linux_dsm_epyc7002/drivers/nvdimm/nd.h

409 lines
11 KiB
C
Raw Normal View History

/*
* Copyright(c) 2013-2015 Intel Corporation. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of version 2 of the GNU General Public License as
* published by the Free Software Foundation.
*
* This program is distributed in the hope that it will be useful, but
* WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* General Public License for more details.
*/
#ifndef __ND_H__
#define __ND_H__
libnvdimm, nfit: regions (block-data-window, persistent memory, volatile memory) A "region" device represents the maximum capacity of a BLK range (mmio block-data-window(s)), or a PMEM range (DAX-capable persistent memory or volatile memory), without regard for aliasing. Aliasing, in the dimm-local address space (DPA), is resolved by metadata on a dimm to designate which exclusive interface will access the aliased DPA ranges. Support for the per-dimm metadata/label arrvies is in a subsequent patch. The name format of "region" devices is "regionN" where, like dimms, N is a global ida index assigned at discovery time. This id is not reliable across reboots nor in the presence of hotplug. Look to attributes of the region or static id-data of the sub-namespace to generate a persistent name. However, if the platform configuration does not change it is reasonable to expect the same region id to be assigned at the next boot. "region"s have 2 generic attributes "size", and "mapping"s where: - size: the BLK accessible capacity or the span of the system physical address range in the case of PMEM. - mappingN: a tuple describing a dimm's contribution to the region's capacity in the format (<nmemX>,<dpa>,<size>). For a PMEM-region there will be at least one mapping per dimm in the interleave set. For a BLK-region there is only "mapping0" listing the starting DPA of the BLK-region and the available DPA capacity of that space (matches "size" above). The max number of mappings per "region" is hard coded per the constraints of sysfs attribute groups. That said the number of mappings per region should never exceed the maximum number of possible dimms in the system. If the current number turns out to not be enough then the "mappings" attribute clarifies how many there are supposed to be. "32 should be enough for anybody...". Cc: Neil Brown <neilb@suse.de> Cc: <linux-acpi@vger.kernel.org> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: Robert Moore <robert.moore@intel.com> Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Acked-by: Christoph Hellwig <hch@lst.de> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Toshi Kani <toshi.kani@hp.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-10 07:13:14 +07:00
#include <linux/libnvdimm.h>
#include <linux/badblocks.h>
#include <linux/blkdev.h>
#include <linux/device.h>
#include <linux/mutex.h>
#include <linux/ndctl.h>
#include <linux/types.h>
#include <linux/nd.h>
#include "label.h"
enum {
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
/*
* Limits the maximum number of block apertures a dimm can
* support and is an input to the geometry/on-disk-format of a
* BTT instance
*/
ND_MAX_LANES = 256,
SECTOR_SHIFT = 9,
INT_LBASIZE_ALIGNMENT = 64,
};
struct nd_poison {
u64 start;
u64 length;
struct list_head list;
};
struct nvdimm_drvdata {
struct device *dev;
int nsindex_size;
struct nd_cmd_get_config_size nsarea;
void *data;
int ns_current, ns_next;
struct resource dpa;
struct kref kref;
};
struct nd_region_data {
int ns_count;
int ns_active;
unsigned int hints_shift;
void __iomem *flush_wpq[0];
};
static inline void __iomem *ndrd_get_flush_wpq(struct nd_region_data *ndrd,
int dimm, int hint)
{
unsigned int num = 1 << ndrd->hints_shift;
unsigned int mask = num - 1;
return ndrd->flush_wpq[dimm * num + (hint & mask)];
}
static inline void ndrd_set_flush_wpq(struct nd_region_data *ndrd, int dimm,
int hint, void __iomem *flush)
{
unsigned int num = 1 << ndrd->hints_shift;
unsigned int mask = num - 1;
ndrd->flush_wpq[dimm * num + (hint & mask)] = flush;
}
static inline struct nd_namespace_index *to_namespace_index(
