| Message ID | 20230724060336.8939-1-nj.shetty@samsung.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | fs/read_write: Enable copy_file_range for block device. | expand |
On Mon, Jul 24, 2023 at 11:33:36AM +0530, Nitesh Shetty wrote: > From: Anuj Gupta <anuj20.g@samsung.com> > > Change generic_copy_file_checks to use ->f_mapping->host for both inode_in > and inode_out. Allow block device in generic_file_rw_checks. Why? copy_file_range() is for copying a range of a regular file to another regular file - why do we want to support block devices for somethign that is clearly intended for copying data files? Also, the copy_file_range man page states: ERRORS ..... EINVAL Either fd_in or fd_out is not a regular file. ..... If we are changing the behavioru of copy_file_range (why?), then man page updates need to be done as well, documenting the change, which kernel versions only support regular files, etc. Cheers, Dave.
> > Change generic_copy_file_checks to use ->f_mapping->host for both inode_in > > and inode_out. Allow block device in generic_file_rw_checks. > > Why? copy_file_range() is for copying a range of a regular file to > another regular file - why do we want to support block devices for > somethign that is clearly intended for copying data files? Nitesh has a series to add block layer copy offload, and uses that to implement copy_file_range on block device nodes, which seems like a sensible use case for copy_file_range on block device nodes, and that series was hiding a change like this deep down in a "block" title patch, so I asked for it to be split out. It still really should be in that series, as there's very little point in changing this check without an actual implementation making use of it.
> { > - struct inode *inode_in = file_inode(file_in); > - struct inode *inode_out = file_inode(file_out); > + struct inode *inode_in = file_in->f_mapping->host; > + struct inode *inode_out = file_out->f_mapping->host; This doesn't directly have anything to do with block devices, as regular files can also have a f_mapping that's different. None of the file systems actually supporting copy offload right now do, but changing the dereference here is a correctness thing totally independent of block device support.
On Mon, Jul 24, 2023 at 06:38:38PM +0200, Christoph Hellwig wrote: > > > Change generic_copy_file_checks to use ->f_mapping->host for both inode_in > > > and inode_out. Allow block device in generic_file_rw_checks. > > > > Why? copy_file_range() is for copying a range of a regular file to > > another regular file - why do we want to support block devices for > > somethign that is clearly intended for copying data files? > > Nitesh has a series to add block layer copy offload, Yes, I know. > and uses that to > implement copy_file_range on block device nodes, Yes, I know. > which seems like a > sensible use case for copy_file_range on block device nodes, Except for the fact it's documented and implemented as for copying data ranges of *regular files*. Block devices are not regular files... There is nothing in this patchset that explains why this syscall feature creep is desired, why it is the best solution, what benefits it provides, how this new feature is discoverable, etc. It also does not mention that user facing documentation needs to change, etc > and that > series was hiding a change like this deep down in a "block" title > patch, I know. > so I asked for it to be split out. It still really should > be in that series, as there's very little point in changing this > check without an actual implementation making use of it. And that's because it's the wrong way to be implementing block device copy offloads. That patchset originally added ioctls to issue block copy offloads to block devices from userspace - that's the way block device specific functionality is generally added and I have no problems with that. However, when I originally suggested that the generic copy_file_range() fallback path that filesystems use (i.e. generic_copy_file_range()) should try to use the block copy offload path before falling back to using splice to copy the data through host memory, things went off the rails. That has turned into "copy_file_range() should support block devices directly" and the ioctl interfaces were removed. The block copy offload patchset still doesn't have a generic path for filesystems to use this block copy offload. This is *not* what I originally suggested, and provides none of the benefit to users that would come from what I originally suggested. Block device copy offload is largely useless to users if file data copies within a filesystem don't make use of it - very few applications ever copy data directly to/from block devices directly... So from a system level point of view, expanding copy_file_range() to do direct block device data copies doesn't make any sense. Expanding the existing copy_file_range() generic fallback to attempt block copy offload (i.e. hardware accel) makes much more sense, and will make copy_file_range() much more useful to users on filesystem like ext4 that don't have reflink support... So, yeah, this patch, regardless of how it is presented, needs to a whole lot more justification that "we want to do this" in the commit message... -Dave.
