512K readahead size with thrashing safe readahead

Andrew,

This is to lift default readahead size to 512KB, which I believe yields more I/O throughput without noticeably increasing I/O latency for today's HDD.

For example, for a 100MB/s and 8ms access time HDD:

io_size KB      access_time   transfer_time  io_latency  util%    throughput KB/s IOPS
4               8             0.04           8.04        0.49%    497.57          124.39
8               8             0.08           8.08        0.97%    990.33          123.79
16              8             0.16           8.16        1.92%    1961.69         122.61
32              8             0.31           8.31        3.76%    3849.62         120.30
64              8             0.62           8.62        7.25%    7420.29         115.94
128             8             1.25           9.25        13.51%   13837.84        108.11
256             8             2.50           10.50       23.81%   24380.95        95.24
512             8             5.00           13.00       38.46%   39384.62        76.92
1024            8             10.00          18.00       55.56%   56888.89        55.56
2048            8             20.00          28.00       71.43%   73142.86        35.71
4096            8             40.00          48.00       83.33%   85333.33        20.83

The 128KB => 512KB readahead size boosts IO throughput from ~13MB/s to ~39MB/s, while merely increases IO latency from 9.25ms to 13.00ms.

As for SSD, I find that Intel X25-M SSD desires large readahead size even for sequential reads (the first patch has benchmark details):

        rasize  first run time/throughput       second run time/throughput
        ------------------------------------------------------------------
          4k    3.40038 s,      123 MB/s        3.42842 s,      122 MB/s
          8k    2.7362 s,       153 MB/s        2.74528 s,      153 MB/s
         16k    2.59808 s,      161 MB/s        2.58728 s,      162 MB/s
         32k    2.50488 s,      167 MB/s        2.49138 s,      168 MB/s
         64k    2.12861 s,      197 MB/s        2.13055 s,      197 MB/s
        128k    1.92905 s,      217 MB/s        1.93176 s,      217 MB/s
        256k    1.75896 s,      238 MB/s        1.78963 s,      234 MB/s
        512k    1.67357 s,      251 MB/s        1.69112 s,      248 MB/s
          1M    1.62115 s,      259 MB/s        1.63206 s,      257 MB/s
          2M    1.56204 s,      269 MB/s        1.58854 s,      264 MB/s
          4M    1.57949 s,      266 MB/s        1.57426 s,      266 MB/s

As suggested by Linus, decrease default readahead size for small devices at the same time.

[PATCH 01/11] readahead: limit readahead size for small devices [PATCH 02/11] readahead: bump up the default readahead size
[PATCH 03/11] readahead: introduce {MAX|MIN}_READAHEAD_PAGES macros for ease of use

The two other impacts of an enlarged readahead size are

- memory footprint (caused by readahead miss)
Sequential readahead hit ratio is pretty high regardless of max readahead size; the extra memory footprint is mainly caused by enlarged mmap read-around.
I measured my desktop:
- under Xwindow:
128KB readahead cache hit ratio = 143MB/230MB = 62% 512KB readahead cache hit ratio = 138MB/248MB = 55%
- under console: (seems more stable than the Xwindow data)
128KB readahead cache hit ratio = 30MB/56MB = 53%
1MB readahead cache hit ratio = 30MB/59MB = 51%
So the impact to memory footprint looks acceptable.

- readahead thrashing
It will now cost 1MB readahead buffer per stream. Memory tight systems typically do not run multiple streams; but if they do so, it should help I/O performance as long as we can avoid thrashing, which can be achieved with the following patches.

[PATCH 04/11] readahead: replace ra->mmap_miss with ra->ra_flags [PATCH 05/11] readahead: retain inactive lru pages to be accessed soon [PATCH 06/11] readahead: thrashing safe context readahead

This is a major rewrite of the readahead algorithm, so I did careful tests with the following tracing/stats patches:

[PATCH 07/11] readahead: record readahead patterns [PATCH 08/11] readahead: add tracing event [PATCH 09/11] readahead: add /debug/readahead/stats

I verified the new readahead behavior on various access patterns, as well as stress tested the thrashing safety, by running 300 streams with mem=128M.

Only 2031/61325=3.3% readahead windows are thrashed (due to workload variation):

# cat /debug/readahead/stats
pattern     readahead    eof_hit  cache_hit         io    sync_io    mmap_io       size async_size    io_size
initial            20          9          4         20         20         12         73         37         35
subsequent          3          3          0          1          0          1          8          8          1
context         61325          1       5479      61325       6788          5         14          2         13
thrash           2031          0       1222       2031       2031          0          9          0          6
around            235         90        142        235        235        235         60          0         19
fadvise             0          0          0          0          0          0          0          0          0
random            223        133          0         91         91          1          1          0          1
all             63837        236       6847      63703       9165          0         14          2         13

And the readahead inside a single stream is working as expected:

# grep streams-3162 /debug/tracing/trace
         streams-3162  [000]  8602.455953: readahead: readahead-context(dev=0:2, ino=0, req=287352+1, ra=287354+10-2, async=1) = 10
         streams-3162  [000]  8602.907873: readahead: readahead-context(dev=0:2, ino=0, req=287362+1, ra=287364+20-3, async=1) = 20
         streams-3162  [000]  8604.027879: readahead: readahead-context(dev=0:2, ino=0, req=287381+1, ra=287384+14-2, async=1) = 14
         streams-3162  [000]  8604.754722: readahead: readahead-context(dev=0:2, ino=0, req=287396+1, ra=287398+10-2, async=1) = 10
         streams-3162  [000]  8605.191228: readahead: readahead-context(dev=0:2, ino=0, req=287406+1, ra=287408+18-3, async=1) = 18
         streams-3162  [000]  8606.831895: readahead: readahead-context(dev=0:2, ino=0, req=287423+1, ra=287426+12-2, async=1) = 12
         streams-3162  [000]  8606.919614: readahead: readahead-thrash(dev=0:2, ino=0, req=287425+1, ra=287425+8-0, async=0) = 1
         streams-3162  [000]  8607.545016: readahead: readahead-context(dev=0:2, ino=0, req=287436+1, ra=287438+9-2, async=1) = 9
         streams-3162  [000]  8607.960039: readahead: readahead-context(dev=0:2, ino=0, req=287445+1, ra=287447+18-3, async=1) = 18
         streams-3162  [000]  8608.790973: readahead: readahead-context(dev=0:2, ino=0, req=287462+1, ra=287465+21-3, async=1) = 21
         streams-3162  [000]  8609.763138: readahead: readahead-context(dev=0:2, ino=0, req=287483+1, ra=287486+15-2, async=1) = 15
         streams-3162  [000]  8611.467401: readahead: readahead-context(dev=0:2, ino=0, req=287499+1, ra=287501+11-2, async=1) = 11
         streams-3162  [000]  8642.512413: readahead: readahead-context(dev=0:2, ino=0, req=288053+1, ra=288056+10-2, async=1) = 10
         streams-3162  [000]  8643.246618: readahead: readahead-context(dev=0:2, ino=0, req=288064+1, ra=288066+22-3, async=1) = 22
         streams-3162  [000]  8644.278613: readahead: readahead-context(dev=0:2, ino=0, req=288085+1, ra=288088+16-3, async=1) = 16
         streams-3162  [000]  8644.395782: readahead: readahead-context(dev=0:2, ino=0, req=288087+1, ra=288087+21-3, async=0) = 5
         streams-3162  [000]  8645.109918: readahead: readahead-context(dev=0:2, ino=0, req=288101+1, ra=288108+8-1, async=1) = 8
         streams-3162  [000]  8645.285078: readahead: readahead-context(dev=0:2, ino=0, req=288105+1, ra=288116+8-1, async=1) = 8
         streams-3162  [000]  8645.731794: readahead: readahead-context(dev=0:2, ino=0, req=288115+1, ra=288122+14-2, async=1) = 13
         streams-3162  [000]  8646.114250: readahead: readahead-context(dev=0:2, ino=0, req=288123+1, ra=288136+8-1, async=1) = 8
         streams-3162  [000]  8646.626320: readahead: readahead-context(dev=0:2, ino=0, req=288134+1, ra=288144+16-3, async=1) = 16
         streams-3162  [000]  8647.035721: readahead: readahead-context(dev=0:2, ino=0, req=288143+1, ra=288160+10-2, async=1) = 10
         streams-3162  [000]  8647.693082: readahead: readahead-context(dev=0:2, ino=0, req=288157+1, ra=288165+12-2, async=1) = 8
         streams-3162  [000]  8648.221368: readahead: readahead-context(dev=0:2, ino=0, req=288168+1, ra=288177+15-2, async=1) = 15
         streams-3162  [000]  8649.280800: readahead: readahead-context(dev=0:2, ino=0, req=288190+1, ra=288192+23-3, async=1) = 23
	 [...]

btw, Linus suggested to disable start-of-file readahead if lseek() has been called:

[PATCH 10/11] readahead: dont do start-of-file readahead after lseek()

At last, the updated context readahead will do more radix tree scans, so need to optimize radix_tree_prev_hole():

[PATCH 11/11] radixtree: speed up next/prev hole search

It will on average reduce 8*64 level-0 slot searches to 32 level-0 slot plus 8 level-1 node searches.

Thanks,
Fengguang

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Linus reports a _really_ small & slow (505kB, 15kB/s) USB device, on which blkid runs unpleasantly slow. He manages to optimize the blkid reads down to 1kB+16kB, but still kernel read-ahead turns it into 48kB.

lseek 0, read 1024 => readahead 4 pages (start of file)
lseek 1536, read 16384 => readahead 8 pages (page contiguous)

The readahead heuristics involved here are reasonable ones in general. So it's good to fix blkid with fadvise(RANDOM), as Linus already did.

For the kernel part, Linus suggests:
So maybe we could be less aggressive about read-ahead when the size of the device is small? Turning a 16kB read into a 64kB one is a big deal, when it's about 15% of the whole device!

This looks reasonable: smaller device tend to be slower (USB sticks as well as micro/mobile/old hard disks).

Given that the non-rotational attribute is not always reported, we can take disk size as a max readahead size hint. We use a formula that generates the following concrete limits:

disk size readahead size
(scale by 4) (scale by 2)
2M 4k
8M 8k
32M 16k
128M 32k
512M 64k
2G 128k
8G 256k
32G 512k
128G 1024k

The formula is determined on the following data, collected by script:

	#!/bin/sh

	# please make sure BDEV is not mounted or opened by others
	BDEV=sdb

	for rasize in 4 16 32 64 128 256 512 1024 2048
	do
		echo $rasize > /sys/block/$BDEV/queue/read_ahead_kb 
		time dd if=/dev/$BDEV of=/dev/null bs=4k count=102400
	done

The principle is, the formula shall not limit readahead size to such a degree that will impact some device's sequential read performance.

The Intel SSD is special in that its throughput increases steadily with larger readahead size. However it may take years for Linux to increase its default readahead size to 2MB, so we don't take it seriously in the formula.

SSD 80G Intel x25-M SSDSA2M080

	rasize	first run time/throughput	second run time/throughput
	------------------------------------------------------------------
	  4k	3.40038 s,	123 MB/s	3.42842 s,	122 MB/s
	  8k	2.7362 s,	153 MB/s	2.74528 s,	153 MB/s
	 16k	2.59808 s,	161 MB/s	2.58728 s,	162 MB/s
	 32k	2.50488 s,	167 MB/s	2.49138 s,	168 MB/s
	 64k	2.12861 s,	197 MB/s	2.13055 s,	197 MB/s
	128k	1.92905 s,	217 MB/s	1.93176 s,	217 MB/s
	256k	1.75896 s,	238 MB/s	1.78963 s,	234 MB/s
	512k	1.67357 s,	251 MB/s	1.69112 s,	248 MB/s
	  1M	1.62115 s,	259 MB/s	1.63206 s,	257 MB/s
==>	  2M	1.56204 s,	269 MB/s	1.58854 s,	264 MB/s
	  4M	1.57949 s,	266 MB/s	1.57426 s,	266 MB/s

Note that ==> points to the readahead size that yields plateau throughput.

