[PATCH 5/6] [RFC] Checkpoint/restart unlinked files

[Oren Laadan]

On 09/29/2010 09:17 PM, Matt Helsley wrote:
> [ Persistent readers may wish to grab some coffee because this is gonna be
>    a long reply. ]

[ Persistent readers have been already warned ... ]

>
> On Wed, Sep 29, 2010 at 06:22:37PM -0400, Oren Laadan wrote:
>>
>> On 09/23/2010 05:53 PM, Matt Helsley wrote:
>>> Implement checkpoint of unlinked files by relinking them into their
>>> filesystem at lost+found/checkpoint/unlink-at-restart-<ktime>-<N>.
>>> We can change the function to generate better paths -- I just need to
>>> know what that path should be. For example, perhaps sys_checkpoint should
>>> take a template string somewhat like mkstemp().
>>
>> /lost+found/checkpoint/.... relative to what ?
>
> The root of the mounted filesystem. Not the root of the mount namespace nor
> The Root Filesystem. Using that "root" ensures that it stays in the
> filesystem and we won't get EXDEV.
>
>> security concerns ?
>
> It occurs to me that the checkpointing user may not have permission to
> add entries there so checkpoint may fail because of that. And if they do
> have permission to create entries there then it could be a means to
> hardlink files and pin quota (just like any time you can hardlink
> visible files that other users own: http://pastebin.com/RybskyyY ). :(
>
>>
>>> Relinking allows userspace to leverage the snapshotting capabilities
>>> of various linux block devices and filesystems. sys_checkpoint relinks
>>> the files and returns. Userspace then checkpoints the filesystem contents
>>> using any backup-like method prior to thawing. That backup would then be
>>> available for use during an optional migration followed by restore and
>>> restart. In the case of network and cluster/distributed filesystems copying
>>> the filesystem contents explicitly for migration may not be necessary at
>>> all -- it would be part of normal file writes. For non-migration uses
>>> of checkpoint/restart filesystems like btrfs a snapshot could simply be
>>> taken during checkpoint and mounted during restart -- again without
>>> requiring IO proportional to the aggregate size of filesystem contents
>>> being checkpointed.
>>>
>>> In addition to the original path of the file we save the newly-linked
>>> path. This newly-linked path is opened during restart instead of the
>>> original path (which is only useful files that were linked at the time
>>> of checkpoint). The newly-linked location is also useful in order to
>>> identify which relinked files in lost+found/checkpoint were created by
>>> this particular invokation of sys_checkpoint. This enables userspace
>>
>> s/invokation/invocation/
>
> Argh, yes. :)
>
>>
>>> to cleanup after checkpoints which failed yet successfully relinked.
>>>
>>> Note that we'd still be restricted by the limitations of hardlinks.
>>> Furthermore, as Aneesh Kumar mentioned in the LKML threads leading up to
>>> the v19 file handle patches, this kind of linking seems to require
>>> CAP_DAC_READ_SEARCH because the files to be linked lack a path to
>>> search in the first place. Aneesh added the check in response to Al Viro's
>>> point that being able to relink open files (passed via SCM_RIGHTS for
>>> example) has non-trivial security ramifications.
>>>
>>> To understand why relinking is extremely useful for checkpoint/restart
>>> consider this simple pseudocode program and a specific example checkpoint
>>> of it:
>>>
>>> 	a_fd = open("a"); /* example: size of the file at "a" is 1GB */
>>> 	link("a", "b");
>>> 	unlink("a");
>>> 	creat("a");
>>> 	<---- example: checkpoint happens here
>>> 	write(a_fd, "bar");
>>>
>>> The file "a" is unlinked and a different file has been placed at that
>>> path. a_fd still refers to the inode shared with "b".
>>>
>>> Without relinking we would need to walk the entire filesystem to find out
>>> that "b" is a path to the same inode (another variation on this case: "b"
>>> would also have been unlinked). We'd need to do this for every
>>> unlinked file that remains open in every task to checkpoint. Even then
>>> there is no guarantee such a "b" exists for every unlinked file -- the
>>> inodes could be "orphans" -- and we'd need to preserve their contents
>>> some other way.
>>>
>>> I considered a couple alternatives to preserving unlinked file contents:
>>> copying and file handles. Each has significant drawbacks.
>>>
>>> First I attempted to copy the file contents into the image and then
>>> recreate and unlink the file during restart. Using a simple version of
>>> that method the write above would not reach "b". One fix would be to search
>>> the filesystem for a file with the same inode number (inode of "b") and
>>> either open it or hardlink it to "a". Another would be to record the inode
>>> number. This either shifts the search from checkpoint time to restart time
>>> or has all the drawbacks of the second method I considered: file handles.
>>
>> Maybe worth to explicitly mention the obvious, just in case:
>>
>> Only re-create and unlink the file if at checkpoint time the
>> link count was 0 (no other references existed); and do it under
>> a temporary name, to not overwrite existing (newer) files.
>
> Yup.
>
>> If the link count was greater than 0, then the only correct
>> solution is to somehow find the other pathname for the same
>
> Nope. We don't need to find the other pathname so long as we have the
> new link.

