BUG: libxl vuart build order

[Stefano Stabellini <stefano.stabellini@xxxxxxxxxx>]

+ xen-devel and libxl maintainers

In short, there is a regression in libxl with the ARM vuart introduced
by moving ARM guests to the PVH build.


On Thu, 29 Oct 2020, Takahiro Akashi wrote:
> On Wed, Oct 28, 2020 at 02:44:16PM -0700, Stefano Stabellini wrote:
> > On Wed, 28 Oct 2020, Takahiro Akashi wrote:
> > > On Tue, Oct 27, 2020 at 09:02:14AM +0900, Takahiro Akashi wrote:
> > > > On Mon, Oct 26, 2020 at 04:37:30PM -0700, Stefano Stabellini wrote:
> > > > > 
> > > > > On Mon, 26 Oct 2020, Takahiro Akashi wrote:
> > > > > > Stefano,
> > > > > > 
> > > > > > # I'm afraid that I have already bothered you with a lot of 
> > > > > > questions.
> > > > > > 
> > > > > > When I looked at Xen's vpl011 implementation, I found
> > > > > > CR (and LCHR) register is not supported. (trap may cause a data 
> > > > > > abort).
> > > > > > On the other hand, for example, linux's pl011 driver surely
> > > > > > accesses CR (and LCHR) register.
> > > > > > So I guess that linux won't be able to use pl011 on a Xen guest vm
> > > > > > if vuart = "sbsa_uart".
> > > > > > 
> > > > > > Is this a known issue or do I miss anything?
> > > > > 
> > > > > Linux should definitely be able to use it, and in fact, I am using it
> > > > > with Linux in my test environment.
> > > > > 
> > > > > I think the confusion comes from the name "vpl011": it is in fact not 
> > > > > a
> > > > > full PL011 UART, but an SBSA UART.
> > > > 
> > > > Yeah, I have noticed it.
> > > > 
> > > > > SBSA UART only implements a subset of
> > > > > the PL011 registers. The compatible string is "arm,sbsa-uart", also 
> > > > > see
> > > > > drivers/tty/serial/amba-pl011.c:sbsa_uart_probe.
> > > > 
> > > > Looking closely into the details of implementation, I found
> > > > that all the accesses to unimplemented registers, including
> > > > CR, are deliberately avoided in sbsa part of linux driver.
> > > 
> > > So I'm now trying to implement "sbsa-uart" driver on U-Boot
> > > by modifying the existing pl011 driver.
> > > (Please note the current xen'ized U-Boot utilises a para-virtualized
> > > console, i.e. with HVM_PARAM_CONSOLE_PFN.)
> > > 
> > > So far my all attempts have failed.
> > > 
> > > There are a couple of problems, and one of them is how we can
> > > access vpl011 port (from dom0).
> > > What I did is:
> > > - modify U-Boot's pl011 driver
> > >   (I'm sure that the driver correctly handle a vpl011 device
> > >   with regard of accessing a proper set of registers.)
> > > - start U-Boot guest with "vuart=sbsa_uart" by
> > >     xl create uboot.cfg -c
> > > 
> > > Then I have seen almost nothing on the screen.
> > > Digging into vpl011 implementation, I found that all the characters
> > > written DR register will be directed to a "backend domain" if a guest
> > > vm is launched by xl command.
> > > (In case of dom0less, the backend seems to be Xen itself.)
> > > 
> > > As a silly experiment, I modified domain_vpl011_init() to always create
> > > a vpl011 device with "backend_in_domain == false".
> > > Then, I could see more boot messages from U-Boot, but still fails
> > > to use the device as a console, I mean, we will lose all the outputs
> > > after at some point and won't be able to type any keys (at a command 
> > > prompt).
> > > (This will be another problem on U-Boot side.)
> > > 
> > > My first question here is how we can configure and connect a console
> > > in this case?
> > > Should "xl create -c" or "xl console -t vuart" simply work?
> > 
> > "xl create -c" creates a guest and connect to the primary console which
> > is the PV console (i.e. HVM_PARAM_CONSOLE_PFN.)
> 
> So in case of vuart, it (console) doesn't work?
> (Apparently, "xl create" doesn't take '-t' option.)
> 
> > To connect to the emulated sbsa uart you need to pass -t vuart. So yes,
> > "xl console -t vuart domain_name" should get you access to the emulated
> > sbsa uart. The sbsa uart can also be exposed to dom0less guests; you get
> > their output by using CTRL-AAA to switch to right domU console.
> > 
> > You can add printks to xen/arch/arm/vpl011.c in Xen to see what's
> > happening on the Xen side. vpl011.c is the emulator.
> 
> I'm sure that write to "REG_DR" register is caught by Xen.
> What I don't understand is
> if back_in_domain -> no outputs
> if !back_in_domain -> can see outputs
> 
> (As you know, if a guest is created by xl command, back_in_domain
> is forcedly set to true.)
> 
> I looked into xenstore and found that "vuart/0/tty" does not exist,
> but "console/tty" does.
> How can this happen for vuart?
> (I clearly specified, vuart = "sbsa_uart" in Xen config.)

It looks like we have a bug :-(

I managed to reproduce the issue. The problem is a race in libxl.

tools/libxc/xc_dom_arm.c:alloc_magic_pages is called first, setting
dom->vuart_gfn.  Then, libxl__build_hvm should be setting
state->vuart_gfn to dom->vuart_gfn (like libxl__build_pv does) but it
doesn't.


---

libxl: set vuart_gfn in libxl__build_hvm

Setting vuart_gfn was missed when switching ARM guests to the PVH build.
Like libxl__build_pv, libxl__build_hvm should set state->vuart_gfn to
dom->vuart_gfn.

Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxx>

diff --git a/tools/libxl/libxl_dom.c b/tools/libxl/libxl_dom.c
index f8661e90d4..36fe8915e7 100644
--- a/tools/libxl/libxl_dom.c
+++ b/tools/libxl/libxl_dom.c
@@ -1184,6 +1184,7 @@ int libxl__build_hvm(libxl__gc *gc, uint32_t domid,
         LOG(ERROR, "hvm build set params failed");
         goto out;
     }
+    state->vuart_gfn = dom->vuart_gfn;
 
     rc = hvm_build_set_xs_values(gc, domid, dom, info);
     if (rc != 0) {