SSD 30G SanDisk SATA 5000

	  4k	14.1593 s,	29.6 MB/s	14.1699 s,	29.6 MB/s	14.1782 s,	29.6 MB/s
	  8k	8.05231 s,	52.1 MB/s	8.04463 s,	52.1 MB/s	8.04758 s,	52.1 MB/s
	 16k	6.81751 s,	61.5 MB/s	6.81564 s,	61.5 MB/s	6.8146 s,	61.5 MB/s
	 32k	6.24176 s,	67.2 MB/s	6.2438 s,	67.2 MB/s	6.24645 s,	67.1 MB/s
	 64k	5.87828 s,	71.4 MB/s	5.87858 s,	71.3 MB/s	5.87481 s,	71.4 MB/s
	128k	5.71649 s,	73.4 MB/s	5.71804 s,	73.4 MB/s	5.72055 s,	73.3 MB/s
==>	256k	5.62466 s,	74.6 MB/s	5.62304 s,	74.6 MB/s	5.62114 s,	74.6 MB/s
	512k	5.61532 s,	74.7 MB/s	5.62098 s,	74.6 MB/s	5.61818 s,	74.7 MB/s
	  1M	5.50888 s,	76.1 MB/s	5.6204 s,	74.6 MB/s	5.62281 s,	74.6 MB/s

USB stick 32G Teclast CoolFlash idVendor=1307, idProduct=0165

	  4k	53.1635 s,	7.9 MB/s 	53.155 s,	7.9 MB/s 	53.107 s,	7.9 MB/s
	  8k	23.4061 s,	17.9 MB/s	23.3955 s,	17.9 MB/s	23.4222 s,	17.9 MB/s
	 16k	17.1077 s,	24.5 MB/s	17.0909 s,	24.5 MB/s	17.0875 s,	24.5 MB/s
	 32k	14.6029 s,	28.7 MB/s	14.5913 s,	28.7 MB/s	14.5951 s,	28.7 MB/s
	 64k	14.5483 s,	28.8 MB/s	14.5344 s,	28.9 MB/s	14.5333 s,	28.9 MB/s
==>	128k	13.7497 s,	30.5 MB/s	13.7364 s,	30.5 MB/s	13.731 s,	30.5 MB/s
	256k	13.5521 s,	30.9 MB/s	13.5415 s,	31.0 MB/s	13.5554 s,	30.9 MB/s
	512k	13.5414 s,	31.0 MB/s	13.5631 s,	30.9 MB/s	13.5654 s,	30.9 MB/s
	  1M	13.574 s,	30.9 MB/s	13.5686 s,	30.9 MB/s	13.5667 s,	30.9 MB/s

USB stick 4G SanDisk Cruzer idVendor=0781, idProduct=5151

	  4k	65.3449 s,	6.4 MB/s 	65.3759 s,	6.4 MB/s 	65.3405 s,	6.4 MB/s
	  8k	31.2002 s,	13.4 MB/s	31.1914 s,	13.4 MB/s	31.6836 s,	13.2 MB/s
	 16k	23.5281 s,	17.8 MB/s	23.4705 s,	17.9 MB/s	23.5859 s,	17.8 MB/s
	 32k	19.6786 s,	21.3 MB/s	19.719 s,	21.3 MB/s	19.7548 s,	21.2 MB/s
	 64k	19.6219 s,	21.4 MB/s	19.6125 s,	21.4 MB/s	19.594 s,	21.4 MB/s
==>	128k	18.021 s,	23.3 MB/s	18.0527 s,	23.2 MB/s	18.0694 s,	23.2 MB/s
	256k	17.978 s,	23.3 MB/s	17.6483 s,	23.8 MB/s	17.9324 s,	23.4 MB/s
	512k	17.659 s,	23.8 MB/s	17.9403 s,	23.4 MB/s	17.986 s,	23.3 MB/s
	  1M	17.9437 s,	23.4 MB/s	18.0634 s,	23.2 MB/s	17.9469 s,	23.4 MB/s

USB stick 2G idVendor=0204, idProduct=6025 SerialNumber: 08082005000113

	  4k	62.6246 s,	6.7 MB/s 	60.5872 s,	6.9 MB/s 	62.2581 s,	6.7 MB/s
	  8k	35.7505 s,	11.7 MB/s	35.764 s,	11.7 MB/s	35.7396 s,	11.7 MB/s
	 16k	33.7949 s,	12.4 MB/s	33.8041 s,	12.4 MB/s	33.8015 s,	12.4 MB/s
-->	 32k	31.3851 s,	13.4 MB/s	31.381 s,	13.4 MB/s	31.3784 s,	13.4 MB/s
	 64k	31.3478 s,	13.4 MB/s	31.3494 s,	13.4 MB/s	31.3486 s,	13.4 MB/s
==>	128k	30.7384 s,	13.6 MB/s	30.7337 s,	13.6 MB/s	30.728 s,	13.6 MB/s
	256k	30.5439 s,	13.7 MB/s	30.544 s,	13.7 MB/s	30.5433 s,	13.7 MB/s
	512k	30.5408 s,	13.7 MB/s	30.543 s,	13.7 MB/s	30.5447 s,	13.7 MB/s
	  1M	30.5919 s,	13.7 MB/s	30.5893 s,	13.7 MB/s	30.5939 s,	13.7 MB/s

Anyone has 512/128MB USB stick? Anyway you get satisfiable performance with >= 32k readahead size.

---
 block/genhd.c |   18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

@@ -518,6 +518,7 @@ void add_disk(struct gendisk *disk)
 	struct backing_dev_info *bdi;
 	dev_t devt;
 	int retval;
+	unsigned long size;
 
 	/* minors == 0 indicates to use ext devt from part0 and should
 	 * be accompanied with EXT_DEVT flag.  Make sure all

@@ -551,6 +552,23 @@ void add_disk(struct gendisk *disk)
 	retval = sysfs_create_link(&disk_to_dev(disk)->kobj, &bdi->dev->kobj,
 				   "bdi");
 	WARN_ON(retval);
+
+	/*
+	 * limit readahead size for small devices
+	 *        disk size    readahead size
+	 *               2M                4k
+	 *               8M                8k
+	 *              32M               16k
+	 *             128M               32k
+	 *             512M               64k
+	 *               2G              128k
+	 *               8G              256k
+	 *              32G              512k
+	 *             128G             1024k
+	 */
+	size = get_capacity(disk) >> 12;
+	size = 1UL << (ilog2(size) / 2);
+	bdi->ra_pages = min(bdi->ra_pages, size);
 }
 
 EXPORT_SYMBOL(add_disk);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Use 512kb max readahead size, and 32kb min readahead size.

The former helps io performance for common workloads. The latter will be used in the thrashing safe context readahead.

---
 include/linux/mm.h |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

@@ -1184,8 +1184,8 @@ int write_one_page(struct page *page, in
 void task_dirty_inc(struct task_struct *tsk);
 
 /* readahead.c */
-#define VM_MAX_READAHEAD	128	/* kbytes */
-#define VM_MIN_READAHEAD	16	/* kbytes (includes current page) */
+#define VM_MAX_READAHEAD	512	/* kbytes */
+#define VM_MIN_READAHEAD	32	/* kbytes (includes current page) */
 
 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 			pgoff_t offset, unsigned long nr_to_read);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

---
 block/blk-core.c   |    3 +--
 fs/fuse/inode.c    |    2 +-
 include/linux/mm.h |    3 +++
 mm/backing-dev.c   |    2 +-
 4 files changed, 6 insertions(+), 4 deletions(-)

@@ -498,8 +498,7 @@ struct request_queue *blk_alloc_queue_no
 
 	q->backing_dev_info.unplug_io_fn = blk_backing_dev_unplug;
 	q->backing_dev_info.unplug_io_data = q;
-	q->backing_dev_info.ra_pages =
-			(VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
+	q->backing_dev_info.ra_pages = MAX_READAHEAD_PAGES;
 	q->backing_dev_info.state = 0;
 	q->backing_dev_info.capabilities = BDI_CAP_MAP_COPY;
 	q->backing_dev_info.name = "block";

@@ -870,7 +870,7 @@ static int fuse_bdi_init(struct fuse_con
 	int err;
 
 	fc->bdi.name = "fuse";
-	fc->bdi.ra_pages = (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
+	fc->bdi.ra_pages = MAX_READAHEAD_PAGES;
 	fc->bdi.unplug_io_fn = default_unplug_io_fn;
 	/* fuse does it's own writeback accounting */
 	fc->bdi.capabilities = BDI_CAP_NO_ACCT_WB;

@@ -1187,6 +1187,9 @@ void task_dirty_inc(struct task_struct *
 #define VM_MAX_READAHEAD	512	/* kbytes */
 #define VM_MIN_READAHEAD	32	/* kbytes (includes current page) */
 
+#define MAX_READAHEAD_PAGES (VM_MAX_READAHEAD*1024 / PAGE_CACHE_SIZE)
+#define MIN_READAHEAD_PAGES DIV_ROUND_UP(VM_MIN_READAHEAD*1024, PAGE_CACHE_SIZE)
+
 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 			pgoff_t offset, unsigned long nr_to_read);
 

@@ -18,7 +18,7 @@ EXPORT_SYMBOL(default_unplug_io_fn);
 
 struct backing_dev_info default_backing_dev_info = {
 	.name		= "default",
-	.ra_pages	= VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
+	.ra_pages	= MAX_READAHEAD_PAGES,
 	.state		= 0,
 	.capabilities	= BDI_CAP_MAP_COPY,
 	.unplug_io_fn	= default_unplug_io_fn,

Introduce a readahead flags field and embed the existing mmap_miss in it (to save space).

It will be possible to lose the flags in race conditions, however the impact should be limited.

---
 include/linux/fs.h |   30 +++++++++++++++++++++++++++++-
 mm/filemap.c       |    7 ++-----
 2 files changed, 31 insertions(+), 6 deletions(-)

@@ -889,10 +889,38 @@ struct file_ra_state {
 					   there are only # of pages ahead */
 
 	unsigned int ra_pages;		/* Maximum readahead window */
-	unsigned int mmap_miss;		/* Cache miss stat for mmap accesses */
+	unsigned int ra_flags;
 	loff_t prev_pos;		/* Cache last read() position */
 };
 
+/* ra_flags bits */
+#define	READAHEAD_MMAP_MISS	0x0000ffff /* cache misses for mmap access */
+
+/*
+ * Don't do ra_flags++ directly to avoid possible overflow:
+ * the ra fields can be accessed concurrently in a racy way.
+ */
+static inline unsigned int ra_mmap_miss_inc(struct file_ra_state *ra)
+{
+	unsigned int miss = ra->ra_flags & READAHEAD_MMAP_MISS;
+
+	if (miss < READAHEAD_MMAP_MISS) {
+		miss++;
+		ra->ra_flags = miss | (ra->ra_flags &~ READAHEAD_MMAP_MISS);
+	}
+	return miss;
+}
+
+static inline void ra_mmap_miss_dec(struct file_ra_state *ra)
+{
+	unsigned int miss = ra->ra_flags & READAHEAD_MMAP_MISS;
+
+	if (miss) {
+		miss--;
+		ra->ra_flags = miss | (ra->ra_flags &~ READAHEAD_MMAP_MISS);
+	}
+}
+
 /*
  * Check if @index falls in the readahead windows.
  */

@@ -1418,14 +1418,12 @@ static void do_sync_mmap_readahead(struc
 		return;
 	}
 
-	if (ra->mmap_miss < INT_MAX)
-		ra->mmap_miss++;
 
 	/*
 	 * Do we miss much more than hit in this file? If so,
 	 * stop bothering with read-ahead. It will only hurt.
 	 */
-	if (ra->mmap_miss > MMAP_LOTSAMISS)
+	if (ra_mmap_miss_inc(ra) > MMAP_LOTSAMISS)
 		return;
 
 	/*

@@ -1455,8 +1453,7 @@ static void do_async_mmap_readahead(stru
 	/* If we don't want any read-ahead, don't bother */
 	if (VM_RandomReadHint(vma))
 		return;
-	if (ra->mmap_miss > 0)
-		ra->mmap_miss--;
+	ra_mmap_miss_dec(ra);
 	if (PageReadahead(page))
 		page_cache_async_readahead(mapping, ra, file,
 					   page, offset, ra->ra_pages);

From: Chris Frost <[email protected]>

Ensure that cached pages in the inactive list are not prematurely evicted; move such pages to lru head when they are covered by
- in-kernel heuristic readahead
- an posix_fadvise(POSIX_FADV_WILLNEED) hint from an application

Before this patch, pages already in core may be evicted before the pages covered by the same prefetch scan but that were not yet in core. Many small read requests may be forced on the disk because of this behavior.

In particular, posix_fadvise(... POSIX_FADV_WILLNEED) on an in-core page has no effect on the page's location in the LRU list, even if it is the next victim on the inactive list.

This change helps address the performance problems we encountered while modifying SQLite and the GIMP to use large file prefetching. Overall these prefetching techniques improved the runtime of large benchmarks by 10-17x for these applications. More in the publication _Reducing Seek Overhead with Application-Directed Prefetching_ in USENIX ATC 2009 and at http://libprefetch.cs.ucla.edu/.

