Files
gem5/src/mem/multi_level_page_table_impl.hh
Gabe Black b8b13206c8 mem: Track TLB entries in the lookup cache as pointers.
Using the architectural page table on x86 and the functional page table
on ARM, both with the twolf benchmark in SE mode, there was no
performance penalty for doing so, and again possibly a performance
improvement. By using a pointer instead of an inline instance, it's
possible for the actual type of the TLB entry to be hidden somewhat,
taking a step towards abstracting away another aspect of the ISAs.

Since the TLB entries are no longer overwritten and now need to be
allocated and freed, this change introduces return types from the
updateCache and eraseCacheEntry functions. These functions will return
the pointer to any entry which has been displaced from the cache which
the caller can either free or ignore, depending on whether the entry
has a purpose outside of the cache.

Because the functional page table stores its entries over a longer time
period, it will generally not delete the pointer returned from those
functions. The "architechtural" page table, ie the one which is backed
by memory, doesn't have any other use for the TlbEntrys and will delete
them. That leads to more news and deletes than there used to be.

To address that, and also to speed up the architectural page table in
general, it would be a good idea to augment the functional page table
with an image of the table in memory, instead of replacing it with one.
The functional page table would provide quick lookups and also avoid
having to translate page table entries to TLB entries, making
performance essentially equivalent to the functional case. The backing
page tables, which are primarily for consumption by the physical
hardware when in KVM, can be updated when mappings change but otherwise
left alone.

If we end up doing that, we could just let the ISA specific process
classes enable whatever additional TLB machinery they need, likely
a backing copy in memory, without any knowledge or involvement from
the ISA agnostic class. We would be able to get rid of the useArchPT
setting and the bits of code in the configs which set it.

Change-Id: I2e21945cd852bb1b3d0740fe6a4c5acbfd9548c5
Reviewed-on: https://gem5-review.googlesource.com/6983
Maintainer: Gabe Black <gabeblack@google.com>
Reviewed-by: Brandon Potter <Brandon.Potter@amd.com>
Reviewed-by: Jason Lowe-Power <jason@lowepower.com>
Reviewed-by: Anthony Gutierrez <anthony.gutierrez@amd.com>
2018-01-15 22:36:41 +00:00

