Received-SPF: pass (google.com: domain of linux-ext4-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: quoted-printable
From:   Andreas Dilger <adilger@dilger.ca>
Mime-Version: 1.0 (1.0)
Subject: Re: [PATCH v5 5/6] ext4: improve cr 0 / cr 1 group scanning
Date:   Mon, 29 Mar 2021 22:10:36 -0600
Message-Id: <1A322CE3-C4FA-4140-BD34-7092F7629FB0@dilger.ca>
References: <CAD+ocbxteFqbHgajxBu9x_Rx4u2vNMdszVjvLvKBH4i47G9fOw@mail.gmail.com>
Cc:     Ext4 Developers List <linux-ext4@vger.kernel.org>,
        kernel test robot <lkp@intel.com>,
        Dan Carpenter <dan.carpenter@oracle.com>,
        Alex Zhuravlev <bzzz@whamcloud.com>
In-Reply-To: <CAD+ocbxteFqbHgajxBu9x_Rx4u2vNMdszVjvLvKBH4i47G9fOw@mail.gmail.com>
To:     harshad shirwadkar <harshadshirwadkar@gmail.com>
Precedence: bulk

On Mar 29, 2021, at 17:03, harshad shirwadkar <harshadshirwadkar@gmail.com> w=
rote:
>=20
> =EF=BB=BFAnother problem that I found in the patch was that when all group=
s are
> not loaded, the optimization concludes that CR0 and CR1 cannot be used
> for allocation (since these structures aren't populated). The problem
> goes away if we also mount with prefetch_block_bitmaps mount option.
> But I think if all the groups are not loaded, we should at least fall
> back to linear search at CR1. Maybe skipping CR0 is okay.

Alex's patch to skip cr0/1 if no groups are loaded is OK, as long as there i=
s
some group prefetch is triggered in the background, IMHO, and is there
for later allocations.=20

That avoids lots of scanning at mount time if the filesystem is mostly full,=

and the first allocation is not critical as long as this triggers group read=
s.

Do you have any idea what the performance impact of this is (eg.
latency to first write)?  Ideally there would be some group prefetch done
right at mount time (itable zero thread?) so that if writes are not
done *right* after mount there would not be any noticeable latency.

Cheers, Andreas

>> On Mon, Mar 29, 2021 at 12:32 PM harshad shirwadkar
>> <harshadshirwadkar@gmail.com> wrote:
>>=20
>> Okay sounds good. I'll do that!
>>=20
>>> On Mon, Mar 29, 2021 at 12:14 PM Andreas Dilger <adilger@dilger.ca> wrot=
e:
>>>=20
>>> On Mar 24, 2021, at 5:19 PM, Harshad Shirwadkar <harshadshirwadkar@gmail=
.com> wrote:
>>>>=20
>>>> Instead of traversing through groups linearly, scan groups in specific
>>>> orders at cr 0 and cr 1. At cr 0, we want to find groups that have the
>>>> largest free order >=3D the order of the request. So, with this patch,
>>>> we maintain lists for each possible order and insert each group into a
>>>> list based on the largest free order in its buddy bitmap. During cr 0
>>>> allocation, we traverse these lists in the increasing order of largest
>>>> free orders. This allows us to find a group with the best available cr
>>>> 0 match in constant time. If nothing can be found, we fallback to cr 1
>>>> immediately.
>>>>=20
>>>> At CR1, the story is slightly different. We want to traverse in the
>>>> order of increasing average fragment size. For CR1, we maintain a rb
>>>> tree of groupinfos which is sorted by average fragment size. Instead
>>>> of traversing linearly, at CR1, we traverse in the order of increasing
>>>> average fragment size, starting at the most optimal group. This brings
>>>> down cr 1 search complexity to log(num groups).
>>>>=20
>>>> For cr >=3D 2, we just perform the linear search as before. Also, in
>>>> case of lock contention, we intermittently fallback to linear search
>>>> even in CR 0 and CR 1 cases. This allows us to proceed during the
>>>> allocation path even in case of high contention.
>>>=20
>>> If you are refreshing this patch anyway, could you rename "mb_linear_lim=
it"
>>> to "mb_linear_groups" or "mb_max_linear_groups" or similar?  It otherwis=
e
>>> isn't clear what units the "limit" is in (could be MB/GB or something el=
se).
>>>=20
>>> Cheers, Andreas
>>>=20
>>>> There is an opportunity to do optimization at CR2 too. That's because
>>>> at CR2 we only consider groups where bb_free counter (number of free
>>>> blocks) is greater than the request extent size. That's left as future
>>>> work.
>>>>=20
>>>> All the changes introduced in this patch are protected under a new
>>>> mount option "mb_optimize_scan".
>>>>=20
>>>> With this patchset, following experiment was performed:
>>>>=20
>>>> Created a highly fragmented disk of size 65TB. The disk had no
>>>> contiguous 2M regions. Following command was run consecutively for 3
>>>> times:
>>>>=20
>>>> time dd if=3D/dev/urandom of=3Dfile bs=3D2M count=3D10
>>>>=20
>>>> Here are the results with and without cr 0/1 optimizations introduced
>>>> in this patch:
>>>>=20
>>>> |---------+------------------------------+---------------------------|
>>>> |         | Without CR 0/1 Optimizations | With CR 0/1 Optimizations |
>>>> |---------+------------------------------+---------------------------|
>>>> | 1st run | 5m1.871s                     | 2m47.642s                 |
>>>> | 2nd run | 2m28.390s                    | 0m0.611s                  |
>>>> | 3rd run | 2m26.530s                    | 0m1.255s                  |
>>>> |---------+------------------------------+---------------------------|
>>>>=20
>>>> Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@gmail.com>
>>>> Reported-by: kernel test robot <lkp@intel.com>
>>>> Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
>>>> Reviewed-by: Andreas Dilger <adilger@dilger.ca>
>>>> ---
>>>> fs/ext4/ext4.h    |  19 ++-
>>>> fs/ext4/mballoc.c | 381 ++++++++++++++++++++++++++++++++++++++++++++--
>>>> fs/ext4/mballoc.h |  17 ++-
>>>> fs/ext4/super.c   |  28 +++-
>>>> fs/ext4/sysfs.c   |   2 +
>>>> 5 files changed, 432 insertions(+), 15 deletions(-)
>>>>=20
>>>> diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
>>>> index 85eeeba3bca3..5930c8cb5159 100644
>>>> --- a/fs/ext4/ext4.h
>>>> +++ b/fs/ext4/ext4.h
>>>> @@ -162,7 +162,10 @@ enum SHIFT_DIRECTION {
>>>> #define EXT4_MB_USE_RESERVED          0x2000
>>>> /* Do strict check for free blocks while retrying block allocation */
>>>> #define EXT4_MB_STRICT_CHECK          0x4000
>>>> -
>>>> +/* Large fragment size list lookup succeeded at least once for cr =3D 0=
 */
>>>> +#define EXT4_MB_CR0_OPTIMIZED                0x8000
>>>> +/* Avg fragment size rb tree lookup succeeded at least once for cr =3D=
 1 */
>>>> +#define EXT4_MB_CR1_OPTIMIZED                0x00010000
>>>> struct ext4_allocation_request {
>>>>      /* target inode for block we're allocating */
>>>>      struct inode *inode;
>>>> @@ -1247,7 +1250,9 @@ struct ext4_inode_info {
>>>> #define EXT4_MOUNT2_JOURNAL_FAST_COMMIT       0x00000010 /* Journal fas=
t commit */
>>>> #define EXT4_MOUNT2_DAX_NEVER         0x00000020 /* Do not allow Direct=
 Access */
>>>> #define EXT4_MOUNT2_DAX_INODE         0x00000040 /* For printing option=
s only */
>>>> -
>>>> +#define EXT4_MOUNT2_MB_OPTIMIZE_SCAN 0x00000080 /* Optimize group
>>>> +                                                 * scanning in mballoc=

