Set/get accessor cleanups
David Sterba
maintainercover Set/get accessor cleanups
This patchset is first in a series to cleanup and optimize the accessors, now removing code that's not needed anymore, explained with references in the last patch.
Overall effects on .ko:
text data bss dec hex filename
1463615 115665 16088 1595368 1857e8 pre/btrfs.ko 1456601 115665 16088 1588354 183c82 post/btrfs.ko
DELTA: -7014
And stack consumption:
__push_leaf_left -32 (176 -> 144) copy_for_split -32 (144 -> 112) fill_inode_item -32 (80 -> 48) btrfs_truncate_item -24 (152 -> 128) btrfs_extend_item -16 (104 -> 88) btrfs_del_items -16 (144 -> 128) setup_items_for_insert -32 (144 -> 112) __push_leaf_right -24 (168 -> 144)
REMOVED (744):
btrfs_get_token_32 88
btrfs_set_token_64 96
btrfs_get_token_64 88
btrfs_set_token_32 96
btrfs_get_token_16 88
btrfs_get_token_8 88
btrfs_set_token_16 96
btrfs_set_token_8 96 btrfs_init_map_token 8
REMOVED/NEW DELTA: -744 PRE/POST DELTA: -952
David Sterba (4):
btrfs: don't use token set/get accessors for btrfs_item members
btrfs: don't use token set/get accessors in inode.c:fill_inode_item()
btrfs: tree-log: don't use token set/get accessors in
fill_inode_item()
btrfs: accessors: delete token versions of set/get helpers
fs/btrfs/accessors.c | 78 -------------------------------------------- fs/btrfs/accessors.h | 37 --------------------- fs/btrfs/ctree.c | 51 ++++++++++------------------- fs/btrfs/inode.c | 50 ++++++++++++---------------- fs/btrfs/tree-log.c | 48 +++++++++++---------------- 5 files changed, 57 insertions(+), 207 deletions(-)
Signature · 1 line
2.49.0
David Sterba
maintainerPATCH 1/4 btrfs: don't use token set/get accessors for btrfs_item members
The token versions of set/get accessors will be removed, use the normal helpers. The btrfs_item members use that interface the most but there are no real benefits anymore.
This reduces stack consumption on x86_64 release config:
setup_items_for_insert -32 (144 -> 112) __push_leaf_left -32 (176 -> 144) btrfs_extend_item -16 (104 -> 88) copy_for_split -32 (144 -> 112) btrfs_del_items -16 (144 -> 128) btrfs_truncate_item -24 (152 -> 128) __push_leaf_right -24 (168 -> 144)
and module size:
text data bss dec hex filename
1463615 115665 16088 1595368 1857e8 pre/btrfs.ko 1463413 115665 16088 1595166 18571e post/btrfs.ko
DELTA: -202
--- fs/btrfs/ctree.c | 51 +++++++++++++++++------------------------------- 1 file changed, 18 insertions(+), 33 deletions(-)
fs/btrfs/ctree.c +18 -33
@@ -3109,7 +3109,6 @@ static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,struct btrfs_fs_info *fs_info = right->fs_info;struct extent_buffer *left = path->nodes[0 ];struct extent_buffer *upper = path->nodes[1 ]; -struct btrfs_map_token token;struct btrfs_disk_key disk_key;int slot;u32 i;
@@ -3183,13 +3182,12 @@ static noinline int __push_leaf_right(struct btrfs_trans_handle *trans,copy_leaf_items (right, left,0 , left_nritems - push_items, push_items);/* update the item pointers */ -btrfs_init_map_token (&token, right); right_nritems += push_items;btrfs_set_header_nritems (right, right_nritems); push_space =BTRFS_LEAF_DATA_SIZE (fs_info);for (i =0 ; i < right_nritems; i++) { - push_space -=btrfs_token_item_size (&token, i); -btrfs_set_token_item_offset (&token, i, push_space); + push_space -=btrfs_item_size (right, i); +btrfs_set_item_offset (right, i, push_space); } left_nritems -= push_items;