Well, this was in the context of your "I considered a couple
alternatives..."...

>
>> inode, make a new temporary link, open the new link, and finally
>> unlink the new link.
>
> Yup. We make a new temporary link during checkpoint. That link becomes
> part of the snapshot/backup saved for restart. During restart we re-open
> the link and then unlink it. It should even work with multiple
> struct file *'s referring to the same file.
>
>>>
>>> Instead of copying contents or recording inodes I also considered using
>>> file handles. We'd need to ensure that the filehandles persist in storage,
>>> can be snapshotted/backed up, and can be migrated. Can handlefs or any
>>> generic file handle system do this? My _guess_ is "no" but folks are
>>> welcome to tell me I'm wrong.
>>>
>>> In contrast, linking the file from a_fd back into its filesystem can avoid
>>> these complexities. Relinking avoids the search for matching inodes and
>>> copying large quantities of data from storage only to write it back (in
>>> fact the data would be read-and-written twice -- once for checkpoint and
>>> once for restart). Like file handles it does require changes to the
>>> filesystem code. Unlike file handles, enabling relinking does not require
>>> every filesystem to support a new kind of filesystem "object" -- only
>>> an operation that is quite similar to one that already exists: link.
>>>
>>> Signed-off-by: Matt Helsley<matthltc at us.ibm.com>
>>> Cc: Eric Sandeen<sandeen at redhat.com>
>>> Cc: Theodore Ts'o<tytso at mit.edu>
>>> Cc: Andreas Dilger<adilger.kernel at dilger.ca>
>>> Cc: linux-ext4 at vger.kernel.org
>>> Cc: Jan Kara<jack at suse.cz>
>>> Cc: containers at lists.linux-foundation.org
>>> Cc: Oren Laadan<orenl at cs.columbia.edu>
>>> Cc: linux-fsdevel at vger.kernel.org
>>> Cc: Al Viro<viro at zeniv.linux.org.uk>
>>> Cc: Christoph Hellwig<hch at infradead.org>
>>> Cc: Jamie Lokier<jamie at shareable.org>
>>> Cc: Amir Goldstein<amir73il at users.sf.net>
>>> Cc: Aneesh Kumar<aneesh.kumar at linux.vnet.ibm.com>
>>> Cc: Miklos Szeredi<miklos at szeredi.hu>
>>> ---
>>>   fs/checkpoint.c                  |   51 ++++++++++++++-----
>>>   fs/namei.c                       |  102 ++++++++++++++++++++++++++++++++++++++
>>>   fs/pipe.c                        |    2 +-
>>>   include/linux/checkpoint.h       |    3 +-
>>>   include/linux/checkpoint_hdr.h   |    3 +
>>>   include/linux/checkpoint_types.h |    3 +
>>>   6 files changed, 149 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/fs/checkpoint.c b/fs/checkpoint.c
>>> index 87d7c6e..9c7caec 100644
>>> --- a/fs/checkpoint.c
>>> +++ b/fs/checkpoint.c
>>> @@ -16,6 +16,7 @@
>>>   #include<linux/sched.h>
>>>   #include<linux/file.h>
>>>   #include<linux/namei.h>
>>> +#include<linux/mount.h>
>>>   #include<linux/fs_struct.h>
>>>   #include<linux/fs.h>
>>>   #include<linux/fdtable.h>
>>> @@ -26,6 +27,7 @@
>>>   #include<linux/checkpoint.h>
>>>   #include<linux/eventpoll.h>
>>>   #include<linux/eventfd.h>
>>> +#include<linux/sys-wrapper.h>
>>>   #include<net/sock.h>
>>>
>>>   /**************************************************************************
>>> @@ -174,6 +176,9 @@ int checkpoint_file_common(struct ckpt_ctx *ctx, struct file *file,
>>>   	h->f_pos = file->f_pos;
>>>   	h->f_version = file->f_version;
>>>
>>> +	if (d_unlinked(file->f_dentry))
>>> +		/* Perform post-checkpoint and post-restart unlink() */
>>> +		h->f_restart_flags |= RESTART_FILE_F_UNLINK;
>>
>> Follow the convention for these ?  s/RESTART_.../CKPT_.../
>
> I'll prefix it with CKPT_RESTART_ since "RESTART_" is kind of essential for
> conveying the purpose of the flag -- it indicates that we should unlink
> the file at restart.
>
>>
>>>   	h->f_credref = checkpoint_obj(ctx, f_cred, CKPT_OBJ_CRED);
>>>   	if (h->f_credref<   0)
>>>   		return h->f_credref;
>>> @@ -197,16 +202,6 @@ int generic_file_checkpoint(struct ckpt_ctx *ctx, struct file *file)
>>>   	struct ckpt_hdr_file_generic *h;
>>>   	int ret;
>>>
>>> -	/*
>>> -	 * FIXME: when we'll add support for unlinked files/dirs, we'll
>>> -	 * need to distinguish between unlinked filed and unlinked dirs.
>>> -	 */
>>> -	if (d_unlinked(file->f_dentry)) {
>>> -		ckpt_err(ctx, -EBADF, "%(T)%(P)Unlinked files unsupported\n",
>>> -			 file);
>>> -		return -EBADF;
>>> -	}
>>
>> We still need to handle unlinked directories separately (see
>> below). And they are actually easier to handle. So you probably
>> also need a flag to indicate that it is a directory.
>>
>> Perhaps rename flag from above, e.g.:
>>    CKPT_FILE_UNLINKED_FILE
>>    CKPT_FILE_UNLINKED_DIR
>
> I did't think we needed to. When I looked at the code I figured it should
> handle them just fine. That said, I don't have a testcase for relinking
> them.
>
> They're relatively simple because unlinked directories must be empty. We're
> essentially saving the filesystem permission, owner/group, acls, and xattrs
> of the directory for use during restart by relinking it.

