delete sync.i in ia64_switch_to()
PATCH 1/1 delete sync.i in ia64_switch_to()
The sync.i instruction in ia64_switch_to() context switch code looks like an old leftover code. sync.i instruction supposedly ensures flush cache operation issued by processor become visible to memory reference. Everywhere I looked where fc instruction is used (ia64_fc), a sync.i is always accompanied. Everywhere in the vicinity of context switch (save_switch_stack/load_switch_stack) don't have any fc instruction in there. I don't see any reason why we need to issue sync.i in context switch code. Patch to remove it.
arch/ia64/kernel/entry.S renamed from arch/ia64/kernel/entry.S.orig +0 -4
@@ -213,10 +213,6 @@ GLOBAL_ENTRY(ia64_switch_to) mov r13=in0 // set "current" pointer ;; DO_LOAD_SWITCH_STACK - -#ifdef CONFIG_SMP - sync.i // ensure "fc"s done by this CPU are visible on other CPUs -#endif br.ret.sptk.many rp // boogie on out in new context .map:
Chen, Kenneth wrote on Wednesday, September 07, 2005 12:22 AM
Quoted 8 lines
The sync.i instruction in
ia64_switch_to()context switch code looks like an old leftover code. sync.i instruction supposedly ensures flush cache operation issued by processor become visible to memory reference. Everywhere I looked where fc instruction is used (ia64_fc), a sync.i is always accompanied. Everywhere in the vicinity of context switch (save_switch_stack/load_switch_stack) don't have any fc instruction in there. I don't see any reason why we need to issue sync.i in context switch code. Patch to remove it.
While I'm at it, the reenabling of psr.ic should really belong to dtr mapping code block. It make the fall through code fast since it doesn't need to execute the predicated-off instruction. Logically make more sense as well since psr.ic was turned off in .map code block.
arch/ia64/kernel/entry.S renamed from arch/ia64/kernel/entry.S.orig +3 -3
@@ -204,9 +204,6 @@ GLOBAL_ENTRY(ia64_switch_to) (p6) br.cond.dpnt .map ;; .done: -(p6) ssm psr.ic // if we had to map, reenable the psr.ic bit FIRST!!! - ;; -(p6) srlz.d ld8 sp=[r21] // load kernel stack pointer of new task mov IA64_KR(CURRENT)=in0 // update "current" application register mov r8=r13 // return pointer to previously running task
@@ -230,6 +227,9 @@ GLOBAL_ENTRY(ia64_switch_to) mov IA64_KR(CURRENT_STACK)=r26 // remember last page we mapped... ;; itr.d dtr[r25]=r23 // wire in new mapping... + ssm psr.ic // reenable the psr.ic bit + ;; + srlz.d br.cond.sptk .done END(ia64_switch_to)
That sync.i is not old code - it is required.
It is not there to handle kernel fc instructions - as you say, they have their own sync instructions. The sync.i in the context switch path is there to guarantee that user-mode fc's become visible.
The case it's covering is one where some user app is modifying code using the required sequence (as described on the sync instruction page in SDM Volume 3), but in the midst of that sequence, it gets interrupted, context switched, and rescheduled onto some other processor. In that case, unless the OS issues the sync.i on the app's behalf on the original processor, then those fc's need not ever become visible. The sync.i that the app (eventually) does as part of it's sequence isn't good enough, because it executes on the second processor, and syncs aren't broadcast between processors.
Section 4.5.2.1 of SDM 2.1, Volume 2, Part II (page 2:428) talks about this requirement. That section also says that an mf instruction is required in the context switch path.
Now I'm no expert on the linux codebase, so maybe these requirements are being met by some other part of the context-switch path. But if not, then rather than a patch to delete the sync instruction, I think you need one to improve the comments, and add the missing mf.
-- Jim
Itanium Processor Architect at HP
Quoted 43 lines
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Chen, Kenneth W
Sent: Wednesday, September 07, 2005 12:22 AM
To: [email protected]
Subject: [patch] delete sync.i inia64_switch_to()The sync.i instruction in
ia64_switch_to()context switch code looks like an old leftover code. sync.i instruction supposedly ensures flush cache operation issued by processor become visible to memory reference. Everywhere I looked where fc instruction is used (ia64_fc), a sync.i is always accompanied. Everywhere in the vicinity of context switch (save_switch_stack/load_switch_stack) don't have any fc instruction in there. I don't see any reason why we need to issue sync.i in context switch code. Patch to remove it.Signed-off-by Ken Chen <[email protected]>---
./arch/ia64/kernel/entry.S.orig2005-09-06
22:36:17.114182129 -0700
+++./arch/ia64/kernel/entry.S2005-09-07
00:20:37.216644505 -0700arch/ia64/kernel/entry.S renamed from arch/ia64/kernel/entry.S.orig +0
-3@@ -213,10 +213,6 @@ GLOBAL_ENTRY(ia64_switch_to) mov r13=in0 // set "current" pointer ;; DO_LOAD_SWITCH_STACK - -#ifdef CONFIG_SMP - sync.i // ensure "fc"s done bythis CPU are visible on other CPUs
arch/ia64/kernel/entry.S renamed from arch/ia64/kernel/entry.S.orig +0
-2-#endif br.ret.sptk.many rp // boogie on out in new context .map: -To unsubscribe from this list: send the line "unsubscribe linux-ia64" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 1 line · Jim Hull
That sync.i is not old code - it is required.
Yes, the sync.i was added by Asit at some point precisely for the reasons you explained. I believe we concluded at the time that the "mf" wasn't needed since we'd be guaranteed to have a fence in the execution-path anyhow (IIRC, it was due to locks being acquired/released on the paths in question). I'm not sure, however, whether this point has been considered when kernel-preemption support was added. The folks caring about CONFIG_PREEMPT may want to reconsider that issue if not.
--david
Jim Hull wrote on Wednesday, September 07, 2005 10:33 AM
Quoted 5 lines
That sync.i is not old code - it is required.
It is not there to handle kernel fc instructions - as you say, they have their own sync instructions. The sync.i in the context switch path is there to guarantee that user-mode fc's become visible.
Thank you for the explanation. Now I see the lights :-)
- Ken