On Mon, Dec 2, 2019 at 1:55 PM Christophe Leroy christophe.leroy@c-s.fr wrote:
Le 27/11/2019 à 12:03, Arnd Bergmann a écrit :
On Thu, Nov 21, 2019 at 5:25 PM Christophe Leroy christophe.leroy@c-s.fr wrote:
Arnd Bergmann arnd@arndb.de a écrit :
On Wed, Nov 20, 2019 at 11:43 PM Ben Hutchings ben.hutchings@codethink.co.uk wrote:
On Fri, 2019-11-08 at 22:07 +0100, Arnd Bergmann wrote:
@@ -192,7 +190,7 @@ V_FUNCTION_BEGIN(__kernel_time) bl __get_datapage@local mr r9, r3 /* datapage ptr in r9 */
lwz r3,STAMP_XTIME+TSPEC_TV_SEC(r9)
lwz r3,STAMP_XTIME_SEC+LOWPART(r9)"LOWPART" should be "LOPART".
Thanks, fixed both instances in a patch on top now. I considered folding it into the original patch, but as it's close to the merge window I'd rather not rebase it, and this way I also give you credit for finding the bug.
Take care, might conflict with https://github.com/linuxppc/linux/commit/5e381d727fe8834ca5a126f510194a7a4ac...
Sorry for my late reply. I see this commit and no other variant of it has made it into linux-next by now, so I assume this is not getting sent for v5.5 and it's not stopping me from sending my own pull request.
Please let me know if I missed something and this will cause problems.
On a related note: are you still working on the generic lib/vdso support for powerpc? Without that, future libc implementations that use 64-bit time_t will have to use the slow clock_gettime64 syscall instead of the vdso, which has a significant performance impact.
I have left this generic lib/vdso subject aside for the moment, because performance is disappointing and its architecture doesn't real fit with powerpc ABI.
From a performance point of view, it is manipulating 64 bits vars where is could use 32 bits vars. Of course I understand that y2038 will anyway force the use of 64 bits for seconds, but at the time being powerpc32 VDSO is using 32 bits vars for both secs and ns, it make the difference.
Do you think we could optimize the common code? This sounds like it could improve things for other architectures as well.
Also, the generic VDSO is playing too much with data on stacks and associated memory read/write/copies, which kills performance on RISC processors like powerpc. Inlining do_hres() for instance significantly improves that as it allow handling the 'struct __kernel_timespec ts' on registers instead of using stack.
That should be easy enough to change in the common code, as long as adding 'inline' does not cause harm on x86 and arm.
Regarding powerpc ABI, the issue is that errors shall be reported by setting the SO bit in CR register, and this cannot be done in C. This means:
- The VDSO entry point must be in ASM and the generic VDSO C function
must be called from there, it cannot be the VDSO entry point.
- The VDSO fallback (ie the system call) cannot be done from the generic
VDSO C function, it must be called from the ASM as well.
As far as I can tell, both the VDSO entry point and the fallback are in architecture specific code on all architectures, so this does not seem to be a show-stopper.
It also seems that they might be combined as long the current powerpc code could be changed to use the generic vdso_data structure definition: the existing code can keep being used for gettimeofday(), clock_gettime(CLOCK_MONOTONIC, ...) and clock_gettime(CLOCK_REALTIME), while the generic implementation can be called for clock_gettime64(), clock_getres() and clock_gettime() with other time clock IDs.
Last point/question, what's the point in using 64 bits for nanoseconds on 32 bits arches ?
The __kernel_timespec structure is defined with two 64-bit members so it has the same layout on both 32-bit and 64-bit architectures, which lets us share the implementation of the compat syscall handlers even on big-endian architectures, and it avoids accidentally leaking four bytes of stack data when copying a timespec from kernel to user space. The high 32 bits of the nanosecond are expected to always be zero when copying to user space, and to be ignored when copied into the kernel (see get_timespec64()).
Note that C99 and POSIX require tv_nsec to be 'long', so 64-bit architectures have to make it 64-bit wide, and 32-bit architectures end up including padding for it.
In the vdso_data, the "nsec" value is shifted, so it actually needs more bits. I don't know if this is a strict requirement, or if we could change it to be 32 bits non-shifted during the update at the cost of losing 1 nanosecond of accuracy.
Arnd