@@ -3332,7 +3330,6 @@ static noinline int __push_leaf_left(struct btrfs_trans_handle *trans,int ret =0 ;u32 this_item_size;u32 old_left_item_size; -struct btrfs_map_token token;if (empty) nr =min (right_nritems, max_slot);
@@ -3380,13 +3377,12 @@ static noinline int __push_leaf_left(struct btrfs_trans_handle *trans, old_left_nritems =btrfs_header_nritems (left);BUG_ON (old_left_nritems <=0 ); -btrfs_init_map_token (&token, left); old_left_item_size =btrfs_item_offset (left, old_left_nritems -1 );for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {u32 ioff; - ioff =btrfs_token_item_offset (&token, i); -btrfs_set_token_item_offset (&token, i, + ioff =btrfs_item_offset (left, i); +btrfs_set_item_offset (left, i, ioff - (BTRFS_LEAF_DATA_SIZE (fs_info) - old_left_item_size)); }btrfs_set_header_nritems (left, old_left_nritems + push_items);
@@ -3407,13 +3403,12 @@ static noinline int __push_leaf_left(struct btrfs_trans_handle *trans, btrfs_header_nritems(right) - push_items); } -btrfs_init_map_token (&token, right); right_nritems -= push_items;btrfs_set_header_nritems (right, right_nritems); push_space =BTRFS_LEAF_DATA_SIZE (fs_info);for (i =0 ; i < right_nritems; i++) { - push_space = push_space -btrfs_token_item_size (&token, i); -btrfs_set_token_item_offset (&token, i, push_space); + push_space = push_space -btrfs_item_size (right, i); +btrfs_set_item_offset (right, i, push_space); }btrfs_mark_buffer_dirty (trans, left);
@@ -3527,7 +3522,6 @@ static noinline int copy_for_split(struct btrfs_trans_handle *trans,int i;int ret;struct btrfs_disk_key disk_key; -struct btrfs_map_token token; nritems = nritems - mid;btrfs_set_header_nritems (right, nritems);
@@ -3540,12 +3534,11 @@ static noinline int copy_for_split(struct btrfs_trans_handle *trans, rt_data_off =BTRFS_LEAF_DATA_SIZE (fs_info) -btrfs_item_data_end (l, mid); -btrfs_init_map_token (&token, right);for (i =0 ; i < nritems; i++) {u32 ioff; - ioff =btrfs_token_item_offset (&token, i); -btrfs_set_token_item_offset (&token, i, ioff + rt_data_off); + ioff =btrfs_item_offset (right, i); +btrfs_set_item_offset (right, i, ioff + rt_data_off); }btrfs_set_header_nritems (l, mid);
@@ -4011,7 +4004,6 @@ void btrfs_truncate_item(struct btrfs_trans_handle *trans,unsigned int old_size;unsigned int size_diff;int i; -struct btrfs_map_token token; leaf = path->nodes[0 ]; slot = path->slots[0 ];
@@ -4034,12 +4026,11 @@ void btrfs_truncate_item(struct btrfs_trans_handle *trans,* item0..itemN ... dataN.offset..dataN.size .. data0.size */ /* first correct the data pointers */ -btrfs_init_map_token (&token, leaf);for (i = slot; i < nritems; i++) {u32 ioff; - ioff =btrfs_token_item_offset (&token, i); -btrfs_set_token_item_offset (&token, i, ioff + size_diff); + ioff =btrfs_item_offset (leaf, i); +btrfs_set_item_offset (leaf, i, ioff + size_diff); }/* shift the data */
@@ -4102,7 +4093,6 @@ void btrfs_extend_item(struct btrfs_trans_handle *trans,unsigned int old_data;unsigned int old_size;int i; -struct btrfs_map_token token; leaf = path->nodes[0 ];
@@ -4128,12 +4118,11 @@ void btrfs_extend_item(struct btrfs_trans_handle *trans,* item0..itemN ... dataN.offset..dataN.size .. data0.size */ /* first correct the data pointers */ -btrfs_init_map_token (&token, leaf);for (i = slot; i < nritems; i++) {u32 ioff; - ioff =btrfs_token_item_offset (&token, i); -btrfs_set_token_item_offset (&token, i, ioff - data_size); + ioff =btrfs_item_offset (leaf, i); +btrfs_set_item_offset (leaf, i, ioff - data_size); }/* shift the data */
