mem: support for gpu-style RMWs in ruby
This patch adds support for GPU-style read-modify-write (RMW) operations in
ruby. Such atomic operations are traditionally executed at the memory controller
(instead of through an L1 cache using cache-line locking).
Currently, this patch works by propogating operation functors through the memory
system.
Just changes the metavar for --debug-start from TIME
to TICK in cset 72046b9b3323 and didn't notice that the
comment "must be in ticks" is now redundant.
For historical reasons, the ExecContext interface had a single
function, readMem(), that did two different things depending on
whether the ExecContext supported atomic memory mode (i.e.,
AtomicSimpleCPU) or timing memory mode (all the other models).
In the former case, it actually performed a memory read; in the
latter case, it merely initiated a read access, and the read
completion did not happen until later when a response packet
arrived from the memory system.
This led to some confusing things, including timing accesses
being required to provide a pointer for the return data even
though that pointer was only used in atomic mode.
This patch splits this interface, adding a new initiateMemRead()
function to the ExecContext interface to replace the timing-mode
use of readMem().
For consistency and clarity, the readMemTiming() helper function
in the ISA definitions is renamed to initiateMemRead() as well.
For x86, where the access size is passed in explicitly, we can
also get rid of the data parameter at this level. For other ISAs,
where the access size is determined from the type of the data
parameter, we have to keep the parameter for that purpose.
The read() function merely initiates a memory read operation; the
data doesn't arrive until the access completes and a response packet
is received from the memory system. Thus there's no need to provide
a data pointer; its existence is historical.
Getting this pointer out of this internal o3 interface sets the
stage for similar cleanup in the ExecContext interface. Also
found that we were pointlessly setting the contents at this pointer
on a store forward (the useful memcpy happens just a few lines
below the deleted one).
The readMemAtomic/writeMemAtomic helper functions were calling
readMemTiming/writeMemTiming respectively. This is functionally
correct, since the *Timing functions are doing the same access
initiation operation as the *Atomic functions (just that the
*Atomic versions also complete the access in line). It also
provides for some (very minimal) code reuse. Unfortunately,
it's potentially pretty confusing, since it makes it look like
the atomic accesses are somehow being converted to timing
accesses. It also gets in the way of specializing the timing
interface (as will be done in a future patch).
By ignoring SIG_TRAP, using --debug-break <N> when not connected to
a debugger becomes a no-op. Apparently this was intended to be a
feature, though the rationale is not clear.
If we don't ignore SIG_TRAP, then using --debug-break <N> when not
connected to a debugger causes the simulation process to terminate
at tick N. This is occasionally useful, e.g., if you just want to
collect a trace for a specific window of execution then you can combine
this with --debug-start to do exactly that.
In addition to not ignoring the signal, this patch also updates
the --debug-break help message and deletes a handful of unprotected
calls to Debug::breakpoint() that relied on the prior behavior.
Add a platform with support for both aarch32 and aarch64. This
platform implements a subset of the devices in a real Versatile
Express and extends it with some gem5-specific functionality. It is in
many ways similar to the old VExpress_EMM64 platform, but supports the
following new features:
* Automatic PCI interrupt assignment
* PCI interrupts allocated in a contiguous range.
* Automatic boot loader selection (32-bit / 64-bit)
* Cleaner memory map where gem5-specific devices live in CS5 which
isn't used by current Versatile Express platforms.
* No fake devices. Devices that were previously faked will be
removed from the device tree instead.
* Support for 510 GiB contiguous memory
Add support for automatic PCI interrupt routing using a device's ID on
the PCI bus. Our current DTBs typically tell the kernel that we do
this or something similar when declaring the PCI controller. This
changeset adds an option to make the simulator behave in the same way.
Interrupt routing can be selected by setting the int_policy parameter
in the GenericArmPciHost. The following values are supported:
* ARM_PCI_INT_STATIC: Use the old static routing policy using the
interrupt line from a device's configurtion space.
* ARM_PCI_INT_DEV: Use device number on the PCI bus to map to an
interrupt in the GIC. The interrupt is computed as:
gic_int = int_base + (pci_dev % int_count)
* ARM_PCI_INT_PIN: Use device interrupt pin on the PCI bus to map to
an interrupt in the GIC. The PCI specification reserves pin ID 0
for devices without interrupts, the interrupt therefore computed
as:
gic_int = int_base + ((pin - 1) % int_count)
The new Packet::setRaw() method incorrectly still contained
an htog() conversion. As a result, calls to the old set()
method (now defined as setRaw(htog(v))) underwent two htog
conversions, which breaks things when htog() is not a no-op.
Interestingly the only test that caught this was a SPARC
boot test, where an IsaFake device with a non-zero return
value was getting swapped twice resulting in a register
getting loaded with 0x100000000000000 instead of 1.
(Good reason for keeping SPARC around, perhaps?)
This patch replaces the gzstream zlib wrapper with the iostream3
wrapper provided as part of zlib contributions. The main reason for
the switch is to avoid including LGPL in the default gem5
build. iostream3 is provided under a more permissive license:
The code is provided "as is", with the permission to use, copy,
modify, distribute and sell it for any purpose without fee.
Distributed gem5 (abbreviated dist-gem5) is the result of the
convergence effort between multi-gem5 and pd-gem5 (from Univ. of
Wisconsin). It relies on the base multi-gem5 infrastructure for packet
forwarding, synchronisation and checkpointing but combines those with
the elaborated network switch model from pd-gem5.
--HG--
rename : src/dev/net/multi_etherlink.cc => src/dev/net/dist_etherlink.cc
rename : src/dev/net/multi_etherlink.hh => src/dev/net/dist_etherlink.hh
rename : src/dev/net/multi_iface.cc => src/dev/net/dist_iface.cc
rename : src/dev/net/multi_iface.hh => src/dev/net/dist_iface.hh
rename : src/dev/net/multi_packet.hh => src/dev/net/dist_packet.hh
Some of the DPRINTFs added to the classic cache in cset 45df88079f04,
while useful to those unfamiliar with the cache code, end up being
noise when you're familiar with the code but are trying to debug tricky
protocol issues. (Particularly getting two messages from each cache
as it receives a snoop request then declares that there was no match.)
This patch introduces a CacheVerbose debug flag, and moves a subset of
the added DPRINTFs into that category, so that Cache by itself returns
to being a more succinct summary of cache activity.
Also added a CacheAll compound flag to turn on all the cache-related
debug flags (other than CacheTags, which you *really* have to want badly
to turn it on, IMO).
This patch removes the NeedsWritable flag for all responses, as it is
really only the request that needs a writable response. The response,
on the other hand, should in these cases always provide the line in a
writable state, as indicated by the hasSharers flag not being set.
When we send requests that has NeedsWritable set, the response will
always have the hasSharers flag not set. Additionally, there are cases
where the request did not have NeedsWritable set, and we still get a
writable response with the hasSharers flag not set. This never happens
on snoops, but is used by downstream caches to pass ownership
upstream.
As part of this patch, the affected response types are updated, and
the snoop filter is similarly modified to check only the hasSharers
flag (as it should). A sanity check is also added to the packet class,
asserting that we never look at the NeedsWritable flag for responses.
No regressions are affected.
This patch looks at the request and response command to determine if
either actually has any data payload, and if not, we do not allocate
any space for packet data.
The only tricky case is where the command type is changed as part of
the MSHR functionality. In these cases where the original packet had
no data, but the new packet does, we need to explicitly call
allocate().
This patch ensures we do not respond with a Modified (dirty and
writable) line if the request is uncacheable, and that the cache
responding retains the line without modifying the state (even if
responding).
This patch changes the name of a bunch of packet flags and MSHR member
functions and variables to make the coherency protocol easier to
understand. In addition the patch adds and updates lots of
descriptions, explicitly spelling out assumptions.
The following name changes are made:
* the packet memInhibit flag is renamed to cacheResponding
* the packet sharedAsserted flag is renamed to hasSharers
* the packet NeedsExclusive attribute is renamed to NeedsWritable
* the packet isSupplyExclusive is renamed responderHadWritable
* the MSHR pendingDirty is renamed to pendingModified
The cache states, Modified, Owned, Exclusive, Shared are also called
out in the cache and MSHR code to make it easier to understand.
This patch is imported from reviewboard patch 2551 by Nilay.
This patch moves from a dynamically defined MachineType to a statically
defined one. The need for this patch was felt since a dynamically defined
type prevents us from having types for which no machine definition may
exist.
The following changes have been made:
i. each machine definition now uses a type from the MachineType enumeration
instead of any random identifier. This required changing the grammar and the
*.sm files.
ii. MachineType enumeration defined statically in RubySlicc_Exports.sm.
* * *
normal protocol fixes for nilay's parser machine type fix
This patch is imported from reviewboard patch 2550 by Nilay.
It was possible to specify multiple machine types with a single state machine.
This seems unnecessary and is being removed.
Add a sanity check to make it explicit that we currently do not allow
an I/O coherent agent to directly issue writes into the coherent part
of the memory system (it has to go via a cache, and get transformed
into a read ex, upgrade or invalidation).
This patch changes how the cache tracks which snoops are forwarded,
and which ones are created locally. Previously the identification was
based on an empty sender state of a specific class, but this method
fails to distinguish which cache actually attached the sender
state. Instead we use the same mechanism as the crossbar, and keep
track of the requests that have outstanding snoops.
This patch addresses a bug in how the cache attached the MSHR as a
sender state. Rather than overwriting any existing sender state it now
pushes a new one. The handling of upward snoops is also clarified.
Currently, the wire format of register values in g- and G-packets is
modelled using a union of uint8/16/32/64 arrays. The offset positions
of each register are expressed as a "register count" scaled according
to the width of the register in question. This results in counter-
intuitive and error-prone "register count arithmetic", and some
formats would even be altogether unrepresentable in such model, e.g.
a 64-bit register following a 32-bit one would have a fractional index
in the regs64 array.
Another difficulty is that the array is allocated before the actual
architecture of the workload is known (and therefore before the correct
size for the array can be calculated).
With this patch I propose a simpler mechanism for expressing the
register set structure. In the new code, GdbRegCache is an abstract
class; its subclasses contain straightforward structs reflecting the
register representation. The determination whether to use e.g. the
AArch32 vs. AArch64 register set (or SPARCv8 vs SPARCv9, etc.) is made
by polymorphically dispatching getregs() to the concrete subclass.
The subclass is not instantiated until it is needed for actual
g-/G-packet processing, when the mode is already known.
This patch is not meant to be merged in on its own, because it changes
the contract between src/base/remote_gdb.* and src/arch/*/remote_gdb.*,
so as it stands right now, it would break the other architectures.
In this patch only the base and the ARM code are provided for review;
once we agree on the structure, I will provide src/arch/*/remote_gdb.*
for the other architectures; those patches could then be merged in
together.
