diff options
author | Linus Torvalds <torvalds@linux-foundation.org> | 2023-03-07 12:16:18 -0800 |
---|---|---|
committer | Linus Torvalds <torvalds@linux-foundation.org> | 2023-03-07 12:16:18 -0800 |
commit | 63355b9884b3d1677de6bd1517cd2b8a9bf53978 (patch) | |
tree | fe4de6d46a84ec554753526d546166f56a3eb72f /include/linux/cpumask.h | |
parent | 8ca09d5fa3549d142c2080a72a4c70ce389163cd (diff) |
cpumask: be more careful with 'cpumask_setall()'
Commit 596ff4a09b89 ("cpumask: re-introduce constant-sized cpumask
optimizations") changed cpumask_setall() to use "bitmap_set()" instead
of "bitmap_fill()", because bitmap_fill() would explicitly set all the
bits of a constant sized small bitmap, and that's exactly what we don't
want: we want to only set bits up to 'nr_cpu_ids', which is what
"bitmap_set()" does.
However, Yury correctly points out that while "bitmap_set()" does indeed
only set bits up to the required bitmap size, it doesn't _clear_ bits
above that size, so the upper bits would still not have well-defined
values.
Now, none of this should really matter, since any bits set past
'nr_cpu_ids' should always be ignored in the first place. Yes, the bit
scanning functions might return them as a result, but since users should
always consider the ">= nr_cpu_ids" condition to mean "no more bits",
that shouldn't have any actual effect (see previous commit 8ca09d5fa354
"cpumask: fix incorrect cpumask scanning result checks").
But let's just do it right, the way the code was _intended_ to work. We
have had enough lazy code that works but bites us in the *rse later
(again, see previous commit) that there's no reason to not just do this
properly.
It turns out that "bitmap_fill()" gets this all right for the complex
case, and really only fails for the inlined optimized case that just
fills the whole word. And while we could just fix bitmap_fill() to use
the proper last word mask, there's two issues with that:
- the cpumask case wants to do the _optimization_ based on "NR_CPUS is
a small constant", but then wants to do the actual bit _fill_ based
on "nr_cpu_ids" that isn't necessarily that same constant
- we have lots of non-cpumask users of bitmap_fill(), and while they
hopefully don't care, and probably would want the proper semantics
anyway ("only set bits up to the limit"), I do not want the cpumask
changes to impact other parts
So this ends up just doing the single-word optimization by hand in the
cpumask code. If our cpumask is fundamentally limited to a single word,
just do the proper "fill in that word" exactly. And if it's the more
complex multi-word case, then the generic bitmap_fill() will DTRT.
This is all an example of how our bitmap function optimizations really
are somewhat broken. They conflate the "this is size of the bitmap"
optimizations with the actual bit(s) we want to set.
In many cases we really want to have the two be separate things:
sometimes we base our optimizations on the size of the whole bitmap ("I
know this whole bitmap fits in a single word, so I'll just use
single-word accesses"), and sometimes we base them on the bit we are
looking at ("this is just acting on bits that are in the first word, so
I'll use single-word accesses").
Notice how the end result of the two optimizations are the same, but the
way we get to them are quite different.
And all our cpumask optimization games are really about that fundamental
distinction, and we'd often really want to pass in both the "this is the
bit I'm working on" (which _can_ be a small constant but might be
variable), and "I know it's in this range even if it's variable" (based
on CONFIG_NR_CPUS).
So this cpumask_setall() implementation just makes that explicit. It
checks the "I statically know the size is small" using the known static
size of the cpumask (which is what that 'small_cpumask_bits' is all
about), but then sets the actual bits using the exact number of cpus we
have (ie 'nr_cpumask_bits')
Of course, in a perfect world, the compiler would have done all the
range analysis (possibly with help from us just telling it that
"this value is always in this range"), and would do all of this for us.
But that is not the world we live in.
While we dream of that perfect world, this does that manual logic to
make it all work out. And this was a very long explanation for a small
code change that shouldn't even matter.
Reported-by: Yury Norov <yury.norov@gmail.com>
Link: https://lore.kernel.org/lkml/ZAV9nGG9e1%2FrV+L%2F@yury-laptop/
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'include/linux/cpumask.h')
-rw-r--r-- | include/linux/cpumask.h | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/include/linux/cpumask.h b/include/linux/cpumask.h index ce8eb7ef2107..63d637d18e79 100644 --- a/include/linux/cpumask.h +++ b/include/linux/cpumask.h @@ -518,14 +518,14 @@ static __always_inline bool cpumask_test_and_clear_cpu(int cpu, struct cpumask * /** * cpumask_setall - set all cpus (< nr_cpu_ids) in a cpumask * @dstp: the cpumask pointer - * - * Note: since we set bits, we should use the tighter 'bitmap_set()' with - * the eact number of bits, not 'bitmap_fill()' that will fill past the - * end. */ static inline void cpumask_setall(struct cpumask *dstp) { - bitmap_set(cpumask_bits(dstp), 0, nr_cpumask_bits); + if (small_const_nbits(small_cpumask_bits)) { + cpumask_bits(dstp)[0] = BITMAP_LAST_WORD_MASK(nr_cpumask_bits); + return; + } + bitmap_fill(cpumask_bits(dstp), nr_cpumask_bits); } /** |