Message ID | 20230519123959.77335-3-hannes@cmpxchg.org (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | mm: compaction: cleanups & simplifications | expand |
On 5/19/23 14:39, Johannes Weiner wrote: > The different branches for retry are unnecessarily complicated. There > are really only three outcomes: progress (retry n times), skipped > (retry if reclaim can help), failed (retry with higher priority). > > Rearrange the branches and the retry counter to make it simpler. > > v2: > - fix trace point build (Mel) > - fix max_retries logic for costly allocs (Huang) > > Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> > --- > mm/page_alloc.c | 53 +++++++++++++++---------------------------------- > 1 file changed, 16 insertions(+), 37 deletions(-) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 5a84a0bebc37..72660e924b95 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -3772,16 +3772,22 @@ should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, > * Compaction managed to coalesce some page blocks, but the > * allocation failed presumably due to a race. Retry some. > */ > - if (compact_result == COMPACT_SUCCESS) > - (*compaction_retries)++; > + if (compact_result == COMPACT_SUCCESS) { > + /* > + * !costly requests are much more important than > + * __GFP_RETRY_MAYFAIL costly ones because they are de > + * facto nofail and invoke OOM killer to move on while > + * costly can fail and users are ready to cope with > + * that. 1/4 retries is rather arbitrary but we would > + * need much more detailed feedback from compaction to > + * make a better decision. > + */ > + if (order > PAGE_ALLOC_COSTLY_ORDER) > + max_retries /= 4; > > - /* > - * All zones were scanned completely and still no result. It > - * doesn't really make much sense to retry except when the > - * failure could be caused by insufficient priority > - */ > - if (compact_result == COMPACT_COMPLETE) > - goto check_priority; > + ret = ++(*compaction_retries) <= max_retries; > + goto out; I think you simplified this part too much, so now once it runs out of retries, it will return false, while previously it would increase the priority. > + } > > /* > * Compaction was skipped due to a lack of free order-0 > @@ -3793,35 +3799,8 @@ should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, > } > > /* > - * If compaction backed due to being deferred, due to > - * contended locks in async mode, or due to scanners meeting > - * after a partial scan, retry with increased priority. > - */ > - if (compact_result == COMPACT_DEFERRED || > - compact_result == COMPACT_CONTENDED || > - compact_result == COMPACT_PARTIAL_SKIPPED) > - goto check_priority; > - > - /* > - * !costly requests are much more important than __GFP_RETRY_MAYFAIL > - * costly ones because they are de facto nofail and invoke OOM > - * killer to move on while costly can fail and users are ready > - * to cope with that. 1/4 retries is rather arbitrary but we > - * would need much more detailed feedback from compaction to > - * make a better decision. > - */ > - if (order > PAGE_ALLOC_COSTLY_ORDER) > - max_retries /= 4; > - if (*compaction_retries <= max_retries) { > - ret = true; > - goto out; > - } > - > - /* > - * Make sure there are attempts at the highest priority if we exhausted > - * all retries or failed at the lower priorities. > + * Compaction failed. Retry with increasing priority. > */ > -check_priority: > min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ? > MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY; >
On Mon, May 29, 2023 at 03:03:52PM +0200, Vlastimil Babka wrote: > On 5/19/23 14:39, Johannes Weiner wrote: > > The different branches for retry are unnecessarily complicated. There > > are really only three outcomes: progress (retry n times), skipped > > (retry if reclaim can help), failed (retry with higher priority). > > > > Rearrange the branches and the retry counter to make it simpler. > > > > v2: > > - fix trace point build (Mel) > > - fix max_retries logic for costly allocs (Huang) > > > > Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> > > --- > > mm/page_alloc.c | 53 +++++++++++++++---------------------------------- > > 1 file changed, 16 insertions(+), 37 deletions(-) > > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > > index 5a84a0bebc37..72660e924b95 100644 > > --- a/mm/page_alloc.c > > +++ b/mm/page_alloc.c > > @@ -3772,16 +3772,22 @@ should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, > > * Compaction managed to coalesce some page blocks, but the > > * allocation failed presumably due to a race. Retry some. > > */ > > - if (compact_result == COMPACT_SUCCESS) > > - (*compaction_retries)++; > > + if (compact_result == COMPACT_SUCCESS) { > > + /* > > + * !costly requests are much more important than > > + * __GFP_RETRY_MAYFAIL costly ones because they are de > > + * facto nofail and invoke OOM killer to move on while > > + * costly can fail and users are ready to cope with > > + * that. 1/4 retries is rather arbitrary but we would > > + * need much more detailed feedback from compaction to > > + * make a better decision. > > + */ > > + if (order > PAGE_ALLOC_COSTLY_ORDER) > > + max_retries /= 4; > > > > - /* > > - * All zones were scanned completely and still no result. It > > - * doesn't really make much sense to retry except when the > > - * failure could be caused by insufficient priority > > - */ > > - if (compact_result == COMPACT_COMPLETE) > > - goto check_priority; > > + ret = ++(*compaction_retries) <= max_retries; > > + goto out; > > I think you simplified this part too much, so now once it runs out of > retries, it will return false, while previously it would increase the priority. Oops, I'll send a delta fix to Andrew tomorrow. Thanks!
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 5a84a0bebc37..72660e924b95 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -3772,16 +3772,22 @@ should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, * Compaction managed to coalesce some page blocks, but the * allocation failed presumably due to a race. Retry some. */ - if (compact_result == COMPACT_SUCCESS) - (*compaction_retries)++; + if (compact_result == COMPACT_SUCCESS) { + /* + * !costly requests are much more important than + * __GFP_RETRY_MAYFAIL costly ones because they are de + * facto nofail and invoke OOM killer to move on while + * costly can fail and users are ready to cope with + * that. 1/4 retries is rather arbitrary but we would + * need much more detailed feedback from compaction to + * make a better decision. + */ + if (order > PAGE_ALLOC_COSTLY_ORDER) + max_retries /= 4; - /* - * All zones were scanned completely and still no result. It - * doesn't really make much sense to retry except when the - * failure could be caused by insufficient priority - */ - if (compact_result == COMPACT_COMPLETE) - goto check_priority; + ret = ++(*compaction_retries) <= max_retries; + goto out; + } /* * Compaction was skipped due to a lack of free order-0 @@ -3793,35 +3799,8 @@ should_compact_retry(struct alloc_context *ac, int order, int alloc_flags, } /* - * If compaction backed due to being deferred, due to - * contended locks in async mode, or due to scanners meeting - * after a partial scan, retry with increased priority. - */ - if (compact_result == COMPACT_DEFERRED || - compact_result == COMPACT_CONTENDED || - compact_result == COMPACT_PARTIAL_SKIPPED) - goto check_priority; - - /* - * !costly requests are much more important than __GFP_RETRY_MAYFAIL - * costly ones because they are de facto nofail and invoke OOM - * killer to move on while costly can fail and users are ready - * to cope with that. 1/4 retries is rather arbitrary but we - * would need much more detailed feedback from compaction to - * make a better decision. - */ - if (order > PAGE_ALLOC_COSTLY_ORDER) - max_retries /= 4; - if (*compaction_retries <= max_retries) { - ret = true; - goto out; - } - - /* - * Make sure there are attempts at the highest priority if we exhausted - * all retries or failed at the lower priorities. + * Compaction failed. Retry with increasing priority. */ -check_priority: min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ? MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
The different branches for retry are unnecessarily complicated. There are really only three outcomes: progress (retry n times), skipped (retry if reclaim can help), failed (retry with higher priority). Rearrange the branches and the retry counter to make it simpler. v2: - fix trace point build (Mel) - fix max_retries logic for costly allocs (Huang) Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> --- mm/page_alloc.c | 53 +++++++++++++++---------------------------------- 1 file changed, 16 insertions(+), 37 deletions(-)