src/soc/fu/ldst/loadstore.py

   1 """LoadStore1 FSM.
   2
   3 based on microwatt loadstore1.vhdl, but conforming to PortInterface.
   4 unlike loadstore1.vhdl this does *not* deal with actual Load/Store
   5 ops: that job is handled by LDSTCompUnit, which talks to LoadStore1
   6 by way of PortInterface.  PortInterface is where things need extending,
   7 such as adding dcbz support, etc.
   8
   9 this module basically handles "pure" load / store operations, and
  10 its first job is to ask the D-Cache for the data.  if that fails,
  11 the second task (if virtual memory is enabled) is to ask the MMU
  12 to perform a TLB, then to go *back* to the cache and ask again.
  13
  14 Links:
  15
  16 * https://bugs.libre-soc.org/show_bug.cgi?id=465
  17
  18 """
  19
  20 from nmigen import (Elaboratable, Module, Signal, Shape, unsigned, Cat, Mux,
  21                     Record, Memory,
  22                     Const)
  23 from nmutil.iocontrol import RecordObject
  24 from nmutil.util import rising_edge
  25 from enum import Enum, unique
  26
  27 from soc.experiment.dcache import DCache
  28 from soc.experiment.pimem import PortInterfaceBase
  29 from soc.experiment.mem_types import LoadStore1ToMMUType
  30 from soc.experiment.mem_types import MMUToLoadStore1Type
  31
  32 from soc.minerva.wishbone import make_wb_layout
  33 from soc.bus.sram import SRAM
  34 from nmutil.util import Display
  35
  36
  37 @unique
  38 class State(Enum):
  39     IDLE = 0       # ready for instruction
  40     ACK_WAIT = 1   # waiting for ack from dcache
  41     MMU_LOOKUP = 2 # waiting for MMU to look up translation
  42     TLBIE_WAIT = 3 # waiting for MMU to finish doing a tlbie
  43
  44
  45 # captures the LDSTRequest from the PortInterface, which "blips" most
  46 # of this at us (pipeline-style).
  47 class LDSTRequest(RecordObject):
  48     def __init__(self, name=None):
  49         RecordObject.__init__(self, name=name)
  50
  51         self.load          = Signal()
  52         self.dcbz          = Signal()
  53         self.addr          = Signal(64)
  54         # self.store_data    = Signal(64) # this is already sync (on a delay)
  55         self.byte_sel      = Signal(8)
  56         self.nc            = Signal()              # non-cacheable access
  57         self.virt_mode     = Signal()
  58         self.priv_mode     = Signal()
  59         self.align_intr    = Signal()
  60
  61 # glue logic for microwatt mmu and dcache
  62 class LoadStore1(PortInterfaceBase):
  63     def __init__(self, pspec):
  64         self.pspec = pspec
  65         self.disable_cache = (hasattr(pspec, "disable_cache") and
  66                               pspec.disable_cache == True)
  67         regwid = pspec.reg_wid
  68         addrwid = pspec.addr_wid
  69
  70         super().__init__(regwid, addrwid)
  71         self.dcache = DCache()
  72         # these names are from the perspective of here (LoadStore1)
  73         self.d_out  = self.dcache.d_in     # in to dcache is out for LoadStore
  74         self.d_in = self.dcache.d_out      # out from dcache is in for LoadStore
  75         self.m_out  = LoadStore1ToMMUType() # out *to* MMU
  76         self.m_in = MMUToLoadStore1Type()   # in *from* MMU
  77         self.req = LDSTRequest(name="ldst_req")
  78
  79         # TODO, convert dcache wb_in/wb_out to "standard" nmigen Wishbone bus
  80         self.dbus = Record(make_wb_layout(pspec))
  81
  82         # for creating a single clock blip to DCache
  83         self.d_valid = Signal()
  84         self.d_w_valid = Signal()
  85         self.d_validblip = Signal()
  86
  87         # DSISR and DAR cached values.  note that the MMU FSM is where
  88         # these are accessed by OP_MTSPR/OP_MFSPR, on behalf of LoadStore1.
