Skip to contentSkip to conversation

PPC64: EEH Recovery

PATCH In review Linas Vepstas 3 participants 7 messages Jan 6, 2005 – Jan 21, 2005

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 PPC64: EEH Recovery

Hi Paul,

The patch below implements hotplug style EEH error recovery. Its split into two pieces: a part that needs to be applied to the PPC64 arch tree, and a part that needs to be applied to the RPA PHP hotplug tree. The PPC64 part needs to go in first.

Assuming this doesn't generate a round of discussion, please forward upstream to akpm/torvalds.

===== arch/ppc64/kernel/eeh.c 1.41 vs edited =====

arch/ppc64/kernel/eeh.c +277 -44
@@ -17,21 +17,19 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  */
 
-#include <linux/bootmem.h>
+#include <linux/delay.h>
 #include <linux/init.h>
 #include <linux/list.h>
-#include <linux/mm.h>
 #include <linux/notifier.h>
 #include <linux/pci.h>
 #include <linux/proc_fs.h>
 #include <linux/rbtree.h>
 #include <linux/seq_file.h>
-#include <linux/spinlock.h>
+#include <asm/atomic.h>
 #include <asm/eeh.h>
 #include <asm/io.h>
 #include <asm/machdep.h>
 #include <asm/rtas.h>
-#include <asm/atomic.h>
 #include "pci.h"
 
 #undef DEBUG
@@ -89,7 +87,6 @@ static struct notifier_block *eeh_notifi
  * attempts we allow before panicking.
  */
 #define EEH_MAX_FAILS	1000
-static atomic_t eeh_fail_count;
 
 /* RTAS tokens */
 static int ibm_set_eeh_option;
@@ -106,6 +103,10 @@ static spinlock_t slot_errbuf_lock = SPI
 static int eeh_error_buf_size;
 
 /* System monitoring statistics */
+static DEFINE_PER_CPU(unsigned long, no_device);
+static DEFINE_PER_CPU(unsigned long, no_dn);
+static DEFINE_PER_CPU(unsigned long, no_cfg_addr);
+static DEFINE_PER_CPU(unsigned long, ignored_check);
 static DEFINE_PER_CPU(unsigned long, total_mmio_ffs);
 static DEFINE_PER_CPU(unsigned long, false_positives);
 static DEFINE_PER_CPU(unsigned long, ignored_failures);
@@ -224,9 +225,9 @@ pci_addr_cache_insert(struct pci_dev *de
 	while (*p) {
 		parent = *p;
 		piar = rb_entry(parent, struct pci_io_addr_range, rb_node);
-		if (alo < piar->addr_lo) {
+		if (ahi < piar->addr_lo) {
 			p = &parent->rb_left;
-		} else if (ahi > piar->addr_hi) {
+		} else if (alo > piar->addr_hi) {
 			p = &parent->rb_right;
 		} else {
 			if (dev != piar->pcidev ||
@@ -244,6 +245,11 @@ pci_addr_cache_insert(struct pci_dev *de
 	piar->addr_hi = ahi;
 	piar->pcidev = dev;
 	piar->flags = flags;
+	
+#ifdef DEBUG 
+	printk (KERN_DEBUG "PIAR: insert range=[%lx:%lx] dev=%s\n", 
+	               alo, ahi, pci_name (dev));
+#endif
 
 	rb_link_node(&piar->rb_node, parent, p);
 	rb_insert_color(&piar->rb_node, &pci_io_addr_cache_root.rb_root);
@@ -368,6 +374,7 @@ void pci_addr_cache_remove_device(struct
  */
 void __init pci_addr_cache_build(void)
 {
+	struct device_node *dn;
 	struct pci_dev *dev = NULL;
 
 	spin_lock_init(&pci_io_addr_cache_root.piar_lock);
@@ -378,6 +385,14 @@ void __init pci_addr_cache_build(void)
 			continue;
 		}
 		pci_addr_cache_insert_device(dev);
+		
+		/* Save the BAR's; firmware doesn't restore these after EEH reset */
+		dn = pci_device_to_OF_node(dev);
+		if (dn) {
+			int i;
+			for (i = 0; i < 16; i++) 
+				pci_read_config_dword(dev, i * 4, &dn->config_space[i]);
+		}
 	}
 
 #ifdef DEBUG
@@ -389,6 +404,32 @@ void __init pci_addr_cache_build(void)
 /* --------------------------------------------------------------- */
 /* Above lies the PCI Address Cache. Below lies the EEH event infrastructure */
 
+void eeh_slot_error_detail (struct device_node *dn, int severity)
+{
+	unsigned long flags;
+	int rc;
+
+	if (!dn) return;
+
+	/* Log the error with the rtas logger */
+	spin_lock_irqsave(&slot_errbuf_lock, flags);
+	memset(slot_errbuf, 0, eeh_error_buf_size);
+
+	rc = rtas_call(ibm_slot_error_detail,
+	               8, 1, NULL, dn->eeh_config_addr,
+	               BUID_HI(dn->phb->buid),
+	               BUID_LO(dn->phb->buid), NULL, 0,
+	               virt_to_phys(slot_errbuf),
+	               eeh_error_buf_size,
+	               severity);
+
+	if (rc == 0)
+		log_error(slot_errbuf, ERR_TYPE_RTAS_LOG, 0);
+	spin_unlock_irqrestore(&slot_errbuf_lock, flags);
+}
+
+EXPORT_SYMBOL(eeh_slot_error_detail);
+
 /**
  * eeh_register_notifier - Register to find out about EEH events.
  * @nb: notifier block to callback on events
@@ -484,11 +525,9 @@ static void eeh_event_handler(void *dumm
 		       "%s %s\n", event->reset_state,
 		       pci_name(event->dev), pci_pretty_name(event->dev));
 
