Starting a full Address Range Scrub (ARS) on hitting a memory error machine check exception may not always be desirable. Provide a way through sysfs to toggle the behavior between just adding the address (cache line) where the MCE happened to the poison list and doing a full scrub. The former (selective insertion of the address) is done unconditionally. Cc: linux-acpi(a)vger.kernel.org Cc: Dan Williams <dan.j.williams(a)intel.com&gt; Cc: Linda Knippers <linda.knippers(a)hpe.com&gt; Cc: Rafael J. Wysocki <rafael.j.wysocki(a)intel.com&gt; Signed-off-by: Vishal Verma <vishal.l.verma(a)intel.com&gt; --- drivers/acpi/nfit/core.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++ drivers/acpi/nfit/mce.c | 24 ++++++++++++++++----- drivers/acpi/nfit/nfit.h | 6 ++++++ 3 files changed, 79 insertions(+), 5 deletions(-) diff --git a/drivers/acpi/nfit/core.c b/drivers/acpi/nfit/core.c index 80cc7c0..6a12bc2 100644 --- a/drivers/acpi/nfit/core.c +++ b/drivers/acpi/nfit/core.c @@ -878,6 +878,58 @@ static ssize_t revision_show(struct device *dev, } static DEVICE_ATTR_RO(revision); +static ssize_t hw_error_scrub_show(struct device *dev, + struct device_attribute *attr, char *buf) +{ + struct nvdimm_bus *nvdimm_bus = to_nvdimm_bus(dev); + struct nvdimm_bus_descriptor *nd_desc = to_nd_desc(nvdimm_bus); + struct acpi_nfit_desc *acpi_desc = to_acpi_desc(nd_desc); + + return sprintf(buf, "%d\n", acpi_desc->scrub_mode); +} + +/* + * The 'hw_error_scrub' attribute can have the following values written to it: + * '1': Enable a full scrub to happen if an exception for a memory error is + * received. + * '2': Switch to the default mode where an exception will only insert + * the address of the memory error into the poison and badblocks lists. + */ +static ssize_t hw_error_scrub_store(struct device *dev, + struct device_attribute *attr, const char *buf, size_t size) +{ + struct nvdimm_bus_descriptor *nd_desc; + ssize_t rc; + long val; + + rc = kstrtol(buf, 0, &val); + if (rc) + return rc; + + device_lock(dev); + nd_desc = dev_get_drvdata(dev); + if (nd_desc) { + struct acpi_nfit_desc *acpi_desc = to_acpi_desc(nd_desc); + + switch (val) { + case MCE_SCRUB_ON: + acpi_desc->scrub_mode = MCE_SCRUB_ON; + break; + case MCE_SCRUB_OFF: + acpi_desc->scrub_mode = MCE_SCRUB_OFF; + break; + default: + rc = -EINVAL; + break; + } + } + device_unlock(dev); + if (rc) + return rc; + return size; +} +static DEVICE_ATTR_RW(hw_error_scrub); + /* * This shows the number of full Address Range Scrubs that have been * completed since driver load time. Userspace can wait on this using @@ -950,6 +1002,7 @@ static umode_t nfit_visible(struct kobject *kobj, struct attribute *a, int n) static struct attribute *acpi_nfit_attributes[] = { &dev_attr_revision.attr, &dev_attr_scrub.attr, + &dev_attr_hw_error_scrub.attr, NULL, }; @@ -2489,6 +2542,7 @@ int acpi_nfit_init(struct acpi_nfit_desc *acpi_desc, void *data, acpi_size sz) goto out_unlock; } + acpi_desc->scrub_mode = MCE_SCRUB_OFF; rc = acpi_nfit_check_deletions(acpi_desc, &prev); if (rc) goto out_unlock; diff --git a/drivers/acpi/nfit/mce.c b/drivers/acpi/nfit/mce.c index 161f915..05005e2 100644 --- a/drivers/acpi/nfit/mce.c +++ b/drivers/acpi/nfit/mce.c @@ -14,6 +14,7 @@ */ #include <linux/notifier.h> #include <linux/acpi.h> +#include <linux/nd.h> #include <asm/mce.h> #include "nfit.h" @@ -62,12 +63,25 @@ static int nfit_handle_mce(struct notifier_block *nb, unsigned long val, } mutex_unlock(&acpi_desc->init_mutex); - /* - * We can ignore an -EBUSY here because if an ARS is already - * in progress, just let that be the last authoritative one - */ - if (found_match) + if (!found_match) + continue; + + /* If this fails due to an -ENOMEM, there is little we can do */ + nvdimm_bus_add_poison(acpi_desc->nvdimm_bus, + ALIGN(mce->addr, L1_CACHE_BYTES), + L1_CACHE_BYTES); + nvdimm_region_notify(nfit_spa->nd_region, + NVDIMM_REVALIDATE_POISON); + + if (acpi_desc->scrub_mode == MCE_SCRUB_ON) { + /* + * We can ignore an -EBUSY here because if an ARS is + * already in progress, just let that be the last + * authoritative one + */ acpi_nfit_ars_rescan(acpi_desc); + } + break; } mutex_unlock(&acpi_desc_lock); diff --git a/drivers/acpi/nfit/nfit.h b/drivers/acpi/nfit/nfit.h index e894ded..487dc0a 100644 --- a/drivers/acpi/nfit/nfit.h +++ b/drivers/acpi/nfit/nfit.h @@ -152,6 +152,7 @@ struct acpi_nfit_desc { struct list_head list; struct kernfs_node *scrub_count_state; unsigned int scrub_count; + unsigned int scrub_mode; unsigned int cancel:1; unsigned long dimm_cmd_force_en; unsigned long bus_cmd_force_en; @@ -159,6 +160,11 @@ struct acpi_nfit_desc { void *iobuf, u64 len, int rw); }; +enum scrub_mode { + MCE_SCRUB_ON = 1, + MCE_SCRUB_OFF = 2, +}; + enum nd_blk_mmio_selector { BDW, DCR, -- 2.7.4

