| Message ID | 20210128130411.6125-1-vvghjk1234@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] mm/compactoin: Fix misbehaviors of fast_find_migrateblock() | expand |
On Thu, Jan 28, 2021 at 10:04:11PM +0900, Wonhyuk Yang wrote: > In the fast_find_migrateblock(), It iterate freelist to find > proper pageblock. But there are some misbehaviors. > > First, if the page we found is equal to cc->migrate_pfn, it > is considered that we didn't found suitable pageblock. Second, > if the loop was terminated because order is less than > PAGE_ALLOC_COSTLY_ORDER, it could be considered that we found > suitable one. Third, if the skip bit is set on the page block > and go continue, it doesn't check the nr_scanned. Fourth, if > the page block's skip bit is set, it check that page block is > the last of list. But it is unnecessary. > > Fixes: 70b44595eafe9 ("mm, compaction: use free lists to quickly locate > a migration source") > > Signed-off-by: Wonhyuk Yang <vvghjk1234@gmail.com> > --- > Changes in v2: > - Follow Vlastimil's suggestions, using bool varable. > - Fix new misbehavior and remove unecessary . > > v1: https://lore.kernel.org/linux-mm/20210123154320.24278-1-vvghjk1234@gmail.com > --- > mm/compaction.c | 27 ++++++++++++--------------- > 1 file changed, 12 insertions(+), 15 deletions(-) > > diff --git a/mm/compaction.c b/mm/compaction.c > index e5acb9714436..8e5cd9d20435 100644 > --- a/mm/compaction.c > +++ b/mm/compaction.c > @@ -1701,6 +1701,7 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > unsigned long pfn = cc->migrate_pfn; > unsigned long high_pfn; > int order; > + bool found_block = false; > > /* Skip hints are relied on to avoid repeats on the fast search */ > if (cc->ignore_skip_hint) > @@ -1743,7 +1744,7 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); > > for (order = cc->order - 1; > - order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit; > + order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; > order--) { > struct free_area *area = &cc->zone->free_area[order]; > struct list_head *freelist; > @@ -1758,7 +1759,11 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > list_for_each_entry(freepage, freelist, lru) { > unsigned long free_pfn; > > - nr_scanned++; > + if (nr_scanned++ >= limit) { > + move_freelist_tail(freelist, freepage); > + break; > + } > + > free_pfn = page_to_pfn(freepage); > if (free_pfn < high_pfn) { > /* As nr_scanned is post-incremented, this will still consider if the page should be used when the limit is reached. ++nr_scanned?
On Thu, Jan 28, 2021 at 10:25 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > As nr_scanned is post-incremented, this will still consider if the page > should be used when the limit is reached. ++nr_scanned? Hi, Let's consider that limit is 4, nr_scanned is 0 and loop until reach the limit. 1) pre-increment: it will search three times.(0,1,2) 2) post-increment: it will search four times. (0,1,2,3) So you mean that searching three times is correct?
On 1/28/21 2:50 PM, Wonhyuk Yang wrote: > On Thu, Jan 28, 2021 at 10:25 PM Mel Gorman <mgorman@techsingularity.net> wrote: >> >> As nr_scanned is post-incremented, this will still consider if the page >> should be used when the limit is reached. ++nr_scanned? > > Hi, > > Let's consider that limit is 4, nr_scanned is 0 and loop until reach the limit. > 1) pre-increment: it will search three times.(0,1,2) > 2) post-increment: it will search four times. (0,1,2,3) > > So you mean that searching three times is correct? 1) will match the outer loop's "nr_scanned < limit;" condition. It doesn't matter that much in practice, but for consistency sake, it should be 1)
On Thu, Jan 28, 2021 at 03:01:27PM +0100, Vlastimil Babka wrote: > On 1/28/21 2:50 PM, Wonhyuk Yang wrote: > > On Thu, Jan 28, 2021 at 10:25 PM Mel Gorman <mgorman@techsingularity.net> wrote: > >> > >> As nr_scanned is post-incremented, this will still consider if the page > >> should be used when the limit is reached. ++nr_scanned? > > > > Hi, > > > > Let's consider that limit is 4, nr_scanned is 0 and loop until reach the limit. > > 1) pre-increment: it will search three times.(0,1,2) > > 2) post-increment: it will search four times. (0,1,2,3) > > > > So you mean that searching three times is correct? > > 1) will match the outer loop's "nr_scanned < limit;" condition. It doesn't > matter that much in practice, but for consistency sake, it should be 1) Ok, while I find it a little strange to enter the loop and then break immediately due to the limit, it's a marginal difference and in general, the patch makes sense.