-		atomic_set(&eeh_fail_count, 0);
-		notifier_call_chain (&eeh_notifier_chain,
-				     EEH_NOTIFY_FREEZE, event);
-
 		__get_cpu_var(slot_resets)++;
+		notifier_call_chain (&eeh_notifier_chain,
+		           EEH_NOTIFY_FREEZE, event);
 
 		pci_dev_put(event->dev);
 		kfree(event);
@@ -496,8 +535,8 @@ static void eeh_event_handler(void *dumm
 }
 
 /**
- * eeh_token_to_phys - convert EEH address token to phys address
- * @token i/o token, should be address in the form 0xE....
+ * eeh_token_to_phys - convert I/O address to phys address
+ * @token i/o address, should be address in the form 0xA....
  */
 static inline unsigned long eeh_token_to_phys(unsigned long token)
 {
@@ -512,6 +551,17 @@ static inline unsigned long eeh_token_to
 	return pa | (token & (PAGE_SIZE-1));
 }
 
+static inline struct pci_dev * eeh_get_pci_dev(struct device_node *dn)
+{
+	struct pci_dev *dev = NULL;
+
+	for_each_pci_dev(dev) {
+		if (pci_device_to_OF_node(dev) == dn)
+			return dev;
+	}
+	return NULL;
+}
+
 /**
  * eeh_dn_check_failure - check if all 1's data is due to EEH slot freeze
  * @dn device node
@@ -532,7 +582,7 @@ int eeh_dn_check_failure(struct device_n
 	int ret;
 	int rets[3];
 	unsigned long flags;
-	int rc, reset_state;
+	int reset_state;
 	struct eeh_event  *event;
 
 	__get_cpu_var(total_mmio_ffs)++;
@@ -540,16 +590,20 @@ int eeh_dn_check_failure(struct device_n
 	if (!eeh_subsystem_enabled)
 		return 0;
 
-	if (!dn)
+	if (!dn) {
+		__get_cpu_var(no_dn)++;
 		return 0;
+	}
 
 	/* Access to IO BARs might get this far and still not want checking. */
 	if (!(dn->eeh_mode & EEH_MODE_SUPPORTED) ||
 	    dn->eeh_mode & EEH_MODE_NOCHECK) {
+		__get_cpu_var(ignored_check)++;
 		return 0;
 	}
 
 	if (!dn->eeh_config_addr) {
+		__get_cpu_var(no_cfg_addr)++;
 		return 0;
 	}
 
@@ -558,8 +612,9 @@ int eeh_dn_check_failure(struct device_n
 	 * slot, we know it's bad already, we don't need to check...
 	 */
 	if (dn->eeh_mode & EEH_MODE_ISOLATED) {
-		atomic_inc(&eeh_fail_count);
-		if (atomic_read(&eeh_fail_count) >= EEH_MAX_FAILS) {
+		dn->eeh_freeze_count ++;
+		if (dn->eeh_freeze_count >= EEH_MAX_FAILS) {
+			dump_stack();
 			/* re-read the slot reset state */
 			if (read_slot_reset_state(dn, rets) != 0)
 				rets[0] = -1;	/* reset state unknown */
@@ -581,34 +636,25 @@ int eeh_dn_check_failure(struct device_n
 		return 0;
 	}
 
-	/* prevent repeated reports of this failure */
+	/* Prevent repeated reports of this failure */
 	dn->eeh_mode |= EEH_MODE_ISOLATED;
 
 	reset_state = rets[0];
+	/* Log the error with the rtas logger */
+	if (dn->eeh_freeze_count < EEH_MAX_ALLOWED_FREEZES) {
+		eeh_slot_error_detail (dn, 1 /* Temporary Error */);
+	} else {
+		eeh_slot_error_detail (dn, 2 /* Permanent Error */);
+   }
 
-	spin_lock_irqsave(&slot_errbuf_lock, flags);
-	memset(slot_errbuf, 0, eeh_error_buf_size);
-
-	rc = rtas_call(ibm_slot_error_detail,
-	               8, 1, NULL, dn->eeh_config_addr,
-	               BUID_HI(dn->phb->buid),
-	               BUID_LO(dn->phb->buid), NULL, 0,
-	               virt_to_phys(slot_errbuf),
-	               eeh_error_buf_size,
-	               1 /* Temporary Error */);
-
-	if (rc == 0)
-		log_error(slot_errbuf, ERR_TYPE_RTAS_LOG, 0);
-	spin_unlock_irqrestore(&slot_errbuf_lock, flags);
-
-	printk(KERN_INFO "EEH: MMIO failure (%d) on device: %s %s\n",
-	       rets[0], dn->name, dn->full_name);
 	event = kmalloc(sizeof(*event), GFP_ATOMIC);
 	if (event == NULL) {
-		eeh_panic(dev, reset_state);
+		printk (KERN_ERR "EEH: out of memory, event not handled\n");
 		return 1;
  	}
 
