1========================== 2PCI Bus EEH Error Recovery 3========================== 4 5Linas Vepstas <linas@austin.ibm.com> 6 712 January 2005 8 9 10Overview: 11--------- 12The IBM POWER-based pSeries and iSeries computers include PCI bus 13controller chips that have extended capabilities for detecting and 14reporting a large variety of PCI bus error conditions. These features 15go under the name of "EEH", for "Enhanced Error Handling". The EEH 16hardware features allow PCI bus errors to be cleared and a PCI 17card to be "rebooted", without also having to reboot the operating 18system. 19 20This is in contrast to traditional PCI error handling, where the 21PCI chip is wired directly to the CPU, and an error would cause 22a CPU machine-check/check-stop condition, halting the CPU entirely. 23Another "traditional" technique is to ignore such errors, which 24can lead to data corruption, both of user data or of kernel data, 25hung/unresponsive adapters, or system crashes/lockups. Thus, 26the idea behind EEH is that the operating system can become more 27reliable and robust by protecting it from PCI errors, and giving 28the OS the ability to "reboot"/recover individual PCI devices. 29 30Future systems from other vendors, based on the PCI-E specification, 31may contain similar features. 32 33 34Causes of EEH Errors 35-------------------- 36EEH was originally designed to guard against hardware failure, such 37as PCI cards dying from heat, humidity, dust, vibration and bad 38electrical connections. The vast majority of EEH errors seen in 39"real life" are due to either poorly seated PCI cards, or, 40unfortunately quite commonly, due to device driver bugs, device firmware 41bugs, and sometimes PCI card hardware bugs. 42 43The most common software bug, is one that causes the device to 44attempt to DMA to a location in system memory that has not been 45reserved for DMA access for that card. This is a powerful feature, 46as it prevents what; otherwise, would have been silent memory 47corruption caused by the bad DMA. A number of device driver 48bugs have been found and fixed in this way over the past few 49years. Other possible causes of EEH errors include data or 50address line parity errors (for example, due to poor electrical 51connectivity due to a poorly seated card), and PCI-X split-completion 52errors (due to software, device firmware, or device PCI hardware bugs). 53The vast majority of "true hardware failures" can be cured by 54physically removing and re-seating the PCI card. 55 56 57Detection and Recovery 58---------------------- 59In the following discussion, a generic overview of how to detect 60and recover from EEH errors will be presented. This is followed 61by an overview of how the current implementation in the Linux 62kernel does it. The actual implementation is subject to change, 63and some of the finer points are still being debated. These 64may in turn be swayed if or when other architectures implement 65similar functionality. 66 67When a PCI Host Bridge (PHB, the bus controller connecting the 68PCI bus to the system CPU electronics complex) detects a PCI error 69condition, it will "isolate" the affected PCI card. Isolation 70will block all writes (either to the card from the system, or 71from the card to the system), and it will cause all reads to 72return all-ff's (0xff, 0xffff, 0xffffffff for 8/16/32-bit reads). 73This value was chosen because it is the same value you would 74get if the device was physically unplugged from the slot. 75This includes access to PCI memory, I/O space, and PCI config 76space. Interrupts; however, will continued to be delivered. 77 78Detection and recovery are performed with the aid of ppc64 79firmware. The programming interfaces in the Linux kernel 80into the firmware are referred to as RTAS (Run-Time Abstraction 81Services). The Linux kernel does not (should not) access 82the EEH function in the PCI chipsets directly, primarily because 83there are a number of different chipsets out there, each with 84different interfaces and quirks. The firmware provides a 85uniform abstraction layer that will work with all pSeries 86and iSeries hardware (and be forwards-compatible). 87 88If the OS or device driver suspects that a PCI slot has been 89EEH-isolated, there is a firmware call it can make to determine if 90this is the case. If so, then the device driver should put itself 91into a consistent state (given that it won't be able to complete any 92pending work) and start recovery of the card. Recovery normally 93would consist of resetting the PCI device (holding the PCI #RST 94line high for two seconds), followed by setting up the device 95config space (the base address registers (BAR's), latency timer, 96cache line size, interrupt line, and so on). This is followed by a 97reinitialization of the device driver. In a worst-case scenario, 98the power to the card can be toggled, at least on hot-plug-capable 99slots. In principle, layers far above the device driver probably 100do not need to know that the PCI card has been "rebooted" in this 101way; ideally, there should be at most a pause in Ethernet/disk/USB 102I/O while the card is being reset. 