So you suggest that an open unlinked directory (which is, of course,
empty) - will be restored as an open unlinked file ?  If so, did you
test that it works correctly ?  I'm not sure you'll get the desired
behavior from calling fstat() on the restored fd, or from calling
readdir() on it.

In fact, we don't even need to relink an unlinked directory - it's
empty anyway and easy to recreate (and re-unlink) at restart time.

So IMHO we need to know if the file pointer referred to a file or
to a directory.

>
>>> -
>>>   	h = ckpt_hdr_get_type(ctx, sizeof(*h), CKPT_HDR_FILE);
>>>   	if (!h)
>>>   		return -ENOMEM;
>>> @@ -220,6 +215,9 @@ int generic_file_checkpoint(struct ckpt_ctx *ctx, struct file *file)
>>>   	if (ret<   0)
>>>   		goto out;
>>>   	ret = checkpoint_fname(ctx,&file->f_path,&ctx->root_fs_path);
>>> +	if (ret<   0)
>>> +		goto out;
>>> +	ret = checkpoint_file_links(ctx, file);
>>>    out:
>>>   	ckpt_hdr_put(ctx, h);
>>>   	return ret;
>>> @@ -570,9 +568,11 @@ static int ckpt_read_fname(struct ckpt_ctx *ctx, char **fname)
>>>   /**
>>>    * restore_open_fname - read a file name and open a file
>>>    * @ctx: checkpoint context
>>> + * @restore_unlinked: unlink the opened file
>>
>> nit: s/restore_unlinked/unlinked/   ?
>
> I'd rather rename it do_unlink since that's the meaning the paremeter
> is meant to convey. Though that name could easily shadow a future
> syscall helper which may be useful...

Sure. Whatever makes it shorted :)

>
>>
>>>    * @flags: file flags
>>>    */
>>> -struct file *restore_open_fname(struct ckpt_ctx *ctx, int flags)
>>> +struct file *restore_open_fname(struct ckpt_ctx *ctx,
>>> +				int restore_unlinked, int flags)
>>>   {
>>>   	struct file *file;
>>>   	char *fname;
>>> @@ -586,8 +586,33 @@ struct file *restore_open_fname(struct ckpt_ctx *ctx, int flags)
>>>   	if (len<   0)
>>>   		return ERR_PTR(len);
>>>   	ckpt_debug("fname '%s' flags %#x\n", fname, flags);
>>> -
>>> +	if (restore_unlinked) {
>>> +		kfree(fname);
>>
>> This is too early to discard the original name (why would we
>> have saved it in the first place anyway ?) because ...
>
> Hmm.
>
>>
>>> +		fname = NULL;
>>> +		len = ckpt_read_payload(ctx, (void **)&fname, PATH_MAX,
>>> +					CKPT_HDR_BUFFER);
>>> +		if (len<   0)
>>> +			return ERR_PTR(len);
>>> +		fname[len] = '\0';
>>> +	}
>>
>> Uhm... for the following to work you need to explicitly handle
>> the case of unlinked directory:
>>
>> 	if (restore_unlinked&  CKPT_FILE_UNLINKED_DIR)
>> 		mkdir(...);
>
> A simple mkdir is insufficient as I pointed out above. We need the
> associated owner, group, permissions, xattrs, and acls to faithfully
> re-create the directory. Again, I *think* the code as it is handles
> all that properly and without extra fuss. I'll review the code to be
> sure and write a new test to make extra sure.