  89         # by contrast microwatt has the spr set/get done *in* loadstore1.vhdl
  90         self.dsisr = Signal(64)
  91         self.dar = Signal(64)
  92
  93         # state info for LD/ST
  94         self.done          = Signal()
  95         # latch most of the input request
  96         self.load          = Signal()
  97         self.tlbie         = Signal()
  98         self.dcbz          = Signal()
  99         self.addr          = Signal(64)
 100         self.store_data    = Signal(64)
 101         self.load_data     = Signal(64)
 102         self.byte_sel      = Signal(8)
 103         #self.xerc         : xer_common_t;
 104         #self.reserve       = Signal()
 105         #self.atomic        = Signal()
 106         #self.atomic_last   = Signal()
 107         #self.rc            = Signal()
 108         self.nc            = Signal()              # non-cacheable access
 109         self.virt_mode     = Signal()
 110         self.priv_mode     = Signal()
 111         self.state        = Signal(State)
 112         self.instr_fault   = Signal()
 113         self.align_intr    = Signal()
 114         self.busy          = Signal()
 115         self.wait_dcache   = Signal()
 116         self.wait_mmu      = Signal()
 117         #self.mode_32bit    = Signal()
 118         #self.intr_vec     : integer range 0 to 16#fff#;
 119         #self.nia           = Signal(64)
 120         #self.srr1          = Signal(16)
 121
 122     def set_wr_addr(self, m, addr, mask, misalign, msr_pr):
 123         m.d.comb += self.req.load.eq(0) # store operation
 124         m.d.comb += self.req.byte_sel.eq(mask)
 125         m.d.comb += self.req.addr.eq(addr)
 126         m.d.comb += self.req.priv_mode.eq(~msr_pr) # not-problem  ==> priv
 127         m.d.comb += self.req.virt_mode.eq(msr_pr) # problem-state ==> virt
 128         m.d.comb += self.req.align_intr.eq(misalign)
 129         # option to disable the cache entirely for write
 130         if self.disable_cache:
 131             m.d.comb += self.req.nc.eq(1)
 132         return None
 133
 134     def set_rd_addr(self, m, addr, mask, misalign, msr_pr):
 135         m.d.comb += self.d_valid.eq(1)
 136         m.d.comb += self.req.load.eq(1) # load operation
 137         m.d.comb += self.req.byte_sel.eq(mask)
 138         m.d.comb += self.req.align_intr.eq(misalign)
 139         m.d.comb += self.req.addr.eq(addr)
 140         m.d.comb += self.req.priv_mode.eq(~msr_pr) # not-problem  ==> priv
 141         m.d.comb += self.req.virt_mode.eq(msr_pr) # problem-state ==> virt
 142         # BAD HACK! disable cacheing on LD when address is 0xCxxx_xxxx
 143         # this is for peripherals. same thing done in Microwatt loadstore1.vhdl
 144         with m.If(addr[28:] == Const(0xc, 4)):
 145             m.d.comb += self.req.nc.eq(1)
 146         # option to disable the cache entirely for read
 147         if self.disable_cache:
 148             m.d.comb += self.req.nc.eq(1)
 149         return None #FIXME return value
 150
 151     def set_wr_data(self, m, data, wen):
 152         # do the "blip" on write data
 153         m.d.comb += self.d_valid.eq(1)
 154         # put data into comb which is picked up in main elaborate()
 155         m.d.comb += self.d_w_valid.eq(1)
 156         m.d.comb += self.store_data.eq(data)
 157         #m.d.sync += self.d_out.byte_sel.eq(wen) # this might not be needed
 158         st_ok = self.done # TODO indicates write data is valid
 159         return st_ok
 160
 161     def get_rd_data(self, m):
 162         ld_ok = self.done     # indicates read data is valid
 163         data = self.load_data # actual read data
 164         return data, ld_ok
 165
 166     def elaborate(self, platform):
 167         m = super().elaborate(platform)
 168         comb, sync = m.d.comb, m.d.sync
 169
 170         # create dcache module
 171         m.submodules.dcache = dcache = self.dcache
 172
 173         # temp vars
 174         d_out, d_in, dbus = self.d_out, self.d_in, self.dbus
 175         m_out, m_in = self.m_out, self.m_in
 176         exc = self.pi.exc_o
 177         exception = exc.happened
 178         mmureq = Signal()
 179
 180         # copy of address, but gets over-ridden for OP_FETCH_FAILED
 181         maddr = Signal(64)
 182         m.d.comb += maddr.eq(self.addr)
 183
 184         # create a blip (single pulse) on valid read/write request
 185         # this can be over-ridden in the FSM to get dcache to re-run
 186         # a request when MMU_LOOKUP completes.
