RE: [f2fs-dev] [PATCH V2 2/2 RESEND] f2fs: read contiguous sit entrypages by merging for mount performance

From: Chao Yu
Date: Wed Nov 20 2013 - 22:20:30 EST

Hi,

> -----Original Message-----
> From: Jaegeuk Kim [mailto:jaegeuk.kim@xxxxxxxxxxx]
> Sent: Thursday, November 21, 2013 9:32 AM
> To: Chao Yu
> Cc: linux-fsdevel@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx; èå
> Subject: Re: [f2fs-dev] [PATCH V2 2/2 RESEND] f2fs: read contiguous sit entry pages by merging for mount performance
>
> Hi,
>
> It seems that ra_sit_pages() is too tightly coupled with
> build_sit_entries().

This code could be improved.

> Is there another way not to use *is_order?

Previously the code is like this:
-build_sit_entries()
next_setp:
for(start = 0; start < TOTAL_SEGS(sbi); start++)
/*step#1 readahead all sit entries blocks*/
if(start % SIT_ENTRY_PER_BLOCK == 0) {
blk_addr = current_sit_addr(sbi, start);
/* grab and submit_read_page */
}
if(start == TOTAL_SEGS(sbi) - 1)
f2fs_submit_read_bio();
continue;
/*step#2 fill sit entries info*/
/*step#3 cover sit entries with journal*/

But I think its weakness is that it will cost lots of memory to read
ahead all sit entry pages when f2fs mount, and also it's serious waste
that we read them again after these pages are released by VM when
out of memory.

>
> The ra_sit_pages() tries to read consecutive sit pages as many as
> possible.
> So then, what about just checking whether its block address is
> contiguous or not?
>
> Something like this:
> -ra_sit_pages()
> blkno = start;
> while (blkno < sit_i->sit_blocks) {
> blk_addr = current_sit_addr(sbi, blkno);
> if (blkno != start && prev_blk_addr + 1 != blk_addr)
> break;
>
> /* grab and submit_read_page */
>
> prev_blk_addr = blk_addr;
> blkno++;
> }

Agreed, this method could remove *order.
Shouldn't we add nrpages for readahead policy as VM?

>
> Thanks,
>
> 2013-11-20 (ì), 14:47 +0800, Chao Yu:
> > Previously we read sit entries page one by one, this method lost the chance
> > of reading contiguous page together. So we read pages as contiguous as
> > possible for better mount performance.
> >
> > v1-->v2:
> > o merge judgements/use 'Continue' or 'Break' instead of 'Goto' as Gu Zheng
> > suggested.
> > o add mark_page_accessed() before release page to delay VM reclaiming them.
> >
> > Signed-off-by: Chao Yu <chao2.yu@xxxxxxxxxxx>
> > ---
> > fs/f2fs/segment.c | 108 ++++++++++++++++++++++++++++++++++++++++-------------
> > fs/f2fs/segment.h | 2 +
> > 2 files changed, 84 insertions(+), 26 deletions(-)
> >
> > diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> > index 8149eba..52c88de 100644
> > --- a/fs/f2fs/segment.c
> > +++ b/fs/f2fs/segment.c
> > @@ -14,6 +14,7 @@
> > #include <linux/blkdev.h>
> > #include <linux/prefetch.h>
> > #include <linux/vmalloc.h>
> > +#include <linux/swap.h>
> >
> > #include "f2fs.h"
> > #include "segment.h"
> > @@ -1488,41 +1489,96 @@ static int build_curseg(struct f2fs_sb_info *sbi)
> > return restore_curseg_summaries(sbi);
> > }
> >
> > +static int ra_sit_pages(struct f2fs_sb_info *sbi, int start,
> > + int nrpages, bool *is_order)
>
> Why do you use nrpages?

nrpages point out expectation of caller, caller cloud control how many
pages want to read this time. It solves the weakness of previous coed I
give above.