103 104If the card cannot be recovered after three or four resets, the 105kernel/device driver should assume the worst-case scenario, that the 106card has died completely, and report this error to the sysadmin. 107In addition, error messages are reported through RTAS and also through 108syslogd (/var/log/messages) to alert the sysadmin of PCI resets. 109The correct way to deal with failed adapters is to use the standard 110PCI hotplug tools to remove and replace the dead card. 111 112 113Current PPC64 Linux EEH Implementation 114-------------------------------------- 115At this time, a generic EEH recovery mechanism has been implemented, 116so that individual device drivers do not need to be modified to support 117EEH recovery. This generic mechanism piggy-backs on the PCI hotplug 118infrastructure, and percolates events up through the userspace/udev 119infrastructure. Following is a detailed description of how this is 120accomplished. 121 122EEH must be enabled in the PHB's very early during the boot process, 123and if a PCI slot is hot-plugged. The former is performed by 124eeh_init() in arch/powerpc/platforms/pseries/eeh.c, and the later by 125drivers/pci/hotplug/pSeries_pci.c calling in to the eeh.c code. 126EEH must be enabled before a PCI scan of the device can proceed. 127Current Power5 hardware will not work unless EEH is enabled; 128although older Power4 can run with it disabled. Effectively, 129EEH can no longer be turned off. PCI devices *must* be 130registered with the EEH code; the EEH code needs to know about 131the I/O address ranges of the PCI device in order to detect an 132error. Given an arbitrary address, the routine 133pci_get_device_by_addr() will find the pci device associated 134with that address (if any). 135 136The default arch/powerpc/include/asm/io.h macros readb(), inb(), insb(), 137etc. include a check to see if the i/o read returned all-0xff's. 138If so, these make a call to eeh_dn_check_failure(), which in turn 139asks the firmware if the all-ff's value is the sign of a true EEH 140error. If it is not, processing continues as normal. The grand 141total number of these false alarms or "false positives" can be 142seen in /proc/ppc64/eeh (subject to change). Normally, almost 143all of these occur during boot, when the PCI bus is scanned, where 144a large number of 0xff reads are part of the bus scan procedure. 145 146If a frozen slot is detected, code in 147arch/powerpc/platforms/pseries/eeh.c will print a stack trace to 148syslog (/var/log/messages). This stack trace has proven to be very 149useful to device-driver authors for finding out at what point the EEH 150error was detected, as the error itself usually occurs slightly 151beforehand. 152 153Next, it uses the Linux kernel notifier chain/work queue mechanism to 154allow any interested parties to find out about the failure. Device 155drivers, or other parts of the kernel, can use 156`eeh_register_notifier(struct notifier_block *)` to find out about EEH 157events. The event will include a pointer to the pci device, the 158device node and some state info. Receivers of the event can "do as 159they wish"; the default handler will be described further in this 160section. 161 162To assist in the recovery of the device, eeh.c exports the 163following functions: 164 165rtas_set_slot_reset() 166 assert the PCI #RST line for 1/8th of a second 167rtas_configure_bridge() 168 ask firmware to configure any PCI bridges 169 located topologically under the pci slot. 170eeh_save_bars() and eeh_restore_bars(): 171 save and restore the PCI 172 config-space info for a device and any devices under it. 173 174 175A handler for the EEH notifier_block events is implemented in 176drivers/pci/hotplug/pSeries_pci.c, called handle_eeh_events(). 177It saves the device BAR's and then calls rpaphp_unconfig_pci_adapter(). 178This last call causes the device driver for the card to be stopped, 179which causes uevents to go out to user space. This triggers 180user-space scripts that might issue commands such as "ifdown eth0" 181for ethernet cards, and so on. This handler then sleeps for 5 seconds, 182hoping to give the user-space scripts enough time to complete. 183It then resets the PCI card, reconfigures the device BAR's, and 184any bridges underneath. It then calls rpaphp_enable_pci_slot(), 185which restarts the device driver and triggers more user-space 186events (for example, calling "ifup eth0" for ethernet cards). 187 188 189Device Shutdown and User-Space Events 190------------------------------------- 191This section documents what happens when a pci slot is unconfigured, 192focusing on how the device driver gets shut down, and on how the 193events get delivered to user-space scripts. 