class as it is part of how the cache identifies
responses to hardware prefetches that are snooped upwards.
This patch aligns how the response routing is done in the RubyPort,
using the SenderState for both memory and I/O accesses. Before this
patch, only the I/O used the SenderState, whereas the memory accesses
relied on the src field in the packet. With this patch we shift to
using SenderState in both cases, thus not relying on the src field any
longer.
This patch removes the need for a source and destination field in the
packet by shifting the onus of the tracking to the crossbar, much like
a real implementation. This change in behaviour also means we no
longer need a SenderState to remember the source/dest when ever we
have multiple crossbars in the system. Thus, the stack that was
created by the SenderState is not needed, and each crossbar locally
tracks the response routing.
The fields in the packet are still left behind as the RubyPort (which
also acts as a crossbar) does routing based on them. In the succeeding
patches the uses of the src and dest field will be removed. Combined,
these patches improve the simulation performance by roughly 2%.
This patch tidies up how we create and set the fields of a Request. In
essence it tries to use the constructor where possible (as opposed to
setPhys and setVirt), thus avoiding spreading the information across a
number of locations. In fact, setPhys is made private as part of this
patch, and a number of places where we callede setVirt instead uses
the appropriate constructor.
This patch fixes a bug where the DRAM controller tried to access the
system cacheline size before the system pointer was initialised. It
also fixes a bug where the granularity is 0 (no interleaving).
The cache's MemSidePacketQueue schedules a sendEvent based upon
nextMSHRReadyTime() which is the time when the next MSHR is ready or whenever
a future prefetch is ready. However, a prefetch being ready does not guarentee
that it can obtain an MSHR. So, when all MSHRs are full,
the simulation ends up unnecessiciarly scheduling a sendEvent every picosecond
until an MSHR is finally freed and the prefetch can happen.
This patch fixes this by not signaling the prefetch ready time if the prefetch
could not be generated. The event is rescheduled as soon as a MSHR becomes
available.
Previously the code commented about an unhandled case where it might be
possible for a writeback to arrive after a prefetch was generated but
before it was sent to the memory system. I hit that case. Luckily
the prefetchSquash() logic already in the code handles dropping prefetch
request in certian circumstances.
Re-organizes the prefetcher class structure. Previously the
BasePrefetcher forced multiple assumptions on the prefetchers that
inherited from it. This patch makes the BasePrefetcher class truly
representative of base functionality. For example, the base class no
longer enforces FIFO order. Instead, prefetchers with FIFO requests
(like the existing stride and tagged prefetchers) now inherit from a
new QueuedPrefetcher base class.
Finally, the stride-based prefetcher now assumes a custimizable lookup table
(sets/ways) rather than the previous fully associative structure.
Adds a new parameter that reserves some number of MSHR entries for demand
accesses. This helps prevent prefetchers from taking all MSHRs, forcing demand
requests from the CPU to stall.
This patch gives the user direct influence over the number of DRAM
ranks to make it easier to tune the memory density without affecting
the bandwidth (previously the only means of scaling the device count
was through the number of channels).
The patch also adds some basic sanity checks to ensure that the number
of ranks is a power of two (since we rely on bit slices in the address
decoding).
This patch addresses an issue seen with the KVM CPU where the refresh
events scheduled by the DRAM controller forces the simulator to switch
out of the KVM mode, thus killing performance.
The current patch works around the fact that we currently have no
proper API to inform a SimObject of the mode switches. Instead we rely
on drainResume being called after any switch, and cache the previous
mode locally to be able to decide on appropriate actions.
The switcheroo regression require a minor stats bump as a result.
This patch adds rank-wise refresh to the controller, as opposed to the
channel-wide refresh currently in place. In essence each rank can be
refreshed independently, and for this to be possible the controller
is extended with a state machine per rank.
Without this patch the data bus is always idle during a refresh, as
all the ranks are refreshing at the same time. With the rank-wise
refresh it is possible to use one rank while another one is
refreshing, and thus the data bus can be kept busy.
The patch introduces a Rank class to encapsulate the state per rank,
and also shifts all the relevant banks, activation tracking etc to the
rank. The arbitration is also updated to consider the state of the rank.
