@@ -426,10 +426,39 @@ static int advk_pcie_wait_pio(struct advk_pcie *pcie)
udelay(PIO_RETRY_DELAY);
}
- dev_err(dev, "config read/write timed out\n");
+ dev_err(dev, "PIO read/write transfer time out\n");
return -ETIMEDOUT;
}
+static bool advk_pcie_pio_is_running(struct advk_pcie *pcie)
+{
+ struct device *dev = &pcie->pdev->dev;
+
+ /*
+ * Trying to start a new PIO transfer when previous has not completed
+ * cause External Abort on CPU which results in kernel panic:
+ *
+ * SError Interrupt on CPU0, code 0xbf000002 -- SError
+ * Kernel panic - not syncing: Asynchronous SError Interrupt
+ *
+ * Functions advk_pcie_rd_conf() and advk_pcie_wr_conf() are protected
+ * by raw_spin_lock_irqsave() at pci_lock_config() level to prevent
+ * concurrent calls at the same time. But because PIO transfer may take
+ * about 1.5s when link is down or card is disconnected, it means that
+ * advk_pcie_wait_pio() does not always have to wait for completion.
+ *
+ * Some versions of ARM Trusted Firmware handles this External Abort at
+ * EL3 level and mask it to prevent kernel panic. Relevant TF-A commit:
+ * https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/commit/?id=3c7dcdac5c50
+ */
+ if (advk_readl(pcie, PIO_START)) {
+ dev_err(dev, "Previous PIO read/write transfer is still running\n");
+ return true;
+ }
+
+ return false;
+}
+
static int advk_pcie_rd_conf(struct pci_bus *bus, u32 devfn,
int where, int size, u32 *val)
{
@@ -442,9 +471,10 @@ static int advk_pcie_rd_conf(struct pci_bus *bus, u32 devfn,
return PCIBIOS_DEVICE_NOT_FOUND;
}
- /* Start PIO */
- advk_writel(pcie, 0, PIO_START);
- advk_writel(pcie, 1, PIO_ISR);
+ if (advk_pcie_pio_is_running(pcie)) {
+ *val = 0xffffffff;
+ return PCIBIOS_SET_FAILED;
+ }
/* Program the control register */
reg = advk_readl(pcie, PIO_CTRL);
@@ -463,7 +493,8 @@ static int advk_pcie_rd_conf(struct pci_bus *bus, u32 devfn,
/* Program the data strobe */
advk_writel(pcie, 0xf, PIO_WR_DATA_STRB);
- /* Start the transfer */
+ /* Clear PIO DONE ISR and start the transfer */
+ advk_writel(pcie, 1, PIO_ISR);
advk_writel(pcie, 1, PIO_START);
ret = advk_pcie_wait_pio(pcie);
@@ -497,9 +528,8 @@ static int advk_pcie_wr_conf(struct pci_bus *bus, u32 devfn,
if (where % size)
return PCIBIOS_SET_FAILED;
- /* Start PIO */
- advk_writel(pcie, 0, PIO_START);
- advk_writel(pcie, 1, PIO_ISR);
+ if (advk_pcie_pio_is_running(pcie))
+ return PCIBIOS_SET_FAILED;
/* Program the control register */
reg = advk_readl(pcie, PIO_CTRL);
@@ -526,7 +556,8 @@ static int advk_pcie_wr_conf(struct pci_bus *bus, u32 devfn,
/* Program the data strobe */
advk_writel(pcie, data_strobe, PIO_WR_DATA_STRB);
- /* Start the transfer */
+ /* Clear PIO DONE ISR and start the transfer */
+ advk_writel(pcie, 1, PIO_ISR);
advk_writel(pcie, 1, PIO_START);
ret = advk_pcie_wait_pio(pcie);
commit f18139966d072dab8e4398c95ce955a9742e04f7 upstream. Trying to start a new PIO transfer by writing value 0 in PIO_START register when previous transfer has not yet completed (which is indicated by value 1 in PIO_START) causes an External Abort on CPU, which results in kernel panic: SError Interrupt on CPU0, code 0xbf000002 -- SError Kernel panic - not syncing: Asynchronous SError Interrupt To prevent kernel panic, it is required to reject a new PIO transfer when previous one has not finished yet. If previous PIO transfer is not finished yet, the kernel may issue a new PIO request only if the previous PIO transfer timed out. In the past the root cause of this issue was incorrectly identified (as it often happens during link retraining or after link down event) and special hack was implemented in Trusted Firmware to catch all SError events in EL3, to ignore errors with code 0xbf000002 and not forwarding any other errors to kernel and instead throw panic from EL3 Trusted Firmware handler. Links to discussion and patches about this issue: https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/commit/?id=3c7dcdac5c50 https://lore.kernel.org/linux-pci/20190316161243.29517-1-repk@triplefau.lt/ https://lore.kernel.org/linux-pci/971be151d24312cc533989a64bd454b4@www.loen.fr/ https://review.trustedfirmware.org/c/TF-A/trusted-firmware-a/+/1541 But the real cause was the fact that during link retraining or after link down event the PIO transfer may take longer time, up to the 1.44s until it times out. This increased probability that a new PIO transfer would be issued by kernel while previous one has not finished yet. After applying this change into the kernel, it is possible to revert the mentioned TF-A hack and SError events do not have to be caught in TF-A EL3. Link: https://lore.kernel.org/r/20210608203655.31228-1-pali@kernel.org Signed-off-by: Pali Rohár <pali@kernel.org> Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com> Signed-off-by: Bjorn Helgaas <bhelgaas@google.com> Reviewed-by: Marek Behún <kabel@kernel.org> Cc: stable@vger.kernel.org # 7fbcb5da811b ("PCI: aardvark: Don't rely on jiffies while holding spinlock") [pali: Backported to 4.14 version] --- This patch is backported to 4.14 version. It depends on commit 7fbcb5da811b as presented on Cc: stable line. Backport for commit 7fbcb5da811b to 4.14 was sent in previous email: https://lore.kernel.org/stable/20210716122033.22568-1-pali@kernel.org/ --- drivers/pci/host/pci-aardvark.c | 49 +++++++++++++++++++++++++++------ 1 file changed, 40 insertions(+), 9 deletions(-)