>>>> +                                                 */
>>>>=20
>>>> #define clear_opt(sb, opt)            EXT4_SB(sb)->s_mount_opt &=3D \
>>>>                                              ~EXT4_MOUNT_##opt
>>>> @@ -1527,9 +1532,14 @@ struct ext4_sb_info {
>>>>      unsigned int s_mb_free_pending;
>>>>      struct list_head s_freed_data_list;     /* List of blocks to be fr=
eed
>>>>                                                 after commit completed *=
/
>>>> +     struct rb_root s_mb_avg_fragment_size_root;
>>>> +     rwlock_t s_mb_rb_lock;
>>>> +     struct list_head *s_mb_largest_free_orders;
>>>> +     rwlock_t *s_mb_largest_free_orders_locks;
>>>>=20
>>>>      /* tunables */
>>>>      unsigned long s_stripe;
>>>> +     unsigned int s_mb_linear_limit;
>>>>      unsigned int s_mb_stream_request;
>>>>      unsigned int s_mb_max_to_scan;
>>>>      unsigned int s_mb_min_to_scan;
>>>> @@ -1553,6 +1563,8 @@ struct ext4_sb_info {
>>>>      atomic_t s_bal_goals;   /* goal hits */
>>>>      atomic_t s_bal_breaks;  /* too long searches */
>>>>      atomic_t s_bal_2orders; /* 2^order hits */
>>>> +     atomic_t s_bal_cr0_bad_suggestions;
>>>> +     atomic_t s_bal_cr1_bad_suggestions;
>>>>      atomic64_t s_bal_cX_groups_considered[4];
>>>>      atomic64_t s_bal_cX_hits[4];
>>>>      atomic64_t s_bal_cX_failed[4];          /* cX loop didn't find blo=
cks */
>>>> @@ -3309,11 +3321,14 @@ struct ext4_group_info {
>>>>      ext4_grpblk_t   bb_free;        /* total free blocks */
>>>>      ext4_grpblk_t   bb_fragments;   /* nr of freespace fragments */
>>>>      ext4_grpblk_t   bb_largest_free_order;/* order of largest frag in B=
G */
>>>> +     ext4_group_t    bb_group;       /* Group number */
>>>>      struct          list_head bb_prealloc_list;
>>>> #ifdef DOUBLE_CHECK
>>>>      void            *bb_bitmap;
>>>> #endif
>>>>      struct rw_semaphore alloc_sem;
>>>> +     struct rb_node  bb_avg_fragment_size_rb;
>>>> +     struct list_head bb_largest_free_order_node;
>>>>      ext4_grpblk_t   bb_counters[];  /* Nr of free power-of-two-block
>>>>                                       * regions, index is order.
>>>>                                       * bb_counters[3] =3D 5 means
>>>> diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
>>>> index 15127d815461..cbf9a89c0ef5 100644
>>>> --- a/fs/ext4/mballoc.c
>>>> +++ b/fs/ext4/mballoc.c
>>>> @@ -127,11 +127,50 @@
>>>> * smallest multiple of the stripe value (sbi->s_stripe) which is
>>>> * greater than the default mb_group_prealloc.
>>>> *
>>>> + * If "mb_optimize_scan" mount option is set, we maintain in memory gr=
oup info
>>>> + * structures in two data structures:
>>>> + *
>>>> + * 1) Array of largest free order lists (sbi->s_mb_largest_free_orders=
)
>>>> + *
>>>> + *    Locking: sbi->s_mb_largest_free_orders_locks(array of rw locks)
>>>> + *
>>>> + *    This is an array of lists where the index in the array represent=
s the
>>>> + *    largest free order in the buddy bitmap of the participating grou=
p infos of
>>>> + *    that list. So, there are exactly MB_NUM_ORDERS(sb) (which means t=
otal
>>>> + *    number of buddy bitmap orders possible) number of lists. Group-i=
nfos are
>>>> + *    placed in appropriate lists.
>>>> + *
>>>> + * 2) Average fragment size rb tree (sbi->s_mb_avg_fragment_size_root)=

>>>> + *
>>>> + *    Locking: sbi->s_mb_rb_lock (rwlock)
>>>> + *
>>>> + *    This is a red black tree consisting of group infos and the tree i=
s sorted
>>>> + *    by average fragment sizes (which is calculated as ext4_group_inf=
o->bb_free
>>>> + *    / ext4_group_info->bb_fragments).
>>>> + *
>>>> + * When "mb_optimize_scan" mount option is set, mballoc consults the a=
bove data
>>>> + * structures to decide the order in which groups are to be traversed f=
or
>>>> + * fulfilling an allocation request.
>>>> + *
>>>> + * At CR =3D 0, we look for groups which have the largest_free_order >=
=3D the order
>>>> + * of the request. We directly look at the largest free order list in t=
he data
>>>> + * structure (1) above where largest_free_order =3D order of the reque=
st. If that
>>>> + * list is empty, we look at remaining list in the increasing order of=

