On Thu, Mar 30, 2023 at 05:59:37PM +0800, Alan Huang wrote:
> > Alan, would you like to drag the counting code kicking and screaming
> > into the 2020s? ;-)
>
> I’d like to!
Very good! Again, this code was written before tools such as LKMM
existed, so you will likely find instances where I used heavier-weight
memory ordering than needed, and, as always, you also will likely find
out-and-out bugs.
Looking forward to seeing what you come up with!
Thanx, Paul
> Thanks,
> Alan
>
> > 2023年3月30日 上午7:18，Paul E. McKenney <paulmck@xxxxxxxxxx> 写道：
> >
> > On Wed, Mar 29, 2023 at 08:43:33AM +0900, Akira Yokosawa wrote:
> >> On Tue, 28 Mar 2023 10:58:24 -0400, Alan Huang wrote:
> >>> This patch add several 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 original code 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)
> >>> {
> >>> WRITE_ONCE(*counter, READ_ONCE(*counter)+1);
> >>> }
> >>> }
> >>>
> >>> P2(int *counter)
> >>> {
> >>> int r0;
> >>>
> >>> r0 = READ_ONCE(*counter);
> >>> }
> >>>
> >>> exists (2:r0=2)
> >>>
> >>> Note that I also changed the reading of theft variable
> >>> to smp_load_acquire in add_count/sub_count's fast path
> >>> to make sure that reading theft happens before reading
> >>> counter.
> >>>
> >>> Signed-off-by: Alan Huang <mmpgouride@xxxxxxxxx>
> >>
> >> Looks good to me.
> >> Reviewed-by: Akira Yokosawa <akiyks@xxxxxxxxx>
> >
> > Thank you both!
> >
> >> Paul, there is a question for you.
> >> Please see below.
> >>
> >>> ---
> >>> CodeSamples/count/count_lim_sig.c | 8 ++++----
> >>> 1 file changed, 4 insertions(+), 4 deletions(-)
> >>>
> >>> diff --git a/CodeSamples/count/count_lim_sig.c b/CodeSamples/count/count_lim_sig.c
> >>> index 023d6215..59da8077 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}
> >>>
> >>> @@ -116,7 +116,7 @@ int add_count(unsigned long delta) //\lnlbl{b}
> >>>
> >>> WRITE_ONCE(counting, 1); //\lnlbl{fast:b}
> >>> barrier(); //\lnlbl{barrier:1}
> >>> - if (READ_ONCE(theft) <= THEFT_REQ && //\lnlbl{check:b}
> >>> + if (smp_load_acquire(&theft) <= THEFT_REQ && //\lnlbl{check:b}
> >>> countermax - counter >= delta) { //\lnlbl{check:e}
> >>> WRITE_ONCE(counter, counter + delta); //\lnlbl{add:f}
> >>> fastpath = 1; //\lnlbl{fasttaken}
> >>> @@ -155,7 +155,7 @@ int sub_count(unsigned long delta)
> >>>
> >>> WRITE_ONCE(counting, 1);
> >>> barrier();
> >>> - if (READ_ONCE(theft) <= THEFT_REQ &&
> >>> + if (smp_load_acquire(&theft) <= THEFT_REQ &&
> >>> counter >= delta) {
> >>> WRITE_ONCE(counter, counter - delta);
> >>> fastpath = 1;
> >>
> >> In add_count() and sub_count(), is the plain accesses to counter
> >> guaranteed to see the value stored in flush_local_count()?
> >>
> >> I'm asking because counter is declared as a thread-local variable:
> >>
> >> In this code, as each of the plain loads of counter is the first
> >> access to the variable in each function, there is not much room
> >> for compilers to exploit the plainness, I suppose.
> >>
> >> What do you think?
> >
> > Well, gitk thinks that the last deliberate change to this code was in
> > 2010, before READ_ONCE() and WRITE_ONCE(), let alone LKMM. So it needs
> > careful scrutiny, along with the other counting algorithms.
> >
> > This commit looks like a step in the right direction, so I am applying
> > it, thank you both!
> >
> > Alan, would you like to drag the counting code kicking and screaming
> > into the 2020s? ;-)
> >
> > Thanx, Paul
>