 187         m.d.comb += self.d_validblip.eq(rising_edge(m, self.d_valid))
 188         ldst_r = LDSTRequest("ldst_r")
 189
 190         # fsm skeleton
 191         with m.Switch(self.state):
 192             with m.Case(State.IDLE):
 193                 with m.If(self.d_validblip & ~exc.happened):
 194                     comb += self.busy.eq(1)
 195                     sync += self.state.eq(State.ACK_WAIT)
 196                     sync += ldst_r.eq(self.req) # copy of LDSTRequest on "blip"
 197                 with m.Else():
 198                     sync += ldst_r.eq(0)
 199
 200             # waiting for completion
 201             with m.Case(State.ACK_WAIT):
 202                 comb += self.busy.eq(~exc.happened)
 203
 204                 with m.If(d_in.error):
 205                     # cache error is not necessarily "final", it could
 206                     # be that it was just a TLB miss
 207                     with m.If(d_in.cache_paradox):
 208                         comb += exception.eq(1)
 209                         sync += self.state.eq(State.IDLE)
 210                         sync += self.dsisr[63 - 38].eq(~self.load)
 211                         # XXX there is no architected bit for this
 212                         # (probably should be a machine check in fact)
 213                         sync += self.dsisr[63 - 35].eq(d_in.cache_paradox)
 214
 215                     with m.Else():
 216                         # Look up the translation for TLB miss
 217                         # and also for permission error and RC error
 218                         # in case the PTE has been updated.
 219                         comb += mmureq.eq(1)
 220                         sync += self.state.eq(State.MMU_LOOKUP)
 221                 with m.If(d_in.valid):
 222                     m.d.comb += self.done.eq(~mmureq) # done if not doing MMU
 223                     with m.If(self.done):
 224                         sync += Display("ACK_WAIT, done %x", self.addr)
 225                     sync += self.state.eq(State.IDLE)
 226                     with m.If(self.load):
 227                         m.d.comb += self.load_data.eq(d_in.data)
 228
 229             # waiting here for the MMU TLB lookup to complete.
 230             # either re-try the dcache lookup or throw MMU exception
 231             with m.Case(State.MMU_LOOKUP):
 232                 comb += self.busy.eq(1)
 233                 with m.If(m_in.done):
 234                     with m.If(~self.instr_fault):
 235                         sync += Display("MMU_LOOKUP, done %x -> %x",
 236                                         self.addr, d_out.addr)
 237                         # retry the request now that the MMU has
 238                         # installed a TLB entry, if not exception raised
 239                         m.d.comb += self.d_out.valid.eq(~exception)
 240                         sync += self.state.eq(State.ACK_WAIT)
 241                     with m.Else():
 242                         sync += Display("MMU_LOOKUP, exception %x", self.addr)
 243                         # instruction lookup fault: store address in DAR
 244                         comb += exc.happened.eq(1)
 245                         sync += self.dar.eq(self.addr)
 246
 247                 with m.If(m_in.err):
 248                     # MMU RADIX exception thrown
 249                     comb += exception.eq(1)
 250                     sync += self.dsisr[63 - 33].eq(m_in.invalid)
 251                     sync += self.dsisr[63 - 36].eq(m_in.perm_error)
 252                     sync += self.dsisr[63 - 38].eq(self.load)
 253                     sync += self.dsisr[63 - 44].eq(m_in.badtree)
 254                     sync += self.dsisr[63 - 45].eq(m_in.rc_error)
 255
 256             with m.Case(State.TLBIE_WAIT):
 257                 pass
 258
 259         # alignment error: store address in DAR
 260         with m.If(self.align_intr):
 261             comb += exc.happened.eq(1)
 262             sync += self.dar.eq(self.addr)
 263
 264         # happened, alignment, instr_fault, invalid.
 265         # note that all of these flow through - eventually to the TRAP
 266         # pipeline, via PowerDecoder2.
 267         comb += exc.invalid.eq(m_in.invalid)
 268         comb += exc.alignment.eq(self.align_intr)
 269         comb += exc.instr_fault.eq(self.instr_fault)
 270         # badtree, perm_error, rc_error, segment_fault
 271         comb += exc.badtree.eq(m_in.badtree)
 272         comb += exc.perm_error.eq(m_in.perm_error)
 273         comb += exc.rc_error.eq(m_in.rc_error)
 274         comb += exc.segment_fault.eq(m_in.segerr)
 275
 276         # TODO, connect dcache wb_in/wb_out to "standard" nmigen Wishbone bus
 277         comb += dbus.adr.eq(dcache.wb_out.adr)
 278         comb += dbus.dat_w.eq(dcache.wb_out.dat)
 279         comb += dbus.sel.eq(dcache.wb_out.sel)
 280         comb += dbus.cyc.eq(dcache.wb_out.cyc)
 281         comb += dbus.stb.eq(dcache.wb_out.stb)
 282         comb += dbus.we.eq(dcache.wb_out.we)
 283
 284         comb += dcache.wb_in.dat.eq(dbus.dat_r)
 285         comb += dcache.wb_in.ack.eq(dbus.ack)
 286         if hasattr(dbus, "stall"):
 287             comb += dcache.wb_in.stall.eq(dbus.stall)
 288
 289         # write out d data only when flag set
 290         with m.If(self.d_w_valid):
 291             m.d.sync += d_out.data.eq(self.store_data)
 292         with m.Else():
 293             m.d.sync += d_out.data.eq(0)
 294
 295         # this must move into the FSM, conditionally noticing that
 296         # the "blip" comes from self.d_validblip.