Yes, the rest of the attributes are important, as is the case for
unlinked files. So it it works for unlinked files, it should work
for unlinked dirs too.

Anyway, the 'mkdir' can be to any pathname, because we didn't have
to re-link the directory in the first place.

>
>>
>>>   	file = filp_open(fname, flags, 0);
>>> +	if (IS_ERR(file)) {
>>> +		ckpt_err(ctx, PTR_ERR(file), "Could not open file \"%s\"\n", fname);
>>> +
>>> +		goto out;
>>> +	}
>>> +	if (!restore_unlinked)
>>> +		goto out;
>>> +	if (S_ISDIR(file->f_mapping->host->i_mode))
>>> +		len = kernel_sys_rmdir(fname);
>>> +	else
>>> +		len = kernel_sys_unlink(fname);
>>
>> ... because now you need to manually modify the dentry to have
>> the original name, so that doing 'ls -l /proc/PID/fd' will give
>>    "NN ->  original_name (deleted)"
>> rather than
>>    "NN ->  relinked_obfuscated_name (deleted)"
>
> A decent point. However the problem is worse than you describe since the
> path with have changed too.
>
> "manually modifying" the dentry name seems like a terrible hack as far
> as organizing the kernel code goes. I'd like to spend some time trying
> to think of alternative solutions unless VFS developers care to
> weigh in on this with either "It's OK" or "Here's a way to avoid the
> hack...".

Maybe make the /proc/PID/fd/NN code aware of restored unlinked
files/dirs ?  (that will also solve the related path-issue).

>>> +	if (len<   0) {
>>> +		ckpt_err(ctx, len, "Could not unlink \"%s\"\n", fname);
>>> +		fput(file);
>>> +		file = ERR_PTR(len);
>>> +	}
>>> +out:
>>>   	kfree(fname);
>>>
>>>   	return file;
>>> @@ -692,7 +717,7 @@ static struct file *generic_file_restore(struct ckpt_ctx *ctx,
>>>   	    ptr->h.len != sizeof(*ptr) || ptr->f_type != CKPT_FILE_GENERIC)
>>>   		return ERR_PTR(-EINVAL);
>>>
>>> -	file = restore_open_fname(ctx, ptr->f_flags);
>>> +	file = restore_open_fname(ctx, !!(ptr->f_restart_flags&   RESTART_FILE_F_UNLINK), ptr->f_flags);
>>
>> Now f_restart_flags can have more than one flag, so remove the "!!"
>> (and I would rename it to f_ckpt_flags)
>
> That doesn't make sense. Yes, it could have more than one flag. But the
> only flag restore_open_fname() is designed to take is the one indicating
> that it should unlink the file. We don't know that any flags added in
> the future will be relevant to restore_open_fname().

IMHO we need to know if it's an unlinked file or unlinked dir
(as I argued above).

>
> As to the field rename: I disagree. the flags are explicitly there to instruct
> "restart" to unlink the file. They aren't "checkpointed file flags" but
> "f_ckpt_flags" sounds like some file flags we're saving. The proposed name
> also incorrectly implies a redundancy in the data structure. Finally, these
> fields are already inside a ckpt_* struct _and_ only used in checkpoint/restart
> code. So there's little chance we're going to forget that and need "_ckpt_"
> in the field name.

Fair enough. How about 'f_unlinked' ?