>>>> + * largest_free_order. This allows us to perform CR =3D 0 lookup in O(=
1) time.
>>>> + *
>>>> + * At CR =3D 1, we only consider groups where average fragment size > r=
equest
>>>> + * size. So, we lookup a group which has average fragment size just ab=
ove or
>>>> + * equal to request size using our rb tree (data structure 2) in O(log=
 N) time.
>>>> + *
>>>> + * If "mb_optimize_scan" mount option is not set, mballoc traverses gr=
oups in
>>>> + * linear order which requires O(N) search time for each CR 0 and CR 1=
 phase.
>>>> + *
>>>> * The regular allocator (using the buddy cache) supports a few tunables=
.
>>>> *
>>>> * /sys/fs/ext4/<partition>/mb_min_to_scan
>>>> * /sys/fs/ext4/<partition>/mb_max_to_scan
>>>> * /sys/fs/ext4/<partition>/mb_order2_req
>>>> + * /sys/fs/ext4/<partition>/mb_linear_limit
>>>> *
>>>> * The regular allocator uses buddy scan only if the request len is powe=
r of
>>>> * 2 blocks and the order of allocation is >=3D sbi->s_mb_order2_reqs. T=
he
>>>> @@ -149,6 +188,16 @@
>>>> * can be used for allocation. ext4_mb_good_group explains how the group=
s are
>>>> * checked.
>>>> *
>>>> + * When "mb_optimize_scan" is turned on, as mentioned above, the group=
s may not
>>>> + * get traversed linearly. That may result in subsequent allocations b=
eing not
>>>> + * close to each other. And so, the underlying device may get filled u=
p in a
>>>> + * non-linear fashion. While that may not matter on non-rotational dev=
ices, for
>>>> + * rotational devices that may result in higher seek times. "mb_linear=
_limit"
>>>> + * tells mballoc how many groups mballoc should search linearly before=

>>>> + * performing consulting above data structures for more efficient look=
ups. For
>>>> + * non rotational devices, this value defaults to 0 and for rotational=
 devices
>>>> + * this is set to MB_DEFAULT_LINEAR_LIMIT.
>>>> + *
>>>> * Both the prealloc space are getting populated as above. So for the fi=
rst
>>>> * request we will hit the buddy cache which will result in this preallo=
c
>>>> * space getting filled. The prealloc space is then later used for the
>>>> @@ -299,6 +348,8 @@
>>>> *  - bitlock on a group      (group)
>>>> *  - object (inode/locality) (object)
>>>> *  - per-pa lock             (pa)
>>>> + *  - cr0 lists lock         (cr0)
>>>> + *  - cr1 tree lock          (cr1)
>>>> *
>>>> * Paths:
>>>> *  - new pa
>>>> @@ -328,6 +379,9 @@
>>>> *    group
>>>> *        object
>>>> *
>>>> + *  - allocation path (ext4_mb_regular_allocator)
>>>> + *    group
>>>> + *    cr0/cr1
>>>> */
>>>> static struct kmem_cache *ext4_pspace_cachep;
>>>> static struct kmem_cache *ext4_ac_cachep;
>>>> @@ -351,6 +405,9 @@ static void ext4_mb_generate_from_freelist(struct s=
uper_block *sb, void *bitmap,
>>>>                                              ext4_group_t group);
>>>> static void ext4_mb_new_preallocation(struct ext4_allocation_context *a=
c);
>>>>=20
>>>> +static bool ext4_mb_good_group(struct ext4_allocation_context *ac,
>>>> +                            ext4_group_t group, int cr);
>>>> +
>>>> /*
>>>> * The algorithm using this percpu seq counter goes below:
>>>> * 1. We sample the percpu discard_pa_seq counter before trying for bloc=
k
>>>> @@ -744,6 +801,251 @@ static void ext4_mb_mark_free_simple(struct super=
_block *sb,
>>>>      }
>>>> }
>>>>=20
>>>> +static void ext4_mb_rb_insert(struct rb_root *root, struct rb_node *ne=
w,
>>>> +                     int (*cmp)(struct rb_node *, struct rb_node *))
>>>> +{
>>>> +     struct rb_node **iter =3D &root->rb_node, *parent =3D NULL;
>>>> +
>>>> +     while (*iter) {
>>>> +             parent =3D *iter;
>>>> +             if (cmp(new, *iter) > 0)
>>>> +                     iter =3D &((*iter)->rb_left);
>>>> +             else
>>>> +                     iter =3D &((*iter)->rb_right);
>>>> +     }
>>>> +
>>>> +     rb_link_node(new, parent, iter);
>>>> +     rb_insert_color(new, root);
>>>> +}
>>>> +
>>>> +static int
>>>> +ext4_mb_avg_fragment_size_cmp(struct rb_node *rb1, struct rb_node *rb2=
)
>>>> +{
>>>> +     struct ext4_group_info *grp1 =3D rb_entry(rb1,
>>>> +                                             struct ext4_group_info,
>>>> +                                             bb_avg_fragment_size_rb);=

>>>> +     struct ext4_group_info *grp2 =3D rb_entry(rb2,
>>>> +                                             struct ext4_group_info,
>>>> +                                             bb_avg_fragment_size_rb);=

