Merge pull request #9415 from liamwhite/dc

memory: correct semantics of data cache management operations
This commit is contained in:
Mai 2022-12-11 21:09:31 +00:00 committed by GitHub
commit d5684dbe7d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 15 additions and 102 deletions

View file

@ -34,8 +34,6 @@ add_library(common STATIC
bit_util.h
cityhash.cpp
cityhash.h
cache_management.cpp
cache_management.h
common_funcs.h
common_precompiled_headers.h
common_types.h

View file

@ -1,59 +0,0 @@
// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <cstdint>
#include <cstring>
#include "common/cache_management.h"
namespace Common {
#if defined(ARCHITECTURE_x86_64)
// Most cache operations are no-ops on x86
void DataCacheLineCleanByVAToPoU(void* start, size_t size) {}
void DataCacheLineCleanAndInvalidateByVAToPoC(void* start, size_t size) {}
void DataCacheLineCleanByVAToPoC(void* start, size_t size) {}
void DataCacheZeroByVA(void* start, size_t size) {
std::memset(start, 0, size);
}
#elif defined(ARCHITECTURE_arm64)
// BS/DminLine is log2(cache size in words), we want size in bytes
#define EXTRACT_DMINLINE(ctr_el0) (1 << ((((ctr_el0) >> 16) & 0xf) + 2))
#define EXTRACT_BS(dczid_el0) (1 << (((dczid_el0)&0xf) + 2))
#define DEFINE_DC_OP(op_name, function_name) \
void function_name(void* start, size_t size) { \
size_t ctr_el0; \
asm volatile("mrs %[ctr_el0], ctr_el0\n\t" : [ctr_el0] "=r"(ctr_el0)); \
size_t cacheline_size = EXTRACT_DMINLINE(ctr_el0); \
uintptr_t va_start = reinterpret_cast<uintptr_t>(start); \
uintptr_t va_end = va_start + size; \
for (uintptr_t va = va_start; va < va_end; va += cacheline_size) { \
asm volatile("dc " #op_name ", %[va]\n\t" : : [va] "r"(va) : "memory"); \
} \
}
#define DEFINE_DC_OP_DCZID(op_name, function_name) \
void function_name(void* start, size_t size) { \
size_t dczid_el0; \
asm volatile("mrs %[dczid_el0], dczid_el0\n\t" : [dczid_el0] "=r"(dczid_el0)); \
size_t cacheline_size = EXTRACT_BS(dczid_el0); \
uintptr_t va_start = reinterpret_cast<uintptr_t>(start); \
uintptr_t va_end = va_start + size; \
for (uintptr_t va = va_start; va < va_end; va += cacheline_size) { \
asm volatile("dc " #op_name ", %[va]\n\t" : : [va] "r"(va) : "memory"); \
} \
}
DEFINE_DC_OP(cvau, DataCacheLineCleanByVAToPoU);
DEFINE_DC_OP(civac, DataCacheLineCleanAndInvalidateByVAToPoC);
DEFINE_DC_OP(cvac, DataCacheLineCleanByVAToPoC);
DEFINE_DC_OP_DCZID(zva, DataCacheZeroByVA);
#endif
} // namespace Common

View file

@ -1,27 +0,0 @@
// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#pragma once
#include <cstddef>
namespace Common {
// Data cache instructions enabled at EL0 by SCTLR_EL1.UCI.
// VA = virtual address
// PoC = point of coherency
// PoU = point of unification
// dc cvau
void DataCacheLineCleanByVAToPoU(void* start, size_t size);
// dc civac
void DataCacheLineCleanAndInvalidateByVAToPoC(void* start, size_t size);
// dc cvac
void DataCacheLineCleanByVAToPoC(void* start, size_t size);
// dc zva
void DataCacheZeroByVA(void* start, size_t size);
} // namespace Common

View file

@ -6,7 +6,6 @@
#include "common/assert.h"
#include "common/atomic_ops.h"
#include "common/cache_management.h"
#include "common/common_types.h"
#include "common/logging/log.h"
#include "common/page_table.h"
@ -340,10 +339,9 @@ struct Memory::Impl {
LOG_ERROR(HW_Memory, "Unmapped cache maintenance @ {:#018X}", current_vaddr);
throw InvalidMemoryException();
},
[&](const std::size_t block_size, u8* const host_ptr) { cb(block_size, host_ptr); },
[&](const std::size_t block_size, u8* const host_ptr) {},
[&](const VAddr current_vaddr, const std::size_t block_size, u8* const host_ptr) {
system.GPU().FlushRegion(current_vaddr, block_size);
cb(block_size, host_ptr);
cb(current_vaddr, block_size);
},
[](const std::size_t block_size) {});
} catch (InvalidMemoryException&) {
@ -354,27 +352,30 @@ struct Memory::Impl {
}
Result InvalidateDataCache(const Kernel::KProcess& process, VAddr dest_addr, std::size_t size) {
auto perform = [&](const std::size_t block_size, u8* const host_ptr) {
// Do nothing; this operation (dc ivac) cannot be supported
// from EL0
auto on_rasterizer = [&](const VAddr current_vaddr, const std::size_t block_size) {
// dc ivac: Invalidate to point of coherency
// GPU flush -> CPU invalidate
system.GPU().FlushRegion(current_vaddr, block_size);
};
return PerformCacheOperation(process, dest_addr, size, perform);
return PerformCacheOperation(process, dest_addr, size, on_rasterizer);
}
Result StoreDataCache(const Kernel::KProcess& process, VAddr dest_addr, std::size_t size) {
auto perform = [&](const std::size_t block_size, u8* const host_ptr) {
auto on_rasterizer = [&](const VAddr current_vaddr, const std::size_t block_size) {
// dc cvac: Store to point of coherency
Common::DataCacheLineCleanByVAToPoC(host_ptr, block_size);
// CPU flush -> GPU invalidate
system.GPU().InvalidateRegion(current_vaddr, block_size);
};
return PerformCacheOperation(process, dest_addr, size, perform);
return PerformCacheOperation(process, dest_addr, size, on_rasterizer);
}
Result FlushDataCache(const Kernel::KProcess& process, VAddr dest_addr, std::size_t size) {
auto perform = [&](const std::size_t block_size, u8* const host_ptr) {
auto on_rasterizer = [&](const VAddr current_vaddr, const std::size_t block_size) {
// dc civac: Store to point of coherency, and invalidate from cache
Common::DataCacheLineCleanAndInvalidateByVAToPoC(host_ptr, block_size);
// CPU flush -> GPU invalidate
system.GPU().InvalidateRegion(current_vaddr, block_size);
};
return PerformCacheOperation(process, dest_addr, size, perform);
return PerformCacheOperation(process, dest_addr, size, on_rasterizer);
}
void MarkRegionDebug(VAddr vaddr, u64 size, bool debug) {