pmem_do_bvec used to kmap_atomic at the begin, and only unmap at the end. Things like nvdimm_clear_poison may want to do nvdimm subsystem bookkeeping operations that may involve taking locks or doing memory allocations, and we can't do that from the atomic context. Reduce the atomic context to just what needs it - the memcpy to/from pmem. Cc: Ross Zwisler <ross.zwisler(a)linux.intel.com&gt; Signed-off-by: Vishal Verma <vishal.l.verma(a)intel.com&gt; --- drivers/nvdimm/pmem.c | 28 +++++++++++++++++++++++----- 1 file changed, 23 insertions(+), 5 deletions(-) diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c index 571a6c7..ca038d8 100644 --- a/drivers/nvdimm/pmem.c +++ b/drivers/nvdimm/pmem.c @@ -66,13 +66,32 @@ static void pmem_clear_poison(struct pmem_device *pmem, phys_addr_t offset, invalidate_pmem(pmem->virt_addr + offset, len); } +static void map_and_copy_to_pmem(void *pmem_addr, struct page *page, + unsigned int off, unsigned int len) +{ + void *mem = kmap_atomic(page); + + memcpy_to_pmem(pmem_addr, mem + off, len); + kunmap_atomic(mem); +} + +static int map_and_copy_from_pmem(struct page *page, unsigned int off, + void *pmem_addr, unsigned int len) +{ + int rc; + void *mem = kmap_atomic(page); + + rc = memcpy_from_pmem(mem + off, pmem_addr, len); + kunmap_atomic(mem); + return rc; +} + static int pmem_do_bvec(struct pmem_device *pmem, struct page *page, unsigned int len, unsigned int off, bool is_write, sector_t sector) { int rc = 0; bool bad_pmem = false; - void *mem = kmap_atomic(page); phys_addr_t pmem_off = sector * 512 + pmem->data_offset; void *pmem_addr = pmem->virt_addr + pmem_off; @@ -83,7 +102,7 @@ static int pmem_do_bvec(struct pmem_device *pmem, struct page *page, if (unlikely(bad_pmem)) rc = -EIO; else { - rc = memcpy_from_pmem(mem + off, pmem_addr, len); + rc = map_and_copy_from_pmem(page, off, pmem_addr, len); flush_dcache_page(page); } } else { @@ -102,14 +121,13 @@ static int pmem_do_bvec(struct pmem_device *pmem, struct page *page, * after clear poison. */ flush_dcache_page(page); - memcpy_to_pmem(pmem_addr, mem + off, len); + map_and_copy_to_pmem(pmem_addr, page, off, len); if (unlikely(bad_pmem)) { pmem_clear_poison(pmem, pmem_off, len); - memcpy_to_pmem(pmem_addr, mem + off, len); + map_and_copy_to_pmem(pmem_addr, page, off, len); } } - kunmap_atomic(mem); return rc; } -- 2.7.4

nvdimm_clear_poison cleared the user-visible badblocks, and sent commands to the NVDIMM to clear the areas marked as 'poison', but it neglected to clear the same areas from the internal poison_list which is used to marshal ARS results before sorting them by namespace. As a result, once on-demand ARS functionality was added: 37b137f nfit, libnvdimm: allow an ARS scrub to be triggered on demand A scrub triggered from either sysfs or an MCE was found to be adding stale entries that had been cleared from gendisk->badblocks, but were still present in nvdimm_bus->poison_list. This adds the missing step of clearing poison_list entries when clearing poison, so that it is in sync with badblocks. Fixes: 37b137f nfit, libnvdimm: allow an ARS scrub to be triggered on demand Cc: Dan Williams <dan.j.williams(a)intel.com&gt; Signed-off-by: Vishal Verma <vishal.l.verma(a)intel.com&gt; --- drivers/nvdimm/bus.c | 2 ++ drivers/nvdimm/core.c | 73 ++++++++++++++++++++++++++++++++++++++++++++--- include/linux/libnvdimm.h | 2 ++ 3 files changed, 73 insertions(+), 4 deletions(-) diff --git a/drivers/nvdimm/bus.c b/drivers/nvdimm/bus.c index 935866f..a8b6949 100644 --- a/drivers/nvdimm/bus.c +++ b/drivers/nvdimm/bus.c @@ -217,6 +217,8 @@ long nvdimm_clear_poison(struct device *dev, phys_addr_t phys, return rc; if (cmd_rc < 0) return