---
 mm/readahead.c |   44 ++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)

@@ -9,7 +9,9 @@
 
 #include <linux/kernel.h>
 #include <linux/fs.h>
+#include <linux/memcontrol.h>
 #include <linux/mm.h>
+#include <linux/mm_inline.h>
 #include <linux/module.h>
 #include <linux/blkdev.h>
 #include <linux/backing-dev.h>

@@ -133,6 +135,40 @@ out:
 }
 
 /*
+ * The file range is expected to be accessed in near future.  Move pages
+ * (possibly in inactive lru tail) to lru head, so that they are retained
+ * in memory for some reasonable time.
+ */
+static void retain_inactive_pages(struct address_space *mapping,
+				  pgoff_t index, int len)
+{
+	int i;
+	struct page *page;
+	struct zone *zone;
+
+	for (i = 0; i < len; i++) {
+		page = find_get_page(mapping, index + i);
+		if (!page)
+			continue;
+
+		zone = page_zone(page);
+		spin_lock_irq(&zone->lru_lock);
+
+		if (PageLRU(page) &&
+		    !PageActive(page) &&
+		    !PageUnevictable(page)) {
+			int lru = page_lru_base_type(page);
+
+			del_page_from_lru_list(zone, page, lru);
+			add_page_to_lru_list(zone, page, lru);
+		}
+
+		spin_unlock_irq(&zone->lru_lock);
+		put_page(page);
+	}
+}
+
+/*
  * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates all
  * the pages first, then submits them all for I/O. This avoids the very bad
  * behaviour which would occur if page allocations are causing VM writeback.

@@ -184,6 +220,14 @@ __do_page_cache_readahead(struct address
 	}
 
 	/*
+	 * Normally readahead will auto stop on cached segments, so we won't
+	 * hit many cached pages. If it does happen, bring the inactive pages
+	 * adjecent to the newly prefetched ones(if any).
+	 */
+	if (ret < nr_to_read)
+		retain_inactive_pages(mapping, offset, page_idx);
+
+	/*
 	 * Now start the IO.  We ignore I/O errors - if the page is not
 	 * uptodate then the caller will launch readpage again, and
 	 * will then handle the error.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Introduce a more complete version of context readahead, which is a full-fledged readahead algorithm by itself. It replaces some of the existing cases.

- oversize read
no behavior change; except in thrashed mode, async_size will be 0
- random read
no behavior change; implies some different internal handling The random read will now be recorded in file_ra_state, which means in an intermixed sequential+random pattern, the sequential part's state will be flushed by random ones, and hence will be serviced by the context readahead instead of the stateful one. Also means that the first readahead for a sequential read in the middle of file will be started by the stateful one, instead of the sequential cache miss.
- sequential cache miss
better
When walking out of a cached page segment, the readahead size will be fully restored immediately instead of ramping up from initial size.
- hit readahead marker without valid state
better in rare cases; costs more radix tree lookups, but won't be a problem with optimized radix_tree_prev_hole(). The added radix tree scan for history pages is to calculate the thrashing safe readahead size and adaptive async size.

The algorithm first looks ahead to find the start point of next read-ahead, then looks backward in the page cache to get an estimation of the thrashing-threshold.

It is able to automatically adapt to the thrashing threshold in a smooth workload. The estimation theory can be illustrated with figure:

chunk A chunk B chunk C head

l01 l11 l12 l21 l22

| |-->|-->|       |------>|-->|                |------>|
| +-------+       +-----------+                +-------------+               |
| |   #   |       |       #   |                |       #     |               |
| +-------+       +-----------+                +-------------+               |
| |<==============|<===========================|<============================|
        L0                     L1                            L2

Let f(l) = L be a map from
l: the number of pages read by the stream
to
L: the number of pages pushed into inactive_list in the mean time
then

     f(l01) <= L0
     f(l11 + l12) = L1
     f(l21 + l22) = L2
     ...
     f(l01 + l11 + ...) <= Sum(L0 + L1 + ...)
                        <= Length(inactive_list) = f(thrashing-threshold)

So the count of continuous history pages left in inactive_list is always a lower estimation of the true thrashing-threshold. Given a stable workload, the readahead size will keep ramping up and then stabilize in range

(thrashing_threshold/2, thrashing_threshold)

This is good because, it's in fact bad to always reach thrashing_threshold. That would not only be more susceptible to fluctuations, but also impose eviction pressure to the cached pages.

---
 include/linux/fs.h |    1 
 mm/readahead.c     |  155 ++++++++++++++++++++++++-------------------
 2 files changed, 88 insertions(+), 68 deletions(-)

@@ -20,6 +20,11 @@
 #include <linux/pagemap.h>
 
 /*
+ * Set async size to 1/# of the thrashing threshold.
+ */
+#define READAHEAD_ASYNC_RATIO	8
+
+/*
  * Initialise a struct file's readahead state.  Assumes that the caller has
  * memset *ra to zero.
  */

@@ -393,39 +398,16 @@ static pgoff_t count_history_pages(struc
 }
 
 /*
- * page cache context based read-ahead
+ * Is @index recently readahead but not yet read by application?
+ * The low boundary is permissively estimated.
  */
-static int try_context_readahead(struct address_space *mapping,
-				 struct file_ra_state *ra,
-				 pgoff_t offset,
-				 unsigned long req_size,
-				 unsigned long max)
+static bool ra_thrashed(struct file_ra_state *ra, pgoff_t index)
 {
-	pgoff_t size;
-
-	size = count_history_pages(mapping, ra, offset, max);
-
-	/*
-	 * no history pages:
-	 * it could be a random read
-	 */
-	if (!size)
-		return 0;
-
-	/*
-	 * starts from beginning of file:
-	 * it is a strong indication of long-run stream (or whole-file-read)
-	 */
-	if (size >= offset)
-		size *= 2;
-
-	ra->start = offset;
-	ra->size = get_init_ra_size(size + req_size, max);
-	ra->async_size = ra->size;
-
-	return 1;
+	return (index >= ra->start - ra->size &&
+		index <  ra->start + ra->size);
 }
 
+
 /*
  * A minimal readahead algorithm for trivial sequential/random reads.
  */

@@ -436,12 +418,26 @@ ondemand_readahead(struct address_space 
 		   unsigned long req_size)
 {
 	unsigned long max = max_sane_readahead(ra->ra_pages);
+	unsigned int size;
+	pgoff_t start;
 
 	/*
 	 * start of file
 	 */
-	if (!offset)
-		goto initial_readahead;
+	if (!offset) {
+		ra->start = offset;
+		ra->size = get_init_ra_size(req_size, max);
+		ra->async_size = ra->size > req_size ?
+				 ra->size - req_size : ra->size;
+		goto readit;
+	}
+
+	/*
+	 * Context readahead is thrashing safe, and can adapt to near the
+	 * thrashing threshold given a stable workload.
+	 */
+	if (ra->ra_flags & READAHEAD_THRASHED)
+		goto context_readahead;
 
 	/*
 	 * It's the expected callback offset, assume sequential access.

@@ -456,58 +452,81 @@ ondemand_readahead(struct address_space 
 	}
 
 	/*
-	 * Hit a marked page without valid readahead state.
-	 * E.g. interleaved reads.
-	 * Query the pagecache for async_size, which normally equals to
-	 * readahead size. Ramp it up and use it as the new readahead size.
+	 * oversize read, no need to query page cache
 	 */
-	if (hit_readahead_marker) {
-		pgoff_t start;
+	if (req_size > max && !hit_readahead_marker) {
+		ra->start = offset;
+		ra->size = max;
+		ra->async_size = max;
+		goto readit;
+	}
 
+	/*
+	 * page cache context based read-ahead
+	 *
+	 *     ==========================_____________..............
+	 *                          [ current window ]
+	 *                               ^offset
+	 * 1)                            |---- A ---->[start
+	 * 2) |<----------- H -----------|
+	 * 3)                            |----------- H ----------->]end
+	 *                                            [ new window ]
+	 *    [=] cached,visited [_] cached,to-be-visited [.] not cached
+	 *
+	 * 1) A = pages ahead = previous async_size
+	 * 2) H = history pages = thrashing safe size
+	 * 3) H - A = new readahead size
+	 */
+context_readahead:
+	if (hit_readahead_marker) {
 		rcu_read_lock();
-		start = radix_tree_next_hole(&mapping->page_tree, offset+1,max);
+		start = radix_tree_next_hole(&mapping->page_tree,
+					     offset + 1, max);
 		rcu_read_unlock();
-
+		/*
+		 * there are enough pages ahead: no readahead
+		 */
 		if (!start || start - offset > max)
 			return 0;
+	} else
+		start = offset;
 
+	size = count_history_pages(mapping, ra, offset,
+				   READAHEAD_ASYNC_RATIO * max);
+	/*
+	 * no history pages cached, could be
+	 * 	- a random read
+	 * 	- a thrashed sequential read
+	 */
+	if (!size && !hit_readahead_marker) {
+		if (!ra_thrashed(ra, offset)) {
+			ra->size = min(req_size, max);
+		} else {
+			retain_inactive_pages(mapping, offset, min(2 * max,
+						ra->start + ra->size - offset));
+			ra->size = max_t(int, ra->size/2, MIN_READAHEAD_PAGES);
+			ra->ra_flags |= READAHEAD_THRASHED;
+		}
+		ra->async_size = 0;
 		ra->start = start;
-		ra->size = start - offset;	/* old async_size */
-		ra->size += req_size;
-		ra->size = get_next_ra_size(ra, max);
-		ra->async_size = ra->size;
 		goto readit;
 	}
-
 	/*
-	 * oversize read
+	 * history pages start from beginning of file:
+	 * it is a strong indication of long-run stream (or whole-file reads)
 	 */
-	if (req_size > max)
-		goto initial_readahead;
-
-	/*
-	 * sequential cache miss
-	 */
-	if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
-		goto initial_readahead;
-
-	/*
-	 * Query the page cache and look for the traces(cached history pages)
-	 * that a sequential stream would leave behind.
-	 */
-	if (try_context_readahead(mapping, ra, offset, req_size, max))
-		goto readit;
-
+	if (size >= offset)
+		size *= 2;
 	/*
-	 * standalone, small random read
-	 * Read as is, and do not pollute the readahead state.
+	 * pages to readahead are already cached
 	 */
-	return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
+	if (size <= start - offset)
+		return 0;
 
-initial_readahead:
-	ra->start = offset;
-	ra->size = get_init_ra_size(req_size, max);
-	ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
+	size -= start - offset;
+	ra->start = start;
+	ra->size = clamp_t(unsigned int, size, MIN_READAHEAD_PAGES, max);
+	ra->async_size = min(ra->size, 1 + size / READAHEAD_ASYNC_RATIO);
 
 readit:
 	/*

@@ -895,6 +895,7 @@ struct file_ra_state {
 
 /* ra_flags bits */
 #define	READAHEAD_MMAP_MISS	0x0000ffff /* cache misses for mmap access */
+#define READAHEAD_THRASHED	0x10000000
 
 /*
  * Don't do ra_flags++ directly to avoid possible overflow:

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Record the readahead pattern in ra_flags. This info can be examined by users via the readahead tracing/stats interfaces.