+	if (!dev)
+		dev = eeh_get_pci_dev (dn);
 	event->dev = dev;
 	event->dn = dn;
 	event->reset_state = reset_state;
@@ -634,7 +680,6 @@ EXPORT_SYMBOL(eeh_dn_check_failure);
  * @token i/o token, should be address in the form 0xA....
  * @val value, should be all 1's (XXX why do we need this arg??)
  *
- * Check for an eeh failure at the given token address.
  * Check for an EEH failure at the given token address.  Call this
  * routine if the result of a read was all 0xff's and you want to
  * find out if this is due to an EEH slot freeze event.  This routine
@@ -642,6 +687,7 @@ EXPORT_SYMBOL(eeh_dn_check_failure);
  *
  * Note this routine is safe to call in an interrupt context.
  */
+
 unsigned long eeh_check_failure(const volatile void __iomem *token, unsigned long val)
 {
 	unsigned long addr;
@@ -651,8 +697,10 @@ unsigned long eeh_check_failure(const vo
 	/* Finding the phys addr + pci device; this is pretty quick. */
 	addr = eeh_token_to_phys((unsigned long __force) token);
 	dev = pci_get_device_by_addr(addr);
-	if (!dev)
+	if (!dev) {
+		__get_cpu_var(no_device)++;
 		return val;
+	}
 
 	dn = pci_device_to_OF_node(dev);
 	eeh_dn_check_failure (dn, dev);
@@ -663,6 +711,172 @@ unsigned long eeh_check_failure(const vo
 
 EXPORT_SYMBOL(eeh_check_failure);
 
+/* ------------------------------------------------------------- */
+/* The code below deals with error recovery */
+
+void
+rtas_set_slot_reset(struct device_node *dn)
+{
+	int token = rtas_token ("ibm,set-slot-reset");
+	int rc;
+
+	if (token == RTAS_UNKNOWN_SERVICE)
+		return;
+	rc = rtas_call(token,4,1, NULL,
+	               dn->eeh_config_addr,
+	               BUID_HI(dn->phb->buid),
+	               BUID_LO(dn->phb->buid),
+	               1);
+	if (rc) {
+		printk (KERN_WARNING "EEH: Unable to reset the failed slot\n");
+		return;
+	}
+	
+	/* The PCI bus requires that the reset be held high for at least
+	 * a 100 milliseconds. We wait a bit longer 'just in case'.
+	 */
+   msleep (200);
+	
+	rc = rtas_call(token,4,1, NULL,
+	               dn->eeh_config_addr,
+	               BUID_HI(dn->phb->buid),
+	               BUID_LO(dn->phb->buid),
+	               0);
+}
+
+EXPORT_SYMBOL(rtas_set_slot_reset);
+
+void
+rtas_configure_bridge(struct device_node *dn)
+{
+	int token = rtas_token ("ibm,configure-bridge");
+	int rc;
+
+	if (token == RTAS_UNKNOWN_SERVICE)
+		return;
+	rc = rtas_call(token,3,1, NULL,
+	               dn->eeh_config_addr,
+	               BUID_HI(dn->phb->buid),
+	               BUID_LO(dn->phb->buid));
+	if (rc) {
+		printk (KERN_WARNING "EEH: Unable to configure device bridge\n");
+	}
+}
+
+EXPORT_SYMBOL(rtas_configure_bridge);
+
+/* ------------------------------------------------------- */
+/** Save and restore of PCI BARs
+ * 
+ * Although firmware will set up BARs during boot, it doesn't
+ * set up device BAR's after a device reset, although it will,
+ * if requested, set up bridge configuration. Thus, we need to 
+ * configure the PCI devices ourselves.  Config-space setup is 
+ * stored in the PCI structures which are normally deleted during
+ * device removal.  Thus, the "save" routine references the
+ * structures so that they aren't deleted. 
+ */
+
+
+struct eeh_cfg_tree
+{
+	struct eeh_cfg_tree *sibling;
+	struct eeh_cfg_tree *child;
+	struct device_node *dn;
+	int is_bridge;
+};
+
+/** 
+ * eeh_save_bars - save the PCI config space info
+ */
+struct eeh_cfg_tree * eeh_save_bars(struct device_node *dn)
+{
+	struct pci_dev *dev;
+	struct eeh_cfg_tree *cnode;
+
+	dev = eeh_get_pci_dev(dn);
+	if (!dev)
+		return NULL;
+	
+	cnode = kmalloc(sizeof(struct eeh_cfg_tree), GFP_KERNEL);
+	if (!cnode) 
+		return NULL;
+	
+	cnode->is_bridge = 0;
+	
+	if (dev->hdr_type == PCI_HEADER_TYPE_BRIDGE) 
+		cnode->is_bridge = 1;
+			  
+	of_node_get(dn);
+	cnode->dn = dn;
+	
+	cnode->sibling = NULL;
+	cnode->child = NULL;
+
+	if (dn->child) {
+		cnode->child = eeh_save_bars (dn->child);
+	}
+	if (dn->sibling) {
+		cnode->sibling = eeh_save_bars (dn->sibling);
+	}
+
+	return cnode;
+}
+EXPORT_SYMBOL(eeh_save_bars);
+
+/**
+ * __restore_bars - Restore the Base Address Registers
+ * Loads the PCI configuration space base address registers, 
+ * the expansion ROM base address, the latency timer, and etc.
+ * from the saved values in the device node.
+ */
+static inline void __restore_bars (struct device_node *dn)
+{
+	int i;
+	for (i=4; i<10; i++) {
+		rtas_write_config(dn, i*4, 4, dn->config_space[i]);
+	}
+
+	/* 12 == Expansion ROM Address */
+	rtas_write_config(dn, 12*4, 4, dn->config_space[12]);
+	
+#define SAVED_BYTE(OFF) (((u8 *)(dn->config_space))[OFF])
+	
+	rtas_write_config (dn, PCI_CACHE_LINE_SIZE, 1, 
+	            SAVED_BYTE(PCI_CACHE_LINE_SIZE));
+	
+	rtas_write_config (dn, PCI_LATENCY_TIMER, 1, 
+	            SAVED_BYTE(PCI_LATENCY_TIMER));
+	
+	rtas_write_config (dn, PCI_INTERRUPT_LINE, 1, 
+	            SAVED_BYTE(PCI_INTERRUPT_LINE));
+}
+
+/** 
+ * eeh_restore_bars - restore the PCI config space info
+ */
+void eeh_restore_bars(struct eeh_cfg_tree *tree)
+{
+	if (!(tree->is_bridge))
+		__restore_bars (tree->dn);
+	
+	if (tree->child)
+		eeh_restore_bars (tree->child);
+
+	if (tree->sibling)
+		eeh_restore_bars (tree->sibling);
+
+	of_node_put (tree->dn);
+	kfree (tree);
+}
+EXPORT_SYMBOL(eeh_restore_bars);
+
+/* ------------------------------------------------------------- */
+/* The code below deals with enabling EEH for devices during  the
+ * early boot sequence.  EEH must be enabled before any PCI probing
+ * can be done.
+ */
+
 struct eeh_early_enable_info {
 	unsigned int buid_hi;
 	unsigned int buid_lo;
@@ -829,7 +1043,9 @@ void eeh_add_device_early(struct device_
 		return;
 	phb = dn->phb;
 	if (NULL == phb || 0 == phb->buid) {
-		printk(KERN_WARNING "EEH: Expected buid but found none\n");
+		printk(KERN_WARNING "EEH: Expected buid but found none for %s\n",
+		                dn->full_name);
+		dump_stack();
 		return;
 	}
 
