* [PATCH 1/2] block: introduce content activity based ioprio
2024-02-21 7:53 [PATCH 0/2] *** introduce content activity based ioprio *** zhaoyang.huang
@ 2024-02-21 7:53 ` zhaoyang.huang
2024-02-21 12:53 ` Matthew Wilcox
2024-02-21 7:53 ` [PATCH 2/2] fs: " zhaoyang.huang
1 sibling, 1 reply; 4+ messages in thread
From: zhaoyang.huang @ 2024-02-21 7:53 UTC (permalink / raw)
To: Jens Axboe, Darrick J . Wong, Alexander Viro, Christian Brauner,
linux-block, linux-fsdevel, linux-kernel, Zhaoyang Huang,
steve.kang
From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
Currently, request's ioprio are set via task's schedule priority(when no
blkcg configured), which has high priority tasks possess the privilege on
both of CPU and IO scheduling. Furthermore, most of the write requestes
are launched asynchronosly from kworker which can't know the submitter's
priorities.
This commit works as a hint of original policy by promoting the request
ioprio based on the page/folio's activity. The original idea comes from
LRU_GEN which provides more precised folio activity than before. This
commit try to adjust the request's ioprio when certain part of its folios
are hot, which indicate that this request carry important contents and
need be scheduled ealier.
The filesystem should call bio_set_active_ioprio_folio() after
calling bio_add_folio. Please be noted that this set of API can not
handle bvec_try_merge_page cases.
This commit is verified on a v6.6 6GB RAM android14 system via 4 test cases
by calling bio_set_active_ioprio in erofs, ext4, f2fs and blkdev(raw
partition of gendisk)
Case 1:
script[a] which get significant improved fault time as expected[b]*
where dd's cost also shrink from 55s to 40s.
(1). fault_latency.bin is an ebpf based test tool which measure all task's
iowait latency during page fault when scheduled out/in.
(2). costmem generate page fault by mmaping a file and access the VA.
(3). dd generate concurrent vfs io.
[a]
./fault_latency.bin 1 5 > /data/dd_costmem &
costmem -c0 -a2048000 -b128000 -o0 1>/dev/null &
costmem -c0 -a2048000 -b128000 -o0 1>/dev/null &
costmem -c0 -a2048000 -b128000 -o0 1>/dev/null &
costmem -c0 -a2048000 -b128000 -o0 1>/dev/null &
dd if=/dev/block/sda of=/data/ddtest bs=1024 count=2048000 &
dd if=/dev/block/sda of=/data/ddtest1 bs=1024 count=2048000 &
dd if=/dev/block/sda of=/data/ddtest2 bs=1024 count=2048000 &
dd if=/dev/block/sda of=/data/ddtest3 bs=1024 count=2048000
[b]
mainline commit
io wait 736us 523us
* provide correct result for test case 1 in v7 which was compared between
EMMC and UFS wrongly.
Case 2:
fio -filename=/dev/block/by-name/userdata -rw=randread -direct=0 -bs=4k -size=2000M -numjobs=8 -group_reporting -name=mytest
mainline: 513MiB/s
READ: bw=531MiB/s (557MB/s), 531MiB/s-531MiB/s (557MB/s-557MB/s), io=15.6GiB (16.8GB), run=30137-30137msec
READ: bw=543MiB/s (569MB/s), 543MiB/s-543MiB/s (569MB/s-569MB/s), io=15.6GiB (16.8GB), run=29469-29469msec
READ: bw=474MiB/s (497MB/s), 474MiB/s-474MiB/s (497MB/s-497MB/s), io=15.6GiB (16.8GB), run=33724-33724msec
READ: bw=535MiB/s (561MB/s), 535MiB/s-535MiB/s (561MB/s-561MB/s), io=15.6GiB (16.8GB), run=29928-29928msec
READ: bw=523MiB/s (548MB/s), 523MiB/s-523MiB/s (548MB/s-548MB/s), io=15.6GiB (16.8GB), run=30617-30617msec
READ: bw=492MiB/s (516MB/s), 492MiB/s-492MiB/s (516MB/s-516MB/s), io=15.6GiB (16.8GB), run=32518-32518msec
READ: bw=533MiB/s (559MB/s), 533MiB/s-533MiB/s (559MB/s-559MB/s), io=15.6GiB (16.8GB), run=29993-29993msec
READ: bw=524MiB/s (550MB/s), 524MiB/s-524MiB/s (550MB/s-550MB/s), io=15.6GiB (16.8GB), run=30526-30526msec
READ: bw=529MiB/s (554MB/s), 529MiB/s-529MiB/s (554MB/s-554MB/s), io=15.6GiB (16.8GB), run=30269-30269msec
READ: bw=449MiB/s (471MB/s), 449MiB/s-449MiB/s (471MB/s-471MB/s), io=15.6GiB (16.8GB), run=35629-35629msec
commit: 633MiB/s
READ: bw=668MiB/s (700MB/s), 668MiB/s-668MiB/s (700MB/s-700MB/s), io=15.6GiB (16.8GB), run=23952-23952msec
READ: bw=589MiB/s (618MB/s), 589MiB/s-589MiB/s (618MB/s-618MB/s), io=15.6GiB (16.8GB), run=27164-27164msec
READ: bw=638MiB/s (669MB/s), 638MiB/s-638MiB/s (669MB/s-669MB/s), io=15.6GiB (16.8GB), run=25071-25071msec
READ: bw=714MiB/s (749MB/s), 714MiB/s-714MiB/s (749MB/s-749MB/s), io=15.6GiB (16.8GB), run=22409-22409msec