Currently 7 patterns are defined:

pattern readahead for
-----------------------------------------------------------
RA_PATTERN_INITIAL start-of-file/oversize read
RA_PATTERN_SUBSEQUENT trivial sequential read
RA_PATTERN_CONTEXT interleaved sequential read
RA_PATTERN_THRASH thrashed sequential read
RA_PATTERN_MMAP_AROUND mmap fault
RA_PATTERN_FADVISE posix_fadvise()
RA_PATTERN_RANDOM random read

---
 include/linux/fs.h |   32 ++++++++++++++++++++++++++++++++
 include/linux/mm.h |    4 +++-
 mm/filemap.c       |    9 +++++++--
 mm/readahead.c     |   17 +++++++++++++----
 4 files changed, 55 insertions(+), 7 deletions(-)

@@ -894,8 +894,40 @@ struct file_ra_state {
 };
 
 /* ra_flags bits */
+#define READAHEAD_PATTERN_SHIFT	20
+#define READAHEAD_PATTERN	0x00f00000
 #define	READAHEAD_MMAP_MISS	0x0000ffff /* cache misses for mmap access */
 #define READAHEAD_THRASHED	0x10000000
+#define	READAHEAD_MMAP		0x20000000
+
+/*
+ * Which policy makes decision to do the current read-ahead IO?
+ */
+enum readahead_pattern {
+	RA_PATTERN_INITIAL,
+	RA_PATTERN_SUBSEQUENT,
+	RA_PATTERN_CONTEXT,
+	RA_PATTERN_THRASH,
+	RA_PATTERN_MMAP_AROUND,
+	RA_PATTERN_FADVISE,
+	RA_PATTERN_RANDOM,
+	RA_PATTERN_ALL,		/* for summary stats */
+	RA_PATTERN_MAX
+};
+
+static inline int ra_pattern(int ra_flags)
+{
+	int pattern = (ra_flags & READAHEAD_PATTERN)
+			       >> READAHEAD_PATTERN_SHIFT;
+
+	return min(pattern, RA_PATTERN_ALL);
+}
+
+static inline void ra_set_pattern(struct file_ra_state *ra, int pattern)
+{
+	ra->ra_flags = (ra->ra_flags & ~READAHEAD_PATTERN) |
+			    (pattern << READAHEAD_PATTERN_SHIFT);
+}
 
 /*
  * Don't do ra_flags++ directly to avoid possible overflow:

@@ -291,7 +291,10 @@ unsigned long max_sane_readahead(unsigne
  * Submit IO for the read-ahead request in file_ra_state.
  */
 unsigned long ra_submit(struct file_ra_state *ra,
-		       struct address_space *mapping, struct file *filp)
+			struct address_space *mapping,
+			struct file *filp,
+			pgoff_t offset,
+			unsigned long req_size)
 {
 	int actual;
 

@@ -425,6 +428,7 @@ ondemand_readahead(struct address_space 
 	 * start of file
 	 */
 	if (!offset) {
+		ra_set_pattern(ra, RA_PATTERN_INITIAL);
 		ra->start = offset;
 		ra->size = get_init_ra_size(req_size, max);
 		ra->async_size = ra->size > req_size ?

@@ -445,6 +449,7 @@ ondemand_readahead(struct address_space 
 	 */
 	if ((offset == (ra->start + ra->size - ra->async_size) ||
 	     offset == (ra->start + ra->size))) {
+		ra_set_pattern(ra, RA_PATTERN_SUBSEQUENT);
 		ra->start += ra->size;
 		ra->size = get_next_ra_size(ra, max);
 		ra->async_size = ra->size;

@@ -455,6 +460,7 @@ ondemand_readahead(struct address_space 
 	 * oversize read, no need to query page cache
 	 */
 	if (req_size > max && !hit_readahead_marker) {
+		ra_set_pattern(ra, RA_PATTERN_INITIAL);
 		ra->start = offset;
 		ra->size = max;
 		ra->async_size = max;

@@ -500,8 +506,10 @@ context_readahead:
 	 */
 	if (!size && !hit_readahead_marker) {
 		if (!ra_thrashed(ra, offset)) {
+			ra_set_pattern(ra, RA_PATTERN_RANDOM);
 			ra->size = min(req_size, max);
 		} else {
+			ra_set_pattern(ra, RA_PATTERN_THRASH);
 			retain_inactive_pages(mapping, offset, min(2 * max,
 						ra->start + ra->size - offset));
 			ra->size = max_t(int, ra->size/2, MIN_READAHEAD_PAGES);

@@ -518,12 +526,13 @@ context_readahead:
 	if (size >= offset)
 		size *= 2;
 	/*
-	 * pages to readahead are already cached
+	 * Pages to readahead are already cached?
 	 */
 	if (size <= start - offset)
 		return 0;
-
 	size -= start - offset;
+
+	ra_set_pattern(ra, RA_PATTERN_CONTEXT);
 	ra->start = start;
 	ra->size = clamp_t(unsigned int, size, MIN_READAHEAD_PAGES, max);
 	ra->async_size = min(ra->size, 1 + size / READAHEAD_ASYNC_RATIO);

@@ -539,7 +548,7 @@ readit:
 		ra->size += ra->async_size;
 	}
 
-	return ra_submit(ra, mapping, filp);
+	return ra_submit(ra, mapping, filp, offset, req_size);
 }
 
 /**

@@ -1209,7 +1209,9 @@ void page_cache_async_readahead(struct a
 unsigned long max_sane_readahead(unsigned long nr);
 unsigned long ra_submit(struct file_ra_state *ra,
 			struct address_space *mapping,
-			struct file *filp);
+			struct file *filp,
+			pgoff_t offset,
+			unsigned long req_size);
 
 /* Do stack extension */
 extern int expand_stack(struct vm_area_struct *vma, unsigned long address);

@@ -1413,6 +1413,7 @@ static void do_sync_mmap_readahead(struc
 
 	if (VM_SequentialReadHint(vma) ||
 			offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
+		ra->ra_flags |= READAHEAD_MMAP;
 		page_cache_sync_readahead(mapping, ra, file, offset,
 					  ra->ra_pages);
 		return;

@@ -1431,10 +1432,12 @@ static void do_sync_mmap_readahead(struc
 	 */
 	ra_pages = max_sane_readahead(ra->ra_pages);
 	if (ra_pages) {
+		ra->ra_flags |= READAHEAD_MMAP;
+		ra_set_pattern(ra, RA_PATTERN_MMAP_AROUND);
 		ra->start = max_t(long, 0, offset - ra_pages/2);
 		ra->size = ra_pages;
 		ra->async_size = 0;
-		ra_submit(ra, mapping, file);
+		ra_submit(ra, mapping, file, offset, 1);
 	}
 }
 

@@ -1454,9 +1457,11 @@ static void do_async_mmap_readahead(stru
 	if (VM_RandomReadHint(vma))
 		return;
 	ra_mmap_miss_dec(ra);
-	if (PageReadahead(page))
+	if (PageReadahead(page)) {
+		ra->ra_flags |= READAHEAD_MMAP;
 		page_cache_async_readahead(mapping, ra, file,
 					   page, offset, ra->ra_pages);
+	}
 }
 
 /**

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Example output:

# echo 1 > /debug/tracing/events/readahead/enable
# cp test-file /dev/null
# cat /debug/tracing/trace  # trimmed output
readahead-initial(dev=0:15, ino=100177, req=0+2, ra=0+4-2, async=0) = 4
readahead-subsequent(dev=0:15, ino=100177, req=2+2, ra=4+8-8, async=1) = 8
readahead-subsequent(dev=0:15, ino=100177, req=4+2, ra=12+16-16, async=1) = 16
readahead-subsequent(dev=0:15, ino=100177, req=12+2, ra=28+32-32, async=1) = 32
readahead-subsequent(dev=0:15, ino=100177, req=28+2, ra=60+60-60, async=1) = 24
readahead-subsequent(dev=0:15, ino=100177, req=60+2, ra=120+60-60, async=1) = 0

---
 include/trace/events/readahead.h |   69 +++++++++++++++++++++++++++++
 mm/readahead.c                   |   22 +++++++++
 2 files changed, 91 insertions(+)

@@ -0,0 +1,69 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM readahead
+
+#if !defined(_TRACE_READAHEAD_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_READAHEAD_H
+
+#include <linux/tracepoint.h>
+
+extern const char * const ra_pattern_names[];
+
+/*
+ * Tracepoint for guest mode entry.
+ */
+TRACE_EVENT(readahead,
+	TP_PROTO(struct address_space *mapping,
+		 pgoff_t offset,
+		 unsigned long req_size,
+		 unsigned int ra_flags,
+		 pgoff_t start,
+		 unsigned int size,
+		 unsigned int async_size,
+		 unsigned int actual),
+
+	TP_ARGS(mapping, offset, req_size,
+		ra_flags, start, size, async_size, actual),
+
+	TP_STRUCT__entry(
+		__field(	dev_t,		dev		)
+		__field(	ino_t,		ino		)
+		__field(	pgoff_t,	offset		)
+		__field(	unsigned long,	req_size	)
+		__field(	unsigned int,	pattern		)
+		__field(	pgoff_t,	start		)
+		__field(	unsigned int,	size		)
+		__field(	unsigned int,	async_size	)
+		__field(	unsigned int,	actual		)
+	),
+
+	TP_fast_assign(
+		__entry->dev		= mapping->host->i_sb->s_dev;
+		__entry->ino		= mapping->host->i_ino;
+		__entry->pattern	= ra_pattern(ra_flags);
+		__entry->offset		= offset;
+		__entry->req_size	= req_size;
+		__entry->start		= start;
+		__entry->size		= size;
+		__entry->async_size	= async_size;
+		__entry->actual		= actual;
+	),
+
+	TP_printk("readahead-%s(dev=%d:%d, ino=%lu, "
+		  "req=%lu+%lu, ra=%lu+%d-%d, async=%d) = %d",
+			ra_pattern_names[__entry->pattern],
+			MAJOR(__entry->dev),
+			MINOR(__entry->dev),
+			__entry->ino,
+			__entry->offset,
+			__entry->req_size,
+			__entry->start,
+			__entry->size,
+			__entry->async_size,
+			__entry->start > __entry->offset,
+			__entry->actual)
+);
+
+#endif /* _TRACE_READAHEAD_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>

@@ -19,11 +19,25 @@
 #include <linux/pagevec.h>
 #include <linux/pagemap.h>
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/readahead.h>
+
 /*
  * Set async size to 1/# of the thrashing threshold.
  */
 #define READAHEAD_ASYNC_RATIO	8
 
+const char * const ra_pattern_names[] = {
+	[RA_PATTERN_INITIAL]		= "initial",
+	[RA_PATTERN_SUBSEQUENT]		= "subsequent",
+	[RA_PATTERN_CONTEXT]		= "context",
+	[RA_PATTERN_THRASH]		= "thrash",
+	[RA_PATTERN_MMAP_AROUND]	= "around",
+	[RA_PATTERN_FADVISE]		= "fadvise",
+	[RA_PATTERN_RANDOM]		= "random",
+	[RA_PATTERN_ALL]		= "all",
+};
+
 /*
  * Initialise a struct file's readahead state.  Assumes that the caller has
  * memset *ra to zero.

@@ -274,6 +288,11 @@ int force_page_cache_readahead(struct ad
 		offset += this_chunk;
 		nr_to_read -= this_chunk;
 	}
+
+	trace_readahead(mapping, offset, nr_to_read,
+			RA_PATTERN_FADVISE << READAHEAD_PATTERN_SHIFT,
+			offset, nr_to_read, 0, ret);
+
 	return ret;
 }
 

@@ -301,6 +320,9 @@ unsigned long ra_submit(struct file_ra_s
 	actual = __do_page_cache_readahead(mapping, filp,
 					ra->start, ra->size, ra->async_size);
 
+	trace_readahead(mapping, offset, req_size, ra->ra_flags,
+			ra->start, ra->size, ra->async_size, actual);
+
 	return actual;
 }
 

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Collect readahead stats when CONFIG_READAHEAD_STATS=y.

This is enabled by default because the added overheads are trivial:
two readahead_stats() calls per readahead.

