Hi,
Just a new version trying to make forward progress on this ;-)
v1: https://lore.kernel.org/lkml/20210930130823.2103688-1-boqun.feng@gmail.com/
Changes since v1:
* Split the patch into three to help resolve the litmus test
addition discussion.
* Add some explanation in patch #2 on the requirement of tests in
litmus-tests directory.
To summarize the change in memory model, we now guarantee in the
following code:
<memory access M>
spin_unlock(A);
spin_lock(B);
<memory access N>
M is ordered against N unless M is a store and N is a load. More
detailed examples of this guarantee can be found in patch #3.
Architecture maintainers, appreciate it that you can take a look at
patch #3 and rest of whole set to confirm this guarantee works on your
architectures.
Alan, I split the patchset into three patches because I do think we need
some sort of patch #2 so that we can have consensus about whether merge
patch #3 or not. I know you want to keep litmus-tests directory as
simple as possible, but it won't hurt to document the requirement.
Looking forwards to your thoughts ;-)
Suggestion and comments are welcome!
Regards,
Boqun
Boqun Feng (3):
tools/memory-model: Provide extra ordering for unlock+lock pair on the
same CPU
tools/memory-model: doc: Describe the requirement of the litmus-tests
directory
tools/memory-model: litmus: Add two tests for unlock(A)+lock(B)
ordering
.../Documentation/explanation.txt | 44 +++++++++++--------
tools/memory-model/README | 12 +++++
tools/memory-model/linux-kernel.cat | 6 +--
...LB+unlocklockonceonce+poacquireonce.litmus | 33 ++++++++++++++
...unlocklockonceonce+fencermbonceonce.litmus | 33 ++++++++++++++
tools/memory-model/litmus-tests/README | 8 ++++
6 files changed, 114 insertions(+), 22 deletions(-)
create mode 100644 tools/memory-model/litmus-tests/LB+unlocklockonceonce+poacquireonce.litmus
create mode 100644 tools/memory-model/litmus-tests/MP+unlocklockonceonce+fencermbonceonce.litmus
--
2.33.0
A recent discussion[1] shows that we are in favor of strengthening the
ordering of unlock + lock on the same CPU: a unlock and a po-after lock
should provide the so-called RCtso ordering, that is a memory access S
po-before the unlock should be ordered against a memory access R
po-after the lock, unless S is a store and R is a load.
The strengthening meets programmers' expection that "sequence of two
locked regions to be ordered wrt each other" (from Linus), and can
reduce the mental burden when using locks. Therefore add it in LKMM.
[1]: https://lore.kernel.org/lkml/20210909185937.GA12379@rowland.harvard.edu/
Co-developed-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Boqun Feng <redacted>
Reviewed-by: Michael Ellerman <mpe@ellerman.id.au> (powerpc)
Acked-by: Palmer Dabbelt <redacted> (RISC-V)
---
.../Documentation/explanation.txt | 44 +++++++++++--------
tools/memory-model/linux-kernel.cat | 6 +--
2 files changed, 28 insertions(+), 22 deletions(-)
@@ -1813,15 +1813,16 @@ spin_trylock() -- we can call these things lock-releases and lock-acquires -- have two properties beyond those of ordinary releases and acquires.-First, when a lock-acquire reads from a lock-release, the LKMM-requires that every instruction po-before the lock-release must-execute before any instruction po-after the lock-acquire. This would-naturally hold if the release and acquire operations were on different-CPUs, but the LKMM says it holds even when they are on the same CPU.-For example:+First, when a lock-acquire reads from or is po-after a lock-release,+the LKMM requires that every instruction po-before the lock-release+must execute before any instruction po-after the lock-acquire. This+would naturally hold if the release and acquire operations were on+different CPUs and accessed the same lock variable, but the LKMM says+it also holds when they are on the same CPU, even if they access+different lock variables. For example: int x, y;- spinlock_t s;+ spinlock_t s, t; P0() {