 297         # task 1: look up in dcache
 298         # task 2: if dcache fails, look up in MMU.
 299         # do **NOT** confuse the two.
 300         with m.If(self.d_validblip):
 301             m.d.comb += self.d_out.valid.eq(~exc.happened)
 302             m.d.comb += d_out.load.eq(self.req.load)
 303             m.d.comb += d_out.byte_sel.eq(self.req.byte_sel)
 304             m.d.comb += self.addr.eq(self.req.addr)
 305             m.d.comb += d_out.nc.eq(self.req.nc)
 306             m.d.comb += d_out.priv_mode.eq(self.req.priv_mode)
 307             m.d.comb += d_out.virt_mode.eq(self.req.virt_mode)
 308             m.d.comb += self.align_intr.eq(self.req.align_intr)
 309         with m.Else():
 310             m.d.comb += d_out.load.eq(ldst_r.load)
 311             m.d.comb += d_out.byte_sel.eq(ldst_r.byte_sel)
 312             m.d.comb += self.addr.eq(ldst_r.addr)
 313             m.d.comb += d_out.nc.eq(ldst_r.nc)
 314             m.d.comb += d_out.priv_mode.eq(ldst_r.priv_mode)
 315             m.d.comb += d_out.virt_mode.eq(ldst_r.virt_mode)
 316             m.d.comb += self.align_intr.eq(ldst_r.align_intr)
 317
 318         # XXX these should be possible to remove but for some reason
 319         # cannot be... yet. TODO, investigate
 320         m.d.comb += self.load_data.eq(d_in.data)
 321         m.d.comb += d_out.addr.eq(self.addr)
 322
 323         # Update outputs to MMU
 324         m.d.comb += m_out.valid.eq(mmureq)
 325         m.d.comb += m_out.iside.eq(self.instr_fault)
 326         m.d.comb += m_out.load.eq(ldst_r.load)
 327         # m_out.priv <= r.priv_mode; TODO
 328         m.d.comb += m_out.tlbie.eq(self.tlbie)
 329         # m_out.mtspr <= mmu_mtspr; # TODO
 330         # m_out.sprn <= sprn; # TODO
 331         m.d.comb += m_out.addr.eq(maddr)
 332         # m_out.slbia <= l_in.insn(7); # TODO: no idea what this is
 333         # m_out.rs <= l_in.data; # nope, probably not needed, TODO investigate
 334
 335         return m
 336
 337     def ports(self):
 338         yield from super().ports()
 339         # TODO: memory ports
 340
 341
 342 class TestSRAMLoadStore1(LoadStore1):
 343     def __init__(self, pspec):
 344         super().__init__(pspec)
 345         pspec = self.pspec
 346         # small 32-entry Memory
 347         if (hasattr(pspec, "dmem_test_depth") and
 348                 isinstance(pspec.dmem_test_depth, int)):
 349             depth = pspec.dmem_test_depth
 350         else:
 351             depth = 32
 352         print("TestSRAMBareLoadStoreUnit depth", depth)
 353
 354         self.mem = Memory(width=pspec.reg_wid, depth=depth)
 355
 356     def elaborate(self, platform):
 357         m = super().elaborate(platform)
 358         comb = m.d.comb
 359         m.submodules.sram = sram = SRAM(memory=self.mem, granularity=8,
 360                                         features={'cti', 'bte', 'err'})
 361         dbus = self.dbus
 362
 363         # directly connect the wishbone bus of LoadStoreUnitInterface to SRAM
 364         # note: SRAM is a target (slave), dbus is initiator (master)
 365         fanouts = ['dat_w', 'sel', 'cyc', 'stb', 'we', 'cti', 'bte']
 366         fanins = ['dat_r', 'ack', 'err']
 367         for fanout in fanouts:
 368             print("fanout", fanout, getattr(sram.bus, fanout).shape(),
 369                   getattr(dbus, fanout).shape())
 370             comb += getattr(sram.bus, fanout).eq(getattr(dbus, fanout))
 371             comb += getattr(sram.bus, fanout).eq(getattr(dbus, fanout))
 372         for fanin in fanins:
 373             comb += getattr(dbus, fanin).eq(getattr(sram.bus, fanin))
 374         # connect address
 375         comb += sram.bus.adr.eq(dbus.adr)
 376
 377         return m
 378