>>>> +     int num_frags_1, num_frags_2;
>>>> +
>>>> +     num_frags_1 =3D grp1->bb_fragments ?
>>>> +             grp1->bb_free / grp1->bb_fragments : 0;
>>>> +     num_frags_2 =3D grp2->bb_fragments ?
>>>> +             grp2->bb_free / grp2->bb_fragments : 0;
>>>> +
>>>> +     return (num_frags_2 - num_frags_1);
>>>> +}
>>>> +
>>>> +/*
>>>> + * Reinsert grpinfo into the avg_fragment_size tree with new average
>>>> + * fragment size.
>>>> + */
>>>> +static void
>>>> +mb_update_avg_fragment_size(struct super_block *sb, struct ext4_group_=
info *grp)
>>>> +{
>>>> +     struct ext4_sb_info *sbi =3D EXT4_SB(sb);
>>>> +
>>>> +     if (!test_opt2(sb, MB_OPTIMIZE_SCAN) || grp->bb_free =3D=3D 0)
>>>> +             return;
>>>> +
>>>> +     write_lock(&sbi->s_mb_rb_lock);
>>>> +     if (!RB_EMPTY_NODE(&grp->bb_avg_fragment_size_rb)) {
>>>> +             rb_erase(&grp->bb_avg_fragment_size_rb,
>>>> +                             &sbi->s_mb_avg_fragment_size_root);
>>>> +             RB_CLEAR_NODE(&grp->bb_avg_fragment_size_rb);
>>>> +     }
>>>> +
>>>> +     ext4_mb_rb_insert(&sbi->s_mb_avg_fragment_size_root,
>>>> +             &grp->bb_avg_fragment_size_rb,
>>>> +             ext4_mb_avg_fragment_size_cmp);
>>>> +     write_unlock(&sbi->s_mb_rb_lock);
>>>> +}
>>>> +
>>>> +/*
>>>> + * Choose next group by traversing largest_free_order lists. Return 0 i=
f next
>>>> + * group was selected optimally. Return 1 if next group was not select=
ed
>>>> + * optimally. Updates *new_cr if cr level needs an update.
>>>> + */
>>>> +static int ext4_mb_choose_next_group_cr0(struct ext4_allocation_contex=
t *ac,
>>>> +             int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
>>>> +{
>>>> +     struct ext4_sb_info *sbi =3D EXT4_SB(ac->ac_sb);
>>>> +     struct ext4_group_info *iter, *grp;
>>>> +     int i;
>>>> +
>>>> +     if (ac->ac_status =3D=3D AC_STATUS_FOUND)
>>>> +             return 1;
>>>> +
>>>> +     if (unlikely(sbi->s_mb_stats && ac->ac_flags & EXT4_MB_CR0_OPTIMI=
ZED))
>>>> +             atomic_inc(&sbi->s_bal_cr0_bad_suggestions);
>>>> +
>>>> +     grp =3D NULL;
>>>> +     for (i =3D ac->ac_2order; i < MB_NUM_ORDERS(ac->ac_sb); i++) {
>>>> +             if (list_empty(&sbi->s_mb_largest_free_orders[i]))
>>>> +                     continue;
>>>> +             read_lock(&sbi->s_mb_largest_free_orders_locks[i]);
>>>> +             if (list_empty(&sbi->s_mb_largest_free_orders[i])) {
>>>> +                     read_unlock(&sbi->s_mb_largest_free_orders_locks[=
i]);
>>>> +                     continue;
>>>> +             }
>>>> +             grp =3D NULL;
>>>> +             list_for_each_entry(iter, &sbi->s_mb_largest_free_orders[=
i],
>>>> +                                 bb_largest_free_order_node) {
>>>> +                     if (sbi->s_mb_stats)
>>>> +                             atomic64_inc(&sbi->s_bal_cX_groups_consid=
ered[0]);
>>>> +                     if (likely(ext4_mb_good_group(ac, iter->bb_group,=
 0))) {
>>>> +                             grp =3D iter;
>>>> +                             break;
>>>> +                     }
>>>> +             }
>>>> +             read_unlock(&sbi->s_mb_largest_free_orders_locks[i]);
>>>> +             if (grp)
>>>> +                     break;
>>>> +     }
>>>> +
>>>> +     if (!grp) {
>>>> +             /* Increment cr and search again */
>>>> +             *new_cr =3D 1;
>>>> +     } else {
>>>> +             *group =3D grp->bb_group;
>>>> +             ac->ac_last_optimal_group =3D *group;
>>>> +             ac->ac_flags |=3D EXT4_MB_CR0_OPTIMIZED;
>>>> +     }
>>>> +     return 0;
>>>> +}
>>>> +
>>>> +/*
>>>> + * Choose next group by traversing average fragment size tree. Return 0=
 if next
>>>> + * group was selected optimally. Return 1 if next group could not sele=
cted
>>>> + * optimally (due to lock contention). Updates *new_cr if cr lvel need=
s an
>>>> + * update.
>>>> + */
>>>> +static int ext4_mb_choose_next_group_cr1(struct ext4_allocation_contex=
t *ac,
>>>> +             int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
>>>> +{
>>>> +     struct ext4_sb_info *sbi =3D EXT4_SB(ac->ac_sb);
>>>> +     int avg_fragment_size, best_so_far;
>>>> +     struct rb_node *node, *found;
>>>> +     struct ext4_group_info *grp;
>>>> +
>>>> +     /*
>>>> +      * If there is contention on the lock, instead of waiting for the=
 lock
>>>> +      * to become available, just continue searching lineraly. We'll r=
esume
>>>> +      * our rb tree search later starting at ac->ac_last_optimal_group=
.
>>>> +      */
>>>> +     if (!read_trylock(&sbi->s_mb_rb_lock))
>>>> +             return 1;
>>>> +
>>>> +     if (unlikely(ac->ac_flags & EXT4_MB_CR1_OPTIMIZED)) {
>>>> +             if (sbi->s_mb_stats)
>>>> +                     atomic_inc(&sbi->s_bal_cr1_bad_suggestions);
>>>> +             /* We have found something at CR 1 in the past */
>>>> +             grp =3D ext4_get_group_info(ac->ac_sb, ac->ac_last_optima=
l_group);
>>>> +             for (found =3D rb_next(&grp->bb_avg_fragment_size_rb); fo=
und !=3D NULL;
>>>> +                  found =3D rb_next(found)) {
>>>> +                     grp =3D rb_entry(found, struct ext4_group_info,
>>>> +                                    bb_avg_fragment_size_rb);
>>>> +                     if (sbi->s_mb_stats)
>>>> +                             atomic64_inc(&sbi->s_bal_cX_groups_consid=
ered[1]);
>>>> +                     if (likely(ext4_mb_good_group(ac, grp->bb_group, 1=
)))
>>>> +                             break;
>>>> +             }
>>>> +
>>>> +             goto done;
>>>> +     }
>>>> +
>>>> +     node =3D sbi->s_mb_avg_fragment_size_root.rb_node;
>>>> +     best_so_far =3D 0;
>>>> +     found =3D NULL;
>>>> +
>>>> +     while (node) {
>>>> +             grp =3D rb_entry(node, struct ext4_group_info,
>>>> +                            bb_avg_fragment_size_rb);
>>>> +             avg_fragment_size =3D 0;
>>>> +             /*
>>>> +              * Perform this check without locking, we'll lock later t=
o confirm.
>>>> +              */
>>>> +             if (ext4_mb_good_group(ac, grp->bb_group, 1)) {
>>>> +                     avg_fragment_size =3D grp->bb_fragments ?
>>>> +                             grp->bb_free / grp->bb_fragments : 0;
>>>> +                     if (!best_so_far || avg_fragment_size < best_so_f=
ar) {
>>>> +                             best_so_far =3D avg_fragment_size;