Review Request: http://reviews.gem5.org/r/3207/
Pushed by Joel Hestness <jthestness@gmail.com>
When adding an option to forward work items to the Python environment,
the new behavior was accidentally enabled by default. Set the value of
exit_on_work_items to False by default to revert to the old behavior
unless the simulation scripts explicitly requests work item
forwarding.
This patch fixes a corner case in the deferred snoop handling, where
requests ended up being used by multiple packets with different
lifetimes, and inadvertently got deleted while they were still in use.
There are cases where we want the Python world to handle work items
instead of the C++ world. However, that's currently not possible. This
changeset adds the forward_work_items option to the System class. Then
it is set to True, work items will generate workbegin/workend
simulation exists with the work item ID as the exit code and the old
C++ handling is completely bypassed.
--HG--
extra : rebase_source : 8de637a744fc4b6ff2bc763f00cdf8ddf2bff885
This patch allows the ruby random tester to use ruby ports that may only
support instr or data requests. This patch is similar to a previous changeset
(8932:1b2c17565ac8) that was unfortunately broken by subsequent changesets.
This current patch implements the support in a more straight-forward way.
Since retries are now tested when running the ruby random tester, this patch
splits up the retry and drain check behavior so that RubyPort children, such
as the GPUCoalescer, can perform those operations correctly without having to
duplicate code. Finally, the patch also includes better DPRINTFs for
debugging the tester.
Move pcidev.(hh|cc) to src/dev/pci/device.(hh|cc) and update existing
devices to use the new header location. This also renames the PCIDEV
debug flag to have a capitalization that is consistent with the PCI
host and other devices.
--HG--
rename : src/dev/Pci.py => src/dev/pci/PciDevice.py
rename : src/dev/pcidev.cc => src/dev/pci/device.cc
rename : src/dev/pcidev.hh => src/dev/pci/device.hh
rename : src/dev/pcireg.h => src/dev/pci/pcireg.h
The writefile pseudo instruction uses OutputDirectory::create and
OutputDirectory::openFile to create the output files. However, by
default these will check the file extention for .gz, and create a gzip
compressed stream if the file ending matches. When writing out files,
we want to write them out exactly as they are in the guest simulation,
and never want to compress them with gzio. Additionally, this causes
m5 writefile to fail when checking the error flags for the output
steam.
With this patch we add an additional no_gz argument to
OutputDirectory::create and OutputDirectory::openFile which allows us
to override the gzip compression. Therefore, for m5 writefile we
disable the filename check, and always create a standard ostream.
Previous ARM-based simulations were limited to 8 cores due to
limitations in GICv2 and earlier. This changeset adds a set of
gem5-specific extensions that enable support for up to 256 cores.
When the gem5 extensions are enabled, the GIC uses CPU IDs instead of
a CPU bitmask in the GIC's register interface. To OS can enable the
extensions by setting bit 0x200 in ICDICTR.
This changeset is based on previous work by Matt Evans.
There's a well-meaning check in Process::allocFD() to return an invalid
target fd (-1) if the incoming host fd is -1. However, this means that
emulated drivers, which want to allocate a target fd that doesn't
correspond to a host fd, can't use -1 to indicate an intentionally
invalid host fd.
It turns out the allocFD() check is redundant, as callers always test
the host fd for validity before calling. Also, callers never test the
return value of allocFD() for validity, so even if the test failed,
it would likely have the undesirable result of returning -1 to the
target app as a file descriptor without setting errno.
Thus the check is pointless and is now getting in the way, so it seems
we should just get rid of it.
This patch adds support to optionally capture the virtual address and asid
for load/store instructions in the elastic traces. If they are present in
the traces, Trace CPU will set those fields of the request during replay.
This patch replaces the booleans that specified the elastic trace record
type with an enum type. The source of change is the proto message for
elastic trace where the enum is introduced. The struct definitions in the
elastic trace probe listener as well as the Trace CPU replace the boleans
with the proto message enum.
The patch does not impact functionality, but traces are not compatible with
previous version. This is preparation for adding new types of records in
subsequent patches.
This patch defines a TraceCPU that replays trace generated using the elastic
trace probe attached to the O3 CPU model. The elastic trace is an execution
trace with data dependencies and ordering dependencies annoted to it. It also
replays fixed timestamp instruction fetch trace that is also generated by the
elastic trace probe.
The TraceCPU inherits from BaseCPU as a result of which some methods need
to be defined. It has two port subclasses inherited from MasterPort for
instruction and data ports. It issues the memory requests deducing the
timing from the trace and without performing real execution of micro-ops.
As soon as the last dependency for an instruction is complete,
its computational delay, also provided in the input trace is added. The
dependency-free nodes are maintained in a list, called 'ReadyList',
ordered by ready time. Instructions which depend on load stall until the
responses for read requests are received thus achieving elastic replay. If
the dependency is not found when adding a new node, it is assumed complete.
Thus, if this node is found to be completely dependency-free its issue time is
calculated and it is added to the ready list immediately. This is encapsulated
in the subclass ElasticDataGen.
If ready nodes are issued in an unconstrained way there can be more nodes
outstanding which results in divergence in timing compared to the O3CPU.
Therefore, the Trace CPU also models hardware resources. A sub-class to model
hardware resources is added which contains the maximum sizes of load buffer,
store buffer and ROB. If resources are not available, the node is not issued.
The 'depFreeQueue' structure holds nodes that are pending issue.
Modeling the ROB size in the Trace CPU as a resource limitation is arguably the
most important parameter of all resources. The ROB occupancy is estimated using
the newly added field 'robNum'. We need to use ROB number as sequence number is
at times much higher due to squashing and trace replay is focused on correct
path modeling.
A map called 'inFlightNodes' is added to track nodes that are not only in
the readyList but also load nodes that are executed (and thus removed from
readyList) but are not complete. ReadyList handles what and when to execute
next node while the inFlightNodes is used for resource modelling. The oldest
ROB number is updated when any node occupies the ROB or when an entry in the
ROB is released. The ROB occupancy is equal to the difference in the ROB number
of the newly dependency-free node and the oldest ROB number in flight.
If no node dependends on a non load/store node then there is no reason to track
it in the dependency graph. We filter out such nodes but count them and add a
weight field to the subsequent node that we do include in the trace. The weight
field is used to model ROB occupancy during replay.
The depFreeQueue is chosen to be FIFO so that child nodes which are in
program order get pushed into it in that order and thus issued in the in
program order, like in the O3CPU. This is also why the dependents is made a
sequential container, std::set to std::vector. We only check head of the
depFreeQueue as nodes are issued in order and blocking on head models that
better than looping the entire queue. An alternative choice would be to inspect
top N pending nodes where N is the issue-width. This is left for future as the
timing correlation looks good as it is.
At the start of an execution event, first we attempt to issue such pending
nodes by checking if appropriate resources have become available. If yes, we
compute the execute tick with respect to the time then. Then we proceed to
complete nodes from the readyList.
When a read response is received, sometimes a dependency on it that was
supposed to be released when it was issued is still not released. This occurs
because the dependent gets added to the graph after the read was sent. So the
check is made less strict and the dependency is marked complete on read
response instead of insisting that it should have been removed on read sent.
There is a check for requests spanning two cache lines as this condition
triggers an assert fail in the L1 cache. If it does then truncate the size
to access only until the end of that line and ignore the remainder.
Strictly-ordered requests are skipped and the dependencies on such requests
are handled by simply marking them complete immediately.
The simulated seconds can be calculated as the difference between the
final_tick stat and the tickOffset stat. A CountedExitEvent that contains
a static int belonging to the Trace CPU class as a down counter is used to
implement multi Trace CPU simulation exit.
The elastic trace is a type of probe listener and listens to probe points
in multiple stages of the O3CPU. The notify method is called on a probe
point typically when an instruction successfully progresses through that
stage.
As different listener methods mapped to the different probe points execute,
relevant information about the instruction, e.g. timestamps and register
accesses, are captured and stored in temporary InstExecInfo class objects.
When the instruction progresses through the commit stage, the timing and the
dependency information about the instruction is finalised and encapsulated in
a struct called TraceInfo. TraceInfo objects are collected in a list instead
of writing them out to the trace file one a time. This is required as the
trace is processed in chunks to evaluate order dependencies and computational
delay in case an instruction does not have any register dependencies. By this
we achieve a simpler algorithm during replay because every record in the
trace can be hooked onto a record in its past. The instruction dependency
trace is written out as a protobuf format file. A second trace containing
fetch requests at absolute timestamps is written to a separate protobuf
format file.
If the instruction is not executed then it is not added to the trace.
The code checks if the instruction had a fault, if it predicated
false and thus previous register values were restored or if it was a
load/store that did not have a request (e.g. when the size of the
request is zero). In all these cases the instruction is set as
executed by the Execute stage and is picked up by the commit probe
listener. But a request is not issued and registers are not written.
So practically, skipping these should not hurt the dependency modelling.
If squashing results in squashing younger instructions, it may happen that
the squash probe discards the inst and removes it from the temporary
store but execute stage deals with the instruction in the next cycle which
results in the execute probe seeing this inst as 'new' inst. A sequence
number of the last processed trace record is used to trap these cases and
not add to the temporary store.
The elastic instruction trace and fetch request trace can be read in and
played back by the TraceCPU.
This patch adds probe points in Fetch, IEW, Rename and Commit stages as follows.
A probe point is added in the Fetch stage for probing when a fetch request is
sent. Notify is fired on the probe point when a request is sent succesfully in
the first attempt as well as on a retry attempt.
Probe points are added in the IEW stage when an instruction begins to execute
and when execution is complete. This points can be used for monitoring the
execution time of an instruction.
Probe points are added in the Rename stage to probe renaming of source and
destination registers and when there is squashing. These probe points can be
used to track register dependencies and remove when there is squashing.
A probe point for squashing is added in Commit to probe squashed instructions.
The gem5's current PCI host functionality is very ad hoc. The current
implementations require PCI devices to be hooked up to the
configuration space via a separate configuration port. Devices query
the platform to get their config-space address range. Un-mapped parts
of the config space are intercepted using the XBar's default port
mechanism and a magic catch-all device (PciConfigAll).
This changeset redesigns the PCI host functionality to improve code
reuse and make config-space and interrupt mapping more
transparent. Existing platform code has been updated to use the new
PCI host and configured to stay backwards compatible (i.e., no
guest-side visible changes). The current implementation does not
expose any new functionality, but it can easily be extended with
features such as automatic interrupt mapping.
PCI devices now register themselves with a PCI host controller. The
host controller interface is defined in the abstract base class
PciHost. Registration is done by PciHost::registerDevice() which takes
the device, its bus position (bus/dev/func tuple), and its interrupt
pin (INTA-INTC) as a parameter. The registration interface returns a
PciHost::DeviceInterface that the PCI device can use to query memory
mappings and signal interrupts.
The host device manages the entire PCI configuration space. Accesses
to devices decoded into the devices bus position and then forwarded to
the correct device.
