Skip to content

Commit

Permalink
update docs and tests for weak cas! and lock! operations
Browse files Browse the repository at this point in the history
The `vector-cas!`, `box-cas!`, and `ftype-lock!` operations are "weak"
in the sense that they may spuriously fail on an architecture (like
AArch64) with a weak memory model. Also, they do not necessarily imply
any overall memory ordering, so using `memory-order-acquire` and
`memory-order-release` may be needed. The documentation did not make
this clear, and some tests did not handle spurious failure.

Closes #720
  • Loading branch information
mflatt committed Oct 18, 2023
1 parent 2451d12 commit 4ba5455
Show file tree
Hide file tree
Showing 8 changed files with 74 additions and 24 deletions.
18 changes: 16 additions & 2 deletions csug/objects.stex
Original file line number Diff line number Diff line change
Expand Up @@ -820,9 +820,16 @@ If the \var{n}th element of \var{vector} that would be replaced
is not \scheme{eq?} to \var{old-obj}, then
\var{vector} is unchanged.

On an architecture with a weak memory model, \scheme{vector-cas!} can
``spuriously'' fail, leaving \var{vector} unchanged and returning
\scheme{#f} even if the current value of element \var{n} is
\var{old-obj}. On success, no memory ordering is implied, which means
that \scheme{memory-order-acquire} and/or \scheme{memory-order-release}
may be needed to complete an intended synchronization.

\schemedisplay
(define v (vector 'old0 'old1 'old2))
(vector-cas! v 1 'old1 'new1) ;=> #t
(vector-cas! v 1 'old1 'new1) ;=> #t, assuming no spurious failure
(vector-ref v 1) ;=> 'new1
(vector-cas! v 2 'old1 'new2) ;=> #f
(vector-ref v 2) ;=> 'old2
Expand Down Expand Up @@ -1852,9 +1859,16 @@ if the replaced content is \scheme{eq?} to \var{old-obj}.
If the content of \var{box} that would be replaced is not \scheme{eq?} to \var{old-obj}, then
\var{box} is unchanged.

On an architecture with a weak memory model, \scheme{box-cas!} can
``spuriously'' fail, leaving \var{box} unchanged and returning
\scheme{#f} even if the current value in \var{box} is \var{old-obj}.
On success, no memory ordering is implied, which means that
\scheme{memory-order-acquire} and/or \scheme{memory-order-release} may be
needed to complete an intended synchronization.

\schemedisplay
(define b (box 'old))
(box-cas! b 'old 'new) ;=> #t
(box-cas! b 'old 'new) ;=> #t, assuming no spurious failure
(unbox b) ;=> 'new
(box-cas! b 'other 'wrong) ;=> #f
(unbox b) ;=> 'new
Expand Down
18 changes: 14 additions & 4 deletions csug/threads.stex
Original file line number Diff line number Diff line change
Expand Up @@ -364,7 +364,10 @@ not just by the process or thread that most recently locked the lock.

The lock mechanism provides little structure, and mistakes
in allocation and use can lead to memory faults, deadlocks,
and other problems.
and other problems. Furthermore, no memory ordering is implied by a lock
operation, which means that \scheme{memory-order-acquire} and
\scheme{memory-order-release} may be needed to complete the intended
synchronization of a lock.
Thus, it is usually advisable to use locks only as part of a
higher-level abstraction that ensures locks are used in a
disciplined manner.
Expand All @@ -375,7 +378,7 @@ disciplined manner.
(foreign-alloc (ftype-sizeof uptr))))

(ftype-init-lock! uptr () lock)
(ftype-lock! uptr () lock) ;=> #t
(ftype-lock! uptr () lock) ;=> #t, assuming no spurious failure
(ftype-lock! uptr () lock) ;=> #f
(ftype-unlock! uptr () lock)
(ftype-spin-lock! uptr () lock)
Expand Down Expand Up @@ -414,9 +417,13 @@ to the use of any of the other operators; if this is not done, the
behavior of the other operators is undefined.

