On 2015-09-15 12:38, Darrick J. Wong wrote:
> On Tue, Sep 15, 2015 at 11:58:04AM -0400, Anna Schumaker wrote:
>> On 09/14/2015 11:32 PM, Darrick J. Wong wrote:
>>> On Fri, Sep 11, 2015 at 04:30:21PM -0400, Anna Schumaker wrote:
>>>> The NFS server will need some kind offallback for filesystems that don't
>>>> have any kind of copy acceleration, and it should be generally useful to
>>>> have an in-kernel copy to avoid lots of switches between kernel and user
>>>> space.
>>>>
>>>> I make this configurable by adding two new flags.  Users who only want a
>>>> reflink can pass COPY_FR_REFLINK, and users who want a full data copy can
>>>> pass COPY_FR_COPY.  The default (flags=0) means to first attempt a
>>>> reflink, but use the pagecache if that fails.
>>>>
>>>> I moved the rw_verify_area() calls into the fallback code since some
>>>> filesystems can handle reflinking a large range.
>>>>
>>>> Signed-off-by: Anna Schumaker <Anna.Schumaker-ZwjVKphTwtPQT0dZR+AlfA@public.gmane.org>
>>>> ---
>>>> v2:
>>>> - Rename COPY_REFLINK -> COPY_FR_REFLINK
>>>> - Introduce COPY_FR_COPY flag
>>>> - Flags == 0 is really COPY_FR_COPY|COPY_FR_REFLINK
>>>> - Drop check for invalid flags
>>>> - Move call to do_splice_direct() into a new function
>>>> - Move rw_verify_area() checks into the new fallback function
>>>> ---
>>>>   fs/read_write.c           | 56 ++++++++++++++++++++++++++++-------------------
>>>>   include/linux/copy.h      |  6 +++++
>>>>   include/uapi/linux/Kbuild |  1 +
>>>>   include/uapi/linux/copy.h |  7 ++++++
>>>>   4 files changed, 48 insertions(+), 22 deletions(-)
>>>>   create mode 100644 include/linux/copy.h
>>>>   create mode 100644 include/uapi/linux/copy.h
>>>>
>>>> diff --git a/fs/read_write.c b/fs/read_write.c
>>>> index 363bd3e..ba24884 100644
>>>> --- a/fs/read_write.c
>>>> +++ b/fs/read_write.c
>>>> @@ -7,6 +7,7 @@
>>>>   #include <linux/slab.h>
>>>>   #include <linux/stat.h>
>>>>   #include <linux/fcntl.h>
>>>> +#include <linux/copy.h>
>>>>   #include <linux/file.h>
>>>>   #include <linux/uio.h>
>>>>   #include <linux/fsnotify.h>
>>>> @@ -1329,6 +1330,29 @@ COMPAT_SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd,
>>>>   }
>>>>   #endif
>>>>
>>>> +static ssize_t vfs_copy_file_pagecache(struct file *file_in, loff_t pos_in,
>>>> +				       struct file *file_out, loff_t pos_out,
>>>> +				       size_t len)
>>>> +{
>>>> +	ssize_t ret;
>>>> +
>>>> +	ret = rw_verify_area(READ, file_in, &pos_in, len);
>>>> +	if (ret >= 0) {
>>>> +		len = ret;
>>>> +		ret = rw_verify_area(WRITE, file_out, &pos_out, len);
>>>> +		if (ret >= 0)
>>>> +			len = ret;
>>>> +	}
>>>> +	if (ret < 0)
>>>> +		return ret;
>>>> +
>>>> +	file_start_write(file_out);
>>>> +	ret = do_splice_direct(file_in, &pos_in, file_out, &pos_out, len, 0);
>>>> +	file_end_write(file_out);
>>>> +
>>>> +	return ret;
>>>> +}
>>>> +
>>>>   /*
>>>>    * copy_file_range() differs from regular file read and write in that it
>>>>    * specifically allows return partial success.  When it does so is up to
>>>> @@ -1338,34 +1362,17 @@ ssize_t vfs_copy_file_range(struct file *file_in, loff_t pos_in,
>>>>   			    struct file *file_out, loff_t pos_out,
>>>>   			    size_t len, int flags)
>>>>   {
>>>> -	struct inode *inode_in;
>>>> -	struct inode *inode_out;
>>>>   	ssize_t ret;
>>>>
>>>> -	if (flags)
>>>> -		return -EINVAL;
>>>> -
>>>> -	/* copy_file_range allows full ssize_t len, ignoring MAX_RW_COUNT  */
>>>> -	ret = rw_verify_area(READ, file_in, &pos_in, len);
>>>> -	if (ret >= 0)
>>>> -		ret = rw_verify_area(WRITE, file_out, &pos_out, len);
>>>> -	if (ret < 0)
>>>> -		return ret;
>>>> +	if (flags == 0)
>>>> +		flags = COPY_FR_COPY | COPY_FR_REFLINK;
>>>
>>> This function must return -EINVAL if any of the undefined flags bits are
>>> set.