This patch adds a stand-alone stack distance calculator. The stack
distance calculator is a passive SimObject that observes the addresses
passed to it. It calculates stack distances (LRU Distances) of
incoming addresses based on the partial sum hierarchy tree algorithm
described by Alamasi et al. http://doi.acm.org/10.1145/773039.773043.
For each transaction a hashtable look-up is performed. At every
non-unique transaction the tree is traversed from the leaf at the
returned index to the root, the old node is deleted from the tree, and
the sums (to the right) are collected and decremented. The collected
sum represets the stack distance of the found node. At every unique
transaction the stack distance is returned as
numeric_limits<uint64>::max().
In addition to the basic stack distance calculation, a feature to mark
an old node in the tree is added. This is useful if it is required to
see the reuse pattern. For example, Writebacks to the lower level
(e.g. membus from L2), can be marked instead of being removed from the
stack (isMarked flag of Node set to True). And then later if this same
address is accessed (by L1), the value of the isMarked flag would be
True. This gives some insight on how the Writeback policy of the
lower level affect the read/write accesses in an application.
Debugging is enabled by setting the verify flag to true. Debugging is
implemented using a dummy stack that behaves in a naive way, using STL
vectors. Note that this has a large impact on run time.
This patch adds the MemChecker and MemCheckerMonitor classes. While
MemChecker can be integrated anywhere in the system and is independent,
the most convenient usage is through the MemCheckerMonitor -- this
however, puts limitations on where the MemChecker is able to observe
read/write transactions.
This patch takes a clean-slate approach to providing WriteInvalidate
(write streaming, full cache line writes without first reading)
support.
Unlike the prior attempt, which took an aggressive approach of directly
writing into the cache before handling the coherence actions, this
approach follows the existing cache flows as closely as possible.
This patch allows objects to get the src/dest of a packet even if it
is not set to a valid port id. This simplifies (ab)using the bridge as
a buffer and latency adapter in situations where the neighbouring
MemObjects are not crossbars.
The checks that were done in the packet are now shifted to the
crossbar where the fields are used to index into the port
arrays. Thus, the carrier of the information is not burdened with
checking, and the crossbar can check not only that the destination is
set, but also that the port index is within limits.
This patch attempts to make the rules for data allocation in the
packet explicit, understandable, and easy to verify. The constructor
that copies a packet is extended with an additional flag "alloc_data"
to enable the call site to explicitly say whether the newly created
packet is short-lived (a zero-time snoop), or has an unknown life-time
and therefore should allocate its own data (or copy a static pointer
in the case of static data).
The tricky case is the static data. In essence this is a
copy-avoidance scheme where the original source of the request (DMA,
CPU etc) does not ask the memory system to return data as part of the
packet, but instead provides a pointer, and then the memory system
carries this pointer around, and copies the appropriate data to the
location itself. Thus any derived packet actually never copies any
data. As the original source does not copy any data from the response
packet when arriving back at the source, we must maintain the copy of
the original pointer to not break the system. We might want to revisit
this one day and pay the price for a few extra memcpy invocations.
All in all this patch should make it easier to grok what is going on
in the memory system and how data is actually copied (or not).
This patch cleans up the use of hasData and checkFunctional in the
packet. The hasData function is unfortunately suggesting that it
checks if the packet has a valid data pointer, when it does in fact
only check if the specific packet type is specified to have a data
payload. The confusion led to a bug in checkFunctional. The latter
function is also tidied up to avoid name overloading.
This adds a basic level of sanity checking to the packet by ensuring
that a request is not modified once the packet is created. The only
issue that had to be worked around is the relaying of
software-prefetches in the cache. The specific situation is now solved
by first copying the request, and then creating a new packet
accordingly.
This patch tidies up the Request class, making all getters const. The
odd one out is incAccessDepth which is called by the memory system as
packets carry the request around. This is also const to enable the
packet to hold on to a const Request.
This patch simplifies how we deal with dynamically allocated data in
the packet, always assuming that it is array allocated, and hence
should be array deallocated (delete[] as opposed to delete). The only
uses of dataDynamic was in the Ruby testers.
