1
0
Commit Graph

526 Commits

Author SHA1 Message Date
Howard Mao
82bbbf908d Fix L2 Writeback deadlock issue
The deadlock condition occurs when the acquire tracker attempts to
request a writeback while the writeback unit is still busy and a
voluntary release for the block to be written back is coming in.

The voluntary release cannot be accepted because it conflicts with the
acquire tracker. The acquire tracker can't merge the voluntary release
because it is waiting to send the writeback. The writeback can't
progress because the release it is waiting on is behind the voluntary
release.

The solution to this is to break the atomicity guarantee between the
acquire tracker and the writeback unit. This allows the voluntary
release tracker to take the voluntary release before the writeback unit
accepts the conflicting request. This causes a potential race condition
for the metadata array. The solution to this is to have the writeback
unit re-read the metadata after accepting a request.
2016-07-26 12:31:08 -07:00
Wesley W. Terpstra
11ec5b2cf4 bram: don't deal with multibeat; rely on the fragmenter 2016-07-22 14:51:05 -07:00
Wesley W. Terpstra
a52d418439 fragmenter: support multi-beat get/put via fragmenting to single-beat operations 2016-07-22 14:48:22 -07:00
Howard Mao
9168f35971 clean up the requirements in StatelessBridge
* No need to check that release ID bits and acquire ID bits the same
 * Check that inner and outer coherence policies match
2016-07-21 19:41:56 -07:00
Howard Mao
12067a3b8d make sure outer probe and finish lines are disconnected 2016-07-21 15:15:44 -07:00
Howard Mao
c38dff0855 add some more warnings about the StatelessBridge 2016-07-21 15:07:10 -07:00
Megan Wachs
eb9e998c08 Add ManagerToClientStatelessBridge 2016-07-21 13:49:16 -07:00
Howard Mao
0a1cd64786 fix number of builtin Acquire types 2016-07-21 13:45:20 -07:00
Howard Mao
86e31be820 fix lockup from back to back releases with data 2016-07-21 12:06:58 -07:00
Wesley W. Terpstra
fa8317fec1 debug: add clock crossing primitives 2016-07-19 14:52:43 -07:00
Henry Cook
e406d1bd73 Make probeCopy have same behavior as probeDowngrade 2016-07-18 18:22:49 -07:00
Howard Mao
9eeb1112d4 fix Bufferless irel_vs_iacq_conflict signal 2016-07-18 17:38:20 -07:00
Howard Mao
e5cccc0526 don't update xact_vol_irel if not a voluntary irel 2016-07-18 17:05:23 -07:00
Howard Mao
84098db81f add a TileLinkTestRAM 2016-07-15 11:03:26 -07:00
Howard Mao
b122a54c32 don't allow more outer IDs than inner IDs 2016-07-13 12:42:28 -07:00
Howard Mao
de1e25f3d1 reduce usage of CAMs in converters 2016-07-13 11:20:50 -07:00
Howard Mao
8aa73915a1 put locking arbiter back into converter 2016-07-08 09:31:33 -07:00
Howard Mao
a50ba39ea7 Revert "add buffering and locking to TL -> Nasti converter"
This reverts commit 2109a48e18719383942d535ff4c1d0a859dcc424.

Conflicts:
	src/main/scala/converters/Nasti.scala
2016-07-08 09:31:33 -07:00
Andrew Waterman
70b677ecda Vec considered harmful; use isOneOf instead (#64)
Vec is heavyweight and really should only be used for I/O and
dynamic indexing.  A recurring pattern in uncore is

    Vec(const1, const2, const3) contains x

which is nice but has a deleterious effect on simulation copilation
and execution time.  This patch proposes an alternative:

    x isOneOf (const1, const2, const3)
    x isOneOf seqOfThings

I think it's also more idiomatic.