331 lines
11 KiB
C++

/*
* Copyright (c) 2014 Advanced Micro Devices, Inc.
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are
* met: redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer;
* redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution;
* neither the name of the copyright holders nor the names of its
* contributors may be used to endorse or promote products derived from
* this software without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
* "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
* OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* Authors: Alexandru Dutu
*/
/**
* @file
* Definitions of page table
*/
#include <string>
#include "arch/isa_traits.hh"
#include "base/trace.hh"
#include "config/the_isa.hh"
#include "debug/MMU.hh"
#include "mem/multi_level_page_table.hh"
#include "mem/page_table.hh"
using namespace std;
using namespace TheISA;
template <class ISAOps>
MultiLevelPageTable<ISAOps>::MultiLevelPageTable(const std::string &__name,
uint64_t _pid, System *_sys,
Addr pageSize)
: PageTableBase(__name, _pid, pageSize), system(_sys),
logLevelSize(PageTableLayout),
numLevels(logLevelSize.size())
{
}
template <class ISAOps>
MultiLevelPageTable<ISAOps>::~MultiLevelPageTable()
{
}
template <class ISAOps>
void
MultiLevelPageTable<ISAOps>::initState(ThreadContext* tc)
{
basePtr = pTableISAOps.getBasePtr(tc);
if (basePtr == 0) basePtr++;
DPRINTF(MMU, "basePtr: %d\n", basePtr);
system->pagePtr = basePtr;
/* setting first level of the page table */
uint64_t log_req_size = floorLog2(sizeof(PageTableEntry)) +
logLevelSize[numLevels-1];
assert(log_req_size >= PageShift);
uint64_t npages = 1 << (log_req_size - PageShift);
Addr paddr = system->allocPhysPages(npages);
PortProxy &p = system->physProxy;
p.memsetBlob(paddr, 0, npages << PageShift);
}
template <class ISAOps>
bool
MultiLevelPageTable<ISAOps>::walk(Addr vaddr, bool allocate, Addr &PTE_addr)
{
std::vector<uint64_t> offsets = pTableISAOps.getOffsets(vaddr);
Addr level_base = basePtr;
for (int i = numLevels - 1; i > 0; i--) {
Addr entry_addr = (level_base<<PageShift) +
offsets[i] * sizeof(PageTableEntry);
PortProxy &p = system->physProxy;
PageTableEntry entry = p.read<PageTableEntry>(entry_addr);
Addr next_entry_pnum = pTableISAOps.getPnum(entry);
if (next_entry_pnum == 0) {
if (!allocate) return false;
uint64_t log_req_size = floorLog2(sizeof(PageTableEntry)) +
logLevelSize[i-1];
assert(log_req_size >= PageShift);
uint64_t npages = 1 << (log_req_size - PageShift);
DPRINTF(MMU, "Allocating %d pages needed for entry in level %d\n",
npages, i - 1);
/* allocate new entry */
Addr next_entry_paddr = system->allocPhysPages(npages);
p.memsetBlob(next_entry_paddr, 0, npages << PageShift);
next_entry_pnum = next_entry_paddr >> PageShift;
pTableISAOps.setPnum(entry, next_entry_pnum);
pTableISAOps.setPTEFields(entry);
p.write<PageTableEntry>(entry_addr, entry);
}
DPRINTF(MMU, "Level %d base: %d offset: %d entry: %d\n",
i, level_base, offsets[i], next_entry_pnum);
level_base = next_entry_pnum;
}
PTE_addr = (level_base<<PageShift) +
offsets[0] * sizeof(PageTableEntry);
DPRINTF(MMU, "Returning PTE_addr: %x\n", PTE_addr);
return true;
}
template <class ISAOps>
void
MultiLevelPageTable<ISAOps>::map(Addr vaddr, Addr paddr,
int64_t size, uint64_t flags)
{
bool clobber = flags & Clobber;
// starting address must be page aligned
assert(pageOffset(vaddr) == 0);
DPRINTF(MMU, "Allocating Page: %#x-%#x\n", vaddr, vaddr + size);
PortProxy &p = system->physProxy;
for (; size > 0; size -= pageSize, vaddr += pageSize, paddr += pageSize) {
Addr PTE_addr;
if (walk(vaddr, true, PTE_addr)) {
PageTableEntry PTE = p.read<PageTableEntry>(PTE_addr);
Addr entry_paddr = pTableISAOps.getPnum(PTE);
if (!clobber && entry_paddr != 0) {
fatal("addr 0x%x already mapped to %x", vaddr, entry_paddr);
}
pTableISAOps.setPnum(PTE, paddr >> PageShift);
uint64_t PTE_flags = 0;
if (flags & NotPresent)
PTE_flags |= TheISA::PTE_NotPresent;
if (flags & Uncacheable)
PTE_flags |= TheISA::PTE_Uncacheable;
if (flags & ReadOnly)
