This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC][PATCH 0/5] arch: atomic rework


On Sun, Mar 02, 2014 at 04:05:52AM -0600, Peter Sewell wrote:
> On 1 March 2014 08:03, Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> > On Sat, Mar 01, 2014 at 04:06:34AM -0600, Peter Sewell wrote:
> >> Hi Paul,
> >>
> >> On 28 February 2014 18:50, Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> >> > On Thu, Feb 27, 2014 at 12:53:12PM -0800, Paul E. McKenney wrote:
> >> >> On Thu, Feb 27, 2014 at 11:47:08AM -0800, Linus Torvalds wrote:
> >> >> > On Thu, Feb 27, 2014 at 11:06 AM, Paul E. McKenney
> >> >> > <paulmck@linux.vnet.ibm.com> wrote:
> >> >> > >
> >> >> > > 3.      The comparison was against another RCU-protected pointer,
> >> >> > >         where that other pointer was properly fetched using one
> >> >> > >         of the RCU primitives.  Here it doesn't matter which pointer
> >> >> > >         you use.  At least as long as the rcu_assign_pointer() for
> >> >> > >         that other pointer happened after the last update to the
> >> >> > >         pointed-to structure.
> >> >> > >
> >> >> > > I am a bit nervous about #3.  Any thoughts on it?
> >> >> >
> >> >> > I think that it might be worth pointing out as an example, and saying
> >> >> > that code like
> >> >> >
> >> >> >    p = atomic_read(consume);
> >> >> >    X;
> >> >> >    q = atomic_read(consume);
> >> >> >    Y;
> >> >> >    if (p == q)
> >> >> >         data = p->val;
> >> >> >
> >> >> > then the access of "p->val" is constrained to be data-dependent on
> >> >> > *either* p or q, but you can't really tell which, since the compiler
> >> >> > can decide that the values are interchangeable.
> >> >> >
> >> >> > I cannot for the life of me come up with a situation where this would
> >> >> > matter, though. If "X" contains a fence, then that fence will be a
> >> >> > stronger ordering than anything the consume through "p" would
> >> >> > guarantee anyway. And if "X" does *not* contain a fence, then the
> >> >> > atomic reads of p and q are unordered *anyway*, so then whether the
> >> >> > ordering to the access through "p" is through p or q is kind of
> >> >> > irrelevant. No?
> >> >>
> >> >> I can make a contrived litmus test for it, but you are right, the only
> >> >> time you can see it happen is when X has no barriers, in which case
> >> >> you don't have any ordering anyway -- both the compiler and the CPU can
> >> >> reorder the loads into p and q, and the read from p->val can, as you say,
> >> >> come from either pointer.
> >> >>
> >> >> For whatever it is worth, hear is the litmus test:
> >> >>
> >> >> T1:   p = kmalloc(...);
> >> >>       if (p == NULL)
> >> >>               deal_with_it();
> >> >>       p->a = 42;  /* Each field in its own cache line. */
> >> >>       p->b = 43;
> >> >>       p->c = 44;
> >> >>       atomic_store_explicit(&gp1, p, memory_order_release);
> >> >>       p->b = 143;
> >> >>       p->c = 144;
> >> >>       atomic_store_explicit(&gp2, p, memory_order_release);
> >> >>
> >> >> T2:   p = atomic_load_explicit(&gp2, memory_order_consume);
> >> >>       r1 = p->b;  /* Guaranteed to get 143. */
> >> >>       q = atomic_load_explicit(&gp1, memory_order_consume);
> >> >>       if (p == q) {
> >> >>               /* The compiler decides that q->c is same as p->c. */
> >> >>               r2 = p->c; /* Could get 44 on weakly order system. */
> >> >>       }
> >> >>
> >> >> The loads from gp1 and gp2 are, as you say, unordered, so you get what
> >> >> you get.
> >> >>
> >> >> And publishing a structure via one RCU-protected pointer, updating it,
> >> >> then publishing it via another pointer seems to me to be asking for
> >> >> trouble anyway.  If you really want to do something like that and still
> >> >> see consistency across all the fields in the structure, please put a lock
> >> >> in the structure and use it to guard updates and accesses to those fields.
> >> >
> >> > And here is a patch documenting the restrictions for the current Linux
> >> > kernel.  The rules change a bit due to rcu_dereference() acting a bit
> >> > differently than atomic_load_explicit(&p, memory_order_consume).
> >> >
> >> > Thoughts?
> >>
> >> That might serve as informal documentation for linux kernel
> >> programmers about the bounds on the optimisations that you expect
> >> compilers to do for common-case RCU code - and I guess that's what you
> >> intend it to be for.   But I don't see how one can make it precise
> >> enough to serve as a language definition, so that compiler people
> >> could confidently say "yes, we respect that", which I guess is what
> >> you really need.  As a useful criterion, we should aim for something
> >> precise enough that in a verified-compiler context you can
> >> mathematically prove that the compiler will satisfy it  (even though
> >> that won't happen anytime soon for GCC), and that analysis tool
> >> authors can actually know what they're working with.   All this stuff > >> about "you should avoid cancellation", and "avoid masking with just a
> >> small number of bits" is just too vague.
> >
> > Understood, and yes, this is intended to document current compiler
> > behavior for the Linux kernel community.  It would not make sense to show
> > it to the C11 or C++11 communities, except perhaps as an informational
> > piece on current practice.
> >
> >> The basic problem is that the compiler may be doing sophisticated
> >> reasoning with a bunch of non-local knowledge that it's deduced from
> >> the code, neither of which are well-understood, and here we have to
> >> identify some envelope, expressive enough for RCU idioms, in which
> >> that reasoning doesn't allow data/address dependencies to be removed
> >> (and hence the hardware guarantee about them will be maintained at the
> >> source level).
> >>
> >> The C11 syntactic notion of dependency, whatever its faults, was at
> >> least precise, could be reasoned about locally (just looking at the
> >> syntactic code in question), and did do that.  The fact that current
> >> compilers do optimisations that remove dependencies and will likely
> >> have many bugs at present is besides the point - this was surely
> >> intended as a *new* constraint on what they are allowed to do.  The
> >> interesting question is really whether the compiler writers think that
> >> they *could* implement it in a reasonable way - I'd like to hear
> >> Torvald and his colleagues' opinion on that.
> >>
> >> What you're doing above seems to be basically a very cut-down version
> >> of that, but with a fuzzy boundary.   If you want it to be precise,
> >> maybe it needs to be much simpler (which might force you into ruling
> >> out some current code idioms).
> >
> > I hope that Torvald Riegel's proposal (https://lkml.org/lkml/2014/2/27/806)
> > can be developed to serve this purpose.
> 
> (I missed that mail when it first came past, sorry)

