On 01/30/2017 03:15 PM, David Miller wrote:
From: Khalid Aziz <khalid.aziz@xxxxxxxxxx>
Date: Wed, 25 Jan 2017 12:57:16 -0700
+static inline void enable_adi(void)
+{
...
+ __asm__ __volatile__(
+ "rdpr %%pstate, %%g1\n\t"
+ "or %%g1, %0, %%g1\n\t"
+ "wrpr %%g1, %%g0, %%pstate\n\t"
+ ".word 0x83438000\n\t" /* rd %mcdper, %g1 */
+ ".word 0xaf900001\n\t" /* wrpr %g0, %g1, %pmcdper */
+ :
+ : "i" (PSTATE_MCDE)
+ : "g1");
+}
This is _crazy_ expensive.
This is 4 privileged register operations, every single one incurs a full
pipline flush and virtual cpu thread yield.
And we do this around _every_ single userspace access from the kernel
when the thread has ADI enabled.
Hi Dave,
Thanks for the feedback. This is very helpful. I checked and it indeed
can cost 50+ cycles even on M7 processor for PSTATE accesses.
I think if the kernel manages the ADI metadata properly, you can get rid
of all of this.
On etrap, you change ESTATE_PSTATE{1,2} to have the MCDE bit enabled.
Then the kernel always runs with ADI enabled.
Running the kernel with PSTATE.mcde=1 can possibly be problematic as we
had discussed earlier in this thread where keeping PSTATE.mcde enabled
might mean kernel having to keep track of which pages still have tags
set on them or flush tags on every page on free. I will go through the
code again to see if it PSTATE.mcde can be turned on in kernel all the
time, which might be the case if we can ensure kernel accesses pages
with TTE.mcd cleared.
Furthermore, since the %mcdper register should be set to whatever the
current task has asked for, you should be able to avoid touching it
as well assuming that traps do not change %mcdper's value.
When running in privileged mode, it is the value of %pmcdper that
matter, not %mcdper, hence I added code to sync %pmcdper with %mcdper
when entering privileged mode. Nevertheless, one of the HW designers has
suggested I might be able to get away without having to futz with
%pmcdper by using membar before exiting privileged mode which might
still get me the same effect I am looking for without the cost.
--
Khalid
Then you don't need to do anything special during userspace accesses
which seems to be the way this was designed to be used.
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