@@ -4173,7 +4162,6 @@ static void setup_items_for_insert(struct btrfs_trans_handle *trans,struct btrfs_disk_key disk_key;struct extent_buffer *leaf;int slot; -struct btrfs_map_token token;u32 total_size;/*
@@ -4201,7 +4189,6 @@ static void setup_items_for_insert(struct btrfs_trans_handle *trans,BUG (); } -btrfs_init_map_token (&token, leaf);if (slot != nritems) {unsigned int old_data =btrfs_item_data_end (leaf, slot);
@@ -4219,8 +4206,8 @@ static void setup_items_for_insert(struct btrfs_trans_handle *trans,for (i = slot; i < nritems; i++) {u32 ioff; - ioff =btrfs_token_item_offset (&token, i); -btrfs_set_token_item_offset (&token, i, + ioff =btrfs_item_offset (leaf, i); +btrfs_set_item_offset (leaf, i, ioff - batch->total_data_size); }/* shift the items */
@@ -4237,8 +4224,8 @@ static void setup_items_for_insert(struct btrfs_trans_handle *trans,btrfs_cpu_key_to_disk (&disk_key, &batch->keys[i]);btrfs_set_item_key (leaf, &disk_key, slot + i); data_end -= batch->data_sizes[i]; -btrfs_set_token_item_offset (&token, slot + i, data_end); -btrfs_set_token_item_size (&token, slot + i, batch->data_sizes[i]); +btrfs_set_item_offset (leaf, slot + i, data_end); +btrfs_set_item_size (leaf, slot + i, batch->data_sizes[i]); }btrfs_set_header_nritems (leaf, nritems + batch->nr);
@@ -4478,7 +4465,6 @@ int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,if (slot + nr != nritems) {const u32 last_off =btrfs_item_offset (leaf, slot + nr -1 );const int data_end =leaf_data_end (leaf); -struct btrfs_map_token token;u32 dsize =0 ;int i;
@@ -4488,12 +4474,11 @@ int btrfs_del_items(struct btrfs_trans_handle *trans, struct btrfs_root *root,memmove_leaf_data (leaf, data_end + dsize, data_end, last_off - data_end); -btrfs_init_map_token (&token, leaf);for (i = slot + nr; i < nritems; i++) {u32 ioff; - ioff =btrfs_token_item_offset (&token, i); -btrfs_set_token_item_offset (&token, i, ioff + dsize); + ioff =btrfs_item_offset (leaf, i); +btrfs_set_item_offset (leaf, i, ioff + dsize); }memmove_leaf_items (leaf, slot, slot + nr, nritems - slot - nr);
Signature · 1 line
2.49.0
David Sterba
maintainerPATCH 2/4 btrfs: don't use token set/get accessors in inode.c:fill_inode_item()
In reply to David SterbaThe token versions of set/get accessors will be removed, use the normal helpers.
There's additional overhead of the token helpers that update the cached address in case it moves to another page/folio. The normal versions don't need to do that.
--- fs/btrfs/inode.c | 50 +++++++++++++++++++----------------------------- 1 file changed, 20 insertions(+), 30 deletions(-)
fs/btrfs/inode.c +20 -30
@@ -4078,45 +4078,35 @@ static void fill_inode_item(struct btrfs_trans_handle *trans,struct btrfs_inode_item *item, struct inode *inode) { - struct btrfs_map_token token; u64 flags; -btrfs_init_map_token (&token, leaf); +btrfs_set_inode_uid (leaf, item,i_uid_read (inode)); +btrfs_set_inode_gid (leaf, item,i_gid_read (inode)); +btrfs_set_inode_size (leaf, item,BTRFS_I (inode)->disk_i_size); +btrfs_set_inode_mode (leaf, item, inode->i_mode); +btrfs_set_inode_nlink (leaf, item, inode->i_nlink); -btrfs_set_token_inode_uid (&token, item,i_uid_read (inode)); -btrfs_set_token_inode_gid (&token, item,i_gid_read (inode)); -btrfs_set_token_inode_size (&token, item,BTRFS_I (inode)->disk_i_size); -btrfs_set_token_inode_mode (&token, item, inode->i_mode); -btrfs_set_token_inode_nlink (&token, item, inode->i_nlink); +btrfs_set_timespec_sec (leaf, &item->atime,inode_get_atime_sec (inode)); +btrfs_set_timespec_nsec (leaf, &item->atime,inode_get_atime_nsec (inode)); -btrfs_set_token_timespec_sec (&token, &item->atime, -inode_get_atime_sec (inode)); -btrfs_set_token_timespec_nsec (&token, &item->atime, -inode_get_atime_nsec (inode)); +btrfs_set_timespec_sec (leaf, &item->mtime,inode_get_mtime_sec (inode)); +btrfs_set_timespec_nsec (leaf, &item->mtime,inode_get_mtime_nsec (inode)); -btrfs_set_token_timespec_sec (&token, &item->mtime, -inode_get_mtime_sec (inode)); -btrfs_set_token_timespec_nsec (&token, &item->mtime, -inode_get_mtime_nsec (inode)); +btrfs_set_timespec_sec (leaf, &item->ctime,inode_get_ctime_sec (inode)); +btrfs_set_timespec_nsec (leaf, &item->ctime,inode_get_ctime_nsec (inode)); -btrfs_set_token_timespec_sec (&token, &item->ctime, -inode_get_ctime_sec (inode)); -btrfs_set_token_timespec_nsec (&token, &item->ctime, -inode_get_ctime_nsec (inode)); +btrfs_set_timespec_sec (leaf, &item->otime,BTRFS_I (inode)->i_otime_sec); +btrfs_set_timespec_nsec (leaf, &item->otime,BTRFS_I (inode)->i_otime_nsec); -btrfs_set_token_timespec_sec (&token, &item->otime,BTRFS_I (inode)->i_otime_sec); -btrfs_set_token_timespec_nsec (&token, &item->otime,BTRFS_I (inode)->i_otime_nsec); - -btrfs_set_token_inode_nbytes (&token, item,inode_get_bytes (inode)); -btrfs_set_token_inode_generation (&token, item, -BTRFS_I (inode)->generation); -btrfs_set_token_inode_sequence (&token, item,inode_peek_iversion (inode)); -btrfs_set_token_inode_transid (&token, item, trans->transid); -btrfs_set_token_inode_rdev (&token, item, inode->i_rdev); +btrfs_set_inode_nbytes (leaf, item,inode_get_bytes (inode)); +btrfs_set_inode_generation (leaf, item,BTRFS_I (inode)->generation); +btrfs_set_inode_sequence (leaf, item,inode_peek_iversion (inode)); +btrfs_set_inode_transid (leaf, item, trans->transid); +btrfs_set_inode_rdev (leaf, item, inode->i_rdev); flags =btrfs_inode_combine_flags (BTRFS_I (inode)->flags,BTRFS_I (inode)->ro_flags); -btrfs_set_token_inode_flags (&token, item, flags); -btrfs_set_token_inode_block_group (&token, item,0 ); +btrfs_set_inode_flags (leaf, item, flags); +btrfs_set_inode_block_group (leaf, item,0 ); }/*
Signature · 1 line
2.49.0
David Sterba
maintainerPATCH 4/4 btrfs: accessors: delete token versions of set/get helpers
In reply to David SterbaOnce upon a time there was a need to cache address of extent buffer pages, as it was a costly operation (map_private_extent_buffer(), cfed81a04eb555 ("Btrfs: add the ability to cache a pointer into the eb")). This was not even due to use of HIGHMEM, this had been removed before that due to possible locking issues ( a65917156e3459 ("Btrfs:
stop using highmem for extent_buffers")).
Over the time the amount of work in the set/get helpers got reduced and became quite straightforward bounds checking with an unaligned read/write, commit db3756c879773c ("btrfs: remove unused map_private_extent_buffer").
The actual caching of the page_address()/folio_address() in the token was more work for very little gain. This depended on subsequent access into the same page/folio, otherwise the cached pointer had to be updated.
For metadata-heavy operations this showed up in the 'perf top' profile where the btrfs_get_token_32 calls were at the top, on my testing machine consuming about 2-3%. The other generic 32/64 bit helpers also appeared in the profile with similar fraction.
After removing use of the token helpers we can remove them completely, this leads to reduction of btrfs.ko by 6.7KiB on release config.
