Asynchronous EEH recovery
Subsystems (5)
cover Asynchronous EEH recovery
Hi,
EEH reocvery is currently serialized and these patches shorten the time taken for EEH recovery by making the recovery to run in parallel. The original author of these patches is Sam Bobroff, I have rebased and tested these patches.
On powervm with 64 VFs and I see approximately 48% reduction in time taken in EEH recovery, Yet to be tested on powernv.
These patches were originally posted as separate RFCs, I think posting them as single series would be more helpful, I know the patches are too big, I will try to logically divide in next iterations.
Thanks
Ganesh Goudar (3):
powerpc/eeh: Synchronization for safety powerpc/eeh: Provide a unique ID for each EEH recovery powerpc/eeh: Asynchronous recovery
arch/powerpc/include/asm/eeh.h | 7 +- arch/powerpc/include/asm/eeh_event.h | 10 +- arch/powerpc/include/asm/pci-bridge.h | 3 + arch/powerpc/include/asm/ppc-pci.h | 2 +- arch/powerpc/kernel/eeh.c | 154 +++-- arch/powerpc/kernel/eeh_driver.c | 578 +++++++++++++++---- arch/powerpc/kernel/eeh_event.c | 71 ++- arch/powerpc/kernel/eeh_pe.c | 33 +- arch/powerpc/platforms/powernv/eeh-powernv.c | 12 +- arch/powerpc/platforms/pseries/eeh_pseries.c | 5 +- arch/powerpc/platforms/pseries/pci_dlpar.c | 5 +- drivers/pci/hotplug/pnv_php.c | 5 +- drivers/pci/hotplug/rpadlpar_core.c | 2 + drivers/vfio/vfio_spapr_eeh.c | 10 +- 14 files changed, 685 insertions(+), 212 deletions(-)
Signature · 1 line
2.37.1
RFC 1/3 powerpc/eeh: Synchronization for safety
Based on the original work from Sam Bobroff.
There is currently little synchronization between EEH error detection (eeh_dev_check_failure()), EEH error recovery (eeh_handle_{normal,special}_event()) and the PCI subsystem (device addition and removal), and so there are race conditions that lead to crashes (often access to free'd memory or LIST_POISON).
However, a solution must consider:
- EEH error detection can occur in interrupt context, which prevents the use of a mutex.
- EEH recovery may need to sleep, which prevents the use of a spinlock.
- EEH recovery uses PCI operations that may require the PCI rescan/remove lock and/or device lock to be held
- PCI operations may hold the rescan/remove and/or device lock when calling into EEH functions.
- Device driver callbacks may perform arbitrary PCI operations during recovery, including device removal.
In this patch the existing mutex and spinlock are combined with the EEH_PE_RECOVERING flag to provide some assurances that are then used to reduce the race conditions.
The fields to be protected are the ones that provide the structure of the trees of struct eeh_pe that are held for each PHB: the parent pointer and child lists and the list of struct eeh_dev, as well as the pe and pdev pointers within struct eeh_dev.
The existing way of using EEH_PE_RECOVERING is kept and slightly
extended: No struct eeh_pe will be removed while it has the flag set on it. Additionally, when adding new PEs, they are marked EEH_PE_RECOVERING if their parent PE is marked: this allows the recovery thread to assume that all PEs underneath the one it's processing will continue to exist during recovery.
Both the mutex and spinlock are held while any protected field is changed or a PE is deleted, so holding either of them (elsewhere) will keep them stable and safe to access. Additionally, if EEH_PE_RECOVERING is set on a PE then the locks can be released and re-acquired safely, as long as the protected fields aren't used while no locks are held. This is used during recovery to release locks for long sleeps (i.e. during eeh_wait_state() when we may sleep up to 5 minutes), or to maintain lock ordering.
