Re: [PATCH] lib/list_sort: introduce list_sort_nonatomic() and remove dummy cmp() calls

[Kuan-Wei Chiu <visitorckw@gmail.com>]

Hi Zhihao,

On Tue, Mar 17, 2026 at 09:22:26PM +0800, Zhihao Cheng wrote:
> 在 2026/3/17 20:32, Kuan-Wei Chiu 写道:
> > Hi Zhihao,
> > 
> > On Tue, Mar 17, 2026 at 12:05:54PM +0800, Zhihao Cheng wrote:
> > > 在 2026/3/16 3:39, Kuan-Wei Chiu 写道:
> > > > Historically, list_sort() implemented a hack in merge_final():
> > > >       if (unlikely(!++count))
> > > >           cmp(priv, b, b);
> > > > 
> > > > This was designed specifically so that callers could periodically
> > > > invoke cond_resched() within their comparison functions when merging
> > > > highly unbalanced lists.
> > > > 
> > > > However, an audit of the kernel tree reveals that only fs/ubifs/ relies
> > > > on this mechanism. For the vast majority of list_sort() users (such as
> > > > block layer IO schedulers and file systems), this results in completely
> > > > wasted function calls. In the worst-case scenario (merging an already
> > > > sorted list where 'a' is exhausted quickly), this results in
> > > > approximately (N/2)/256 unnecessary cmp() calls.
> > > > 
> > > > To clean up this API while ensuring behavior compatibility:
> > > > 1. Introduce list_sort_nonatomic(), which explicitly calls
> > > >      cond_resched() internally when count overflows.
> > > > 2. Remove the dummy cmp(priv, b, b) fallback for standard list_sort(),
> > > >      saving unnecessary function calls and improving determinism.
> > > > 3. Convert the sole user (fs/ubifs/) to the new API.
> > > > 
> > > > Note that ubifs still maintains cond_resched() inside its own
> > > > comparison functions. This patch does not alter the frequency or timing
> > > > of those scheduling points, guaranteeing no regressions for ubifs,
> > > > while benefiting all other kernel users.
> > > > 
> > > > Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
> > > > ---
> > > >    fs/ubifs/gc.c             |   4 +-
> > > >    fs/ubifs/replay.c         |   2 +-
> > > >    include/linux/list_sort.h |   3 +
> > > >    lib/list_sort.c           | 166 +++++++++++++++++++++-----------------
> > > >    4 files changed, 100 insertions(+), 75 deletions(-)
> > > 
> > > lgtm for UBIFS.
> > > 
> > > Reviewed-by: Zhihao Cheng <chengzhihao1@huawei.com>
> > 
> > Thanks for your review!
> > 
> > > 
> > > one small nit below.
> > > 
> > > > 
> > > [...]
> > > > diff --git a/include/linux/list_sort.h b/include/linux/list_sort.h
> > > > index 453105f74e05..f7af29073d48 100644
> > > > --- a/include/linux/list_sort.h
> > > > +++ b/include/linux/list_sort.h
> > > > @@ -11,4 +11,7 @@ typedef int __attribute__((nonnull(2,3))) (*list_cmp_func_t)(void *,
> > > >    __attribute__((nonnull(2,3)))
> > > >    void list_sort(void *priv, struct list_head *head, list_cmp_func_t cmp);
> > > > +
> > > > +__attribute__((nonnull(2, 3)))
> > > > +void list_sort_nonatomic(void *priv, struct list_head *head, list_cmp_func_t cmp);
> > > >    #endif
> > > > diff --git a/lib/list_sort.c b/lib/list_sort.c
> > > > index a310ecb7ccc0..788bfc26cf7b 100644
> > > > --- a/lib/list_sort.c
> > > > +++ b/lib/list_sort.c
> > > > @@ -3,6 +3,7 @@
> > > >    #include <linux/export.h>
> > > >    #include <linux/list_sort.h>
> > > >    #include <linux/list.h>
> > > > +#include <linux/sched.h>
> > > >    /*
> > > >     * Returns a list organized in an intermediate format suited
> > > > @@ -47,7 +48,7 @@ static struct list_head *merge(void *priv, list_cmp_func_t cmp,
> > > >     */
> > > >    __attribute__((nonnull(2,3,4,5)))
> > > >    static void merge_final(void *priv, list_cmp_func_t cmp, struct list_head *head,
> > > > -			struct list_head *a, struct list_head *b)
> > > > +			struct list_head *a, struct list_head *b, bool may_schedule)
> > > >    {
> > > >    	struct list_head *tail = head;
> > > >    	u8 count = 0;
> > > > @@ -79,12 +80,11 @@ static void merge_final(void *priv, list_cmp_func_t cmp, struct list_head *head,
> > > >    		/*
> > > >    		 * If the merge is highly unbalanced (e.g. the input is
> > > >    		 * already sorted), this loop may run many iterations.
> > > > -		 * Continue callbacks to the client even though no
> > > > -		 * element comparison is needed, so the client's cmp()
> > > > -		 * routine can invoke cond_resched() periodically.
> > > > +		 * If may_schedule is true, periodically invoke cond_resched()
> > > > +		 * to avoid soft lockups.
> > > >    		 */
> > > > -		if (unlikely(!++count))
> > > > -			cmp(priv, b, b);
> > > > +		if (may_schedule && unlikely(!++count))
> > > > +			cond_resched();
> > > The cond_resched() already has a judgment on whether to schedule out, so the
> > > 'count' could be removed?
> > 
> > However, I think keeping the u8 count rate-limiter makes more sense
> > here due to the overhead difference.
> > 
> > Evaluating unlikely(!++count) is essentially a single ALU instruction
> > (register increment) and a zero-flag check, which has virtually zero
> > cost. On the other hand, cond_resched() is a macro that does much more
> > than a simple flag check. Depending on the kernel config, it often
> > invokes __might_resched() (which reads current to check task_struct
> > states, irq flags, etc.) and makes a call to __cond_resched().
> > Evaluating all of this heavy machinery on every single iteration of
> > such a tight loop would probably introduce noticeable overhead.
> > 
> > Actually, your comment brings up another thought I wanted to discuss.
> > 
> > Since we are introducing list_sort_nonatomic(), I wonder if we should
> > eventually move the cond_resched() out of UBIFS's cmp() functions
> > entirely and handle it inside list_sort_nonatomic().
> > 
> > Right now, because the cmp() callback is inherently invoked at every
> > step of the merge process, UBIFS ends up evaluating the cond_resched()
> > macro every 3 or 4 pointer assignments during the main sort. While
> > UBIFS needs to prevent soft lockups on huge lists, checking for resched
> > at such a micro-granularity still feels excessive and likely leaves
> > performance on the table.
> 
> In my humble opinion, I don't think frequent 'cond_resched' calling will
> bring observable performance impact, and there are many examples in kernel
> hotspot paths(eg.
> blk_mq_prealloc_tag_set_tags/blk_rq_poll_completion/__blk_mq_alloc_rq_maps
> ...). For list_sort(), I prefer the aim of code cleanup is to make the code
> more readable. I am neutral on code cleanup for the current implementation
> of list_sort.

OK. Since this only affects UBIFS, if performance isn't a concern for
UBIFS, I'll drop the !++count check in v2.

Additionally, I plan to remove cond_resched() from UBIFS's cmp()
functions in v2 and move it directly into merge() and merge_final()
inside list_sort. I think this will make the code much more readable
compared to hiding the scheduling points inside the comparison
callbacks.

Regards,
Kuan-Wei