READ: bw=600MiB/s (629MB/s), 600MiB/s-600MiB/s (629MB/s-629MB/s), io=15.6GiB (16.8GB), run=26669-26669msec
READ: bw=592MiB/s (621MB/s), 592MiB/s-592MiB/s (621MB/s-621MB/s), io=15.6GiB (16.8GB), run=27036-27036msec
READ: bw=691MiB/s (725MB/s), 691MiB/s-691MiB/s (725MB/s-725MB/s), io=15.6GiB (16.8GB), run=23150-23150msec
READ: bw=569MiB/s (596MB/s), 569MiB/s-569MiB/s (596MB/s-596MB/s), io=15.6GiB (16.8GB), run=28142-28142msec
READ: bw=563MiB/s (590MB/s), 563MiB/s-563MiB/s (590MB/s-590MB/s), io=15.6GiB (16.8GB), run=28429-28429msec
READ: bw=712MiB/s (746MB/s), 712MiB/s-712MiB/s (746MB/s-746MB/s), io=15.6GiB (16.8GB), run=22478-22478msec
Case 3:
This commit is also verified by the case of launching camera APP which is
usually considered as heavy working load on both of memory and IO, which
shows 12%-24% improvement.
ttl = 0 ttl = 50 ttl = 100
mainline 2267ms 2420ms 2316ms
commit 1992ms 1806ms 1998ms
case 4:
androbench has no improvment as well as regression in RD/WR test item
while make a 3% improvement in sqlite items.
Suggested-by: Matthew Wilcox <willy@infradead.org>
Suggested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
---
change of v2: calculate page's activity via helper function
change of v3: solve layer violation by move API into mm
change of v4: keep block clean by removing the page related API
change of v5: introduce the macros of bio_add_folio/page for read dir.
change of v6: replace the macro of bio_add_xxx by submit_bio which
iterating the bio_vec before launching bio to block layer
change of v7: introduce the function bio_set_active_ioprio
provide updated test result
change of v8: provide two sets of APIs for bio_set_active_ioprio_xxx
change of v9: modify the code according to Matthew's opinion, leave
bio_set_active_ioprio_folio only
---
---
block/Kconfig | 15 +++++++++++++++
block/bio.c | 34 ++++++++++++++++++++++++++++++++++
include/linux/bio.h | 1 +
3 files changed, 50 insertions(+)
diff --git a/block/Kconfig b/block/Kconfig
index f1364d1c0d93..fb3a888194c0 100644
--- a/block/Kconfig
+++ b/block/Kconfig
@@ -228,6 +228,21 @@ config BLOCK_HOLDER_DEPRECATED
config BLK_MQ_STACKING
bool
+config BLK_CONT_ACT_BASED_IOPRIO
+ bool "Enable content activity based ioprio"
+ depends on LRU_GEN
+ default n
+ help
+ This item enable the feature of adjust bio's priority by
+ calculating its content's activity.
+ This feature works as a hint of original bio_set_ioprio
+ which means rt task get no change of its bio->bi_ioprio
+ while other tasks have the opportunity to raise the ioprio
+ if the bio take certain numbers of active pages.
+ The file system should use the API after bio_add_folio for
+ their buffered read/write/sync function to adjust the
+ bio->bi_ioprio.
+
source "block/Kconfig.iosched"
endif # BLOCK
diff --git a/block/bio.c b/block/bio.c
index 816d412c06e9..532001ee08bc 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1476,6 +1476,40 @@ void bio_set_pages_dirty(struct bio *bio)
}
EXPORT_SYMBOL_GPL(bio_set_pages_dirty);
+/*
+ * bio_set_active_ioprio_folio is helper function to count the bio's
+ * content's activities which measured by MGLRU.
+ * The file system should call this function after bio_add_page/folio for
+ * the buffered read/write/sync.
+ */
+#ifdef CONFIG_BLK_CONT_ACT_BASED_IOPRIO
+void bio_set_active_ioprio_folio(struct bio *bio, struct folio *folio)
+{
+ int class, level, hint;
+ int activities;
+
+ /*
+ * use bi_ioprio to record the activities, assume no one will set it
+ * before submit_bio
+ */
+ bio->bi_ioprio += folio_test_workingset(folio) ? 1 : 0;
+ activities = IOPRIO_PRIO_DATA(bio->bi_ioprio);
+ class = IOPRIO_PRIO_CLASS(bio->bi_ioprio);
+ level = IOPRIO_PRIO_LEVEL(bio->bi_ioprio);
+ hint = IOPRIO_PRIO_HINT(bio->bi_ioprio);
+
+ if (activities > bio->bi_vcnt / 2)
+ class = IOPRIO_CLASS_RT;
+ else if (activities > bio->bi_vcnt / 4)
+ class = max(IOPRIO_PRIO_CLASS(get_current_ioprio()), IOPRIO_CLASS_BE);
+
+ bio->bi_ioprio = IOPRIO_PRIO_VALUE_HINT(class, level, hint);
+}
+#else
+void bio_set_active_ioprio_folio(struct bio *bio, struct folio *folio) {}
+#endif
+EXPORT_SYMBOL_GPL(bio_set_active_ioprio_folio);
+
/*
* bio_check_pages_dirty() will check that all the BIO's pages are still dirty.