No worries!

> That's also going to be tricky, I'm afraid.  The key condition there is:
> 
> "* at the time of execution of E, L       [PS: I assume that L is a
> typo and should be E]

I believe it really is "L".  As I understand it (and Torvald will correct
me if I am wrong), the idea is that the implementation is prohibited
from guessing the value of "L" -- it must assume that any value from
L's type might be returned, regardless of what it might otherwise know.

However, after L's value is loaded, the implementation -is- permitted
to learn constraint's on this value based on "if" statements and the
like between the load from "L" and the execution of "E".

Does that help?

>         can possibly have returned at
>         least two different values under the assumption that L itself
>         could have returned any value allowed by L's type."
> 
> First, the evaluation of E might be nondeterministic  - e.g., for an
> artificial example, if it's just a nondeterministic value obtained
> from the result of a race on SC atomics.  The above doesn't
> distinguish between that (which doesn't have a real dependency on L)
> and that XOR'd with L (which does).  And it does so in the wrong
> direction: it'll say there the former has a dependency on L.

Right, it is only any dependency that E has on L that would be
constrained.  If E also depends on other quantities obtained some
other way than a memory_order_consume load into a value_dep_preserving,
variable, then as I understand it, the compiler is within its rights
to optimize these other quantities to within an inch of their lives.

It is quite possible that E depends on L only sometimes.  For example:

	p = atomic_load_explicit(&gp, memory_order_consume);
	p = random() & 0x8 ? p : &default_structure;
	E(p);

My guess is that in this case, the ordering would be guaranteed only
for those executions where there is a value dependency.  In my naive
view, this should be no different than something like this:

	if (random() & 0x10)
		p = atomic_load_explicit(&gp, memory_order_acquire);
	else
		p = &default_structure;
	E(p);

Or am I missing your point?