Basic PCI host functionality is implemented in the GenericPciHost base
class. Most platforms can use this class as a basic PCI controller. It
provides the following functionality:
* Configurable configuration space decoding. The number of bits
dedicated to a device is a prameter, making it possible to support
both CAM, ECAM, and legacy mappings.
* Basic interrupt mapping using the interruptLine value from a
device's configuration space. This behavior is the same as in the
old implementation. More advanced controllers can override the
interrupt mapping method to dynamically assign host interrupts to
PCI devices.
* Simple (base + addr) remapping from the PCI bus's address space to
physical addresses for PIO, memory, and DMA.
The assert in lsq_unit_impl.hh line 963 needs pktPending to be initialized to
NULL (I got the assertion failure several times without the fix).
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
The last SimObject using the legacy serialize API with non-const
methods has now been transitioned to the new API. This changeset
removes the serializeOld() methods from the serialization base class
as they are no longer used.
Add support for automatically discover available platforms. The
Python-side uses functionality similar to what we use when
auto-detecting available CPU models. The machine IDs have been updated
to match the platform configurations. If there isn't a matching
machine ID, the configuration scripts default to -1 which Linux uses
for device tree only platforms.
The HDLCD model implements a workaround that swaps the red and blue
channels. This works around an issue in certain old kernels. The new
driver doesn't seem to have this behavior, so disable the workaround
by default and enable it in the affected platforms.
Devices behind the Versatile Express configuration controllers are
currently all lumped into one SimObject. This will make DTB generation
challenging since the DTB assumes them to be in different parts of the
hierarchy. It also makes it hard to model other CoreTiles without also
replicating devices from the motherboard.
This changeset splits the VExpressCoreTileCtrl into two subsystems:
VExpressMCC for all motherboard-related devices and CoreTile2A15DCC
for Core Tile specific devices.
Add functionality to generate a back trace if gem5 crashes (SIGABRT or
SIGSEGV). The current implementation uses glibc's stack traversal
support if available and stubs out the call to print_backtrace()
otherwise.
Add support for automatically selecting a boot loader that matches the
guest system's kernel. Instead of accepting a single boot loader, the
ArmSystem class now accepts a vector of boot loaders. When
initializing a system, the we now look for the first boot loader with
an architecture that matches the kernel.
This changeset makes it possible to use the same system for both
64-bit and 32-bit kernels.
The MaltaPChip class is currently unused and identical (except for the
class name) to the TsunamiPChip. If someone decides to implement PCI
for Malta, they should make sure to share code with the Tsunami
implementation if they are similar.
The gem5 option '--list-sim-objects' is supposed to list all available
SimObjects and their parameters. It currently chokes on SimObjects
with parameters that have an object instance as their default
value. This is caused by __str__ in SimObject trying to resolve its
complete path. When the path resolution method reaches the parent
object (a MetaSimObject since it hasn't been instantiated), it dies
with a Python exception.
This changeset adds a guard to stop path resolution if the parent
object is a MetaSimObject.
Added the missing types EthernetAddr and Current to the JSON/INI file
reader example configs/example/read_config.py.
Also added __str__ to EthernetAddr to make values appear in the same form
in JSON an INI files.
The flash model has typos in its serialization code for
unknownPages, locationTable, blockValidEntries, and blockEmptyEntries
arrays where it would save each entry in the array under the same
name in the checkpoint. This patch fixes these typos.
As per the x86 architecture specification, matching TLB entries need to be
invalidated on a page fault. For instance, after a page fault due to inadequate
protection bits on a TLB hit, the TLB entry needs to be invalidated. This
behavior is clearly specified in the x86 architecture manuals from both AMD and
Intel. This invalidation is missing currently in gem5, due to which linux
kernel versions 3.8 and up cannot be simulated efficiently. This is exposed by
a linux optimisation in commit e4a1cc56e4d728eb87072c71c07581524e5160b1, which
removes a tlb flush on updating page table entries in x86.
Testing: Linux kernel versions 3.8 onwards were booting very slowly in FS mode,
due to repeated page faults (~300000 before the first print statement in a
bash file). Ensured that page fault rate drops drastically and observed
reduction in boot time from order of hours to minutes for linux kernel v3.8
and v3.11
doCpuid() has to identical warn messages about unimplemented functions. Add
the family to the log message to make them distinguishable.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
Remove sparc V8 TBR register from list of registers since it is not part of
sparc V9. This brings the number of registers in sync with what gdb expects
Without this patch gdb complains about receoved packet too long.
with this patch gdb is able to work properly with gem5 for remote debugging.
Note: gdb is version 7.8
Note: gdb is configured with --target=sparc64-sun-solaris2.8
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
the sanity check, while generally useful for exposing memory system bugs,
may be spurious with respect to GPU workloads, which may generate many more
requests than typical CPU workloads. the large number of requests generated
by the GPU may cause the req/resp queues to back up, thus queueing more than
100 packets.
The IICRPR register in the GIC is currently not being initialized when
the GIC is instantiated. Initialize to the value mandated by the
architecture specification.
This patch adds very basic checkpoint support for the VirtIO9PProxy
device. Previously, attempts to checkpoint gem5 with a present 9P
device caused gem5 to fatal as none of the state is tracked. We still
do not track any state, but we replace the fatal with a warning which
is triggered if the device has been used by the guest system. In the
event that it has not been used, we assume that no state is lost
during checkpointing. The warning is triggered on both a serialize and
an unserialize to ensure maximum visibility for the user.
Cleanup PCI devices to avoid using the PciDevice::platform pointer
directly. The PCI-specific functionality provided by the Platform
should be accessed through the wrappers in PciDevice.
This patch adds the necessary commands and cache functionality to
allow clean writebacks. This functionality is crucial, especially when
having exclusive (victim) caches. For example, if read-only L1
instruction caches are not sending clean writebacks, there will never
be any spills from the L1 to the L2. At the moment the cache model
defaults to not sending clean writebacks, and this should possibly be
re-evaluated.
The implementation of clean writebacks relies on a new packet command
WritebackClean, which acts much like a Writeback (renamed
WritebackDirty), and also much like a CleanEvict. On eviction of a
clean block the cache either sends a clean evict, or a clean
writeback, and if any copies are still cached upstream the clean
evict/writeback is dropped. Similarly, if a clean evict/writeback
reaches a cache where there are outstanding MSHRs for the block, the
packet is dropped. In the typical case though, the clean writeback
allocates a block in the downstream cache, and marks it writable if
the evicted block was writable.
The patch changes the O3_ARM_v7a L1 cache configuration and the
default L1 caches in config/common/Caches.py
This patch adds a parameter to control the cache clusivity, that is if
the cache is mostly inclusive or exclusive. At the moment there is no
intention to support strict policies, and thus the options are: 1)
mostly inclusive, or 2) mostly exclusive.
The choice of policy guides the behaviuor on a cache fill, and a new
helper function, allocOnFill, is created to encapsulate the decision
making process. For the timing mode, the decision is annotated on the
MSHR on sending out the downstream packet, and in atomic we directly
pass the decision to handleFill. We (ab)use the tempBlock in cases
where we are not allocating on fill, leaving the rest of the cache
unaffected. Simple and effective.
This patch also makes it more explicit that multiple caches are
allowed to consider a block writable (this is the case
also before this patch). That is, for a mostly inclusive cache,
multiple caches upstream may also consider the block exclusive. The
caches considering the block writable/exclusive all appear along the
same path to memory, and from a coherency protocol point of view it
works due to the fact that we always snoop upwards in zero time before
querying any downstream cache.
Note that this patch does not introduce clean writebacks. Thus, for
clean lines we are essentially removing a cache level if it is made
mostly exclusive. For example, lines from the read-only L1 instruction
cache or table-walker cache are always clean, and simply get dropped
rather than being passed to the L2. If the L2 is mostly exclusive and
does not allocate on fill it will thus never hold the line. A follow
on patch adds the clean writebacks.
The patch changes the L2 of the O3_ARM_v7a CPU configuration to be
mostly exclusive (and stats are affected accordingly).
This patch optimises the handling of writebacks and clean evictions
when using a snoop filter. Instead of snooping into the caches to
determine if the block is cached or not, simply set the status based
on the snoop-filter result.
Instead of conservatively enforcing order for all packets, which may
negatively impact the simulated-system performance, this patch updates
the packet queue such that it only applies the restriction if there
are already packets with the same address in the queue.
The basic need for the order enforcement is due to coherency
interactions where requests/responses to the same cache line must not
over-take each other. We rely on the fact that any packet that needs
order enforcement will have a block-aligned address. Thus, there is no
need for the queue to know about the cacheline size.
This patch enforces insertion order transmission of packets on the
response path in the cache. Note that the logic to enforce order is
already present in the packet queue, this patch simply turns it on for
queues in the response path.
Without this patch, there are corner cases where a request-response is
faster than a response-response forwarded through the cache. This
violation of queuing order causes problems in the snoop filter leaving
it with inaccurate information. This causes assert failures in the
snoop filter later on.
A follow on patch relaxes the order enforcement in the packet queue to
limit the performance impact.
This patch updates the I/O devices, bridge and simple memory to take
the packet header and payload delay into account in their latency
calculations. In all cases we add the header delay, i.e. the
accumulated pipeline delay of any crossbars, and the payload delay
needed for deserialisation of any payload.
Due to the additional unknown latency contribution, the packet queue
of the simple memory is changed to use insertion sorting based on the
time stamp. Moreover, since the memory hands out exclusive (non
shared) responses, we also need to ensure ordering for reads to the
same address.
This patch aligns how the memory-system slaves, i.e. the various
memory controllers and the bridge, identify and deal with sinking of
inhibited packets that are only useful within the coherent part of the
memory system.
In the future we could shift the onus to the crossbar, and add a
parameter "is_point_of_coherence" that would allow it to sink the
aforementioned packets.
This patch changes the CleanEvict command type to not be considered a
write. Initially it was made a zero-sized write to match the writeback
command, but as things developed it became clear that it causes more
problems than it solves. For example, the memory modules (and bridge)
should not consider the CleanEvict as a write, but instead discard
it. With this patch it will be neither a read, nor write, and as it
does not need a response the slave will simply sink it.
This patch unifies how we deal with delayed packet deletion, where the
receiving slave is responsible for deleting the packet, but the
sending agent (e.g. a cache) is still relying on the pointer until the
call to sendTimingReq completes. Previously we used a mix of a
deletion vector and a construct using unique_ptr. With this patch we
ensure all slaves use the latter approach.
The CoherentXBar currently doesn't check its queued slave ports when
receiving a functional snoop. This caused data corruption in cases
when a modified cache lines is forwarded between two caches.
Add the required functional calls into the queued slave ports.
This changeset adds a serial link model for the Hybrid Memory Cube (HMC).
