linux/drivers/nvdimm/btt.h

240 lines
7.2 KiB
C
Raw Normal View History

/* SPDX-License-Identifier: GPL-2.0-only */
/*
* Block Translation Table library
* Copyright (c) 2014-2015, Intel Corporation.
*/
#ifndef _LINUX_BTT_H
#define _LINUX_BTT_H
#include <linux/types.h>
#define BTT_SIG_LEN 16
#define BTT_SIG "BTT_ARENA_INFO\0"
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 08:20:32 +00:00
#define MAP_ENT_SIZE 4
#define MAP_TRIM_SHIFT 31
#define MAP_TRIM_MASK (1 << MAP_TRIM_SHIFT)
#define MAP_ERR_SHIFT 30
#define MAP_ERR_MASK (1 << MAP_ERR_SHIFT)
#define MAP_LBA_MASK (~((1 << MAP_TRIM_SHIFT) | (1 << MAP_ERR_SHIFT)))
#define MAP_ENT_NORMAL 0xC0000000
#define LOG_GRP_SIZE sizeof(struct log_group)
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 08:20:32 +00:00
#define LOG_ENT_SIZE sizeof(struct log_entry)
#define ARENA_MIN_SIZE (1UL << 24) /* 16 MB */
#define ARENA_MAX_SIZE (1ULL << 39) /* 512 GB */
#define RTT_VALID (1UL << 31)
#define RTT_INVALID 0
#define BTT_PG_SIZE 4096
#define BTT_DEFAULT_NFREE ND_MAX_LANES
#define LOG_SEQ_INIT 1
#define IB_FLAG_ERROR 0x00000001
#define IB_FLAG_ERROR_MASK 0x00000001
#define ent_lba(ent) (ent & MAP_LBA_MASK)
#define ent_e_flag(ent) (!!(ent & MAP_ERR_MASK))
#define ent_z_flag(ent) (!!(ent & MAP_TRIM_MASK))
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 01:36:03 +00:00
#define set_e_flag(ent) (ent |= MAP_ERR_MASK)
libnvdimm/btt: Fix LBA masking during 'free list' population The Linux BTT implementation assumes that log entries will never have the 'zero' flag set, and indeed it never sets that flag for log entries itself. However, the UEFI spec is ambiguous on the exact format of the LBA field of a log entry, specifically as to whether it should include the additional flag bits or not. While a zero bit doesn't make sense in the context of a log entry, other BTT implementations might still have it set. If an implementation does happen to have it set, we would happily read it in as the next block to write to for writes. Since a high bit is set, it pushes the block number out of the range of an 'arena', and we fail such a write with an EIO. Follow the robustness principle, and tolerate such implementations by stripping out the zero flag when populating the free list during initialization. Additionally, use the same stripped out entries for detection of incomplete writes and map restoration that happens at this stage. Add a sysfs file 'log_zero_flags' that indicates the ability to accept such a layout to userspace applications. This enables 'ndctl check-namespace' to recognize whether the kernel is able to handle zero flags, or whether it should attempt a fix-up under the --repair option. Cc: Dan Williams <dan.j.williams@intel.com> Reported-by: Dexuan Cui <decui@microsoft.com> Reported-by: Pedro d'Aquino Filocre F S Barbuda <pbarbuda@microsoft.com> Tested-by: Dexuan Cui <decui@microsoft.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2019-02-28 00:06:27 +00:00
/* 'normal' is both e and z flags set */
#define ent_normal(ent) (ent_e_flag(ent) && ent_z_flag(ent))
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 08:20:32 +00:00
enum btt_init_state {
INIT_UNCHECKED = 0,
INIT_NOTFOUND,
INIT_READY
};
/*
* A log group represents one log 'lane', and consists of four log entries.
* Two of the four entries are valid entries, and the remaining two are
* padding. Due to an old bug in the padding location, we need to perform a
* test to determine the padding scheme being used, and use that scheme
* thereafter.