The ARRAY_DATA flag in the packet is removed accordingly. No
defragmentation of the flags is done at this point, leaving a gap in
the bit masks.
As the last part the patch, it renames dataDynamicArray to dataDynamic.
This patch cleans up the packet memory allocation confusion. The data
is always allocated at the requesting side, when a packet is created
(or copied), and there is never a need for any device to allocate any
space if it is merely responding to a paket. This behaviour is in line
with how SystemC and TLM works as well, thus increasing
interoperability, and matching established conventions.
The redundant calls to Packet::allocate are removed, and the checks in
the function are tightened up to make sure data is only ever allocated
once. There are still some oddities in the packet copy constructor
where we copy the data pointer if it is static (without ownership),
and allocate new space if the data is dynamic (with ownership). The
latter is being worked on further in a follow-on patch.
This patch changes the various write functions in the port proxies
to use const pointers for all sources (similar to how memcpy works).
The one unfortunate aspect is the need for a const_cast in the packet,
to avoid having to juggle a const and a non-const data pointer. This
design decision can always be re-evaluated at a later stage.
This patch takes a first step in tightening up how we use the data
pointer in write packets. A const getter is added for the pointer
itself (getConstPtr), and a number of member functions are also made
const accordingly. In a range of places throughout the memory system
the new member is used.
The patch also removes the unused isReadWrite function.
This patch removes the parameter that enables bypassing the null check
in the Packet::getPtr method. A number of call sites assume the value
to be non-null.
The one odd case is the RubyTester, which issues zero-sized
prefetches(!), and despite being reads they had no valid data
pointer. This is now fixed, but the size oddity remains (unless anyone
object or has any good suggestions).
Finally, in the Ruby Sequencer, appropriate checks are made for flush
packets as they have no valid data pointer.
This patch adds a first cut GDDR5 config to accommodate the users
combining gem5 and GPUSim. The config is based on a SK Hynix
datasheet, and the Nvidia GTX580 specification. Someone from the
GPUSim user-camp should tweak the default page-policy and static
frontend and backend latencies.
This patch adds uncacheable/cacheable and read-only/read-write attributes to
the map method of PageTableBase. It also modifies the constructor of TlbEntry
structs for all architectures to consider the new attributes.
Ruby's functional accesses are not guaranteed to succeed as of now. While
this is not a problem for the protocols that are currently in the mainline
repo, it seems that coherence protocols for gpus rely on a backing store to
supply the correct data. The aim of this patch is to make this backing store
configurable i.e. it comes into play only when a particular option:
--access-backing-store is invoked.
The backing store has been there since M5 and GEMS were integrated. The only
difference is that earlier the system used to maintain the backing store and
ruby's copy was write-only. Sometime last year, we moved to data being
supplied supplied by ruby in SE mode simulations. And now we have patches on
the reviewboard, which remove ruby's copy of memory altogether and rely
completely on the system's memory to supply data. This patch adds back a
SimpleMemory member to RubySystem. This member is used only if the option:
access-backing-store is set to true. By default, the memory would not be
accessed.
This patch is the final in the series. The whole series and this patch in
particular were written with the aim of interfacing ruby's directory controller
with the memory controller in the classic memory system. This is being done
since ruby's memory controller has not being kept up to date with the changes
going on in DRAMs. Classic's memory controller is more up to date and
supports multiple different types of DRAM. This also brings classic and
ruby ever more close. The patch also changes ruby's memory controller to
expose the same interface.
This function was added when I had incorrectly arrived at the conclusion
that such a function can improve the chances of a functional read succeeding.
As was later realized, this is not possible in the current setup. While the
code using this function was dropped long back, this function was not. Hence
the patch.
This patch removes the data block present in the directory entry structure
of each protocol in gem5's mainline. Firstly, this is required for moving
towards common set of memory controllers for classic and ruby memory systems.
Secondly, the data block was being misused in several places. It was being
used for having free access to the physical memory instead of calling on the
memory controller.
From now on, the directory controller will not have a direct visibility into
the physical memory. The Memory Vector object now resides in the
Memory Controller class. This also means that some significant changes are
being made to the functional accesses in ruby.
