s390/pci: implement minimal PCI error recovery

When the platform detects an error on a PCI function or a service action
has been performed it is put in the error state and an error event
notification is provided to the OS.

Currently we treat all error event notifications the same and simply set
pdev->error_state = pci_channel_io_perm_failure requiring user
intervention such as use of the recover attribute to get the device
usable again. Despite requiring a manual step this also has the
disadvantage that the device is completely torn down and recreated
resulting in higher level devices such as a block or network device
being recreated. In case of a block device this also means that it may
need to be removed and added to a software raid even if that could
otherwise survive with a temporary degradation.

This is of course not ideal more so since an error notification with PEC
0x3A indicates that the platform already performed error recovery
successfully or that the error state was caused by a service action that
is now finished.

At least in this case we can assume that the error state can be reset
and the function made usable again. So as not to have the disadvantage
of a full tear down and recreation we need to coordinate this recovery
with the driver. Thankfully there is already a well defined recovery
flow for this described in Documentation/PCI/pci-error-recovery.rst.

The implementation of this is somewhat straight forward and simplified
by the fact that our recovery flow is defined per PCI function. As
a reset we use the newly introduced zpci_hot_reset_device() which also
takes the PCI function out of the error state.

Reviewed-by: Pierre Morel <pmorel@linux.ibm.com>
Acked-by: Matthew Rosato <mjrosato@linux.ibm.com>
Signed-off-by: Niklas Schnelle <schnelle@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>

2 files changed, 274 insertions, 3 deletions

diff --git a/arch/s390/pci/pci.c b/arch/s390/pci/pci.c
index e95b08e5cc5b..2f9b78fa82a5 100644
--- a/arch/s390/pci/pci.c
+++ b/arch/s390/pci/pci.c
@@ -990,6 +990,59 @@ int zpci_report_error(struct pci_dev *pdev,
 }
 EXPORT_SYMBOL(zpci_report_error);
 
+/**
+ * zpci_clear_error_state() - Clears the zPCI error state of the device
+ * @zdev: The zdev for which the zPCI error state should be reset
+ *
+ * Clear the zPCI error state of the device. If clearing the zPCI error state
+ * fails the device is left in the error state. In this case it may make sense
+ * to call zpci_io_perm_failure() on the associated pdev if it exists.
+ *
+ * Returns: 0 on success, -EIO otherwise
+ */
+int zpci_clear_error_state(struct zpci_dev *zdev)
+{
+	u64 req = ZPCI_CREATE_REQ(zdev->fh, 0, ZPCI_MOD_FC_RESET_ERROR);
+	struct zpci_fib fib = {0};
+	u8 status;
+	int cc;
+
+	cc = zpci_mod_fc(req, &fib, &status);
+	if (cc) {
+		zpci_dbg(3, "ces fid:%x, cc:%d, status:%x\n", zdev->fid, cc, status);
+		return -EIO;
+	}
+
+	return 0;
+}
+
+/**
+ * zpci_reset_load_store_blocked() - Re-enables L/S from error state
+ * @zdev: The zdev for which to unblock load/store access
+ *
+ * Re-enables load/store access for a PCI function in the error state while
+ * keeping DMA blocked. In this state drivers can poke MMIO space to determine
+ * if error recovery is possible while catching any rogue DMA access from the
+ * device.
+ *
+ * Returns: 0 on success, -EIO otherwise
+ */
+int zpci_reset_load_store_blocked(struct zpci_dev *zdev)
+{
+	u64 req = ZPCI_CREATE_REQ(zdev->fh, 0, ZPCI_MOD_FC_RESET_BLOCK);
+	struct zpci_fib fib = {0};
+	u8 status;
+	int cc;
+
+	cc = zpci_mod_fc(req, &fib, &status);
+	if (cc) {
+		zpci_dbg(3, "rls fid:%x, cc:%d, status:%x\n", zdev->fid, cc, status);
+		return -EIO;
+	}
+
+	return 0;
+}
+
 static int zpci_mem_init(void)
 {
 	BUILD_BUG_ON(!is_power_of_2(__alignof__(struct zpci_fmb)) ||
diff --git a/arch/s390/pci/pci_event.c b/arch/s390/pci/pci_event.c
index 8df8b3210c5b..2e3e5b278925 100644
--- a/arch/s390/pci/pci_event.c
+++ b/arch/s390/pci/pci_event.c
@@ -47,18 +47,223 @@ struct zpci_ccdf_avail {
 	u16 pec;			/* PCI event code */
 } __packed;
 
