| Message ID | 20191001091214.3287-1-paul.durrant@citrix.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [PATCH-for-4.13] libxl: choose an appropriate default for passthrough... | expand |
On 01.10.19 11:12, Paul Durrant wrote: > ...if there is no IOMMU or it is globally disabled. > > Without this patch, the following assertion may be hit: > > xl: libxl_create.c:589: libxl__domain_make: Assertion `info->passthrough != LIBXL_PASSTHROUGH_ENABLED' failed. > > This is because libxl__domain_create_info_setdefault() currently only sets > an appropriate value for 'passthrough' in the case that 'cap_hvm_directio' > is true, which is not the case unless an IOMMU is present and enabled in > the system. This is normally masked by xl choosing a default value, but > that will not happen if xl is not used (e.g. when using libvirt) or when > a stub domain is being created. > > Signed-off-by: Paul Durrant <paul.durrant@citrix.com> Release-acked-by: Juergen Gross <jgross@suse.com> Juergen
Paul Durrant writes ("[PATCH-for-4.13] libxl: choose an appropriate default for passthrough..."):
> ...if there is no IOMMU or it is globally disabled.
>
> Without this patch, the following assertion may be hit:
>
> xl: libxl_create.c:589: libxl__domain_make: Assertion `info->passthrough != LIBXL_PASSTHROUGH_ENABLED' failed.
>
> This is because libxl__domain_create_info_setdefault() currently only sets
> an appropriate value for 'passthrough' in the case that 'cap_hvm_directio'
> is true, which is not the case unless an IOMMU is present and enabled in
> the system. This is normally masked by xl choosing a default value, but
> that will not happen if xl is not used (e.g. when using libvirt) or when
> a stub domain is being created.
It's weird that after this patch "enabled" can mean DISABLED. Surely
if you say `passthrough="enabled"' and the host has no PT support (eg
it's disabled in the bios) it should fail ?
Normally libxl config options have an "unknown" or "default" option.
Also it is anomalous that xl is doing the complex work of choosing a
default. I think all the complex code
+ switch (c_info->passthrough) {
+ case LIBXL_PASSTHROUGH_ENABLED:
in xl_parse.c should be in libxl. (Some of it is there already.)
I'm sorry that I wasn't didn't review babde47a3fed...
Ian.
> -----Original Message----- > From: Ian Jackson <ian.jackson@citrix.com> > Sent: 01 October 2019 11:39 > To: Paul Durrant <Paul.Durrant@citrix.com> > Cc: xen-devel@lists.xenproject.org; Wei Liu <wl@xen.org>; Anthony Perard <anthony.perard@citrix.com>; > Juergen Gross <jgross@suse.com> > Subject: Re: [PATCH-for-4.13] libxl: choose an appropriate default for passthrough... > > Paul Durrant writes ("[PATCH-for-4.13] libxl: choose an appropriate default for passthrough..."): > > ...if there is no IOMMU or it is globally disabled. > > > > Without this patch, the following assertion may be hit: > > > > xl: libxl_create.c:589: libxl__domain_make: Assertion `info->passthrough != > LIBXL_PASSTHROUGH_ENABLED' failed. > > > > This is because libxl__domain_create_info_setdefault() currently only sets > > an appropriate value for 'passthrough' in the case that 'cap_hvm_directio' > > is true, which is not the case unless an IOMMU is present and enabled in > > the system. This is normally masked by xl choosing a default value, but > > that will not happen if xl is not used (e.g. when using libvirt) or when > > a stub domain is being created. > > It's weird that after this patch "enabled" can mean DISABLED. Surely > if you say `passthrough="enabled"' and the host has no PT support (eg > it's disabled in the bios) it should fail ? Indeed, and xl will do exactly that. > > Normally libxl config options have an "unknown" or "default" option. > > Also it is anomalous that xl is doing the complex work of choosing a > default. I think all the complex code > > + switch (c_info->passthrough) { > + case LIBXL_PASSTHROUGH_ENABLED: > > in xl_parse.c should be in libxl. (Some of it is there already.) > > I'm sorry that I wasn't didn't review babde47a3fed... > So, would you advocate an 'unknown' value then? That's essentially just a rename operation on 'enabled'. The code in xl_parse.c is there for a reason though; the appropriate amount of extra memory for the IOMMU page tables has to be determined there because the 'build' parts of libxl seem to be largely firewalled from the 'create' parts and thus the relevant information is not available to decide the appropriate overhead. Paul > Ian.