text data bss dec hex filename
1463289 115665 16088 1595042 1856a2 pre/btrfs.ko 1456601 115665 16088 1588354 183c82 post/btrfs.ko
DELTA: -6688
--- fs/btrfs/accessors.c | 78 -------------------------------------------- fs/btrfs/accessors.h | 37 --------------------- 2 files changed, 115 deletions(-)
fs/btrfs/accessors.c +0 -78
@@ -25,13 +25,6 @@ static bool check_setget_bounds(const struct extent_buffer *eb,return true; } -void btrfs_init_map_token (struct btrfs_map_token *token,struct extent_buffer *eb) -{ - token->eb = eb; - token->kaddr =folio_address (eb->folios[0 ]); - token->offset =0 ; -} -/* * Macro templates that define helpers to read/write extent buffer data of a * given size, that are also used via ctree.h for access to item members by
@@ -41,11 +34,6 @@ void btrfs_init_map_token(struct btrfs_map_token *token, struct extent_buffer *e* - btrfs_set_8 (for 8/16/32/64) * - btrfs_get_8 (for 8/16/32/64) * -* Generic helpers with a token (cached address of the most recently accessed -* page): -* - btrfs_set_token_8 (for 8/16/32/64) -* - btrfs_get_token_8 (for 8/16/32/64) -* * The set/get functions handle data spanning two pages transparently, in case * metadata block size is larger than page. Every pointer to metadata items is * an offset into the extent buffer page array, cast to a specific type. This
@@ -57,37 +45,6 @@ void btrfs_init_map_token(struct btrfs_map_token *token, struct extent_buffer *e*/ #define DEFINE_BTRFS_SETGET_BITS (bits) \ -u##bits btrfs_get_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off) \ -{ \ - const unsigned long member_offset = (unsigned long)ptr + off; \ - const unsigned long idx = get_eb_folio_index(token->eb, member_offset); \ - const unsigned long oil = get_eb_offset_in_folio(token->eb, \ - member_offset);\ - const int unit_size = token->eb->folio_size; \ - const int unit_shift = token->eb->folio_shift; \ - const int size = sizeof(u##bits); \ - u8 lebytes[sizeof(u##bits)]; \ - const int part = unit_size - oil; \ - \ - ASSERT(token); \ - ASSERT(token->kaddr); \ - ASSERT(check_setget_bounds(token->eb, ptr, off, size)); \ - if (token->offset <= member_offset && \ - member_offset + size <= token->offset + unit_size) { \ - return get_unaligned_le##bits(token->kaddr + oil); \ - } \ - token->kaddr = folio_address(token->eb->folios[idx]); \ - token->offset = idx << unit_shift; \ - if (INLINE_EXTENT_BUFFER_PAGES == 1 || oil + size <= unit_size) \ - return get_unaligned_le##bits(token->kaddr + oil); \ - \ - memcpy(lebytes, token->kaddr + oil, part); \ - token->kaddr = folio_address(token->eb->folios[idx + 1]); \ - token->offset = (idx + 1) << unit_shift; \ - memcpy(lebytes + part, token->kaddr, size - part); \ - return get_unaligned_le##bits(lebytes); \ -} \ u##bits btrfs_get_##bits(const struct extent_buffer *eb, \ const void *ptr, unsigned long off) \ { \
@@ -110,41 +67,6 @@ u##bits btrfs_get_##bits(const struct extent_buffer *eb, \memcpy (lebytes + part, kaddr, size - part); \ return get_unaligned_le##bits(lebytes); \ } \ -void btrfs_set_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off, \ - u##bits val) \ -{ \ - const unsigned long member_offset = (unsigned long)ptr + off; \ - const unsigned long idx = get_eb_folio_index(token->eb, member_offset); \ - const unsigned long oil = get_eb_offset_in_folio(token->eb, \ - member_offset);\ - const int unit_size = token->eb->folio_size; \ - const int unit_shift = token->eb->folio_shift; \ - const int size = sizeof(u##bits); \ - u8 lebytes[sizeof(u##bits)]; \ - const int part = unit_size - oil; \ - \ - ASSERT(token); \ - ASSERT(token->kaddr); \ - ASSERT(check_setget_bounds(token->eb, ptr, off, size)); \ - if (token->offset <= member_offset && \ - member_offset + size <= token->offset + unit_size) { \ - put_unaligned_le##bits(val, token->kaddr + oil); \ - return; \ - } \ - token->kaddr = folio_address(token->eb->folios[idx]); \ - token->offset = idx << unit_shift; \ - if (INLINE_EXTENT_BUFFER_PAGES == 1 || \ - oil + size <= unit_size) { \ - put_unaligned_le##bits(val, token->kaddr + oil); \ - return; \ - } \ - put_unaligned_le##bits(val, lebytes); \ - memcpy(token->kaddr + oil, lebytes, part); \ - token->kaddr = folio_address(token->eb->folios[idx + 1]); \ - token->offset = (idx + 1) << unit_shift; \ - memcpy(token->kaddr, lebytes + part, size - part); \ -} \ void btrfs_set_##bits(const struct extent_buffer *eb, void *ptr, \ unsigned long off, u##bits val) \ { \