> Second, it involves reasoning about counterfactual executions.  That
> doesn't necessarily make it wrong, per se, but probably makes it hard
> to work with.  For example, suppose that in all the actual
> whole-program executions, a runtime occurrence of L only ever returns
> one particular value (perhaps because of some simple #define'd
> configuration), and that the code used in the evaluation of E depends
> on some invariant which is related to that configuration.  The
> hypothetical execution used above in which a different value is used
> is one in the code is being run in a situation with broken invariants.
>    Then there will be technical difficulties in using the definition:
> I don't see how one would persuade oneself that a compiler always
> satisfies it, because these hypothetical executions are far removed
> from what it's actually working on.

The developer answer would be something like "all it really means is that
the implementation is required to actually emit the memory_order_consume
load and actually use the value," which is probably not much comfort
to someone trying to model it.  Maybe there is a better way of wording
this constraint so as to avoid the counterfactuals?

> (Aside: The notion of a thread "observing" another thread's load,
> dating back a long time and adopted in the Power and ARM architecture
> texts, relies on counterfactual executions in a broadly similar way;
> we're happy to have escaped that now :-)

Here is hoping that there is a way to escape it in this case as well.  ;-)

							Thanx, Paul

> Peter
> 
> 
> 
> 
> >                                                         Thanx, Paul
> >
> >> best,
> >> Peter
> >>
> >>
> >>
> >> >                                                         Thanx, Paul
> >> >
> >> > ------------------------------------------------------------------------
> >> >
> >> > documentation: Record rcu_dereference() value mishandling
> >> >
> >> > Recent LKML discussings (see http://lwn.net/Articles/586838/ and
> >> > http://lwn.net/Articles/588300/ for the LWN writeups) brought out
> >> > some ways of misusing the return value from rcu_dereference() that
> >> > are not necessarily completely intuitive.  This commit therefore
> >> > documents what can and cannot safely be done with these values.
> >> >
> >> > Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> >> >
> >> > diff --git a/Documentation/RCU/00-INDEX b/Documentation/RCU/00-INDEX
> >> > index fa57139f50bf..f773a264ae02 100644
> >> > --- a/Documentation/RCU/00-INDEX
> >> > +++ b/Documentation/RCU/00-INDEX
> >> > @@ -12,6 +12,8 @@ lockdep-splat.txt
> >> >         - RCU Lockdep splats explained.
> >> >  NMI-RCU.txt
> >> >         - Using RCU to Protect Dynamic NMI Handlers
> >> > +rcu_dereference.txt
> >> > +       - Proper care and feeding of return values from rcu_dereference()
> >> >  rcubarrier.txt
> >> >         - RCU and Unloadable Modules
> >> >  rculist_nulls.txt
> >> > diff --git a/Documentation/RCU/checklist.txt b/Documentation/RCU/checklist.txt
> >> > index 9d10d1db16a5..877947130ebe 100644
> >> > --- a/Documentation/RCU/checklist.txt
> >> > +++ b/Documentation/RCU/checklist.txt
> >> > @@ -114,12 +114,16 @@ over a rather long period of time, but improvements are always welcome!
> >> >                         http://www.openvms.compaq.com/wizard/wiz_2637.html
> >> >
> >> >                 The rcu_dereference() primitive is also an excellent
> >> > -               documentation aid, letting the person reading the code
> >> > -               know exactly which pointers are protected by RCU.
> >> > +               documentation aid, letting the person reading the
> >> > +               code know exactly which pointers are protected by RCU.
> >> >                 Please note that compilers can also reorder code, and
> >> >                 they are becoming increasingly aggressive about doing
> >> > -               just that.  The rcu_dereference() primitive therefore
> >> > -               also prevents destructive compiler optimizations.
> >> > +               just that.  The rcu_dereference() primitive therefore also
> >> > +               prevents destructive compiler optimizations.  However,
> >> > +               with a bit of devious creativity, it is possible to
> >> > +               mishandle the return value from rcu_dereference().
> >> > +               Please see rcu_dereference.txt in this directory for
> >> > +               more information.
> >> >
> >> >                 The rcu_dereference() primitive is used by the
> >> >                 various "_rcu()" list-traversal primitives, such
> >> > diff --git a/Documentation/RCU/rcu_dereference.txt b/Documentation/RCU/rcu_dereference.txt