>>>> +                             found =3D node;
>>>> +                     }
>>>> +             }
>>>> +             if (avg_fragment_size > ac->ac_g_ex.fe_len)
>>>> +                     node =3D node->rb_right;
>>>> +             else
>>>> +                     node =3D node->rb_left;
>>>> +     }
>>>> +
>>>> +done:
>>>> +     if (found) {
>>>> +             grp =3D rb_entry(found, struct ext4_group_info,
>>>> +                            bb_avg_fragment_size_rb);
>>>> +             *group =3D grp->bb_group;
>>>> +             ac->ac_flags |=3D EXT4_MB_CR1_OPTIMIZED;
>>>> +     } else {
>>>> +             *new_cr =3D 2;
>>>> +     }
>>>> +
>>>> +     read_unlock(&sbi->s_mb_rb_lock);
>>>> +     ac->ac_last_optimal_group =3D *group;
>>>> +     return 0;
>>>> +}
>>>> +
>>>> +/*
>>>> + * ext4_mb_choose_next_group: choose next group for allocation.
>>>> + *
>>>> + * @ac        Allocation Context
>>>> + * @new_cr    This is an output parameter. If the there is no good gro=
up
>>>> + *            available at current CR level, this field is updated to i=
ndicate
>>>> + *            the new cr level that should be used.
>>>> + * @group     This is an input / output parameter. As an input it indi=
cates the
>>>> + *            last group used for allocation. As output, this field in=
dicates
>>>> + *            the next group that should be used.
>>>> + * @ngroups   Total number of groups
>>>> + */
>>>> +static void ext4_mb_choose_next_group(struct ext4_allocation_context *=
ac,
>>>> +             int *new_cr, ext4_group_t *group, ext4_group_t ngroups)
>>>> +{
>>>> +     int ret;
>>>> +
>>>> +     *new_cr =3D ac->ac_criteria;
>>>> +
>>>> +     if (!test_opt2(ac->ac_sb, MB_OPTIMIZE_SCAN) ||
>>>> +         *new_cr >=3D 2 ||
>>>> +         !ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS))
>>>> +             goto inc_and_return;
>>>> +
>>>> +     if (ac->ac_groups_linear_remaining) {
>>>> +             ac->ac_groups_linear_remaining--;
>>>> +             goto inc_and_return;
>>>> +     }
>>>> +
>>>> +     if (*new_cr =3D=3D 0) {
>>>> +             ret =3D ext4_mb_choose_next_group_cr0(ac, new_cr, group, n=
groups);
>>>> +             if (ret)
>>>> +                     goto inc_and_return;
>>>> +     }
>>>> +     if (*new_cr =3D=3D 1) {
>>>> +             ret =3D ext4_mb_choose_next_group_cr1(ac, new_cr, group, n=
groups);
>>>> +             if (ret)
>>>> +                     goto inc_and_return;
>>>> +     }
>>>> +     return;
>>>> +
>>>> +inc_and_return:
>>>> +     /*
>>>> +      * Artificially restricted ngroups for non-extent
>>>> +      * files makes group > ngroups possible on first loop.
>>>> +      */
>>>> +     *group =3D *group + 1;
>>>> +     if (*group >=3D ngroups)
>>>> +             *group =3D 0;
>>>> +}
>>>> +
>>>> /*
>>>> * Cache the order of the largest free extent we have available in this b=
lock
>>>> * group.
>>>> @@ -751,18 +1053,33 @@ static void ext4_mb_mark_free_simple(struct supe=
r_block *sb,
>>>> static void
>>>> mb_set_largest_free_order(struct super_block *sb, struct ext4_group_inf=
o *grp)
>>>> {
>>>> +     struct ext4_sb_info *sbi =3D EXT4_SB(sb);
>>>>      int i;
>>>> -     int bits;
>>>>=20
>>>> +     if (test_opt2(sb, MB_OPTIMIZE_SCAN) && grp->bb_largest_free_order=
 >=3D 0) {
>>>> +             write_lock(&sbi->s_mb_largest_free_orders_locks[
>>>> +                                           grp->bb_largest_free_order]=
);
>>>> +             list_del_init(&grp->bb_largest_free_order_node);
>>>> +             write_unlock(&sbi->s_mb_largest_free_orders_locks[
>>>> +                                           grp->bb_largest_free_order]=
);
>>>> +     }
>>>>      grp->bb_largest_free_order =3D -1; /* uninit */
>>>>=20
>>>> -     bits =3D MB_NUM_ORDERS(sb) - 1;
>>>> -     for (i =3D bits; i >=3D 0; i--) {
>>>> +     for (i =3D MB_NUM_ORDERS(sb) - 1; i >=3D 0; i--) {
>>>>              if (grp->bb_counters[i] > 0) {
>>>>                      grp->bb_largest_free_order =3D i;
>>>>                      break;
>>>>              }
>>>>      }
>>>> +     if (test_opt2(sb, MB_OPTIMIZE_SCAN) &&
>>>> +         grp->bb_largest_free_order >=3D 0 && grp->bb_free) {
>>>> +             write_lock(&sbi->s_mb_largest_free_orders_locks[
>>>> +                                           grp->bb_largest_free_order]=
);
>>>> +             list_add_tail(&grp->bb_largest_free_order_node,
>>>> +                   &sbi->s_mb_largest_free_orders[grp->bb_largest_free=
_order]);
>>>> +             write_unlock(&sbi->s_mb_largest_free_orders_locks[
>>>> +                                           grp->bb_largest_free_order]=
);
>>>> +     }
>>>> }
>>>>=20
>>>> static noinline_for_stack
>>>> @@ -818,6 +1135,7 @@ void ext4_mb_generate_buddy(struct super_block *sb=
,
>>>>      period =3D get_cycles() - period;
>>>>      atomic_inc(&sbi->s_mb_buddies_generated);
>>>>      atomic64_add(period, &sbi->s_mb_generation_time);
>>>> +     mb_update_avg_fragment_size(sb, grp);
>>>> }
>>>>=20
>>>> /* The buddy information is attached the buddy cache inode
>>>> @@ -1517,6 +1835,7 @@ static void mb_free_blocks(struct inode *inode, s=
truct ext4_buddy *e4b,
>>>>=20
>>>> done:
>>>>      mb_set_largest_free_order(sb, e4b->bd_info);
>>>> +     mb_update_avg_fragment_size(sb, e4b->bd_info);
>>>>      mb_check_buddy(e4b);
>>>> }
>>>>=20
>>>> @@ -1653,6 +1972,7 @@ static int mb_mark_used(struct ext4_buddy *e4b, s=
truct ext4_free_extent *ex)
>>>>      }
>>>>      mb_set_largest_free_order(e4b->bd_sb, e4b->bd_info);
>>>>=20
>>>> +     mb_update_avg_fragment_size(e4b->bd_sb, e4b->bd_info);
>>>>      ext4_set_bits(e4b->bd_bitmap, ex->fe_start, len0);
>>>>      mb_check_buddy(e4b);
>>>>=20
>>>> @@ -2347,17 +2667,21 @@ ext4_mb_regular_allocator(struct ext4_allocatio=
n_context *ac)
>>>>               * from the goal value specified
>>>>               */
>>>>              group =3D ac->ac_g_ex.fe_group;
>>>> +             ac->ac_last_optimal_group =3D group;
>>>> +             ac->ac_groups_linear_remaining =3D sbi->s_mb_linear_limit=
;
>>>>              prefetch_grp =3D group;
>>>>=20
>>>> -             for (i =3D 0; i < ngroups; group++, i++) {
>>>> -                     int ret =3D 0;
>>>> +             for (i =3D 0; i < ngroups; i++) {
>>>> +                     int ret =3D 0, new_cr;
>>>> +
>>>>                      cond_resched();
>>>> -                     /*
>>>> -                      * Artificially restricted ngroups for non-extent=