>
> > +{
> > + struct address_space *mapping = sbi->meta_inode->i_mapping;
> > + struct sit_info *sit_i = SIT_I(sbi);
> > + struct page *page;
> > + block_t blk_addr;
> > + int blkno = start, readcnt = 0;
> > + int sit_blk_cnt = SIT_BLK_CNT(sbi);
> > +
> > + for (; blkno < start + nrpages && blkno < sit_blk_cnt; blkno++) {
> > +
> > + if ((!f2fs_test_bit(blkno, sit_i->sit_bitmap) ^ !*is_order)) {
> > + *is_order = !*is_order;
> > + break;
> > + }
> > +
> > + blk_addr = sit_i->sit_base_addr + blkno;
> > + if (*is_order)
> > + blk_addr += sit_i->sit_blocks;
> > +repeat:
> > + page = grab_cache_page(mapping, blk_addr);
> > + if (!page) {
> > + cond_resched();
> > + goto repeat;
> > + }
> > + if (PageUptodate(page)) {
> > + mark_page_accessed(page);
> > + f2fs_put_page(page, 1);
> > + readcnt++;
> > + continue;
> > + }
> > +
> > + submit_read_page(sbi, page, blk_addr, READ_SYNC);
> > +
> > + mark_page_accessed(page);
> > + f2fs_put_page(page, 0);
> > + readcnt++;
> > + }
> > +
> > + f2fs_submit_read_bio(sbi, READ_SYNC);
> > + return readcnt;
> > +}
> > +
> > static void build_sit_entries(struct f2fs_sb_info *sbi)
> > {
> > struct sit_info *sit_i = SIT_I(sbi);
> > struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_COLD_DATA);
> > struct f2fs_summary_block *sum = curseg->sum_blk;
> > - unsigned int start;
> > -
> > - for (start = 0; start < TOTAL_SEGS(sbi); start++) {
> > - struct seg_entry *se = &sit_i->sentries[start];
> > - struct f2fs_sit_block *sit_blk;
> > - struct f2fs_sit_entry sit;
> > - struct page *page;
> > - int i;
> > + bool is_order = f2fs_test_bit(0, sit_i->sit_bitmap) ? true : false;
> > + int sit_blk_cnt = SIT_BLK_CNT(sbi);
> > + unsigned int i, start, end;
> > + unsigned int readed, start_blk = 0;
> >
> > - mutex_lock(&curseg->curseg_mutex);
> > - for (i = 0; i < sits_in_cursum(sum); i++) {
> > - if (le32_to_cpu(segno_in_journal(sum, i)) == start) {
> > - sit = sit_in_journal(sum, i);
> > - mutex_unlock(&curseg->curseg_mutex);
> > - goto got_it;
> > + do {
> > + readed = ra_sit_pages(sbi, start_blk, sit_blk_cnt, &is_order);
> > +
> > + start = start_blk * sit_i->sents_per_block;
> > + end = (start_blk + readed) * sit_i->sents_per_block;
> > +
> > + for (; start < end && start < TOTAL_SEGS(sbi); start++) {
> > + struct seg_entry *se = &sit_i->sentries[start];
> > + struct f2fs_sit_block *sit_blk;
> > + struct f2fs_sit_entry sit;
> > + struct page *page;
> > +
> > + mutex_lock(&curseg->curseg_mutex);
> > + for (i = 0; i < sits_in_cursum(sum); i++) {
> > + if (le32_to_cpu(segno_in_journal(sum, i)) == start) {
> > + sit = sit_in_journal(sum, i);
> > + mutex_unlock(&curseg->curseg_mutex);
> > + goto got_it;
> > + }
> > }
> > - }
> > - mutex_unlock(&curseg->curseg_mutex);
> > - page = get_current_sit_page(sbi, start);
> > - sit_blk = (struct f2fs_sit_block *)page_address(page);
> > - sit = sit_blk->entries[SIT_ENTRY_OFFSET(sit_i, start)];
> > - f2fs_put_page(page, 1);
> > + mutex_unlock(&curseg->curseg_mutex);
> > +
> > + page = get_current_sit_page(sbi, start);
> > + sit_blk = (struct f2fs_sit_block *)page_address(page);
> > + sit = sit_blk->entries[SIT_ENTRY_OFFSET(sit_i, start)];
> > + f2fs_put_page(page, 1);
> > got_it:
> > - check_block_count(sbi, start, &sit);
> > - seg_info_from_raw_sit(se, &sit);
> > - if (sbi->segs_per_sec > 1) {
> > - struct sec_entry *e = get_sec_entry(sbi, start);
> > - e->valid_blocks += se->valid_blocks;
> > + check_block_count(sbi, start, &sit);
> > + seg_info_from_raw_sit(se, &sit);
> > + if (sbi->segs_per_sec > 1) {
> > + struct sec_entry *e = get_sec_entry(sbi, start);
> > + e->valid_blocks += se->valid_blocks;
> > + }
> > }
> > - }
> > + start_blk += readed;
> > + } while (start_blk < sit_blk_cnt);
> > }
> >
> > static void init_free_segmap(struct f2fs_sb_info *sbi)
> > diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> > index 269f690..ad5b9f1 100644
> > --- a/fs/f2fs/segment.h
> > +++ b/fs/f2fs/segment.h
> > @@ -83,6 +83,8 @@
> > (segno / SIT_ENTRY_PER_BLOCK)
> > #define START_SEGNO(sit_i, segno) \
> > (SIT_BLOCK_OFFSET(sit_i, segno) * SIT_ENTRY_PER_BLOCK)
> > +#define SIT_BLK_CNT(sbi) \
> > + ((TOTAL_SEGS(sbi) + SIT_ENTRY_PER_BLOCK - 1) / SIT_ENTRY_PER_BLOCK)
> > #define f2fs_bitmap_size(nr) \
> > (BITS_TO_LONGS(nr) * sizeof(unsigned long))
> > #define TOTAL_SEGS(sbi) (SM_I(sbi)->main_segments)
>
> --
> Jaegeuk Kim
> Samsung

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/