On Wed, Jun 20, 2018 at 12:53 PM, David Howells dhowells@redhat.com wrote:
Arnd Bergmann arnd@arndb.de wrote:
which avoids the y2038 overflow
No it doesn't. The AFS protocol is limited.
time64_t mtime_client; /* last time client changed data */time64_t mtime_server; /* last time server changed data */...
time_t creation; /* volume creation time */
time64_t creation; /* volume creation time */Unless you can change the AFS protocol, this is a waste of memory. It might be better to change them to u32 as they are protocol values rather than system values.
AFS uses 'unsigned' seconds, right? What I was trying to say there is that with the patch, the 32-bit overflow gets moved from 2038 to 2106, so at least the nearer problem is solved.
On 64-bit machines, we already waste a little memory here, the usual tradeoff I took was to use time64_t for all time storage when possible for clarity reasons, but that is easily changed if you prefer.
inode->i_ctime.tv_sec = get_seconds();inode->i_ctime.tv_nsec = 0;inode->i_atime = inode->i_mtime = inode->i_ctime;
inode->i_ctime = inode->i_atime = inode->i_mtime = current_time(inode);Surely, the tv_nsec should be zero since anything else cannot be represented in the AFS protocol.
current_time() truncates the nanoseconds to the granularity of the filesystem. Since AFS doesn't set s_time_gran, it gets the default 1000000000 value leads to tv_nsec being zero. Once Deepa's patch to truncate the tv_sec range lands, it will also ensure that this is within the range (this is less of a problem for setting the current time than it is for utimensat() which can set arbitrary future timestamps of course).
I will grant, however, I should be consistently using them as unsigned values.
Note that the answers to the above may change if and when I start supporting the YFS protocol extensions, but for the AFS protocol, this is simply not there.
Ok, good to know this exists.
Arnd