aio on ramfs

aio on ramfs

[William Lee Irwin III]

I was trying to run an aio benchmark on pgcl with ramfs in order to
exercise aio codepaths. It didn't have the aio methods, and so was
failing to do so. The following, suggested by Suparna Bhattacharya,
resolved the issue.

The reasoning is as follows:
(1) There's nowhere to write the things to, so nop out write functions.
(2) Since there's nowhere to write the things to, setting it dirty and
	going through writeback logic is also useless. Hence nop out
	->set_page_dirty().

An equivalent patch was successfully tested on pgcl running an aio
benchmark in combination with recently-posted aio fixes for pgcl.

New nop methods are necessary, for without the nop methods, an oops
like the following will occur when the mpage.c writeback code attempts
to invoke NULL method pointers:

Script started on Mon Dec  1 20:15:46 2003
Unable to handle kernel NULL pointer dereference at virtual address 00000000
 printing eip:
00000000
*pdpte = 30050001
*pde = 0
Oops: 0000 [#1]
CPU:    7
EIP:    0060:[<00000000>]    Not tainted
EFLAGS: 00010246
EIP is at 0x0
eax: 00000000   ebx: 00000000   ecx: 00000001   edx: f4ae8188
esi: ec03abe0   edi: f10cba4c   ebp: f10cbc98   esp: f10cb9ec
ds: 007b   es: 007b   ss: 0068
Process OraSim (pid: 2996, threadinfo=f10ca000 task=eec5f100)
Stack: c017bc54 ec03abe0 00000000 00000000 f10cba4c 00000001 f2772ac4 f2773ddc 
       00000000 00000000 00000000 00000000 00000000 00000001 00000000 00000000 
       00000000 00007cdf 00000000 00000001 000a8000 0000000f ec03abe0 ec03ac74 
Call Trace:
 [<c017bc54>] mpage_writepage+0x5a4/0x820
 [<c015d18f>] __getblk+0x2f/0x60
 [<c015d246>] __bread+0x26/0x40
 [<c018fe62>] read_block_bitmap+0x52/0x90
 [<c0190844>] ext2_new_block+0x324/0x570
 [<c0193000>] ext2_alloc_block+0x80/0xc0
 [<c0193395>] ext2_alloc_branch+0x35/0x210
 [<c01935e7>] ext2_get_block+0x77/0x390
 [<c015c08d>] __find_get_block_slow+0x4d/0x160
 [<c014b520>] handle_mm_fault+0x110/0x210
 [<c015c785>] __set_page_dirty_buffers+0xf5/0x120
 [<c015e993>] nobh_prepare_write+0x1b3/0x410
 [<c017c14a>] mpage_writepages+0x27a/0x2ac
 [<c013f607>] do_writepages+0x37/0x40
 [<c017a3a4>] __sync_single_inode+0xc4/0x210
 [<c017aaf6>] write_inode_now+0x46/0x80
 [<c017abe1>] generic_osync_inode+0xb1/0x110
 [<c013bacd>] generic_file_aio_write_nolock+0x5dd/0xbe0
 [<c0150068>] rmap_add_folio+0xd8/0x110
 [<c013c148>] generic_file_write_nolock+0x78/0x90
 [<c0118f36>] do_page_fault+0x176/0xa8f
 [<c011d050>] default_wake_function+0x0/0x20
 [<c011d050>] default_wake_function+0x0/0x20
 [<c013c246>] generic_file_write+0x56/0x70
 [<c015a69a>] vfs_write+0xaa/0x130
 [<c010fcf0>] do_gettimeofday+0x30/0xd0
 [<c015a8b4>] sys_pwrite64+0x54/0x80
 [<c01095e3>] syscall_call+0x7/0xb

Code:  Bad EIP value.
Script done on Mon Dec  1 20:32:13 2003


-- wli

diff -prauN linux-2.6.0-test11/fs/ramfs/inode.c ramfs-2.6.0-test11-1/fs/ramfs/inode.c
--- linux-2.6.0-test11/fs/ramfs/inode.c	2003-11-26 12:43:32.000000000 -0800
+++ ramfs-2.6.0-test11-1/fs/ramfs/inode.c	2003-12-07 00:26:29.000000000 -0800
@@ -31,6 +31,7 @@
 #include <linux/string.h>
 #include <linux/smp_lock.h>
 #include <linux/backing-dev.h>
+#include <linux/writeback.h>
 
 #include <asm/uaccess.h>
 
@@ -140,10 +141,28 @@ static int ramfs_symlink(struct inode * 
 	return error;
 }
 