>
>>>   	if (IS_ERR(file))
>>>   		return file;
>>>
>>> diff --git a/fs/namei.c b/fs/namei.c
>>> index 8c9663d..69c4f4e 100644
>>> --- a/fs/namei.c
>>> +++ b/fs/namei.c
>>> @@ -32,6 +32,9 @@
>>>   #include<linux/fcntl.h>
>>>   #include<linux/device_cgroup.h>
>>>   #include<linux/fs_struct.h>
>>> +#ifdef CONFIG_CHECKPOINT
>>> +#include<linux/checkpoint.h>
>>> +#endif
>>>   #include<asm/uaccess.h>
>>>
>>>   #include "internal.h"
>>> @@ -2543,6 +2546,105 @@ SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname
>>>   	return sys_linkat(AT_FDCWD, oldname, AT_FDCWD, newname, 0);
>>>   }
>>>
>>> +#ifdef CONFIG_CHECKPOINT
>>> +
>>> +/* Path relative to the mounted filesystem's root -- not a "global" root or even a namespace root. The unique_name_count is unique for the entire checkpoint. */
>>> +#define CKPT_RELINKAT_FMT "lost+found/checkpoint/unlink-at-restart-%08llx-%u"
>>
>> A few existential questions:
>>
>> * Do we want to hard-code to /lost+found/checkpoint/... or allow
>
> (nit: sans the first / )
>
>>   userspace to set it per checkpoint ?
>
> Good question. I was hoping to get a discussion on this from
> everyone. lost+found/checkpoint is a poor but simple solution
> which we'll likely need to enhance or change. It highlights the
> issues without requiring everyone to read alot of code.

Ok, it's not a bad start. Let's then leave a big TODO comment
that we need to address it.

Once we get to resolve it, one idea is to make it a property of
the monut point, so that when mounting (e.g. --bind, or loopback)
filesystems for containers, the user could specify a desired
location through a mount option.

>>
>> * If we put it all in some place, could it cause some security
>>   or privacy issues, if we hold data from multiple users on the
>>   same filesystem ?
>>
>> * This doesn't guarantee unique-ness among checkpoints. One option
>>   is to add the ctx->crid to make it unique - but I'm unsure that
>>   we'll keep it in the future.
>
> Yes, the issue is multiple simultaneous checkpoints involving the
> same filesystem but not the same processes. That's why I added the
> checkpoint time in microseconds to the name. That doesn't eliminate
> the possibility but it does make it very unlikely. In the very unlikely
> event that two checkpoints try to make a hardlink with the same
> name we'll get EEXIST and checkpoint will fail. That seems quite
> acceptable to me given that it saves us from needing much more complex
> naming schemes and kernel code to implement it. Of course there are other
> issues related to the location we link to but I don't believe name
> collisions are one of the real problems.
>
>>
>> * For clean-up purposes, maybe get a subdir for each checkpoint ?
>
> Except as you pointed out we may need a per-filesystem-per-user
> directory for security/privacy.

Sure. And in there - per checkpoint subdir. Moreover, you suggest
that we defer those concerns until later; and I suggest that we
should do the subdir per checkpoint already now,

>
>>
>> * If checkpoint fails, who is responsible for the cleanup ?  I'd
>>   prefer that sys_checkpoint() won't leave garbage behind (so using
>>   the subdir approach above, we could just delete that dir ?)
>
> Argh, you're right. I should probably add an "error cleanup" deferqueue
> for that.
>
>>
>> * If restart fails, e.g. because of lack of memory, then another
>>   restart would fail because some (or all) relinked files will
>>   have already been re-unlinked and gone.  Maybe defer all unlinks
>>   to the end of the restart ?
>
> Userspace had to restore the filesystem to the proper state before
> restart anyway. Part of that restoration must be designed to ensure that
> the originally preserved state will be preserved even well after a
> _successful_ restart. Otherwise we could encounter an therwise-recoverable
> hardware failure shortly after restart but prior to another checkpoint --
> one that doesn't affect the system used to store the checkpoint -- and we'd
> have no hope of using the previous checkpoint.

A hardware failure is not a good example - let me make the following
analogy to explain why:

When you use your laptop, you can suspend (to ram, to disk) and later
resume. No filesystem snapshots are required. If the battery dies
soon after resume, your state is lost forever. The scenario I described
is analogous to laptop suspend resume: a user runs a VNC session, then
suspends it to disk by taking a checkpoint (no need for filesystem
snapshot, because it is guaranteed not to be modified). Later the user
resumes the session by doing a restart. If thereafter the hardware
fails, that's "ok", tough luck. But if the restart goes only half way
and there are not enough resources, then the resume never succeeded
and restart should not have destructed the state.

> So the core point here is that restart ought to be non-destructive in
> the sense that it does not destroy the state stored during checkpoint.

Exactly: and it restart fails then it did destroy that state - in
particular, the unlinked files saved during checkpoint.

