add MemMMap class
authorJacob Lifshay <programmerjake@gmail.com>
Sat, 23 Sep 2023 01:23:22 +0000 (18:23 -0700)
committerLuke Kenneth Casson Leighton <lkcl@lkcl.net>
Tue, 26 Sep 2023 21:08:32 +0000 (22:08 +0100)
https://bugs.libre-soc.org/show_bug.cgi?id=1173

src/openpower/decoder/isa/caller.py
src/openpower/decoder/isa/mem.py
src/openpower/test/runner.py

index c460742941d4b86abc96c82e7dba0def0a7ef956..eaf0af47e6ffe97ac32d158a03c97ae3aa1c9fe0 100644 (file)
@@ -22,7 +22,7 @@ from openpower.consts import (MSRb, PIb,  # big-endian (PowerISA versions)
                               SVP64CROffs, SVP64MODEb)
 from openpower.decoder.helpers import (ISACallerHelper, ISAFPHelpers, exts,
                                        gtu, undefined)
-from openpower.decoder.isa.mem import Mem, MemException
+from openpower.decoder.isa.mem import Mem, MemMMap, MemException
 from openpower.decoder.isa.radixmmu import RADIX
 from openpower.decoder.isa.svshape import SVSHAPE
 from openpower.decoder.isa.svstate import SVP64State
@@ -1148,7 +1148,8 @@ class ISACaller(ISACallerHelper, ISAFPHelpers, StepLoop):
                  mmu=False,
                  icachemmu=False,
                  initial_fpscr=0,
-                 insnlog=None):
+                 insnlog=None,
+                 use_mmap_mem=False):
 
         # trace log file for model output. if None do nothing
         self.insnlog = insnlog
@@ -1214,9 +1215,18 @@ class ISACaller(ISACallerHelper, ISAFPHelpers, StepLoop):
         self.last_op_svshape = False
 
         # "raw" memory
-        self.mem = Mem(row_bytes=8, initial_mem=initial_mem, misaligned_ok=True)
-        self.mem.log_fancy(kind=LogKind.InstrInOuts)
-        self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
+        if use_mmap_mem:
+            self.mem = MemMMap(row_bytes=8,
+                               initial_mem=initial_mem,
+                               misaligned_ok=True)
+            self.imem = self.mem
+            self.mem.initialize(row_bytes=4, initial_mem=initial_insns)
+            self.mem.log_fancy(kind=LogKind.InstrInOuts)
+        else:
+            self.mem = Mem(row_bytes=8, initial_mem=initial_mem,
+                           misaligned_ok=True)
+            self.mem.log_fancy(kind=LogKind.InstrInOuts)
+            self.imem = Mem(row_bytes=4, initial_mem=initial_insns)
         # MMU mode, redirect underlying Mem through RADIX
         if mmu:
             self.mem = RADIX(self.mem, self)
index 0f4639b6b0331262332abfdbf9a6bf7d208da601..fccd9c1b6c855f20464d826749cf1c050e0b3c39 100644 (file)
@@ -18,6 +18,9 @@ from openpower.util import log, LogKind
 import math
 import enum
 from cached_property import cached_property
+import mmap
+from pickle import PicklingError
+import ctypes
 
 
 def swap_order(x, nbytes):
@@ -65,6 +68,12 @@ class _ReadReason(enum.Enum):
             return 0
         return None
 
+    @cached_property
+    def needed_mmap_page_flag(self):
+        if self is self.Execute:
+            return _MMapPageFlags.X
+        return _MMapPageFlags.R
+
 
 class MemCommon:
     def __init__(self, row_bytes, initial_mem, misaligned_ok):
@@ -77,6 +86,9 @@ class MemCommon:
         if not initial_mem:
             return
 
