Skip to contentSkip to conversation

Coda and Ext3

DISCUSSION Jan Harkes 2 participants 5 messages Sep 19, 2001 – Sep 20, 2001 1 message not in the archive Coda file system

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Jan Harkes

author maintainer
#1 lore raw
Quoted 10 lines · Sujal Shah

The Linux Coda drivers and the ext3 patches don't seem to get along
very well, at least in Linux 2.4.7. I've got a stock 2.4.7 kernel with a patch applied to the USB drivers (for a sony digital camera; see
http://www.sujal.net/tech/linux/ just a change in unusual_devs.h).

After I applied the ext3 patches from
http://www.uow.edu.au/~andrewm/linux/ext3/ . Basically, when an application tries to write to a file system mounted via coda, the application terminates with "Memory Fault" returned to the terminal. THe file system still thinks it's busy (can't umount).

Yeah, I know, and it will probably work when you don't enable data-journalling. Coda's kernelmodule currently uses generic_file_read and generic_file_write on it's containerfiles, which works for many filesystems. However, ext3fs (and tmpfs and several others) have a filesystem specific write implementation and don't really like being called with the generic functions.

The patch is simple, but I haven't made it yet. Basically we need to wrap the read/write calls and call cii->c_cfile->f_op->file_write or something.

Quoted 2 lines

loaded, however. Also, I backed out the patches for ext3 and the problem went away.

ext2 uses the generic file read/write functions, so whenever ext2 is the underlying filesystem it all works fine.

Jan

author maintainer
#2 lore raw
In reply to a message not in the archive
Quoted 2 lines · Sujal Shah

The Linux Coda drivers and the ext3 patches don't seem to get along
very well, at least in Linux 2.4.7. I've got a stock 2.4.7 kernel with

The attached patch should fix it. I haven't tested it against ext3, but with tmpfs which used to have the same problem, i.e. not using generic_file_ functions to access the container files. I'll pass it on to Linus and Alan once I get some feedback on whether this solves all problems. I believe this will fix the whole batch of problems that are have been reported with ext3fs, XFS, and tmpfs.

It should apply fine for kernel versions 2.4.4 and higher (both the AC and the Linus trees).

btw. my comment about disabling ext3 data journalling fixing it was bogus, there was a BUG() in Coda that got triggered because f_op->write wasn't generic_file_write.

Jan

fs/coda/file.c +47 -10
diff -urN --exclude-from=dontdiff linux-2.4.10-pre9/fs/coda/file.c linux/fs/coda/file.c
@@ -31,28 +31,65 @@
 int use_coda_close;
 
 static ssize_t
-coda_file_write(struct file *file,const char *buf,size_t count,loff_t *ppos)
+coda_file_read(struct file *file, char *buf, size_t count, loff_t *ppos)
 {
+	struct inode *inode = file->f_dentry->d_inode;
+	struct coda_inode_info *cii = ITOC(inode);
 	struct file *cfile;
+
+	cfile = cii->c_container;
+	if (!cfile) BUG();
+
+	if (!cfile->f_op || !cfile->f_op->read)
+		return -EINVAL;
+
+	return cfile->f_op->read(cfile, buf, count, ppos);
+}
+
+static ssize_t
+coda_file_write(struct file *file,const char *buf,size_t count,loff_t *ppos)
+{
 	struct inode *cinode, *inode = file->f_dentry->d_inode;
 	struct coda_inode_info *cii = ITOC(inode);
-	ssize_t n;
+	struct file *cfile;
+	ssize_t ret;
+	int flags;
 
 	cfile = cii->c_container;
 	if (!cfile) BUG();
 
-	if (!cfile->f_op || cfile->f_op->write != generic_file_write)
-		BUG();
+	if (!cfile->f_op || !cfile->f_op->write)
+		return -EINVAL;
 
 	cinode = cfile->f_dentry->d_inode;
-	down(&cinode->i_sem);
+	down(&inode->i_sem);
+	flags = cfile->f_flags;
+        cfile->f_flags |= file->f_flags & (O_APPEND | O_SYNC);
+
+	ret = cfile->f_op->write(cfile, buf, count, ppos);
 
-	n = generic_file_write(file, buf, count, ppos);
+	cfile->f_flags = flags;
 	inode->i_size = cinode->i_size;
+	up(&inode->i_sem);
+
+	return ret;
+}
+
+static int
+coda_file_mmap(struct file *file, struct vm_area_struct *vma)
+{
+	struct inode *inode = file->f_dentry->d_inode;
+	struct coda_inode_info *cii = ITOC(inode);
+	struct file *cfile;
+
+	cfile = cii->c_container;
+
+	if (!cfile) BUG();
 
-	up(&cinode->i_sem);
+	if (!cfile->f_op || !cfile->f_op->mmap)
+		return -ENODEV;
 
-	return n;
+	return cfile->f_op->mmap(cfile, vma);
 }
 
 int coda_open(struct inode *i, struct file *f)
@@ -237,9 +274,9 @@
 
 struct file_operations coda_file_operations = {
 	llseek:		generic_file_llseek,
-	read:		generic_file_read,
+	read:		coda_file_read,
 	write:		coda_file_write,
-	mmap:		generic_file_mmap,
+	mmap:		coda_file_mmap,
 	open:		coda_open,
 	flush:  	coda_flush,
 	release:	coda_release,
fs/coda/psdev.c +1 -1
diff -urN --exclude-from=dontdiff linux-2.4.10-pre9/fs/coda/psdev.c linux/fs/coda/psdev.c
@@ -414,7 +414,7 @@
 static int __init init_coda(void)
 {
 	int status;
-	printk(KERN_INFO "Coda Kernel/Venus communications, v5.3.14, [email protected]\n");
+	printk(KERN_INFO "Coda Kernel/Venus communications, v5.3.15, [email protected]\n");
 
 	status = init_coda_psdev();
 	if ( status ) {
#3 lore raw
Quoted 10 lines · Jan Harkes
Quoted 2 lines · Sujal Shah