* If they are, then fine. If, however, some pages are clean then they must
diff --git a/include/linux/bio.h b/include/linux/bio.h
index 41d417ee1349..6c36546f6b9b 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -487,6 +487,7 @@ void bio_iov_bvec_set(struct bio *bio, struct iov_iter *iter);
void __bio_release_pages(struct bio *bio, bool mark_dirty);
extern void bio_set_pages_dirty(struct bio *bio);
extern void bio_check_pages_dirty(struct bio *bio);
+void bio_set_active_ioprio_folio(struct bio *bio, struct folio *folio);
extern void bio_copy_data_iter(struct bio *dst, struct bvec_iter *dst_iter,
struct bio *src, struct bvec_iter *src_iter);
--
2.25.1
^ permalink raw reply related [flat|nested] 4+ messages in thread* [PATCH 2/2] fs: introduce content activity based ioprio
2024-02-21 7:53 [PATCH 0/2] *** introduce content activity based ioprio *** zhaoyang.huang
2024-02-21 7:53 ` [PATCH 1/2] block: introduce content activity based ioprio zhaoyang.huang
@ 2024-02-21 7:53 ` zhaoyang.huang
1 sibling, 0 replies; 4+ messages in thread
From: zhaoyang.huang @ 2024-02-21 7:53 UTC (permalink / raw)
To: Jens Axboe, Darrick J . Wong, Alexander Viro, Christian Brauner,
linux-block, linux-fsdevel, linux-kernel, Zhaoyang Huang,
steve.kang
From: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
This commit would like to introduce content activity based ioprio into
general aops and def_blk_aops read/write API, which account the
content's(folio) activity and set the ioprio_class accordingly. This
change do NOT violate previous ioprio policy but only promote the
value if the activities raise to certain proportion, that can be
deemed as both of the IO launcher and the content's are important
for raising the priority.
Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com>
---
fs/iomap/buffered-io.c | 3 +++
fs/mpage.c | 2 ++
2 files changed, 5 insertions(+)
diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index 5db54ca29a35..5079395d6823 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -390,6 +390,7 @@ static loff_t iomap_readpage_iter(const struct iomap_iter *iter,
ctx->bio->bi_iter.bi_sector = sector;
ctx->bio->bi_end_io = iomap_read_end_io;
bio_add_folio_nofail(ctx->bio, folio, plen, poff);
+ bio_set_active_ioprio_folio(ctx->bio, folio);
}
done:
@@ -624,6 +625,7 @@ static int iomap_read_folio_sync(loff_t block_start, struct folio *folio,
bio_init(&bio, iomap->bdev, &bvec, 1, REQ_OP_READ);
bio.bi_iter.bi_sector = iomap_sector(iomap, block_start);
bio_add_folio_nofail(&bio, folio, plen, poff);
+ bio_set_active_ioprio_folio(&bio, folio);
return submit_bio_wait(&bio);
}
@@ -1742,6 +1744,7 @@ iomap_add_to_ioend(struct inode *inode, loff_t pos, struct folio *folio,
if (!bio_add_folio(wpc->ioend->io_bio, folio, len, poff)) {
wpc->ioend->io_bio = iomap_chain_bio(wpc->ioend->io_bio);
bio_add_folio_nofail(wpc->ioend->io_bio, folio, len, poff);
+ bio_set_active_ioprio_folio(wpc->ioend->io_bio, folio);
}
if (ifs)
diff --git a/fs/mpage.c b/fs/mpage.c
index 242e213ee064..f209e5860423 100644
--- a/fs/mpage.c
+++ b/fs/mpage.c
@@ -308,6 +308,7 @@ static struct bio *do_mpage_readpage(struct mpage_readpage_args *args)
goto alloc_new;
}
+ bio_set_active_ioprio_folio(args->bio, folio);
relative_block = block_in_file - args->first_logical_block;
nblocks = map_bh->b_size >> blkbits;
if ((buffer_boundary(map_bh) && relative_block == nblocks) ||
@@ -626,6 +627,7 @@ static int __mpage_writepage(struct folio *folio, struct writeback_control *wbc,
goto alloc_new;
}
+ bio_set_active_ioprio_folio(bio, folio);
clean_buffers(&folio->page, first_unmapped);
BUG_ON(folio_test_writeback(folio));
--
2.25.1
^ permalink raw reply related [flat|nested] 4+ messages in thread