> This is certainly true of the checkpoint image itself and yes, we should
> try to make sure the same is true with respect to how we restore the
> filesystem.
>
> That said, I'm worried we'll find ourselves in a familiar pickle if
> we start tacking on lists of things to do "at the end" that cannot fail.
> Not only does thinking like that subtly change the definition of
> "at the end" in subtly-broken ways but the "cannot fail" aspect of the
> operations could easily be overlooked. And it might wind up making
> promises we won't be able to deliver some day due to other code changes.

We can still strive to minimize the chances of undesired outcome.
Specifically, make sure that the deferred _destructive_ operations
are really last.

I can't think of other _destructive_ operations that we may need.
(Which doesn't mean there won't be any). But for now we can make
sure these (unlink) are done at the end, and there is really no
reason for them to fail (because they don't allocate anything, do
they ?). So once we get to do them, we are quite safe at the end of
a restart, and extremely unlikely to fail.

(Technically, we can make deferq enqueue at the front by default,
and API to enqueue at the tail for such cases).

>
> Perhaps we're better off saying that if restart fails the changes to
> the filesystem are boundedly undefined. Bounded because we still obey
> security policy (and it'd be a serious bug if we didn't!) and any of
> the changes are "described" by a subset of what's in the checkpoint image.
> Undefined because unlinked files may or may not have been deleted (in
> this case). Userspace can take steps to save and then restore the filesystem
> based on the contents of the checkpoint image. Indeed, that's a big part
> of what userspace is supposed to do during checkpoint anyway. And this
> supports my first point well.

I prefer to provide maximum guarantees. See above.

However, since you mention userspace, here is another idea: if
indeed we put everything in a per-checkpoint subdirectory, then
we can let restart do away with unlinking, and let userspace do
that when restart completes.  To avoid races, userspace can
choose to have the restarted tasks frozen when restart completes
thus ensuring that they don't run too early.

The above can be an optional behavior, e.g. RESTART_DONT_UNLINK
flag to instruct restart to leave that task to userspace.

>
> So I'm fine with adding the deferqueue to cleanup the links on
> the checkpoint side but I don't think we should defer the unlink
> on the restart side.

:(

>
>>   (This is less of a problem when we had a filesystem snapshot,
>
> Not just "less of a problem when..". It wouldn't be a problem with
> snapshots. We'd just remount the snapshot and the relinked file will
> still be there.
>
>>   just go back to that snapshot. But when we, e.g. suspend a vnc
>>   session to disk using checkpoint, and then try to resume - it
>>   will be a problem).
>>
>> * By using the unique counter, you assume that the sequence of
>>   file-open's (or mkdir) will never change at restart. But if a
>>   future user tool will alter the checkpoint image (assume it is
>>   in a valid and desirable way) and drop one file, then we are
>>   doomed ...  Instead, you should save the unique unlink id
>>   explcitily in the ckpt_hdr_file.
>
> I don't use the counter at all during restart. I use the
> second path/fname stored specially in the checkpoint image for
> unlinked files. That means we can change where we link without
> changing the restart code :).

My bad - I somehow assumes you use checkpoint_fill_relink_fname()
during restart too.

>
>>
>>> +
>>> +static int checkpoint_fill_relink_fname(struct ckpt_ctx *ctx,
>>> +					struct file *for_file,
>>> +					char relink_dir_pathname[PATH_MAX],
>>> +					int *lenp)
>>> +{
>>> +	struct path relink_dir_path;
>>> +	char *tmp;
>>> +	int len;
>>> +
>>> +	/* Find path to mount */
>>> +	relink_dir_path.mnt = for_file->f_path.mnt;
>>> +	relink_dir_path.dentry = relink_dir_path.mnt->mnt_root;
>>> +	tmp = d_path(&relink_dir_path, relink_dir_pathname, PATH_MAX);
>>
>> This path is relative to whom ?
>
> Relative to the root of the filesystem that stores the file.

In that case all you need is a hard-coded "lost+found/checkpoint".
But here, this is as seen by the checkpointer task, no ?