@@ -848,6 +1064,9 @@ EXPORT_SYMBOL(eeh_add_device_early);
  */
 void eeh_add_device_late(struct pci_dev *dev)
 {
+	int i;
+	struct device_node *dn;
+
 	if (!dev || !eeh_subsystem_enabled)
 		return;
 
@@ -857,6 +1076,11 @@ void eeh_add_device_late(struct pci_dev 
 #endif
 
 	pci_addr_cache_insert_device (dev);
+
+	/* Save the BAR's; firmware doesn't restore these after EEH reset */
+	dn = pci_device_to_OF_node(dev);
+	for (i = 0; i < 16; i++)
+		pci_read_config_dword(dev, i * 4, &dn->config_space[i]);
 }
 EXPORT_SYMBOL(eeh_add_device_late);
 
@@ -886,12 +1110,17 @@ static int proc_eeh_show(struct seq_file
 	unsigned int cpu;
 	unsigned long ffs = 0, positives = 0, failures = 0;
 	unsigned long resets = 0;
+	unsigned long no_dev = 0, no_dn = 0, no_cfg = 0, no_check = 0;
 
 	for_each_cpu(cpu) {
 		ffs += per_cpu(total_mmio_ffs, cpu);
 		positives += per_cpu(false_positives, cpu);
 		failures += per_cpu(ignored_failures, cpu);
 		resets += per_cpu(slot_resets, cpu);
+		no_dev += per_cpu(no_device, cpu);
+		no_dn += per_cpu(no_dn, cpu);
+		no_cfg += per_cpu(no_cfg_addr, cpu);
+		no_check += per_cpu(ignored_check, cpu);
 	}
 
 	if (0 == eeh_subsystem_enabled) {
@@ -899,13 +1128,17 @@ static int proc_eeh_show(struct seq_file
 		seq_printf(m, "eeh_total_mmio_ffs=%ld\n", ffs);
 	} else {
 		seq_printf(m, "EEH Subsystem is enabled\n");
-		seq_printf(m, "eeh_total_mmio_ffs=%ld\n"
+		seq_printf(m, 
+				"no device=%ld\n"
+				"no device node=%ld\n"
+				"no config address=%ld\n"
+				"check not wanted=%ld\n"
+				"eeh_total_mmio_ffs=%ld\n"
 			   "eeh_false_positives=%ld\n"
 			   "eeh_ignored_failures=%ld\n"
-			   "eeh_slot_resets=%ld\n"
-				"eeh_fail_count=%d\n",
-			   ffs, positives, failures, resets,
-				eeh_fail_count.counter);
+			   "eeh_slot_resets=%ld\n",
+				no_dev, no_dn, no_cfg, no_check,
+			   ffs, positives, failures, resets);
 	}
 
 	return 0;

===== arch/ppc64/kernel/pSeries_pci.c 1.59 vs edited =====

arch/ppc64/kernel/pSeries_pci.c +1 -1
@@ -102,7 +102,7 @@ static int rtas_pci_read_config(struct p
 	return PCIBIOS_DEVICE_NOT_FOUND;
 }
 
-static int rtas_write_config(struct device_node *dn, int where, int size, u32 val)
+int rtas_write_config(struct device_node *dn, int where, int size, u32 val)
 {
 	unsigned long buid, addr;
 	int ret;