struct nvdimm_drvdata *ndd, int i)
{
if (i < 0)
return NULL;
return ndd->data + sizeof_namespace_index(ndd) * i;
}
static inline struct nd_namespace_index *to_current_namespace_index(
struct nvdimm_drvdata *ndd)
{
return to_namespace_index(ndd, ndd->ns_current);
}
static inline struct nd_namespace_index *to_next_namespace_index(
struct nvdimm_drvdata *ndd)
{
return to_namespace_index(ndd, ndd->ns_next);
}
#define nd_dbg_dpa(r, d, res, fmt, arg...) \
dev_dbg((r) ? &(r)->dev : (d)->dev, "%s: %.13s: %#llx @ %#llx " fmt, \
(r) ? dev_name((d)->dev) : "", res ? res->name : "null", \
(unsigned long long) (res ? resource_size(res) : 0), \
(unsigned long long) (res ? res->start : 0), ##arg)
#define for_each_dpa_resource(ndd, res) \
for (res = (ndd)->dpa.child; res; res = res->sibling)
#define for_each_dpa_resource_safe(ndd, res, next) \
for (res = (ndd)->dpa.child, next = res ? res->sibling : NULL; \
res; res = next, next = next ? next->sibling : NULL)
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
struct nd_percpu_lane {
int count;
spinlock_t lock;
};
struct nd_label_ent {
struct list_head list;
struct nd_namespace_label *label;
};
enum nd_mapping_lock_class {
ND_MAPPING_CLASS0,
ND_MAPPING_UUID_SCAN,
};
struct nd_mapping {
struct nvdimm *nvdimm;
u64 start;
u64 size;
struct list_head labels;
struct mutex lock;
/*
* @ndd is for private use at region enable / disable time for
* get_ndd() + put_ndd(), all other nd_mapping to ndd
* conversions use to_ndd() which respects enabled state of the
* nvdimm.
*/
struct nvdimm_drvdata *ndd;
};
libnvdimm, nfit: regions (block-data-window, persistent memory, volatile memory) A "region" device represents the maximum capacity of a BLK range (mmio block-data-window(s)), or a PMEM range (DAX-capable persistent memory or volatile memory), without regard for aliasing. Aliasing, in the dimm-local address space (DPA), is resolved by metadata on a dimm to designate which exclusive interface will access the aliased DPA ranges. Support for the per-dimm metadata/label arrvies is in a subsequent patch. The name format of "region" devices is "regionN" where, like dimms, N is a global ida index assigned at discovery time. This id is not reliable across reboots nor in the presence of hotplug. Look to attributes of the region or static id-data of the sub-namespace to generate a persistent name. However, if the platform configuration does not change it is reasonable to expect the same region id to be assigned at the next boot. "region"s have 2 generic attributes "size", and "mapping"s where: - size: the BLK accessible capacity or the span of the system physical address range in the case of PMEM. - mappingN: a tuple describing a dimm's contribution to the region's capacity in the format (<nmemX>,<dpa>,<size>). For a PMEM-region there will be at least one mapping per dimm in the interleave set. For a BLK-region there is only "mapping0" listing the starting DPA of the BLK-region and the available DPA capacity of that space (matches "size" above). The max number of mappings per "region" is hard coded per the constraints of sysfs attribute groups. That said the number of mappings per region should never exceed the maximum number of possible dimms in the system. If the current number turns out to not be enough then the "mappings" attribute clarifies how many there are supposed to be. "32 should be enough for anybody...". Cc: Neil Brown <neilb@suse.de> Cc: <linux-acpi@vger.kernel.org> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: Robert Moore <robert.moore@intel.com> Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Acked-by: Christoph Hellwig <hch@lst.de> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Toshi Kani <toshi.kani@hp.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-10 07:13:14 +07:00
struct nd_region {
struct device dev;
struct ida ns_ida;
struct ida btt_ida;
struct ida pfn_ida;
struct ida dax_ida;
unsigned long flags;
struct device *ns_seed;
struct device *btt_seed;
struct device *pfn_seed;
struct device *dax_seed;