The spinlock is used in error detection (which cannot use a mutex, see above) and also where it's possible that the mutex is already held. The mutex is used in areas that don't have that restriction, and where blocking may be required. Care must be taken when ordering these locks against the PCI rescan/remove lock and the device locks to avoid deadlocking.
--- arch/powerpc/include/asm/eeh.h | 6 +- arch/powerpc/kernel/eeh.c | 112 ++++++-- arch/powerpc/kernel/eeh_driver.c | 287 ++++++++++++++----- arch/powerpc/kernel/eeh_pe.c | 30 +- arch/powerpc/platforms/powernv/eeh-powernv.c | 12 +- arch/powerpc/platforms/pseries/eeh_pseries.c | 5 +- arch/powerpc/platforms/pseries/pci_dlpar.c | 5 +- drivers/pci/hotplug/pnv_php.c | 5 +- drivers/pci/hotplug/rpadlpar_core.c | 2 + drivers/vfio/vfio_spapr_eeh.c | 10 +- 10 files changed, 364 insertions(+), 110 deletions(-)
arch/powerpc/include/asm/eeh.h +5 -1
@@ -271,11 +271,15 @@ static inline bool eeh_state_active(int state) == (EEH_STATE_MMIO_ACTIVE | EEH_STATE_DMA_ACTIVE); } +void eeh_recovery_lock (void ); +void eeh_recovery_unlock (void ); +void eeh_recovery_must_be_locked (void ); +typedef void (*eeh_edev_traverse_func )(struct eeh_dev *edev,void *flag);typedef void *(*eeh_pe_traverse_func )(struct eeh_pe *pe,void *flag);void eeh_set_pe_aux_size (int size);int eeh_phb_pe_create (struct pci_controller *phb); -int eeh_wait_state (struct eeh_pe *pe,int max_wait); +int eeh_wait_state (struct eeh_pe *pe,int max_waiti,bool unlock);struct eeh_pe *eeh_phb_pe_get (struct pci_controller *phb);struct eeh_pe *eeh_pe_next (struct eeh_pe *pe,struct eeh_pe *root);struct eeh_pe *eeh_pe_get (struct pci_controller *phb,int pe_no);
arch/powerpc/kernel/eeh.c +84 -28
334 lines not on this page · raw message
arch/powerpc/kernel/eeh_driver.c +217 -70
528 lines not on this page · raw message
arch/powerpc/kernel/eeh_pe.c +28 -2
@@ -103,8 +103,15 @@ int eeh_phb_pe_create(struct pci_controller *phb)* * Wait for the state of associated PE. It might take some time * to retrieve the PE's state. +* +* Note that if this function sleeps and unlock is true, the EEH recovery lock +* will be released and re-acquired. +* It is only safe to do this when the PE has the recovering flag set on it. +* This is important because the sleep may be +* very long (300 seconds) and device removal will be blocked by the recovery +* mutex. See eeh_device_notifier(). */ -int eeh_wait_state (struct eeh_pe *pe,int max_wait) +int eeh_wait_state (struct eeh_pe *pe,int max_wait,bool unlock) {int ret;int mwait;
@@ -120,6 +127,8 @@ int eeh_wait_state(struct eeh_pe *pe, int max_wait)#define EEH_STATE_MIN_WAIT_TIME (1000)#define EEH_STATE_MAX_WAIT_TIME (300 * 1000) +WARN_ON_ONCE (unlock && !(pe->state &EEH_PE_RECOVERING )); +while (1 ) { ret = eeh_ops->get_state (pe, &mwait);
@@ -141,8 +150,11 @@ int eeh_wait_state(struct eeh_pe *pe, int max_wait) __func__, mwait); mwait =EEH_STATE_MAX_WAIT_TIME ; } - +if (unlock) +eeh_recovery_unlock ();msleep (min (mwait, max_wait)); +if (unlock) +eeh_recovery_lock (); max_wait -= mwait; } }