> >> > new file mode 100644
> >> > index 000000000000..6e72cd8622df
> >> > --- /dev/null
> >> > +++ b/Documentation/RCU/rcu_dereference.txt
> >> > @@ -0,0 +1,365 @@
> >> > +PROPER CARE AND FEEDING OF RETURN VALUES FROM rcu_dereference()
> >> > +
> >> > +Most of the time, you can use values from rcu_dereference() or one of
> >> > +the similar primitives without worries.  Dereferencing (prefix "*"),
> >> > +field selection ("->"), assignment ("="), address-of ("&"), addition and
> >> > +subtraction of constants, and casts all work quite naturally and safely.
> >> > +
> >> > +It is nevertheless possible to get into trouble with other operations.
> >> > +Follow these rules to keep your RCU code working properly:
> >> > +
> >> > +o      You must use one of the rcu_dereference() family of primitives
> >> > +       to load an RCU-protected pointer, otherwise CONFIG_PROVE_RCU
> >> > +       will complain.  Worse yet, your code can see random memory-corruption
> >> > +       bugs due to games that compilers and DEC Alpha can play.
> >> > +       Without one of the rcu_dereference() primitives, compilers
> >> > +       can reload the value, and won't your code have fun with two
> >> > +       different values for a single pointer!  Without rcu_dereference(),
> >> > +       DEC Alpha can load a pointer, dereference that pointer, and
> >> > +       return data preceding initialization that preceded the store of
> >> > +       the pointer.
> >> > +
> >> > +       In addition, the volatile cast in rcu_dereference() prevents the
> >> > +       compiler from deducing the resulting pointer value.  Please see
> >> > +       the section entitled "EXAMPLE WHERE THE COMPILER KNOWS TOO MUCH"
> >> > +       for an example where the compiler can in fact deduce the exact
> >> > +       value of the pointer, and thus cause misordering.
> >> > +
> >> > +o      Do not use single-element RCU-protected arrays.  The compiler
> >> > +       is within its right to assume that the value of an index into
> >> > +       such an array must necessarily evaluate to zero.  The compiler
> >> > +       could then substitute the constant zero for the computation, so
> >> > +       that the array index no longer depended on the value returned
> >> > +       by rcu_dereference().  If the array index no longer depends
> >> > +       on rcu_dereference(), then both the compiler and the CPU
> >> > +       are within their rights to order the array access before the
> >> > +       rcu_dereference(), which can cause the array access to return
> >> > +       garbage.
> >> > +
> >> > +o      Avoid cancellation when using the "+" and "-" infix arithmetic
> >> > +       operators.  For example, for a given variable "x", avoid
> >> > +       "(x-x)".  There are similar arithmetic pitfalls from other
> >> > +       arithmetic operatiors, such as "(x*0)", "(x/(x+1))" or "(x%1)".
> >> > +       The compiler is within its rights to substitute zero for all of
> >> > +       these expressions, so that subsequent accesses no longer depend
> >> > +       on the rcu_dereference(), again possibly resulting in bugs due
> >> > +       to misordering.
> >> > +
> >> > +       Of course, if "p" is a pointer from rcu_dereference(), and "a"
> >> > +       and "b" are integers that happen to be equal, the expression
> >> > +       "p+a-b" is safe because its value still necessarily depends on
> >> > +       the rcu_dereference(), thus maintaining proper ordering.
> >> > +
> >> > +o      Avoid all-zero operands to the bitwise "&" operator, and
> >> > +       similarly avoid all-ones operands to the bitwise "|" operator.
> >> > +       If the compiler is able to deduce the value of such operands,
> >> > +       it is within its rights to substitute the corresponding constant
> >> > +       for the bitwise operation.  Once again, this causes subsequent
> >> > +       accesses to no longer depend on the rcu_dereference(), causing
> >> > +       bugs due to misordering.
> >> > +
> >> > +       Please note that single-bit operands to bitwise "&" can also
> >> > +       be dangerous.  At this point, the compiler knows that the
> >> > +       resulting value can only take on one of two possible values.
> >> > +       Therefore, a very small amount of additional information will
> >> > +       allow the compiler to deduce the exact value, which again can
> >> > +       result in misordering.
> >> > +
> >> > +o      If you are using RCU to protect JITed functions, so that the
> >> > +       "()" function-invocation operator is applied to a value obtained
> >> > +       (directly or indirectly) from rcu_dereference(), you may need to
