[patch 0/2] adaptive readahead fixes

[patch 0/2] adaptive readahead fixes

[Fengguang Wu]

Andrew,

Here are two bug fix patches against -mm, with their recommended placement:


 readahead-events-accounting.patch
 readahead-rescue_pages.patch
 readahead-sysctl-parameters.patch
+readahead-sysctl-parameters-fix.patch
 readahead-min-max-sizes.patch
 readahead-state-based-method-aging-accounting.patch
 readahead-state-based-method-routines.patch

 readahead-loop-case.patch
 readahead-nfsd-case.patch
 readahead-nfsd-case-fix.patch
+readahead-nfsd-case-fix-uninitialized-ra_min-ra_max.patch
 readahead-turn-on-by-default.patch
 readahead-remove-size-limit-on-read_ahead_kb.patch
 readahead-remove-size-limit-of-max_sectors_kb-on-read_ahead_kb.patch


readahead-sysctl-parameters-fix.patch
=====================================

Two trivial fixes.
// Was three, but you were so swift on CTL_UNNUMBERED :-)


readahead-nfsd-case-fix-uninitialized-ra_min-ra_max.patch
=========================================================

The NFS benchmark is ran again with the bug fixed.
The summary is included in the patch as changelog.
The raw numbers are here:

pattern=f rsize=8k

ratio=1 	3.04s clock  0.64s kernel  0.01s user  224+1987 cs
ratio=1 	3.06s clock  0.66s kernel  0.02s user  199+2340 cs
ratio=1 	3.10s clock  0.63s kernel  0.02s user  254+2264 cs

ratio=50 	2.59s clock  0.60s kernel  0.02s user  106+1180 cs
ratio=50 	2.43s clock  0.59s kernel  0.01s user  94+1189 cs
ratio=50 	2.50s clock  0.59s kernel  0.02s user  98+1174 cs

pattern=f rsize=32k

ratio=1 	2.39s clock  0.29s kernel  0.02s user  1046+59 cs
ratio=1 	2.40s clock  0.29s kernel  0.01s user  1040+69 cs
ratio=1 	2.41s clock  0.28s kernel  0.02s user  1114+55 cs

ratio=50 	2.18s clock  0.32s kernel  0.02s user  227+213 cs
ratio=50 	2.15s clock  0.32s kernel  0.02s user  230+215 cs
ratio=50 	2.17s clock  0.33s kernel  0.01s user  225+208 cs

pattern=f rsize=128k

ratio=1 	2.42s clock  0.32s kernel  0.01s user  436+131 cs
ratio=1 	2.38s clock  0.33s kernel  0.02s user  441+128 cs
ratio=1 	2.39s clock  0.31s kernel  0.01s user  448+122 cs

ratio=50 	2.36s clock  0.30s kernel  0.01s user  202+67 cs
ratio=50 	2.21s clock  0.31s kernel  0.02s user  194+72 cs
ratio=50 	2.47s clock  0.30s kernel  0.01s user  201+63 cs

pattern=ff rsize=8k

ratio=1 	12.19s clock  1.18s kernel  0.38s user  4152+13042 cs
ratio=1 	12.88s clock  1.21s kernel  0.33s user  4564+12574 cs
ratio=1 	12.26s clock  1.22s kernel  0.38s user  4857+12453 cs

ratio=50 	6.53s clock  1.27s kernel  0.32s user  174+2256 cs
ratio=50 	6.33s clock  1.27s kernel  0.33s user  164+2252 cs
ratio=50 	6.35s clock  1.24s kernel  0.35s user  151+2264 cs

pattern=ff rsize=32k

ratio=1 	14.49s clock  0.90s kernel  0.37s user  2904+9906 cs
ratio=1 	14.55s clock  1.00s kernel  0.34s user  2899+9803 cs
ratio=1 	14.81s clock  1.00s kernel  0.31s user  2910+10147 cs

ratio=50 	5.48s clock  0.65s kernel  0.30s user  177+512 cs
ratio=50 	5.42s clock  0.68s kernel  0.29s user  181+509 cs
ratio=50 	5.47s clock  0.65s kernel  0.29s user  175+521 cs

pattern=ff rsize=128k

ratio=1 	15.87s clock  0.90s kernel  0.32s user  1496+8738 cs
ratio=1 	15.33s clock  0.89s kernel  0.34s user  1718+8350 cs
ratio=1 	16.17s clock  0.91s kernel  0.32s user  1706+7959 cs