>>>> -                      * files makes group > ngroups possible on first l=
oop.
>>>> -                      */
>>>> -                     if (group >=3D ngroups)
>>>> -                             group =3D 0;
>>>> +
>>>> +                     ext4_mb_choose_next_group(ac, &new_cr, &group, ng=
roups);
>>>> +
>>>> +                     if (new_cr !=3D cr) {
>>>> +                             cr =3D new_cr;
>>>> +                             goto repeat;
>>>> +                     }
>>>>=20
>>>>                      /*
>>>>                       * Batch reads of the block allocation bitmaps
>>>> @@ -2578,6 +2902,8 @@ int ext4_seq_mb_stats_show(struct seq_file *seq, v=
oid *offset)
>>>>                 atomic64_read(&sbi->s_bal_cX_groups_considered[0]));
>>>>      seq_printf(seq, "\t\tuseless_loops: %llu\n",
>>>>                 atomic64_read(&sbi->s_bal_cX_failed[0]));
>>>> +     seq_printf(seq, "\t\tbad_suggestions: %u\n",
>>>> +                atomic_read(&sbi->s_bal_cr0_bad_suggestions));
>>>>=20
>>>>      seq_puts(seq, "\tcr1_stats:\n");
>>>>      seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_h=
its[1]));
>>>> @@ -2585,6 +2911,8 @@ int ext4_seq_mb_stats_show(struct seq_file *seq, v=
oid *offset)
>>>>                 atomic64_read(&sbi->s_bal_cX_groups_considered[1]));
>>>>      seq_printf(seq, "\t\tuseless_loops: %llu\n",
>>>>                 atomic64_read(&sbi->s_bal_cX_failed[1]));
>>>> +     seq_printf(seq, "\t\tbad_suggestions: %u\n",
>>>> +                atomic_read(&sbi->s_bal_cr1_bad_suggestions));
>>>>=20
>>>>      seq_puts(seq, "\tcr2_stats:\n");
>>>>      seq_printf(seq, "\t\thits: %llu\n", atomic64_read(&sbi->s_bal_cX_h=
its[2]));
>>>> @@ -2719,7 +3047,10 @@ int ext4_mb_add_groupinfo(struct super_block *sb=
, ext4_group_t group,
>>>>      INIT_LIST_HEAD(&meta_group_info[i]->bb_prealloc_list);
>>>>      init_rwsem(&meta_group_info[i]->alloc_sem);
>>>>      meta_group_info[i]->bb_free_root =3D RB_ROOT;
>>>> +     INIT_LIST_HEAD(&meta_group_info[i]->bb_largest_free_order_node);
>>>> +     RB_CLEAR_NODE(&meta_group_info[i]->bb_avg_fragment_size_rb);
>>>>      meta_group_info[i]->bb_largest_free_order =3D -1;  /* uninit */
>>>> +     meta_group_info[i]->bb_group =3D group;
>>>>=20
>>>>      mb_group_bb_bitmap_alloc(sb, meta_group_info[i], group);
>>>>      return 0;
>>>> @@ -2909,6 +3240,26 @@ int ext4_mb_init(struct super_block *sb)
>>>>              i++;
>>>>      } while (i < MB_NUM_ORDERS(sb));
>>>>=20
>>>> +     sbi->s_mb_avg_fragment_size_root =3D RB_ROOT;
>>>> +     sbi->s_mb_largest_free_orders =3D
>>>> +             kmalloc_array(MB_NUM_ORDERS(sb), sizeof(struct list_head)=
,
>>>> +                     GFP_KERNEL);
>>>> +     if (!sbi->s_mb_largest_free_orders) {
>>>> +             ret =3D -ENOMEM;
>>>> +             goto out;
>>>> +     }
>>>> +     sbi->s_mb_largest_free_orders_locks =3D
>>>> +             kmalloc_array(MB_NUM_ORDERS(sb), sizeof(rwlock_t),
>>>> +                     GFP_KERNEL);
>>>> +     if (!sbi->s_mb_largest_free_orders_locks) {
>>>> +             ret =3D -ENOMEM;
>>>> +             goto out;
>>>> +     }
>>>> +     for (i =3D 0; i < MB_NUM_ORDERS(sb); i++) {
>>>> +             INIT_LIST_HEAD(&sbi->s_mb_largest_free_orders[i]);
>>>> +             rwlock_init(&sbi->s_mb_largest_free_orders_locks[i]);
>>>> +     }
>>>> +     rwlock_init(&sbi->s_mb_rb_lock);
>>>>=20
>>>>      spin_lock_init(&sbi->s_md_lock);
>>>>      sbi->s_mb_free_pending =3D 0;
>>>> @@ -2961,6 +3312,10 @@ int ext4_mb_init(struct super_block *sb)
>>>>              spin_lock_init(&lg->lg_prealloc_lock);
>>>>      }
>>>>=20
>>>> +     if (blk_queue_nonrot(bdev_get_queue(sb->s_bdev)))
>>>> +             sbi->s_mb_linear_limit =3D 0;
>>>> +     else
>>>> +             sbi->s_mb_linear_limit =3D MB_DEFAULT_LINEAR_LIMIT;
>>>>      /* init file for buddy data */
>>>>      ret =3D ext4_mb_init_backend(sb);
>>>>      if (ret !=3D 0)
>>>> @@ -2972,6 +3327,8 @@ int ext4_mb_init(struct super_block *sb)
>>>>      free_percpu(sbi->s_locality_groups);
>>>>      sbi->s_locality_groups =3D NULL;
>>>> out:
>>>> +     kfree(sbi->s_mb_largest_free_orders);
>>>> +     kfree(sbi->s_mb_largest_free_orders_locks);
>>>>      kfree(sbi->s_mb_offsets);
>>>>      sbi->s_mb_offsets =3D NULL;
>>>>      kfree(sbi->s_mb_maxs);
>>>> @@ -3028,6 +3385,8 @@ int ext4_mb_release(struct super_block *sb)
>>>>              kvfree(group_info);
>>>>              rcu_read_unlock();
>>>>      }
>>>> +     kfree(sbi->s_mb_largest_free_orders);
>>>> +     kfree(sbi->s_mb_largest_free_orders_locks);
>>>>      kfree(sbi->s_mb_offsets);
>>>>      kfree(sbi->s_mb_maxs);
>>>>      iput(sbi->s_buddy_cache);
>>>> diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
>>>> index 68111a10cfee..02585e3cbcad 100644
>>>> --- a/fs/ext4/mballoc.h
>>>> +++ b/fs/ext4/mballoc.h
>>>> @@ -78,6 +78,18 @@
>>>> */
>>>> #define MB_DEFAULT_MAX_INODE_PREALLOC 512
>>>>=20
>>>> +/*