Ping? Can I get a response on this (w.r.t. 'enabled' vs. 'unknown') before doing a v2? This issue is currently blocking a push, I believe. On Tue, 1 Oct 2019 at 11:48, Paul Durrant <Paul.Durrant@citrix.com> wrote: > > > -----Original Message----- > > From: Ian Jackson <ian.jackson@citrix.com> > > Sent: 01 October 2019 11:39 > > To: Paul Durrant <Paul.Durrant@citrix.com> > > Cc: xen-devel@lists.xenproject.org; Wei Liu <wl@xen.org>; Anthony Perard <anthony.perard@citrix.com>; > > Juergen Gross <jgross@suse.com> > > Subject: Re: [PATCH-for-4.13] libxl: choose an appropriate default for passthrough... > > > > Paul Durrant writes ("[PATCH-for-4.13] libxl: choose an appropriate default for passthrough..."): > > > ...if there is no IOMMU or it is globally disabled. > > > > > > Without this patch, the following assertion may be hit: > > > > > > xl: libxl_create.c:589: libxl__domain_make: Assertion `info->passthrough != > > LIBXL_PASSTHROUGH_ENABLED' failed. > > > > > > This is because libxl__domain_create_info_setdefault() currently only sets > > > an appropriate value for 'passthrough' in the case that 'cap_hvm_directio' > > > is true, which is not the case unless an IOMMU is present and enabled in > > > the system. This is normally masked by xl choosing a default value, but > > > that will not happen if xl is not used (e.g. when using libvirt) or when > > > a stub domain is being created. > > > > It's weird that after this patch "enabled" can mean DISABLED. Surely > > if you say `passthrough="enabled"' and the host has no PT support (eg > > it's disabled in the bios) it should fail ? > > Indeed, and xl will do exactly that. > > > > > Normally libxl config options have an "unknown" or "default" option. > > > > Also it is anomalous that xl is doing the complex work of choosing a > > default. I think all the complex code > > > > + switch (c_info->passthrough) { > > + case LIBXL_PASSTHROUGH_ENABLED: > > > > in xl_parse.c should be in libxl. (Some of it is there already.) > > > > I'm sorry that I wasn't didn't review babde47a3fed... > > > > So, would you advocate an 'unknown' value then? That's essentially just a rename operation on 'enabled'. > > The code in xl_parse.c is there for a reason though; the appropriate amount of extra memory for the IOMMU page tables has to be determined there because the 'build' parts of libxl seem to be largely firewalled from the 'create' parts and thus the relevant information is not available to decide the appropriate overhead. > > Paul > > > Ian. > > _______________________________________________ > Xen-devel mailing list > Xen-devel@lists.xenproject.org > https://lists.xenproject.org/mailman/listinfo/xen-devel
Paul Durrant writes ("Re: [Xen-devel] [PATCH-for-4.13] libxl: choose an appropriate default for passthrough..."):
> Ping? Can I get a response on this (w.r.t. 'enabled' vs. 'unknown')
> before doing a v2? This issue is currently blocking a push, I believe.
I definitely think we should introduce "unknown".
> > So, would you advocate an 'unknown' value then? That's essentially just a rename operation on 'enabled'.
I think we probably want "enabled" as well but that can wait. If for
now you rename "unknown" that will do.
> > The code in xl_parse.c is there for a reason though; the appropriate amount of extra memory for the IOMMU page tables has to be determined there because the 'build' parts of libxl seem to be largely firewalled from the 'create' parts and thus the relevant information is not available to decide the appropriate overhead.
I want to look into this more deeply.
Thanks,
Ian.
Ian Jackson writes ("Re: [PATCH-for-4.13 v2 0/2] libxl: fix assertion failure"):
> I am continuing to look at the defaulting and config management here
> with a view to getting rid of some of the duplicated code and moving
> it all into libxl.
Specifically, Paul Durrant writes:
> Ok. Specifically libxl_domain_need_memory() only has access to the
> build info, but libxl__domain_build_info_setdefault() does not have
> access to the create info, so it can't choose a sensible default. To
> avoid re-writing lots of code I went with having xl calulate a
> default. (Prior to my patch there was no calculated overhead anyway so
> a host without shared EPT and autoballooning on was already playing
> russian roulette).
So some restructuring may be needed. I'll see how invasive it looks.
Ian.
diff --git a/tools/libxl/libxl_create.c b/tools/libxl/libxl_create.c index b58e030376..3bdb6c51b6 100644 --- a/tools/libxl/libxl_create.c +++ b/tools/libxl/libxl_create.c @@ -68,7 +68,11 @@ int libxl__domain_create_info_setdefault(libxl__gc *gc, c_info->passthrough = ((c_info->type == LIBXL_DOMAIN_TYPE_PV) || !info.cap_iommu_hap_pt_share) ? LIBXL_PASSTHROUGH_SYNC_PT : LIBXL_PASSTHROUGH_SHARE_PT; - } + } else if (!info.cap_hvm_directio) + c_info->passthrough = LIBXL_PASSTHROUGH_DISABLED; + + /* An explicit setting should now have been chosen */ + assert(c_info->passthrough != LIBXL_PASSTHROUGH_ENABLED); return 0; }
...if there is no IOMMU or it is globally disabled. Without this patch, the following assertion may be hit: xl: libxl_create.c:589: libxl__domain_make: Assertion `info->passthrough != LIBXL_PASSTHROUGH_ENABLED' failed. This is because libxl__domain_create_info_setdefault() currently only sets an appropriate value for 'passthrough' in the case that 'cap_hvm_directio' is true, which is not the case unless an IOMMU is present and enabled in the system. This is normally masked by xl choosing a default value, but that will not happen if xl is not used (e.g. when using libvirt) or when a stub domain is being created. Signed-off-by: Paul Durrant <paul.durrant@citrix.com> --- Cc: Ian Jackson <ian.jackson@eu.citrix.com> Cc: Wei Liu <wl@xen.org> Cc: Anthony PERARD <anthony.perard@citrix.com> Cc: Juergen Gross <jgross@suse.com> --- tools/libxl/libxl_create.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)