Message-ID: <20160203234026.GS20964@wotan.suse.de>
Date:	Thu, 4 Feb 2016 00:40:26 +0100
From:	"Luis R. Rodriguez" <mcgrof@...e.com>
To:	Boris Ostrovsky <boris.ostrovsky@...cle.com>
Cc:	"Luis R. Rodriguez" <mcgrof@...nel.org>,
	David Vrabel <david.vrabel@...rix.com>, konrad.wilk@...cle.com,
	xen-devel@...ts.xenproject.org, linux-kernel@...r.kernel.org,
	roger.pau@...rix.com, x86@...nel.org, GLin@...e.coma,
	bblanco@...mgrid.com, pmonclus@...mgrid.com, bp@...e.de,
	hpa@...or.com
Subject: Re: [PATCH v2 02/11] xen/hvmlite: Bootstrap HVMlite guest

On Wed, Feb 03, 2016 at 03:11:56PM -0500, Boris Ostrovsky wrote:
> On 02/03/2016 01:55 PM, Luis R. Rodriguez wrote:
> >I saw no considerations for the recommendations I had made last on your v1:
> >
> >https://lkml.kernel.org/r/CAB=NE6XPA0YzbnM8=rspkKai6d3GkXXO00Gr0VZUYoyzNy6thw@mail.gmail.com
> >
> >Of importance:
> >
> >1) Using pv_info.paravirt_enabled = 1 is wrong unless you mean to say this
> >    is for legacy x86:
> >
> >Your patch #3 keeps on setting pv_info.paravirt_enabled = 1 and as discussed
> >this is wrong. It will be renamed to x86_legacy_free() to align with what folks
> >are pushing for a BIOS flag to annotate if a system requires legacy x86 stuff.
> >This also means re-thinking all use cases and ensuring subarch is used then
> >instead when the goal was to avoid Xen from entering that code. Today Xen does
> >not use this but with my work it does and it helps clean and brush up a lot of
> >these checks with future prospects to even help unify entry points.
> 
> As I said earlier, I am not sure I understand what subarch buys us
> for HVMlite guests.

I accepted subarch may not be the right thing, so proposed a hypervisor type.
What it buys you is a strong semantics association between code designed
for a purpose.

> As for using paravirt_enabled -- this is really only used to
> differentiate HVM from HVMlite and I think (although I'd need to
> check) is only needed by Xen-specific code in a couple of places.

That sounds like a Xen specific use case as such an interface that is
pointed out as going to renamed to reflect its actual use case should not
be abused for that purpose.

> So if/when it is removed we will switch to something else. Since your work is
> WIP I decided to keep using it until it's clear what other options may be
> available.

And your work is not WIP? I'll be splitting my patches up and the rename
will be atomic, it likely can go in first than yours, so not sure why you
are simply brushing this off.

> >2) We should avoid more hypervisor type hacks, and just consider a new
> >    hypervisor type to close the gap:
> >
> >Using x86_legacy_free() and friends in a unified way for all systems means it
> >should only be used after init_hypervisor_platform() which is called during
> >setup_arch().  This means we have a semantic gap for checks on "are we on
> >hypervisor type and which one?".
> 
> In this particular case we don't need any information about
> hypervisor until init_hypervisor_platform().

I pointed out in your v1 patchset how microcode loading was not blocked, you
then asked how KVM does it, and that was explained as well, and that they
don't enable it as well. You need a solution for this.

> >There are drivers now using these sorts of
> >checks as well, for instance snd_intel8x0_inside_vm(). We should avoid having
> >these hacks but that also means cleaning up a well define grammar here for what
> >we want.  I'm doing work to help with this by streamlining use of the subarch
> >type, that should help with PV code, but your use case seems different but yet
> >related, what I had suggested last was to consider we add a new hypervisor type
> >to the x86 boot protocol which would be available early on. This would have a
> >few purposes, one of which deserves its own section below on dead code:
> >
> >     a) clean up hacks as with snd_intel8x0_inside_vm()
> >     b) enable a generic way and clean way to distinguish what hypervisor
> >        type you're on
> >     c) since it would be set early and if we can ensure its accessible
> >        early on boot it would mean avoiding having to add yet-another
> >        asm entry point for Linux, you could just use startup_32() and
> >        the hypervisor type could easily just have an early branch call
> >        and post branch call very similar to how we deal with the subarch
> >        currently on 32-bit. Your calls then just become early stubs and
> >        we'd have a solution for other PV types that want a similar solution
> >        later
> 
> Which calls?
> 
> If you are referring to xen_prepare_hvmlite/hvmlite_bootparams

Even before, hvmlite_start_xen(). 

> then these are needed to prepare boot_params. And we should not enter
> startup_32() without them ready.

As-is the x86 boot protocol would not allow an easy way for this, I'm
suggesting we consider extending the boot protocol to add a hypervisor
type and data pointer much as with subarch and subarch_data for the
particular purpose of both enabling entry into the same startup_32()
but also a clean way for modifications of stubs both at the beginning
and at the end of startup_32().