>>>> + * Number of groups to search linearly before performing group scannin=
g
>>>> + * optimization.
>>>> + */
>>>> +#define MB_DEFAULT_LINEAR_LIMIT              4
>>>> +
>>>> +/*
>>>> + * Minimum number of groups that should be present in the file system t=
o perform
>>>> + * group scanning optimizations.
>>>> + */
>>>> +#define MB_DEFAULT_LINEAR_SCAN_THRESHOLD     16
>>>> +
>>>> /*
>>>> * Number of valid buddy orders
>>>> */
>>>> @@ -166,11 +178,14 @@ struct ext4_allocation_context {
>>>>      /* copy of the best found extent taken before preallocation effort=
s */
>>>>      struct ext4_free_extent ac_f_ex;
>>>>=20
>>>> +     ext4_group_t ac_last_optimal_group;
>>>> +     __u32 ac_groups_considered;
>>>> +     __u32 ac_flags;         /* allocation hints */
>>>>      __u16 ac_groups_scanned;
>>>> +     __u16 ac_groups_linear_remaining;
>>>>      __u16 ac_found;
>>>>      __u16 ac_tail;
>>>>      __u16 ac_buddy;
>>>> -     __u16 ac_flags;         /* allocation hints */
>>>>      __u8 ac_status;
>>>>      __u8 ac_criteria;
>>>>      __u8 ac_2order;         /* if request is to allocate 2^N blocks an=
d
>>>> diff --git a/fs/ext4/super.c b/fs/ext4/super.c
>>>> index 3a2cd9fb7e73..6116640081c0 100644
>>>> --- a/fs/ext4/super.c
>>>> +++ b/fs/ext4/super.c
>>>> @@ -1687,7 +1687,7 @@ enum {
>>>>      Opt_dioread_nolock, Opt_dioread_lock,
>>>>      Opt_discard, Opt_nodiscard, Opt_init_itable, Opt_noinit_itable,
>>>>      Opt_max_dir_size_kb, Opt_nojournal_checksum, Opt_nombcache,
>>>> -     Opt_prefetch_block_bitmaps,
>>>> +     Opt_prefetch_block_bitmaps, Opt_mb_optimize_scan,
>>>> #ifdef CONFIG_EXT4_DEBUG
>>>>      Opt_fc_debug_max_replay, Opt_fc_debug_force
>>>> #endif
>>>> @@ -1788,6 +1788,7 @@ static const match_table_t tokens =3D {
>>>>      {Opt_nombcache, "nombcache"},
>>>>      {Opt_nombcache, "no_mbcache"},  /* for backward compatibility */
>>>>      {Opt_prefetch_block_bitmaps, "prefetch_block_bitmaps"},
>>>> +     {Opt_mb_optimize_scan, "mb_optimize_scan=3D%d"},
>>>>      {Opt_removed, "check=3Dnone"},    /* mount option from ext2/3 */
>>>>      {Opt_removed, "nocheck"},       /* mount option from ext2/3 */
>>>>      {Opt_removed, "reservation"},   /* mount option from ext2/3 */
>>>> @@ -1820,6 +1821,8 @@ static ext4_fsblk_t get_sb_block(void **data)
>>>> }
>>>>=20
>>>> #define DEFAULT_JOURNAL_IOPRIO (IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, 3))
>>>> +#define DEFAULT_MB_OPTIMIZE_SCAN     (-1)
>>>> +
>>>> static const char deprecated_msg[] =3D
>>>>      "Mount option \"%s\" will be removed by %s\n"
>>>>      "Contact linux-ext4@vger.kernel.org if you think we should keep it=
.\n";
>>>> @@ -2008,6 +2011,7 @@ static const struct mount_opts {
>>>>      {Opt_nombcache, EXT4_MOUNT_NO_MBCACHE, MOPT_SET},
>>>>      {Opt_prefetch_block_bitmaps, EXT4_MOUNT_PREFETCH_BLOCK_BITMAPS,
>>>>       MOPT_SET},
>>>> +     {Opt_mb_optimize_scan, EXT4_MOUNT2_MB_OPTIMIZE_SCAN, MOPT_GTE0},
>>>> #ifdef CONFIG_EXT4_DEBUG
>>>>      {Opt_fc_debug_force, EXT4_MOUNT2_JOURNAL_FAST_COMMIT,
>>>>       MOPT_SET | MOPT_2 | MOPT_EXT4_ONLY},
>>>> @@ -2092,6 +2096,7 @@ static int ext4_set_test_dummy_encryption(struct s=
uper_block *sb,
>>>> struct ext4_parsed_options {
>>>>      unsigned long journal_devnum;
>>>>      unsigned int journal_ioprio;
>>>> +     int mb_optimize_scan;
>>>> };
>>>>=20
>>>> static int handle_mount_opt(struct super_block *sb, char *opt, int toke=
n,
>>>> @@ -2388,6 +2393,13 @@ static int handle_mount_opt(struct super_block *=
sb, char *opt, int token,
>>>>              sbi->s_mount_opt |=3D m->mount_opt;
>>>>      } else if (token =3D=3D Opt_data_err_ignore) {
>>>>              sbi->s_mount_opt &=3D ~m->mount_opt;
>>>> +     } else if (token =3D=3D Opt_mb_optimize_scan) {
>>>> +             if (arg !=3D 0 && arg !=3D 1) {
>>>> +                     ext4_msg(sb, KERN_WARNING,
>>>> +                              "mb_optimize_scan should be set to 0 or 1=
.");
>>>> +                     return -1;
>>>> +             }
>>>> +             parsed_opts->mb_optimize_scan =3D arg;
>>>>      } else {
>>>>              if (!args->from)
>>>>                      arg =3D 1;
>>>> @@ -4034,6 +4046,7 @@ static int ext4_fill_super(struct super_block *sb=
, void *data, int silent)
>>>>      /* Set defaults for the variables that will be set during parsing *=
/
>>>>      parsed_opts.journal_ioprio =3D DEFAULT_JOURNAL_IOPRIO;
>>>>      parsed_opts.journal_devnum =3D 0;
>>>> +     parsed_opts.mb_optimize_scan =3D DEFAULT_MB_OPTIMIZE_SCAN;
>>>>=20
>>>>      if ((data && !orig_data) || !sbi)
>>>>              goto out_free_base;
>>>> @@ -4984,6 +4997,19 @@ static int ext4_fill_super(struct super_block *s=
b, void *data, int silent)
>>>>      ext4_fc_replay_cleanup(sb);
>>>>=20
>>>>      ext4_ext_init(sb);
>>>> +
>>>> +     /*
>>>> +      * Enable optimize_scan if number of groups is > threshold. This c=
an be
>>>> +      * turned off by passing "mb_optimize_scan=3D0". This can also be=