@@ -308,6 +320,7 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent) {struct pci_controller *hose = edev->controller;struct eeh_pe *pe, *parent; +unsigned long flags;/* * Search the PE has been existing or not according
@@ -315,6 +328,8 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent)* PE should be composed of PCI bus and its subordinate * components. */ +eeh_recovery_must_be_locked (); +eeh_serialize_lock (&flags); pe =eeh_pe_get (hose, edev->pe_config_addr);if (pe) {if (pe->type &EEH_PE_INVALID ) {
@@ -343,8 +358,10 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent)list_add_tail (&edev->entry, &pe->edevs);eeh_edev_dbg (edev,"Added to bus PE\n" ); } +eeh_serialize_unlock (flags);return 0 ; } +eeh_serialize_unlock (flags);/* Create a new EEH PE */ if (edev->physfn)
@@ -364,6 +381,7 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent)* to PHB directly. Otherwise, we have to associate the * PE with its parent. */ +eeh_serialize_lock (&flags);if (!new_pe_parent) { new_pe_parent =eeh_phb_pe_get (hose);if (!new_pe_parent) {
@@ -371,6 +389,7 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent) __func__, hose->global_number); edev->pe =NULL ;kfree (pe); +eeh_serialize_unlock (flags);return -EEXIST ; } }
@@ -385,6 +404,7 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent)*/ list_add_tail (&edev->entry, &pe->edevs); edev->pe = pe; +eeh_serialize_unlock (flags);eeh_edev_dbg (edev,"Added to new (parent: PE#%x)\n" , new_pe_parent->addr);
@@ -402,13 +422,18 @@ int eeh_pe_tree_insert(struct eeh_dev *edev, struct eeh_pe *new_pe_parent)*/ int eeh_pe_tree_remove (struct eeh_dev *edev) { +unsigned long flags;struct eeh_pe *pe, *parent, *child;bool keep, recover;int cnt; +/* Both locks are required to make changes */ +eeh_recovery_must_be_locked (); +eeh_serialize_lock (&flags); pe =eeh_dev_to_pe (edev);if (!pe) {eeh_edev_dbg (edev,"No PE found for device.\n" ); +eeh_serialize_unlock (flags);return -EEXIST ; }
@@ -475,6 +500,7 @@ int eeh_pe_tree_remove(struct eeh_dev *edev) pe = parent; } +eeh_serialize_unlock (flags);return 0 ; }
arch/powerpc/platforms/powernv/eeh-powernv.c +10 -2
@@ -72,6 +72,7 @@ static ssize_t pnv_eeh_ei_write(struct file *filp,unsigned long addr, mask;char buf[50 ];int ret; +unsigned long flags;if (!eeh_ops || !eeh_ops->err_inject)return -ENXIO ;
@@ -86,14 +87,21 @@ static ssize_t pnv_eeh_ei_write(struct file *filp, &pe_no, &type, &func, &addr, &mask);if (ret !=5 )return -EINVAL ; - +/* +* Use the spinlock rather than the mutex so that errors can be +* injected during slow recovery operations (for testing). +*/ +eeh_serialize_lock (&flags);/* Retrieve PE */ pe =eeh_pe_get (hose, pe_no); -if (!pe) +if (!pe) { +eeh_serialize_unlock (flags);return -ENODEV ; + }/* Do error injection */ ret = eeh_ops->err_inject (pe, type, func, addr, mask); +eeh_serialize_unlock (flags);return ret <0 ? ret : count; }
arch/powerpc/platforms/pseries/eeh_pseries.c +4 -1
