Re: [PATCH v3 13/16] bcache: add bucket_size_hi into struct cache_sb_disk for large bucket

[Coly Li <colyli@suse.de>]

On 2020/7/16 15:02, Hannes Reinecke wrote:
> On 7/16/20 8:41 AM, Coly Li wrote:
>> On 2020/7/16 14:15, Hannes Reinecke wrote:
>>> On 7/15/20 4:30 PM, colyli@suse.de wrote:
>>>> From: Coly Li <colyli@suse.de>
>>>>
>>>> The large bucket feature is to extend bucket_size from 16bit to 32bit.
>>>>
>>>> When create cache device on zoned device (e.g. zoned NVMe SSD), making
>>>> a single bucket cover one or more zones of the zoned device is the
>>>> simplest way to support zoned device as cache by bcache.
>>>>
>>>> But current maximum bucket size is 16MB and a typical zone size of
>>>> zoned
>>>> device is 256MB, this is the major motiviation to extend bucket size to
>>>> a larger bit width.
>>>>
>>>> This patch is the basic and first change to support large bucket size,
>>>> the major changes it makes are,
>>>> - Add BCH_FEATURE_INCOMPAT_LARGE_BUCKET for the large bucket feature,
>>>>     INCOMPAT means it introduces incompatible on-disk format change.
>>>> - Add BCH_FEATURE_INCOMPAT_FUNCS(large_bucket, LARGE_BUCKET) routines.
>>>> - Adds __le32 bucket_size_hi into struct cache_sb_disk at offset 0x8d0
>>>>     for the on-disk super block format.
>>>> - For the in-memory super block struct cache_sb, member bucket_size is
>>>>     extended from __u16 to __32.
>>>> - Add get_bucket_size() to combine the bucket_size and bucket_size_hi
>>>>     from struct cache_sb_disk into an unsigned int value.
>>>>
>>>> Since we already have large bucket size helpers meta_bucket_pages(),
>>>> meta_bucket_bytes() and alloc_meta_bucket_pages(), they make sure when
>>>> bucket size > 8MB, the memory allocation for bcache meta data bucket
>>>> won't fail no matter how large the bucket size extended. So these meta
>>>> data buckets are handled properly when the bucket size width increase
>>>> from 16bit to 32bit, we don't need to worry about them.
>>>>
>>>> Signed-off-by: Coly Li <colyli@suse.de>
>>>> ---
>>>>    drivers/md/bcache/alloc.c    |  2 +-
>>>>    drivers/md/bcache/features.c | 22 ++++++++++++++++++++++
>>>>    drivers/md/bcache/features.h |  9 ++++++---
>>>>    drivers/md/bcache/movinggc.c |  4 ++--
>>>>    drivers/md/bcache/super.c    | 23 +++++++++++++++++++----
>>>>    include/uapi/linux/bcache.h  |  3 ++-
>>>>    6 files changed, 52 insertions(+), 11 deletions(-)
>>>>    create mode 100644 drivers/md/bcache/features.c
>>>>
>>>> diff --git a/drivers/md/bcache/alloc.c b/drivers/md/bcache/alloc.c
>>>> index a1df0d95151c..52035a78d836 100644
>>>> --- a/drivers/md/bcache/alloc.c
>>>> +++ b/drivers/md/bcache/alloc.c
>>>> @@ -87,7 +87,7 @@ void bch_rescale_priorities(struct cache_set *c, int
>>>> sectors)
>>>>    {
>>>>        struct cache *ca;
>>>>        struct bucket *b;
>>>> -    unsigned int next = c->nbuckets * c->sb.bucket_size / 1024;
>>>> +    unsigned long next = c->nbuckets * c->sb.bucket_size / 1024;
>>>>        unsigned int i;
>>>>        int r;
>>>>    diff --git a/drivers/md/bcache/features.c
>>>> b/drivers/md/bcache/features.c
>>>> new file mode 100644
>>>> index 000000000000..ba53944bb390
>>>> --- /dev/null
>>>> +++ b/drivers/md/bcache/features.c
>>>> @@ -0,0 +1,22 @@
>>>> +// SPDX-License-Identifier: GPL-2.0
>>>> +/*
>>>> + * Feature set bits and string conversion.
>>>> + * Inspired by ext4's features compat/incompat/ro_compat related code.
>>>> + *
>>>> + * Copyright 2020 Coly Li <colyli@suse.de>
>>>> + *
>>>> + */
>>>> +#include <linux/bcache.h>
>>>> +#include "bcache.h"
>>>> +
>>>> +struct feature {
>>>> +    int        compat;
>>>> +    unsigned int    mask;
>>>> +    const char    *string;
>>>> +};
>>>> +
>>>> +static struct feature feature_list[] = {
>>>> +    {BCH_FEATURE_INCOMPAT, BCH_FEATURE_INCOMPAT_LARGE_BUCKET,
>>>> +        "large_bucket"},
>>>> +    {0, 0, 0 },
>>>> +};
>>>> diff --git a/drivers/md/bcache/features.h
>>>> b/drivers/md/bcache/features.h
>>>> index ae7df37b9862..dca052cf5203 100644
>>>> --- a/drivers/md/bcache/features.h
>>>> +++ b/drivers/md/bcache/features.h
>>>> @@ -11,9 +11,13 @@
>>>>    #define BCH_FEATURE_INCOMPAT        2
>>>>    #define BCH_FEATURE_TYPE_MASK        0x03
>>>>    +/* Feature set definition */
>>>> +/* Incompat feature set */
>>>> +#define BCH_FEATURE_INCOMPAT_LARGE_BUCKET    0x0001 /* 32bit bucket
>>>> size */
>>>> +
>>>>    #define BCH_FEATURE_COMPAT_SUUP        0
>>>>    #define BCH_FEATURE_RO_COMPAT_SUUP    0
>>>> -#define BCH_FEATURE_INCOMPAT_SUUP    0