*
* In kernels prior to 4.15, 'log group' would have actual log entries at
* indices (0, 2) and padding at indices (1, 3), where as the correct/updated
* format has log entries at indices (0, 1) and padding at indices (2, 3).
*
* Old (pre 4.15) format:
* +-----------------+-----------------+
* | ent[0] | ent[1] |
* | 16B | 16B |
* | lba/old/new/seq | pad |
* +-----------------------------------+
* | ent[2] | ent[3] |
* | 16B | 16B |
* | lba/old/new/seq | pad |
* +-----------------+-----------------+
*
* New format:
* +-----------------+-----------------+
* | ent[0] | ent[1] |
* | 16B | 16B |
* | lba/old/new/seq | lba/old/new/seq |
* +-----------------------------------+
* | ent[2] | ent[3] |
* | 16B | 16B |
* | pad | pad |
* +-----------------+-----------------+
*
* We detect during start-up which format is in use, and set
* arena->log_index[(0, 1)] with the detected format.
*/
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 08:20:32 +00:00
struct log_entry {
__le32 lba;
__le32 old_map;
__le32 new_map;
__le32 seq;
};
struct log_group {
struct log_entry ent[4];
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 08:20:32 +00:00
};
struct btt_sb {
u8 signature[BTT_SIG_LEN];
u8 uuid[16];
u8 parent_uuid[16];
__le32 flags;
__le16 version_major;
__le16 version_minor;
__le32 external_lbasize;
__le32 external_nlba;
__le32 internal_lbasize;
__le32 internal_nlba;
__le32 nfree;
__le32 infosize;
__le64 nextoff;
__le64 dataoff;
__le64 mapoff;
__le64 logoff;
__le64 info2off;
u8 padding[3968];
__le64 checksum;
};
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 08:20:32 +00:00
struct free_entry {
u32 block;
u8 sub;
u8 seq;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 01:36:03 +00:00
u8 has_err;
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 08:20:32 +00:00
};
struct aligned_lock {
union {
spinlock_t lock;
u8 cacheline_padding[L1_CACHE_BYTES];
};
};
/**
* struct arena_info - handle for an arena
* @size: Size in bytes this arena occupies on the raw device.
* This includes arena metadata.
* @external_lba_start: The first external LBA in this arena.
* @internal_nlba: Number of internal blocks available in the arena
* including nfree reserved blocks
* @internal_lbasize: Internal and external lba sizes may be different as
* we can round up 'odd' external lbasizes such as 520B
* to be aligned.
* @external_nlba: Number of blocks contributed by the arena to the number
* reported to upper layers. (internal_nlba - nfree)
* @external_lbasize: LBA size as exposed to upper layers.
* @nfree: A reserve number of 'free' blocks that is used to
* handle incoming writes.
* @version_major: Metadata layout version major.
* @version_minor: Metadata layout version minor.
* @sector_size: The Linux sector size - 512 or 4096
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 08:20:32 +00:00
* @nextoff: Offset in bytes to the start of the next arena.
* @infooff: Offset in bytes to the info block of this arena.
* @dataoff: Offset in bytes to the data area of this arena.
* @mapoff: Offset in bytes to the map area of this arena.
* @logoff: Offset in bytes to the log area of this arena.
* @info2off: Offset in bytes to the backup info block of this arena.
* @freelist: Pointer to in-memory list of free blocks
* @rtt: Pointer to in-memory "Read Tracking Table"
* @map_locks: Spinlocks protecting concurrent map writes
* @nd_btt: Pointer to parent nd_btt structure.
* @list: List head for list of arenas
* @debugfs_dir: Debugfs dentry
* @flags: Arena flags - may signify error states.
* @err_lock: Mutex for synchronizing error clearing.
* @log_index: Indices of the valid log entries in a log_group
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 08:20:32 +00:00
*
* arena_info is a per-arena handle. Once an arena is narrowed down for an
* IO, this struct is passed around for the duration of the IO.