+static inline bool ers_result_indicates_abort(pci_ers_result_t ers_res)
+{
+	switch (ers_res) {
+	case PCI_ERS_RESULT_CAN_RECOVER:
+	case PCI_ERS_RESULT_RECOVERED:
+	case PCI_ERS_RESULT_NEED_RESET:
+		return false;
+	default:
+		return true;
+	}
+}
+
+static bool is_passed_through(struct zpci_dev *zdev)
+{
+	return zdev->s390_domain;
+}
+
+static bool is_driver_supported(struct pci_driver *driver)
+{
+	if (!driver || !driver->err_handler)
+		return false;
+	if (!driver->err_handler->error_detected)
+		return false;
+	if (!driver->err_handler->slot_reset)
+		return false;
+	if (!driver->err_handler->resume)
+		return false;
+	return true;
+}
+
+static pci_ers_result_t zpci_event_notify_error_detected(struct pci_dev *pdev,
+							 struct pci_driver *driver)
+{
+	pci_ers_result_t ers_res = PCI_ERS_RESULT_DISCONNECT;
+
+	ers_res = driver->err_handler->error_detected(pdev,  pdev->error_state);
+	if (ers_result_indicates_abort(ers_res))
+		pr_info("%s: Automatic recovery failed after initial reporting\n", pci_name(pdev));
+	else if (ers_res == PCI_ERS_RESULT_NEED_RESET)
+		pr_debug("%s: Driver needs reset to recover\n", pci_name(pdev));
+
+	return ers_res;
+}
+
+static pci_ers_result_t zpci_event_do_error_state_clear(struct pci_dev *pdev,
+							struct pci_driver *driver)
+{
+	pci_ers_result_t ers_res = PCI_ERS_RESULT_DISCONNECT;
+	struct zpci_dev *zdev = to_zpci(pdev);
+	int rc;
+
+	pr_info("%s: Unblocking device access for examination\n", pci_name(pdev));
+	rc = zpci_reset_load_store_blocked(zdev);
+	if (rc) {
+		pr_err("%s: Unblocking device access failed\n", pci_name(pdev));
+		/* Let's try a full reset instead */
+		return PCI_ERS_RESULT_NEED_RESET;
+	}
+
+	if (driver->err_handler->mmio_enabled) {
+		ers_res = driver->err_handler->mmio_enabled(pdev);
+		if (ers_result_indicates_abort(ers_res)) {
+			pr_info("%s: Automatic recovery failed after MMIO re-enable\n",
+				pci_name(pdev));
+			return ers_res;
+		} else if (ers_res == PCI_ERS_RESULT_NEED_RESET) {
+			pr_debug("%s: Driver needs reset to recover\n", pci_name(pdev));
+			return ers_res;
+		}
+	}
+
+	pr_debug("%s: Unblocking DMA\n", pci_name(pdev));
+	rc = zpci_clear_error_state(zdev);
+	if (!rc) {
+		pdev->error_state = pci_channel_io_normal;
+	} else {
+		pr_err("%s: Unblocking DMA failed\n", pci_name(pdev));
+		/* Let's try a full reset instead */
+		return PCI_ERS_RESULT_NEED_RESET;
+	}
+
+	return ers_res;
+}
+
+static pci_ers_result_t zpci_event_do_reset(struct pci_dev *pdev,
+					    struct pci_driver *driver)
+{
+	pci_ers_result_t ers_res = PCI_ERS_RESULT_DISCONNECT;
+
+	pr_info("%s: Initiating reset\n", pci_name(pdev));
+	if (zpci_hot_reset_device(to_zpci(pdev))) {
+		pr_err("%s: The reset request failed\n", pci_name(pdev));
+		return ers_res;
+	}
+	pdev->error_state = pci_channel_io_normal;
+	ers_res = driver->err_handler->slot_reset(pdev);
+	if (ers_result_indicates_abort(ers_res)) {
+		pr_info("%s: Automatic recovery failed after slot reset\n", pci_name(pdev));
+		return ers_res;
+	}
+
+	return ers_res;
+}
+
+/* zpci_event_attempt_error_recovery - Try to recover the given PCI function
+ * @pdev: PCI function to recover currently in the error state
+ *
+ * We follow the scheme outlined in Documentation/PCI/pci-error-recovery.rst.
+ * With the simplification that recovery always happens per function
+ * and the platform determines which functions are affected for
+ * multi-function devices.
+ */
+static pci_ers_result_t zpci_event_attempt_error_recovery(struct pci_dev *pdev)
+{
+	pci_ers_result_t ers_res = PCI_ERS_RESULT_DISCONNECT;
+	struct pci_driver *driver;
+
+	/*
+	 * Ensure that the PCI function is not removed concurrently, no driver
+	 * is unbound or probed and that userspace can't access its
+	 * configuration space while we perform recovery.
+	 */
+	pci_dev_lock(pdev);
+	if (pdev->error_state == pci_channel_io_perm_failure) {
+		ers_res = PCI_ERS_RESULT_DISCONNECT;
+		goto out_unlock;
+	}
+	pdev->error_state = pci_channel_io_frozen;
+
+	if (is_passed_through(to_zpci(pdev))) {
+		pr_info("%s: Cannot be recovered in the host because it is a pass-through device\n",
+			pci_name(pdev));
+		goto out_unlock;
+	}
+
+	driver = to_pci_driver(pdev->dev.driver);
+	if (!is_driver_supported(driver)) {
+		if (!driver)
+			pr_info("%s: Cannot be recovered because no driver is bound to the device\n",
+				pci_name(pdev));
+		else
+			pr_info("%s: The %s driver bound to the device does not support error recovery\n",
+				pci_name(pdev),
+				driver->name);
+		goto out_unlock;
+	}
+
+	ers_res = zpci_event_notify_error_detected(pdev, driver);
+	if (ers_result_indicates_abort(ers_res))
+		goto out_unlock;
+
+	if (ers_res == PCI_ERS_RESULT_CAN_RECOVER) {
+		ers_res = zpci_event_do_error_state_clear(pdev, driver);
+		if (ers_result_indicates_abort(ers_res))
+			goto out_unlock;
+	}
+
+	if (ers_res == PCI_ERS_RESULT_NEED_RESET)
+		ers_res = zpci_event_do_reset(pdev, driver);
+
+	if (ers_res != PCI_ERS_RESULT_RECOVERED) {
+		pr_err("%s: Automatic recovery failed; operator intervention is required\n",
+		       pci_name(pdev));
+		goto out_unlock;
+	}
+
+	pr_info("%s: The device is ready to resume operations\n", pci_name(pdev));
+	if (driver->err_handler->resume)
+		driver->err_handler->resume(pdev);
+out_unlock:
+	pci_dev_unlock(pdev);
+
+	return ers_res;
+}
+
+/* zpci_event_io_failure - Report PCI channel failure state to driver
+ * @pdev: PCI function for which to report
+ * @es: PCI channel failure state to report
+ */
+static void zpci_event_io_failure(struct pci_dev *pdev, pci_channel_state_t es)
+{
+	struct pci_driver *driver;
+
+	pci_dev_lock(pdev);
+	pdev->error_state = es;
+	/**
+	 * While vfio-pci's error_detected callback notifies user-space QEMU
+	 * reacts to this by freezing the guest. In an s390 environment PCI
+	 * errors are rarely fatal so this is overkill. Instead in the future
+	 * we will inject the error event and let the guest recover the device
+	 * itself.
+	 */
+	if (is_passed_through(to_zpci(pdev)))
+		goto out;
+	driver = to_pci_driver(pdev->dev.driver);
+	if (driver && driver->err_handler && driver->err_handler->error_detected)
+		driver->err_handler->error_detected(pdev, pdev->error_state);
+out:
+	pci_dev_unlock(pdev);
+}
+
 static void __zpci_event_error(struct zpci_ccdf_err *ccdf)
 {
 	struct zpci_dev *zdev = get_zdev_by_fid(ccdf->fid);
 	struct pci_dev *pdev = NULL;
+	pci_ers_result_t ers_res;
 
