Re: [PATCH v2] PCI: aardvark: Use LTSSM state to build link training flag

[Bjorn Helgaas]

On Fri, Apr 26, 2019 at 12:27:57AM +0200, Remi Pommarel wrote:
> On Thu, Apr 25, 2019 at 04:04:39PM -0500, Bjorn Helgaas wrote:
> > On Sat, Mar 16, 2019 at 05:12:43PM +0100, Remi Pommarel wrote:

> > It sounds like reading and/or writing some registers during a retrain
> > causes some sort of EL1 error?  Is this a separate erratum?  Is there
> > a list of the registers and operations (read/write) that are affected?
> > The backtrace below suggests that it's actually a read of LNKCAP or
> > LNKCTL (not LNKSTA) that caused the error.
> 
> IIUC, the backtrace below produces an EL1 error when doing a PIO
> transfer while the link is still retraining. See my comment below for
> more about that. But accessing any root complex's register seems fine.
> > 
> > It sounds like there are really two problems:
> > 
> >   1) Reading PCI_EXP_LNKSTA (or the Aardvark equivalent) doesn't give
> >      valid data for PCI_EXP_LNKSTA_LT.
> 
> The 1) is correct.
> 
> >   2) Sometimes config reads cause EL1 errors.
> 
> Actually EL1 error happens when we try to access device's register with
> a PIO transfer, which is when we try to use the link while it is being
> retrained.
> 
> IMHO, 1) and 2) are linked. ASPM core tries to use the link too early
> because it has read invalid data for PCI_EXP_LNKSTA_LT.

>From the software point of view, there is no such thing as "using the
link too early".  The pattern of:

  - Verify that link is up
  - Access device on other end of link

is always racy because the link can go down at any time due to hotplug
or other issues.  In particular, the link can go down after we verify
that the link is up, but before we access the device.

Software must be able to deal with that gracefully.  I don't know
whether that means catching and recovering from that EL1 error, or
masking it, or what.  This is architecture-specific stuff that's
outside the scope of PCIe itself.

But a link going down should never directly cause a kernel panic.

> > > This fixes boot hang or kernel panic with the following callstack due to
> > > ASPM setup doing a link re-train and polling for PCI_EXP_LNKSTA_LT flag
> > > to be cleared before using it.
> > > 
> > > -------------------- 8< -------------------
> > > 	[    0.915389]  dump_backtrace+0x0/0x140
> > > 	[    0.915391]  show_stack+0x14/0x20
> > > 	[    0.915393]  dump_stack+0x90/0xb4
> > > 	[    0.915394]  panic+0x134/0x2c0
> > > 	[    0.915396]  nmi_panic+0x6c/0x70
> > > 	[    0.915398]  arm64_serror_panic+0x74/0x80
> > > 	[    0.915400]  is_valid_bugaddr+0x0/0x8
> > > 	[    0.915402]  el1_error+0x7c/0xe4
> > > 	[    0.915404]  advk_pcie_rd_conf+0x4c/0x250
> > > 	[    0.915406]  pci_bus_read_config_word+0x7c/0xd0
> > > 	[    0.915408]  pcie_capability_read_word+0x90/0xc8
> > > 	[    0.915410]  pcie_get_aspm_reg+0x68/0x118
> > > 	[    0.915412]  pcie_aspm_init_link_state+0x460/0xa98

> > > +	case PCI_EXP_LNKCTL: {
> > 
> > Don't you mean PCI_EXP_LNKSTA here?
> 
> No, PCI_EXP_LNKSTA and PCI_EXP_LNKCTL are consecutive 16bit registers
> but bridge emulation accesses those registers by 32bit chunk. So when
> one wants to read PCI_EXP_LNKSTA register, pci bridge reads 32bit data
> from PCI_EXP_LNKCTL and 16 bit shift the result to the right.
> 
> This is why I use (PCI_EXP_LNKSTA_LT << 16) below.

Ah, that makes sense.  A comment along the lines of "u32 contains both
PCI_EXP_LNKCTL and PCI_EXP_LNKSTA" would be a nice hint.

Bjorn