@@ -1842,10 +1843,10 @@ For example: WRITE_ONCE(x, 1); }-Here the second spin_lock() reads from the first spin_unlock(), and-therefore the load of x must execute before the load of y. Thus we-cannot have r1 = 1 and r2 = 0 at the end (this is an instance of the-MP pattern).+Here the second spin_lock() is po-after the first spin_unlock(), and+therefore the load of x must execute before the load of y, even though+the two locking operations use different locks. Thus we cannot have+r1 = 1 and r2 = 0 at the end (this is an instance of the MP pattern). This requirement does not apply to ordinary release and acquire fences, only to lock-related operations. For instance, suppose P0()
@@ -1872,13 +1873,13 @@ instructions in the following order: and thus it could load y before x, obtaining r2 = 0 and r1 = 1.-Second, when a lock-acquire reads from a lock-release, and some other-stores W and W' occur po-before the lock-release and po-after the-lock-acquire respectively, the LKMM requires that W must propagate to-each CPU before W' does. For example, consider:+Second, when a lock-acquire reads from or is po-after a lock-release,+and some other stores W and W' occur po-before the lock-release and+po-after the lock-acquire respectively, the LKMM requires that W must+propagate to each CPU before W' does. For example, consider: int x, y;- spinlock_t x;+ spinlock_t s; P0() {
@@ -1908,7 +1909,12 @@ each CPU before W' does. For example, consider: If r1 = 1 at the end then the spin_lock() in P1 must have read from the spin_unlock() in P0. Hence the store to x must propagate to P2-before the store to y does, so we cannot have r2 = 1 and r3 = 0.+before the store to y does, so we cannot have r2 = 1 and r3 = 0. But+if P1 had used a lock variable different from s, the writes could have+propagated in either order. (On the other hand, if the code in P0 and+P1 had all executed on a single CPU, as in the example before this+one, then the writes would have propagated in order even if the two+critical sections used different lock variables.) These two special requirements for lock-release and lock-acquire do not arise from the operational model. Nevertheless, kernel developers
@@ -27,7 +27,7 @@ include "lock.cat" (* Release Acquire *) let acq-po = [Acquire] ; po ; [M] let po-rel = [M] ; po ; [Release]-let po-unlock-rf-lock-po = po ; [UL] ; rf ; [LKR] ; po+let po-unlock-lock-po = po ; [UL] ; (po|rf) ; [LKR] ; po (* Fences *) let R4rmb = R \ Noreturn (* Reads for which rmb works *)
It's better that we have some "standard" about which test should be put
in the litmus-tests directory because it helps future contributors
understand whether they should work on litmus-tests in kernel or Paul's
GitHub repo. Therefore explain a little bit on what a "representative"
litmus test is.
Signed-off-by: Boqun Feng <redacted>
---
tools/memory-model/README | 12 ++++++++++++
1 file changed, 12 insertions(+)
@@ -195,6 +195,18 @@ litmus-tests are listed in litmus-tests/README. A great deal more litmus tests are available at https://github.com/paulmckrcu/litmus.+ By "representative", it means the one in the litmus-tests+ directory is:++ 1) simple, the number of threads should be relatively+ small and each thread function should be relatively+ simple.+ 2) orthogonal, there should be no two litmus tests+ describing the same aspect of the memory model.+ 3) textbook, developers can easily copy-paste-modify+ the litmus tests to use the patterns on their own+ code.+ lock.cat Provides a front-end analysis of lock acquisition and release, for example, associating a lock acquisition with the preceding
The memory model has been updated to provide a stronger ordering
guarantee for unlock(A)+lock(B) on the same CPU/thread. Therefore add
two litmus tests describing this new guarantee, these tests are simple
yet can clearly show the usage of the new guarantee, also they can serve
as the self tests for the modification in the model.
