This patch add two necessary partial memory barriers, the first one change READ_ONCE to smp_load_acquire to makes sure the reading from theftp[t] happens before the reading from counterp[t]. The litmus testing below represents the pattern, and the result is "Sometimes": C counter_sig {} P0(int *theft, int *counter) { int r0; int r1; r0 = READ_ONCE(*theft); r1 = READ_ONCE(*counter); } P1(int *theft, int *counter) { WRITE_ONCE(*counter, 1); smp_mb(); WRITE_ONCE(*theft, 1); } exists (0:r0=1 /\ 0:r1=0) Second one change WRITE_ONCE to smp_store_release to make sure that setting counterp[t] happens before the setting theftp[p] to THEFT_IDLE. Here is the litmus testing, The result is "Sometimes": C counter_sig_2 { int theft = 1; int counter = 1; } P0(int *theft, int *counter) { WRITE_ONCE(*counter, 0); WRITE_ONCE(*theft, 0); } P1(int *theft, int *counter) { if (READ_ONCE(*theft) == 0) { smp_mb(); WRITE_ONCE(*counter, READ_ONCE(*counter)+1); } } P2(int *counter) { int r0; r0 = READ_ONCE(*counter); } exists (2:r0=2) Note that I added one smp_mb() in P1's "if" statement, because in add_count/sub_count's fast path, there is a control dependency. Signed-off-by: Alan Huang <mmpgouride@xxxxxxxxx> --- CodeSamples/count/count_lim_sig.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/CodeSamples/count/count_lim_sig.c b/CodeSamples/count/count_lim_sig.c index 023d6215..09582429 100644 --- a/CodeSamples/count/count_lim_sig.c +++ b/CodeSamples/count/count_lim_sig.c @@ -81,7 +81,7 @@ static void flush_local_count(void) //\lnlbl{flush:b} for_each_tid(t, tid) { //\lnlbl{flush:loop2:b} if (theftp[t] == NULL) //\lnlbl{flush:skip:nonexist} continue; //\lnlbl{flush:next2} - while (READ_ONCE(*theftp[t]) != THEFT_READY) {//\lnlbl{flush:loop3:b} + while (smp_load_acquire(theftp[t]) != THEFT_READY) {//\lnlbl{flush:loop3:b} poll(NULL, 0, 1); //\lnlbl{flush:block} if (READ_ONCE(*theftp[t]) == THEFT_REQ)//\lnlbl{flush:check:REQ} pthread_kill(tid, SIGUSR1);//\lnlbl{flush:signal2} @@ -90,7 +90,7 @@ static void flush_local_count(void) //\lnlbl{flush:b} *counterp[t] = 0; globalreserve -= *countermaxp[t]; *countermaxp[t] = 0; //\lnlbl{flush:thiev:e} - WRITE_ONCE(*theftp[t], THEFT_IDLE); //\lnlbl{flush:IDLE} + smp_store_release(theftp[t], THEFT_IDLE); //\lnlbl{flush:IDLE} } //\lnlbl{flush:loop2:e} } //\lnlbl{flush:e} -- 2.34.1