Hi Dave, On 23/07/25 08:08AM, Dave Chinner wrote: >On Mon, Jul 24, 2023 at 06:38:38PM +0200, Christoph Hellwig wrote: >> > > Change generic_copy_file_checks to use ->f_mapping->host for both inode_in >> > > and inode_out. Allow block device in generic_file_rw_checks. >> > >> > Why? copy_file_range() is for copying a range of a regular file to >> > another regular file - why do we want to support block devices for >> > somethign that is clearly intended for copying data files? >> >> Nitesh has a series to add block layer copy offload, > >Yes, I know. > >> and uses that to >> implement copy_file_range on block device nodes, > >Yes, I know. > >> which seems like a >> sensible use case for copy_file_range on block device nodes, > >Except for the fact it's documented and implemented as for copying >data ranges of *regular files*. Block devices are not regular >files... > >There is nothing in this patchset that explains why this syscall >feature creep is desired, why it is the best solution, what benefits >it provides, how this new feature is discoverable, etc. It also does >not mention that user facing documentation needs to change, etc > Agreed. I missed adding context in description. >> and that >> series was hiding a change like this deep down in a "block" title >> patch, > >I know. > >> so I asked for it to be split out. It still really should >> be in that series, as there's very little point in changing this >> check without an actual implementation making use of it. > >And that's because it's the wrong way to be implementing block >device copy offloads. > >That patchset originally added ioctls to issue block copy offloads >to block devices from userspace - that's the way block device >specific functionality is generally added and I have no problems >with that. > We moved to copy_file_range, so that we can reuse the existing infra instead of adding another ioctl[1]. >However, when I originally suggested that the generic >copy_file_range() fallback path that filesystems use (i.e. >generic_copy_file_range()) should try to use the block copy offload >path before falling back to using splice to copy the data through >host memory, things went off the rails. > >That has turned into "copy_file_range() should support block devices >directly" and the ioctl interfaces were removed. The block copy >offload patchset still doesn't have a generic path for filesystems >to use this block copy offload. This is *not* what I originally >suggested, and provides none of the benefit to users that would come >from what I originally suggested. Block device copy offload is >largely useless to users if file data copies within a filesystem >don't make use of it - very few applications ever copy data directly >to/from block devices directly... > >So from a system level point of view, expanding copy_file_range() to >do direct block device data copies doesn't make any sense. Expanding >the existing copy_file_range() generic fallback to attempt block >copy offload (i.e. hardware accel) makes much more sense, and will >make copy_file_range() much more useful to users on filesystem like >ext4 that don't have reflink support... > Agreed. But adding all the cases is making the series heavier and harder to iterate. So we trimmed some of the patches and feature. Hopefully we can get to filesystems, if the current series gets in. >So, yeah, this patch, regardless of how it is presented, needs to a >whole lot more justification that "we want to do this" in the commit >message... > Agreed, the commit description was not conveying the things we wanted to do. It makes sense to send block related relaxation as part of copy offload series instead of doing it here. However, the change to get inode pointer using mapping_host is still independent and can go as a separate fix patch. Thank you, Nitesh Shetty [1] https://lore.kernel.org/all/Y3607N6lDRK6WU7%2F@T590/
diff --git a/fs/read_write.c b/fs/read_write.c index b07de77ef126..eaeb481477f4 100644 --- a/fs/read_write.c +++ b/fs/read_write.c @@ -1405,8 +1405,8 @@ static int generic_copy_file_checks(struct file *file_in, loff_t pos_in, struct file *file_out, loff_t pos_out, size_t *req_count, unsigned int flags) { - struct inode *inode_in = file_inode(file_in); - struct inode *inode_out = file_inode(file_out); + struct inode *inode_in = file_in->f_mapping->host; + struct inode *inode_out = file_out->f_mapping->host; uint64_t count = *req_count; loff_t size_in; int ret; @@ -1708,7 +1708,9 @@ int generic_file_rw_checks(struct file *file_in, struct file *file_out) /* Don't copy dirs, pipes, sockets... */ if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) return -EISDIR; - if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) + if (!S_ISREG(inode_in->i_mode) && !S_ISBLK(inode_in->i_mode)) + return -EINVAL; + if ((inode_in->i_mode & S_IFMT) != (inode_out->i_mode & S_IFMT)) return -EINVAL; if (!(file_in->f_mode & FMODE_READ) ||