Thanks Artem for reviewing the patches.
On Fri, Aug 28, 2020 at 6:57 AM Благодаренко Артём
<artem.blagodarenko@xxxxxxxxx> wrote:
>
> Hello Harshad,
>
> Thank you for these useful patches. I am still reviewing them. This one looks good for me, but one place looks strange. See bellow.
>
> > On 19 Aug 2020, at 10:31, Harshad Shirwadkar <harshadshirwadkar@xxxxxxxxx> wrote:
> >
> > Freespace trees can occupy a lot of memory with as the fragmentation
> > increases. This patch adds a sysfs file to monitor the memory usage of
> > the freespace tree allocator. Also, added a sysfs config to control
> > maximum memory that the allocator can use. If the allocator exceeds
> > this threshold, file system enters "FRSP_MEM_CRUNCH" state. The next
> > patch in the series performs LRU eviction when this state is reached.
> >
> > Signed-off-by: Harshad Shirwadkar <harshadshirwadkar@xxxxxxxxx>
> > ---
> > fs/ext4/ext4.h | 8 ++++++++
> > fs/ext4/mballoc.c | 20 ++++++++++++++++++++
> > fs/ext4/mballoc.h | 4 ++++
> > fs/ext4/sysfs.c | 11 +++++++++++
> > 4 files changed, 43 insertions(+)
> >
> > diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
> > index 8cfe089ebea6..45fc3b230357 100644
> > --- a/fs/ext4/ext4.h
> > +++ b/fs/ext4/ext4.h
> > @@ -1206,6 +1206,12 @@ struct ext4_inode_info {
> > * allocator off)
> > */
> >
> > +#define EXT4_MOUNT2_FRSP_MEM_CRUNCH 0x00000040 /*
> > + * Freespace tree allocator
> > + * is in a tight memory
> > + * situation.
> > + */
> > +
> > #define clear_opt(sb, opt) EXT4_SB(sb)->s_mount_opt &= \
> > ~EXT4_MOUNT_##opt
> > #define set_opt(sb, opt) EXT4_SB(sb)->s_mount_opt |= \
> > @@ -1589,6 +1595,8 @@ struct ext4_sb_info {
> > atomic_t s_mb_num_frsp_trees_cached;
> > struct list_head s_mb_uncached_trees;
> > u32 s_mb_frsp_cache_aggression;
> > + atomic_t s_mb_num_fragments;
> > + u32 s_mb_frsp_mem_limit;
> >
> > /* workqueue for reserved extent conversions (buffered io) */
> > struct workqueue_struct *rsv_conversion_wq;
> > diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
> > index fa027b626abe..aada6838cafd 100644
> > --- a/fs/ext4/mballoc.c
> > +++ b/fs/ext4/mballoc.c
> > @@ -869,6 +869,7 @@ void ext4_mb_frsp_print_tree_len(struct super_block *sb,
> > static struct ext4_frsp_node *ext4_mb_frsp_alloc_node(struct super_block *sb)
> > {
> > struct ext4_frsp_node *node;
> > + struct ext4_sb_info *sbi = EXT4_SB(sb);
> >
> > node = kmem_cache_alloc(ext4_freespace_node_cachep, GFP_NOFS);
> > if (!node)
> > @@ -877,13 +878,31 @@ static struct ext4_frsp_node *ext4_mb_frsp_alloc_node(struct super_block *sb)
> > RB_CLEAR_NODE(&node->frsp_node);
> > RB_CLEAR_NODE(&node->frsp_len_node);
> >
> > + atomic_inc(&sbi->s_mb_num_fragments);
> > +
> > + if (sbi->s_mb_frsp_mem_limit &&
> > + atomic_read(&sbi->s_mb_num_fragments) >
> > + EXT4_FRSP_MEM_LIMIT_TO_NUM_NODES(sb))
> > + set_opt2(sb, FRSP_MEM_CRUNCH);
> > + else
> > + clear_opt2(sb, FRSP_MEM_CRUNCH);
> > +
> > +
>
> Why FRSP_MEM_CRUNCH is cleared here? Are any cases when a node allocating can reduce fragments numbers?
The reason why I have this here is to handle the case when the memory
limit gets updated by a sysfs tunable. So, once the sysfs tunable
tunable "mb_frsm_max_mem" was increased, then the next allocation
request would clear the MEM_CRUNCH flag. Without this, the FS would
evict a tree in LRU fashion (implemented in the next patch) which
shouldn't be necessary. Other option would be to update the MEM_CRUNCH
option right at the time when sysfs tunable is updated.
>
> > return node;
> > }
> >
> > static void ext4_mb_frsp_free_node(struct super_block *sb,
> > struct ext4_frsp_node *node)
> > {
> > + struct ext4_sb_info *sbi = EXT4_SB(sb);
> > +
> > kmem_cache_free(ext4_freespace_node_cachep, node);
> > + atomic_dec(&sbi->s_mb_num_fragments);
> > +
> > + if (!sbi->s_mb_frsp_mem_limit ||
> > + atomic_read(&sbi->s_mb_num_fragments) <
> > + EXT4_FRSP_MEM_LIMIT_TO_NUM_NODES(sb))
> > + clear_opt2(sb, FRSP_MEM_CRUNCH);
> > }
>
> If there are some reasons to clear FRSP_MEM_CRUNCH in ext4_mb_frsp_alloc_node, should we also set FRSP_MEM_CRUNCH here?