cmd_rc; + + nvdimm_clear_from_poison_list(nvdimm_bus, phys, len); return clear_err.cleared; } EXPORT_SYMBOL_GPL(nvdimm_clear_poison); diff --git a/drivers/nvdimm/core.c b/drivers/nvdimm/core.c index 715583f..42e40db 100644 --- a/drivers/nvdimm/core.c +++ b/drivers/nvdimm/core.c @@ -541,11 +541,12 @@ void nvdimm_badblocks_populate(struct nd_region *nd_region, } EXPORT_SYMBOL_GPL(nvdimm_badblocks_populate); -static int add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length) +static int add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length, + gfp_t flags) { struct nd_poison *pl; - pl = kzalloc(sizeof(*pl), GFP_KERNEL); + pl = kzalloc(sizeof(*pl), flags); if (!pl) return -ENOMEM; @@ -561,7 +562,7 @@ static int bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length) struct nd_poison *pl; if (list_empty(&nvdimm_bus->poison_list)) - return add_poison(nvdimm_bus, addr, length); + return add_poison(nvdimm_bus, addr, length, GFP_KERNEL); /* * There is a chance this is a duplicate, check for those first. @@ -581,7 +582,7 @@ static int bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length) * as any overlapping ranges will get resolved when the list is consumed * and converted to badblocks */ - return add_poison(nvdimm_bus, addr, length); + return add_poison(nvdimm_bus, addr, length, GFP_KERNEL); } int nvdimm_bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length) @@ -596,6 +597,70 @@ int nvdimm_bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length) } EXPORT_SYMBOL_GPL(nvdimm_bus_add_poison); +void nvdimm_clear_from_poison_list(struct nvdimm_bus *nvdimm_bus, + phys_addr_t start, unsigned int len) +{ + struct list_head *poison_list = &nvdimm_bus->poison_list; + u64 clr_end = start + len - 1; + struct nd_poison *pl, *next; + + nvdimm_bus_lock(&nvdimm_bus->dev); + WARN_ON_ONCE(list_empty(poison_list)); + + /* + * [start, clr_end] is the poison interval being cleared. + * [pl->start, pl_end] is the poison_list entry we're comparing + * the above interval against. The poison list entry may need + * to be modified (update either start or length), deleted, or + * split into two based on the overlap characteristics + */ + + list_for_each_entry_safe(pl, next, poison_list, list) { + u64 pl_end = pl->start + pl->length - 1; + + /* Skip intervals with no intersection */ + if (pl_end < start) + continue; + if (pl->start > clr_end) + continue; + /* Delete completely overlapped poison entries */ + if ((pl->start >= start) && (pl_end <= clr_end)) { + list_del(&pl->list); + kfree(pl); + continue; + } + /* Adjust start point of partially cleared entries */ + if ((start <= pl->start) && (clr_end > pl->start)) { + pl->length -= clr_end - pl->start + 1; + pl->start = clr_end + 1; + continue; + } + /* Adjust pl->length for partial clearing at the tail end */ + if ((pl->start < start) && (pl_end <= clr_end)) { + /* pl->start remains the same */ + pl->length = start - pl->start; + continue; + } + /* + * If clearing in the middle of an entry, we split it into + * two by modifying the current entry to represent one half of + * the split, and adding a new entry for the second half. + */ + if ((pl->start < start) && (pl_end > clr_end)) { + u64 new_start = clr_end + 1; + u64 new_len = pl_end - new_start + 1; + + /* Add new entry covering the right half */ + add_poison(nvdimm_bus, new_start, new_len, GFP_NOIO); + /* Adjust this entry to cover the left half */ + pl->length = start - pl->start; + continue; + } + } + nvdimm_bus_unlock(&nvdimm_bus->dev); +} +EXPORT_SYMBOL_GPL(nvdimm_clear_from_poison_list); + #ifdef CONFIG_BLK_DEV_INTEGRITY int nd_integrity_init(struct gendisk *disk, unsigned long meta_size) { diff --git a/include/linux/libnvdimm.h b/include/linux/libnvdimm.h index b519e13..bbfce62 100644 --- a/include/linux/libnvdimm.h +++ b/include/linux/libnvdimm.h @@ -129,6 +129,8 @@ static inline struct nd_blk_region_desc *to_blk_region_desc( } int nvdimm_bus_add_poison(struct nvdimm_bus *nvdimm_bus, u64 addr, u64 length); +void nvdimm_clear_from_poison_list(struct nvdimm_bus *nvdimm_bus, + phys_addr_t start, unsigned int len); struct nvdimm_bus *nvdimm_bus_register(struct device *parent, struct nvdimm_bus_descriptor *nfit_desc); void nvdimm_bus_unregister(struct nvdimm_bus *nvdimm_bus); -- 2.7.4