> >> > +       interact directly with the hardware to flush instruction caches.
> >> > +       This issue arises on some systems when a newly JITed function is
> >> > +       using the same memory that was used by an earlier JITed function.
> >> > +
> >> > +o      Do not use the results from the boolean "&&" and "||" when
> >> > +       dereferencing.  For example, the following (rather improbable)
> >> > +       code is buggy:
> >> > +
> >> > +               int a[2];
> >> > +               int index;
> >> > +               int force_zero_index = 1;
> >> > +
> >> > +               ...
> >> > +
> >> > +               r1 = rcu_dereference(i1)
> >> > +               r2 = a[r1 && force_zero_index];  /* BUGGY!!! */
> >> > +
> >> > +       The reason this is buggy is that "&&" and "||" are often compiled
> >> > +       using branches.  While weak-memory machines such as ARM or PowerPC
> >> > +       do order stores after such branches, they can speculate loads,
> >> > +       which can result in misordering bugs.
> >> > +
> >> > +o      Do not use the results from relational operators ("==", "!=",
> >> > +       ">", ">=", "<", or "<=") when dereferencing.  For example,
> >> > +       the following (quite strange) code is buggy:
> >> > +
> >> > +               int a[2];
> >> > +               int index;
> >> > +               int flip_index = 0;
> >> > +
> >> > +               ...
> >> > +
> >> > +               r1 = rcu_dereference(i1)
> >> > +               r2 = a[r1 != flip_index];  /* BUGGY!!! */
> >> > +
> >> > +       As before, the reason this is buggy is that relational operators
> >> > +       are often compiled using branches.  And as before, although
> >> > +       weak-memory machines such as ARM or PowerPC do order stores
> >> > +       after such branches, but can speculate loads, which can again
> >> > +       result in misordering bugs.
> >> > +
> >> > +o      Be very careful about comparing pointers obtained from
> >> > +       rcu_dereference() against non-NULL values.  As Linus Torvalds
> >> > +       explained, if the two pointers are equal, the compiler could
> >> > +       substitute the pointer you are comparing against for the pointer
> >> > +       obtained from rcu_dereference().  For example:
> >> > +
> >> > +               p = rcu_dereference(gp);
> >> > +               if (p == &default_struct)
> >> > +                       do_default(p->a);
> >> > +
> >> > +       Because the compiler now knows that the value of "p" is exactly
> >> > +       the address of the variable "default_struct", it is free to
> >> > +       transform this code into the following:
> >> > +
> >> > +               p = rcu_dereference(gp);
> >> > +               if (p == &default_struct)
> >> > +                       do_default(default_struct.a);
> >> > +
> >> > +       On ARM and Power hardware, the load from "default_struct.a"
> >> > +       can now be speculated, such that it might happen before the
> >> > +       rcu_dereference().  This could result in bugs due to misordering.
> >> > +
> >> > +       However, comparisons are OK in the following cases:
> >> > +
> >> > +       o       The comparison was against the NULL pointer.  If the
> >> > +               compiler knows that the pointer is NULL, you had better
> >> > +               not be dereferencing it anyway.  If the comparison is
> >> > +               non-equal, the compiler is none the wiser.  Therefore,
> >> > +               it is safe to compare pointers from rcu_dereference()
> >> > +               against NULL pointers.
> >> > +
> >> > +       o       The pointer is never dereferenced after being compared.
> >> > +               Since there are no subsequent dereferences, the compiler
> >> > +               cannot use anything it learned from the comparison
> >> > +               to reorder the non-existent subsequent dereferences.
> >> > +               This sort of comparison occurs frequently when scanning
> >> > +               RCU-protected circular linked lists.
> >> > +
> >> > +       o       The comparison is against a pointer pointer that
> >> > +               references memory that was initialized "a long time ago."
> >> > +               The reason this is safe is that even if misordering
> >> > +               occurs, the misordering will not affect the accesses
> >> > +               that follow the comparison.  So exactly how long ago is
> >> > +               "a long time ago"?  Here are some possibilities:
> >> > +
> >> > +               o       Compile time.
> >> > +
> >> > +               o       Boot time.
> >> > +
> >> > +               o       Module-init time for module code.
> >> > +