===== include/asm-ppc64/eeh.h 1.23 vs edited =====

include/asm-ppc64/eeh.h +40 -1

66 lines not on this page · raw message

===== include/asm-ppc64/prom.h 1.24 vs edited =====

include/asm-ppc64/prom.h +2 -0
@@ -164,8 +164,10 @@ struct device_node {
 	int	status;			/* Current device status (non-zero is bad) */
 	int	eeh_mode;		/* See eeh.h for possible EEH_MODEs */
 	int	eeh_config_addr;
+	int	eeh_freeze_count;   /* number of times this device froze up. */
 	struct  pci_controller *phb;	/* for pci devices */
 	struct	iommu_table *iommu_table;	/* for phb's or bridges */
+	u32      config_space[16]; /* saved PCI config space */
 
 	struct	property *properties;
 	struct	device_node *parent;

===== include/asm-ppc64/rtas.h 1.25 vs edited =====

include/asm-ppc64/rtas.h +2 -0
@@ -241,4 +241,6 @@ extern void rtas_stop_self(void);
 /* RMO buffer reserved for user-space RTAS use */
 extern unsigned long rtas_rmo_buf;
 
+extern int rtas_write_config(struct device_node *dn, int where, int size, u32 val);
+
 #endif /* _PPC64_RTAS_H */

===== drivers/pci/hotplug/rpaphp.h 1.11 vs edited =====

drivers/pci/hotplug/rpaphp.h +2 -0
@@ -126,6 +126,8 @@ extern int register_pci_slot(struct slot
 extern int rpaphp_unconfig_pci_adapter(struct slot *slot);
 extern int rpaphp_get_pci_adapter_status(struct slot *slot, int is_init, u8 * value);
 extern struct hotplug_slot *rpaphp_find_hotplug_slot(struct pci_dev *dev);
+extern void init_eeh_handler (void);
+extern void exit_eeh_handler (void);
 
 /* rpaphp_core.c */
 extern int rpaphp_add_slot(struct device_node *dn);

===== drivers/pci/hotplug/rpaphp_core.c 1.18 vs edited =====

drivers/pci/hotplug/rpaphp_core.c +6 -0
@@ -443,12 +443,18 @@ static int __init rpaphp_init(void)
 {
 	info(DRIVER_DESC " version: " DRIVER_VERSION "\n");
 
+	/* Get set to handle EEH events. */
+	init_eeh_handler();
+
 	/* read all the PRA info from the system */
 	return init_rpa();
 }
 
 static void __exit rpaphp_exit(void)
 {
+	/* Let EEH know we are going away. */
+	exit_eeh_handler();
+
 	cleanup_slots();
 }
 

===== drivers/pci/hotplug/rpaphp_pci.c 1.17 vs edited =====

drivers/pci/hotplug/rpaphp_pci.c +153 -21

247 lines not on this page · raw message

#1 lore raw

Andrew,

The attached file describes PCI bus EEH "Extended Error Handling" concepts and operation; could you drop this into the kernel documentation tree, at linux-2.6/Documentation/powerpc/eeh-pci-error-recovery.txt ?

--linas

p.s. It was not clear to me if the EEH patch previously sent (6 January 2005, same subject line) will be wending its way into the main Torvalds kernel tree, or not. I hadn't really gotten confirmation one way or another.


                      PCI Bus EEH Error Recovery
                      --------------------------
                           Linas Vepstas
                       <[email protected]>
                          12 January 2005

Overview:
---------
The IBM POWER-based pSeries and iSeries computers include PCI bus controller chips that have extended capabilities for detecting and reporting a large variety of PCI bus error conditions. These features go under the name of "EEH", for "Extended Error Handling". The EEH hardware features allow PCI bus errors to be cleared and a PCI card to be "rebooted", without also having to reboot the operating system.

This is in contrast to traditional PCI error handling, where the PCI chip is wired directly to the CPU, and an error would cause a CPU machine-check/check-stop condition, halting the CPU entirely. Another "traditional" technique is to ignore such errors, which can lead to data corruption, both of user data or of kernel data, hung/unresponsive adapters, or system crashes/lockups. Thus, the idea behind EEH is that the operating system can become more reliable and robust by protecting it from PCI errors, and giving the OS the ability to "reboot"/recover individual PCI devices.

Future systems from other vendors, based on the PCI-E specification, may contain similar features.

Causes of EEH Errors
--------------------
EEH was originally designed to guard against hardware failure, such as PCI cards dying from heat, humidity, dust, vibration and bad electrical connections. The vast majority of EEH errors seen in "real life" are due to eithr poorly seated PCI cards, or, unfortunately quite commonly, due device driver bugs, device firmware bugs, and sometimes PCI card hardware bugs.

The most common software bug, is one that causes the device to attempt to DMA to a location in system memory that has not been reserved for DMA access for that card. This is a powerful feature, as it prevents what; otherwise, would have been silent memory corruption caused by the bad DMA. A number of device driver bugs have been found and fixed in this way over the past few years. Other possible causes of EEH errors include data or address line parity errors (for example, due to poor electrical connectivity due to a poorly seated card), and PCI-X split-completion errors (due to software, device firmware, or device PCI hardware bugs). The vast majority of "true hardware failures" can be cured by physically removing and re-seating the PCI card.

Detection and Recovery
----------------------
In the following discussion, a generic overview of how to detect and recover from EEH errors will be presented. This is followed by an overview of how the current implementation in the Linux kernel does it. The actual implementation is subject to change, and some of the finer points are still being debated. These may in turn be swayed if or when other architectures implement similar functionality.

When a PCI Host Bridge (PHB, the bus controller connecting the PCI bus to the system CPU electronics complex) detects a PCI error condition, it will "isolate" the affected PCI card. Isolation will block all writes (either to the card from the system, or from the card to the system), and it will cause all reads to return all-ff's (0xff, 0xffff, 0xffffffff for 8/16/32-bit reads). This value was chosen because it is the same value you would get if the device was physically unplugged from the slot. This includes access to PCI memory, I/O space, and PCI config space. Interrupts; however, will continued to be delivered.