@@ -47,11 +47,13 @@ static void pseries_eeh_init_edev(struct pci_dn *pdn);static void pseries_pcibios_bus_add_device (struct pci_dev *pdev) { -struct pci_dn *pdn =pci_get_pdn (pdev); +struct pci_dn *pdn;if (eeh_has_flag (EEH_FORCE_DISABLED ))return ; +eeh_recovery_lock (); + pdn =pci_get_pdn (pdev);dev_dbg (&pdev->dev,"EEH: Setting up device\n" );#ifdef CONFIG_PCI_IOV if (pdev->is_virtfn) {
@@ -82,6 +84,7 @@ static void pseries_pcibios_bus_add_device(struct pci_dev *pdev) }#endif eeh_probe_device (pdev); +eeh_recovery_unlock (); }
arch/powerpc/platforms/pseries/pci_dlpar.c +4 -1
@@ -38,8 +38,11 @@ struct pci_controller *init_phb_dynamic(struct device_node *dn)/* Create EEH devices for the PHB */ eeh_phb_pe_create (phb); -if (dn->child) +if (dn->child) { +eeh_recovery_lock ();pseries_eeh_init_edev_recursive (PCI_DN (dn)); +eeh_recovery_unlock (); + }pcibios_scan_phb (phb);pcibios_finish_adding_to_bus (phb->bus);
drivers/pci/hotplug/pnv_php.c +2 -3
@@ -769,7 +769,6 @@ static irqreturn_t pnv_php_interrupt(int irq, void *data)u16 sts, lsts;u8 presence;bool added; -unsigned long flags;int ret;pcie_capability_read_word (pdev,PCI_EXP_SLTSTA , &sts);
@@ -807,10 +806,10 @@ static irqreturn_t pnv_php_interrupt(int irq, void *data) edev = pchild ?pci_dev_to_eeh_dev (pchild) :NULL ; pe = edev ? edev->pe :NULL ;if (pe) { -eeh_serialize_lock (&flags); +eeh_recovery_lock ();eeh_pe_mark_isolated (pe); -eeh_serialize_unlock (flags);eeh_pe_set_option (pe,EEH_OPT_FREEZE_PE ); +eeh_recovery_unlock (); } }
drivers/pci/hotplug/rpadlpar_core.c +2 -0
@@ -141,7 +141,9 @@ static void dlpar_pci_add_bus(struct device_node *dn)struct pci_controller *phb = pdn->phb;struct pci_dev *dev =NULL ; +eeh_recovery_lock ();pseries_eeh_init_edev_recursive (pdn); +eeh_recovery_unlock ();/* Add EADS device to PHB bus, adding new entry to bus->devices */ dev =of_create_pci_dev (dn, phb->bus, pdn->devfn);
drivers/vfio/vfio_spapr_eeh.c +8 -2
@@ -54,6 +54,7 @@ long vfio_spapr_iommu_eeh_ioctl(struct iommu_group *group,if (op.argsz < minsz || op.flags)return -EINVAL ; +eeh_recovery_lock ();switch (op.op) {case VFIO_EEH_PE_DISABLE : ret =eeh_pe_set_option (pe,EEH_OPT_DISABLE );
@@ -84,10 +85,14 @@ long vfio_spapr_iommu_eeh_ioctl(struct iommu_group *group,break ;case VFIO_EEH_PE_INJECT_ERR : minsz = offsetofend(struct vfio_eeh_pe_op, err.mask); -if (op.argsz < minsz) +if (op.argsz < minsz) { +eeh_recovery_unlock ();return -EINVAL ; -if (copy_from_user (&op, (void __user *)arg, minsz)) + } +if (copy_from_user (&op, (void __user *)arg, minsz)) { +eeh_recovery_unlock ();return -EFAULT ; + } ret =eeh_pe_inject_err (pe, op.err.type, op.err.func, op.err.addr, op.err.mask);
@@ -95,6 +100,7 @@ long vfio_spapr_iommu_eeh_ioctl(struct iommu_group *group,default : ret = -EINVAL ; } +eeh_recovery_unlock (); }return ret;
Signature · 1 line
2.37.1
RFC 2/3 powerpc/eeh: Provide a unique ID for each EEH recovery
In reply to Ganesh GoudarBased on the original work from Sam Bobroff.