+        self.initialize(row_bytes, initial_mem)
+
+    def initialize(self, row_bytes, initial_mem):
         for addr, (val, width) in process_mem(initial_mem, row_bytes).items():
             # val = swap_order(val, width)
             self.st(addr, val, width, swap=False)
@@ -270,3 +282,129 @@ class Mem(MemCommon):
 
     def word_idxs(self):
         return self.mem.keys()
+
+
+class _MMapPageFlags(enum.IntFlag):
+    """ flags on each mmap-ped page
+
+    Note: these are *not* PowerISA MMU pages, but instead internal to Mem so
+    it can detect invalid accesses and assert rather than segfaulting.
+    """
+    R = 1
+    W = 2
+    X = 4
+    "readable when instr_fetch=True"
+
+    RWX = R | W | X
+
+
+_MMAP_PAGE_SIZE = 1 << 16  # size of chunk that we track
+_PAGE_COUNT = (1 << 48) // _MMAP_PAGE_SIZE  # 48-bit address space
+_NEG_PG_IDX_START = _PAGE_COUNT // 2  # start of negative half of address space
+BLOCK_SIZE = 1 << 32  # code assumes it's a power of two
+assert BLOCK_SIZE % _MMAP_PAGE_SIZE == 0
+DEFAULT_BLOCK_ADDRS = (
+    0,  # low end of user space
+    2 ** 47 - BLOCK_SIZE,  # high end of user space
+)
+
+
+class MemMMap(MemCommon):
+    def __init__(self, row_bytes=8, initial_mem=None, misaligned_ok=False,
+                 block_addrs=DEFAULT_BLOCK_ADDRS, emulating_mmap=False):
+        # we can't allocate the entire 2 ** 47 byte address space, so split
+        # it into smaller blocks
+        self.mem_blocks = {
+            addr: mmap.mmap(-1, BLOCK_SIZE) for addr in block_addrs}
+        assert all(addr % BLOCK_SIZE == 0 for addr in self.mem_blocks), \
+            "misaligned block address not supported"
+        self.page_flags = {}
+        self.modified_pages = set()
+        if not emulating_mmap:
+            # mark blocks as readable/writable
+            for addr, block in self.mem_blocks.items():
+                start_page_idx = addr // _MMAP_PAGE_SIZE
+                end_page_idx = start_page_idx + len(block) // _MMAP_PAGE_SIZE
+                for page_idx in range(start_page_idx, end_page_idx):
+                    self.page_flags[page_idx] = _MMapPageFlags.RWX
+
+        super().__init__(row_bytes, initial_mem, misaligned_ok)
+
+    def mmap_page_idx_to_addr(self, page_idx):
+        assert 0 <= page_idx < _PAGE_COUNT
+        if page_idx >= _NEG_PG_IDX_START:
+            page_idx -= _PAGE_COUNT
+        return (page_idx * _MMAP_PAGE_SIZE) % 2 ** 64
+
+    def addr_to_mmap_page_idx(self, addr):
+        page_idx, offset = divmod(addr, _MMAP_PAGE_SIZE)
+        page_idx %= _PAGE_COUNT
+        expected = self.mmap_page_idx_to_addr(page_idx) + offset
+        if addr != expected:
+            exc = MemException("not sign extended",
+                               f"address not sign extended: 0x{addr:X} "
+                               f"expected 0x{expected:X}")
+            exc.dar = addr
+            raise exc
+        return page_idx
+
+    def __reduce_ex__(self, protocol):
+        raise PicklingError("MemMMap can't be deep-copied or pickled")
+
+    def __access_addr_range_err(self, start_addr, size, needed_flag):
+        assert needed_flag != _MMapPageFlags.W, \
+            f"can't write to address 0x{start_addr:X} size 0x{size:X}"
+        return None, 0
+
+    def __access_addr_range(self, start_addr, size, needed_flag):
+        assert size > 0, "invalid size"
+        page_idx = self.addr_to_mmap_page_idx(start_addr)
+        last_addr = start_addr + size - 1
+        last_page_idx = self.addr_to_mmap_page_idx(last_addr)
+        block_addr = start_addr % BLOCK_SIZE
+        block_k = start_addr - block_addr
+        last_block_addr = last_addr % BLOCK_SIZE
+        last_block_k = last_addr - last_block_addr
+        if block_k != last_block_k:
+            return self.__access_addr_range_err(start_addr, size, needed_flag)
+        for i in range(page_idx, last_page_idx + 1):
+            flags = self.page_flags.get(i, 0)
+            if flags & needed_flag == 0:
+                return self.__access_addr_range_err(
+                    start_addr, size, needed_flag)
+            if needed_flag is _MMapPageFlags.W:
+                self.modified_pages.add(page_idx)
+        return self.mem_blocks[block_k], block_addr
+
+    def get_ctypes(self, start_addr, size, is_write):
+        """ returns a ctypes ubyte array referring to the memory at
+        `start_addr` with size `size`
+        """
+        flag = _MMapPageFlags.W if is_write else _MMapPageFlags.R
+        block, block_addr = self.__access_addr_range(start_addr, size, flag)
+        assert block is not None, \
+            f"can't read from address 0x{start_addr:X} size 0x{size:X}"
+        return (ctypes.c_ubyte * size).from_buffer(block, block_addr)
+
+    def _read_word(self, word_idx, reason):
+        block, block_addr = self.__access_addr_range(
+            word_idx * self.bytes_per_word, self.bytes_per_word,
+            reason.needed_mmap_page_flag)
+        if block is None:
+            return reason.read_default
+        bytes_ = block[block_addr:block_addr + self.bytes_per_word]
+        return int.from_bytes(bytes_, 'little')
+
+    def _write_word(self, word_idx, value):
+        block, block_addr = self.__access_addr_range(
+            word_idx * self.bytes_per_word, self.bytes_per_word,
+            _MMapPageFlags.W)
+        bytes_ = value.to_bytes(self.bytes_per_word, 'little')
+        block[block_addr:block_addr + self.bytes_per_word] = bytes_
+
+    def word_idxs(self):
+        for page_idx in self.modified_pages:
+            start = self.mmap_page_idx_to_addr(page_idx)
+            end = start + _MMAP_PAGE_SIZE
+            yield from range(start // self.bytes_per_word,
+                             end // self.bytes_per_word)
index 656e2b5bf948f15162c6d37ead2df22201f62582..00de46914db1a2591212079834ee81cf9fd85bb8 100644 (file)
@@ -43,7 +43,7 @@ class SimRunner(StateRunner):
     running of tests using ISACaller simulation
     """
 