>>>> +#define BCH_FEATURE_INCOMPAT_SUUP    BCH_FEATURE_INCOMPAT_LARGE_BUCKET
>>>>      #define BCH_HAS_COMPAT_FEATURE(sb, mask) \
>>>>            ((sb)->feature_compat & (mask))
>>>> @@ -22,8 +26,6 @@
>>>>    #define BCH_HAS_INCOMPAT_FEATURE(sb, mask) \
>>>>            ((sb)->feature_incompat & (mask))
>>>>    -/* Feature set definition */
>>>> -
>>>>    #define BCH_FEATURE_COMPAT_FUNCS(name, flagname) \
>>>>    static inline int bch_has_feature_##name(struct cache_sb *sb) \
>>>>    { \
>>>> @@ -75,4 +77,5 @@ static inline void bch_clear_feature_##name(struct
>>>> cache_sb *sb) \
>>>>            ~BCH##_FEATURE_INCOMPAT_##flagname; \
>>>>    }
>>>>    +BCH_FEATURE_INCOMPAT_FUNCS(large_bucket, LARGE_BUCKET);
>>>>    #endif
>>>> diff --git a/drivers/md/bcache/movinggc.c
>>>> b/drivers/md/bcache/movinggc.c
>>>> index b7dd2d75f58c..5872d6470470 100644
>>>> --- a/drivers/md/bcache/movinggc.c
>>>> +++ b/drivers/md/bcache/movinggc.c
>>>> @@ -206,8 +206,8 @@ void bch_moving_gc(struct cache_set *c)
>>>>        mutex_lock(&c->bucket_lock);
>>>>          for_each_cache(ca, c, i) {
>>>> -        unsigned int sectors_to_move = 0;
>>>> -        unsigned int reserve_sectors = ca->sb.bucket_size *
>>>> +        unsigned long sectors_to_move = 0;
>>>> +        unsigned long reserve_sectors = ca->sb.bucket_size *
>>>>                     fifo_used(&ca->free[RESERVE_MOVINGGC]);
>>>>              ca->heap.used = 0;
>>>> diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
>>>> index 02901d0ae8e2..e0da52f8e8c9 100644
>>>> --- a/drivers/md/bcache/super.c
>>>> +++ b/drivers/md/bcache/super.c
>>>> @@ -60,6 +60,17 @@ struct workqueue_struct *bch_journal_wq;
>>>>      /* Superblock */
>>>>    +static unsigned int get_bucket_size(struct cache_sb *sb, struct
>>>> cache_sb_disk *s)
>>>> +{
>>>> +    unsigned int bucket_size = le16_to_cpu(s->bucket_size);
>>>> +
>>>> +    if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES &&
>>>> +         bch_has_feature_large_bucket(sb))
>>>> +        bucket_size |= le32_to_cpu(s->bucket_size_hi) << 16;
>>>> +
>>>> +    return bucket_size;
>>>> +}
>>>> +
>>>>    static const char *read_super_common(struct cache_sb *sb,  struct
>>>> block_device *bdev,
>>>>                         struct cache_sb_disk *s)
>>>>    {
>>> That is a bit unfortunate; bucket_size_hi is 32 bits, so we might end up
>>> with an overflow here if bucket_size_hi is larger that USHRT_MAX.
>>>
>>
>> In bcache-tools, the maximum value of bucket_size is restricted to
>> UINT_MAX in make_bcache(),
>>     case 'b':
>>         bucket_size =
>>             hatoi_validate(optarg, "bucket size", UINT_MAX);
>>         break;
>>
>> So the overflow won't happen if people use bcache program to make the
>> cache device.
>>
>> If people want to modify bucket_size_hi themselves, in
>> read_super_common(), there are several checks to make sure the hacked
>> cache_sb->bucket_size composed from the hacked bucket_size_hi won't mess
>> up whole kernel,
>> 1) should be power of 2
>>     if (!is_power_of_2(sb->bucket_size))
>> 2) should large than page size
>>     if (sb->bucket_size < PAGE_SECTORS)
>> 3) should match sb->nbuckets
>>     if (get_capacity(bdev->bd_disk) <
>>         sb->bucket_size * sb->nbuckets)
>> 4) no overlap with super block
>>     if (sb->first_bucket * sb->bucket_size < 16)
>>
>> Therefore the overflow won't happen, and even happen by hacking it won't
>> panic kernel.
>>
>>> So to avoid overflow either make bucket_size_hi 16 bit, too, or define
>>> this feature such that the original bucket_size field is ignored and
>>> just the new size field is used.
>>>
>>
>> Now cache_sb_disk is for on-disk format, it contains 16bit bucket_size
>> and 32bit bucket_size_hi. cache_sb is for in-memory object only, it
>> simply has a 32bit bucket_size.
>>
>> Indeed I planed to set cache_sb->bucket_size to be uint64_t, but I feel
>> people won't use bucket size > 1TB, and finally make it to be a 32bit
>> value.
>>
>> I will add more code comments to explain why there won't be overflow in
>> get_bucket_size() because user space tool limits the maximum size.
>>
> So why not make bucket_size_hi 16-bit?
> That would clear up the ambiguity, and you wouldn't need to add comments
> ...

I did seriously think maybe many years later for some crazy reason
people do want >4TB bucket size.... also I though 48bit bucket size
should be large enough in whole Linux life cycle....

Maybe I can set bucket_size_hi as 16bit and pad the following 2 bytes
for the on-disk format, or just make buket_size_hi as 16bit if I was
over thoughtful.

Coly Li