libnvdimm, nfit: regions (block-data-window, persistent memory, volatile memory) A "region" device represents the maximum capacity of a BLK range (mmio block-data-window(s)), or a PMEM range (DAX-capable persistent memory or volatile memory), without regard for aliasing. Aliasing, in the dimm-local address space (DPA), is resolved by metadata on a dimm to designate which exclusive interface will access the aliased DPA ranges. Support for the per-dimm metadata/label arrvies is in a subsequent patch. The name format of "region" devices is "regionN" where, like dimms, N is a global ida index assigned at discovery time. This id is not reliable across reboots nor in the presence of hotplug. Look to attributes of the region or static id-data of the sub-namespace to generate a persistent name. However, if the platform configuration does not change it is reasonable to expect the same region id to be assigned at the next boot. "region"s have 2 generic attributes "size", and "mapping"s where: - size: the BLK accessible capacity or the span of the system physical address range in the case of PMEM. - mappingN: a tuple describing a dimm's contribution to the region's capacity in the format (<nmemX>,<dpa>,<size>). For a PMEM-region there will be at least one mapping per dimm in the interleave set. For a BLK-region there is only "mapping0" listing the starting DPA of the BLK-region and the available DPA capacity of that space (matches "size" above). The max number of mappings per "region" is hard coded per the constraints of sysfs attribute groups. That said the number of mappings per region should never exceed the maximum number of possible dimms in the system. If the current number turns out to not be enough then the "mappings" attribute clarifies how many there are supposed to be. "32 should be enough for anybody...". Cc: Neil Brown <neilb@suse.de> Cc: <linux-acpi@vger.kernel.org> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: Robert Moore <robert.moore@intel.com> Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Acked-by: Christoph Hellwig <hch@lst.de> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Toshi Kani <toshi.kani@hp.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-10 07:13:14 +07:00
u16 ndr_mappings;
u64 ndr_size;
u64 ndr_start;
int id, num_lanes, ro, numa_node;
libnvdimm, nfit: regions (block-data-window, persistent memory, volatile memory) A "region" device represents the maximum capacity of a BLK range (mmio block-data-window(s)), or a PMEM range (DAX-capable persistent memory or volatile memory), without regard for aliasing. Aliasing, in the dimm-local address space (DPA), is resolved by metadata on a dimm to designate which exclusive interface will access the aliased DPA ranges. Support for the per-dimm metadata/label arrvies is in a subsequent patch. The name format of "region" devices is "regionN" where, like dimms, N is a global ida index assigned at discovery time. This id is not reliable across reboots nor in the presence of hotplug. Look to attributes of the region or static id-data of the sub-namespace to generate a persistent name. However, if the platform configuration does not change it is reasonable to expect the same region id to be assigned at the next boot. "region"s have 2 generic attributes "size", and "mapping"s where: - size: the BLK accessible capacity or the span of the system physical address range in the case of PMEM. - mappingN: a tuple describing a dimm's contribution to the region's capacity in the format (<nmemX>,<dpa>,<size>). For a PMEM-region there will be at least one mapping per dimm in the interleave set. For a BLK-region there is only "mapping0" listing the starting DPA of the BLK-region and the available DPA capacity of that space (matches "size" above). The max number of mappings per "region" is hard coded per the constraints of sysfs attribute groups. That said the number of mappings per region should never exceed the maximum number of possible dimms in the system. If the current number turns out to not be enough then the "mappings" attribute clarifies how many there are supposed to be. "32 should be enough for anybody...". Cc: Neil Brown <neilb@suse.de> Cc: <linux-acpi@vger.kernel.org> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: Robert Moore <robert.moore@intel.com> Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Acked-by: Christoph Hellwig <hch@lst.de> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Toshi Kani <toshi.kani@hp.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-10 07:13:14 +07:00