Example output:
(taken from a fresh booted NFS-ROOT box with rsize=16k)

$ cat /debug/readahead/stats
pattern     readahead    eof_hit  cache_hit         io    sync_io    mmap_io       size async_size    io_size
initial           524        216         26        498        498         18          7          4          4
subsequent        181         80          1        130         13         60         25         25         24
context            94         28          3         85         64          8          7          2          5
thrash              0          0          0          0          0          0          0          0          0
around            162        121         33        162        162        162         60          0         21
fadvise             0          0          0          0          0          0          0          0          0
random            137          0          0        137        137          0          1          0          1
all              1098        445         63       1012        874          0         17          6          9

The two most important columns are
- io number of readahead IO
- io_size average readahead IO size

---
 mm/Kconfig     |   13 +++
 mm/readahead.c |  177 ++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 188 insertions(+), 2 deletions(-)

@@ -38,6 +38,179 @@ const char * const ra_pattern_names[] = 
 	[RA_PATTERN_ALL]		= "all",
 };
 
+#ifdef CONFIG_READAHEAD_STATS
+#include <linux/seq_file.h>
+#include <linux/debugfs.h>
+enum ra_account {
+	/* number of readaheads */
+	RA_ACCOUNT_COUNT,	/* readahead request */
+	RA_ACCOUNT_EOF,		/* readahead request contains/beyond EOF page */
+	RA_ACCOUNT_CHIT,	/* readahead request covers some cached pages */
+	RA_ACCOUNT_IOCOUNT,	/* readahead IO */
+	RA_ACCOUNT_SYNC,	/* readahead IO that is synchronous */
+	RA_ACCOUNT_MMAP,	/* readahead IO by mmap accesses */
+	/* number of readahead pages */
+	RA_ACCOUNT_SIZE,	/* readahead size */
+	RA_ACCOUNT_ASIZE,	/* readahead async size */
+	RA_ACCOUNT_ACTUAL,	/* readahead actual IO size */
+	/* end mark */
+	RA_ACCOUNT_MAX,
+};
+
+static unsigned long ra_stats[RA_PATTERN_MAX][RA_ACCOUNT_MAX];
+
+static void readahead_stats(struct address_space *mapping,
+			    pgoff_t offset,
+			    unsigned long req_size,
+			    unsigned int ra_flags,
+			    pgoff_t start,
+			    unsigned int size,
+			    unsigned int async_size,
+			    int actual)
+{
+	unsigned int pattern = ra_pattern(ra_flags);
+
+	ra_stats[pattern][RA_ACCOUNT_COUNT]++;
+	ra_stats[pattern][RA_ACCOUNT_SIZE] += size;
+	ra_stats[pattern][RA_ACCOUNT_ASIZE] += async_size;
+	ra_stats[pattern][RA_ACCOUNT_ACTUAL] += actual;
+
+	if (actual < size) {
+		if (start + size >
+		    (i_size_read(mapping->host) - 1) >> PAGE_CACHE_SHIFT)
+			ra_stats[pattern][RA_ACCOUNT_EOF]++;
+		else
+			ra_stats[pattern][RA_ACCOUNT_CHIT]++;
+	}
+
+	if (!actual)
+		return;
+
+	ra_stats[pattern][RA_ACCOUNT_IOCOUNT]++;
+
+	if (start <= offset && start + size > offset)
+		ra_stats[pattern][RA_ACCOUNT_SYNC]++;
+
+	if (ra_flags & READAHEAD_MMAP)
+		ra_stats[pattern][RA_ACCOUNT_MMAP]++;
+}
+
+static int readahead_stats_show(struct seq_file *s, void *_)
+{
+	unsigned long i;
+	unsigned long count, iocount;
+
+	seq_printf(s, "%-10s %10s %10s %10s %10s %10s %10s %10s %10s %10s\n",
+			"pattern",
+			"readahead", "eof_hit", "cache_hit",
+			"io", "sync_io", "mmap_io",
+			"size", "async_size", "io_size");
+
+	for (i = 0; i < RA_PATTERN_MAX; i++) {
+		count = ra_stats[i][RA_ACCOUNT_COUNT];
+		iocount = ra_stats[i][RA_ACCOUNT_IOCOUNT];
+		/*
+		 * avoid division-by-zero
+		 */
+		if (count == 0)
+			count = 1;
+		if (iocount == 0)
+			iocount = 1;
+
+		seq_printf(s, "%-10s %10lu %10lu %10lu %10lu %10lu %10lu "
+			   "%10lu %10lu %10lu\n",
+				ra_pattern_names[i],
+				ra_stats[i][RA_ACCOUNT_COUNT],
+				ra_stats[i][RA_ACCOUNT_EOF],
+				ra_stats[i][RA_ACCOUNT_CHIT],
+				ra_stats[i][RA_ACCOUNT_IOCOUNT],
+				ra_stats[i][RA_ACCOUNT_SYNC],
+				ra_stats[i][RA_ACCOUNT_MMAP],
+				ra_stats[i][RA_ACCOUNT_SIZE]   / count,
+				ra_stats[i][RA_ACCOUNT_ASIZE]  / count,
+				ra_stats[i][RA_ACCOUNT_ACTUAL] / iocount);
+	}
+
+	return 0;
+}
+
+static int readahead_stats_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, readahead_stats_show, NULL);
+}
+
+static ssize_t readahead_stats_write(struct file *file, const char __user *buf,
+				     size_t size, loff_t *offset)
+{
+	memset(ra_stats, 0, sizeof(ra_stats));
+	return size;
+}
+
+static struct file_operations readahead_stats_fops = {
+	.owner		= THIS_MODULE,
+	.open		= readahead_stats_open,
+	.write		= readahead_stats_write,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static struct dentry *ra_debug_root;
+
+static int debugfs_create_readahead(void)
+{
+	struct dentry *debugfs_stats;
+
+	ra_debug_root = debugfs_create_dir("readahead", NULL);
+	if (!ra_debug_root)
+		goto out;
+
+	debugfs_stats = debugfs_create_file("stats", 0644, ra_debug_root,
+					    NULL, &readahead_stats_fops);
+	if (!debugfs_stats)
+		goto out;
+
+	return 0;
+out:
+	printk(KERN_ERR "readahead: failed to create debugfs entries\n");
+	return -ENOMEM;
+}
+
+static int __init readahead_init(void)
+{
+	debugfs_create_readahead();
+	return 0;
+}
+
+static void __exit readahead_exit(void)
+{
+	debugfs_remove_recursive(ra_debug_root);
+}
+
+module_init(readahead_init);
+module_exit(readahead_exit);
+#endif
+
+static void readahead_event(struct address_space *mapping,
+			    pgoff_t offset,
+			    unsigned long req_size,
+			    unsigned int ra_flags,
+			    pgoff_t start,
+			    unsigned int size,
+			    unsigned int async_size,
+			    unsigned int actual)
+{
+#ifdef CONFIG_READAHEAD_STATS
+	readahead_stats(mapping, offset, req_size, ra_flags,
+			start, size, async_size, actual);
+	readahead_stats(mapping, offset, req_size,
+			RA_PATTERN_ALL << READAHEAD_PATTERN_SHIFT,
+			start, size, async_size, actual);
+#endif
+	trace_readahead(mapping, offset, req_size, ra_flags,
+			start, size, async_size, actual);
+}
+
 /*
  * Initialise a struct file's readahead state.  Assumes that the caller has
  * memset *ra to zero.

@@ -289,7 +462,7 @@ int force_page_cache_readahead(struct ad
 		nr_to_read -= this_chunk;
 	}
 
-	trace_readahead(mapping, offset, nr_to_read,
+	readahead_event(mapping, offset, nr_to_read,
 			RA_PATTERN_FADVISE << READAHEAD_PATTERN_SHIFT,
 			offset, nr_to_read, 0, ret);
 

@@ -320,7 +493,7 @@ unsigned long ra_submit(struct file_ra_s
 	actual = __do_page_cache_readahead(mapping, filp,
 					ra->start, ra->size, ra->async_size);
 
-	trace_readahead(mapping, offset, req_size, ra->ra_flags,
+	readahead_event(mapping, offset, req_size, ra->ra_flags,
 			ra->start, ra->size, ra->async_size, actual);
 
 	return actual;

@@ -283,3 +283,16 @@ config NOMMU_INITIAL_TRIM_EXCESS
 	  of 1 says that all excess pages should be trimmed.
 
 	  See Documentation/nommu-mmap.txt for more information.
+
+config READAHEAD_STATS
+	bool "Collect page-cache readahead stats"
+	depends on DEBUG_FS
+	default y
+	help
+	  Enable readahead events accounting. Usage:
+
+	  # mount -t debugfs none /debug
+
+	  # echo > /debug/readahead/stats  # reset counters
+	  # do benchmarks
+	  # cat /debug/readahead/stats     # check counters

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Some applications (eg. blkid, id3tool etc.) seek around the file to get information. For example, blkid does

	     seek to	0
	     read	1024
	     seek to	1536
	     read	16384

The start-of-file readahead heuristic is wrong for them, whose access pattern can be identified by lseek() calls.

So test-and-set a READAHEAD_LSEEK flag on lseek() and don't do start-of-file readahead on seeing it. Proposed by Linus.

---
 fs/read_write.c    |    3 +++
 include/linux/fs.h |    1 +
 mm/readahead.c     |    5 +++++
 3 files changed, 9 insertions(+)

@@ -625,6 +625,11 @@ ondemand_readahead(struct address_space 
 	if (!offset) {
 		ra_set_pattern(ra, RA_PATTERN_INITIAL);
 		ra->start = offset;
+		if ((ra->ra_flags & READAHEAD_LSEEK) && req_size <= max) {
+			ra->size = req_size;
+			ra->async_size = 0;
+			goto readit;
+		}
 		ra->size = get_init_ra_size(req_size, max);
 		ra->async_size = ra->size > req_size ?
 				 ra->size - req_size : ra->size;

@@ -71,6 +71,9 @@ generic_file_llseek_unlocked(struct file
 		file->f_version = 0;
 	}
 
+	if (!(file->f_ra.ra_flags & READAHEAD_LSEEK))
+		file->f_ra.ra_flags |= READAHEAD_LSEEK;
+
 	return offset;
 }
 EXPORT_SYMBOL(generic_file_llseek_unlocked);

@@ -899,6 +899,7 @@ struct file_ra_state {
 #define	READAHEAD_MMAP_MISS	0x0000ffff /* cache misses for mmap access */
 #define READAHEAD_THRASHED	0x10000000
 #define	READAHEAD_MMAP		0x20000000
+#define	READAHEAD_LSEEK		0x40000000 /* be conservative after lseek() */
 
 /*
  * Which policy makes decision to do the current read-ahead IO?

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Replace the hole scan functions with more fast versions:
- radix_tree_next_hole(root, index, max_scan)
- radix_tree_prev_hole(root, index, max_scan)

---
 lib/radix-tree.c |   85 +++++++++++++++++++++++++++++++++++++++------
 1 file changed, 74 insertions(+), 11 deletions(-)

@@ -609,6 +609,24 @@ int radix_tree_tag_get(struct radix_tree
 }
 EXPORT_SYMBOL(radix_tree_tag_get);
 
+/*
+ * Find the bottom radix tree node that contains @index.
+ * Return NULL if @index is hole, or is the special root node.
+ */
+static struct radix_tree_node *
+radix_tree_lookup_node(struct radix_tree_root *root, unsigned long index)
+{
+	void *slot;
+
+	slot = radix_tree_lookup_element(root, index, 1);
+	if (!slot || slot == &root->rnode)
+		return NULL;
+
+	slot -= (index & RADIX_TREE_MAP_MASK) * sizeof(void *);
+
+	return container_of(slot, struct radix_tree_node, slots);
+}
+
 /**
  *	radix_tree_next_hole    -    find the next hole (not-present entry)
  *	@root:		tree root

@@ -630,18 +648,41 @@ EXPORT_SYMBOL(radix_tree_tag_get);
  *	under rcu_read_lock.
  */
 unsigned long radix_tree_next_hole(struct radix_tree_root *root,
-				unsigned long index, unsigned long max_scan)
+				   unsigned long index, unsigned long max_scan)
 {
-	unsigned long i;
+	struct radix_tree_node *node;
+	unsigned long origin = index;
+	int i;
+
+	node = rcu_dereference(root->rnode);
+	if (node == NULL)
+		return index;
+
+	if (!radix_tree_is_indirect_ptr(node))
+		return index ? index : 1;
 
-	for (i = 0; i < max_scan; i++) {
-		if (!radix_tree_lookup(root, index))
+	while (index - origin < max_scan) {
+		node = radix_tree_lookup_node(root, index);
+		if (!node)
 			break;
-		index++;
-		if (index == 0)
+
+		if (node->count == RADIX_TREE_MAP_SIZE) {
+			index = (index | RADIX_TREE_MAP_MASK) + 1;
+			goto check_overflow;
+		}
+
+		for (i = index & RADIX_TREE_MAP_MASK;
+		     i < RADIX_TREE_MAP_SIZE;
+		     i++, index++)
+			if (rcu_dereference(node->slots[i]) == NULL)
+				goto out;
+
+check_overflow:
+		if (unlikely(index == 0))
 			break;
 	}
 
+out:
 	return index;
 }
 EXPORT_SYMBOL(radix_tree_next_hole);

@@ -669,16 +710,38 @@ EXPORT_SYMBOL(radix_tree_next_hole);
 unsigned long radix_tree_prev_hole(struct radix_tree_root *root,
 				   unsigned long index, unsigned long max_scan)
 {
-	unsigned long i;
+	struct radix_tree_node *node;
+	unsigned long origin = index;
+	int i;
+
+	node = rcu_dereference(root->rnode);
+	if (node == NULL)
+		return index;
+
+	if (!radix_tree_is_indirect_ptr(node))
+		return index ? index : ULONG_MAX;
 
-	for (i = 0; i < max_scan; i++) {
-		if (!radix_tree_lookup(root, index))
+	while (origin - index < max_scan) {
+		node = radix_tree_lookup_node(root, index);
+		if (!node)
 			break;
-		index--;
-		if (index == LONG_MAX)
+
+		if (node->count == RADIX_TREE_MAP_SIZE) {
+			index = (index - RADIX_TREE_MAP_SIZE) |
+					 RADIX_TREE_MAP_MASK;
+			goto check_underflow;
+		}
+
+		for (i = index & RADIX_TREE_MAP_MASK; i >= 0; i--, index--)
+			if (rcu_dereference(node->slots[i]) == NULL)
+				goto out;
+
+check_underflow:
+		if (unlikely(index == ULONG_MAX))
 			break;
 	}
 
+out:
 	return index;
 }
 EXPORT_SYMBOL(radix_tree_prev_hole);

Linus

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Wouldn't that trigger on lseeks to end of file to get the size?