+static int ramfs_writepage(struct page *page, struct writeback_control *wbc)
+{
+	return 0;
+}
+
+static int ramfs_writepages(struct address_space *mapping, struct writeback_control *wbc)
+{
+	return 0;
+}
+
+static int ramfs_set_page_dirty(struct page *page)
+{
+	return 0;
+}
+
 static struct address_space_operations ramfs_aops = {
 	.readpage	= simple_readpage,
 	.prepare_write	= simple_prepare_write,
-	.commit_write	= simple_commit_write
+	.commit_write	= simple_commit_write,
+	.set_page_dirty	= ramfs_set_page_dirty,
+	.writepage	= ramfs_writepage,
+	.writepages	= ramfs_writepages,
 };
 
 static struct file_operations ramfs_file_operations = {
@@ -153,6 +172,8 @@ static struct file_operations ramfs_file
 	.fsync		= simple_sync_file,
 	.sendfile	= generic_file_sendfile,
 	.llseek		= generic_file_llseek,
+	.aio_read	= generic_file_aio_read,
+	.aio_write	= generic_file_aio_write,
 };
 
 static struct inode_operations ramfs_file_inode_operations = {

Re: aio on ramfs

[OGAWA Hirofumi]

William Lee Irwin III <wli@holomorphy.com> writes:

> +static int ramfs_writepage(struct page *page, struct writeback_control *wbc)
> +{
> +	return 0;
> +}

Doesn't this break the magic of shrink_list()? I think it need the
"return WRITEPAGE_ACTIVATE;" at least.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: aio on ramfs

[William Lee Irwin III]

William Lee Irwin III <wli@holomorphy.com> writes:
>> +static int ramfs_writepage(struct page *page, struct writeback_control *wbc)
>> +{
>> +	return 0;
>> +}

On Mon, Dec 08, 2003 at 02:40:03AM +0900, OGAWA Hirofumi wrote:
> Doesn't this break the magic of shrink_list()? I think it need the
> "return WRITEPAGE_ACTIVATE;" at least.

In truth these things shouldn't be on the LRU at all, though they're
probably blindly plopped down there. My handwavy argument was that it
makes no sense to do anything with it on the LRU and that I'd nopped
out ->set_page_dirty() anyhow (i.e. PG_dirty should never get set). Does
that hold enough water or should I still hand back WRITEPAGE_ACTIVATE?


-- wli

Re: aio on ramfs

[OGAWA Hirofumi]

William Lee Irwin III <wli@holomorphy.com> writes:

> William Lee Irwin III <wli@holomorphy.com> writes:
> >> +static int ramfs_writepage(struct page *page, struct writeback_control *wbc)
> >> +{
> >> +	return 0;
> >> +}
> 
> On Mon, Dec 08, 2003 at 02:40:03AM +0900, OGAWA Hirofumi wrote:
> > Doesn't this break the magic of shrink_list()? I think it need the
> > "return WRITEPAGE_ACTIVATE;" at least.
> 
> In truth these things shouldn't be on the LRU at all, though they're
> probably blindly plopped down there. My handwavy argument was that it
> makes no sense to do anything with it on the LRU and that I'd nopped
> out ->set_page_dirty() anyhow (i.e. PG_dirty should never get set). Does
> that hold enough water or should I still hand back WRITEPAGE_ACTIVATE?

I see. Sorry for noise.
-- 
OGAWA Hirofumi <hirofumi@mail.parknet.co.jp>

Re: aio on ramfs

[Rob Landley]

On Sunday 07 December 2003 02:34, William Lee Irwin III wrote:

> -- wli
>
> diff -prauN linux-2.6.0-test11/fs/ramfs/inode.c
> ramfs-2.6.0-test11-1/fs/ramfs/inode.c ---
> linux-2.6.0-test11/fs/ramfs/inode.c	2003-11-26 12:43:32.000000000 -0800 +++
> ramfs-2.6.0-test11-1/fs/ramfs/inode.c	2003-12-07 00:26:29.000000000 -0800
> @@ -31,6 +31,7 @@
>  #include <linux/string.h>
>  #include <linux/smp_lock.h>
>  #include <linux/backing-dev.h>
> +#include <linux/writeback.h>
>
>  #include <asm/uaccess.h>
>
> @@ -140,10 +141,28 @@ static int ramfs_symlink(struct inode *
>  	return error;
>  }

Can I suggest a comment right here, "writepage/writepages/set_page_dirty stubs 
needed for aio"?

> +static int ramfs_writepage(struct page *page, struct writeback_control
> *wbc) +{
> +	return 0;
> +}
> +
> +static int ramfs_writepages(struct address_space *mapping, struct
> writeback_control *wbc) +{
> +	return 0;
> +}
> +
> +static int ramfs_set_page_dirty(struct page *page)
> +{
> +	return 0;
> +}

Rob