draft actual CMO operations - riscvarchive/riscv-CMOs-discuss GitHub Wiki

Actual CMO operations .<cmo_specifier>.<cmo_operation>

Actual CMO operations- flushes and prefetches, etc.

This proposal includes the following actual CMO operations. Short names are listed here - more cvomplete deascriptions in a section below.

Traditional CMOs: CLEAN, FLUSH, INVALIDATE-I$, DISCARD
Less Common: INVALIDATE-CLEAN, SET-LRU, LOCK-LINE.

Space should be reserved for more operations, included SAFER_DISCARD_1 and SAFER_DISCARD_2, that remedy the security deficiences of the DISCARD operation (the well known PowerPC DCBA) while preserving much of the performance advantage.

In addition to these CBOs that perform various forms of flushes and invalidates, this proposal includes operations that are often not called CMOs.

Prefetches: PREFETCH-R, PREFETCH-EW, PREFETCH-X - using the variable address range approach.
Destructive: ZALLOC - allocate a zero-filled-cache line.

Some have requested locking versions: ZALLOC-and-LOCK, and FETCH-R/W/X-and-LOCK.

COUNT: 13 encodings: 4 bits.

Security / Timing Channel Bit

Requirement: in addition to flushing caches, it is also required, for timing channel mitigation such as in Spectre, to flush microarchitecture mechanisms that can provide timing channekls, such as LRU bits, predictors and prefetchers. Some of these are associated with cache entries - hence the security/timing channel "bit". Not actually a bit - applied only to 2 CMOs.

The security property is applied to the CMO.UR variants that leave no data behind: FLUSH and INVALIDATE.

This increases the COUNT to 15 encodings: 4 bits.

Detailed description of CMO operations

Unfortunately, there is no widespread agreement as to what CMO names should be. It is therefore necessary to define their behavior more completely according to cache states.

Without loss of generality we will mention only tywo cache states, Clean and Dirty, relevant to writeback caches. Writethrough and instruction caches contain only clean data, so may map to more than one operation that handles dirty data.

Traditi0nal CMOs

CLEAN
- Dirty-→WB-→Clean
- Clean-→Clean
FLUSH
- Dirty-→WB-→Invalid
- Clean-→Invalid
- Alternate names
- Intel calls this WBINVD
- Special considerations: security/timing channel variant for CMO.UR
DISCARD
- Dirty-→no WB-→Invalid
- Clean-→Invalid
- Alternate names
- Intel calls this INVD
- Special considerations:
  - security/timing channel variant for CMO.UR
  - security hole
    
    there are several safedr variants of DISCARD, reserving space for bit not actually part of this proposal
DISCARD-CLEAN
- Dirty-→unaffected
- Clean-→Invalid
- Special considerations:
  - can be used in some incoherehnt I/O use cases
  - remedies the security problems of DISCARD - safe for user mode
SET-LRU
- CMO.VAR only
- most useful special case of the class of replacement algorithm manipulation CMOs

Operations not typically considered CMOs:

PREFETCH-R
PREFETCH-W
- prefetches in exclusive clean or dirty state - ready for writes with least possible expense
PREFETCH-X
- prefetch code, to execute
- like PREFETCH-R, except targetting I$ level(s)

Destructive

ZALLOC
- allocate cache line with reading - zero filling
- PowerPC DCBZ
ALLOC
- allocate cache line with reading - using whatever was there before
- security hole - but still sometimes used
- PowerPC DCBA

Locking variants of the above * FETCH-R-and-LOCK * FETCH-W-and-LOCK * FETCH-X-and-LOCK * ZALLOC-and-LOCK * ALLOC-and-LOCK

*Count*: 15 operations - 4 bits