The Linux Coda drivers and the ext3 patches don't seem to get along
very well, at least in Linux 2.4.7. I've got a stock 2.4.7 kernel with

The attached patch should fix it. I haven't tested it against ext3, but with tmpfs which used to have the same problem, i.e. not using generic_file_ functions to access the container files. I'll pass it on to Linus and Alan once I get some feedback on whether this solves all problems. I believe this will fix the whole batch of problems that are have been reported with ext3fs, XFS, and tmpfs.

I just tried your patch together with coda-debug-client-5.3.15-1, kernel-2.4.9 and ext3-2.4-0.9.6-249 and still get the BUG message. The ksymoops output is attached.

Ciao
Florian

author maintainer
#4 lore raw
In reply to a message not in the archive
Quoted 12 lines · Florian Schaefer
Quoted 10 lines · Jan Harkes
Quoted 2 lines · Sujal Shah

The Linux Coda drivers and the ext3 patches don't seem to get along
very well, at least in Linux 2.4.7. I've got a stock 2.4.7 kernel with

The attached patch should fix it. I haven't tested it against ext3, but with tmpfs which used to have the same problem, i.e. not using generic_file_ functions to access the container files. I'll pass it on to Linus and Alan once I get some feedback on whether this solves all problems. I believe this will fix the whole batch of problems that are have been reported with ext3fs, XFS, and tmpfs.

Quoted 3 lines

I just tried your patch together with coda-debug-client-5.3.15-1, kernel-2.4.9 and ext3-2.4-0.9.6-249 and still get the BUG message. The ksymoops output is attached.

No, different bug,

Fixed problem is basically what everyone is seeing when they are using Coda on top of a filesystem that doesn't use generic_file_write. It was simply there to avoid corrupting the underlying fs.

Your bug is unusual, because it happens during the mount, everybody else successfully mounted Coda and could read files, it would just die when trying to write anything.

The bug is triggered because the inode we just got passed from the VFS has a superblock that has no pointer to the Coda specific info.

If you look at the trace, parts of it look right, but most of it doesn't make sense, printk never calls back into FS specific code, coda_iget does not call coda_cnode make, coda_inocmp isn't calling iget4, etc.

If you look at the logic of the mount operation,

coda_read_super

Here we allocate the Coda info and link it to the superblock. If the allocation fails we return early. An upcall is made to venus to get the rootfid. Then we call coda_cnode_make, passing it the rootfid and the initialized superblock.

coda_cnode_make

Makes a second upcall, getting the attributes of the root inode, if it fails we return. Else it calls coda_iget, passing the superblock, the rootfid and the attributes.

coda_iget
Calls iget4 passing it the initialized superblock, a hash of the fid, and a callback function + rootfid to find the real inode.

iget4
Should fail to find the inode, and calls get_new_inode again passing it the initialized superblock.

get_new_inode
Allocates an inode, initializes inode->i_sb with the passed in superblock and calls the fs-specific read_inode.

coda_read_inode

    Craps out because the inode->i_sb.u.generic_sbp is NULL, wtf?

The only way I can see this happening is if the sb.u.generic_sbp field is cleared, but it looks like even coda_put_super isn't doing that.

Summary, I'm clueless how this could have happened, and I can't see how it could happen. Perhaps a patch got botched, or your kernel is mixing up objects from before the patch with later ones. All those ksyms errors and the illogical jumps in the trace don't make it look to healthy either.

Try to start from a fresh 2.4.9 tree, apply the ext3 patch and my Coda patch, add -fs1 to EXTRAVERSION just to force this tree to install it's modules in a separate directory. Copy the config of your existing tree, make oldconfig ; make dep ; etc...

Jan

#5 lore raw
Quoted 3 lines · Jan Harkes

If you look at the trace, parts of it look right, but most of it doesn't make sense, printk never calls back into FS specific code, coda_iget does not call coda_cnode make, coda_inocmp isn't calling iget4, etc.

Thank you for looking at my problem. :-)

Quoted 1 line

Try to start from a fresh 2.4.9 tree, apply the ext3 patch and my Coda

This is exactly what I did just some hours ago (oh, I had to add an aditional patch for the latest acpi daemon, but it just patches one single acpi file).

Quoted 1 line

patch, add -fs1 to EXTRAVERSION just to force this tree to install it's

Sorry, but I've never heard of this extraversion stuff. I found it in the main Makefile of the kernel, do I have to add the option there?

Quoted 1 line

modules in a separate directory.

Till now, I always tried to compile it directly into the kernel to be able to boot from an ext3 partition. Does this make a difference? This way I never had to touch my working modules.

Quoted 2 lines

Copy the config of your existing tree,
make oldconfig ; make dep ; etc...

make oldconfig? Doesn't make dep use the .config file directly?

As you can see, I have to admit that I'm quite confused. But I'll try to sum up what I've understood: You want me to take a new copy of the kernel, patch ext3 and your last patch into it, add some fancy options to prevent me from ruining my current system and try it again. But this is quite exactly what I did (aside from not compiling it as a module).
One more thing: I can reproduce the BUG on my second computer with the a similar setup.

Ciao
Florian

PS: I know that I haven't made your life easier with this mail, but that's the way it is. Sorry.