Re: [RFC v2 03/83] Add super.h.

From: Darrick J. Wong
Date: Thu Mar 15 2018 - 01:00:28 EST


On Sat, Mar 10, 2018 at 10:17:44AM -0800, Andiry Xu wrote:
> From: Andiry Xu <jix024@xxxxxxxxxxx>
>
> This header file defines NOVA persistent and volatile superblock
> data structures.
>
> It also defines NOVA block layout:
>
> Page 0: Superblock
> Page 1: Reserved inodes
> Page 2 - 15: Reserved
> Page 16 - 31: Inode table pointers
> Page 32 - 47: Journal address pointers
> Page 48 - 63: Reserved
> Pages n-2: Replicate reserved inodes
> Pages n-1: Replicate superblock
>
> Other pages are for normal inodes, logs and data.
>
> Signed-off-by: Andiry Xu <jix024@xxxxxxxxxxx>
> ---
> fs/nova/super.h | 149 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 149 insertions(+)
> create mode 100644 fs/nova/super.h
>
> diff --git a/fs/nova/super.h b/fs/nova/super.h
> new file mode 100644
> index 0000000..cb53908
> --- /dev/null
> +++ b/fs/nova/super.h
> @@ -0,0 +1,149 @@
> +#ifndef __SUPER_H
> +#define __SUPER_H
> +/*
> + * Structure of the NOVA super block in PMEM
> + *
> + * The fields are partitioned into static and dynamic fields. The static fields
> + * never change after file system creation. This was primarily done because
> + * nova_get_block() returns NULL if the block offset is 0 (helps in catching
> + * bugs). So if we modify any field using journaling (for consistency), we
> + * will have to modify s_sum which is at offset 0. So journaling code fails.
> + * This (static+dynamic fields) is a temporary solution and can be avoided
> + * once the file system becomes stable and nova_get_block() returns correct
> + * pointers even for offset 0.
> + */
> +struct nova_super_block {
> + /* static fields. they never change after file system creation.
> + * checksum only validates up to s_start_dynamic field below
> + */
> + __le32 s_sum; /* checksum of this sb */
> + __le32 s_magic; /* magic signature */
> + __le32 s_padding32;
> + __le32 s_blocksize; /* blocksize in bytes */
> + __le64 s_size; /* total size of fs in bytes */
> + char s_volume_name[16]; /* volume name */
> +
> + /* all the dynamic fields should go here */
> + __le64 s_epoch_id; /* Epoch ID */
> +
> + /* s_mtime and s_wtime should be together and their order should not be
> + * changed. we use an 8 byte write to update both of them atomically
> + */
> + __le32 s_mtime; /* mount time */
> + __le32 s_wtime; /* write time */

Hmmm, 32-bit timestamps? 2038 isn't that far away...

> +} __attribute((__packed__));
> +
> +#define NOVA_SB_SIZE 512 /* must be power of two */
> +
> +/* ======================= Reserved blocks ========================= */
> +
> +/*
> + * Page 0 contains super blocks;
> + * Page 1 contains reserved inodes;
> + * Page 2 - 15 are reserved.
> + * Page 16 - 31 contain pointers to inode tables.
> + * Page 32 - 47 contain pointers to journal pages.
> + */
> +#define HEAD_RESERVED_BLOCKS 64
> +#define NUM_JOURNAL_PAGES 16
> +
> +#define SUPER_BLOCK_START 0 // Superblock
> +#define RESERVE_INODE_START 1 // Reserved inodes
> +#define INODE_TABLE_START 16 // inode table pointers
> +#define JOURNAL_START 32 // journal pointer table
> +
> +/* For replica super block and replica reserved inodes */
> +#define TAIL_RESERVED_BLOCKS 2
> +
> +/* ======================= Reserved inodes ========================= */
> +
> +/* We have space for 31 reserved inodes */
> +#define NOVA_ROOT_INO (1)
> +#define NOVA_INODETABLE_INO (2) /* Fake inode associated with inode
> + * stroage. We need this because our
> + * allocator requires inode to be
> + * associated with each allocation.
> + * The data actually lives in linked
> + * lists in INODE_TABLE_START. */
> +#define NOVA_BLOCKNODE_INO (3) /* Storage for allocator state */
> +#define NOVA_LITEJOURNAL_INO (4) /* Storage for lightweight journals */
> +#define NOVA_INODELIST_INO (5) /* Storage for Inode free list */
> +
> +
> +/* Normal inode starts at 32 */
> +#define NOVA_NORMAL_INODE_START (32)

I've been wondering this whole time, why not make the inode number the
byte offset into the pmem? Then you don't have to lose the last 8 bytes
of each inode block to point to the next one.

--D

> +
> +
> +
> +/*
> + * NOVA super-block data in DRAM
> + */
> +struct nova_sb_info {
> + struct super_block *sb; /* VFS super block */
> + struct nova_super_block *nova_sb; /* DRAM copy of SB */
> + struct block_device *s_bdev;
> + struct dax_device *s_dax_dev;
> +
> + /*
> + * base physical and virtual address of NOVA (which is also
> + * the pointer to the super block)
> + */
> + phys_addr_t phys_addr;
> + void *virt_addr;
> + void *replica_reserved_inodes_addr;
> + void *replica_sb_addr;
> +
> + unsigned long num_blocks;
> +
> + /* Mount options */
> + unsigned long bpi;
> + unsigned long blocksize;
> + unsigned long initsize;
> + unsigned long s_mount_opt;
> + kuid_t uid; /* Mount uid for root directory */
> + kgid_t gid; /* Mount gid for root directory */
> + umode_t mode; /* Mount mode for root directory */
> + atomic_t next_generation;
> + /* inode tracking */
> + unsigned long s_inodes_used_count;
> + unsigned long head_reserved_blocks;
> + unsigned long tail_reserved_blocks;
> +
> + struct mutex s_lock; /* protects the SB's buffer-head */
> +
> + int cpus;
> +
> + /* Current epoch. volatile guarantees visibility */
> + volatile u64 s_epoch_id;
> +
> + /* ZEROED page for cache page initialized */
> + void *zeroed_page;
> +};
> +
> +static inline struct nova_sb_info *NOVA_SB(struct super_block *sb)
> +{
> + return sb->s_fs_info;
> +}
> +
> +static inline struct nova_super_block
> +*nova_get_redund_super(struct super_block *sb)
> +{
> + struct nova_sb_info *sbi = NOVA_SB(sb);
> +
> + return (struct nova_super_block *)(sbi->replica_sb_addr);
> +}
> +
> +
> +/* If this is part of a read-modify-write of the super block,
> + * nova_memunlock_super() before calling!
> + */
> +static inline struct nova_super_block *nova_get_super(struct super_block *sb)
> +{
> + struct nova_sb_info *sbi = NOVA_SB(sb);
> +
> + return (struct nova_super_block *)sbi->virt_addr;
> +}
> +
> +extern void nova_error_mng(struct super_block *sb, const char *fmt, ...);
> +
> +#endif
> --
> 2.7.4
>