In my opinion, it creates needless complications in rest of the code.
Also, this structure hinders the move towards common set of code for
physical memory controllers.
Both ruby and the system used to maintain memory copies. With the changes
carried for programmed io accesses, only one single memory is required for
fs simulations. This patch sets the copy of memory that used to reside
with the system to null, so that no space is allocated, but address checks
can still be carried out. All the memory accesses now source and sink values
to the memory maintained by ruby.
As of now DMASequencer inherits from the RubyPort class. But the code in
RubyPort class is heavily tailored for the CPU Sequencer. There are parts of
the code that are not required at all for the DMA sequencer. Moreover, the
next patch uses the dma sequencer for carrying out memory accesses for all the
io devices. Hence, it is better to have a leaner dma sequencer.
WriteInvalidate semantics depend on the unconditional writeback
or they won't complete. Also, there's no point in deferring snoops
on their MSHRs, as they don't get new data at the end of their life
cycle the way other transactions do.
Add comment in the cache about a minor inefficiency re: WriteInvalidate.
Since WriteInvalidate directly writes into the cache, it can
create tricky timing interleavings with reads and writes to the
same cache line that haven't yet completed. This patch ensures
that these requests, when completed, don't overwrite the newer
data from the WriteInvalidate.
This patch adds the size of the DRAM device to the DRAM config. It
also compares the actual DRAM size (calculated using information from
the config) to the size defined in the system. If these two values do
not match gem5 will print a warning. In order to do correct DRAM
research the size of the memory defined in the system should match the
size of the DRAM in the config. The timing and current parameters
found in the DRAM configs are defined for a DRAM device with a
specific size and would differ for another device with a different
size.
This patch adds two MemoryObject's: ExternalMaster and ExternalSlave.
Each object has a single port which can be bound to an externally-
provided bridge to a port of another simulation system at
initialisation.
This patch transitions the Ruby Message and its derived classes from
the ad-hoc RefCountingPtr to the c++11 shared_ptr. There are no
changes in behaviour, and the code modifications are mainly replacing
"new" with "make_shared".
The cloning of derived messages is slightly changed as they previously
relied on overriding the base-class through covariant return types.
This patch makes the memory system ISA-agnostic by enabling the Ruby
Sequencer to dynamically determine if it has to do a store check. To
enable this check, the ISA is encoded as an enum, and the system
is able to provide the ISA to the Sequencer at run time.
--HG--
rename : src/arch/x86/insts/microldstop.hh => src/arch/x86/ldstflags.hh
This patch takes a step towards an ISA-agnostic memory
system by enabling the components to establish the page size after
instantiation. The swap operation in the memory is now also allowing
any granularity to avoid depending on the IntReg of the ISA.
This patch adds some statistics to garnet that record the activity
of certain structures in the on-chip network. These statistics, in a later
patch, will be used for computing the energy consumed by the on-chip network.
Orion is being dropped from ruby. It would be replaced with DSENT
which has better models. Note that the power / energy numbers reported
after this patch has been applied are not for use.
This patch takes the final step in integrating DRAMPower and adds the
appropriate calls in the DRAM controller to provide the command trace
and extract the power and energy stats. The debug printouts are still
left in place, but will eventually be removed.
At the moment the DRAM power calculation is always on when using the
DRAM controller model. The run-time impact of this addition is around
1.5% when looking at the total host seconds of the regressions. We
deem this a sensible trade-off to avoid the complication of adding an
enable/disable mechanism.
This patch adds a class to wrap DRAMPower Library in gem5.
This class initiates an object of class MemorySpecification
of the DRAMPower Library, passes the parameters from DRAMCtrl.py
to this object and creates an object of drampower library using
the memory specification.
This patch adds missing timing and current parameters to the existing
DRAM configs. These missing timing and current parameters are required
by DRAMPower for the DRAM power calculations. The missing values are
datasheet values of the specified DRAMs, and the appropriate
references are added for the variuos configs.
This patch prunes the DDR3 config that was initially created to match
the default config of DRAMSim2. The config is not complete as it is,
and to avoid having to maintain it, the easiest way forward is to
simply prune it. Going forward we are adding power number etc to the
other configurations.
