This update includes the changes to whole-line writes, the refinement
of Read to ReadClean and ReadShared, the introduction of CleanEvict
for snoop-filter tracking, and updates to the DRAM command scheduler
for bank-group-aware scheduling.
Needless to say, almost every regression is affected.
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.
