Use memory barriers defined by xen headers - #28
Conversation
This gives us the necessary definitions for arm64. Those live in xen-barrier.h starting with 4.20, but we won't be able to use it for some time. Before this they live in xenctrl.h, but it pulls too much stuff. So this is a sync with the 4.20 headers contents. Since the code only uses mb, and no [rw]mb, the latter are not imported. The original macro name from the Xen header is preserved for easier future synchro, so this adjusts the callsites accordingly. Signed-off-by: Yann Dirson <yann.dirson@vates.tech>
| * back to those provided by the kernel */ | ||
| /* from <xen-barrier.h> in 4.20.0 */ | ||
|
|
||
| #define xen_barrier() asm volatile ( "" : : : "memory") |
There was a problem hiding this comment.
Why is this added if it is not used?
Having said that, I think it might be preferable to simply include a copy of xen-barrier.h and #include it. It would be easier to keep up-to-date and mean less tat in this file.
| #elif defined(__arm__) | ||
| #define xen_mb() asm volatile ("dmb" : : : "memory") | ||
| #elif defined(__aarch64__) | ||
| #define xen_mb() asm volatile ("dmb sy" : : : "memory") |
There was a problem hiding this comment.
This is a monster of a barrier. This orders against all observers: every core in every cluster, plus outer-shareable and non-shareable agents, GPU, other DMA masters, the interconnect fabric. This brings the entire system to a halt.
In the Linux kernel there a number of barriers defined for core coherency:
#define __smp_mb() dmb(ish)
#define __smp_rmb() dmb(ishld)
#define __smp_wmb() dmb(ishst)
dmb ish (inner-shareable) is for ordering with respect to other cores in the same coherency domain, with no device/DMA involvement. This is what Linux's arm64 barrier.h defines. The dmb sy is only used when talking to MMIO or DMA devices.
|
|
||
| /* Get the correct defines for memory barriers - do not fall | ||
| * back to those provided by the kernel */ | ||
| /* from <xen-barrier.h> in 4.20.0 */ |
There was a problem hiding this comment.
Can't we do:
#if __has_include(<xen-barrier.h>)
#include <xen-barrier.h>
#else
/* local fallback copy for pre-4.20 Xen headers */
...
#endif
There was a problem hiding this comment.
Actually I wouldn't even wait for including the xen-barrier.h header. Because after discussing with @andyhhp I would just add my own barrier code, which doesn't do dmb sy for performance.
3 participants
This gives us the necessary definitions for arm64.
Those live in xen-barrier.h starting with 4.20, but we won't be able to use it for some time. Before this they live in xenctrl.h, but it pulls too much stuff. So this is a sync with the 4.20 headers contents.
Since the code only uses mb, and no [rw]mb, the latter are not imported. The original macro name from the Xen header is preserved for easier future synchro, so this adjusts the callsites accordingly.