ratio=50 	5.18s clock  0.56s kernel  0.28s user  175+255 cs
ratio=50 	5.22s clock  0.57s kernel  0.30s user  172+258 cs
ratio=50 	5.16s clock  0.54s kernel  0.30s user  168+260 cs

pattern=d rsize=8k

ratio=1 	2.86s clock  0.45s kernel  0.07s user  7961+729 cs
ratio=1 	2.87s clock  0.50s kernel  0.07s user  8036+627 cs
ratio=1 	2.80s clock  0.51s kernel  0.07s user  7986+640 cs

ratio=50 	2.58s clock  0.51s kernel  0.07s user  8873+177 cs
ratio=50 	2.39s clock  0.51s kernel  0.05s user  8782+194 cs
ratio=50 	2.48s clock  0.51s kernel  0.06s user  8884+171 cs

pattern=d rsize=32k

ratio=1 	2.49s clock  0.40s kernel  0.05s user  7428+83 cs
ratio=1 	2.57s clock  0.40s kernel  0.06s user  7425+86 cs
ratio=1 	2.54s clock  0.41s kernel  0.06s user  7427+83 cs

ratio=50 	2.02s clock  0.40s kernel  0.08s user  7478+169 cs
ratio=50 	1.99s clock  0.43s kernel  0.05s user  7488+162 cs
ratio=50 	1.95s clock  0.39s kernel  0.05s user  7483+169 cs

pattern=d rsize=128k

ratio=1 	2.17s clock  0.41s kernel  0.05s user  7248+120 cs
ratio=1 	2.13s clock  0.42s kernel  0.05s user  7242+133 cs
ratio=1 	2.23s clock  0.43s kernel  0.07s user  7247+121 cs

ratio=50 	2.09s clock  0.39s kernel  0.06s user  7315+97 cs
ratio=50 	1.93s clock  0.41s kernel  0.05s user  7314+131 cs
ratio=50 	1.97s clock  0.43s kernel  0.07s user  7317+132 cs

pattern=dd rsize=8k

ratio=1 	8.23s clock  1.04s kernel  0.18s user  15564+1306 cs
ratio=1 	8.20s clock  1.06s kernel  0.17s user  15465+1251 cs
ratio=1 	8.04s clock  1.03s kernel  0.15s user  15487+1226 cs

ratio=50 	8.06s clock  1.03s kernel  0.18s user  16992+291 cs
ratio=50 	8.14s clock  0.98s kernel  0.19s user  17011+297 cs
ratio=50 	7.77s clock  1.01s kernel  0.16s user  16920+292 cs

pattern=dd rsize=32k

ratio=1 	8.46s clock  0.91s kernel  0.16s user  14248+593 cs
ratio=1 	8.27s clock  0.85s kernel  0.17s user  14266+564 cs
ratio=1 	8.09s clock  0.86s kernel  0.18s user  14255+551 cs

ratio=50 	6.85s clock  0.89s kernel  0.18s user  14304+301 cs
ratio=50 	6.79s clock  0.93s kernel  0.14s user  14302+274 cs
ratio=50 	6.99s clock  0.93s kernel  0.17s user  14296+303 cs

pattern=dd rsize=128k

ratio=1 	7.87s clock  0.88s kernel  0.17s user  13845+368 cs
ratio=1 	7.81s clock  0.89s kernel  0.17s user  13853+393 cs
ratio=1 	7.57s clock  0.90s kernel  0.17s user  13838+433 cs

ratio=50 	6.97s clock  0.84s kernel  0.19s user  13816+235 cs
ratio=50 	6.90s clock  0.88s kernel  0.19s user  13816+225 cs
ratio=50 	7.03s clock  0.84s kernel  0.18s user  13813+222 cs


Regards,
Fengguang Wu

PS.
About the partial sendfile problem mentioned in the previous benchmark mail:
The stock readahead is not handling it correctly, but not so destructive.
So I choose not to submit the patch.

[patch 1/2] readahead sysctl parameters fix

[Fengguang Wu]

[-- Attachment #1: readahead-sysctl-parameters-fix.patch --]
[-- Type: text/plain, Size: 1350 bytes --]

- do no extra readahead when (readahead_ratio == 1)
- define readahead_hit_rate inside CONFIG_ADAPTIVE_READAHEAD ifdefs

Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
---

--- linux-2.6.19-rc6-mm1.orig/include/linux/mm.h
+++ linux-2.6.19-rc6-mm1/include/linux/mm.h
@@ -1074,7 +1074,7 @@ extern int readahead_ratio;
 
 static inline int prefer_adaptive_readahead(void)
 {
-	return readahead_ratio > 1;
+	return readahead_ratio != 1;
 }
 
