On Sun, Nov 24, 2019 at 11:49 AM Al Viro viro@zeniv.linux.org.uk wrote:
On Sun, Nov 24, 2019 at 09:31:45PM +0200, Amir Goldstein wrote:
Push clamping timestamps down the call stack into notify_change(), so in-kernel callers like nfsd and overlayfs will get similar timestamp set behavior as utimes.
Makes sense; said that, shouldn't we go through ->setattr() instances and get rid of that there, now that notify_change() is made to do it?
I mean, if (ia_valid & ATTR_ATIME) sd_iattr->ia_atime = timestamp_truncate(iattr->ia_atime, inode); in configfs_setattr() looks like it should be reverted to if (ia_valid & ATTR_ATIME) sd_iattr->ia_atime = iattr->ia_atime; with that, etc.
Moreover, does that leave any valid callers of timestamp_truncate() outside of notify_change() and current_time()? IOW, is there any point having it exported? Look: fs/attr.c:187: inode->i_atime = timestamp_truncate(attr->ia_atime, fs/attr.c:191: inode->i_mtime = timestamp_truncate(attr->ia_mtime, fs/attr.c:195: inode->i_ctime = timestamp_truncate(attr->ia_ctime, setattr_copy(), called downstream of your changes. fs/configfs/inode.c:79: sd_iattr->ia_atime = timestamp_truncate(iattr->ia_atime, fs/configfs/inode.c:82: sd_iattr->ia_mtime = timestamp_truncate(iattr->ia_mtime, fs/configfs/inode.c:85: sd_iattr->ia_ctime = timestamp_truncate(iattr->ia_ctime, configfs_setattr(); ditto. fs/f2fs/file.c:755: inode->i_atime = timestamp_truncate(attr->ia_atime, fs/f2fs/file.c:759: inode->i_mtime = timestamp_truncate(attr->ia_mtime, fs/f2fs/file.c:763: inode->i_ctime = timestamp_truncate(attr->ia_ctime, __setattr_copy() from f2fs_setattr(); ditto. fs/inode.c:2224: return timestamp_truncate(now, inode); current_time() fs/kernfs/inode.c:163: inode->i_atime = timestamp_truncate(attrs->ia_atime, inode); fs/kernfs/inode.c:164: inode->i_mtime = timestamp_truncate(attrs->ia_mtime, inode); fs/kernfs/inode.c:165: inode->i_ctime = timestamp_truncate(attrs->ia_ctime, inode); ->s_time_max and ->s_time_min are left TIME64_MAX and TIME64_MIN resp., so timestamp_truncate() should be a no-op there. fs/ntfs/inode.c:2903: vi->i_atime = timestamp_truncate(attr->ia_atime, fs/ntfs/inode.c:2907: vi->i_mtime = timestamp_truncate(attr->ia_mtime, fs/ntfs/inode.c:2911: vi->i_ctime = timestamp_truncate(attr->ia_ctime, ntfs_setattr(); downstream from your changes fs/ubifs/file.c:1082: inode->i_atime = timestamp_truncate(attr->ia_atime, fs/ubifs/file.c:1086: inode->i_mtime = timestamp_truncate(attr->ia_mtime, fs/ubifs/file.c:1090: inode->i_ctime = timestamp_truncate(attr->ia_ctime, do_attr_changes(), from do_truncation() or do_setattr(), both from ubifs_setattr(); ditto. fs/utimes.c:39: newattrs.ia_atime = timestamp_truncate(times[0], inode); fs/utimes.c:46: newattrs.ia_mtime = timestamp_truncate(times[1], inode); disappears in your patch.
We also want to replace all uses of timespec64_trunc() with timestamp_truncate() for all fs cases.
In that case we have a few more:
fs/ceph/mds_client.c: req->r_stamp = timespec64_trunc(ts, mdsc->fsc->sb->s_time_gran); fs/cifs/inode.c: fattr->cf_mtime = timespec64_trunc(fattr->cf_mtime, sb->s_time_gran); fs/cifs/inode.c: fattr->cf_atime = timespec64_trunc(fattr->cf_atime, sb->s_time_gran); fs/fat/misc.c:static inline struct timespec64 fat_timespec64_trunc_2secs(struct timespec64 ts) fs/fat/misc.c: inode->i_ctime = timespec64_trunc(*now, 10000000); fs/fat/misc.c: inode->i_ctime = fat_timespec64_trunc_2secs(*now); fs/fat/misc.c: inode->i_mtime = fat_timespec64_trunc_2secs(*now); fs/ubifs/sb.c: ts = timespec64_trunc(ts, DEFAULT_TIME_GRAN);
These do not follow from notify_change(), so these might still need timestamp_truncate() exported. I will post a cleanup series for timespec64_trunc() also, then we can decide.
-Deepa