PTE_flags |= TheISA::PTE_ReadOnly;
pTableISAOps.setPTEFields(PTE, PTE_flags);
p.write<PageTableEntry>(PTE_addr, PTE);
DPRINTF(MMU, "New mapping: %#x-%#x\n", vaddr, paddr);
delete eraseCacheEntry(vaddr);
delete updateCache(vaddr, new TlbEntry(pid, vaddr, paddr,
flags & Uncacheable,
flags & ReadOnly));
}
}
}
template <class ISAOps>
void
MultiLevelPageTable<ISAOps>::remap(Addr vaddr, int64_t size, Addr new_vaddr)
{
assert(pageOffset(vaddr) == 0);
assert(pageOffset(new_vaddr) == 0);
DPRINTF(MMU, "moving pages from vaddr %08p to %08p, size = %d\n", vaddr,
new_vaddr, size);
PortProxy &p = system->physProxy;
for (; size > 0;
size -= pageSize, vaddr += pageSize, new_vaddr += pageSize)
{
Addr PTE_addr;
if (walk(vaddr, false, PTE_addr)) {
PageTableEntry PTE = p.read<PageTableEntry>(PTE_addr);
Addr paddr = pTableISAOps.getPnum(PTE);
if (paddr == 0) {
fatal("Page fault while remapping");
} else {
/* unmapping vaddr */
pTableISAOps.setPnum(PTE, 0);
p.write<PageTableEntry>(PTE_addr, PTE);
/* maping new_vaddr */
Addr new_PTE_addr;
walk(new_vaddr, true, new_PTE_addr);
PageTableEntry new_PTE = p.read<PageTableEntry>(new_PTE_addr);
pTableISAOps.setPnum(new_PTE, paddr>>PageShift);
pTableISAOps.setPTEFields(new_PTE);
p.write<PageTableEntry>(new_PTE_addr, new_PTE);
DPRINTF(MMU, "Remapping: %#x-%#x\n", vaddr, new_PTE_addr);
}
delete eraseCacheEntry(vaddr);
delete updateCache(new_vaddr,
new TlbEntry(pid, new_vaddr, paddr,
pTableISAOps.isUncacheable(PTE),
pTableISAOps.isReadOnly(PTE)));
} else {
fatal("Page fault while remapping");
}
}
}
template <class ISAOps>
void
MultiLevelPageTable<ISAOps>::unmap(Addr vaddr, int64_t size)
{
assert(pageOffset(vaddr) == 0);
DPRINTF(MMU, "Unmapping page: %#x-%#x\n", vaddr, vaddr+ size);
PortProxy &p = system->physProxy;
for (; size > 0; size -= pageSize, vaddr += pageSize) {
Addr PTE_addr;
if (walk(vaddr, false, PTE_addr)) {
PageTableEntry PTE = p.read<PageTableEntry>(PTE_addr);
Addr paddr = pTableISAOps.getPnum(PTE);
if (paddr == 0) {
fatal("PageTable::allocate: address 0x%x not mapped", vaddr);
} else {
pTableISAOps.setPnum(PTE, 0);
p.write<PageTableEntry>(PTE_addr, PTE);
DPRINTF(MMU, "Unmapping: %#x\n", vaddr);
}
delete eraseCacheEntry(vaddr);
} else {
fatal("Page fault while unmapping");
}
}
}
template <class ISAOps>
bool
MultiLevelPageTable<ISAOps>::isUnmapped(Addr vaddr, int64_t size)
{
// starting address must be page aligned
assert(pageOffset(vaddr) == 0);
PortProxy &p = system->physProxy;
for (; size > 0; size -= pageSize, vaddr += pageSize) {
Addr PTE_addr;
if (walk(vaddr, false, PTE_addr)) {
PageTableEntry PTE = p.read<PageTableEntry>(PTE_addr);
if (pTableISAOps.getPnum(PTE) != 0)
return false;
}
}
return true;
}
template <class ISAOps>
bool
MultiLevelPageTable<ISAOps>::lookup(Addr vaddr, TlbEntry &entry)
{
Addr page_addr = pageAlign(vaddr);
if (pTableCache[0].entry && pTableCache[0].vaddr == page_addr) {
entry = *pTableCache[0].entry;
return true;
}
if (pTableCache[1].entry && pTableCache[1].vaddr == page_addr) {
entry = *pTableCache[1].entry;
return true;
}
if (pTableCache[2].entry && pTableCache[2].vaddr == page_addr) {
entry = *pTableCache[2].entry;
return true;
}
DPRINTF(MMU, "lookup page_addr: %#x\n", page_addr);
Addr PTE_addr;
if (walk(page_addr, false, PTE_addr)) {
PortProxy &p = system->physProxy;
PageTableEntry PTE = p.read<PageTableEntry>(PTE_addr);
Addr pnum = pTableISAOps.getPnum(PTE);
if (pnum == 0)
return false;
TlbEntry *new_entry = new TlbEntry(pid, vaddr, pnum << PageShift,
pTableISAOps.isUncacheable(PTE),
pTableISAOps.isReadOnly(PTE));
entry = *new_entry;
delete updateCache(page_addr, new_entry);
} else {
return false;
}
return true;
}
template <class ISAOps>
void
MultiLevelPageTable<ISAOps>::serialize(CheckpointOut &cp) const
{
/** Since, the page table is stored in system memory
* which is serialized separately, we will serialize
* just the base pointer
*/
paramOut(cp, "ptable.pointer", basePtr);
}
template <class ISAOps>
void
MultiLevelPageTable<ISAOps>::unserialize(CheckpointIn &cp)
{
paramIn(cp, "ptable.pointer", basePtr);
}