-    def __init__(self, dut, m, pspec):
+    def __init__(self, dut, m, pspec, use_mmap_mem=False):
         super().__init__("sim", SimRunner)
         self.dut = dut
 
@@ -53,6 +53,7 @@ class SimRunner(StateRunner):
         self.simdec2 = simdec2 = PowerDecode2(
             None, regreduce_en=regreduce_en, fp_en=fp_en)
         m.submodules.simdec2 = simdec2  # pain in the neck
+        self.use_mmap_mem = use_mmap_mem
 
     def prepare_for_test(self, test):
         self.test = test
@@ -75,7 +76,8 @@ class SimRunner(StateRunner):
                   initial_svstate=test.svstate,
                   mmu=self.mmu,
                   fpregfile=test.fpregs,
-                  initial_fpscr=test.initial_fpscr)
+                  initial_fpscr=test.initial_fpscr,
+                  use_mmap_mem=self.use_mmap_mem)
 
         # run the loop of the instructions on the current test
         index = sim.pc.CIA.value//4
@@ -131,7 +133,8 @@ class TestRunnerBase(FHDLTestCase):
 
     def __init__(self, tst_data, microwatt_mmu=False, rom=None,
                  svp64=True, run_hdl=None, run_sim=True,
-                 allow_overlap=False, inorder=False, fp=False):
+                 allow_overlap=False, inorder=False, fp=False,
+                 use_mmap_mem=False):
         super().__init__("run_all")
         self.test_data = tst_data
         self.microwatt_mmu = microwatt_mmu
@@ -142,6 +145,7 @@ class TestRunnerBase(FHDLTestCase):
         self.run_hdl = run_hdl
         self.run_sim = run_sim
         self.fp = fp
+        self.use_mmap_mem = use_mmap_mem
 
     def run_all(self):
         m = Module()
@@ -195,7 +199,7 @@ class TestRunnerBase(FHDLTestCase):
             state_list.append(hdlrun)
 
         if self.run_sim:
-            simrun = SimRunner(self, m, pspec)
+            simrun = SimRunner(self, m, pspec, use_mmap_mem=self.use_mmap_mem)
             state_list.append(simrun)
 
         # run core clock at same rate as test clock