SerialLink is a simple variation of the Bridge class, with the ability to
account for the latency of packet serialization. Also trySendTiming has been
modified to correctly model bandwidth.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
This patch models a simple HMC Controller. It simply schedules the incoming
packets to HMC Serial Links using a round robin mechanism. This patch should
be applied in series with other patches modeling a complete HMC device.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
Fix a bug in which the flash device would write out of bounds and
could either trigger a segfault and corrupt the memory of other
objects. This was caused by using pageSize in the place of
pagesPerBlock when running the garbage collector.
Also, added an assert to flag this condition in the future.
This patch fixes the drain logic for the UFSHostDevice and the
FlashDevice. In the case of the FlashDevice, the logic for CheckDrain
needed to be reversed, whilst in the case of the UFSHostDevice check
drain was never being called. In both cases the system would never
complete draining if the initial attampt to drain failed.
This patch addresses the upgrading of deferred targets in the MSHR,
and makes it clearer by explicitly calling out what is happening
(deferred targets are promoted if we get exclusivity without asking
for it).
This patch adds explicit overrides as this is now required when using
"-Wall" with clang >= 3.5, the latter now part of the most recent
XCode. The patch consequently removes "virtual" for those methods
where "override" is added. The latter should be enough of an
indication.
As part of this patch, a few minor issues that clang >= 3.5 complains
about are also resolved (unused methods and variables).
This patch moves away from using M5_ATTR_OVERRIDE and the m5::hashmap
(and similar) abstractions, as these are no longer needed with gcc 4.7
and clang 3.1 as minimum compiler versions.
ARM uses UDelayEvents to emulate kernel __*udelay functions and speed up
simulation. UDelayEvents call Pseudoinst::quiesceNs to quiesce the system for
a specified delay. Changeset 10341:0b4d10f53c2d introduced the requirement
that any quiesce process that is started must also be completed by scheduling
an EndQuiesceEvent. This change causes the CPU to hang if an IsQuiesce
instruction is executed, but the corresponding EndQuiesceEvent is not
scheduled.
Changeset 11058:d0934b57735a introduces a fix for uses of PseudoInst::quiesce*
that would conditionally execute the EndQuiesceEvent. ARM UDelayEvents specify
quiesce period of 0 ns (src/arch/arm/linux/system.cc), so changeset 11058
causes these events to now execute full quiesce processes, greatly increasing
the total instructions executed in kernel delay loops and slowing simulation.
This patch updates the UDelayEvent to conditionally execute
PseudoInst::quiesceNs (**a quiesce operation**) only if the specified
delay is >0 ns. The result is ARM delay loops no longer execute instructions
for quiesce handling, and regression time returns to normal.
The decoder is responsible for splitting instructions in micro
operations (uops). Given that different micro architectures may split
operations differently, this patch allows to specify which micro
architecture each isa implements, so different cores in the system can
split instructions differently, also decoupling uop splitting
(microArch) from ISA (Arch). This is done making the decodification
calls templates that receive a type 'DecoderFlavour' that maps the
name of the operation to the class that implements it. This way there
is only one selection point (converting the command line enum to the
appropriate DecodeFeatures object). In addition, there is no explicit
code replication: template instantiation hides that, and the compiler
should be able to resolve a number of things at compile-time.
Although some decent error messages were getting generated inside
isa_parser.py, they weren't always getting printed because of the
screwy way we were handling exceptions. (Basically an inner
exception would get hidden by an outer exception, and the more
informative inner error message would not get printed.)
Also line numbers were messed up, since they were taken from the
lexer, which is typically a token (or more) ahead of the grammar
rule that's being matched. Using the 'lineno' attribute that
PLY associates with the grammar production is more accurate.
The new LineTracker class extends lineno to track filenames as
well as line numbers.
This information is useful if you have a bunch of simulations running
and want to know which one to kill, but you've neglected to take
advantage of the previous patch and embed the pid in your output path.
These are packed single-precision approximate reciprocal operations,
vector and scalar versions, respectively.
This code was basically developed by copying the code for
sqrtps and sqrtss. The mrcp micro-op was simplified relative to
msqrt since there are no double-precision versions of this operation.
fild loads an integer value into the x87 top of stack register.
fucomi/fucomip compare two x87 register values (the latter
also doing a stack pop).
These instructions are used by some versions of GNU libstdc++.
The DTRACE() macro tests both Trace::enabled and the specific flag. This
change uses the same administrative interface for enabling/disabling
tracing, but masks the SimpleFlags settings directly. This eliminates a
load for every DTRACE() test, e.g. DPRINTF.
In ARM, certain variables are only updated when a necessary change is
detected. Having 2 SMT threads share a TLB resulted in these not being
updated as required. This patch adds a thread context identifer to
assist in the invalidation of these variables.
Adds SMT support to the "simple" CPU models so that they can be
used with other SMT-supported CPUs. Example usage: this enables
the TimingSimpleCPU to be used to warmup caches before swapping to
detailed mode with the in-order or out-of-order based CPU models.
Trying to run an SE system with varying threads per core (SMT cores + Non-SMT
cores) caused failures due to the CPU id assignment logic. The comment
about thread assignment (worrying about core 0 not having tid 0) seems
not to be valid given that our configuration scripts initialize them in
order.
This removes that constraint so a heterogenously threaded sytem can work.
If a cache entry permission was previously set to NotPresent, but the entry was
not deleted, a following cache allocation can cause the entry to be leaked by
setting the entry pointer to a newly allocated entry. To eliminate this
possibility, check if the new entry is different from the old one, and if so,
delete the old one.
IntDevice::recvResponse is called from two places in current mainline: (1) the
short circuit path of X86ISA::IntDevice::IntMasterPort::sendMessage for atomic
mode, and (2) the full request->response path to and from the x86 interrupts
device (finally called from MessageMasterPort::recvTimingResp). In the former
case, the packet was deleted correctly, but in the latter case, the packet and
request leak. To fix the leak, move request and packet deletion into IntDevice
inherited class implementations of recvResponse.
In RubyPort::ruby_eviction_callback, prior changes fixed a memory leak caused
by instantiating separate packets for each port that the eviction was forwarded
to. That change, however, left the instantiated request to also leak. Allocate
it on the stack to avoid the leak.
Recent changes to memory access queuing allocate requests for packets sent to
memory controllers, but did not free the requests. Delete them to avoid leaks.
Changes to the RubyMemoryControl removed the dequeue function, which deleted
MemoryNode instances. This results in leaked MemoryNode instances. Correctly
delete these instances.
The recent changeset to readlink() to handle reading the /proc/self/exe link
introduces a number of problems. This patch fixes two:
1) Because readlink() called on /proc/self/exe now uses LiveProcess::progName()
to find the binary path, it will only get the zeroth parameter of the simulated
system command line. However, if a config script also specifies the process'
executable, the executable parameter is used to create the LiveProcess rather
than the zeroth command line parameter. Thus, the zeroth command line parameter
is not necessarily the correct path to the binary executing in the simulated
system. To fix this, add a LiveProcess data member, 'executable', which is
correctly set during instantiation and returned from progName().
2) If a config script allows a user to pass a relative path as the zeroth
simulated system command line parameter or process executable, readlink() will
incorrecly return a relative path when called on '/proc/self/exe'.
/proc/self/exe is always set to a full path, so running benchmarks can fail if
a relative path is returned. To fix this, clean up the handling of
LiveProcess::progName() within readlink() to get the full binary path.
NOTE: This patch still leaves the potential problem that host full path to the
binary bleeds into the simulated system, potentially causing the appearance of
non-deterministic simulated system execution.
This patch fixes a use-after-delete issue in the packet probe points
by adding a PacketInfo struct to retain the key fields before passing
the packet onwards. We want to probe the packet after it is
successfully sent, but by that time the fields may be modified, and
the packet may even be deleted.
Amazingly enough the issue has gone undetected for months, and only
recently popped up in our regressions.
This patch fixes issues in the interactions between deferred snoops
and WriteLineReq. More specifically, the patch addresses an issue
where deferred snoops caused assertion failures when being serviced on
the arrival of an InvalidateResp. The response packet was perceived to
be invalidating, when actually it is not for the cache that sent out
the original invalidation request.
This patch changes the tracking of ports in the snoop filter to use
local dense port IDs so that we can have 64 snooping ports (rather
than crossbar slave ports). This is achieved by adding a simple
remapping vector that translates the actal port IDs into the local
slave IDs used in the SnoopMask.
Ultimately this patch allows us to scale to much larger systems
without introducing a hierarchy of crossbars.
This patch adds a snoop filter to the L2XBar. For now we refrain from
globally adding a snoop filter to the SystemXBar, since the latter is
also used in systems without caches. In scenarios without caches the
snoop filter will not see any writeback/clean evicts from the CPU
ports, despite the fact that they are snooping. To avoid inadvertent
use of the snoop filter in these cases we leave it out for now.
A size check is added to the snoop filter, merely to ensure it does
not grow beyond the total capacity of the caches above it. The size
has to be set manually, and a value of 8 MByte is choosen as suitably
high default.
This patch introduces a private member storing the iterator from the
lookupRequest call, such that it can be re-used when the request
eventually finishes. The method previously called updateRequest is
renamed finishRequest to make it more clear that the two functions
must be called together.
This patch mirrors the logic in timing mode which sends up snoops to
check for cached copies before sending CleanEvicts and Writebacks down
the memory hierarchy. In case there is a copy in a cache above,
discard CleanEvicts and set the BLOCK_CACHED flag in Writebacks so
that writebacks do not reset the cache residency bit in the snoop
filter below.
This patch adds the functionality to properly track CleanEvicts and
Writebacks in the snoop filter. Previously there were no CleanEvicts, and
Writebacks did not send up snoops to ensure there were no copies in
caches above. Hence a writeback could never erase an entry from the
snoop filter.
When a CleanEvict message reaches a snoop filter, it confirms that the
BLOCK_CACHED flag is not set and resets the bits corresponding to the
CleanEvict address and port it arrived on. If none of the other peer
caches have (or have requested) the block, the snoop filter forwards
the CleanEvict to lower levels of memory. In case of a Writeback
message, the snoop filter checks if the BLOCK_CACHED flag is not set
and only then resets the bits corresponding to the Writeback
address. If any of the other peer caches have (or has requested) the
same block, the snoop filter sets the BLOCK_CACHED flag in the
Writeback before forwarding it to lower levels of memory heirarachy.
This patch prevents the snoop filter from creating items for requests
originating from non-snooping ports. The allocation decision is thus
based both on the cacheability of the line, and the snooping status of
the source port. Ultimately we should check if the source of the
packet is caching, since also the CPU ports are snooping (but not
allocating). Thus, at the moment we rely on the snoop filter being
used together with caches.
The patch also transitions to use the Packet::getBlockAddr in
determining the line address.
This patch introduces the concept of a snoop latency. Given the
requirement to snoop and forward packets in zero time (due to the
coherency mechanism), the latency is accounted for later.