> >> > +               o       Prior to kthread creation for kthread code.
> >> > +
> >> > +               o       During some prior acquisition of the lock that
> >> > +                       we now hold.
> >> > +
> >> > +               o       Before mod_timer() time for a timer handler.
> >> > +
> >> > +               There are many other possibilities involving the Linux
> >> > +               kernel's wide array of primitives that cause code to
> >> > +               be invoked at a later time.
> >> > +
> >> > +       o       The pointer being compared against also came from
> >> > +               rcu_dereference().  In this case, both pointers depend
> >> > +               on one rcu_dereference() or another, so you get proper
> >> > +               ordering either way.
> >> > +
> >> > +               That said, this situation can make certain RCU usage
> >> > +               bugs more likely to happen.  Which can be a good thing,
> >> > +               at least if they happen during testing.  An example
> >> > +               of such an RCU usage bug is shown in the section titled
> >> > +               "EXAMPLE OF AMPLIFIED RCU-USAGE BUG".
> >> > +
> >> > +       o       All of the accesses following the comparison are stores,
> >> > +               so that a control dependency preserves the needed ordering.
> >> > +               That said, it is easy to get control dependencies wrong.
> >> > +               Please see the "CONTROL DEPENDENCIES" section of
> >> > +               Documentation/memory-barriers.txt for more details.
> >> > +
> >> > +       o       The pointers compared not-equal -and- the compiler does
> >> > +               not have enough information to deduce the value of the
> >> > +               pointer.  Note that the volatile cast in rcu_dereference()
> >> > +               will normally prevent the compiler from knowing too much.
> >> > +
> >> > +o      Disable any value-speculation optimizations that your compiler
> >> > +       might provide, especially if you are making use of feedback-based
> >> > +       optimizations that take data collected from prior runs.  Such
> >> > +       value-speculation optimizations reorder operations by design.
> >> > +
> >> > +       There is one exception to this rule:  Value-speculation
> >> > +       optimizations that leverage the branch-prediction hardware are
> >> > +       safe on strongly ordered systems (such as x86), but not on weakly
> >> > +       ordered systems (such as ARM or Power).  Choose your compiler
> >> > +       command-line options wisely!
> >> > +
> >> > +
> >> > +EXAMPLE OF AMPLIFIED RCU-USAGE BUG
> >> > +
> >> > +Because updaters can run concurrently with RCU readers, RCU readers can
> >> > +see stale and/or inconsistent values.  If RCU readers need fresh or
> >> > +consistent values, which they sometimes do, they need to take proper
> >> > +precautions.  To see this, consider the following code fragment:
> >> > +
> >> > +       struct foo {
> >> > +               int a;
> >> > +               int b;
> >> > +               int c;
> >> > +       };
> >> > +       struct foo *gp1;
> >> > +       struct foo *gp2;
> >> > +
> >> > +       void updater(void)
> >> > +       {
> >> > +               struct foo *p;
> >> > +
> >> > +               p = kmalloc(...);
> >> > +               if (p == NULL)
> >> > +                       deal_with_it();
> >> > +               p->a = 42;  /* Each field in its own cache line. */
> >> > +               p->b = 43;
> >> > +               p->c = 44;
> >> > +               rcu_assign_pointer(gp1, p);
> >> > +               p->b = 143;
> >> > +               p->c = 144;
> >> > +               rcu_assign_pointer(gp2, p);
> >> > +       }
> >> > +
> >> > +       void reader(void)
> >> > +       {
> >> > +               struct foo *p;
> >> > +               struct foo *q;
> >> > +               int r1, r2;
> >> > +
> >> > +               p = rcu_dereference(gp2);
> >> > +               r1 = p->b;  /* Guaranteed to get 143. */
> >> > +               q = rcu_dereference(gp1);
> >> > +               if (p == q) {
> >> > +                       /* The compiler decides that q->c is same as p->c. */
> >> > +                       r2 = p->c; /* Could get 44 on weakly order system. */
> >> > +               }
> >> > +       }
> >> > +
> >> > +You might be surprised that the outcome (r1 == 143 && r2 == 44) is possible,
> >> > +but you should not be.  After all, the updater might have been invoked
> >> > +a second time between the time reader() loaded into "r1" and the time
> >> > +that it loaded into "r2".  The fact that this same result can occur due
> >> > +to some reordering from the compiler and CPUs is beside the point.