\scheme{ftype-lock!} can be used to lock the lock.
If it finds the lock unlocked at the time of the operation, it locks
If it finds the lock unlocked at the time of the operation, it (normally) locks
the lock and returns \scheme{#t}; if it finds the lock already locked,
it returns \scheme{#f} without changing the lock.
it returns \scheme{#f} without changing the lock. On an architecture with a weak memory model,
\scheme{ftype-lock!} can ``spuriously'' fail, leaving a lock unchanged and returning
\scheme{#f} even if the lock is currently unlocked. On success, no memory ordering is implied,
which means that \scheme{memory-order-acquire} may be
needed to complete an intended synchronization.

\scheme{ftype-spin-lock!} can also be used to lock the lock.
If it finds the lock unlocked at the time of the operation, it locks the
Expand All @@ -432,6 +439,9 @@ the lock.
\scheme{ftype-unlock!} is used to unlock a lock.
If it finds the lock locked, it unlocks the lock and returns.
Otherwise, it returns without changing the lock.
On an architecture with a weak memory model,
no memory ordering is implied, and \scheme{memory-order-release} may be
needed to complete an intended synchronization.

\section{Locked increment and decrement\label{SECTTHREADLOCKEDINCRDECR}}

Expand Down
10 changes: 5 additions & 5 deletions mats/5_6.ms
Original file line number Diff line number Diff line change
Expand Up @@ -1378,22 +1378,22 @@
(eq? 1 (vector-ref vec1 0)))
(not (vector-cas! vec1 0 0 1))
(eq? 1 (vector-ref vec1 0))
(vector-cas! vec1 0 1 4)
(retry-for-spurious (vector-cas! vec1 0 1 4))
(eq? 4 (vector-ref vec1 0))
(not (vector-cas! vec1 0 1 5))

(not (vector-cas! vec1 1 0 1))
(eq? 2 (vector-ref vec1 1))
(vector-cas! vec1 1 2 5)
(retry-for-spurious (vector-cas! vec1 1 2 5))
(eq? 5 (vector-ref vec1 1))

(not (vector-cas! vec2 0 'banana 'donut))
(vector-cas! vec2 0 'apple 'donut)
(retry-for-spurious (vector-cas! vec2 0 'apple 'donut))
(not (vector-cas! vec2 0 'apple 'eclair))
(eq? 'donut (vector-ref vec2 0))

(not (vector-cas! vec2 1 'apple 'fig))
(vector-cas! vec2 1 'banana 'fig)
(retry-for-spurious (vector-cas! vec2 1 'banana 'fig))
(not (vector-cas! vec2 1 'banana 'grape))
(eq? 'fig (vector-ref vec2 1))

Expand All @@ -1414,7 +1414,7 @@
(begin
(collect 0)
(let ([g1 (gensym)])
(and (vector-cas! vec2 2 'coconut g1)
(and (retry-for-spurious (vector-cas! vec2 2 'coconut g1))
(begin
(collect 0)
(eq? g1 (vector-ref vec2 2))))))
Expand Down
8 changes: 3 additions & 5 deletions mats/5_8.ms
Original file line number Diff line number Diff line change
Expand Up @@ -38,11 +38,11 @@
(eq? 1 (unbox bx1)))
(not (box-cas! bx1 0 1))
(eq? 1 (unbox bx1))
(box-cas! bx1 1 2)
(retry-for-spurious (box-cas! bx1 1 2))
(eq? 2 (unbox bx1))

(not (box-cas! bx2 #f 'banana))
(box-cas! bx2 'apple 'banana)
(retry-for-spurious (box-cas! bx2 'apple 'banana))
(not (box-cas! bx2 'apple 'banana))
(eq? 'banana (unbox bx2))