Thanks, I think we can do that. If the memory limit decreases, it'd be
good to set MEM_CRUNCH here. In my testing, I have found that
decreasing the memory limit to be very slow to take effect. Perhaps
this might be one of the reasons.
Thanks,
Harshad
>
> >
> > /* Evict a tree from memory */
> > @@ -1607,6 +1626,7 @@ int ext4_mb_init_freespace_trees(struct super_block *sb)
> > }
> > rwlock_init(&sbi->s_mb_frsp_lock);
> > atomic_set(&sbi->s_mb_num_frsp_trees_cached, 0);
> > + atomic_set(&sbi->s_mb_num_fragments, 0);
> >
> > return 0;
> > }
> > diff --git a/fs/ext4/mballoc.h b/fs/ext4/mballoc.h
> > index ac65f7eac611..08cac358324d 100644
> > --- a/fs/ext4/mballoc.h
> > +++ b/fs/ext4/mballoc.h
> > @@ -88,6 +88,10 @@ struct ext4_frsp_node {
> > struct rb_node frsp_node;
> > struct rb_node frsp_len_node;
> > };
> > +
> > +#define EXT4_FRSP_MEM_LIMIT_TO_NUM_NODES(__sb) \
> > + ((sbi->s_mb_frsp_mem_limit / sizeof(struct ext4_frsp_node)))
> > +
> > struct ext4_free_data {
> > /* this links the free block information from sb_info */
> > struct list_head efd_list;
> > diff --git a/fs/ext4/sysfs.c b/fs/ext4/sysfs.c
> > index 31e0db726d21..d23cb51635c3 100644
> > --- a/fs/ext4/sysfs.c
> > +++ b/fs/ext4/sysfs.c
> > @@ -8,6 +8,7 @@
> > *
> > */
> >
> > +#include "mballoc.h"
> > #include <linux/time.h>
> > #include <linux/fs.h>
> > #include <linux/seq_file.h>
> > @@ -24,6 +25,7 @@ typedef enum {
> > attr_session_write_kbytes,
> > attr_lifetime_write_kbytes,
> > attr_reserved_clusters,
> > + attr_frsp_tree_usage,
> > attr_inode_readahead,
> > attr_trigger_test_error,
> > attr_first_error_time,
> > @@ -205,6 +207,7 @@ EXT4_ATTR_FUNC(delayed_allocation_blocks, 0444);
> > EXT4_ATTR_FUNC(session_write_kbytes, 0444);
> > EXT4_ATTR_FUNC(lifetime_write_kbytes, 0444);
> > EXT4_ATTR_FUNC(reserved_clusters, 0644);
> > +EXT4_ATTR_FUNC(frsp_tree_usage, 0444);
> >
> > EXT4_ATTR_OFFSET(inode_readahead_blks, 0644, inode_readahead,
> > ext4_sb_info, s_inode_readahead_blks);
> > @@ -242,6 +245,7 @@ EXT4_ATTR(last_error_time, 0444, last_error_time);
> > EXT4_ATTR(journal_task, 0444, journal_task);
> > EXT4_RW_ATTR_SBI_UI(mb_prefetch, s_mb_prefetch);
> > EXT4_RW_ATTR_SBI_UI(mb_prefetch_limit, s_mb_prefetch_limit);
> > +EXT4_RW_ATTR_SBI_UI(mb_frsp_max_mem, s_mb_frsp_mem_limit);
> >
> > static unsigned int old_bump_val = 128;
> > EXT4_ATTR_PTR(max_writeback_mb_bump, 0444, pointer_ui, &old_bump_val);
> > @@ -251,6 +255,7 @@ static struct attribute *ext4_attrs[] = {
> > ATTR_LIST(session_write_kbytes),
> > ATTR_LIST(lifetime_write_kbytes),
> > ATTR_LIST(reserved_clusters),
> > + ATTR_LIST(frsp_tree_usage),
> > ATTR_LIST(inode_readahead_blks),
> > ATTR_LIST(inode_goal),
> > ATTR_LIST(mb_stats),
> > @@ -287,6 +292,7 @@ static struct attribute *ext4_attrs[] = {
> > #endif
> > ATTR_LIST(mb_prefetch),
> > ATTR_LIST(mb_prefetch_limit),
> > + ATTR_LIST(mb_frsp_max_mem),
> > NULL,
> > };
> > ATTRIBUTE_GROUPS(ext4);
> > @@ -369,6 +375,11 @@ static ssize_t ext4_attr_show(struct kobject *kobj,
> > return snprintf(buf, PAGE_SIZE, "%llu\n",
> > (unsigned long long)
> > atomic64_read(&sbi->s_resv_clusters));
> > + case attr_frsp_tree_usage:
> > + return snprintf(buf, PAGE_SIZE, "%llu\n",
> > + (unsigned long long)
> > + atomic_read(&sbi->s_mb_num_fragments) *
> > + sizeof(struct ext4_frsp_node));
> > case attr_inode_readahead:
> > case attr_pointer_ui:
> > if (!ptr)
> > --
> > 2.28.0.220.ged08abb693-goog
> >
>
> Best regards,
> Artem Blagodarenko.
>
>