fs/btrfs/accessors.h +0 -37
@@ -16,14 +16,6 @@struct extent_buffer ; -struct btrfs_map_token { -struct extent_buffer *eb; -char *kaddr; -unsigned long offset; -}; - -void btrfs_init_map_token (struct btrfs_map_token *token,struct extent_buffer *eb); -/* * Some macros to generate set/get functions for the struct fields. This * assumes there is a lefoo_to_cpu for every type, so lets make a simple one
@@ -56,11 +48,6 @@ static inline void put_unaligned_le8(u8 val, void *p) sizeof_field(type, member)))#define DECLARE_BTRFS_SETGET_BITS (bits) \ -u##bits btrfs_get_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off); \ -void btrfs_set_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off, \ - u##bits val); \ u##bits btrfs_get_##bits(const struct extent_buffer *eb, \ const void *ptr, unsigned long off); \ void btrfs_set_##bits(const struct extent_buffer *eb, void *ptr, \
@@ -83,18 +70,6 @@ static inline void btrfs_set_##name(const struct extent_buffer *eb, type *s, \ { \ static_assert(sizeof(u##bits) == sizeof_field(type, member)); \ btrfs_set_##bits(eb, s, offsetof(type, member), val); \ -} \ -static inline u##bits btrfs_token_##name(struct btrfs_map_token *token, \ - const type *s) \ -{ \ - static_assert(sizeof(u##bits) == sizeof_field(type, member)); \ - return btrfs_get_token_##bits(token, s, offsetof(type, member));\ -} \ -static inline void btrfs_set_token_##name(struct btrfs_map_token *token,\ - type *s, u##bits val) \ -{ \ - static_assert(sizeof(u##bits) == sizeof_field(type, member)); \ - btrfs_set_token_##bits(token, s, offsetof(type, member), val); \ }#define BTRFS_SETGET_HEADER_FUNCS (name, type, member, bits) \
@@ -479,18 +454,6 @@ static inline void btrfs_set_item_##member(const struct extent_buffer *eb, \ int slot, u32 val) \ { \ btrfs_set_raw_item_##member(eb, btrfs_item_nr(eb, slot), val); \ -} \ -static inline u32 btrfs_token_item_##member(struct btrfs_map_token *token, \ - int slot) \ -{ \ - struct btrfs_item *item = btrfs_item_nr(token->eb, slot); \ - return btrfs_token_raw_item_##member(token, item); \ -} \ -static inline void btrfs_set_token_item_##member(struct btrfs_map_token *token, \ - int slot, u32 val) \ -{ \ - struct btrfs_item *item = btrfs_item_nr(token->eb, slot); \ - btrfs_set_token_raw_item_##member(token, item, val); \ }BTRFS_ITEM_SETGET_FUNCS (offset)
Signature · 1 line
2.49.0
David Sterba
maintainerPATCH 3/4 btrfs: tree-log: don't use token set/get accessors in fill_inode_item()
In reply to David SterbaThe token versions of set/get accessors will be removed, use the normal helpers.
There's additional overhead of the token helpers that update the cached address in case it moves to another page/folio. The normal versions don't need to do that.
Note this is similar to fill_inode_item() in inode.c but with slight differences. The two functions could be deduplicated eventually.
--- fs/btrfs/tree-log.c | 48 ++++++++++++++++++--------------------------- 1 file changed, 19 insertions(+), 29 deletions(-)
fs/btrfs/tree-log.c +19 -29
@@ -4204,44 +4204,34 @@ static void fill_inode_item(struct btrfs_trans_handle *trans, struct inode *inode, int log_inode_only, u64 logged_isize) { - struct btrfs_map_token token; u64 flags; -btrfs_init_map_token (&token, leaf); -if (log_inode_only) {/* set the generation to zero so the recover code * can tell the difference between an logging * just to say 'this inode exists' and a logging * to say 'update this inode with these values' */ -btrfs_set_token_inode_generation (&token, item,0 ); -btrfs_set_token_inode_size (&token, item, logged_isize); +btrfs_set_inode_generation (leaf, item,0 ); +btrfs_set_inode_size (leaf, item, logged_isize); }else { -btrfs_set_token_inode_generation (&token, item, -BTRFS_I (inode)->generation); -btrfs_set_token_inode_size (&token, item, inode->i_size); +btrfs_set_inode_generation (leaf, item,BTRFS_I (inode)->generation); +btrfs_set_inode_size (leaf, item, inode->i_size); } -btrfs_set_token_inode_uid (&token, item,i_uid_read (inode)); -btrfs_set_token_inode_gid (&token, item,i_gid_read (inode)); -btrfs_set_token_inode_mode (&token, item, inode->i_mode); -btrfs_set_token_inode_nlink (&token, item, inode->i_nlink); +btrfs_set_inode_uid (leaf, item,i_uid_read (inode)); +btrfs_set_inode_gid (leaf, item,i_gid_read (inode)); +btrfs_set_inode_mode (leaf, item, inode->i_mode); +btrfs_set_inode_nlink (leaf, item, inode->i_nlink); -btrfs_set_token_timespec_sec (&token, &item->atime, -inode_get_atime_sec (inode)); -btrfs_set_token_timespec_nsec (&token, &item->atime, -inode_get_atime_nsec (inode)); +btrfs_set_timespec_sec (leaf, &item->atime,inode_get_atime_sec (inode)); +btrfs_set_timespec_nsec (leaf, &item->atime,inode_get_atime_nsec (inode)); -btrfs_set_token_timespec_sec (&token, &item->mtime, -inode_get_mtime_sec (inode)); -btrfs_set_token_timespec_nsec (&token, &item->mtime, -inode_get_mtime_nsec (inode)); +btrfs_set_timespec_sec (leaf, &item->mtime,inode_get_mtime_sec (inode)); +btrfs_set_timespec_nsec (leaf, &item->mtime,inode_get_mtime_nsec (inode)); -btrfs_set_token_timespec_sec (&token, &item->ctime, -inode_get_ctime_sec (inode)); -btrfs_set_token_timespec_nsec (&token, &item->ctime, -inode_get_ctime_nsec (inode)); +btrfs_set_timespec_sec (leaf, &item->ctime,inode_get_ctime_sec (inode)); +btrfs_set_timespec_nsec (leaf, &item->ctime,inode_get_ctime_nsec (inode));/* * We do not need to set the nbytes field, in fact during a fast fsync
@@ -4252,13 +4242,13 @@ static void fill_inode_item(struct btrfs_trans_handle *trans,* inode item in subvolume tree as needed (see overwrite_item()). */ -btrfs_set_token_inode_sequence (&token, item,inode_peek_iversion (inode)); -btrfs_set_token_inode_transid (&token, item, trans->transid); -btrfs_set_token_inode_rdev (&token, item, inode->i_rdev); +btrfs_set_inode_sequence (leaf, item,inode_peek_iversion (inode)); +btrfs_set_inode_transid (leaf, item, trans->transid); +btrfs_set_inode_rdev (leaf, item, inode->i_rdev); flags =btrfs_inode_combine_flags (BTRFS_I (inode)->flags,BTRFS_I (inode)->ro_flags); -btrfs_set_token_inode_flags (&token, item, flags); -btrfs_set_token_inode_block_group (&token, item,0 ); +btrfs_set_inode_flags (leaf, item, flags); +btrfs_set_inode_block_group (leaf, item,0 ); } static int log_inode_item(struct btrfs_trans_handle *trans,
Signature · 1 line
2.49.0
Re: [PATCH 4/4] btrfs: accessors: delete token versions of set/get helpers
In reply to David SterbaQuoted 11 lines · David Sterba
Once upon a time there was a need to cache address of extent buffer pages, as it was a costly operation (
map_private_extent_buffer(),cfed81a04eb555("Btrfs: add the ability to cache a pointer into the eb")). This was not even due to use of HIGHMEM, this had been removed before that due to possible locking issues (a65917156e3459("Btrfs:
stop using highmem forextent_buffers")).Over the time the amount of work in the set/get helpers got reduced and became quite straightforward bounds checking with an unaligned read/write, commit
db3756c879773c("btrfs: remove unusedmap_private_extent_buffer").
Thanks a lot for the history of the token accessors.
And this is also a great chance to us to sync the removal of token eb accessors to the progs.
If you haven't yet started that progs removal, I'm totally fine to do that.
Thanks,
Qu
Quoted 205 lines
The actual caching of the page_address()/folio_address() in the token was more work for very little gain. This depended on subsequent access into the same page/folio, otherwise the cached pointer had to be updated.
For metadata-heavy operations this showed up in the 'perf top' profile where the
btrfs_get_token_32calls were at the top, on my testing machine consuming about 2-3%. The other generic 32/64 bit helpers also appeared in the profile with similar fraction.After removing use of the token helpers we can remove them completely, this leads to reduction of btrfs.ko by 6.7KiB on release config.