void *provider_data;
2015-05-02 00:11:27 +07:00
struct nd_interleave_set *nd_set;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
struct nd_percpu_lane __percpu *lane;
libnvdimm, nfit: regions (block-data-window, persistent memory, volatile memory) A "region" device represents the maximum capacity of a BLK range (mmio block-data-window(s)), or a PMEM range (DAX-capable persistent memory or volatile memory), without regard for aliasing. Aliasing, in the dimm-local address space (DPA), is resolved by metadata on a dimm to designate which exclusive interface will access the aliased DPA ranges. Support for the per-dimm metadata/label arrvies is in a subsequent patch. The name format of "region" devices is "regionN" where, like dimms, N is a global ida index assigned at discovery time. This id is not reliable across reboots nor in the presence of hotplug. Look to attributes of the region or static id-data of the sub-namespace to generate a persistent name. However, if the platform configuration does not change it is reasonable to expect the same region id to be assigned at the next boot. "region"s have 2 generic attributes "size", and "mapping"s where: - size: the BLK accessible capacity or the span of the system physical address range in the case of PMEM. - mappingN: a tuple describing a dimm's contribution to the region's capacity in the format (<nmemX>,<dpa>,<size>). For a PMEM-region there will be at least one mapping per dimm in the interleave set. For a BLK-region there is only "mapping0" listing the starting DPA of the BLK-region and the available DPA capacity of that space (matches "size" above). The max number of mappings per "region" is hard coded per the constraints of sysfs attribute groups. That said the number of mappings per region should never exceed the maximum number of possible dimms in the system. If the current number turns out to not be enough then the "mappings" attribute clarifies how many there are supposed to be. "32 should be enough for anybody...". Cc: Neil Brown <neilb@suse.de> Cc: <linux-acpi@vger.kernel.org> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: Robert Moore <robert.moore@intel.com> Cc: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Acked-by: Christoph Hellwig <hch@lst.de> Acked-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Toshi Kani <toshi.kani@hp.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-10 07:13:14 +07:00
struct nd_mapping mapping[0];
};
struct nd_blk_region {
int (*enable)(struct nvdimm_bus *nvdimm_bus, struct device *dev);
int (*do_io)(struct nd_blk_region *ndbr, resource_size_t dpa,
void *iobuf, u64 len, int rw);
void *blk_provider_data;
struct nd_region nd_region;
};
/*
* Lookup next in the repeating sequence of 01, 10, and 11.
*/
static inline unsigned nd_inc_seq(unsigned seq)
{
static const unsigned next[] = { 0, 2, 3, 1 };
return next[seq & 3];
}
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
struct btt;
struct nd_btt {
struct device dev;
struct nd_namespace_common *ndns;
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
struct btt *btt;
unsigned long lbasize;
u64 size;
u8 *uuid;
int id;
};
enum nd_pfn_mode {
PFN_MODE_NONE,
PFN_MODE_RAM,
PFN_MODE_PMEM,
};
struct nd_pfn {
int id;
u8 *uuid;
struct device dev;
unsigned long align;
unsigned long npfns;
enum nd_pfn_mode mode;
struct nd_pfn_sb *pfn_sb;
struct nd_namespace_common *ndns;
};
struct nd_dax {
struct nd_pfn nd_pfn;
};
enum nd_async_mode {
ND_SYNC,
ND_ASYNC,
};
int nd_integrity_init(struct gendisk *disk, unsigned long meta_size);
void wait_nvdimm_bus_probe_idle(struct device *dev);
void nd_device_register(struct device *dev);
void nd_device_unregister(struct device *dev, enum nd_async_mode mode);
void nd_device_notify(struct device *dev, enum nvdimm_event event);
