arch-riscv: reduced lr/sc implementation
In gem5::RiscvISA::ISA, handleLocked* functions maintain an address stack(i.e. locked_addrs) to check whether each SC matches the most recent LR. However, there are some problems with this implementation. First, the elements in the stack may only be popped when the handleLockedSnoop function is invoked. In other cases, the elements in the stack will not be popped even if the SC and LR match. This makes the `locked_addrs` get bigger and bigger as gem5 runs. Second, LR/SC does not always match. For example, in Linux's __cmpxchg[1], after executing LR, if the value read is not equal to the old value, the subsequent SC is skipped. For gem5's current implementation, this would cause the address to be pushed into `locked_addrs` every time __cmpxchg is failed. But these addresses are never popped. This also makes the `locked_addrs` get bigger and bigger. Third, existing emulator implementations (spike, qemu) do not use the stack, but only record the last address accessed by LR. Afterward, when executing SC, these implementations determine whether the address accessed by SC is the same as the one recorded. This patch modifies gem5's handleLocked* function by referring to other existing RISC-V implementations. It eliminates `locked_addrs` and simplifies the related code. Thus, it fixes the "memory leak"-like error that can occur on `locked_addrs` when executing LR/SC. Related links: [1] Linux's cmpxchg implementation for RISC-V: + https://github.com/torvalds/linux/blob/master/arch/riscv/include/asm/cmpxchg.h [2] spike lr/sc implementation: + https://github.com/riscv-software-src/riscv-isa-sim/blob/master/riscv/insns/sc_d.h + https://github.com/riscv-software-src/riscv-isa-sim/blob/master/riscv/insns/lr_d.h + https://github.com/riscv-software-src/riscv-isa-sim/blob/master/riscv/mmu.h [3] rocket lr/sc implementation: + https://github.com/chipsalliance/rocket-chip/blob/master/src/main/scala/rocket/NBDcache.scala [4] QEMU lr/sc implementation: + https://gitlab.com/qemu-project/qemu/-/blob/master/target/riscv/insn_trans/trans_rva.c.inc Change-Id: Ic79444cace62e39b7fe9e01f665cb13e4d990d0a Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/55663 Reviewed-by: Bobby Bruce <bbruce@ucdavis.edu> Reviewed-by: Jason Lowe-Power <power.jg@gmail.com> Maintainer: Jason Lowe-Power <power.jg@gmail.com> Tested-by: kokoro <noreply+kokoro@google.com>
This commit is contained in:
@@ -504,30 +504,29 @@ ISA::unserialize(CheckpointIn &cp)
|
||||
|
||||
const int WARN_FAILURE = 10000;
|
||||
|
||||
// RISC-V allows multiple locks per hart, but each SC has to unlock the most
|
||||
// recent one, so we use a stack here.
|
||||
std::unordered_map<int, std::stack<Addr>> locked_addrs;
|
||||
const Addr INVALID_RESERVATION_ADDR = (Addr) -1;
|
||||
std::unordered_map<int, Addr> load_reservation_addrs;
|
||||
|
||||
void
|
||||
ISA::handleLockedSnoop(PacketPtr pkt, Addr cacheBlockMask)
|
||||
{
|
||||
std::stack<Addr>& locked_addr_stack = locked_addrs[tc->contextId()];
|
||||
Addr& load_reservation_addr = load_reservation_addrs[tc->contextId()];
|
||||
|
||||
if (locked_addr_stack.empty())
|
||||
if (load_reservation_addr == INVALID_RESERVATION_ADDR)
|
||||
return;
|
||||
Addr snoop_addr = pkt->getAddr() & cacheBlockMask;
|
||||
DPRINTF(LLSC, "Locked snoop on address %x.\n", snoop_addr);
|
||||
if ((locked_addr_stack.top() & cacheBlockMask) == snoop_addr)
|
||||
locked_addr_stack.pop();
|
||||
if ((load_reservation_addr & cacheBlockMask) == snoop_addr)
|
||||
load_reservation_addr = INVALID_RESERVATION_ADDR;
|
||||
}
|
||||
|
||||
|
||||
void
|
||||
ISA::handleLockedRead(const RequestPtr &req)
|
||||
{
|
||||
std::stack<Addr>& locked_addr_stack = locked_addrs[tc->contextId()];
|
||||
Addr& load_reservation_addr = load_reservation_addrs[tc->contextId()];
|
||||
|
||||
locked_addr_stack.push(req->getPaddr() & ~0xF);
|
||||
load_reservation_addr = req->getPaddr() & ~0xF;
|
||||
DPRINTF(LLSC, "[cid:%d]: Reserved address %x.\n",
|
||||
req->contextId(), req->getPaddr() & ~0xF);
|
||||
}
|
||||
@@ -535,23 +534,25 @@ ISA::handleLockedRead(const RequestPtr &req)
|
||||
bool
|
||||
ISA::handleLockedWrite(const RequestPtr &req, Addr cacheBlockMask)
|
||||
{
|
||||
std::stack<Addr>& locked_addr_stack = locked_addrs[tc->contextId()];
|
||||
Addr& load_reservation_addr = load_reservation_addrs[tc->contextId()];
|
||||
bool lr_addr_empty = (load_reservation_addr == INVALID_RESERVATION_ADDR);
|
||||
|
||||
// Normally RISC-V uses zero to indicate success and nonzero to indicate
|
||||
// failure (right now only 1 is reserved), but in gem5 zero indicates
|
||||
// failure and one indicates success, so here we conform to that (it should
|
||||
// be switched in the instruction's implementation)
|
||||
|
||||
DPRINTF(LLSC, "[cid:%d]: locked_addrs empty? %s.\n", req->contextId(),
|
||||
locked_addr_stack.empty() ? "yes" : "no");
|
||||
if (!locked_addr_stack.empty()) {
|
||||
DPRINTF(LLSC, "[cid:%d]: load_reservation_addrs empty? %s.\n",
|
||||
req->contextId(),
|
||||
lr_addr_empty ? "yes" : "no");
|
||||
if (!lr_addr_empty) {
|
||||
DPRINTF(LLSC, "[cid:%d]: addr = %x.\n", req->contextId(),
|
||||
req->getPaddr() & ~0xF);
|
||||
DPRINTF(LLSC, "[cid:%d]: last locked addr = %x.\n", req->contextId(),
|
||||
locked_addr_stack.top());
|
||||
load_reservation_addr);
|
||||
}
|
||||
if (locked_addr_stack.empty()
|
||||
|| locked_addr_stack.top() != ((req->getPaddr() & ~0xF))) {
|
||||
if (lr_addr_empty
|
||||
|| load_reservation_addr != ((req->getPaddr() & ~0xF))) {
|
||||
req->setExtraData(0);
|
||||
int stCondFailures = tc->readStCondFailures();
|
||||
tc->setStCondFailures(++stCondFailures);
|
||||
@@ -564,6 +565,7 @@ ISA::handleLockedWrite(const RequestPtr &req, Addr cacheBlockMask)
|
||||
if (req->isUncacheable()) {
|
||||
req->setExtraData(2);
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user