>> Don't we need similar precautions as in ckpt_fill_name(), in
>> particular verify that that path is "reachable" from the
>> checkpointer ?    Can it be made to reuse the code from there ?
>
> That's not quite what I'm doing here. I'm making "from scratch" a
> name that, if d_path succeeds, will be reachable.
>
>>
>>> +	if (IS_ERR(tmp))
>>> +		return PTR_ERR(tmp);
>>> +
>>> +	/* Append path to relinked file. */
>>> +	len = strlen(tmp);
>>> +	if (len<= 0)
>>> +		return -ENOENT;
>>> +	memmove(relink_dir_pathname, tmp, len);
>>> +	tmp = relink_dir_pathname + len - 1;
>>> +	/* Ensure we've got a single dir separator */
>>> +	if (*tmp == '/')
>>> +		tmp++;
>>> +	else {
>>> +		tmp++;
>>> +		*tmp = '/';
>>> +		tmp++;
>>> +		len++;
>>> +	}
>>> +	len += snprintf(tmp, PATH_MAX - len, CKPT_RELINKAT_FMT,
>>> +			ctx->ktime_begin.tv64,
>>> +			 ++ctx->unique_name_count);
>>
>> Should test for truncation, and bail...
>
> Oops! You're quite correct!
>
>>
>>> +	relink_dir_pathname[len] = '\0';
>>> +	*lenp = len;
>>> +	return 0;
>>> +}
>>> +
>>> +static int checkpoint_file_relink(struct ckpt_ctx *ctx,
>>> +				  struct file *file,
>>> +				  char new_path[PATH_MAX])
>>> +{
>>> +	int ret, len;
>>> +
>>> +	/*
>>> +	 * Relinking arbitrary files without searching a path
>>> +	 * (which non-existent if the file is unlinked) requires
>>> +	 * special privileges.
>>> +	 */
>>> +	if (!capable(CAP_DAC_OVERRIDE|CAP_DAC_READ_SEARCH)) {
>>> +		ckpt_err(ctx, -EPERM, "%(T)Relinking unlinked files requires CAP_DAC_{OVERRIDE,READ_SEARCH}\n");
>>> +		return -EPERM;
>>> +	}
>>> +	ret = checkpoint_fill_relink_fname(ctx, file, new_path,&len);
>>> +	if (ret)
>>> +		return ret;
>>> +	ret = do_kern_linkat(&file->f_path, file->f_dentry,
>>> +			     AT_FDCWD, new_path, 0);
>>> +	if (ret)
>>> +		ckpt_err(ctx, ret, "%(T)%(P)%(V)Failed to relink unlinked file.\n", file, file->f_op);
>>> +	return ret;
>>> +}
>>> +
>>> +int checkpoint_file_links(struct ckpt_ctx *ctx, struct file *file)
>>> +{
>>> +	char *new_link_path;
>>> +	int ret, len;
>>> +
>>> +	if (!d_unlinked(file->f_dentry))
>>> +		return 0;
>>> +
>>> +	/*
>>> +	 * Unlinked files need at least one hardlink for the post-sys_checkpoint
>>> +	 * filesystem backup/snapshot.
>>> +	 */
>>> +	new_link_path = kmalloc(PATH_MAX, GFP_KERNEL);
>>> +	if (!new_link_path)
>>> +		return -ENOMEM;
>>> +	ret = checkpoint_file_relink(ctx, file, new_link_path);
>>> +	if (ret<   0)
>>> +		goto out_free;
>>> +	len = strlen(new_link_path);
>>> +	ret = ckpt_write_obj_type(ctx, NULL, len + 1, CKPT_HDR_BUFFER);
>>
>> Do we need to save the entire pathname ?  Only the pathname to
>> the relevant mount point suffices, no ?  (the timestamp and the
>> unique id are available at restart).
>
> No because, as you pointed out earlier, multiple checkpoints could
> be in progress.
>
>>
>> Also worth to think (and at least comment with a big TODO) that
>> when we work on c/r support for mount points, we could just
>> refer to a (shared object) mount point instead of carrying the
>> path.
>
> I haven't decided yet. :)

I know - so let's leave a big TODO comment there :)

>
> On the one hand, yes we could do that. It certainly would help us
> handle the "overmount" case (I think that's what the openvz folks call
> it) that we don't right now:
>
> 	mount /dev/foo /bar
> 	open /bar/baz
> 	mount /dev/qux /bar
> 	<checkpoint saves "/bar/baz" path>
> 	...
> 	<restart restores the mounts>
> 	filp_open("/bar/baz")
>
> During restart we'll open the wrong "/bar/baz" if we've done a simple
> restoration of the mount tree prior to re-opening. In fact we can't
> cleanly divide mount tree restoration from re-opening files unless
> we move away from using filp_open().
>
> That said, full paths would be useful if we don't care about exactly
> how the (file) namespace is constructed -- only that it's contents
> are correct. For example, suppose we migrate the checkpoint image
> from machine X to machine Y. X might have "/bar/baz" on the root
> filesystem but Y might have it on a 'different' filesystem mounted
> on "/bar". If we attempt to recreate the mount tree then restart
> won't work.

This scenario would also break the current unlinked files/dir scheme...