Give a unique ID to each recovery event, to ease log parsing and prepare for parallel recovery.
Also add some new messages with a very simple format that may be useful to log-parsers.
--- arch/powerpc/include/asm/eeh_event.h | 3 +- arch/powerpc/include/asm/ppc-pci.h | 2 +- arch/powerpc/kernel/eeh.c | 42 +++--- arch/powerpc/kernel/eeh_driver.c | 188 ++++++++++++++++----------- arch/powerpc/kernel/eeh_event.c | 12 +- 5 files changed, 146 insertions(+), 101 deletions(-)
arch/powerpc/include/asm/eeh_event.h +2 -1
@@ -17,13 +17,14 @@struct eeh_event {struct list_head list;/* to form event queue */ struct eeh_pe *pe;/* EEH PE */ +unsigned int id;/* Event ID */ };int eeh_event_init (void );int eeh_send_failure_event (struct eeh_pe *pe);int __eeh_send_failure_event (struct eeh_pe *pe);void eeh_remove_event (struct eeh_pe *pe,bool force); -void eeh_handle_normal_event (struct eeh_pe *pe); +void eeh_handle_normal_event (unsigned int event_id,struct eeh_pe *pe);void eeh_handle_special_event (void );#endif /* __KERNEL__ */
arch/powerpc/include/asm/ppc-pci.h +1 -1
@@ -40,7 +40,7 @@ extern int rtas_setup_phb(struct pci_controller *phb);void eeh_addr_cache_insert_dev (struct pci_dev *dev);void eeh_addr_cache_rmv_dev (struct pci_dev *dev);struct eeh_dev *eeh_addr_cache_get_dev (unsigned long addr); -void eeh_slot_error_detail (struct eeh_pe *pe,int severity); +void eeh_slot_error_detail (unsigned int event_id,struct eeh_pe *pe,int severity);int eeh_pci_enable (struct eeh_pe *pe,int function);int eeh_pe_reset_full (struct eeh_pe *pe,bool include_passed);void eeh_save_bars (struct eeh_dev *edev);
arch/powerpc/kernel/eeh.c +23 -19
122 lines not on this page · raw message
arch/powerpc/kernel/eeh_driver.c +109 -79
475 lines not on this page · raw message
arch/powerpc/kernel/eeh_event.c +11 -1
@@ -26,6 +26,9 @@ static DEFINE_SPINLOCK(eeh_eventlist_lock);static DECLARE_COMPLETION (eeh_eventlist_event);static LIST_HEAD (eeh_eventlist); +/* Event ID 0 is reserved for special events */ +static atomic_t eeh_event_id =ATOMIC_INIT (1 ); +/** * eeh_event_handler - Dispatch EEH events. * @dummy - unused
@@ -59,7 +62,7 @@ static int eeh_event_handler(void * dummy)/* We might have event without binding PE */ if (event->pe) -eeh_handle_normal_event (event->pe); +eeh_handle_normal_event (event->id, event->pe);else eeh_handle_special_event ();
@@ -110,6 +113,13 @@ int __eeh_send_failure_event(struct eeh_pe *pe)return -ENOMEM ; } event->pe = pe; +do { +/* Skip over the special value (0) */ + event->id = (unsigned int )atomic_inc_return (&eeh_event_id); + }while (!event->id); + +pr_err ("EEH(%u): EVENT=ERROR_DETECTED PHB=%#x PE=%#x\n" , + event->id, pe->phb->global_number, pe->addr);/* * Mark the PE as recovering before inserting it in the queue.
Signature · 1 line
2.37.1
RFC 3/3 powerpc/eeh: Asynchronous recovery
In reply to Ganesh GoudarBased on the original work from Sam Bobroff.
Currently, EEH recovery is entirely serialized and takes place within a single kernel thread. This can cause recovery to take a long time when there are many devices.