>>>> +      * turned on forcefully by passing "mb_optimize_scan=3D1".
>>>> +      */
>>>> +     if (parsed_opts.mb_optimize_scan =3D=3D 1)
>>>> +             set_opt2(sb, MB_OPTIMIZE_SCAN);
>>>> +     else if (parsed_opts.mb_optimize_scan =3D=3D 0)
>>>> +             clear_opt2(sb, MB_OPTIMIZE_SCAN);
>>>> +     else if (sbi->s_groups_count >=3D MB_DEFAULT_LINEAR_SCAN_THRESHOL=
D)
>>>> +             set_opt2(sb, MB_OPTIMIZE_SCAN);
>>>> +
>>>>      err =3D ext4_mb_init(sb);
>>>>      if (err) {
>>>>              ext4_msg(sb, KERN_ERR, "failed to initialize mballoc (%d)"=
,
>>>> diff --git a/fs/ext4/sysfs.c b/fs/ext4/sysfs.c
>>>> index 59ca9d73b42f..16b8a838f631 100644
>>>> --- a/fs/ext4/sysfs.c
>>>> +++ b/fs/ext4/sysfs.c
>>>> @@ -213,6 +213,7 @@ EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs=
);
>>>> EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
>>>> EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
>>>> EXT4_RW_ATTR_SBI_UI(mb_max_inode_prealloc, s_mb_max_inode_prealloc);
>>>> +EXT4_RW_ATTR_SBI_UI(mb_linear_limit, s_mb_linear_limit);
>>>> EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
>>>> EXT4_ATTR(trigger_fs_error, 0200, trigger_test_error);
>>>> EXT4_RW_ATTR_SBI_UI(err_ratelimit_interval_ms, s_err_ratelimit_state.in=
terval);
>>>> @@ -260,6 +261,7 @@ static struct attribute *ext4_attrs[] =3D {
>>>>      ATTR_LIST(mb_stream_req),
>>>>      ATTR_LIST(mb_group_prealloc),
>>>>      ATTR_LIST(mb_max_inode_prealloc),
>>>> +     ATTR_LIST(mb_linear_limit),
>>>>      ATTR_LIST(max_writeback_mb_bump),
>>>>      ATTR_LIST(extent_max_zeroout_kb),
>>>>      ATTR_LIST(trigger_fs_error),
>>>> --
>>>> 2.31.0.291.g576ba9dcdaf-goog
>>>>=20
>>>=20
>>>=20
>>> Cheers, Andreas
>>>=20
>>>=20
>>>=20
>>>=20
>>>=20