On a snoop, we establish the latency, and later add it to the header
delay of the packet. To allow multiple caches to contribute to the
snoop latency, we use a separate variable in the packet, and then take
the maximum before adding it to the header delay.
This patch ensures that the snoop-filter latency only contributes to
the packet latency, and not to the crossbar throughput/occupancy. In
essence we treat the snoop-filter lookup as pipelined.
Created the following HBM configurations:
1) HBM gen1 (x128/CH), 2Gb die, 4H stack, 1Gbps, 8 channels
2) HBM gen2 (x64/PC), 8Gb die, 4H stack, 1Gbps, 16 pseudo-channels
The configuration values are based on:
- The HBM gen1 public JEDEC spec
- Publically released data from MemCon presentations
- Timing extrapolated from existing LPDDR configurations
Will adjust once specs become available.
Changeset 4872dbdea907 replaced Address by Addr, but did not make changes to
print statements. So the addresses which were being printed in hex earlier
along with their line address, were now being printed in decimals. This patch
adds a function printAddress(Addr) that can be used to print the address in hex
along with the lines address. This function has been put to use in some of the
places. At other places, change has been made to print just the address in
hex.
The DataMember class in Type.py was being derived from PairContainer. A
separate Var object was also created for the DataMember. This meant some
duplication of across the members of these two classes (Var and DataMember).
This patch changes DataMember from Var instead. There is no obvious reason to
derive from PairContainer which can only hold pairs, something that Var class
already supports. The only thing that DataMember has over Var is init_code,
which is being retained. This change would later on help in having pointers
in DataMembers.
Some blocks in MOESI hammer were not getting deallocated when they were set to
an idle state (e.g. by invalidate or other_getx/s messages). While
functionally correct, this caused some bad effects on performance, such as
blocks in I in the L1s getting sent to the L2 upon eviction, in turn evicting
valid blocks. Also, if a valid block was in LRU, that block could be evicted
rather than a block in I. This patch adds in the missing deallocations.
Committed by: Nilay Vaish<nilay@cs.wisc.edu>
The recent changes to make MessageBuffers SimObjects required them to be
initialized in a particular order, which could break some protocols. Fix this
by calling initNetQueues on the external nodes of each external link in the
constructor of Network.
This patch also refactors the duplicated code for checking network allocation
and setting net queues (which are called by initNetQueues) from the simple and
garnet networks to be in Network.
This patch changes MessageBuffer and TimerTable, two structures used for
buffering messages by components in ruby. These structures would no longer
maintain pointers to clock objects. Functions in these structures have been
changed to take as input current time in Tick. Similarly, these structures
will not operate on Cycle valued latencies for different operations. The
corresponding functions would need to be provided with these latencies by
components invoking the relevant functions. These latencies should also be
in Ticks.
I felt the need for these changes while trying to speed up ruby. The ultimate
aim is to eliminate Consumer class and replace it with an EventManager object in
the MessageBuffer and TimerTable classes. This object would be used for
scheduling events. The event itself would contain information on the object and
function to be invoked.
In hindsight, it seems I should have done this while I was moving away from use
of a single global clock in the memory system. That change led to introduction
of clock objects that replaced the global clock object. It never crossed my
mind that having clock object pointers is not a good design. And now I really
don't like the fact that we have separate consumer, receiver and sender
pointers in message buffers.
The eventual aim of this change is to pass RubySystem pointers through to
objects generated from the SLICC protocol code.
Because some of these objects need to dereference their RubySystem pointers,
they need access to the System.hh header file.
In src/mem/ruby/SConscript, the MakeInclude function creates single-line header
files in the build directory that do nothing except include the corresponding
header file from the source tree.
However, SLICC also generates a list of header files from its symbol table, and
writes it to mem/protocol/Types.hh in the build directory. This code assumes
that the header file name is the same as the class name.
The end result of this is the many of the generated slicc files try to include
RubySystem.hh, when the file they really need is System.hh. The path of least
resistence is just to rename System.hh to RubySystem.hh.
--HG--
rename : src/mem/ruby/system/System.cc => src/mem/ruby/system/RubySystem.cc
rename : src/mem/ruby/system/System.hh => src/mem/ruby/system/RubySystem.hh
This register is writable according to UA2005
Tried to boot NetBSD which starts the kernel by writing to the tick_cmpr
register. Without the patch gem5 crashes with a panic. With the patch NetBSD
starts to boot normally (although sun4v support in NetBSD is not complete yet)
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
Handle bad IDE disk image size 0. When image size is 0, gem5 will cause an
exception with log "Floating point exception (core dumped)".
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
When a branch gets squashed, it's speculative branch predictor state should get
rolled back in squash(). However, only the globalHistory state was being
rolled back. This patch adds (at least some) support for rolling back the
local predictor state also.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
This patch enables instructions in LSQ to track two physical addresses for
corresponding two split requests. Later, the information is used in
checksnoop() to search for/invalidate the corresponding LD instructions.
The current implementation has kept track of only the physical address that is
referenced by the first split request. Thus, for checksnoop(), the line
accessed by the second request has not been considered, causing potential
correctness issues.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
Refactored the code in operateVnet(), moved partly to a new function
operateMessageBuffer(). This is required since a later patch moves to having a
wakeup event per MessageBuffer instead of one event for the entire Switch.
There are two reasons for doing so:
a. provide a source of clock to PerfectSwitch. A follow on patch removes sender
and receiver pointers from MessageBuffer means that the object owning the
buffer should have some way of providing timing info.
b. schedule events. A follow on patch removes the consumer class. So the
PerfectSwitch needs some EventManager object to schedule events on its own.
Add a stat that counts buffer underruns in the HDLCD controller. The
stat counts at most one underrun per frame since the controller aborts
the current frame if it underruns.
Rewrite the HDLCD controller to use the new DMA engine and pixel
pump. This fixes several bugs in the current implementation:
* Broken/missing interrupt support (VSync, underrun, DMA end)
* Fragile resolution changes (changing resolutions used
to cause assertion errors).
* Support for resolutions with a width that isn't divisible by 32.
* The pixel clock can now be set dynamically.
This breaks checkpoint compatibility. Checkpoints can be upgraded with
the checkpoint conversion script. However, upgraded checkpoints won't
contain the state of the current frame. That means that HDLCD
controllers restoring from a converted checkpoint immediately start
drawing a new frame (i.e, expect timing differences).
Currently the sequencer calls the function setMRU that updates the replacement
policy structures with the first level caches. While functionally this is
correct, the problem is that this requires calling findTagInSet() which is an
expensive function. This patch removes the calls to setMRU from the sequencer.
All controllers should now update the replacement policy on their own.
The set and the way index for a given cache entry can be found within the
AbstractCacheEntry structure. Use these indicies to update the replacement
policy structures.
The current Set data structure is slow and therefore is being reimplemented
using std::bitset. A maximum limit of 64 is being set on the number of
controllers of each type. This means that for simulating a system with more
controllers of a given type, one would need to change the value of the variable
NUMBER_BITS_PER_SET
MessageBuffer is a SimObject now. There were protocols that still declared
some of the message buffers are variables of the controller, but not as input
parameters. Special handling was required for these variables in the SLICC
compiler. This patch changes this. Now all message buffers are declared as
input parameters.
In cases where a newly added target does not have any upstream MSHR to
mark as downstreamPending, remember that nothing is marked. This
allows us to avoid attempting to find the MSHR as part of the clearing
of downstreamPending.
This commit addresses gem5 checkpoints' linear versioning bottleneck.
Since development is distributed across many private trees, there exists
a sort of 'race' for checkpoint version numbers: internally a checkpoint
version may be used but then resynchronizing with the external tree causes
a conflict on that version. This change replaces the linear version number
with a set of unique strings called tags. Now the only conflicts that can
arise are of tag names, where collisions are much easier to avoid.
The checkpoint upgrader (util/cpt_upgrader.py) upgrades the version
representation, as one would expect. Each tag version implements its
upgrader code in a python file in the util/cpt_upgraders directory
rather than adding a function to the upgrader script itself.
The version tags are stored in the 'Globals' section rather than 'root'
(as the version was previously) because 'Globals' gets unserialized
first and can provide a warning before any other unserialization errors
can occur.
This is in support of tag-based checkpoint versioning. It should be
possible to examine an optional parameter in a checkpoint during
unserialization and not have it throw a warning.
We no longer use the C library based random number generator: random().
Instead we use the C++ library provided rng. So setting the random seed for
the RubySystem class has no effect. Hence the variable and the corresponding
option are being dropped.
Event auto-serialization no longer in use and has been broken ever
since the introduction of PDES support almost two years
ago. Additionally, serializing the individual event queues is
undesirable since it exposes the thread structure of the
simulator. What this means in practice is that the number of threads
in the simulator must be the same when taking a checkpoint and when
loading the checkpoint.
This changeset removes support for the AutoSerialize event flag and
the associated serialization code.
EtherLink currently uses a fire-and-forget link delay event that
delays sending of packets by a fixed number of ticks. In order to
serialize this event, it relies on the event queue's auto
serialization support. However, support for event auto serialization
has been broken for more than two years, which means that checkpoints
of multi-system setups are likely to drop in-flight packets.
This changeset the replaces rewrites this part of the EtherLink to use
a packet queue instead. The queue contains a (tick, packet) tuple. The
tick indicates when the packet will be ready. Instead of relying on
event autoserialization, we now explicitly serialize the packet queue
in the EhterLink::Link class.
Note that this changeset changes the way in-flight packages are
serialized. Old checkpoints will still load, but in-flight packets
will be dropped (just as before). There has been no attempt to upgrade
checkpoints since this would actually change the behavior of existing
checkpoints.
This changeset removes the support for the autoserialize parameter in
GlobalSimLoopExitEvent (including exitSimLoop()) and
LocalSimLoopExitEvent.
Auto-serialization of the LocalSimLoopExitEvent was never used, so
this is not expected to affect anything. However, it was sometimes
used for GlobalSimLoopExitEvent. Unfortunately, serialization of
global events has never been supported, so checkpoints with such
events will currently cause simulation panics.
The serialize parameter to exitSimLoop() has been left in-place to
maintain API compatibility (removing it would affect m5ops). Instead
of just dropping it, we now print a warning if the parameter is set
and the exit event is scheduled in the future (i.e., not at the
current tick).
The object resolver isn't serialization specific and shouldn't live in
serialize.hh. Move it to sim_object.hh since it queries to the
SimObject hierarchy.
This member indicates whether or not a particular virtual network is in use.
Instead of having a default big value for the number of virtual networks and
then checking whether a virtual network is in use, the next patch removes the
default value and the protocol configuration file would now specify the
number of virtual networks it requires.
Additionally, the patch also refactors some of the code used for computing the
virtual channel next in the round robin order.
Both FuncCallExprAST and MethodCallExprAST had code for checking the arguments
with which a function is being called. The patch does away with this
duplication. Now the code for checking function call arguments resides in the
Func class.