To shorten recovery time, this change allows recovery to proceed in parallel in two ways:
- Each PHB is given it's own recovery event queue and can be recovered independently from other PHBs.
- Driver handlers are called in parallel, but with the constraint that handlers higher up (closer to the PHB) in the PE hierarchy must be called before those lower down.
To maintain the constraint, above, the driver handlers are called by traversing the tree of affected PEs from the top, stopping to call handlers (in parallel) when a PE with devices is discovered. When the calls for that PE are complete, traversal continues at each child PE.
--- arch/powerpc/include/asm/eeh.h | 1 + arch/powerpc/include/asm/eeh_event.h | 7 + arch/powerpc/include/asm/pci-bridge.h | 3 + arch/powerpc/kernel/eeh_driver.c | 323 +++++++++++++++++++------- arch/powerpc/kernel/eeh_event.c | 65 +++--- arch/powerpc/kernel/eeh_pe.c | 3 + 6 files changed, 288 insertions(+), 114 deletions(-)
arch/powerpc/include/asm/eeh.h +1 -0
@@ -128,6 +128,7 @@ static inline bool eeh_pe_passed(struct eeh_pe *pe)#define EEH_DEV_NO_HANDLER (1 << 8)/* No error handler */ #define EEH_DEV_SYSFS (1 << 9)/* Sysfs created */ #define EEH_DEV_REMOVED (1 << 10)/* Removed permanently */ +#define EEH_DEV_RECOVERING (1 << 11)/* Recovering */ struct eeh_dev {int mode;/* EEH mode */
arch/powerpc/include/asm/eeh_event.h +7 -0
@@ -8,6 +8,8 @@#define ASM_POWERPC_EEH_EVENT_H #ifdef __KERNEL__ +#include <linux/workqueue.h> +/* * structure holding pci controller data that describes a * change in the isolation status of a PCI slot. A pointer
@@ -15,16 +17,21 @@* callback. */ struct eeh_event { +struct work_struct work;struct list_head list;/* to form event queue */ struct eeh_pe *pe;/* EEH PE */ unsigned int id;/* Event ID */ }; +extern spinlock_t eeh_eventlist_lock; +int eeh_event_init (void ); +int eeh_phb_event (struct eeh_pe *pe);int eeh_send_failure_event (struct eeh_pe *pe);int __eeh_send_failure_event (struct eeh_pe *pe);void eeh_remove_event (struct eeh_pe *pe,bool force);void eeh_handle_normal_event (unsigned int event_id,struct eeh_pe *pe); +void eeh_handle_normal_event_work (struct work_struct *work);void eeh_handle_special_event (void );#endif /* __KERNEL__ */
arch/powerpc/include/asm/pci-bridge.h +3 -0
@@ -131,6 +131,9 @@ struct pci_controller {struct irq_domain *dev_domain;struct irq_domain *msi_domain;struct fwnode_handle *fwnode; + +bool eeh_in_progress; +struct list_head eeh_eventlist; };/* These are used for config access before all the PCI probing
arch/powerpc/kernel/eeh_driver.c +235 -88
485 lines not on this page · raw message
arch/powerpc/kernel/eeh_event.c +39 -26
104 lines not on this page · raw message
arch/powerpc/kernel/eeh_pe.c +3 -0
@@ -81,6 +81,9 @@ int eeh_phb_pe_create(struct pci_controller *phb) {struct eeh_pe *pe; + phb->eeh_in_progress = false;/* TODO: Necessary? */ +INIT_LIST_HEAD (&phb->eeh_eventlist); +/* Allocate PHB PE */ pe =eeh_pe_alloc (phb,EEH_PE_PHB );if (!pe) {
Signature · 1 line
2.37.1
Oliver O'Halloran
reviewerQuoted 10 lines · Ganesh Goudar
Hi,
EEH reocvery is currently serialized and these patches shorten the time taken for EEH recovery by making the recovery to run in parallel. The original author of these patches is Sam Bobroff, I have rebased and tested these patches.
