On Thu, Sep 28, 2017 at 09:57:19PM -0500, Bjorn Helgaas wrote:
> On Thu, Sep 28, 2017 at 04:19:12PM +0200, Thierry Reding wrote:
> > On Mon, Sep 25, 2017 at 10:44:29AM +0530, vidya sagar wrote:
> > > Inline...
> > >
> > > On Sat, Sep 23, 2017 at 11:47 AM, Thierry Reding
> > > <thierry.reding@xxxxxxxxx> wrote:
> > > >
> > > > The Tegra20 PCIe controller has a different address range for MSI, so
> > > > select a different target address.
> > > >
> > > > Fixes: d7bd554f27c9 ("PCI: tegra: Do not allocate MSI target memory")
> > > > Signed-off-by: Thierry Reding <treding@xxxxxxxxxx>
> > > > ---
> > > > drivers/pci/host/pci-tegra.c | 12 +++++++++++-
> > > > 1 file changed, 11 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/pci/host/pci-tegra.c b/drivers/pci/host/pci-tegra.c
> > > > index e8e1ddbaabc9..5b02ea59524b 100644
> > > > --- a/drivers/pci/host/pci-tegra.c
> > > > +++ b/drivers/pci/host/pci-tegra.c
> > > > @@ -1563,8 +1563,18 @@ static int tegra_pcie_enable_msi(struct tegra_pcie *pcie)
> > > > * none of the Tegra SoCs that contain this PCI host bridge can
> > > > * address more than 16 GiB of system memory, the last 4 KiB of
> > > > * these 1012 GiB is a good candidate.
> > > > + *
> > > > + * Unfortunately, Tegra20 is slightly different in that the physical
> > > > + * address for this MSI region is limited to the lower 32 bits of the
> > > > + * address map, so the address that we pick is going to have to be
> > > > + * located somewhere within the region addressable by the CPU and
> > > > + * on-SoC controllers. To be on the safe side, we select an address
> > > > + * from a region that is marked unused (0xf0010000 - 0xfffeffff).
> > > > */
> > > > - msi->phys = 0xfcfffff000;
> > > > + if (soc->msi_base_shift > 0)
> > > > + msi->phys = 0xfcfffff000;
> > > > + else
> > > > + msi->phys = 0x00f0010000;
> > >
> > > Can we use this for later Tegra chip versions as well?
> > > Reason being, if an end point's config space says that it cannot
> > > support >32-bit MSI addresses (Marvel SATA controller being one
> > > example), with the current code, we still allocate an address of
> > > >32-bits, resulting in end point not being able to generate MSI
> > > interrupt (as it discards >32-bits where generating upstream memory
> > > write transaction for MSI)
> >
> > Looks like the universal fix for this may not be trivial, so perhaps in
> > the meantime this one could go in in order to restore MSI for TrimSlice
> > (and other Tegra20 boards)?
> >
> > Bjorn: do you have any objections to taking this into v4.13 via a stable
> > backport? 32-bit only MSI were never guaranteed to work because prior to
> > this patch the address was taken from an allocated page, which may or
> > may not have been > 4 GiB.
> >
> > We could then go and fix 32-bit MSI hopefully in time for 4.14, but it'd
> > be a shame if v4.13 keeps being broken just because we're trying to
> > settle on the right solution for 32-bit MSI.
>
> I don't know your hardware, so I don't have any suggestions about your
> strategy for choosing addresses.
>
> To me it looks slightly obscure to use "soc->msi_base_shift == 0" to
> identify SoCs that can only use 32-bit MSI, but maybe that makes sense
> if you know the hardware.
I should probably clarify this change. msi_base_shift == 0 applies only
to Tegra20. This is because the implementation of the PCI host bridge on
Tegra20 has a bug that doesn't actually shift the address written to the
AFI_MSI_FPCI_BAR_ST by 8 bits to the left for matching. This was fixed
in Tegra30 (and later), hence why we have this special case in the
driver. It works around the issue on Tegra20.
A side-effect of that is that MSIs on Tegra20 can only go to the lower
32 bits of the physical address space, whereas later chips can address
the full 40 bits.