 /* Do stack extension */
--- linux-2.6.19-rc6-mm1.orig/Documentation/sysctl/vm.txt
+++ linux-2.6.19-rc6-mm1/Documentation/sysctl/vm.txt
@@ -234,7 +234,7 @@ plenty of memory(>>2MB per reader), a bi
 readahead_ratio also selects the readahead logic:
 	VALUE	CODE PATH
 	-------------------------------------------
-	    0	disable readahead totally
+	    0	read as is, no extra readahead
 	    1	select the stock readahead logic
 	2-100	select the adaptive readahead logic
 
--- linux-2.6.19-rc6-mm1.orig/mm/readahead.c
+++ linux-2.6.19-rc6-mm1/mm/readahead.c
@@ -40,10 +40,10 @@
 /* Set read-ahead size to ##% of the thrashing-threshold. */
 int readahead_ratio = 50;
 EXPORT_SYMBOL_GPL(readahead_ratio);
-#endif
 
 /* Readahead as long as cache hit ratio keeps above 1/##. */
 int readahead_hit_rate = 0;
+#endif /* CONFIG_ADAPTIVE_READAHEAD */
 
 /*
  * Detailed classification of read-ahead behaviors.

--

[patch 2/2] readahead nfsd case: fix uninitialized ra_min/ra_max

[Fengguang Wu]

[-- Attachment #1: readahead-nfsd-case-fix-uninitialized-ra_min-ra_max.patch --]
[-- Type: text/plain, Size: 2256 bytes --]

ra_min/ra_max are not initialized, fix it.

New benchmark numbers on

	hda: ST3250820A, ATA DISK drive
	hda: max request size: 512KiB
	hda: 488397168 sectors (250059 MB) w/8192KiB Cache, CHS=30401/255/63, UDMA(100)

with nfs mount
	mount -o tcp,rsize=$((rsize<<10)) localhost:$EXPORT $MNT

   rsize      stock  adaptive
   ==================================
f
      8k       3.07    2.51    -18.2%
     32k       2.40    2.17     -9.6%
    128k       2.40    2.35     -2.1%
ff
      8k      12.44    6.40    -48.6%
     32k      14.62    5.46    -62.7%
    128k      15.79    5.19    -67.1%
d
      8k       2.84    2.48    -12.7%
     32k       2.53    1.99    -21.3%
    128k       2.18    2.00     -8.3%
dd
      8k       8.16    7.99     -2.1%
     32k       8.27    6.88    -16.8%
    128k       7.75    6.97    -10.1%

The 4 patterns tested are:
	 f: single file read
	ff: double file parallel read
	 d: single dir read
	dd: double dir parallel read

Signed-off-by: Fengguang Wu <wfg@mail.ustc.edu.cn>
---

--- linux-2.6.19-rc6-mm1.orig/mm/readahead.c
+++ linux-2.6.19-rc6-mm1/mm/readahead.c
@@ -1616,6 +1616,10 @@ page_cache_readahead_adaptive(struct add
 	unsigned long ra_max;
 	int ret;
 
+	/* no read-ahead */
+	if (!ra->ra_pages)
+		return 0;
+
 	if (page) {
 		ClearPageReadahead(page);
 
@@ -1635,9 +1639,6 @@ page_cache_readahead_adaptive(struct add
 						offset + LAPTOP_POLL_INTERVAL))
 				return 0;
 		}
-	} else if (ra->flags & RA_FLAG_NFSD) { /* nfsd read */
-		ra_size = max_sane_readahead(req_size);
-		goto readit;
 	}
 
 	if (page)
@@ -1647,11 +1648,13 @@ page_cache_readahead_adaptive(struct add
 
 	get_readahead_bounds(ra, &ra_min, &ra_max);
 
-	/* read-ahead disabled? */
-	if (unlikely(!ra_max || !readahead_ratio)) {
-		ra_size = max_sane_readahead(req_size);
+	/* read as is */
+	if (!readahead_ratio)
+		goto readit;
+
+	/* nfsd read */
+	if (!page && (ra->flags & RA_FLAG_NFSD))
 		goto readit;
-	}
 
 	/*
 	 * Start of file.
@@ -1698,11 +1701,11 @@ page_cache_readahead_adaptive(struct add
 		return 0;
 	}
 
+readit:
 	/*
 	 * Random read.
 	 */
 	ra_size = min(req_size, ra_max);
-readit:
 	ra_size = __do_page_cache_readahead(mapping, filp, offset, ra_size, 0);
 
 	ra_account(ra, RA_EVENT_RANDOM_READ, ra_size);

--