[Xen-ia64-devel] [PATCH 0/6] Add full evtchn mechanism for xen/ia64

["Tian, Kevin" <kevin.tian@xxxxxxxxx>]

Hi, all,
        This patch sets will port full event channel mechanism to
xen/ia64, 
based on which all physical irqs/virtual irqs/ipis are now bound to a 
specific event port. From now on, a typical /proc/interrupts of dom0
will 
look like:
           CPU0
 34:         12        Phys-irq  ide0
 39:          0        Phys-irq  acpi
 45:        322        Phys-irq  serial
 48:     115006        Phys-irq  peth0
 49:      16269        Phys-irq  ioc0
 50:         31        Phys-irq  ioc1
 51:          2        Phys-irq  ehci_hcd:usb1
 52:          0        Phys-irq  uhci_hcd:usb2
 53:         55        Phys-irq  uhci_hcd:usb3
256:          0     Dynamic-irq  RESCHED0
257:          0     Dynamic-irq  IPI0
258:      44572     Dynamic-irq  timer0
259:       2316     Dynamic-irq  xenbus
260:       8304     Dynamic-irq  blkif-backend
261:      25947     Dynamic-irq  vif3.0
ERR:          0

Then for SMP domU:
           CPU0       CPU1
256:       1417          0     Dynamic-irq  RESCHED0
257:         40          0     Dynamic-irq  IPI0
258:       4937          0     Dynamic-irq  timer0
259:          0       1691     Dynamic-irq  RESCHED1
260:          0        165     Dynamic-irq  IPI1
261:          0       4953     Dynamic-irq  timer1
262:        220          0     Dynamic-irq  xenbus
263:        189          0     Dynamic-irq  xencons
264:       3493          0     Dynamic-irq  blkif
265:        128          0     Dynamic-irq  eth0

        This patch set is tested upon current tip (Rev10021 "[IA64]
pte_xchg 
added"), including:
Non-vp + UP domU + domVTI
Non-vp + SMP domU + domVTI
VP + UP domU + domVTI
VP + SMP domU + domVTI

        I also tested network performance by wget with several combos:
(The result is the average of three tests)
                             TIP             With patches
Dom0 before 'xend start'       11.21Mb         11.21Mb
Dom0 after 'xend start'         11.21Mb         11.21Mb
Single domU                  5.49Mb          5.63Mb
Single dom0 (domU is up)      11.21Mb         11.21Mb
Both domU + dom0            9.62Mb/1.28Mb   9.63Mb/1.45Mb

        Based on above result, dom0's net performance is not affected, 
while domU's is seen with several percents increase. Actually even 
dom0 should be better because I observed the ITC cycle for single 
eth0 irq handling decreased by 35%. But because previously dom0's 
wget is already equal to native, so we don't observe difference on the 
result. 
        That means this patch can even help performance a bit at the
same 
time when it mainly targets to add new feature to support driver 
domain. :-)

Following is the patch set briefs:
[PATCH 1/6] 0_pull_header_files.patch
[PATCH 2/6] 1_add_callback_ops.patch
[PATCH 3/6] 2_add_evt_handle_path.patch
[PATCH 4/6] 3_clean_pirq_bind_logic.patch
Above 4 patches only serve as prep steps with added logics functionally 
disabled because no place uses them. 
(I've also tested with tip to find nothing broken with them.)

[PATCH 5/6] evtchn_common.patch
This patch modifies common code and I've sent to xen-devel which has 
been checked into xen-unstable.hg now. So Alex, could you do a favor 
to make a sync with xen-unstable.hg first? :-)

[PATCH 6/6] 4_final_evtchn_support.patch
This patch is the very one to move xen/ia64 to full evtchn mechanism, 
and once after applying it, you could observe above example output.

Cheers,
Kevin

_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel