On Tue, 2018-06-19 at 21:42 +0200, Arnd Bergmann wrote:
On Tue, Jun 19, 2018 at 7:03 PM, Viacheslav Dubeyko <slava@dubeyko.co m> wrote:
On Tue, 2018-06-19 at 18:02 +0200, Arnd Bergmann wrote:
The native HFS timestamps overflow in year 2040, two years after the Unix y2038 overflow. However, the way that the conversion between on- disk timestamps and in-kernel timestamps was implemented, 64-bit machines actually ended up converting negative UTC timestamps (1902 through 1969) into times between 2038 and 2106.
Rather than making all machines faithfully represent timestamps in the ancient past but break after 2040, this changes the file system to always use the unsigned UTC interpretation, reading back times between 1970 and 2106.
The trouble with HFS and HFS+ that the specification [1] declares this:
"HFS Plus stores dates in several data structures, including the volume header and catalog records. These dates are stored in unsigned 32- bit integers (UInt32) containing the number of seconds since midnight, January 1, 1904, GMT. This is slightly different from HFS, where the value represents local time. The maximum representable date is February 6, 2040 at 06:28:15 GMT."
So, I am not sure that we are able to support later dates because such timestamps cannot be stored on HFS/HFS+ volumes and will be incompatible with Mac OS X.
We never followed that interpretation in Linux though. As I wrote, on 64-bit machines, these two calculations (hfs and hfs+, respectively)
#define __hfs_m_to_utime(sec) (be32_to_cpu(sec) - 2082844800U + sys_tz.tz_minuteswest * 60) #define __hfsp_mt2ut(t) (be32_to_cpu(t) - 2082844800U)
just wrap around when reading the timestamps before 1970 from disk. On 32-bit machines they get wrapped another time when we assign them to a signed 32-bit time_t.
The whole patchset looks reasonable for me. I simply guess what the correct behaviour of HFS/HFS+ file system driver could look like for the case of achieving 2040 year. So, maybe the good way could be to mount in the READ-ONLY mode. What do you think?
Also, I am not sure that anybody will use HFS/HFS+ after 2040.
I'm trying to fix all file systems to be unambiguous regarding inode timestamps. This means it should behave the same way on 32-bit and 64-bit kernels, and if possible in a sane way.
Even if you don't care about running HFS in the future, you can trivially create files with arbitrary timestamps, just try
touch -d "Jan 1 1901" 1901 touch -d "Jan 1 1905" 1905 touch -d "Jan 1 1969" 1969 touch -d "Jan 1 2038" 2038 touch -d "Jan 1 2040" 2040 touch -d "Jan 1 2106" 2106 touch -d "Jan 1 2107" 2107
on HFS and do an 'ls -l' after an unmount/remount.
If you think it's important that we change the current behavior to be compatible with MacOS and represent the 1904..2040 time range rather than 1970..2106, we can definitely do that as well, using this patch:
diff --git a/fs/hfs/hfs_fs.h b/fs/hfs/hfs_fs.h index ff432931a5b1..2c7366342656 100644 --- a/fs/hfs/hfs_fs.h +++ b/fs/hfs/hfs_fs.h @@ -249,7 +249,7 @@ extern void hfs_mark_mdb_dirty(struct super_block *sb); * actually works until year 2106 */ #define __hfs_u_to_mtime(sec) cpu_to_be32(sec + 2082844800U - sys_tz.tz_minuteswest * 60) -#define __hfs_m_to_utime(sec) (be32_to_cpu(sec) - 2082844800U + sys_tz.tz_minuteswest * 60) +#define __hfs_m_to_utime(sec) ((time64_t)be32_to_cpu(sec) - 2082844800U + sys_tz.tz_minuteswest * 60)
#define HFS_I(inode) (container_of(inode, struct hfs_inode_info, vfs_inode)) #define HFS_SB(sb) ((struct hfs_sb_info *)(sb)->s_fs_info) diff --git a/fs/hfsplus/hfsplus_fs.h b/fs/hfsplus/hfsplus_fs.h index 1a6b469f8d22..4eaee8bdfcb2 100644 --- a/fs/hfsplus/hfsplus_fs.h +++ b/fs/hfsplus/hfsplus_fs.h @@ -534,7 +534,7 @@ int hfsplus_read_wrapper(struct super_block *sb);
/* time macros: convert between 1904-2040 and 1970-2106 range, * pre-1970 timestamps are interpreted as post-2038 times after wrap-around */ -#define __hfsp_mt2ut(t) (be32_to_cpu(t) - 2082844800U) +#define __hfsp_mt2ut(t) ((time64_t)be32_to_cpu(t) - 2082844800U) #define __hfsp_ut2mt(t) (cpu_to_be32(t + 2082844800U))
/* compatibility */
I can submit that separately so that it can get backported into stable kernels if you like, with the type changes as a follow-up on top.
Sounds good.
Thanks, Vyacheslav Dubeyko.