lkml.org 
[lkml]   [2006]   [Mar]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC, PATCH 3/24] i386 Vmi interface definition
Chris Wright wrote:
> * Zachary Amsden (zach@vmware.com) wrote:
>
>> Master definition of VMI interface, including calls, constants, and
>> interface version.
>>
>
>
>> +/* VROM call table definitions */
>> +#define VROM_CALL_LEN 32
>> +
>> +typedef struct VROMCallEntry {
>> + char f[VROM_CALL_LEN];
>> +} VROMCallEntry;
>>
>
> And the call entry is meant to be handled in whatever mechanism hypervisor
> prefers for its entry points (ABI constraints notwithstanding) as in,
> arbitrary software interrupt, or call gate, etc? I guess for transparent
> it has to, since those would be local calls. Quite similar to the
> hypercall entry point that Xen places on the hypercall_page, so easily
> compatible.

See below.

> The document is slightly more descriptive. The above reserved slots
> are shown as:
>
> char reserved[32];
> char elfHeader[64];
>
> But that's only 3 (0-2). I think I'm missing some small bit of magic.
>
>
>> +typedef struct VROMCallTable {
>> + VROMCallEntry vromCall[128]; // @ 0x80: ROM calls 4-127
>> +} VROMCallTable;
>>
>
> That comment eludes me. Are 0-3 special somehow (IOW, I thought it was
> just 0-2 as per above), and is it suggesting int 0x80?
>

Yeah, most of this is rather crufty - it is in transition. We had a
full blown ROM image with 32-byte aligned stub points at one point for
all of the VMI calls. In fact, it still is in that form, and it was
required to overlap exactly with a native ROM that was built into
Linux. See patch 6, VMI magic fixes for the details.

The machinery is already in place to do this, and it is a very nice
thing that Xen has decided to adopt a similar approach (to the ROM) of
publishing hypervisor code. I think they even use 32-byte alignment as
well. The power of an indirection at this layer is just too attractive,
and once you decide on a single binary image, I think it is inevitable
that everyone will converge on the same idea. The same concept surfaces
over and over - vsyscall being another example. You're basically
dynamically linking in code at runtime, which is a pretty common thing
to do, and gives you a very powerful redirection interface.

But we discovered that we couldn't achieve native performance, even
using directly linked calls into the ROM. Calling out for sti / cli /
popf / pushf on fast paths is just too expensive. We had to inline the
native code to match native performance. And this leaves an opportunity
for flexibility and performance in the hypervisor implementation.

You don't want direct calls to 32-byte stubs; in fact, as Joshua found,
you can't get optimal performance in a hypervisor that way. What you
need is the ability to preferentially inline certain hot calls into the
VMI layer by using NOP padding. The non-hot calls can call out to the
hypercall page.

Publishing code is the first step - inlining is the second, and it gets
you back the hit you took when indirecting your fast paths.

Zach
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-03-14 04:19    [W:0.071 / U:0.460 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site