Detection and recovery are performed with the aid of ppc64 firmware. The programming interfaces in the Linux kernel into the firmware are referred to as RTAS (Run-Time Abstraction Services). The Linux kernel does not (should not) access the EEH function in the PCI chipsets directly, primarily because there are a number of different chipsets out there, each with different interfaces and quirks. The firmware provides a uniform abstraction layer that will work with all pSeries and iSeries hardware (and be forwards-compatible).

If the OS or device driver suspects that a PCI slot has been EEH-isolated, there is a firmware call it can make to determine if this is the case. If so, then the device driver should put itself into a consistent state (given that it won't be able to complete any pending work) and start recovery of the card. Recovery normally would consist of reseting the PCI device (holding the PCI #RST line high for two seconds), followed by setting up the device config space (the base address registers (BAR's), latency timer, cache line size, interrupt line, and so on). This is followed by a reinitialization of the device driver. In a worst-case scenario, the power to the card can be toggled, at least on hot-plug-capable slots. In principle, layers far above the device driver probably do not need to know that the PCI card has been "rebooted" in this way; ideally, there should be at most a pause in Ethernet/disk/USB I/O while the card is being reset.

If the card cannot be recovered after three or four resets, the kernel/device driver should assume the worst-case scenario, that the card has died completely, and report this error to the sysadmin. In addition, error messages are reported through RTAS and also through syslogd (/var/log/messages) to alert the sysadmin of PCI resets. The correct way to deal with failed adapters is to use the standard PCI hotplug tools to remove and replace the dead card.

Current PPC64 Linux EEH Implementation
--------------------------------------
At this time, a generic EEH recovery mechanism has been implemented, so that individual device drivers do not need to be modified to support EEH recovery. This generic mechanism piggy-backs on the PCI hotplug infrastructure, and percolates events up through the hotplug/udev infrastructure. Followiing is a detailed description of how this is accomplished.

EEH must be enabled in the PHB's very early during the boot process, and if a PCI slot is hot-plugged. The former is performed by eeh_init() in arch/ppc64/kernel/eeh.c, and the later by drivers/pci/hotplug/pSeries_pci.c calling in to the eeh.c code. EEH must be enabled before a PCI scan of the device can proceed. Current Power5 hardware will not work unless EEH is enabled; although older Power4 can run with it disabled. Effectively, EEH can no longer be turned off. PCI devices *must* be registered with the EEH code; the EEH code needs to know about the I/O address ranges of the PCI device in order to detect an error. Given an arbitrary address, the routine pci_get_device_by_addr() will find the pci device associated with that address (if any).

The default include/asm-ppc64/io.h macros readb(), inb(), insb(), etc. include a check to see if the the i/o read returned all-0xff's. If so, these make a call to eeh_dn_check_failure(), which in turn asks the firmware if the all-ff's value is the sign of a true EEH error. If it is not, processing continues as normal. The grand total number of these false alarms or "false positives" can be seen in /proc/ppc64/eeh (subject to change). Normally, almost all of these occur during boot, when the PCI bus is scanned, where a large number of 0xff reads are part of the bus scan procedure.

If a frozen slot is detected, code in arch/ppc64/kernel/eeh.c will print a stack trace to syslog (/var/log/messages). This stack trace has proven to be very useful to device-driver authors for finding out at what point the EEH error was detected, as the error itself usually occurs slightly beforehand.

Next, it uses the Linux kernel notifier chain/work queue mechanism to allow any interested parties to find out about the failure. Device drivers, or other parts of the kernel, can use eeh_register_notifier(struct notifier_block *) to find out about EEH events. The event will include a pointer to the pci device, the device node and some state info. Receivers of the event can "do as they wish"; the default handler will be described further in this section.

To assist in the recovery of the device, eeh.c exports the following functions:

rtas_set_slot_reset() -- assert the PCI #RST line for 1/8th of a second rtas_configure_bridge() -- ask firmware to configure any PCI bridges
located topologically under the pci slot.
eeh_save_bars() and eeh_restore_bars(): save and restore the PCI
config-space info for a device and any devices under it.

A handler for the EEH notifier_block events is implemented in drivers/pci/hotplug/pSeries_pci.c, called handle_eeh_events(). It saves the device BAR's and then calls rpaphp_unconfig_pci_adapter(). This last call causes the device driver for the card to be stopped, which causes hotplug events to go out to user space. This triggers user-space scripts that might issue commands such as "ifdown eth0" for ethernet cards, and so on. This handler then sleeps for 5 seconds, hoping to give the user-space scripts enough time to complete. It then resets the PCI card, reconfigures the device BAR's, and any bridges underneath. It then calls rpaphp_enable_pci_slot(), which restarts the device driver and triggers more user-space events (for example, calling "ifup eth0" for ethernet cards).

Device Shutdown and User-Space Events
-------------------------------------
This section documents what happens when a pci slot is unconfigured, focusing on how the device driver gets shut down, and on how the events get delivered to user-space scripts.