int nd_uuid_store(struct device *dev, u8 **uuid_out, const char *buf,
size_t len);
ssize_t nd_sector_size_show(unsigned long current_lbasize,
const unsigned long *supported, char *buf);
ssize_t nd_sector_size_store(struct device *dev, const char *buf,
unsigned long *current_lbasize, const unsigned long *supported);
int __init nvdimm_init(void);
int __init nd_region_init(void);
void nvdimm_exit(void);
void nd_region_exit(void);
struct nvdimm;
struct nvdimm_drvdata *to_ndd(struct nd_mapping *nd_mapping);
int nvdimm_check_config_data(struct device *dev);
int nvdimm_init_nsarea(struct nvdimm_drvdata *ndd);
int nvdimm_init_config_data(struct nvdimm_drvdata *ndd);
int nvdimm_set_config_data(struct nvdimm_drvdata *ndd, size_t offset,
void *buf, size_t len);
long nvdimm_clear_poison(struct device *dev, phys_addr_t phys,
unsigned int len);
void nvdimm_set_aliasing(struct device *dev);
struct nd_btt *to_nd_btt(struct device *dev);
struct nd_gen_sb {
char reserved[SZ_4K - 8];
__le64 checksum;
};
u64 nd_sb_checksum(struct nd_gen_sb *sb);
#if IS_ENABLED(CONFIG_BTT)
int nd_btt_probe(struct device *dev, struct nd_namespace_common *ndns);
bool is_nd_btt(struct device *dev);
struct device *nd_btt_create(struct nd_region *nd_region);
#else
static inline int nd_btt_probe(struct device *dev,
struct nd_namespace_common *ndns)
{
return -ENODEV;
}
static inline bool is_nd_btt(struct device *dev)
{
return false;
}
static inline struct device *nd_btt_create(struct nd_region *nd_region)
{
return NULL;
}
#endif
struct nd_pfn *to_nd_pfn(struct device *dev);
#if IS_ENABLED(CONFIG_NVDIMM_PFN)
int nd_pfn_probe(struct device *dev, struct nd_namespace_common *ndns);
bool is_nd_pfn(struct device *dev);
struct device *nd_pfn_create(struct nd_region *nd_region);
struct device *nd_pfn_devinit(struct nd_pfn *nd_pfn,
struct nd_namespace_common *ndns);
int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig);
extern struct attribute_group nd_pfn_attribute_group;
#else
static inline int nd_pfn_probe(struct device *dev,
struct nd_namespace_common *ndns)
{
return -ENODEV;
}
static inline bool is_nd_pfn(struct device *dev)
{
return false;
}
static inline struct device *nd_pfn_create(struct nd_region *nd_region)
{
return NULL;
}
static inline int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
{
return -ENODEV;
}
#endif
struct nd_dax *to_nd_dax(struct device *dev);
#if IS_ENABLED(CONFIG_NVDIMM_DAX)
int nd_dax_probe(struct device *dev, struct nd_namespace_common *ndns);
bool is_nd_dax(struct device *dev);
struct device *nd_dax_create(struct nd_region *nd_region);
#else
static inline int nd_dax_probe(struct device *dev,
struct nd_namespace_common *ndns)
{
return -ENODEV;
}
static inline bool is_nd_dax(struct device *dev)
{
return false;
}
static inline struct device *nd_dax_create(struct nd_region *nd_region)
{
return NULL;
}
#endif
struct nd_region *to_nd_region(struct device *dev);
int nd_region_to_nstype(struct nd_region *nd_region);
int nd_region_register_namespaces(struct nd_region *nd_region, int *err);
u64 nd_region_interleave_set_cookie(struct nd_region *nd_region);
void nvdimm_bus_lock(struct device *dev);
void nvdimm_bus_unlock(struct device *dev);
bool is_nvdimm_bus_locked(struct device *dev);
int nvdimm_revalidate_disk(struct gendisk *disk);
void nvdimm_drvdata_release(struct kref *kref);
void put_ndd(struct nvdimm_drvdata *ndd);
int nd_label_reserve_dpa(struct nvdimm_drvdata *ndd);
void nvdimm_free_dpa(struct nvdimm_drvdata *ndd, struct resource *res);
struct resource *nvdimm_allocate_dpa(struct nvdimm_drvdata *ndd,
struct nd_label_id *label_id, resource_size_t start,
resource_size_t n);
resource_size_t nvdimm_namespace_capacity(struct nd_namespace_common *ndns);
struct nd_namespace_common *nvdimm_namespace_common_probe(struct device *dev);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
int nvdimm_namespace_attach_btt(struct nd_namespace_common *ndns);
int nvdimm_namespace_detach_btt(struct nd_btt *nd_btt);
