On 10/18/2023 2:17 AM, Bart Van Assche wrote:
> - * Write life time hint values.
> - * Stored in struct inode as u8.
> - */
> -enum rw_hint {
> - WRITE_LIFE_NOT_SET = 0,
> - WRITE_LIFE_NONE = RWH_WRITE_LIFE_NONE,
> - WRITE_LIFE_SHORT = RWH_WRITE_LIFE_SHORT,
> - WRITE_LIFE_MEDIUM = RWH_WRITE_LIFE_MEDIUM,
> - WRITE_LIFE_LONG = RWH_WRITE_LIFE_LONG,
> - WRITE_LIFE_EXTREME = RWH_WRITE_LIFE_EXTREME,
> -};
> -
> /* Match RWF_* bits to IOCB bits */
> #define IOCB_HIPRI (__force int) RWF_HIPRI
> #define IOCB_DSYNC (__force int) RWF_DSYNC
> @@ -677,7 +665,7 @@ struct inode {
> spinlock_t i_lock; /* i_blocks, i_bytes, maybe i_size */
> unsigned short i_bytes;
> u8 i_blkbits;
> - u8 i_write_hint;
> + enum rw_hint i_write_hint;
> blkcnt_t i_blocks;
>
> #ifdef __NEED_I_SIZE_ORDERED
> diff --git a/include/linux/rw_hint.h b/include/linux/rw_hint.h
> new file mode 100644
> index 000000000000..4a7d28945973
> --- /dev/null
> +++ b/include/linux/rw_hint.h
> @@ -0,0 +1,20 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _LINUX_RW_HINT_H
> +#define _LINUX_RW_HINT_H
> +
> +#include <linux/build_bug.h>
> +#include <linux/compiler_attributes.h>
> +
> +/* Block storage write lifetime hint values. */
> +enum rw_hint {
> + WRITE_LIFE_NOT_SET = 0, /* RWH_WRITE_LIFE_NOT_SET */
> + WRITE_LIFE_NONE = 1, /* RWH_WRITE_LIFE_NONE */
> + WRITE_LIFE_SHORT = 2, /* RWH_WRITE_LIFE_SHORT */
> + WRITE_LIFE_MEDIUM = 3, /* RWH_WRITE_LIFE_MEDIUM */
> + WRITE_LIFE_LONG = 4, /* RWH_WRITE_LIFE_LONG */
> + WRITE_LIFE_EXTREME = 5, /* RWH_WRITE_LIFE_EXTREME */
> +} __packed;
> +
> +static_assert(sizeof(enum rw_hint) == 1);
Does it make sense to do away with these, and have temperature-neutral
names instead e.g., WRITE_LIFE_1, WRITE_LIFE_2?
With the current choice:
- If the count goes up (beyond 5 hints), infra can scale fine but these
names do not. Imagine ULTRA_EXTREME after EXTREME.
- Applications or in-kernel users can specify LONG hint with data that
actually has a SHORT lifetime. Nothing really ensures that LONG is
really LONG.
Temperature-neutral names seem more generic/scalable and do not present
the unnecessary need to be accurate with relative temperatures.