Following is an example sequence of events that cause a device driver close function to be called during the first phase of an EEH reset. The following sequence is an example of the pcnet32 device driver.

    rpa_php_unconfig_pci_adapter (struct slot *)  // in rpaphp_pci.c
    {
      calls
      pci_remove_bus_device (struct pci_dev *) // in /drivers/pci/remove.c
      { 
        calls
        pci_destroy_dev (struct pci_dev *) 
        {
          calls 
          device_unregister (&dev->dev) // in /drivers/base/core.c
          {
            calls
            device_del (struct device *)
            {
              calls 
              bus_remove_device() // in /drivers/base/bus.c
              {
                calls 
                device_release_driver()
                {
                  calls 
                  struct device_driver->remove() which is just
                  pci_device_remove()  // in /drivers/pci/pci_driver.c
                  {
                    calls
                    struct pci_driver->remove() which is just
                    pcnet32_remove_one() // in /drivers/net/pcnet32.c  
                    {
                      calls
                      unregister_netdev() // in /net/core/dev.c
                      {
                        calls 
                        dev_close()  // in /net/core/dev.c
                        { 
                           calls dev->stop();
                           which is just pcnet32_close() // in pcnet32.c
                           {
                             which does what you wanted
                             to stop the device
                           }
                        }
                     }
                   which
                   frees pcnet32 device driver memory
                }
     }}}}}}


    in drivers/pci/pci_driver.c, 
    struct device_driver->remove() is just pci_device_remove() 
    which calls struct pci_driver->remove() which is pcnet32_remove_one()
    which calls unregister_netdev()  (in net/core/dev.c)
    which calls dev_close()  (in net/core/dev.c) 
    which calls dev->stop() which is pcnet32_close() 
    which then does the appropriate shutdown. 

---
Following is the analogous stack trace for events sent to user-space when the pci device is unconfigured.

rpa_php_unconfig_pci_adapter() { // in rpaphp_pci.c
calls
pci_remove_bus_device (struct pci_dev *) { // in /drivers/pci/remove.c
calls
pci_destroy_dev (struct pci_dev *) {
calls
device_unregister (&dev->dev) { // in /drivers/base/core.c
calls
device_del(struct device * dev) { // in /drivers/base/core.c
calls
kobject_del() { //in /libs/kobject.c
calls
kobject_hotplug() { // in /libs/kobject.c
calls
kset_hotplug() { // in /lib/kobject.c
calls
kset->hotplug_ops->hotplug() which is really just a call to
dev_hotplug() { // in /drivers/base/core.c
calls
dev->bus->hotplug() which is really just a call to
pci_hotplug () { // in drivers/pci/hotplug.c
which prints device name, etc....
}
}
then kset_hotplug() calls
call_usermodehelper () with
argv[0]=hotplug_path[] which is "/sbin/hotplug"
--> event to userspace,
}
}
kobject_del() then calls sysfs_remove_dir(), which would trigger any user-space daemon that was watching /sysfs, and notice the delete event.

Pro's and Con's of the Current Design
-------------------------------------
There are several issues with the current EEH software recovery design, which may be addressed in future revisions. But first, note that the big plus of the current design is that no changes need to be made to individual device drivers, so that the current design throws a wide net. The biggest negative of the design is that it potentially disturbs network daemons and file systems that didn't need to be disturbed.

-- A minor complaint is that resetting the network card causes
user-space back-to-back ifdown/ifup burps that potentially disturb network daemons, that didn't need to even know that the pci card was being rebooted.

-- A more serious concern is that the same reset, for SCSI devices,
causes havoc to mounted file systems. Scripts cannot post-facto unmount a file system without flushing pending buffers, but this is impossible, because I/O has already been stopped. Thus, ideally, the reset should happen at or below the block layer, so that the file systems are not disturbed.

Reiserfs does not tolerate errors returned from the block device. Ext3fs seems to be tolerant, retrying reads/writes until it does succeed. Both have been only lightly tested in this scenario.

The SCSI-generic subsystem already has built-in code for performing SCSI device resets, SCSI bus resets, and SCSI host-bus-adapter (HBA) resets. These are cascaded into a chain of attempted resets if a SCSI command fails. These are completely hidden from the block layer. It would be very natural to add an EEH reset into this chain of events.

-- If a SCSI error occurs for the root device, all is lost unless
the sysadmin had the foresight to run /bin, /sbin, /etc, /var and so on, out of ramdisk/tmpfs.

Conclusions
-----------
There's forward progress ...

#2 lore raw
Quoted 4 lines · Linas Vepstas

p.s. It was not clear to me if the EEH patch previously sent (6 January 2005, same subject line) will be wending its way into the main Torvalds kernel tree, or not. I hadn't really gotten confirmation one way or another.

I'm not really totally happy with it yet, on a number of fronts:

1. You're adding more PCI-specific stuff to the device_node struct, which I don't like. I would prefer that the device_node tree contains basically just what we get from OF, and that we have a separate struct for storing ppc64-specific information for each PCI device. Fixing that is outside the scope of your patch, though.

2. I don't see why the device nodes for the PCI subtree being reset would go away, and thus I don't see the need for your eeh_cfg_tree struct.

3. Is there a good reason why we can't use the assigned-addresses property on the relevant device tree nodes to tell us what to set the BARs to?

4. I think the 5 second sleep is quite bogus, and shows that we have the flow of control wrong. In particular I think it should be a userland write to a sysfs file that kicks off the restart process rather than it just happening after 5 seconds. Anyway, what process or thread is executing that 5 second sleep? Is it keventd or something?

5. AFAICS userland will get an unplug notification for the device, but nothing to indicate that is due to an EEH slot isolation event. I think userland should be told about EEH events.

Regards,
Paul.

#3 lore raw
Quoted 4 lines · Paul Mackerras

5. AFAICS userland will get an unplug notification for the device, but nothing to indicate that is due to an EEH slot isolation event. I think userland should be told about EEH events.