>>
>> Sure, I'll add that.
>>
>>>
>>>>
>>>>   	if (!(file_in->f_mode & FMODE_READ) ||
>>>>   	    !(file_out->f_mode & FMODE_WRITE) ||
>>>>   	    (file_out->f_flags & O_APPEND) ||
>>>> -	    !file_out->f_op || !file_out->f_op->copy_file_range)
>>>> +	    !file_in->f_op)
>>>>   		return -EBADF;
>>>>
>>>> -	inode_in = file_inode(file_in);
>>>> -	inode_out = file_inode(file_out);
>>>> -
>>>> -	/* make sure offsets don't wrap and the input is inside i_size */
>>>> -	if (pos_in + len < pos_in || pos_out + len < pos_out ||
>>>> -	    pos_in + len > i_size_read(inode_in))
>>>> -		return -EINVAL;
>>>> -
>>>>   	if (len == 0)
>>>>   		return 0;
>>>>
>>>> @@ -1373,8 +1380,13 @@ ssize_t vfs_copy_file_range(struct file *file_in, loff_t pos_in,
>>>>   	if (ret)
>>>>   		return ret;
>>>>
>>>> -	ret = file_out->f_op->copy_file_range(file_in, pos_in, file_out, pos_out,
>>>> -					      len, flags);
>>>> +	ret = -EOPNOTSUPP;
>>>> +	if (file_out->f_op->copy_file_range)
>>>> +		ret = file_out->f_op->copy_file_range(file_in, pos_in, file_out,
>>>> +						      pos_out, len, flags);
>>>> +	if ((ret < 0) && (flags & COPY_FR_COPY))
>>>> +		ret = vfs_copy_file_pagecache(file_in, pos_in, file_out,
>>>> +					      pos_out, len);
>>>>   	if (ret > 0) {
>>>>   		fsnotify_access(file_in);
>>>>   		add_rchar(current, ret);
>>>> diff --git a/include/linux/copy.h b/include/linux/copy.h
>>>> new file mode 100644
>>>> index 0000000..fd54543
>>>> --- /dev/null
>>>> +++ b/include/linux/copy.h
>>>> @@ -0,0 +1,6 @@
>>>> +#ifndef _LINUX_COPY_H
>>>> +#define _LINUX_COPY_H
>>>> +
>>>> +#include <uapi/linux/copy.h>
>>>> +
>>>> +#endif /* _LINUX_COPY_H */
>>>> diff --git a/include/uapi/linux/Kbuild b/include/uapi/linux/Kbuild
>>>> index 70ff1d9..d46830a 100644
>>>> --- a/include/uapi/linux/Kbuild
>>>> +++ b/include/uapi/linux/Kbuild
>>>> @@ -90,6 +90,7 @@ header-y += coda_psdev.h
>>>>   header-y += coff.h
>>>>   header-y += connector.h
>>>>   header-y += const.h
>>>> +header-y += copy.h
>>>>   header-y += cramfs_fs.h
>>>>   header-y += cuda.h
>>>>   header-y += cyclades.h
>>>> diff --git a/include/uapi/linux/copy.h b/include/uapi/linux/copy.h
>>>> new file mode 100644
>>>> index 0000000..2da59a8
>>>> --- /dev/null
>>>> +++ b/include/uapi/linux/copy.h
>>>> @@ -0,0 +1,7 @@
>>>> +#ifndef _UAPI_LINUX_COPY_H
>>>> +#define _UAPI_LINUX_COPY_H
>>>> +
>>>> +#define COPY_FR_COPY		(1 << 0)  /* Only do a pagecache copy.  */
>>>> +#define COPY_FR_REFLINK		(1 << 1)  /* Only make a reflink.       */
>>>
>>> Could I have a COPY_FR_DEDUPE flag too, please?
>>>
>>> (I don't mind adding it myself when I get around to hooking up XFS, but I
>>> was hoping to get it in during the first round).
>>
>> I guess I can, but only iff everybody has agreed on using copy for dedupes
>> instead of somethink like fallocate.
>
> I don't see how fallocate could even handle userspace-directed deduplication
> since the dedupe operation compares two file ranges and reflinks them if the
> contents match whereas fallocate only operates on a single file range.
>
> A single file range would work if userspace was telling the kernel to start a
> lengthy search for all duplicates of a particular file's range, but ...
> yuck.
>
> (Granted, the btrfs extent_same ioctl lets userspace call dedupe on multiple
> files simultaneously, but that's for another day.)
>
Agreed, having something equivalent to reflinking would be the most 
sensible interface for deduplicating two files.  I can however see 
fallocate possibly having some value when in-band deduplication is 
considered (ie, use fallocate to tell the system to include/exclude 
certain file ranges from in-band deduplication).