On Thu, Jan 28, 2021 at 11:23 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > On Thu, Jan 28, 2021 at 03:01:27PM +0100, Vlastimil Babka wrote: > > On 1/28/21 2:50 PM, Wonhyuk Yang wrote: > > > On Thu, Jan 28, 2021 at 10:25 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > >> > > >> As nr_scanned is post-incremented, this will still consider if the page > > >> should be used when the limit is reached. ++nr_scanned? > > > > > > Hi, > > > > > > Let's consider that limit is 4, nr_scanned is 0 and loop until reach the limit. > > > 1) pre-increment: it will search three times.(0,1,2) > > > 2) post-increment: it will search four times. (0,1,2,3) > > > > > > So you mean that searching three times is correct? > > > > 1) will match the outer loop's "nr_scanned < limit;" condition. It doesn't > > matter that much in practice, but for consistency sake, it should be 1) > > Ok, while I find it a little strange to enter the loop and then break > immediately due to the limit, it's a marginal difference and in general, > the patch makes sense. > Then, should I change post-increment to pre-increment? If so, should I send a new version? It seems like it's too tiny change...
On 1/28/21 2:04 PM, Wonhyuk Yang wrote: > In the fast_find_migrateblock(), It iterate freelist to find > proper pageblock. But there are some misbehaviors. > > First, if the page we found is equal to cc->migrate_pfn, it > is considered that we didn't found suitable pageblock. Second, > if the loop was terminated because order is less than > PAGE_ALLOC_COSTLY_ORDER, it could be considered that we found > suitable one. Third, if the skip bit is set on the page block > and go continue, it doesn't check the nr_scanned. Fourth, if > the page block's skip bit is set, it check that page block is > the last of list. But it is unnecessary. > > Fixes: 70b44595eafe9 ("mm, compaction: use free lists to quickly locate > a migration source") > > Signed-off-by: Wonhyuk Yang <vvghjk1234@gmail.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> > --- > Changes in v2: > - Follow Vlastimil's suggestions, using bool varable. > - Fix new misbehavior and remove unecessary . > > v1: https://lore.kernel.org/linux-mm/20210123154320.24278-1-vvghjk1234@gmail.com > --- > mm/compaction.c | 27 ++++++++++++--------------- > 1 file changed, 12 insertions(+), 15 deletions(-) > > diff --git a/mm/compaction.c b/mm/compaction.c > index e5acb9714436..8e5cd9d20435 100644 > --- a/mm/compaction.c > +++ b/mm/compaction.c > @@ -1701,6 +1701,7 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > unsigned long pfn = cc->migrate_pfn; > unsigned long high_pfn; > int order; > + bool found_block = false; > > /* Skip hints are relied on to avoid repeats on the fast search */ > if (cc->ignore_skip_hint) > @@ -1743,7 +1744,7 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); > > for (order = cc->order - 1; > - order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit; > + order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; > order--) { > struct free_area *area = &cc->zone->free_area[order]; > struct list_head *freelist; > @@ -1758,7 +1759,11 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > list_for_each_entry(freepage, freelist, lru) { > unsigned long free_pfn; > > - nr_scanned++; > + if (nr_scanned++ >= limit) { > + move_freelist_tail(freelist, freepage); > + break; > + } > + > free_pfn = page_to_pfn(freepage); > if (free_pfn < high_pfn) { > /* > @@ -1767,12 +1772,8 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > * the list assumes an entry is deleted, not > * reordered. > */ > - if (get_pageblock_skip(freepage)) { > - if (list_is_last(freelist, &freepage->lru)) > - break; > - > + if (get_pageblock_skip(freepage)) > continue; > - } > > /* Reorder to so a future search skips recent pages */ > move_freelist_tail(freelist, freepage); > @@ -1780,15 +1781,10 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > update_fast_start_pfn(cc, free_pfn); > pfn = pageblock_start_pfn(free_pfn); > cc->fast_search_fail = 0; > + found_block = true; > set_pageblock_skip(freepage); > break; > } > - > - if (nr_scanned >= limit) { > - cc->fast_search_fail++; > - move_freelist_tail(freelist, freepage); > - break; > - } > } > spin_unlock_irqrestore(&cc->zone->lock, flags); > } > @@ -1799,9 +1795,10 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) > * If fast scanning failed then use a cached entry for a page block > * that had free pages as the basis for starting a linear scan. > */ > - if (pfn == cc->migrate_pfn) > + if (!found_block) { > + cc->fast_search_fail++; > pfn = reinit_migrate_pfn(cc); > - > + } > return pfn; > } > >
On 1/28/21 3:31 PM, Wonhyuk Yang wrote: > On Thu, Jan 28, 2021 at 11:23 PM Mel Gorman <mgorman@techsingularity.net> wrote: >> >> On Thu, Jan 28, 2021 at 03:01:27PM +0100, Vlastimil Babka wrote: >> > On 1/28/21 2:50 PM, Wonhyuk Yang wrote: >> > > On Thu, Jan 28, 2021 at 10:25 PM Mel Gorman <mgorman@techsingularity.net> wrote: >> > >> >> > >> As nr_scanned is post-incremented, this will still consider if the page >> > >> should be used when the limit is reached. ++nr_scanned? >> > > >> > > Hi, >> > > >> > > Let's consider that limit is 4, nr_scanned is 0 and loop until reach the limit. >> > > 1) pre-increment: it will search three times.(0,1,2) >> > > 2) post-increment: it will search four times. (0,1,2,3) >> > > >> > > So you mean that searching three times is correct? >> > >> > 1) will match the outer loop's "nr_scanned < limit;" condition. It doesn't >> > matter that much in practice, but for consistency sake, it should be 1) Got confused there, it's 2) that matches the outer loop condition, so your patch is fine. >> Ok, while I find it a little strange to enter the loop and then break >> immediately due to the limit, it's a marginal difference and in general, >> the patch makes sense. >> > > Then, should I change post-increment to pre-increment? IMHO, nope. > If so, should I send a new version? It seems like it's too tiny change...