On powervm with 64 VFs and I see approximately 48% reduction in time taken in EEH recovery,
Can you elaborate on how you're testing this? A VM with 64 separate VFs on 64 separate PHBs I guess? How are you injecting errors?
Quoted 1 line
Yet to be tested on powernv.
If you're not testing on powernv you might as well not be testing. EEH support on pseries is relatively straightforward since managing PCI bridges, etc is all hidden away in the hypervisor. Most of the things about EEH which give me headaches simply don't happen on pseries since the PCI topology seen by the guest is flat.
Quoted 9 lines
These patches were originally posted as separate RFCs, I think posting them as single series would be more helpful, I know the patches are too big, I will try to logically divide in next iterations.
Thanks
Ganesh Goudar (3):
powerpc/eeh: Synchronization for safety
I didn't pick that patch up after Sam left for a reason. I don't recall the exact details, but I think I decided that the lock being added had the same (or mostly the same) scope as the pci rescan lock. All modifications to the EEH PE tree have to occur with the rescan lock held since the per-device EEH setup occurs while configuring the pci_dev (similarly the teardown happens when we remove the pci_dev from its bus). I don't think that was true when Same wrote the patch originally though since it pre-dates my big EEH init re-work.
It's possible (probable even!) I got that wrong, or that it was something I was trying to make true through re-works rather than an accurate assessment of the current state of things. I'll try page some of this back into my brain later. I think I left some comments on Sam's original RFC patch which might still be valid.
Quoted 1 line
powerpc/eeh: Asynchronous recovery
I remember not hating the idea, but I think the implementation leaves a lot to be desired. If you only really care about pseries then just moving to a model where each PHB has its own recovery thread would get you most of the benefit without needing to turn the already complicated recovery path into an async trainwreck. IIRC Sam's motivation for that patch was to make recovery faster for powernv systems when an error was injected on a PF with a lot of child VFs. Certain drivers took forever to recover each VF (might have been mlx5) so doing it in parallel would have sped things up considerably.
Oliver
Quoted 6 lines · Ganesh Goudar
Hi,
EEH reocvery is currently serialized and these patches shorten the time taken for EEH recovery by making the recovery to run in parallel. The original author of these patches is Sam Bobroff, I have rebased and tested these patches.
How did you test this?
I understand that VFIO on 6.0 does not work at all on power?
I am waiting for power maintainers to pick up this series to fix it:
https://lore.kernel.org/kvm/[email protected]/
Jason
Quoted 8 lines · Jason Gunthorpe
Quoted 6 lines · Ganesh Goudar
Hi,
EEH reocvery is currently serialized and these patches shorten the time taken for EEH recovery by making the recovery to run in parallel. The original author of these patches is Sam Bobroff, I have rebased and tested these patches.
How did you test this?
This is tested on SRIOV VFs.
Quoted 8 lines
I understand that VFIO on 6.0 does not work at all on power?
I am waiting for power maintainers to pick up this series to fix it:
https://lore.kernel.org/kvm/[email protected]/
Jason
Hi,
Le 15/09/2022 à 12:15, Ganesh a écrit :
Quoted 5 lines · Jason Gunthorpe
Quoted 4 lines · Ganesh Goudar
Quoted 2 lines
Hi,
Quoted 14 lines
Quoted 5 lines
Quoted 4 lines
EEH reocvery is currently serialized and these patches shorten the time taken for EEH recovery by making the recovery to run in parallel. The original author of these patches is Sam Bobroff, I have rebased and tested these patches.
How did you test this?
This is tested on SRIOV VFs.
Quoted 6 lines
I understand that VFIO on 6.0 does not work at all on power?
I am waiting for power maintainers to pick up this series to fix it:
This series doesn't apply anymore, for instance drivers/vfio/vfio_spapr_eeh.c doesn't exist anymore.
If you series is still relevant, please resubmit.
Christophe