Co-developed-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Boqun Feng <redacted>
---
...LB+unlocklockonceonce+poacquireonce.litmus | 33 +++++++++++++++++++
...unlocklockonceonce+fencermbonceonce.litmus | 33 +++++++++++++++++++
tools/memory-model/litmus-tests/README | 8 +++++
3 files changed, 74 insertions(+)
create mode 100644 tools/memory-model/litmus-tests/LB+unlocklockonceonce+poacquireonce.litmus
create mode 100644 tools/memory-model/litmus-tests/MP+unlocklockonceonce+fencermbonceonce.litmus
@@ -0,0 +1,33 @@+C LB+unlocklockonceonce+poacquireonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, all accesses+ * in the first must execute before any accesses in the second, even if+ * the critical sections are protected by different locks.+ *)++{}++P0(spinlock_t *s, spinlock_t *t, int *x, int *y)+{+ int r1;++ spin_lock(s);+ r1 = READ_ONCE(*x);+ spin_unlock(s);+ spin_lock(t);+ WRITE_ONCE(*y, 1);+ spin_unlock(t);+}++P1(int *x, int *y)+{+ int r2;++ r2 = smp_load_acquire(y);+ WRITE_ONCE(*x, 1);+}++exists (0:r1=1 /\ 1:r2=1)
@@ -0,0 +1,33 @@+C MP+unlocklockonceonce+fencermbonceonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, stores in the+ * first must propagate to each CPU before stores in the second do, even if+ * the critical sections are protected by different locks.+ *)++{}++P0(spinlock_t *s, spinlock_t *t, int *x, int *y)+{+ spin_lock(s);+ WRITE_ONCE(*x, 1);+ spin_unlock(s);+ spin_lock(t);+ WRITE_ONCE(*y, 1);+ spin_unlock(t);+}++P1(int *x, int *y)+{+ int r1;+ int r2;++ r1 = READ_ONCE(*y);+ smp_rmb();+ r2 = READ_ONCE(*x);+}++exists (1:r1=1 /\ 1:r2=0)
@@ -63,6 +63,10 @@ LB+poonceonces.litmus As above, but with store-release replaced with WRITE_ONCE() and load-acquire replaced with READ_ONCE().+LB+unlocklockonceonce+poacquireonce.litmus+ Does a unlock+lock pair provides ordering guarantee between a+ load and a store?+ MP+onceassign+derefonce.litmus As below, but with rcu_assign_pointer() and an rcu_dereference().
@@ -90,6 +94,10 @@ MP+porevlocks.litmus As below, but with the first access of the writer process and the second access of reader process protected by a lock.+MP+unlocklockonceonce+fencermbonceonce.litmus+ Does a unlock+lock pair provides ordering guarantee between a+ store and another store?+ MP+fencewmbonceonce+fencermbonceonce.litmus Does a smp_wmb() (between the stores) and an smp_rmb() (between the loads) suffice for the message-passing litmus test, where one
@@ -0,0 +1,33 @@+C LB+unlocklockonceonce+poacquireonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, all accesses+ * in the first must execute before any accesses in the second, even if+ * the critical sections are protected by different locks.
One small nit; the above "all accesses" reads as if:
spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(t);
r1 = READ_ONCE(*y);
spin_unlock(t);
would also work, except of course that's the one reorder allowed by TSO.
From: Peter Zijlstra <peterz@infradead.org> Date: 2021-10-26 07:03:12
On Mon, Oct 25, 2021 at 10:54:13PM +0800, Boqun Feng wrote:
Boqun Feng (3):
tools/memory-model: Provide extra ordering for unlock+lock pair on the
same CPU
tools/memory-model: doc: Describe the requirement of the litmus-tests
directory
tools/memory-model: litmus: Add two tests for unlock(A)+lock(B)
ordering
I'm obviously all in favour of this :-)
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
@@ -0,0 +1,33 @@+C LB+unlocklockonceonce+poacquireonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, all accesses+ * in the first must execute before any accesses in the second, even if+ * the critical sections are protected by different locks.
One small nit; the above "all accesses" reads as if:
spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(t);
r1 = READ_ONCE(*y);
spin_unlock(t);
would also work, except of course that's the one reorder allowed by TSO.
I applied this series with Peter's Acked-by, and with the above comment
reading as follows:
+(*
+ * Result: Never
+ *
+ * If two locked critical sections execute on the same CPU, all accesses
+ * in the first must execute before any accesses in the second, even if the
+ * critical sections are protected by different locks. The one exception
+ * to this rule is that (consistent with TSO) a prior write can be reordered
+ * with a later read from the viewpoint of a process not holding both locks.
+ *)
Thank you all!
Thanx, Paul
@@ -0,0 +1,33 @@+C LB+unlocklockonceonce+poacquireonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, all accesses+ * in the first must execute before any accesses in the second, even if+ * the critical sections are protected by different locks.