Pseudo code:

startup_32()                         startup_64()
       |                                  |
       |                                  |
       V                                  V
pre_hypervisor_stub_32()	pre_hypervisor_stub_64()
       |                                  |
       |                                  |
       V                                  V
 [existing startup_32()]       [existing startup_64()]
       |                                  |
       |                                  |
       V                                  V
post_hypervisor_stub_32()	post_hypervisor_stub_64()

The pre_hypervisor_stub_32() would have much of the code in
hvmlite_start_xen() but for 32-bit, pre_hypervisor_stub_64()
would have the 64-bits.

> >3) Dead code concerns and unifying entry points:
> >
> >Addressing the semantics for the gray areas I am highlighting are critical for
> >ensuring one does not run code or even exposes code as a available for the type
> >of run time system booted, some folks call this "dead code". This is critical
> >for Linux distributions which need to rely on the flexibility of having one
> >kernel work for different use cases. The resolution to this problem was pvops
> >but pvops has shortcoming for dead code, it didn't address the problem likely as
> >it was not considered serious. It also didn't address the issue of different
> >hypervisors wanting different entry points and that this fact alone also contributes
> >to more dead code concerns, case in point the regressions introduced by cr4 shadow
> >and the latest one is Kasan which to this day breaks Xen! Dead code topics are
> >not easy to grasp, its why I've started on my own crusade to talk to people and
> >write about it [0], and as of late propose some changes to avoid these in a
> >clean way without extending pvops. Adding yet another entry point will not help
> >here *specially* if we do not take semantics seriously over the different hypervisors
> >and hypervisor types.
> >
> >[0] http://www.do-not-panic.com/2015/12/avoiding-dead-code-pvops-not-silver-bullet.html
> >[1] http://www.do-not-panic.com/2015/12/xen-and-x86-linux-zero-page.html
> 
> I don't understand what this has to do with HVMlite guests.

We have no formal semantics to ensure correctness and avoid dead code for HVMLite,
as with PV. The microcode loader is one example of code that should not run. Sure,
an adhoc solution might be possible, but I'm advocating for a generic solution.

> >My recommendation then:
> >
> >We add new hypervisor type to close the semantic gap for hypervisor types, and
> >much like subarch enable also a subarch_data to let you pass and use your
> >hvmlite_start_info. This would not only help with the semantics but also help
> >avoid yet-another-entry point and force us to provide a well define structure
> >for considering code that should not run by pegging it as required or supported
> >for different early x86 code stubs.
> 
> As I said before, I don't see how we can avoid having another entry
> point without making Xen change its load procedure. Which is highly
> unlikely to happen.

Why would it not be possible if this is about a new guest type? Surely there
are good reasons to consider extending the load procedure protocol?

> >I had hinted perhaps we might be able to piggy back on top of the ELF loader
> >protocol as well, and since that's standard do wonder if that could instead
> >be extended to help unify a mechanism for different OSes instead of making
> >this just a solution for Linux.

If it was possible to extend it, would this be a reasonable venue to consider?

> >Code review below.
> >
> >On Mon, Feb 01, 2016 at 10:38:48AM -0500, Boris Ostrovsky wrote:
> >>Start HVMlite guest at XEN_ELFNOTE_PHYS32_ENTRY address. Setup hypercall
> >>page, initialize boot_params, enable early page tables.
> >>
> >>Since this stub is executed before kernel entry point we cannot use
> >>variables in .bss which is cleared by kernel. We explicitly place
> >>variables that are initialized here into .data.
> >>
> >>Signed-off-by: Boris Ostrovsky <boris.ostrovsky@...cle.com>
> >>---
> >>diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
> >>index 5774800..5f05fa2 100644
> >>--- a/arch/x86/xen/enlighten.c
> >>+++ b/arch/x86/xen/enlighten.c
> >>@@ -171,6 +172,17 @@ struct tls_descs {
> >>   */
> >>  static DEFINE_PER_CPU(struct tls_descs, shadow_tls_desc);
> >>+#ifdef CONFIG_XEN_PVHVM
> >>+/*
> >>+ * HVMlite variables. These need to live in data segment since they are
> >>+ * initialized before startup_{32|64}, which clear .bss, are invoked.
> >>+ */
> >>+int xen_hvmlite __attribute__((section(".data"))) = 0;
> >>+struct hvm_start_info hvmlite_start_info __attribute__((section(".data")));
> >>+uint hvmlite_start_info_sz = sizeof(hvmlite_start_info);
> >>+struct boot_params xen_hvmlite_boot_params __attribute__((section(".data")));
> >>+#endif
> >>+
> >The section annotations seems very special use case but likely worth documenting
> >and defining a new macro for in include/linux/compiler.h. This would make it
> >easier to change should we want to change the section used here later and
> >enable others to easily look for the reason for these annotations in a
> >single place.
> 
> I wonder whether __initdata would be a good attribute. We only need
> this early in the boot.

I could not find other users of .data other than some specific driver.
Using anything with *init* alludes you can free the data later but if we
want to keep it I suggest a different prefix, up to you.

  Luis

Powered by blists - more mailing lists