 	zpci_dbg(3, "err fid:%x, fh:%x, pec:%x\n",
 		 ccdf->fid, ccdf->fh, ccdf->pec);
 	zpci_err("error CCDF:\n");
 	zpci_err_hex(ccdf, sizeof(*ccdf));
 
-	if (zdev)
-		pdev = pci_get_slot(zdev->zbus->bus, zdev->devfn);
+	if (zdev) {
+		zpci_update_fh(zdev, ccdf->fh);
+		if (zdev->zbus->bus)
+			pdev = pci_get_slot(zdev->zbus->bus, zdev->devfn);
+	}
 
 	pr_err("%s: Event 0x%x reports an error for PCI function 0x%x\n",
 	       pdev ? pci_name(pdev) : "n/a", ccdf->pec, ccdf->fid);
@@ -66,7 +271,20 @@ static void __zpci_event_error(struct zpci_ccdf_err *ccdf)
 	if (!pdev)
 		return;
 
-	pdev->error_state = pci_channel_io_perm_failure;
+	switch (ccdf->pec) {
+	case 0x003a: /* Service Action or Error Recovery Successful */
+		ers_res = zpci_event_attempt_error_recovery(pdev);
+		if (ers_res != PCI_ERS_RESULT_RECOVERED)
+			zpci_event_io_failure(pdev, pci_channel_io_perm_failure);
+		break;
+	default:
+		/*
+		 * Mark as frozen not permanently failed because the device
+		 * could be subsequently recovered by the platform.
+		 */
+		zpci_event_io_failure(pdev, pci_channel_io_frozen);
+		break;
+	}
 	pci_dev_put(pdev);
 }