One small nit; the above "all accesses" reads as if:
spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(t);
r1 = READ_ONCE(*y);
spin_unlock(t);
would also work, except of course that's the one reorder allowed by TSO.
I applied this series with Peter's Acked-by, and with the above comment
Thanks!
reading as follows:
+(*
+ * Result: Never
+ *
+ * If two locked critical sections execute on the same CPU, all accesses
+ * in the first must execute before any accesses in the second, even if the
+ * critical sections are protected by different locks. The one exception
+ * to this rule is that (consistent with TSO) a prior write can be reordered
+ * with a later read from the viewpoint of a process not holding both locks.
Just want to be accurate, in our memory model "execute" means a CPU
commit an memory access instruction to the Memory Subsystem, so if we
have a store W and a load R, where W executes before R, it doesn't mean
the memory effect of W is observed before the memory effect of R by
other CPUs, consider the following case
CPU0 Memory Subsystem CPU1
==== ====
WRITE_ONCE(*x,1); // W ---------->|
spin_unlock(s); |
spin_lock(t); |
r1 = READ_ONCE(*y); // R -------->|
// R reads 0 |
|<----------------WRITR_ONCE(*y, 1); // W'
W' propagates to CPU0 |
<-------------------------|
| smp_mb();
|<----------------r1 = READ_ONCE(*x); // R' reads 0
|
| W progrates to CPU 1
|----------------->
The "->" from CPU0 to the Memory Subsystem shows that W executes before
R, however the memory effect of a store can be observed only after the
Memory Subsystem propagates it to another CPU, as a result CPU1 doesn't
observe W before R is executed. So the original version of the comments
is correct in our memory model terminology, at least that's how I
understand it, Alan can correct me if I'm wrong.
Maybe it's better to replace the sentence starting with "The one
exception..." into:
One thing to notice is that even though a write executes by a read, the
memory effects can still be reordered from the viewpoint of a process
not holding both locks, similar to TSO ordering.
Thoughts?
Apologies for responsing late...
("Memory Subsystem" is an abstraction in our memory model, which doesn't
mean hardware implements things in the same way.).
Regards,
Boqun
@@ -0,0 +1,33 @@+C LB+unlocklockonceonce+poacquireonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, all accesses+ * in the first must execute before any accesses in the second, even if+ * the critical sections are protected by different locks.
One small nit; the above "all accesses" reads as if:
spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(t);
r1 = READ_ONCE(*y);
spin_unlock(t);
would also work, except of course that's the one reorder allowed by TSO.
I applied this series with Peter's Acked-by, and with the above comment
Thanks!
quoted
reading as follows:
+(*
+ * Result: Never
+ *
+ * If two locked critical sections execute on the same CPU, all accesses
+ * in the first must execute before any accesses in the second, even if the
+ * critical sections are protected by different locks. The one exception
+ * to this rule is that (consistent with TSO) a prior write can be reordered
+ * with a later read from the viewpoint of a process not holding both locks.
Just want to be accurate, in our memory model "execute" means a CPU
commit an memory access instruction to the Memory Subsystem, so if we
have a store W and a load R, where W executes before R, it doesn't mean
the memory effect of W is observed before the memory effect of R by
other CPUs, consider the following case
CPU0 Memory Subsystem CPU1
==== ====
WRITE_ONCE(*x,1); // W ---------->|
spin_unlock(s); |
spin_lock(t); |
r1 = READ_ONCE(*y); // R -------->|
// R reads 0 |
|<----------------WRITR_ONCE(*y, 1); // W'
W' propagates to CPU0 |
<-------------------------|
| smp_mb();
|<----------------r1 = READ_ONCE(*x); // R' reads 0
|
| W progrates to CPU 1
|----------------->
The "->" from CPU0 to the Memory Subsystem shows that W executes before
R, however the memory effect of a store can be observed only after the
Memory Subsystem propagates it to another CPU, as a result CPU1 doesn't
observe W before R is executed. So the original version of the comments
is correct in our memory model terminology, at least that's how I
understand it, Alan can correct me if I'm wrong.
Indeed, that is correct.
