On Mon, Sep 17, 2018 at 9:46 PM Helge Deller deller@gmx.de wrote:
On 13.09.2018 17:59, Arnd Bergmann wrote:
There are only two 64-bit architecture ports that have a 32-bit suseconds_t: sparc64 and parisc64. I've encountered a number of problems with this, while trying to get a proper 64-bit time_t working on 32-bit architectures. Having a 32-bit suseconds_t combined with a 64-bit time_t means that we get extra padding in data structures that may leak kernel stack data to user space, and it breaks all code that assumes that timespec and timeval have the same layout.
While we can't change sparc64, it seems that glibc on parisc64 has always set suseconds_t to 'long', and the current version would give incorrect results for gettimeofday() and many other interfaces: timestamps passed from user space into the kernel result in tv_usec being always zero (the lower bits contain the intended value but are ignored) while data passed from the kernel to user space contains either zeroes or random data in tv_usec.
[back from traveling now, sorry for the delay in replying]
Should this wrong behavior be visible with 32-bit userspace or with 64-bit userspace (or both)? I didn't noticed such wrong behavior yet.
Only 64-bit user space.
Can you suggest some test programs to verify? LTP seems to be correct.
A simple 64-bit gettimeofday() should report incorrect nanoseconds using the upstream glibc implementation.
Based on that, it seems best to change the user API in the kernel in an incompatible way to match what glibc expects.
Basically that sounds the right way to go, but as said before, I didn't see such issues.
Note that the distros I could find (gentoo and debian) all just have 32-bit user space, which does not suffer from this problem.
That's correct. Kernel can be 32- or 64-bit, but userspace is currentl 32-bit only.
So, breaking any 64-bit userspace is OK, since we don't have it yet. Breaking 32-bit userspace needs some thoughts...
Ok.
Arnd