nd_btt: atomic sector updates BTT stands for Block Translation Table, and is a way to provide power fail sector atomicity semantics for block devices that have the ability to perform byte granularity IO. It relies on the capability of libnvdimm namespace devices to do byte aligned IO. The BTT works as a stacked blocked device, and reserves a chunk of space from the backing device for its accounting metadata. It is a bio-based driver because all IO is done synchronously, and there is no queuing or asynchronous completions at either the device or the driver level. The BTT uses 'lanes' to index into various 'on-disk' data structures, and lanes also act as a synchronization mechanism in case there are more CPUs than available lanes. We did a comparison between two lane lock strategies - first where we kept an atomic counter around that tracked which was the last lane that was used, and 'our' lane was determined by atomically incrementing that. That way, for the nr_cpus > nr_lanes case, theoretically, no CPU would be blocked waiting for a lane. The other strategy was to use the cpu number we're scheduled on to and hash it to a lane number. Theoretically, this could block an IO that could've otherwise run using a different, free lane. But some fio workloads showed that the direct cpu -> lane hash performed faster than tracking 'last lane' - my reasoning is the cache thrash caused by moving the atomic variable made that approach slower than simply waiting out the in-progress IO. This supports the conclusion that the driver can be a very simple bio-based one that does synchronous IOs instead of queuing. Cc: Andy Lutomirski <luto@amacapital.net> Cc: Boaz Harrosh <boaz@plexistor.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Jens Axboe <axboe@fb.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Christoph Hellwig <hch@lst.de> Cc: Neil Brown <neilb@suse.de> Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Dave Chinner <david@fromorbit.com> Cc: Greg KH <gregkh@linuxfoundation.org> [jmoyer: fix nmi watchdog timeout in btt_map_init] [jmoyer: move btt initialization to module load path] [jmoyer: fix memory leak in the btt initialization path] [jmoyer: Don't overwrite corrupted arenas] Signed-off-by: Vishal Verma <vishal.l.verma@linux.intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2015-06-25 15:20:32 +07:00
const char *nvdimm_namespace_disk_name(struct nd_namespace_common *ndns,
char *name);
void nvdimm_badblocks_populate(struct nd_region *nd_region,
struct badblocks *bb, const struct resource *res);
#if IS_ENABLED(CONFIG_ND_CLAIM)
struct vmem_altmap *nvdimm_setup_pfn(struct nd_pfn *nd_pfn,
struct resource *res, struct vmem_altmap *altmap);
int devm_nsio_enable(struct device *dev, struct nd_namespace_io *nsio);
void devm_nsio_disable(struct device *dev, struct nd_namespace_io *nsio);
#else
static inline struct vmem_altmap *nvdimm_setup_pfn(struct nd_pfn *nd_pfn,
struct resource *res, struct vmem_altmap *altmap)
{
return ERR_PTR(-ENXIO);
}
static inline int devm_nsio_enable(struct device *dev,
struct nd_namespace_io *nsio)
{
return -ENXIO;
}
static inline void devm_nsio_disable(struct device *dev,
struct nd_namespace_io *nsio)
{
}
#endif
int nd_blk_region_init(struct nd_region *nd_region);
int nd_region_activate(struct nd_region *nd_region);
void __nd_iostat_start(struct bio *bio, unsigned long *start);
static inline bool nd_iostat_start(struct bio *bio, unsigned long *start)
{
struct gendisk *disk = bio->bi_bdev->bd_disk;
if (!blk_queue_io_stat(disk->queue))
return false;
*start = jiffies;
generic_start_io_acct(bio_data_dir(bio),
bio_sectors(bio), &disk->part0);
return true;
}
static inline void nd_iostat_end(struct bio *bio, unsigned long start)
{
struct gendisk *disk = bio->bi_bdev->bd_disk;
generic_end_io_acct(bio_data_dir(bio), &disk->part0, start);
}
static inline bool is_bad_pmem(struct badblocks *bb, sector_t sector,
unsigned int len)
{
if (bb->count) {
sector_t first_bad;
int num_bad;
return !!badblocks_check(bb, sector, len / 512, &first_bad,
&num_bad);
}
return false;
}
resource_size_t nd_namespace_blk_validate(struct nd_namespace_blk *nsblk);
const u8 *nd_dev_to_uuid(struct device *dev);
bool pmem_should_map_pages(struct device *dev);
#endif /* __ND_H__ */