It is an unfortunate inconsistency with the terminology in
Documentation/memory-barriers.txt. I suspect most people think of a
write as executing when it is observed by another CPU, even though that
really isn't a coherent concept. (For example, it could easily lead
somebody to think that a write observed at different times by different
CPUs has executed more than once!)
Maybe it's better to replace the sentence starting with "The one
exception..." into:
One thing to notice is that even though a write executes by a read, the
memory effects can still be reordered from the viewpoint of a process
not holding both locks, similar to TSO ordering.
Thoughts?
Or more briefly:
Note: Even when a write executes before a read, their memory
effects can be reordered from the viewpoint of another CPU (the
kind of reordering allowed by TSO).
Alan
Apologies for responsing late...
("Memory Subsystem" is an abstraction in our memory model, which doesn't
mean hardware implements things in the same way.).
Regards,
Boqun
@@ -0,0 +1,33 @@+C LB+unlocklockonceonce+poacquireonce++(*+ * Result: Never+ *+ * If two locked critical sections execute on the same CPU, all accesses+ * in the first must execute before any accesses in the second, even if+ * the critical sections are protected by different locks.
One small nit; the above "all accesses" reads as if:
spin_lock(s);
WRITE_ONCE(*x, 1);
spin_unlock(s);
spin_lock(t);
r1 = READ_ONCE(*y);
spin_unlock(t);
would also work, except of course that's the one reorder allowed by TSO.
I applied this series with Peter's Acked-by, and with the above comment
Thanks!
quoted
reading as follows:
+(*
+ * Result: Never
+ *
+ * If two locked critical sections execute on the same CPU, all accesses
+ * in the first must execute before any accesses in the second, even if the
+ * critical sections are protected by different locks. The one exception
+ * to this rule is that (consistent with TSO) a prior write can be reordered
+ * with a later read from the viewpoint of a process not holding both locks.
Just want to be accurate, in our memory model "execute" means a CPU
commit an memory access instruction to the Memory Subsystem, so if we
have a store W and a load R, where W executes before R, it doesn't mean
the memory effect of W is observed before the memory effect of R by
other CPUs, consider the following case
CPU0 Memory Subsystem CPU1
==== ====
WRITE_ONCE(*x,1); // W ---------->|
spin_unlock(s); |
spin_lock(t); |
r1 = READ_ONCE(*y); // R -------->|
// R reads 0 |
|<----------------WRITR_ONCE(*y, 1); // W'
W' propagates to CPU0 |
<-------------------------|
| smp_mb();
|<----------------r1 = READ_ONCE(*x); // R' reads 0
|
| W progrates to CPU 1
|----------------->
The "->" from CPU0 to the Memory Subsystem shows that W executes before
R, however the memory effect of a store can be observed only after the
Memory Subsystem propagates it to another CPU, as a result CPU1 doesn't
observe W before R is executed. So the original version of the comments
is correct in our memory model terminology, at least that's how I
understand it, Alan can correct me if I'm wrong.
Indeed, that is correct.
It is an unfortunate inconsistency with the terminology in
Documentation/memory-barriers.txt. I suspect most people think of a
write as executing when it is observed by another CPU, even though that
really isn't a coherent concept. (For example, it could easily lead
somebody to think that a write observed at different times by different
CPUs has executed more than once!)
Agreed, the terminology is odd. But the fact that different CPUs can
see writes in different orders is probably always going to be a bit
counter-intuitive, so it is good to avoid giving that intuition any
support.
quoted
Maybe it's better to replace the sentence starting with "The one
exception..." into:
One thing to notice is that even though a write executes by a read, the
memory effects can still be reordered from the viewpoint of a process
not holding both locks, similar to TSO ordering.
Thoughts?
Or more briefly:
Note: Even when a write executes before a read, their memory
effects can be reordered from the viewpoint of another CPU (the
kind of reordering allowed by TSO).
Very good! I took this verbatim in a fixup patch to be combined
with the original on my next rebase.
Thanx, Paul
Alan
quoted
Apologies for responsing late...
("Memory Subsystem" is an abstraction in our memory model, which doesn't
mean hardware implements things in the same way.).
Regards,
Boqun