OG.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Well, you'd only ever do that with a raw block device, no (if even that:
more "raw block device" tools just use the BLKSIZE64 ioctl etc)? Any sane regular file accessor will do 'fstat()' instead.

And do we care about startup speed of ramping up read-ahead from the beginning? In fact, the problem case that caused this was literally 'blkid' on a block device - and the fact that the kernel tried to read-ahead TOO MUCh rather than too little.

If somebody is really doing lots of serial reading, the read-ahead code will figure it out very quickly. The case this worries about is just the _first_ read, where the question is one of "do we think it might be seeking around, or does it look like the user is going to just read the whole thing"?

IOW, if you start off with a SEEK_END, I think it's reasonable to expect it to _not_ read the whole thing.

Linus

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I've seen a lot of:
int fd = open(...);
size = lseek(fd, 0, SEEK_END);
lseek(fd, 0, SEEK_SET);

  data = malloc(size);
  read(fd, data, size);
  close(fd);

Why not fstat? I don't know. Perhaps a case of cargo culting, perhaps a case of "other unixes suck for portability"[1]. But it's probably still there a lot in real code.

OG.

[1] In the hpux, dgux, sunos, etc sense. Not to be taken as a comment on modern BSDs.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Well, the above will work perfectly with or without the patch, since it does the read of the full size. There is no read-ahead hint necessary for that kind of single read behavior.

Rememebr: read-ahead is about filling the empty IO spaces _between_ reads, and turning many smaller reads into one bigger one. If you only have a single big read, read-ahead cannot help.

Also, keep in mind that read-ahead is not always a win. It can be a huge loss too. Which is why we have _heuristics_. They fundamentally cannot catch every case, but what they aim for is to do a good job on average.

Linus

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I'm not sure the size part makes a ton of sense. You can have really fast small devices, and large slow devices. One real world example are the Sun FMod SSD devices, which are only 22GB in size but are faster than the Intel X25-E SLC disks.

What makes it even worse for these devices is that they are often attached to fatter controllers than ahci, where command overhead is larger.

Running your script on such a device yields (I enlarged the read-count by 2, makes it more reproducible):

MARVELL SD88SA02 MP1F

rasize 1st 2nd
------------------------------------------------------------------
4k 41 MB/s 41 MB/s
16k 85 MB/s 81 MB/s
32k 102 MB/s 109 MB/s
64k 125 MB/s 144 MB/s
128k 183 MB/s 185 MB/s
256k 216 MB/s 216 MB/s
512k 216 MB/s 236 MB/s
1024k 251 MB/s 252 MB/s
2M 258 MB/s 258 MB/s
4M 266 MB/s 266 MB/s
8M 266 MB/s 266 MB/s

So for that device, 1M-2M looks like the sweet spot, with even needing 4-8M to fully reach full throughput.

I don't think this is atypical of bigger systems. Only very recently have controller started to slim down the command overhead for real, because of the SSD devices. What probably is atypical is a device that is this small yet pretty fast.

Jens Axboe

as a note from the field, I just had an application that needed to be changed because it did excessive read-ahead. it turned a 2 min reporting run into a 20 min reporting run because for this report the access was really random and the app forced large read-ahead.

David Lang

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Yeah. And the reason Wu did this patch is similar: something that _should_ have taken just quarter of a second took about 7 seconds, because read-ahead triggered on this really slow device that only feeds about 15kB/s (yes, _kilo_byte, not megabyte).

You can always use POSIX_FADVISE_RANDOM to disable it, but it's seldom something that people do. And there are real loads that have random components to them without being _entirely_ random, so in an optimal world we should just have heuristics that work well.

The problem is, it's often easier to test/debug the "good" cases, ie the cases where we _want_ read-ahead to trigger. So that probably means that we have a tendency to read-ahead too aggressively, because those cases are the ones where people can most easily look at it and say "yeah, this improves throughput of a 'dd bs=8192'".

So then when we find loads where read-ahead hurts, I think we need to take _that_ case very seriously. Because otherwise our selection bias for testing read-ahead will fail.

Linus

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Hi Fengguang,

I was doing a quick test with the patches. I was using fio to run some sequential reader threads. I have got one access to one Lun from an HP EVA. In my case it looks like with the patches throughput has come down.

Folllowing are the results.

Kernel=2.6.33-rc5 Workload=bsr iosched=cfq Filesz=1G bs=32K
AVERAGE
-------

job       Set NR  ReadBW(KB/s)   MaxClat(us)    WriteBW(KB/s)  MaxClat(us)
---       --- --  ------------   -----------    -------------  -----------
bsr       3   1   141768         130965         0              0
bsr       3   2   131979         135402         0              0
bsr       3   4   132351         420733         0              0
bsr       3   8   133152         455434         0              0
bsr       3   16  130316         674499         0              0

Kernel=2.6.33-rc5-readahead Workload=bsr iosched=cfq Filesz=1G bs=32K
AVERAGE
-------

job       Set NR  ReadBW(KB/s)   MaxClat(us)    WriteBW(KB/s)  MaxClat(us)
---       --- --  ------------   -----------    -------------  -----------
bsr       3   1   84749.3        53213          0              0
bsr       3   2   83189.7        157473         0              0
bsr       3   4   77583.3        330030         0              0
bsr       3   8   88545.7        378201         0              0
bsr       3   16  95331.7        482657         0              0

I run increasing number of sequential readers. File system is ext3 and filesize is 1G.

I have run the tests 3 times (3sets) and taken the average of it.

Thanks
Vivek

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I ran same test on a different piece of hardware. There are few SATA disks (5-6) in striped configuration behind a hardware RAID controller. Here I do see improvement in sequenetial reader performance with the patches.

Kernel=2.6.33-rc5 Workload=bsr iosched=cfq Filesz=1G bs=32K
=========================================================================
AVERAGE
-------

job       Set NR  ReadBW(KB/s)   MaxClat(us)    WriteBW(KB/s)  MaxClat(us)    
---       --- --  ------------   -----------    -------------  -----------    
bsr       3   1   147569         14369.7        0              0              
bsr       3   2   124716         243932         0              0              
bsr       3   4   123451         327665         0              0              
bsr       3   8   122486         455102         0              0              
bsr       3   16  117645         1.03957e+06    0              0              

Kernel=2.6.33-rc5-readahead Workload=bsr iosched=cfq Filesz=1G bs=32K
=========================================================================
AVERAGE
-------

job       Set NR  ReadBW(KB/s)   MaxClat(us)    WriteBW(KB/s)  MaxClat(us)    
---       --- --  ------------   -----------    -------------  -----------    
bsr       3   1   160191         22752          0              0
bsr       3   2   149343         184698         0              0
bsr       3   4   147183         430875         0              0
bsr       3   8   144568         484045         0              0
bsr       3   16  137485         1.06257e+06    0              0

Vivek

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Ah, good to know about this fast 22GB SSD.

Thanks for the data! I updated the formula to (16GB device => 1MB readahead). However the limit in this patch is only true for <4GB devices, since the default readahead size is merely 512KB.

IOW, this patch only limits the default readahead size (which is now 512KB in general and 4MB for btrfs). The user can always set any readahead size.

Right. I didn't expect such small yet fast SSD..

Thanks,
Fengguang
---
readahead: limit readahead size for small devices

Linus reports a _really_ small & slow (505kB, 15kB/s) USB device, on which blkid runs unpleasantly slow. He manages to optimize the blkid reads down to 1kB+16kB, but still kernel read-ahead turns it into 48kB.

lseek 0, read 1024 => readahead 4 pages (start of file)
lseek 1536, read 16384 => readahead 8 pages (page contiguous)

The readahead heuristics involved here are reasonable ones in general. So it's good to fix blkid with fadvise(RANDOM), as Linus already did.

For the kernel part, Linus suggests:
So maybe we could be less aggressive about read-ahead when the size of the device is small? Turning a 16kB read into a 64kB one is a big deal, when it's about 15% of the whole device!

This looks reasonable: smaller device tend to be slower (USB sticks as well as micro/mobile/old hard disks).

Given that the non-rotational attribute is not always reported, we can take disk size as a max readahead size hint. This patch uses a formula that generates the following concrete limits:

        disk size    readahead size
     (scale by 4)      (scale by 2)
               1M                8k
               4M               16k
              16M               32k
              64M               64k
             256M              128k
               1G              256k
        --------------------------- (*)
               4G              512k
              16G             1024k
              64G             2048k
             256G             4096k

(*) Since the default readahead size is 512k, this limit only takes effect for devices whose size is less than 4G.

The formula is determined on the following data, collected by script:

	#!/bin/sh

	# please make sure BDEV is not mounted or opened by others
	BDEV=sdb

	for rasize in 4 16 32 64 128 256 512 1024 2048 4096 8192
	do
		echo $rasize > /sys/block/$BDEV/queue/read_ahead_kb 
		time dd if=/dev/$BDEV of=/dev/null bs=4k count=102400
	done

The principle is, the formula shall not limit readahead size to such a degree that will impact some device's sequential read performance.

The Intel SSD is special in that its throughput increases steadily with larger readahead size. However it may take years for Linux to increase its default readahead size to 2MB, so we don't take it seriously in the formula.

SSD 80G Intel x25-M SSDSA2M080 (reported by Li Shaohua)

	rasize	1st run		2nd run
	----------------------------------
	  4k	123 MB/s	122 MB/s
	 16k  	153 MB/s	153 MB/s
	 32k	161 MB/s	162 MB/s
	 64k	167 MB/s	168 MB/s
	128k	197 MB/s	197 MB/s
	256k	217 MB/s	217 MB/s
	512k	238 MB/s	234 MB/s
	  1M	251 MB/s	248 MB/s
	  2M	259 MB/s	257 MB/s
==>	  4M	269 MB/s	264 MB/s
	  8M	266 MB/s	266 MB/s

Note that ==> points to the readahead size that yields plateau throughput.

SSD 22G MARVELL SD88SA02 MP1F (reported by Jens Axboe)

	rasize  1st             2nd
	--------------------------------
	  4k     41 MB/s         41 MB/s
	 16k     85 MB/s         81 MB/s
	 32k    102 MB/s        109 MB/s
	 64k    125 MB/s        144 MB/s
	128k    183 MB/s        185 MB/s
	256k    216 MB/s        216 MB/s
	512k    216 MB/s        236 MB/s
	1024k   251 MB/s        252 MB/s
	  2M    258 MB/s        258 MB/s
==>       4M    266 MB/s        266 MB/s
	  8M    266 MB/s        266 MB/s

SSD 30G SanDisk SATA 5000

	  4k	29.6 MB/s	29.6 MB/s	29.6 MB/s
	 16k	52.1 MB/s	52.1 MB/s	52.1 MB/s
	 32k	61.5 MB/s	61.5 MB/s	61.5 MB/s
	 64k	67.2 MB/s	67.2 MB/s	67.1 MB/s
	128k	71.4 MB/s	71.3 MB/s	71.4 MB/s
	256k	73.4 MB/s	73.4 MB/s	73.3 MB/s
==>	512k	74.6 MB/s	74.6 MB/s	74.6 MB/s
	  1M	74.7 MB/s	74.6 MB/s	74.7 MB/s
	  2M	76.1 MB/s	74.6 MB/s	74.6 MB/s

USB stick 32G Teclast CoolFlash idVendor=1307, idProduct=0165

	  4k	7.9 MB/s 	7.9 MB/s 	7.9 MB/s
	 16k	17.9 MB/s	17.9 MB/s	17.9 MB/s
	 32k	24.5 MB/s	24.5 MB/s	24.5 MB/s
	 64k	28.7 MB/s	28.7 MB/s	28.7 MB/s
	128k	28.8 MB/s	28.9 MB/s	28.9 MB/s
==>	256k	30.5 MB/s	30.5 MB/s	30.5 MB/s
	512k	30.9 MB/s	31.0 MB/s	30.9 MB/s
	  1M	31.0 MB/s	30.9 MB/s	30.9 MB/s
	  2M	30.9 MB/s	30.9 MB/s	30.9 MB/s