On Thu, Jan 28, 2021 at 11:40 PM Vlastimil Babka <vbabka@suse.cz> wrote: > > > > Then, should I change post-increment to pre-increment? > > IMHO, nope. Ok, Thanks :-)
diff --git a/mm/compaction.c b/mm/compaction.c index e5acb9714436..8e5cd9d20435 100644 --- a/mm/compaction.c +++ b/mm/compaction.c @@ -1701,6 +1701,7 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) unsigned long pfn = cc->migrate_pfn; unsigned long high_pfn; int order; + bool found_block = false; /* Skip hints are relied on to avoid repeats on the fast search */ if (cc->ignore_skip_hint) @@ -1743,7 +1744,7 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) high_pfn = pageblock_start_pfn(cc->migrate_pfn + distance); for (order = cc->order - 1; - order >= PAGE_ALLOC_COSTLY_ORDER && pfn == cc->migrate_pfn && nr_scanned < limit; + order >= PAGE_ALLOC_COSTLY_ORDER && !found_block && nr_scanned < limit; order--) { struct free_area *area = &cc->zone->free_area[order]; struct list_head *freelist; @@ -1758,7 +1759,11 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) list_for_each_entry(freepage, freelist, lru) { unsigned long free_pfn; - nr_scanned++; + if (nr_scanned++ >= limit) { + move_freelist_tail(freelist, freepage); + break; + } + free_pfn = page_to_pfn(freepage); if (free_pfn < high_pfn) { /* @@ -1767,12 +1772,8 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) * the list assumes an entry is deleted, not * reordered. */ - if (get_pageblock_skip(freepage)) { - if (list_is_last(freelist, &freepage->lru)) - break; - + if (get_pageblock_skip(freepage)) continue; - } /* Reorder to so a future search skips recent pages */ move_freelist_tail(freelist, freepage); @@ -1780,15 +1781,10 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) update_fast_start_pfn(cc, free_pfn); pfn = pageblock_start_pfn(free_pfn); cc->fast_search_fail = 0; + found_block = true; set_pageblock_skip(freepage); break; } - - if (nr_scanned >= limit) { - cc->fast_search_fail++; - move_freelist_tail(freelist, freepage); - break; - } } spin_unlock_irqrestore(&cc->zone->lock, flags); } @@ -1799,9 +1795,10 @@ static unsigned long fast_find_migrateblock(struct compact_control *cc) * If fast scanning failed then use a cached entry for a page block * that had free pages as the basis for starting a linear scan. */ - if (pfn == cc->migrate_pfn) + if (!found_block) { + cc->fast_search_fail++; pfn = reinit_migrate_pfn(cc); - + } return pfn; }
In the fast_find_migrateblock(), It iterate freelist to find proper pageblock. But there are some misbehaviors. First, if the page we found is equal to cc->migrate_pfn, it is considered that we didn't found suitable pageblock. Second, if the loop was terminated because order is less than PAGE_ALLOC_COSTLY_ORDER, it could be considered that we found suitable one. Third, if the skip bit is set on the page block and go continue, it doesn't check the nr_scanned. Fourth, if the page block's skip bit is set, it check that page block is the last of list. But it is unnecessary. Fixes: 70b44595eafe9 ("mm, compaction: use free lists to quickly locate a migration source") Signed-off-by: Wonhyuk Yang <vvghjk1234@gmail.com> --- Changes in v2: - Follow Vlastimil's suggestions, using bool varable. - Fix new misbehavior and remove unecessary . v1: https://lore.kernel.org/linux-mm/20210123154320.24278-1-vvghjk1234@gmail.com --- mm/compaction.c | 27 ++++++++++++--------------- 1 file changed, 12 insertions(+), 15 deletions(-)