The new serialization code (kudos to Tim Jones) moves all of the state
mangling in RubySystem to memWriteback. This makes it possible to use
the new const serialization interface.
This changeset moves the cache recorder cleanup from the checkpoint()
method to drainResume() to make checkpointing truly constant and
updates the checkpointing code to use the new interface.
The sequencer takes care of llsc accesses by calling upon functions
from the CacheMemory. This is unnecessary once the required CacheEntry object
is available. Thus some of the calls to findTagInSet() are avoided.
The O3CPU blocks the Fetch when it sees a quiesce instruction (IsQuiesce flag).
When the inst. is executed, a quiesce event is created to reactivate the
context and unblock the Fetch.
If the quiesceNs or quiesceCycles are called with a value of 0, the
QuiesceEvent will not be created and the Fetch stage will remain blocked.
Committed by Joel Hestness <jthestness@gmail.com>
This patch reverts part of (842f56345a42), as apparently there are
use-cases outside the main repository relying on the late setting of
the physical address.
This patch simplifies the packet, and removes the possibility of
creating a packet without a valid address and/or size. Under no
circumstances are these fields set at a later point, and thus they
really have to be provided at construction time.
The patch also fixes a case there the MinorCPU creates a packet
without a valid address and size, only to later delete it.
Cleaning up dead code. The CLREX stores zero directly to
MISCREG_LOCKFLAG and so the request flag is no longer needed. The
corresponding functionality in the cache tags is also removed.
Open up for other subclasses to BaseCache and transition to using the
explicit Cache subclass.
--HG--
rename : src/mem/cache/BaseCache.py => src/mem/cache/Cache.py
This patch serves to avoid name clashes with the classic cache. For
some reason having two 'SimObject' files with the same name creates
problems.
--HG--
rename : src/mem/ruby/structures/Cache.py => src/mem/ruby/structures/RubyCache.py
We no longer use the C library based random number generator: random().
Instead we use the C++ library provided rng. So setting the random seed for
the RubySystem class has no effect. Hence the variable and the corresponding
option are being dropped.
Currently the sequencer calls the function setMRU that updates the replacement
policy structures with the first level caches. While functionally this is
correct, the problem is that this requires calling findTagInSet() which is an
expensive function. This patch removes the calls to setMRU from the sequencer.
All controllers should now update the replacement policy on their own.
The set and the way index for a given cache entry can be found within the
AbstractCacheEntry structure. Use these indicies to update the replacement
policy structures.
Before this patch, while one could declare / define a function with default
argument values, but the actual function call would require one to specify
all the arguments. This patch changes the check for function arguments.
Now a function call needs to specify arguments that are at least as much as
those with default values and at most the total number of arguments taken
as input by the function.
Both FuncCallExprAST and MethodCallExprAST had code for checking the arguments
with which a function is being called. The patch does away with this
duplication. Now the code for checking function call arguments resides in the
Func class.
This is in preparation for adding a second arugment to the lookup
function for the CacheMemory class. The change to *.sm files was made using
the following sed command:
sed -i 's/\[\([0-9A-Za-z._()]*\)\]/.lookup(\1)/' src/mem/protocol/*.sm
The sequencer takes care of llsc accesses by calling upon functions
from the CacheMemory. This is unnecessary once the required CacheEntry object
is available. Thus some of the calls to findTagInSet() are avoided.
This patch eliminates the type Address defined by the ruby memory system.
This memory system would now use the type Addr that is in use by the
rest of the system.
Expose MessageBuffers from SLICC controllers as SimObjects that can be
manipulated in Python. This patch has numerous benefits:
1) First and foremost, it exposes MessageBuffers as SimObjects that can be
manipulated in Python code. This allows parameters to be set and checked in
Python code to avoid obfuscating parameters within protocol files. Further, now
as SimObjects, MessageBuffer parameters are printed to config output files as a
way to track parameters across simulations (e.g. buffer sizes)
2) Cleans up special-case code for responseFromMemory buffers, and aligns their
instantiation and use with mandatoryQueue buffers. These two special buffers
are the only MessageBuffers that are exposed to components outside of SLICC
controllers, and they're both slave ends of these buffers. They should be
exposed outside of SLICC in the same way, and this patch does it.
3) Distinguishes buffer-specific parameters from buffer-to-network parameters.
Specifically, buffer size, randomization, ordering, recycle latency, and ports
are all specific to a MessageBuffer, while the virtual network ID and type are
intrinsics of how the buffer is connected to network ports. The former are
specified in the Python object, while the latter are specified in the
controller *.sm files. Unlike buffer-specific parameters, which may need to
change depending on the simulated system structure, buffer-to-network
parameters can be specified statically for most or all different simulated
systems.
CacheMemory and DirectoryMemory lookup functions return pointers to entries
stored in the memory. Bring PerfectCacheMemory in line with this convention,
and clean up SLICC code generation that was in place solely to handle
references like that which was returned by PerfectCacheMemory::lookup.
The RubyCache (CacheMemory) latency parameter is only used for top-level caches
instantiated for Ruby coherence protocols. However, the top-level cache hit
latency is assessed by the Sequencer as accesses flow through to the cache
hierarchy. Further, protocol state machines should be enforcing these cache hit
latencies, but RubyCaches do not expose their latency to any existng state
machines through the SLICC/C++ interface. Thus, the RubyCache latency parameter
is superfluous for all caches. This is confusing for users.
As a step toward pushing L0/L1 cache hit latency into the top-level cache
controllers, move their latencies out of the RubyCache declarations and over to
their Sequencers. Eventually, these Sequencer parameters should be exposed as
parameters to the top-level cache controllers, which should assess the latency.
NOTE: Assessing these latencies in the cache controllers will require modifying
each to eliminate instantaneous Ruby hit callbacks in transitions that finish
accesses, which is likely a large undertaking.
The Packet::get() and Packet::set() methods both have very strange
semantics. Currently, they automatically convert between the guest
system's endianness and the host system's endianness. This behavior is
usually undesired and unexpected.
This patch introduces three new method pairs to access data:
* getLE() / setLE() - Get data stored as little endian.
* getBE() / setBE() - Get data stored as big endian.
* get(ByteOrder) / set(v, ByteOrder) - Configurable endianness
For example, a little endian device that is receiving a write request
will use teh getLE() method to get the data from the packet.
The old interface will be deprecated once all existing devices have
been ported to the new interface.
Timing generator for a pixel-based display. The timing generator is
intended for display processors driving a standard rasterized
display. The simplest possible display processor needs to derive from
this class and override the nextPixel() method to feed the display
with pixel data.
Pixels are ordered relative to the top left corner of the
display. Scan lines appear in the following order:
* Vertical Sync (starting at line 0)
* Vertical back porch
* Visible lines
* Vertical front porch
Pixel order within a scan line:
* Horizontal Sync
* Horizontal Back Porch
* Visible pixels
* Horizontal Front Porch
All events in the timing generator are automatically suspended on a
drain() request and restarted on drainResume(). This is conceptually
equivalent to clock gating when the pixel clock while the system is
draining. By gating the pixel clock, we prevent display controllers
from disturbing a memory system that is about to drain.
Add support for oscillators that can be programmed using the RealView
/ Versatile Express configuration interface. These oscillators are
typically used for things like the pixel clock in the display
controller.
The default configurations support the oscillators from a Versatile
Express motherboard (V2M-P1) with a CoreTile Express A15x2.
Add a simple DMA engine that sits behind a FIFO. This engine can be
used by devices that need to read large amounts of data (e.g., display
controllers). Most aspects of the controller, such as FIFO size,
maximum number of in-flight accesses, and maximum request sizes can be
configured.
The DMA copies blocks of data into its FIFO. Transfers are initiated
with a call to startFill() command that takes a start address and a
size. Advanced users can create a derived class that overrides the
onEndOfBlock() callback that is triggered when the last request to a
block has been issued. At this point, the DMA engine is ready to start
fetching a new block of data, potentially from a different address
range.
The DMA engine stops issuing new requests while it is draining. Care
must be taken to ensure that devices that are fed by a DMA engine are
suspended while the system is draining to avoid buffer underruns.
Split ClockedObject into two classes: Clocked that provides the basic
clock functionality, and ClockedObject that inherits from Clocked and
SimObject to provide the functionality of the old ClockedObject.
The CircleBuf class has at least one bug causing it to overwrite the
wrong elements when wrapping. The current code has a lot of unused
functionality and duplicated code. This changeset replaces the old
implementation with a new version that supports serialization and
arbitrary types in the buffer (not just char).
The i8042 device drops the contents of a PS2 device's buffer when
serializing, which results in corrupted PS2 state when continuing
simulation after a checkpoint. This changeset fixes this bug and
transitions the i8042 model to use the new serialization API that
requires the serialize() method to be const.
Declare the constructor and all of the operators that don't change the
state of a Cycles instance as constexpr. This makes it possible to use
Cycles as a static constant and allows the compiler to evaulate simple
expressions at compile time. An unfortunate side-effect of this is
that we cannot use assertions since C++11 doesn't support them in
constexpr functions. As a workaround, we throw an invalid_argument
exception when the assert would have triggered. A nice side-effect of
this is that the compiler will evaluate the "assertion" at compile
time when an expression involving Cycles can be statically evaluated.
This patch removes the extraneous flags and attributes from the
request and packet, and simply leaves the new commands. The change
introduced when adding acquire/release breaks all compatibility with
existing traces, and there is really no need for any new flags and
attributes. The commands should be sufficient.
This patch fixes packet tracing (urgent), and also removes the
unnecessary complexity.
It is sometimes desirable to be able to instantiate Drainable objects
when the simulator isn't in the Running state. Currently, we always
initialize Drainable objects to the Running state. However, this
confuses many of the sanity checks in the base class since objects
aren't expected to be in the Running state if the system is in the
Draining or Drained state.
Instead of always initializing the state variable in Drainable to
DrainState::Running, initialize it to the state the DrainManager is
in.
Note: This means an object can be created in the Draining/Drained
state without first calling drain().
This changeset moves the access trace functionality from the
CommMonitor into a separate probe. The probe can be hooked up to any
component that exports probe points of the type ProbePoints::Packet.
This patch moves the dependency on Google's Protocol Buffers library
from the CommMonitor to the MemTraceProbe, which means that the
CommMonitor (including stack distance profiling) no long depends on
it.
This changeset removes the stack distance calculator hooks from the
CommMonitor class and implements a stack distance calculator as a
memory system probe instead. The probe can be hooked up to any
component that exports probe points of the type ProbePoints::Packet.
This changeset adds a standardized probe point type to monitor packets
in the memory system and adds two probe points to the CommMonitor
class. These probe points enable monitoring of successfully delivered
requests and successfully delivered responses.
Memory system probe listeners should use the BaseMemProbe base class
to provide a unified configuration interface and reuse listener
registration code. Unlike the ProbeListenerObject class, the
BaseMemProbe allows objects to be wired to multiple ProbeManager
instances as long as they use the same probe point name.