Expand All @@ -59,9 +59,7 @@
(begin
(collect 0)
(let ([g1 (gensym)])
(and (let loop ()
(or (box-cas! bx2 'banana g1)
(loop)))
(and (retry-for-spurious (box-cas! bx2 'banana g1))
(begin
(collect 0)
(eq? g1 (unbox bx2))))))
Expand Down
4 changes: 2 additions & 2 deletions mats/foreign.ms
Original file line number Diff line number Diff line change
Expand Up @@ -2234,13 +2234,13 @@
(define ff-locked-decr (foreign-procedure "locked_decr" ((* uptr)) boolean))
#t)
(eq? (ff-init-lock (ftype-&ref A (x) a)) (void))
(ftype-lock! A (x) a)
(retry-for-spurious (ftype-lock! A (x) a))
(not (ftype-lock! A (x) a))
(eq? (ftype-unlock! A (x) a) (void))
(eq? (ff-spinlock (ftype-&ref A (x) a)) (void))
(not (ftype-lock! A (x) a))
(eq? (ff-unlock (ftype-&ref A (x) a)) (void))
(ftype-lock! A (x) a)
(retry-for-spurious (ftype-lock! A (x) a))
(eq? (ff-unlock (ftype-&ref A (x) a)) (void))
(eq? (ff-spinlock (ftype-&ref A (x) a)) (void))
(not (ftype-lock! A (x) a))
Expand Down
21 changes: 16 additions & 5 deletions mats/ftype.ms
Original file line number Diff line number Diff line change
Expand Up @@ -2797,11 +2797,11 @@
(ftype-init-lock! A (g *) x)
#t)

(ftype-lock! A (c) x)
(ftype-lock! A (f f1) x)
(ftype-lock! A (f f2 1 f3b) x)
(ftype-lock! A (f f2 $idx f3b) x)
(ftype-lock! A (g *) x)
(retry-for-spurious (ftype-lock! A (c) x))
(retry-for-spurious (ftype-lock! A (f f1) x))
(retry-for-spurious (ftype-lock! A (f f2 1 f3b) x))
(retry-for-spurious (ftype-lock! A (f f2 $idx f3b) x))
(retry-for-spurious (ftype-lock! A (g *) x))

(not (ftype-lock! A (c) x))
(not (ftype-lock! A (f f1) x))
Expand Down Expand Up @@ -2831,6 +2831,17 @@
(fptr-free x)
(fptr-free g)
#t)

;; This is intended mainly as a test of `retry-for-spurious`:
(let ([m (make-ftype-pointer iptr (foreign-alloc (ftype-sizeof iptr)))])
(ftype-init-lock! iptr () m)
(let loop ([i 10000000])
(or (fx= i 0)
(and (retry-for-spurious (ftype-lock! iptr () m))
(begin
(ftype-unlock! iptr () m)
(loop (fx- i 1)))))))

)

(mat ftype-compile-file
Expand Down
17 changes: 17 additions & 0 deletions mats/mat.ss
Original file line number Diff line number Diff line change
Expand Up @@ -615,3 +615,20 @@
(set! counter n)
(fx= n 0))))])
(collect)))))

(define-syntax retry-for-spurious
(let ([mt (symbol->string (machine-type))])
(if (or (memq (substring mt 0 2) '("a6" "i3"))
(memq (substring mt 0 3) '("ta6" "ti3")))
;; no retry loop needed on x86
(lambda (stx)
(syntax-case stx ()
[(_ e) #'e]))
;; add retry loop
(lambda (stx)
(syntax-case stx ()
[(_ e) #'(let loop ([n 10])
;; 10 sprious failures in a row is vanishingly unlikely?
(or e
(and (> n 0)
(loop (- n 1)))))])))))
2 changes: 1 addition & 1 deletion mats/primvars.ms
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@
heap-check-interval
preexisting-profile-dump-entry?
coverage-table record-run-coverage load-coverage-files combine-coverage-files coverage-percent
parameters)))
parameters retry-for-spurious)))
(define (canonical-label x)
(let ([s (symbol->string x)])
(#%$intern3 (format "~a*top*:~a" s s) (string-length s) (+ 6 (* 2 (string-length s))))))
Expand Down

0 comments on commit 4ba5455

Please sign in to comment.