As Vidya points out, using any MSI target address > 4 GiB causes devices
with 32-bit MSI only capability to not work because they will truncate
the address and therefore the MSI controller won't be able to match at
any time.
So a proper fix is to select an MSI target address < 4 GiB for all the
chips, which would fix the regression *and* enable 32-bit MSI. However,
currently the only known regression is on TrimSlice (Tegra20), where a
change in MSI target address will fix the regression.
For all other SoCs, commit d7bd554f27c9 ("PCI: tegra: Do not allocate
MSI target memory") technically regresses for 32-bit MSI only capable
endpoints. However, since before that commit the MSI target address was
derived from an allocated page, 32-bit MSI were not guaranteed to work
either. On any system with > 2 GiB of RAM the MSI target page is likely
to be located in a region that 32-bit MSI could not address.
> The stable backport policy is that the fix must first exist in Linus'
> tree. The normal path is for fixes to be merged for v4.15, so they
> wouldn't appear in Linus' tree until after v4.14 releases.
>
> If I merge a fix via for-linus, it could be merged to Linus' tree
> before v4.14 and could be backported to v4.13 before v4.14. But we'd
> have to make a case for doing that, e.g., this fixes a regression or
> something we broke during the v4.14 merge window, or it's some other
> serious bug fix.
The offending commit d7bd554f27c9 ("PCI: tegra: Do not allocate MSI
target memory") was introduced in v4.13 and it causes a regression on
Tegra20, in particular the TrimSlice board. Other devices could be
affected, but the TrimSlice is the only one I'm aware of that people
still care about and use.
I think this meets all the criteria for merging via for-linus. I'd like
to see this fixed in v4.14 and backported to v4.13. I think we have two
options at this point: one is to apply this fix, which addresses the
TrimSlice regression but doesn't fix a potential regression on Tegra30
and later (I'm not aware of any being reported, though). I have a local
patch (attached), which should also fix that. An alternative would be to
revert the offending commit (it reverts cleanly on linux-next). In both
cases the fix/revert should be backported to v4.13.
Vidya, Jon, Mikko: can you test the attached patch? It works for me on
Tegra20 (TrimSlice), Tegra30 (Beaver), Tegra124 (Jetson TK1) and
Tegra210 (Jetson TX1, with a Realtek NIC attached). It'd be good to get
some broader coverage, though.
Thierry
--- >8 ---
From 36108b8facc443170f91d64bb8d6bed79f5de0ba Mon Sep 17 00:00:00 2001
From: Thierry Reding <treding@xxxxxxxxxx>
Date: Wed, 27 Sep 2017 11:42:31 +0200
Subject: [PATCH] PCI: tegra: Enable the use of 32-bit MSI
Using a value above the 32-bit boundary as MSI target address prevents
devices that are only 32-bit MSI capable from using MSI for signalling
interrupts.
Fix this by relocating the MSI target address to somewhere within the
region in the address map reserved for PCI. Addresses were chosen to
be located in unused regions in the default configurations. While some
configuration could redefine the regions to include the new MSI target
addresses, doing so shouldn't cause any issues because this memory is
not actually accessed by MSI.
Reported-by: Vidya Sagar <vidias@xxxxxxxxxx>
Signed-off-by: Thierry Reding <treding@xxxxxxxxxx>
---
drivers/pci/host/pci-tegra.c | 38 ++++++++++++++++++++------------------
1 file changed, 20 insertions(+), 18 deletions(-)
diff --git a/drivers/pci/host/pci-tegra.c b/drivers/pci/host/pci-tegra.c
index 5b02ea59524b..1026f9dbbbd7 100644
--- a/drivers/pci/host/pci-tegra.c
+++ b/drivers/pci/host/pci-tegra.c
@@ -242,6 +242,7 @@ struct tegra_msi {
struct tegra_pcie_soc {
unsigned int num_ports;
unsigned int msi_base_shift;
+ u64 msi_target;
u32 pads_pll_ctl;
u32 tx_ref_sel;
u32 pads_refclk_cfg0;
@@ -1554,27 +1555,24 @@ static int tegra_pcie_enable_msi(struct tegra_pcie *pcie)
* MSI writes, which means that the MSI target address doesn't have
* to point to actual physical memory. Rather than allocating one 4
* KiB page of system memory that's never used, we can simply pick
- * an arbitrary address within an area reserved for system memory
- * in the FPCI address map.