There are 2 problems with the existing checkpoint and restore code in ruby.
The first is that when the event queue is altered by ruby during serialization,
some events that are currently scheduled cannot be found (e.g. the event to
stop simulation that always lives on the queue), causing a panic.
The second is that ruby is sometimes serialized after the memory system,
meaning that the dirty data in its cache is flushed back to memory too late
and so isn't included in the checkpoint.
These are fixed by implementing memory writeback in ruby, using the same
technique of hijacking the event queue, but first descheduling all events that
are currently on it. They are saved, along with their scheduled time, so that
the event queue can be faithfully reconstructed after writeback has finished.
Events with the AutoDelete flag set will delete themselves when they
are descheduled, causing an error when attempting to schedule them again.
This is fixed by simply not recording them when taking them off the queue.
Writeback is still implemented using flushing, so the cache recorder object,
that is created to generate the trace and manage flushing, is kept
around and used during serialization to write the trace to disk.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
1. Eliminate state NP in L0 and L1 Caches: The two states 'NP' and 'I' both
mean that the cache block is not present in the cache. 'I' also means that the
cache entry has been allocated. This causes problems when we do not correctly
initialize the cache entry when it is re-used. Hence, this patch eliminates
the state NP altogether. Everytime a new block comes into the cache, a cache
entry is allocated. Everytime a block leaves, the corresponding entry is
deallocated.
2. Separate transient state for instruction fetches: purely for accouting
purposes.
3. Drop state IS_I in L1 Cache and the message type STALE_DATA: when
invalidation is received for a block in IS, the block used to be moved to IS_I.
This meant that the data that would arrive in future would be used but not
stored since the controller lost the permissions after gaining them. This
state is being dropped and now invalidation messages would not processed till
the data has arrived. This also means that STALE_DATA type is not longer
required.
The level 2 controller has a bug. In one particular action, the data block was
copied from a message irrespective whether the block is dirty or not. In cases
when L1 sends no data, the data value copied was incorrect.
For many years the slicc symbol table has supported overloaded functions in
external classes. This patch extends that support to functions that are not
part of classes (a.k.a. no parent). For example, this support allows slicc
to understand that mapAddressToRange is overloaded and the NodeID is an
optional parameter.
This patch changes the router pipeline stages from 4 to 2. The
canonical 4-stage router is conservative while a lower-latency router
with look ahead routing and speculative allocation is well acknowledged.
Sets m_stage.second to the second parameter of the function.
Then, for every place where advance_stage is called, adds
a cycle to the argument being passed.
Adds features to allow protocols to reschedule controllers when conditionally
stalling within inport logic or actions. Also insures that resource and
protocol stalls are re-evaluated the next cycle.
This patch adds support that allows the replacement policy to identify each
cache block's access permission. This information can be useful when making
replacement decisions.
The Ruby banked array resource checks (initiated from SLICC) did a check and
allocate at the same time. If a transition needs more than one resource, then
it might check/allocate resource #1, then fail to get resource #2. Another
transition might then try to get the same resources, but in reverse order.
Deadlock.
This patch separates resource checking and resource reservation into two
steps to avoid deadlock.
It was previously possible for a stalled message to be reordered after an
incomming message. This patch ensures that any stalled message stays in its
original request order.
This patch adds a few helpful functions that allow .sm files to directly
invalidate all cache blocks using a trigger queue rather than rely on each
individual cache block to be invalidated via requests from the mandatory
queue.
This patch allows DPRINTFs to be used in SLICC state machines similar to how
they are used by the rest of gem5. Previously all DPRINTFs in the .sm files
had to use the RubySlicc flag.
This patch exposes the tag and data array latencies to the SLICC state machines
so that it can be used to determine the correct enqueue latency for response
messages.
To have multiple Entry types (e.g., a cache Entry type and
a directory Entry type), just declare one of them as a secondary
type by using the pair 'main="false"', e.g.:
structure(DirEntry, desc="...", interface="AbstractCacheEntry",
main="false") {
...and the primary type would be declared:
structure(Entry, desc="...", interface="AbstractCacheEntry") {
This patch fixes the type handling when prefix operations are used. Previously
prefix operators would assume a void return type, which made it impossible to
combine prefix operations with other expressions. This patch allows SLICC
programmers to use prefix operations more naturally.
This patches adds support for transitions of the form:
transition(START, EVENTS, *) { ACTIONS }
This allows a machine to collapse states that differ only in the next state
transition to collapse into one, and can help shorten/simplfy some protocols
significantly.
When * is encountered as an end state of a transition, the next state is
determined by calling the machine-specific getNextState function. The next
state is determined before any actions of the transition execute, and
therefore the next state calculation cannot depend on any of the transition
actions.
This patch allows SLICC protocols to use more than one message type with a
message buffer. For example, you can declare two in ports as such:
in_port(ResponseQueue_in, ResponseMsg, responseFromDir, rank=3) { ... }
in_port(tgtResponseQueue_in, TgtResponseMsg, responseFromDir, rank=2) { ... }
This patch was created by Bihn Pham during his internship at AMD.
There is no need to delay hit callback response messages by a cycle because
the response latency is already incurred in the Ruby protocol. This ensures
correct timing of memory instructions.
The Minor CPU currently doesn't drain properly when it is switched
out. This happens because Fetch 1 expects to be in the FetchHalted
state when it is drained. However, because the CPU is switched out, it
is stuck in the FetchWaitingForPC state. Fix this by ignoring drain
requests and returning DrainState::Drained from MinorCPU::drain() if
the CPU is switched out. This is always safe since a switched out CPU,
by definition, doesn't have any instructions in flight.
Minor currently activates thread 0 in startup() to work around an
issue where activateContext() is called from LiveProcess before the
process entry point is known. When activateContext() is called, Minor
creates a branch instruction to the process's entry point. The first
time it is called, the branch points to an undefined location (0). The
call in startup() updates the branch to point to the actual entry
point.
When instantiating a switched out Minor CPU, it still tries to
activate thread 0. This is clearly incorrect since a switched out CPU
can't have any active threads. This changeset adds a check to ensure
that the thread is active before reactivating it.
The drain refactor patches introduced a couple of bugs in the way
Minor handles draining. This patch fixes an incorrect assert and a
case of infinite recursion when the CPU signals drain done.
This patch removes the RequestCause, and also simplifies how we
schedule the sending of packets through the memory-side port. The
deassertion of bus requests is removed as it is not used.
This patch makes cache sets aware of the way number. This enables
some nice features such as the ablity to restrict way allocation. The
implemented mechanism allows to set a maximum way number to be
allocated 'k' which must fulfill 0 < k <= N (where N is the number of
ways). In the future more sophisticated mechasims can be implemented.
This patch changes how writebacks communicate whether the line is
passed as modified or owned. Previously we relied on the
isSupplyExclusive mechanism, which was originally designed to avoid
unecessary snoops.
For normal cache requests we use the sharedAsserted mechanism to
determine if a block should be marked writeable or not, and with this
patch we transition the writebacks to also use this
mechanism. Conceptually this is cleaner and more consistent.
Some minor fixes and removal of dead code. Changing the flags to be
enums rather than static const (to avoid any linking issues caused by
the latter). Also adding a getBlockAddr member which hopefully can
slowly finds its way into caches, snoop filters etc.
This adds a vector register type. The type is defined as a std::array of a
fixed number of uint64_ts. The isa_parser.py has been modified to parse vector
register operands and generate the required code. Different cpus have vector
register files now.
The Process class methods were using an improper style and this subsequently
bled into the system call code. The following regular expressions should be
helpful if someone transitions private system call patches on top of these
changesets:
s/alloc_fd/allocFD/
s/sim_fd(/simFD(/
s/sim_fd_obj/getFDEntry/
s/fix_file_offsets/fixFileOffsets/
s/find_file_offsets/findFileOffsets/
The patch clarifies whether file descriptors are host file descriptors or
target file descriptors in the system call code. (Host file descriptors
are file descriptors which have been allocated through real system calls
where target file descriptors are allocated from an array in the Process
class.)
This patch extends the previous patch's alterations around fd_map. It cleans
up some of the uglier code in the process file and replaces it with a more
concise C++11 version. As part of the changes, the FdMap class is pulled out
of the Process class and receives its own file.
This patch gets rid of unused Process::dup_fd method and does minor
refactoring in the process class files. The file descriptor max has been
changed to be the number of file descriptors since this clarifies the loop
boundary condition and cleans up the code a bit. The fd_map field has been
altered to be dynamically allocated as opposed to being an array; the
intention here is to build on this is subsequent patches to allow processes
to share their file descriptors with the clone system call.
This patch updates the x86 decoder so that it can decode instructions with vex
prefix. It also updates the isa with opcodes from vex opcode maps 1, 2 and 3.
Note that none of the instructions have been implemented yet. The
implementations would be provided in due course of time.
Multi gem5 is an extension to gem5 to enable parallel simulation of a
distributed system (e.g. simulation of a pool of machines
connected by Ethernet links). A multi gem5 run consists of seperate gem5
processes running in parallel (potentially on different hosts/slots on
a cluster). Each gem5 process executes the simulation of a component of the
simulated distributed system (e.g. a multi-core board with an Ethernet NIC).
The patch implements the "distributed" Ethernet link device
(dev/src/multi_etherlink.[hh.cc]). This device will send/receive
(simulated) Ethernet packets to/from peer gem5 processes. The interface
to talk to the peer gem5 processes is defined in dev/src/multi_iface.hh and
in tcp_iface.hh.
There is also a central message server process (util/multi/tcp_server.[hh,cc])
which acts like an Ethernet switch and transfers messages among the gem5 peers.
A multi gem5 simulations can be kicked off by the util/multi/gem5-multi.sh
wrapper script.
Checkpoints are supported by multi-gem5. The checkpoint must be
initiated by a single gem5 process. E.g., the gem5 process with rank 0
can take a checkpoint from the bootscript just before it invokes
'mpirun' to launch an MPI test. The message server process will notify
all the other peer gem5 processes and make them take a checkpoint, too
(after completing a global synchronisation to ensure that there are no
inflight messages among gem5).
This is another step in the process of removing global variables
from Ruby to enable multiple RubySystem instances in a single simulation.
The list of abstract controllers is per-RubySystem and should be
represented that way, rather than as a global.
Since this is the last remaining Ruby global variable, the
src/mem/ruby/Common/Global.* files are also removed.
This is another step in the process of removing global variables
from Ruby to enable multiple RubySystem instances in a single simulation.
With possibly multiple RubySystem objects, we can no longer use a global
variable to find "the" RubySystem object. Instead, each Ruby component
has to carry a pointer to the RubySystem object to which it belongs.
This patch begins the process of removing global variables from the Ruby
source with the goal of eventually allowing users to create multiple Ruby
instances in a single simulation. Currently, users cannot do so because
several global variables and static members are referenced by the RubySystem
object in a way that assumes that there will only ever be a single RubySystem.
