riscv/rocket-chip
1
0
Fork 0
Code Releases Activity

TLToAXI4: block TL early source re-use before it goes to AXI4 (#1110)

Browse Source 
This is a follow-up to PR #1108.

Rather than increasing the number of transactions we allow to be inflight,
instead just block TL when early source re-use happens. This is a better
fix since it means we don't pay mostly wasted downstream hardware to handle
an additional transaction inflight that almost never happens.
This commit is contained in:
Wesley W. Terpstra 2017-11-14 16:08:43 -08:00 committed by GitHub
parent 147fad6387
commit 509a48c9c9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 9 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
src/main/scala/tilelink/ToAXI4.scala
View File

@ -100,7 +100,7 @@ class TLToAXI4(val combinational: Boolean = true, val adapterName: Option[String
val a_source = in.a.bits.source
val a_size = edgeIn.size(in.a.bits)
val a_isPut = edgeIn.hasData(in.a.bits)
val a_last = edgeIn.last(in.a)
val (a_first, a_last, _) = edgeIn.firstlast(in.a)
// Make sure the fields are within the bounds we assumed
assert (a_source < UInt(BigInt(1) << sourceBits))
@ -154,7 +154,7 @@ class TLToAXI4(val combinational: Boolean = true, val adapterName: Option[String
arw.qos := UInt(0) // no QoS
arw.user.foreach { _ := a_state }
val stall = sourceStall(in.a.bits.source)
val stall = sourceStall(in.a.bits.source) && a_first
in.a.ready := !stall && Mux(a_isPut, (doneAW || out_arw.ready) && out_w.ready, out_arw.ready)
out_arw.valid := !stall && in.a.valid && Mux(a_isPut, !doneAW && out_w.ready, Bool(true))
@ -191,14 +191,14 @@ class TLToAXI4(val combinational: Boolean = true, val adapterName: Option[String
val d_sel = UIntToOH(Mux(r_wins, out.r.bits.id, out.b.bits.id), edgeOut.master.endId).toBools
val d_last = Mux(r_wins, out.r.bits.last, Bool(true))
// If FIFO was requested, ensure that R+W ordering is preserved
(a_sel zip d_sel zip idStall zip idCount) filter { case (_, n) => n.isDefined } foreach { case (((as, ds), s), n) =>
(a_sel zip d_sel zip idStall zip idCount) foreach { case (((as, ds), s), n) =>
// AXI does not guarantee read vs. write ordering. In particular, if we
// are in the middle of receiving a read burst and then issue a write,
// the write might affect the read burst. This violates FIFO behaviour.
// To solve this, we must wait until the last beat of a burst, but this
// means that there can be idCount+1 operations counted due to a TileLink
// master which performs early source reuse.
val maxCount = n.get + 1
// means that a TileLink master which performs early source reuse can
// have one more transaction inflight than we promised AXI; stall it too.
val maxCount = n.getOrElse(1)
val count = RegInit(UInt(0, width = log2Ceil(maxCount + 1)))
val write = Reg(Bool())
val idle = count === UInt(0)
@ -211,7 +211,9 @@ class TLToAXI4(val combinational: Boolean = true, val adapterName: Option[String
assert (!inc || count =/= UInt(maxCount)) // overflow
when (inc) { write := arw.wen }
s := !idle && write =/= arw.wen
// If only one transaction can be inflight, it can't mismatch
val mismatch = if (maxCount > 1) { write =/= arw.wen } else { Bool(false) }
s := (!idle && mismatch) || (count === UInt(maxCount))
}
// Tie off unused channels