Currently there is a way for userland to determine if a hotplug event they receive is due to an EEH slot isolation event. It's not very pretty and requires the rtas_errd daemon to be running.

The RTAS event generated from the EEH event is logged to /var/log/platform by rtas_errd. Userland scripts would have to search the file for a recent EEH event matching their device to make this determination. This isn't as nice as a direct notification but is what we have at this point.

Signature · 1 line
Nathan Fontenot
#4 lore raw
In reply to Paul Mackerras

On Wed, Jan 19, 2005 at 05:06:05PM +1100, Paul Mackerras was heard to remark:

Quoted 7 lines
Quoted 4 lines · Linas Vepstas

p.s. It was not clear to me if the EEH patch previously sent (6 January 2005, same subject line) will be wending its way into the main Torvalds kernel tree, or not. I hadn't really gotten confirmation one way or another.

Quoted 7 lines

I'm not really totally happy with it yet, on a number of fronts:

1. You're adding more PCI-specific stuff to the device_node struct, which I don't like. I would prefer that the device_node tree contains basically just what we get from OF, and that we have a separate struct for storing ppc64-specific information for each PCI device. Fixing that is outside the scope of your patch, though.

I wrote this down on my to-do list. Its the sort of thing that evaporates from my consciousness when other things come along, but I'll give it a shot.

Quoted 3 lines

2. I don't see why the device nodes for the PCI subtree being reset would go away, and thus I don't see the need for your eeh_cfg_tree struct.

Its not the reset, its the hot-plug remove. The hot plug code assumes that you are going to physically remove the device from the slot, so it removes the device_node as part of the "unconfig".

Of course, I found this out only after performing a null-pointer deref. Note only does the node go away, but all of the various pointers it holds are zeroed in the process.

The cfg tree holds on to those pointers, so that I wouldn't have to muck with the device_node removal code to do something tricky.

Quoted 3 lines

3. Is there a good reason why we can't use the assigned-addresses property on the relevant device tree nodes to tell us what to set the BARs to?

Yes, the reason is that after a reset, that property doesn't hold any
decent data. I discussed this with the firmware developers, and thier
response was that it is the kernel's responsibility to compute
(or save/restore) such values. (Except for bridges, which they will do for us).

Quoted 2 lines

4. I think the 5 second sleep is quite bogus, and shows that we have the flow of control wrong.

:) Yes, well, indeed it is. Don't look at me, not my idea.

Quoted 5 lines

In particular I think it should be a
userland write to a sysfs file that kicks off the restart process rather than it just happening after 5 seconds. Anyway, what process or thread is executing that 5 second sleep? Is it keventd or something?

Its a workqueue.

Quoted 3 lines

5. AFAICS userland will get an unplug notification for the device, but nothing to indicate that is due to an EEH slot isolation event. I think userland should be told about EEH events.

In principle, I'd agree. In practice, this would seem to require changes or additions or enhancements to udev that I don't quite understand, as well as potential changes to udev scripts. Maybe I don't understand sysfs sufficiently well. I am very tempted to punt on this, and wait for the Intel-backed PCI-E code to get to this point, and then do whatever they're doing.

--linas

#5 lore raw
Quoted 7 lines · Linas Vepstas
Quoted 3 lines

2. I don't see why the device nodes for the PCI subtree being reset would go away, and thus I don't see the need for your eeh_cfg_tree struct.

Its not the reset, its the hot-plug remove. The hot plug code assumes that you are going to physically remove the device from the slot, so it removes the device_node as part of the "unconfig".

OK, I missed that. It seems a bit bogus to me. Could you point me at where in the code this happens?

Quoted 8 lines
Quoted 3 lines

3. Is there a good reason why we can't use the assigned-addresses property on the relevant device tree nodes to tell us what to set the BARs to?

Yes, the reason is that after a reset, that property doesn't hold any
decent data. I discussed this with the firmware developers, and thier
response was that it is the kernel's responsibility to compute
(or save/restore) such values. (Except for bridges, which they will do for us).

The not holding any decent data is a consequence of the device nodes getting thrown away, isn't it? I fail to see how resetting the device can of itself affect our copy of the device tree.

Quoted 7 lines
Quoted 5 lines

In particular I think it should be a
userland write to a sysfs file that kicks off the restart process rather than it just happening after 5 seconds. Anyway, what process or thread is executing that 5 second sleep? Is it keventd or something?

Its a workqueue.

Which get run in keventd's context. In other words no other workqueues will get run during the 5 second sleep, or at least not on that cpu.

Paul.

#6 lore raw
In reply to Paul Mackerras

On Wed, Jan 19, 2005 at 05:06:05PM +1100, Paul Mackerras was heard to remark:

Quoted 8 lines
Quoted 4 lines · Linas Vepstas

p.s. It was not clear to me if the EEH patch previously sent (6 January 2005, same subject line) will be wending its way into the main Torvalds kernel tree, or not. I hadn't really gotten confirmation one way or another.

I'm not really totally happy with it yet, on a number of fronts:

[...]

I forgot to mention: while I agree with some/many of these points, especially with regards to recovery, I'd also like to note that the patch was mailed in two independent parts:

-- a number of generic infrastructure routines, all in a ppc64 patch, and -- the code that actually performs the recovery, as a patch to
the drivers/pci/hotplug subsystem.

While the actual recovery code is controversial (e.g. no support of scsi recovery), I'd like to at least get in the the generic infrastructure pieces.

--linas