+ * an arbitrary address within a reserved area in the FPCI address
+ * map.
*
- * However, in order to avoid confusion, we pick an address that
- * doesn't map to physical memory. The FPCI address map reserves a
- * 1012 GiB region for system memory and memory-mapped I/O. Since
- * none of the Tegra SoCs that contain this PCI host bridge can
- * address more than 16 GiB of system memory, the last 4 KiB of
- * these 1012 GiB is a good candidate.
+ * Note that MSI writes are never committed to memory, so it doesn't
+ * really matter which address we pick. However, since the address
+ * may show up at some point and potentially confuse users, the best
+ * solution would be to pick an address that is obviously not within
+ * system memory. The FPCI address map contains a range that cannot
+ * be accessed by any Tegra CPU, but unfortunately that excludes the
+ * usage for 32-bit MSI capable devices.
*
- * Unfortunately, Tegra20 is slightly different in that the physical
- * address for this MSI region is limited to the lower 32 bits of the
- * address map, so the address that we pick is going to have to be
- * located somewhere within the region addressable by the CPU and
- * on-SoC controllers. To be on the safe side, we select an address
- * from a region that is marked unused (0xf0010000 - 0xfffeffff).
+ * Instead, in order to support both 32-bit and 64-bit MSI, choose a
+ * memory address within the PCI MMIO region that is currently not
+ * used for configuration space, downstream I/O, prefetchable or non
+ * prefetchable memory. This is hard-coded on a per-chip basis and
+ * is hopefully clear enough not to confuse anyone.
*/
- if (soc->msi_base_shift > 0)
- msi->phys = 0xfcfffff000;
- else
- msi->phys = 0x00f0010000;
+ msi->phys = soc->msi_target;
afi_writel(pcie, msi->phys >> soc->msi_base_shift, AFI_MSI_FPCI_BAR_ST);
afi_writel(pcie, msi->phys, AFI_MSI_AXI_BAR_ST);
@@ -2097,6 +2095,7 @@ static void tegra_pcie_enable_ports(struct tegra_pcie *pcie)
static const struct tegra_pcie_soc tegra20_pcie = {
.num_ports = 2,
.msi_base_shift = 0,
+ .msi_target = 0x81fff000,
.pads_pll_ctl = PADS_PLL_CTL_TEGRA20,
.tx_ref_sel = PADS_PLL_CTL_TXCLKREF_DIV10,
.pads_refclk_cfg0 = 0xfa5cfa5c,
@@ -2111,6 +2110,7 @@ static const struct tegra_pcie_soc tegra20_pcie = {
static const struct tegra_pcie_soc tegra30_pcie = {
.num_ports = 3,
.msi_base_shift = 8,
+ .msi_target = 0x01fff000,
.pads_pll_ctl = PADS_PLL_CTL_TEGRA30,
.tx_ref_sel = PADS_PLL_CTL_TXCLKREF_BUF_EN,
.pads_refclk_cfg0 = 0xfa5cfa5c,
@@ -2126,6 +2126,7 @@ static const struct tegra_pcie_soc tegra30_pcie = {
static const struct tegra_pcie_soc tegra124_pcie = {
.num_ports = 2,
.msi_base_shift = 8,
+ .msi_target = 0x01fff000,
.pads_pll_ctl = PADS_PLL_CTL_TEGRA30,
.tx_ref_sel = PADS_PLL_CTL_TXCLKREF_BUF_EN,
.pads_refclk_cfg0 = 0x44ac44ac,
@@ -2140,6 +2141,7 @@ static const struct tegra_pcie_soc tegra124_pcie = {
static const struct tegra_pcie_soc tegra210_pcie = {
.num_ports = 2,
.msi_base_shift = 8,
+ .msi_target = 0x01fff000,
.pads_pll_ctl = PADS_PLL_CTL_TEGRA30,
.tx_ref_sel = PADS_PLL_CTL_TXCLKREF_BUF_EN,
.pads_refclk_cfg0 = 0x90b890b8,
--
2.14.1
Attachment:
signature.asc
Description: PGP signature