*/
struct arena_info {
u64 size; /* Total bytes for this arena */
u64 external_lba_start;
u32 internal_nlba;
u32 internal_lbasize;
u32 external_nlba;
u32 external_lbasize;
u32 nfree;
u16 version_major;
u16 version_minor;
u32 sector_size;
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 08:20:32 +00:00
/* Byte offsets to the different on-media structures */
u64 nextoff;
u64 infooff;
u64 dataoff;
u64 mapoff;
u64 logoff;
u64 info2off;
/* Pointers to other in-memory structures for this arena */
struct free_entry *freelist;
u32 *rtt;
struct aligned_lock *map_locks;
struct nd_btt *nd_btt;
struct list_head list;
struct dentry *debugfs_dir;
/* Arena flags */
u32 flags;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 01:36:03 +00:00
struct mutex err_lock;
int log_index[2];
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 08:20:32 +00:00
};
struct badblocks;
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 08:20:32 +00:00
/**
* struct btt - handle for a BTT instance
* @btt_disk: Pointer to the gendisk for BTT device
* @arena_list: Head of the list of arenas
* @debugfs_dir: Debugfs dentry
* @nd_btt: Parent nd_btt struct
* @nlba: Number of logical blocks exposed to the upper layers
* after removing the amount of space needed by metadata
* @rawsize: Total size in bytes of the available backing device
* @lbasize: LBA size as requested and presented to upper layers.
* This is sector_size + size of any metadata.
* @sector_size: The Linux sector size - 512 or 4096
* @lanes: Per-lane spinlocks
* @init_lock: Mutex used for the BTT initialization
* @init_state: Flag describing the initialization state for the BTT
* @num_arenas: Number of arenas in the BTT instance
* @phys_bb: Pointer to the namespace's badblocks structure
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 08:20:32 +00:00
*/
struct btt {
struct gendisk *btt_disk;
struct list_head arena_list;
struct dentry *debugfs_dir;
struct nd_btt *nd_btt;
u64 nlba;
unsigned long long rawsize;
u32 lbasize;
u32 sector_size;
struct nd_region *nd_region;
struct mutex init_lock;
int init_state;
int num_arenas;
libnvdimm, btt: rework error clearing Clearing errors or badblocks during a BTT write requires sending an ACPI DSM, which means potentially sleeping. Since a BTT IO happens in atomic context (preemption disabled, spinlocks may be held), we cannot perform error clearing in the course of an IO. Due to this error clearing for BTT IOs has hitherto been disabled. In this patch we move error clearing out of the atomic section, and thus re-enable error clearing with BTTs. When we are about to add a block to the free list, we check if it was previously marked as an error, and if it was, we add it to the freelist, but also set a flag that says error clearing will be required. We then drop the lane (ending the atomic context), and send a zero buffer so that the error can be cleared. The error flag in the free list is protected by the nd 'lane', and is set only be a thread while it holds that lane. When the error is cleared, the flag is cleared, but while holding a mutex for that freelist index. When writing, we check for two things - 1/ If the freelist mutex is held or if the error flag is set. If so, this is an error block that is being (or about to be) cleared. 2/ If the block is a known badblock based on nsio->bb The second check is required because the BTT map error flag for a map entry only gets set when an error LBA is read. If we write to a new location that may not have the map error flag set, but still might be in the region's badblock list, we can trigger an EIO on the write, which is undesirable and completely avoidable. Cc: Jeff Moyer <jmoyer@redhat.com> Cc: Toshi Kani <toshi.kani@hpe.com> Cc: Dan Williams <dan.j.williams@intel.com> Signed-off-by: Vishal Verma <vishal.l.verma@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com>
2017-08-31 01:36:03 +00:00
struct badblocks *phys_bb;
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 08:20:32 +00:00
};
bool nd_btt_arena_is_valid(struct nd_btt *nd_btt, struct btt_sb *super);
int nd_btt_version(struct nd_btt *nd_btt, struct nd_namespace_common *ndns,
struct btt_sb *btt_sb);
#endif