194 195Following is an example sequence of events that cause a device driver 196close function to be called during the first phase of an EEH reset. 197The following sequence is an example of the pcnet32 device driver:: 198 199 rpa_php_unconfig_pci_adapter (struct slot *) // in rpaphp_pci.c 200 { 201 calls 202 pci_remove_bus_device (struct pci_dev *) // in /drivers/pci/remove.c 203 { 204 calls 205 pci_destroy_dev (struct pci_dev *) 206 { 207 calls 208 device_unregister (&dev->dev) // in /drivers/base/core.c 209 { 210 calls 211 device_del (struct device *) 212 { 213 calls 214 bus_remove_device() // in /drivers/base/bus.c 215 { 216 calls 217 device_release_driver() 218 { 219 calls 220 struct device_driver->remove() which is just 221 pci_device_remove() // in /drivers/pci/pci_driver.c 222 { 223 calls 224 struct pci_driver->remove() which is just 225 pcnet32_remove_one() // in /drivers/net/pcnet32.c 226 { 227 calls 228 unregister_netdev() // in /net/core/dev.c 229 { 230 calls 231 dev_close() // in /net/core/dev.c 232 { 233 calls dev->stop(); 234 which is just pcnet32_close() // in pcnet32.c 235 { 236 which does what you wanted 237 to stop the device 238 } 239 } 240 } 241 which 242 frees pcnet32 device driver memory 243 } 244 }}}}}} 245 246 247in drivers/pci/pci_driver.c, 248struct device_driver->remove() is just pci_device_remove() 249which calls struct pci_driver->remove() which is pcnet32_remove_one() 250which calls unregister_netdev() (in net/core/dev.c) 251which calls dev_close() (in net/core/dev.c) 252which calls dev->stop() which is pcnet32_close() 253which then does the appropriate shutdown. 254 255--- 256 257Following is the analogous stack trace for events sent to user-space 258when the pci device is unconfigured:: 259 260 rpa_php_unconfig_pci_adapter() { // in rpaphp_pci.c 261 calls 262 pci_remove_bus_device (struct pci_dev *) { // in /drivers/pci/remove.c 263 calls 264 pci_destroy_dev (struct pci_dev *) { 265 calls 266 device_unregister (&dev->dev) { // in /drivers/base/core.c 267 calls 268 device_del(struct device * dev) { // in /drivers/base/core.c 269 calls 270 kobject_del() { //in /libs/kobject.c 271 calls 272 kobject_uevent() { // in /libs/kobject.c 273 calls 274 kset_uevent() { // in /lib/kobject.c 275 calls 276 kset->uevent_ops->uevent() // which is really just 277 a call to 278 dev_uevent() { // in /drivers/base/core.c 279 calls 280 dev->bus->uevent() which is really just a call to 281 pci_uevent () { // in drivers/pci/hotplug.c 282 which prints device name, etc.... 283 } 284 } 285 then kobject_uevent() sends a netlink uevent to userspace 286 --> userspace uevent 287 (during early boot, nobody listens to netlink events and 288 kobject_uevent() executes uevent_helper[], which runs the 289 event process /sbin/hotplug) 290 } 291 } 292 kobject_del() then calls sysfs_remove_dir(), which would 293 trigger any user-space daemon that was watching /sysfs, 294 and notice the delete event. 295 296 297Pro's and Con's of the Current Design 298------------------------------------- 299There are several issues with the current EEH software recovery design, 300which may be addressed in future revisions. But first, note that the 301big plus of the current design is that no changes need to be made to 302individual device drivers, so that the current design throws a wide net. 303The biggest negative of the design is that it potentially disturbs 304network daemons and file systems that didn't need to be disturbed. 305 306- A minor complaint is that resetting the network card causes 307 user-space back-to-back ifdown/ifup burps that potentially disturb 308 network daemons, that didn't need to even know that the pci 309 card was being rebooted. 310 311- A more serious concern is that the same reset, for SCSI devices, 312 causes havoc to mounted file systems. Scripts cannot post-facto 313 unmount a file system without flushing pending buffers, but this 314 is impossible, because I/O has already been stopped. Thus, 315 ideally, the reset should happen at or below the block layer, 316 so that the file systems are not disturbed. 317 318 Reiserfs does not tolerate errors returned from the block device. 319 Ext3fs seems to be tolerant, retrying reads/writes until it does 320 succeed. Both have been only lightly tested in this scenario. 321 322 The SCSI-generic subsystem already has built-in code for performing 323 SCSI device resets, SCSI bus resets, and SCSI host-bus-adapter 324 (HBA) resets. These are cascaded into a chain of attempted 325 resets if a SCSI command fails. These are completely hidden 326 from the block layer. It would be very natural to add an EEH 327 reset into this chain of events. 328 329- If a SCSI error occurs for the root device, all is lost unless 330 the sysadmin had the foresight to run /bin, /sbin, /etc, /var 331 and so on, out of ramdisk/tmpfs. 332 333 334Conclusions 335----------- 336There's forward progress ... 337