These need to be replaced with per-RubySystem equivalents.
This specific patch replaces the global var g_ruby_start, which is used
to calculate throughput statistics for Throttles in simple networks and
links in Garnet networks, with a RubySystem instance var m_start_cycle.
Add a simple device shim that interfaces with the NoMali model
library. The gem5 side of the interface supports Mali T60x/T62x/T760
GPUs. This device model pretends to be a Mali GPU, but doesn't render
anything and executes in zero time.
The drain() call currently passes around a DrainManager pointer, which
is now completely pointless since there is only ever one global
DrainManager in the system. It also contains vestiges from the time
when SimObjects had to keep track of their child objects that needed
draining.
This changeset moves all of the DrainState handling to the Drainable
base class and changes the drain() and drainResume() calls to reflect
this. Particularly, the drain() call has been updated to take no
parameters (the DrainManager argument isn't needed) and return a
DrainState instead of an unsigned integer (there is no point returning
anything other than 0 or 1 any more). Drainable objects should return
either DrainState::Draining (equivalent to returning 1 in the old
system) if they need more time to drain or DrainState::Drained
(equivalent to returning 0 in the old system) if they are already in a
consistent state. Returning DrainState::Running is considered an
error.
Drain done signalling is now done through the signalDrainDone() method
in the Drainable class instead of using the DrainManager directly. The
new call checks if the state of the object is DrainState::Draining
before notifying the drain manager. This means that it is safe to call
signalDrainDone() without first checking if the simulator has
requested draining. The intention here is to reduce the code needed to
implement draining in simple objects.
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 memWriteback() and memInvalidate() calls used to live in the
Serializable interface. In this series of patches, the Serializable
interface will be redesigned to make serialization independent of the
object graph and always work on the entire simulator. This means that
the Serialization interface won't be useful to perform maintenance of
the caches in a sub-graph of the entire SimObject graph. This
changeset moves these memory maintenance methods to the SimObject
interface instead.
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.
When the Python helper code switches CPU models, it sometimes also
needs to change the memory mode of the simulator. When this happens,
it accidentally tried to drain the simulator despite having done so
already. This changeset removes the redundant drain.
Serialize pixels as unsigned 32 bit integers by adding the required
to_number() and stream operators. This is used by the FrameBuffer,
which now implements the Serializable interface. Users of frame
buffers are expected to serialize it into its own section by calling
serializeSection().
Events expected to be unserialized using an event-specific
unserializeEvent call. This call was never actually used, which meant
the events relying on it never got unserialized (or scheduled after
unserialization).
Instead of relying on a custom call, we now use the normal
serialization code again. In order to schedule the event correctly,
the parrent object is expected to use the
EventQueue::checkpointReschedule() call. This happens automatically
for events that are serialized using the AutoSerialize mechanism.
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.
All x87 misc registers are implemented in an array of 64 bit values
but in real hardware the size of some of these registers is smaller.
Previsouly all 64 bits where incorrectly set and then later read. To
ensure correctness we mask the value in setMiscRegNoEffect to write
only the valid bits.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
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.
All the object loaders directly examine the (already completely loaded
by object_file.cc) memory image. There is no current motivation to
keep the fd around.
This patch updates the compiler minimum requirement to gcc 4.7 and
clang 3.1, thus allowing:
1. Explicit virtual overrides (no need for M5_ATTR_OVERRIDE)
2. Non-static data member initializers
3. Template aliases
4. Delegating constructors
This patch also enables a transition from --std=c++0x to --std=c++11.
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.
There seems to have been a debug print left in when the original ARMv8
support was merged in. This printout is performed every time you
initialize a hardware thread, and it prints raw pointers, so it always
causes diffs in the regression. This patch removes the debug print.
The flush() method in CircleBuf resets the state of the circular
buffer, but fails to set size to zero. This obviously confuses code
that tries to determine the amount of data in the buffer. Set the size
to zero on flush.
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.
put O_DIRECT under ifdefs -- this fixes build for MacOSX.
Also use correct class for arm64 openFlagTable.
Committed by: Nilay Vaish <nilay@cs.wisc.edu>
This changeset adds support for aarch64 in kvm. The CPU module
supports both checkpointing and online CPU model switching as long as
no devices are simulated by the host kernel. It currently has the
following limitations:
* The system register based generic timer can only be simulated by
the host kernel. Workaround: Use a memory mapped timer instead to
simulate the timer in gem5.
* Simulating devices (e.g., the generic timer) in the host kernel
requires that the host kernel also simulates the GIC.
* ID registers in the host and in gem5 must match for switching
between simulated CPUs and KVM. This is particularly important
for ID registers describing memory system capabilities (e.g.,
ASID size, physical address size).
* Switching between a virtualized CPU and a simulated CPU is
currently not supported if in-kernel device emulation is
used. This could be worked around by adding support for switching
to the gem5 (e.g., the KvmGic) side of the device models. A
simpler workaround is to avoid in-kernel device models
altogether.
This changeset adds a GIC implementation that uses the kernel's
built-in support for simulating the interrupt controller. Since there
is currently no support for state transfer between gem5 and the
kernel, the device model does not support serialization and CPU
switching (which would require switching to a gem5-simulated GIC).
There are cases (particularly when attaching GDB) when instruction
events are scheduled at the current instruction tick. This used to
trigger an assertion error in kvm. This changeset adds a check for
this condition and forces KVM to do a quick entry that completes any
pending IO operations, but does not execute any new instructions,
before servicing the event. We could check if we need to enter KVM at
all, but forcing a quick entry is makes the code slightly cleaner and
does not hurt correctness (performance is hardly an issue in these
cases).
This changeset moves the ARM-specific KVM CPU implementation to
arch/arm/kvm/. This change is expected to keep the source tree
somewhat cleaner as we start adding support for ARMv8 and KVM
in-kernel interrupt controller simulation.
--HG--
rename : src/cpu/kvm/ArmKvmCPU.py => src/arch/arm/kvm/ArmKvmCPU.py
rename : src/cpu/kvm/arm_cpu.cc => src/arch/arm/kvm/arm_cpu.cc
rename : src/cpu/kvm/arm_cpu.hh => src/arch/arm/kvm/arm_cpu.hh
This patch changes how the address range calculates intersection such
that a system can have a number of non-overlapping interleaved ranges
without complaining. Without this patch we end up with a panic.
The MinorCPU would count bubbles in Execute::issue as part of
the num_insts_issued and so sometimes reach the instruction
issue limit incorrectly.
Fixed by checking for a bubble in one new place.
A step towards removing RubyMemoryControl and shift users to
DRAMCtrl. The latter is faster, more representative, very versatile,
and is integrated with power models.
There are cases when we don't want to use a system register mapped
generic timer, but can't use the SP804. For example, when using KVM on
aarch64, we want to intercept accesses to the generic timer, but can't
do so if it is using the system register interface. In such cases,
we need to use a memory-mapped generic timer.
This changeset adds a device model that implements the memory mapped
generic timer interface. The current implementation only supports a
single frame (i.e., one virtual timer and one physical timer).
The ArmSystem class has a parameter to indicate whether it is
configured to use the generic timer extension or not. This parameter
doesn't affect any feature flags in the current implementation and is
therefore completely unnecessary. In fact, we usually don't set it
even if a system has a generic timer. If we ever need to check if
there is a generic timer present, we should just request a pointer and
check if it is non-null instead.
The generic timer model currently does not support virtual
counters. Virtual and physical counters both tick with the same
frequency. However, virtual timers allow a hypervisor to set an offset
that is subtracted from the counter when it is read. This enables the
hypervisor to present a time base that ticks with virtual time in the
VM (i.e., doesn't tick when the VM isn't running). Modern Linux
kernels generally assume that virtual counters exist and try to use
them by default.
This changeset cleans up the generic timer a bit and moves most of the
register juggling from the ISA code into a separate class in the same
source file as the rest of the generic timer. It also removes the
assumption that there is always 8 or fewer CPUs in the system. Instead
of having a fixed limit, we now instantiate per-core timers as they
are requested. This is all in preparation for other patches that add
support for virtual timers and a memory mapped interface.
The register dumping code in kvm tries to print the bytes in large
registers (128 bits and larger) instead of printing them as hex. This
changeset fixes that.
Some versions of the kernel incorrectly swap the red and blue color
select registers. This changeset adds a workaround for that by
swapping them when instantiating a PixelConverter.
Currently, frame buffer handling in gem5 is quite ad hoc. In practice,
we pass around naked pointers to raw pixel data and expect consumers
to convert frame buffers using the (broken) VideoConverter.
This changeset completely redesigns the way we handle frame buffers
internally. In summary, it fixes several color conversion bugs, adds
support for more color formats (e.g., big endian), and makes the code
base easier to follow.
In the new world, gem5 always represents pixel data using the Pixel
struct when pixels need to be passed between different classes (e.g.,
a display controller and the VNC server). Producers of entire frames
(e.g., display controllers) should use the FrameBuffer class to
represent a frame.
Frame producers are expected to create one instance of the FrameBuffer
class in their constructors and register it with its consumers
once. Consumers are expected to check the dimensions of the frame
buffer when they consume it.
Conversion between the external representation and the internal
representation is supported for all common "true color" RGB formats of
up to 32-bit color depth. The external pixel representation is
expected to be between 1 and 4 bytes in either big endian or little
endian. Color channels are assumed to be contiguous ranges of bits
within each pixel word. The external pixel value is scaled to an 8-bit
internal representation using a floating multiplication to map it to
the entire 8-bit range.
The bitmap generation code is hard to follow and incorrectly uses the
size of an enum member to calculate the size of a pixel. This
changeset cleans up the code and adds some documentation.
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.
Three minor issues are resolved:
1. Apparently gcc 5.1 does not like negation of booleans followed by
bitwise AND.
2. Somehow the compiler also gets confused and warns about
NoopMachInst being unused (removing it causes compilation errors
though). Most likely a compiler bug.
3. There seems to be a number of instances where loop unrolling causes
false positives for the array-bounds check. For now, switch to
std::array. Potentially we could disable the warning for newer gcc
versions, but switching to std::array is probably a good move in
any case.
The system class currently clears the vector of active CPUs in
initState(). CPUs are added to the list by registerThreadContext()
which is called from BaseCPU::init(). This obviously breaks when the
System object is initialized after the CPUs. This changeset removes
the offending clear() call since the list will be empty after it has
been instantiated anyway.
The current ignoreWarnOnceFunc doesn't really work as expected,
since it will only generate one warning total, for whichever
"warn-once" syscall is invoked first. This patch fixes that
behavior by keeping a "warned" flag in the SyscallDesc object,
allowing suitably flagged syscalls to warn exactly once per
syscall.
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.
We currently assume that all uncacheable memory accesses are strictly
ordered. Instead of always enforcing strict ordering, we now only
enforce it if the required memory type is device memory or strongly
ordered memory.
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.