> >> > +
> >> > +But suppose that the reader needs a consistent view?
> >> > +
> >> > +Then one approach is to use locking, for example, as follows:
> >> > +
> >> > +       struct foo {
> >> > +               int a;
> >> > +               int b;
> >> > +               int c;
> >> > +               spinlock_t lock;
> >> > +       };
> >> > +       struct foo *gp1;
> >> > +       struct foo *gp2;
> >> > +
> >> > +       void updater(void)
> >> > +       {
> >> > +               struct foo *p;
> >> > +
> >> > +               p = kmalloc(...);
> >> > +               if (p == NULL)
> >> > +                       deal_with_it();
> >> > +               spin_lock(&p->lock);
> >> > +               p->a = 42;  /* Each field in its own cache line. */
> >> > +               p->b = 43;
> >> > +               p->c = 44;
> >> > +               spin_unlock(&p->lock);
> >> > +               rcu_assign_pointer(gp1, p);
> >> > +               spin_lock(&p->lock);
> >> > +               p->b = 143;
> >> > +               p->c = 144;
> >> > +               spin_unlock(&p->lock);
> >> > +               rcu_assign_pointer(gp2, p);
> >> > +       }
> >> > +
> >> > +       void reader(void)
> >> > +       {
> >> > +               struct foo *p;
> >> > +               struct foo *q;
> >> > +               int r1, r2;
> >> > +
> >> > +               p = rcu_dereference(gp2);
> >> > +               spin_lock(&p->lock);
> >> > +               r1 = p->b;  /* Guaranteed to get 143. */
> >> > +               q = rcu_dereference(gp1);
> >> > +               if (p == q) {
> >> > +                       /* The compiler decides that q->c is same as p->c. */
> >> > +                       r2 = p->c; /* Could get 44 on weakly order system. */
> >> > +               }
> >> > +               spin_unlock(&p->lock);
> >> > +       }
> >> > +
> >> > +As always, use the right tool for the job!
> >> > +
> >> > +
> >> > +EXAMPLE WHERE THE COMPILER KNOWS TOO MUCH
> >> > +
> >> > +If a pointer obtained from rcu_dereference() compares not-equal to some
> >> > +other pointer, the compiler normally has no clue what the value of the
> >> > +first pointer might be.  This lack of knowledge prevents the compiler
> >> > +from carrying out optimizations that otherwise might destroy the ordering
> >> > +guarantees that RCU depends on.  And the volatile cast in rcu_dereference()
> >> > +should prevent the compiler from guessing the value.
> >> > +
> >> > +But without rcu_dereference(), the compiler knows more than you might
> >> > +expect.  Consider the following code fragment:
> >> > +
> >> > +       struct foo {
> >> > +               int a;
> >> > +               int b;
> >> > +       };
> >> > +       static struct foo variable1;
> >> > +       static struct foo variable2;
> >> > +       static struct foo *gp = &variable1;
> >> > +
> >> > +       void updater(void)
> >> > +       {
> >> > +               initialize_foo(&variable2);
> >> > +               rcu_assign_pointer(gp, &variable2);
> >> > +               /*
> >> > +                * The above is the only store to gp in this translation unit,
> >> > +                * and the address of gp is not exported in any way.
> >> > +                */
> >> > +       }
> >> > +
> >> > +       int reader(void)
> >> > +       {
> >> > +               struct foo *p;
> >> > +
> >> > +               p = gp;
> >> > +               barrier();
> >> > +               if (p == &variable1)
> >> > +                       return p->a; /* Must be variable1.a. */
> >> > +               else
> >> > +                       return p->b; /* Must be variable2.b. */
> >> > +       }
> >> > +
> >> > +Because the compiler can see all stores to "gp", it knows that the only
> >> > +possible values of "gp" are "variable1" on the one hand and "variable2"
> >> > +on the other.  The comparison in reader() therefore tells the compiler
> >> > +the exact value of "p" even in the not-equals case.  This allows the
> >> > +compiler to make the return values independent of the load from "gp",
> >> > +in turn destroying the ordering between this load and the loads of the
> >> > +return values.  This can result in "p->b" returning pre-initialization
> >> > +garbage values.
> >> > +
> >> > +In short, rcu_dereference() is -not- optional when you are going to
> >> > +dereference the resulting pointer.
> >> >
> >> > --
> >> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> >> > the body of a message to majordomo@vger.kernel.org
> >> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> >> > Please read the FAQ at  http://www.tux.org/lkml/
> >>
> >
> 


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]