fix(plat/marvell/a3k): update information about PCIe abort hack
A3700 plat_ea_handler was introduced into TF-A codebase just because of
bugs in U-Boot and Linux kernel PCIe controller driver pci-aardvark.c.
These bugs were finally fixed in both U-Boot and Linux kernel drivers:
https://source.denx.de/u-boot/u-boot/-/commit/eccbd4ad8e4e182638eafbfb87ac139c04f24a01
https://git.kernel.org/stable/c/f18139966d072dab8e4398c95ce955a9742e04f7
Add all these information into comments, including printing error
message into a3k plat_ea_handler. Also check that abort is really
asynchronous and comes from lower level than EL3.
Signed-off-by: Pali Rohár <pali@kernel.org>
Change-Id: I46318d221b39773d5e25b3a0221d7738736ffdf1
diff --git a/plat/marvell/armada/a3k/common/a3700_ea.c b/plat/marvell/armada/a3k/common/a3700_ea.c
index 3a4f720..dde92e3 100644
--- a/plat/marvell/armada/a3k/common/a3700_ea.c
+++ b/plat/marvell/armada/a3k/common/a3700_ea.c
@@ -8,14 +8,75 @@
#include <common/debug.h>
#include <arch_helpers.h>
#include <plat/common/platform.h>
+#include <bl31/ea_handle.h>
-#define ADVK_SERROR_SYNDROME 0xbf000002
+#define A53_SERR_INT_AXI_SLVERR_ON_EXTERNAL_ACCESS 0xbf000002
+
+#if !ENABLE_BACKTRACE
+static const char *get_el_str(unsigned int el)
+{
+ if (el == MODE_EL3) {
+ return "EL3";
+ } else if (el == MODE_EL2) {
+ return "EL2";
+ }
+ return "S-EL1";
+}
+#endif /* !ENABLE_BACKTRACE */
void plat_ea_handler(unsigned int ea_reason, uint64_t syndrome, void *cookie,
void *handle, uint64_t flags)
{
- if (syndrome == ADVK_SERROR_SYNDROME)
+ unsigned int level = (unsigned int)GET_EL(read_spsr_el3());
+
+ /*
+ * Asynchronous External Abort with syndrome 0xbf000002 on Cortex A53
+ * core means SError interrupt caused by AXI SLVERR on external access.
+ *
+ * In most cases this indicates a bug in U-Boot or Linux kernel driver
+ * pci-aardvark.c which implements access to A3700 PCIe config space.
+ * Driver does not wait for PCIe PIO transfer completion and try to
+ * start a new PCIe PIO transfer while previous has not finished yet.
+ * A3700 PCIe controller in this case sends SLVERR via AXI which results
+ * in a fatal Asynchronous SError interrupt on Cortex A53 CPU.
+ *
+ * Following patches fix that bug in U-Boot and Linux kernel drivers:
+ * https://source.denx.de/u-boot/u-boot/-/commit/eccbd4ad8e4e182638eafbfb87ac139c04f24a01
+ * https://git.kernel.org/stable/c/f18139966d072dab8e4398c95ce955a9742e04f7
+ *
+ * As a hacky workaround for unpatched U-Boot and Linux kernel drivers
+ * ignore all asynchronous aborts with that syndrome value received on
+ * CPU from level lower than EL3.
+ *
+ * Because these aborts are delivered on CPU asynchronously, they are
+ * imprecise and we cannot check the real reason of abort and neither
+ * who and why sent this abort. We expect that on A3700 it is always
+ * PCIe controller.
+ *
+ * Hence ignoring all aborts with this syndrome value is just a giant
+ * hack that we need only because of bugs in old U-Boot and Linux kernel
+ * versions and because it was decided that TF-A would implement this
+ * hack for U-Boot and Linux kernel it in this way. New patched U-Boot
+ * and kernel versions do not need it anymore.
+ *
+ * Links to discussion about this workaround:
+ * 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
+ */
+ if (level < MODE_EL3 && ea_reason == ERROR_EA_ASYNC &&
+ syndrome == A53_SERR_INT_AXI_SLVERR_ON_EXTERNAL_ACCESS) {
+ ERROR_NL();
+ ERROR("Ignoring Asynchronous External Abort with"
+ " syndrome 0x%llx received on 0x%lx from %s\n",
+ syndrome, read_mpidr_el1(), get_el_str(level));
+ ERROR("SError interrupt: AXI SLVERR on external access\n");
+ ERROR("This indicates a bug in pci-aardvark.c driver\n");
+ ERROR("Please update U-Boot/Linux to the latest version\n");
+ ERROR_NL();
+ console_flush();
return;
+ }
plat_default_ea_handler(ea_reason, syndrome, cookie, handle, flags);
}