This is just a prototype; I'm not wed to the name or implementation.
2016-07-07 19:25:57 -07:00
Howard Mao
16a6b11081 fix bug in AXI -> TL converter 2016-07-07 14:34:24 -07:00
Howard Mao
7cc64011fb simplify amo_mask generation 2016-07-07 12:14:45 -07:00
Howard Mao
1c5e7be75b make sure Nasti write channel id is set correctly 2016-07-07 12:14:02 -07:00
Howard Mao
8ccc50a8f0 fix IdMapper and TL -> NASTI converter 2016-07-07 10:16:44 -07:00
Howard Mao
5d8d5e598b add buffering and locking to TL -> Nasti converter 2016-07-06 16:51:45 -07:00
Howard Mao
b10d306b4a add option to log L2 cache transactions for easier debugging 2016-07-06 14:59:09 -07:00
Howard Mao
64afc795fd make sure voluntary releases don't get allocated to L2WritebackUnit 2016-07-06 14:10:45 -07:00
Howard Mao
b105076996 fix ID mapper to disallow two in-flight requests with the same inner ID 2016-07-05 17:41:46 -07:00
Howard Mao
af76837970 conform to new NastiWriteDataChannel interface 2016-07-05 17:41:46 -07:00
Albert Ou
4c07aedfad Rewrite BRAMSlave to infer a single BRAM instance 2016-07-05 14:21:21 -07:00
Howard Mao
702444709a make sure pending bits updated for all releases 2016-07-05 12:08:22 -07:00
Howard Mao
06ed9c5794 add a single-entry queue in front of acquire and release for bufferless broadcast hub 2016-07-05 12:08:22 -07:00
Howard Mao
67bac383e3 hopefully fixed last bugs in Bufferless 2016-07-05 12:08:22 -07:00
Howard Mao
a35388bc27 fix merging of same xact ID puts/gets 2016-07-05 12:08:22 -07:00
Howard Mao
51f7bf1511 fix Bufferless voluntary release issue 2016-07-05 12:08:22 -07:00
Howard Mao
afc51c4a35 make sure TL -> NASTI converter handles multibeat transactions properly 2016-07-05 12:08:22 -07:00
Andrew Waterman
85808f8cbb Clean up PseudoLRU code 2016-07-02 15:09:12 -07:00
Howard Mao
caa9ca24b9 NASTI -> TL converter also uses ID mapper 2016-07-01 18:11:29 -07:00
Wesley W. Terpstra
39bee5198d Nasti Puts: decode wmask to determine addr_byte() and op_size()
This change is TL0 specific; TL2 knows the op_size, and can use
this to do a much simpler one-hot decode of the address.
2016-07-01 16:49:32 -07:00
Howard Mao
e163a23583 fix another bug in Widener 2016-07-01 16:24:48 -07:00
Howard Mao
10a46a36ae fix full_addr() function in TileLink 2016-07-01 15:17:41 -07:00
Howard Mao
61e3e5b45a more WIP on fixing Bufferless 2016-06-30 18:29:51 -07:00
Howard Mao
0eedffa82f WIP: Fix BufferlessBroadcastHub 2016-06-30 18:29:51 -07:00
Howard Mao
ce46f523c9 make sure Widener uses proper parameters to generate acquire/grant 2016-06-30 18:17:16 -07:00
Howard Mao
a0b1772404 change TileLinkWidthAdapter interface 2016-06-30 15:50:23 -07:00
Howard Mao
9feca99d5d make PutBlock wmask argument match Put 2016-06-28 13:10:46 -07:00
Howard Mao
b936aa9826 refactor uncore files into separate packages 2016-06-28 13:10:46 -07:00
Andrew Waterman
97e74aec3a Merge RTC and PRCI 2016-06-27 23:06:07 -07:00
Howard Mao
ec5b9dfc86 make sure trackers can handle case where there are no caching clients 2016-06-27 16:29:51 -07:00
Howard Mao
a93a70c8ec make sure merged voluntary releases are handled properly 2016-06-27 11:40:32 -07:00
Andrew Waterman
354b81c8fe Remove legacy HTIF things
The SCR file is gone, too, because it is tightly coupled.  The
general concept could be revived as a module that somehow connects
to (or is contained by) the debug module.
2016-06-23 13:17:11 -07:00