USB stick 4G SanDisk Cruzer idVendor=0781, idProduct=5151

	  4k	6.4 MB/s 	6.4 MB/s 	6.4 MB/s
	 16k	13.4 MB/s	13.4 MB/s	13.2 MB/s
	 32k	17.8 MB/s	17.9 MB/s	17.8 MB/s
	 64k	21.3 MB/s	21.3 MB/s	21.2 MB/s
	128k	21.4 MB/s	21.4 MB/s	21.4 MB/s
==>	256k	23.3 MB/s	23.2 MB/s	23.2 MB/s
	512k	23.3 MB/s	23.8 MB/s	23.4 MB/s
	  1M	23.8 MB/s	23.4 MB/s	23.3 MB/s
	  2M	23.4 MB/s	23.2 MB/s	23.4 MB/s

USB stick 2G idVendor=0204, idProduct=6025 SerialNumber: 08082005000113

	  4k	6.7 MB/s 	6.9 MB/s 	6.7 MB/s
	 16k	11.7 MB/s	11.7 MB/s	11.7 MB/s
	 32k	12.4 MB/s	12.4 MB/s	12.4 MB/s
   	 64k	13.4 MB/s	13.4 MB/s	13.4 MB/s
	128k	13.4 MB/s	13.4 MB/s	13.4 MB/s
==>	256k	13.6 MB/s	13.6 MB/s	13.6 MB/s
	512k	13.7 MB/s	13.7 MB/s	13.7 MB/s
	  1M	13.7 MB/s	13.7 MB/s	13.7 MB/s
	  2M	13.7 MB/s	13.7 MB/s	13.7 MB/s

Anyone has 128MB USB stick? Anyway you get satisfiable performance with >= 64k readahead size :)

---
 block/genhd.c |   23 +++++++++++++++++++++++
 1 file changed, 23 insertions(+)

@@ -518,6 +518,7 @@ void add_disk(struct gendisk *disk)
 	struct backing_dev_info *bdi;
 	dev_t devt;
 	int retval;
+	unsigned long size;
 
 	/* minors == 0 indicates to use ext devt from part0 and should
 	 * be accompanied with EXT_DEVT flag.  Make sure all

@@ -551,6 +552,28 @@ void add_disk(struct gendisk *disk)
 	retval = sysfs_create_link(&disk_to_dev(disk)->kobj, &bdi->dev->kobj,
 				   "bdi");
 	WARN_ON(retval);
+
+	/*
+	 * Limit default readahead size for small devices.
+	 *        disk size    readahead size
+	 *               1M                8k
+	 *               4M               16k
+	 *              16M               32k
+	 *              64M               64k
+	 *             256M              128k
+	 *               1G              256k
+	 *        ---------------------------
+	 *               4G              512k
+	 *              16G             1024k
+	 *              64G             2048k
+	 *             256G             4096k
+	 * Since the default readahead size is 512k, this limit
+	 * only takes effect for devices whose size is less than 4G.
+	 */
+
+	size = get_capacity(disk) >> 9;
+	size = 1UL << (ilog2(size) / 2);
+	bdi->ra_pages = min(bdi->ra_pages, size);
 }
 
 EXPORT_SYMBOL(add_disk);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Vivek,

Thank you for the quick testing!

This patchset does 3 things:

1) 512K readahead size
2) new readahead algorithms
3) new readahead tracing/stats interfaces

(1) will impact performance, while (2) _might_ impact performance in case of bugs.

Would you kindly retest the patchset with readahead size manually set to 128KB? That would help identify the root cause of the performance drop:

        DEV=sda
        echo 128 > /sys/block/$DEV/queue/read_ahead_kb

The readahead stats provided by the patchset are very useful for analyzing the problem:

mount -t debugfs none /debug

# for each benchmark:
echo > /debug/readahead/stats # reset counters # do benchmark
cat /debug/readahead/stats # check counters

Thanks,
Fengguang

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Thanks Wu, you can add my acked-by.

Jens Axboe

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I have got two paths to the HP EVA and got multipath device setup(dm-3). I noticed with vanilla kernel read_ahead_kb=128 after boot but with your patches applied it is set to 4. So looks like something went wrong with device size/capacity detection hence wrong defaults. Manually setting read_ahead_kb=512, got me better performance as compare to vanilla kernel.

AVERAGE[bsr]
-------

job       Set NR  ReadBW(KB/s)   MaxClat(us)    WriteBW(KB/s)  MaxClat(us)    
---       --- --  ------------   -----------    -------------  -----------    
bsr       3   1   190302         97937.3        0              0              
bsr       3   2   185636         223286         0              0              
bsr       3   4   185986         363658         0              0              
bsr       3   8   184352         428478         0              0              
bsr       3   16  185646         594311         0              0              

Thanks
Vivek

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

I put a printk in add_disk and noticed that for multipath device get_capacity() is returning 0 and that's why ra_pages is being set to 1.

Thanks
Vivek

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

FYI, wanted to get this on our radar... seems latest DM isn't allowing RFC readahead code to set a sane readahead default for DM devices? get_capacity() is returning 0 for DM devices (not just multipath). Vivek did share that fdisk -l does show the proper capacity for the DM device.

I haven't had a chance to look at the relevant code yet.

I've asked Vivek to cc dm-devel on any further messages he might send in response to this thread.

---------- Forwarded message ----------
From: Vivek Goyal <[email protected]>
Date: Wed, Feb 3, 2010 at 10:58 AM
Subject: Re: [PATCH 00/11] [RFC] 512K readahead size with thrashing safe readahead
To: Wu Fengguang <[email protected]>
Cc: Andrew Morton <[email protected]>, Jens Axboe <[email protected]>, Peter Zijlstra <[email protected]>, Linux Memory Management List <[email protected]>, "[email protected]" <[email protected]>, LKML <[email protected]>

I put a printk in add_disk and noticed that for multipath device get_capacity() is returning 0 and that's why ra_pages is being set to 1.

Thanks
Vivek

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected].  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

64 MB, USB full speed:
Bus 003 Device 003: ID 08ec:0011 M-Systems Flash Disk Pioneers DiskOnKey

4KB:    139.339 s, 376 kB/s
16KB:   81.0427 s, 647 kB/s
32KB:   71.8513 s, 730 kB/s
64KB:   67.3872 s, 778 kB/s
128KB:  67.5434 s, 776 kB/s
256KB:  65.9019 s, 796 kB/s
512KB:  66.2282 s, 792 kB/s

1024KB: 67.4632 s, 777 kB/s
2048KB: 69.9759 s, 749 kB/s

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Clemens,

Thanks for the data!

It seems to reach good throughput at 64KB readahead size :)

Thanks,
Fengguang

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Vivek,

Good catch, Thanks!

It makes no sense to limit readahead size for multipath or other compound devices. So we may just ignore the get_capacity() == 0 case, as in the following updated patch.

Thanks,
Fengguang
---
readahead: limit readahead size for small devices

Linus reports a _really_ small & slow (505kB, 15kB/s) USB device, on which blkid runs unpleasantly slow. He manages to optimize the blkid reads down to 1kB+16kB, but still kernel read-ahead turns it into 48kB.

lseek 0, read 1024 => readahead 4 pages (start of file)
lseek 1536, read 16384 => readahead 8 pages (page contiguous)

The readahead heuristics involved here are reasonable ones in general. So it's good to fix blkid with fadvise(RANDOM), as Linus already did.

For the kernel part, Linus suggests:
So maybe we could be less aggressive about read-ahead when the size of the device is small? Turning a 16kB read into a 64kB one is a big deal, when it's about 15% of the whole device!

This looks reasonable: smaller device tend to be slower (USB sticks as well as micro/mobile/old hard disks).

Given that the non-rotational attribute is not always reported, we can take disk size as a max readahead size hint. This patch uses a formula that generates the following concrete limits:

        disk size    readahead size
     (scale by 4)      (scale by 2)
               1M                8k
               4M               16k
              16M               32k
              64M               64k
             256M              128k
               1G              256k
        --------------------------- (*)
               4G              512k
              16G             1024k
              64G             2048k
             256G             4096k

(*) Since the default readahead size is 512k, this limit only takes effect for devices whose size is less than 4G.

The formula is determined on the following data, collected by script:

	#!/bin/sh

	# please make sure BDEV is not mounted or opened by others
	BDEV=sdb

	for rasize in 4 16 32 64 128 256 512 1024 2048 4096 8192
	do
		echo $rasize > /sys/block/$BDEV/queue/read_ahead_kb 
		time dd if=/dev/$BDEV of=/dev/null bs=4k count=102400
	done

The principle is, the formula shall not limit readahead size to such a degree that will impact some device's sequential read performance.

The Intel SSD is special in that its throughput increases steadily with larger readahead size. However it may take years for Linux to increase its default readahead size to 2MB, so we don't take it seriously in the formula.

SSD 80G Intel x25-M SSDSA2M080 (reported by Li Shaohua)

	rasize	1st run		2nd run
	----------------------------------
	  4k	123 MB/s	122 MB/s
	 16k  	153 MB/s	153 MB/s
	 32k	161 MB/s	162 MB/s
	 64k	167 MB/s	168 MB/s
	128k	197 MB/s	197 MB/s
	256k	217 MB/s	217 MB/s
	512k	238 MB/s	234 MB/s
	  1M	251 MB/s	248 MB/s
	  2M	259 MB/s	257 MB/s
==>	  4M	269 MB/s	264 MB/s
	  8M	266 MB/s	266 MB/s

Note that ==> points to the readahead size that yields plateau throughput.

SSD 22G MARVELL SD88SA02 MP1F (reported by Jens Axboe)

	rasize  1st             2nd
	--------------------------------
	  4k     41 MB/s         41 MB/s
	 16k     85 MB/s         81 MB/s
	 32k    102 MB/s        109 MB/s
	 64k    125 MB/s        144 MB/s
	128k    183 MB/s        185 MB/s
	256k    216 MB/s        216 MB/s
	512k    216 MB/s        236 MB/s
	1024k   251 MB/s        252 MB/s
	  2M    258 MB/s        258 MB/s
==>       4M    266 MB/s        266 MB/s
	  8M    266 MB/s        266 MB/s

SSD 30G SanDisk SATA 5000

	  4k	29.6 MB/s	29.6 MB/s	29.6 MB/s
	 16k	52.1 MB/s	52.1 MB/s	52.1 MB/s
	 32k	61.5 MB/s	61.5 MB/s	61.5 MB/s
	 64k	67.2 MB/s	67.2 MB/s	67.1 MB/s
	128k	71.4 MB/s	71.3 MB/s	71.4 MB/s
	256k	73.4 MB/s	73.4 MB/s	73.3 MB/s
==>	512k	74.6 MB/s	74.6 MB/s	74.6 MB/s
	  1M	74.7 MB/s	74.6 MB/s	74.7 MB/s
	  2M	76.1 MB/s	74.6 MB/s	74.6 MB/s

USB stick 32G Teclast CoolFlash idVendor=1307, idProduct=0165

	  4k	7.9 MB/s 	7.9 MB/s 	7.9 MB/s
	 16k	17.9 MB/s	17.9 MB/s	17.9 MB/s
	 32k	24.5 MB/s	24.5 MB/s	24.5 MB/s
	 64k	28.7 MB/s	28.7 MB/s	28.7 MB/s
	128k	28.8 MB/s	28.9 MB/s	28.9 MB/s
==>	256k	30.5 MB/s	30.5 MB/s	30.5 MB/s
	512k	30.9 MB/s	31.0 MB/s	30.9 MB/s
	  1M	31.0 MB/s	30.9 MB/s	30.9 MB/s
	  2M	30.9 MB/s	30.9 MB/s	30.9 MB/s

USB stick 4G SanDisk Cruzer idVendor=0781, idProduct=5151

	  4k	6.4 MB/s 	6.4 MB/s 	6.4 MB/s
	 16k	13.4 MB/s	13.4 MB/s	13.2 MB/s
	 32k	17.8 MB/s	17.9 MB/s	17.8 MB/s
	 64k	21.3 MB/s	21.3 MB/s	21.2 MB/s
	128k	21.4 MB/s	21.4 MB/s	21.4 MB/s
==>	256k	23.3 MB/s	23.2 MB/s	23.2 MB/s
	512k	23.3 MB/s	23.8 MB/s	23.4 MB/s
	  1M	23.8 MB/s	23.4 MB/s	23.3 MB/s
	  2M	23.4 MB/s	23.2 MB/s	23.4 MB/s