>
> So whether to use a full path or mount-relative path all depends on
> what method the system administrator wants to use to enable
> migration of checkpoint/restart while meeting performance requirements.
>
> And that's why I say I haven't decided. To some extent this is one
> area where we'd find ourselves inadvertently defining policy for
> userspace which could be "harmful" to their use of c/r and thus
> require future c/r interface changes. Ideally we'd find a way to
> give userspace the ability to choose policy and avoid those interface
> changes.
>
> So I was thinking of using setns() to enable checkpoint and restart
> of mount namespaces. Usespace will set up the mount namespaces prior
> to calling sys_restart. Then we pass in references to the mount
> namespaces in the form of file descriptors opened by user-cr's restart
> binary. This way it's all shell-scriptable too.
>
> Anyhow -- I can post those as an RFC patch and this all belongs in
> a separate thread for those :).

Yes, please !

>
>>
>>> +	if (ret<   0)
>>> +		goto out_free;
>>> +	ret = ckpt_kwrite(ctx, new_link_path, len + 1);
>>> +out_free:
>>> +	kfree(new_link_path);
>>> +
>>> +	return ret;
>>> +}
>>> +#endif /* CONFIG_CHECKPOINT */
>>> +
>>>   /*
>>>    * The worst of all namespace operations - renaming directory. "Perverted"
>>>    * doesn't even start to describe it. Somebody in UCB had a heck of a trip...
>>> diff --git a/fs/pipe.c b/fs/pipe.c
>>> index 7f00e58..1325e84 100644
>>> --- a/fs/pipe.c
>>> +++ b/fs/pipe.c
>>> @@ -1021,7 +1021,7 @@ struct file *fifo_file_restore(struct ckpt_ctx *ctx, struct ckpt_hdr_file *ptr)
>>>   	 * To avoid blocking, always open the fifo with O_RDWR;
>>>   	 * then fix flags below.
>>>   	 */
>>> -	file = restore_open_fname(ctx, (ptr->f_flags&   ~O_ACCMODE) | O_RDWR);
>>> +	file = restore_open_fname(ctx, 0, (ptr->f_flags&   ~O_ACCMODE) | O_RDWR);
>>>   	if (IS_ERR(file))
>>>   		return file;
>>>
>>> diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
>>> index 4e25042..6ca7b24 100644
>>> --- a/include/linux/checkpoint.h
>>> +++ b/include/linux/checkpoint.h
>>> @@ -258,7 +258,8 @@ extern int restore_obj_file_table(struct ckpt_ctx *ctx, int files_objref);
>>>   /* files */
>>>   extern int checkpoint_fname(struct ckpt_ctx *ctx,
>>>   			    struct path *path, struct path *root);
>>> -extern struct file *restore_open_fname(struct ckpt_ctx *ctx, int flags);
>>> +extern int checkpoint_file_links(struct ckpt_ctx *ctx, struct file *file);
>>> +extern struct file *restore_open_fname(struct ckpt_ctx *ctx, int restore_unlinked, int flags);
>>>
>>>   extern int ckpt_collect_file(struct ckpt_ctx *ctx, struct file *file);
>>>
>>> diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
>>> index f4f9577..ea50e7d 100644
>>> --- a/include/linux/checkpoint_hdr.h
>>> +++ b/include/linux/checkpoint_hdr.h
>>> @@ -575,6 +575,9 @@ struct ckpt_hdr_file {
>>>   	__u64 f_pos;
>>>   	__u64 f_version;
>>>   	__s32 f_secref;
>>> +
>>> +	__u32 f_restart_flags;
>>> +#define RESTART_FILE_F_UNLINK (1<<0)
>>>   } __attribute__((aligned(8)));
>>>
>>>   struct ckpt_hdr_file_generic {
>>> diff --git a/include/linux/checkpoint_types.h b/include/linux/checkpoint_types.h
>>> index 3ffe9bd..ceaa671 100644
>>> --- a/include/linux/checkpoint_types.h
>>> +++ b/include/linux/checkpoint_types.h
>>> @@ -57,6 +57,9 @@ struct ckpt_ctx {
>>>
>>>   	struct path root_fs_path;     /* container root (FIXME) */
>>>
>>> +	/* relink unlinked files to<mnt_root>/<unique_name>   */
>>> +	unsigned int unique_name_count;
>>
>> [nit] why not straightforward "unique_relink_count" ?
>
> Sure, that does look better.
>
>>
>>> +
>>>   	struct task_struct *tsk;/* checkpoint: current target task */
>>>   	char err_string[256];	/* checkpoint: error string */
>>>
>>
>> Phew ... end of round one :)
>
> If you read my long-winded responses above then you know that you spoke
> too soon! ;)

Ready for round two ...

Oren