USB stick 2G idVendor=0204, idProduct=6025 SerialNumber: 08082005000113

	  4k	6.7 MB/s 	6.9 MB/s 	6.7 MB/s
	 16k	11.7 MB/s	11.7 MB/s	11.7 MB/s
	 32k	12.4 MB/s	12.4 MB/s	12.4 MB/s
   	 64k	13.4 MB/s	13.4 MB/s	13.4 MB/s
	128k	13.4 MB/s	13.4 MB/s	13.4 MB/s
==>	256k	13.6 MB/s	13.6 MB/s	13.6 MB/s
	512k	13.7 MB/s	13.7 MB/s	13.7 MB/s
	  1M	13.7 MB/s	13.7 MB/s	13.7 MB/s
	  2M	13.7 MB/s	13.7 MB/s	13.7 MB/s

64 MB, USB full speed (collected by Clemens Ladisch)
Bus 003 Device 003: ID 08ec:0011 M-Systems Flash Disk Pioneers DiskOnKey

	4KB:    139.339 s, 376 kB/s
	16KB:   81.0427 s, 647 kB/s
	32KB:   71.8513 s, 730 kB/s
==>	64KB:   67.3872 s, 778 kB/s
	128KB:  67.5434 s, 776 kB/s
	256KB:  65.9019 s, 796 kB/s
	512KB:  66.2282 s, 792 kB/s
	1024KB: 67.4632 s, 777 kB/s
	2048KB: 69.9759 s, 749 kB/s

---
 block/genhd.c |   24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)

@@ -518,6 +518,7 @@ void add_disk(struct gendisk *disk)
 	struct backing_dev_info *bdi;
 	dev_t devt;
 	int retval;
+	unsigned long size;
 
 	/* minors == 0 indicates to use ext devt from part0 and should
 	 * be accompanied with EXT_DEVT flag.  Make sure all

@@ -551,6 +552,29 @@ void add_disk(struct gendisk *disk)
 	retval = sysfs_create_link(&disk_to_dev(disk)->kobj, &bdi->dev->kobj,
 				   "bdi");
 	WARN_ON(retval);
+
+	/*
+	 * Limit default readahead size for small devices.
+	 *        disk size    readahead size
+	 *               1M                8k
+	 *               4M               16k
+	 *              16M               32k
+	 *              64M               64k
+	 *             256M              128k
+	 *               1G              256k
+	 *        ---------------------------
+	 *               4G              512k
+	 *              16G             1024k
+	 *              64G             2048k
+	 *             256G             4096k
+	 * Since the default readahead size is 512k, this limit
+	 * only takes effect for devices whose size is less than 4G.
+	 */
+	if (get_capacity(disk)) {
+		size = get_capacity(disk) >> 9;
+		size = 1UL << (ilog2(size) / 2);
+		bdi->ra_pages = min(bdi->ra_pages, size);
+	}
 }
 
 EXPORT_SYMBOL(add_disk);

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"[email protected]"> [email protected] </a>

Vivek,

This looks good, thank you for the data! I added them to the changelog :)

Thanks,
Fengguang
---
readahead: bump up the default readahead size

Use 512kb max readahead size, and 32kb min readahead size.

The former helps io performance for common workloads. The latter will be used in the thrashing safe context readahead.

-- Rationals on the 512kb size --

I believe it yields more I/O throughput without noticeably increasing I/O latency for today's HDD.

For example, for a 100MB/s and 8ms access time HDD:

io_size KB  access_time  transfer_time  io_latency   util%   throughput KB/s
4           8             0.04           8.04        0.49%    497.57  
8           8             0.08           8.08        0.97%    990.33  
16          8             0.16           8.16        1.92%   1961.69 
32          8             0.31           8.31        3.76%   3849.62 
64          8             0.62           8.62        7.25%   7420.29 
128         8             1.25           9.25       13.51%  13837.84
256         8             2.50          10.50       23.81%  24380.95
512         8             5.00          13.00       38.46%  39384.62
1024        8            10.00          18.00       55.56%  56888.89
2048        8            20.00          28.00       71.43%  73142.86
4096        8            40.00          48.00       83.33%  85333.33

The 128KB => 512KB readahead size boosts IO throughput from ~13MB/s to ~39MB/s, while merely increases (minimal) IO latency from 9.25ms to 13ms.

As for SSD, I find that Intel X25-M SSD desires large readahead size even for sequential reads:

	rasize	1st run		2nd run
	----------------------------------
	  4k	123 MB/s	122 MB/s
	 16k  	153 MB/s	153 MB/s
	 32k	161 MB/s	162 MB/s
	 64k	167 MB/s	168 MB/s
	128k	197 MB/s	197 MB/s
	256k	217 MB/s	217 MB/s
	512k	238 MB/s	234 MB/s
	  1M	251 MB/s	248 MB/s
	  2M	259 MB/s	257 MB/s
   	  4M	269 MB/s	264 MB/s
	  8M	266 MB/s	266 MB/s

The two other impacts of an enlarged readahead size are

- memory footprint (caused by readahead miss)
Sequential readahead hit ratio is pretty high regardless of max readahead size; the extra memory footprint is mainly caused by enlarged mmap read-around.
I measured my desktop:
- under Xwindow:
128KB readahead hit ratio = 143MB/230MB = 62% 512KB readahead hit ratio = 138MB/248MB = 55%
1MB readahead hit ratio = 130MB/253MB = 51%
- under console: (seems more stable than the Xwindow data)
128KB readahead hit ratio = 30MB/56MB = 53%
1MB readahead hit ratio = 30MB/59MB = 51%
So the impact to memory footprint looks acceptable.

- readahead thrashing
It will now cost 1MB readahead buffer per stream. Memory tight systems typically do not run multiple streams; but if they do so, it should help I/O performance as long as we can avoid thrashing, which can be achieved with the following patches.

-- Benchmarks by Vivek Goyal --

I have got two paths to the HP EVA and got multipath device setup(dm-3). I run increasing number of sequential readers. File system is ext3 and filesize is 1G.
I have run the tests 3 times (3sets) and taken the average of it.

Workload=bsr iosched=cfq Filesz=1G bs=32K
======================================================================

                    2.6.33-rc5                2.6.33-rc5-readahead
job   Set NR  ReadBW(KB/s)   MaxClat(us)    ReadBW(KB/s)   MaxClat(us)
---   --- --  ------------   -----------    ------------   -----------
bsr   3   1   141768         130965         190302         97937.3    
bsr   3   2   131979         135402         185636         223286     
bsr   3   4   132351         420733         185986         363658     
bsr   3   8   133152         455434         184352         428478     
bsr   3   16  130316         674499         185646         594311     

I ran same test on a different piece of hardware. There are few SATA disks (5-6) in striped configuration behind a hardware RAID controller.

Workload=bsr iosched=cfq Filesz=1G bs=32K
======================================================================

                    2.6.33-rc5                2.6.33-rc5-readahead
job   Set NR  ReadBW(KB/s)   MaxClat(us)    ReadBW(KB/s)   MaxClat(us)    
---   --- --  ------------   -----------    ------------   -----------    
bsr   3   1   147569         14369.7        160191         22752          
bsr   3   2   124716         243932         149343         184698         
bsr   3   4   123451         327665         147183         430875         
bsr   3   8   122486         455102         144568         484045         
bsr   3   16  117645         1.03957e+06    137485         1.06257e+06    

---
 include/linux/mm.h |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

@@ -1184,8 +1184,8 @@ int write_one_page(struct page *page, in
 void task_dirty_inc(struct task_struct *tsk);
 
 /* readahead.c */
-#define VM_MAX_READAHEAD	128	/* kbytes */
-#define VM_MIN_READAHEAD	16	/* kbytes (includes current page) */
+#define VM_MAX_READAHEAD	512	/* kbytes */
+#define VM_MIN_READAHEAD	32	/* kbytes (includes current page) */
 
 int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 			pgoff_t offset, unsigned long nr_to_read);

Thanks. This patch fixes the issue of read_ahead_kb being set to 4kb on device mapper targets.

Thanks
Vivek

Thank you! Updated patch as follows.

To avoid unnecessary dependency, EXTRACT_TRACE_SYMBOL() calls are leaved out for now.

Thanks,
Fengguang
---
readahead: add tracing event

Example output:

# echo 1 > /debug/tracing/events/readahead/enable
# cp test-file /dev/null
# cat /debug/tracing/trace  # trimmed output
readahead-initial(dev=0:15, ino=100177, req=0+2, ra=0+4-2, async=0) = 4
readahead-subsequent(dev=0:15, ino=100177, req=2+2, ra=4+8-8, async=1) = 8
readahead-subsequent(dev=0:15, ino=100177, req=4+2, ra=12+16-16, async=1) = 16
readahead-subsequent(dev=0:15, ino=100177, req=12+2, ra=28+32-32, async=1) = 32
readahead-subsequent(dev=0:15, ino=100177, req=28+2, ra=60+60-60, async=1) = 24
readahead-subsequent(dev=0:15, ino=100177, req=60+2, ra=120+60-60, async=1) = 0

---
 include/trace/events/readahead.h |   78 +++++++++++++++++++++++++++++
 mm/readahead.c                   |   11 ++++
 2 files changed, 89 insertions(+)

@@ -0,0 +1,78 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM readahead
+
+#if !defined(_TRACE_READAHEAD_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_READAHEAD_H
+
+#include <linux/tracepoint.h>
+
+#define show_pattern_name(val)						   \
+	__print_symbolic(val,						   \
+			{ RA_PATTERN_INITIAL,		"initial"	}, \
+			{ RA_PATTERN_SUBSEQUENT,	"subsequent"	}, \
+			{ RA_PATTERN_CONTEXT,		"context"	}, \
+			{ RA_PATTERN_THRASH,		"thrash"	}, \
+			{ RA_PATTERN_MMAP_AROUND,	"around"	}, \
+			{ RA_PATTERN_FADVISE,		"fadvise"	}, \
+			{ RA_PATTERN_RANDOM,		"random"	}, \
+			{ RA_PATTERN_ALL,		"all"		})
+
+/*
+ * Tracepoint for guest mode entry.
+ */
+TRACE_EVENT(readahead,
+	TP_PROTO(struct address_space *mapping,
+		 pgoff_t offset,
+		 unsigned long req_size,
+		 unsigned int ra_flags,
+		 pgoff_t start,
+		 unsigned int size,
+		 unsigned int async_size,
+		 unsigned int actual),
+
+	TP_ARGS(mapping, offset, req_size,
+		ra_flags, start, size, async_size, actual),
+
+	TP_STRUCT__entry(
+		__field(	dev_t,		dev		)
+		__field(	ino_t,		ino		)
+		__field(	pgoff_t,	offset		)
+		__field(	unsigned long,	req_size	)
+		__field(	unsigned int,	pattern		)
+		__field(	pgoff_t,	start		)
+		__field(	unsigned int,	size		)
+		__field(	unsigned int,	async_size	)
+		__field(	unsigned int,	actual		)
+	),
+
+	TP_fast_assign(
+		__entry->dev		= mapping->host->i_sb->s_dev;
+		__entry->ino		= mapping->host->i_ino;
+		__entry->pattern	= ra_pattern(ra_flags);
+		__entry->offset		= offset;
+		__entry->req_size	= req_size;
+		__entry->start		= start;
+		__entry->size		= size;
+		__entry->async_size	= async_size;
+		__entry->actual		= actual;
+	),
+
+	TP_printk("readahead-%s(dev=%d:%d, ino=%lu, "
+		  "req=%lu+%lu, ra=%lu+%d-%d, async=%d) = %d",
+			show_pattern_name(__entry->pattern),
+			MAJOR(__entry->dev),
+			MINOR(__entry->dev),
+			__entry->ino,
+			__entry->offset,
+			__entry->req_size,
+			__entry->start,
+			__entry->size,
+			__entry->async_size,
+			__entry->start > __entry->offset,
+			__entry->actual)
+);
+
+#endif /* _TRACE_READAHEAD_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>

@@ -19,6 +19,9 @@
 #include <linux/pagevec.h>
 #include <linux/pagemap.h>
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/readahead.h>
+
 /*
  * Set async size to 1/# of the thrashing threshold.
  */

@@ -274,6 +277,11 @@ int force_page_cache_readahead(struct ad
 		offset += this_chunk;
 		nr_to_read -= this_chunk;
 	}
+
+	trace_readahead(mapping, offset, nr_to_read,
+			RA_PATTERN_FADVISE << READAHEAD_PATTERN_SHIFT,
+			offset, nr_to_read, 0, ret);
+
 	return ret;
 }
 

@@ -301,6 +309,9 @@ unsigned long ra_submit(struct file_ra_s
 	actual = __do_page_cache_readahead(mapping, filp,
 					ra->start, ra->size, ra->async_size);
 
+	trace_readahead(mapping, offset, req_size, ra->ra_flags,
+			ra->start, ra->size, ra->async_size, actual);
+
 	return actual;
 }