text data bss dec hex filename
1463289 115665 16088 1595042 1856a2 pre/btrfs.ko 1456601 115665 16088 1588354 183c82 post/btrfs.koDELTA: -6688
Signed-off-by David Sterba <[email protected]>--- fs/btrfs/accessors.c | 78 -------------------------------------------- fs/btrfs/accessors.h | 37 --------------------- 2 files changed, 115 deletions(-)fs/btrfs/accessors.c +0
-78@@ -25,13 +25,6 @@ static bool check_setget_bounds(const struct extent_buffer *eb,return true; } -void btrfs_init_map_token (struct btrfs_map_token *token,struct extent_buffer *eb) -{ - token->eb = eb; - token->kaddr =folio_address (eb->folios[0 ]); - token->offset =0 ; -} -/* * Macro templates that define helpers to read/write extent buffer data of a * given size, that are also used via ctree.h for access to item members by @@ -41,11 +34,6 @@ void btrfs_init_map_token(struct btrfs_map_token *token, struct extent_buffer *e* - btrfs_set_8 (for 8/16/32/64) * - btrfs_get_8 (for 8/16/32/64) * -* Generic helpers with a token (cached address of the most recently accessed -* page): -* - btrfs_set_token_8 (for 8/16/32/64) -* - btrfs_get_token_8 (for 8/16/32/64) -* * The set/get functions handle data spanning two pages transparently, in case * metadata block size is larger than page. Every pointer to metadata items is * an offset into the extent buffer page array, cast to a specific type. This @@ -57,37 +45,6 @@ void btrfs_init_map_token(struct btrfs_map_token *token, struct extent_buffer *e*/ #define DEFINE_BTRFS_SETGET_BITS (bits) \ -u##bits btrfs_get_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off) \ -{ \ - const unsigned long member_offset = (unsigned long)ptr + off; \ - const unsigned long idx = get_eb_folio_index(token->eb, member_offset); \ - const unsigned long oil = get_eb_offset_in_folio(token->eb, \ - member_offset);\ - const int unit_size = token->eb->folio_size; \ - const int unit_shift = token->eb->folio_shift; \ - const int size = sizeof(u##bits); \ - u8 lebytes[sizeof(u##bits)]; \ - const int part = unit_size - oil; \ - \ - ASSERT(token); \ - ASSERT(token->kaddr); \ - ASSERT(check_setget_bounds(token->eb, ptr, off, size)); \ - if (token->offset <= member_offset && \ - member_offset + size <= token->offset + unit_size) { \ - return get_unaligned_le##bits(token->kaddr + oil); \ - } \ - token->kaddr = folio_address(token->eb->folios[idx]); \ - token->offset = idx << unit_shift; \ - if (INLINE_EXTENT_BUFFER_PAGES == 1 || oil + size <= unit_size) \ - return get_unaligned_le##bits(token->kaddr + oil); \ - \ - memcpy(lebytes, token->kaddr + oil, part); \ - token->kaddr = folio_address(token->eb->folios[idx + 1]); \ - token->offset = (idx + 1) << unit_shift; \ - memcpy(lebytes + part, token->kaddr, size - part); \ - return get_unaligned_le##bits(lebytes); \ -} \ u##bits btrfs_get_##bits(const struct extent_buffer *eb, \ const void *ptr, unsigned long off) \ { \@@ -110,41 +67,6 @@ u##bits btrfs_get_##bits(const struct extent_buffer *eb, \memcpy (lebytes + part, kaddr, size - part); \ return get_unaligned_le##bits(lebytes); \ } \ -void btrfs_set_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off, \ - u##bits val) \ -{ \ - const unsigned long member_offset = (unsigned long)ptr + off; \ - const unsigned long idx = get_eb_folio_index(token->eb, member_offset); \ - const unsigned long oil = get_eb_offset_in_folio(token->eb, \ - member_offset);\ - const int unit_size = token->eb->folio_size; \ - const int unit_shift = token->eb->folio_shift; \ - const int size = sizeof(u##bits); \ - u8 lebytes[sizeof(u##bits)]; \ - const int part = unit_size - oil; \ - \ - ASSERT(token); \ - ASSERT(token->kaddr); \ - ASSERT(check_setget_bounds(token->eb, ptr, off, size)); \ - if (token->offset <= member_offset && \ - member_offset + size <= token->offset + unit_size) { \ - put_unaligned_le##bits(val, token->kaddr + oil); \ - return; \ - } \ - token->kaddr = folio_address(token->eb->folios[idx]); \ - token->offset = idx << unit_shift; \ - if (INLINE_EXTENT_BUFFER_PAGES == 1 || \ - oil + size <= unit_size) { \ - put_unaligned_le##bits(val, token->kaddr + oil); \ - return; \ - } \ - put_unaligned_le##bits(val, lebytes); \ - memcpy(token->kaddr + oil, lebytes, part); \ - token->kaddr = folio_address(token->eb->folios[idx + 1]); \ - token->offset = (idx + 1) << unit_shift; \ - memcpy(token->kaddr, lebytes + part, size - part); \ -} \ void btrfs_set_##bits(const struct extent_buffer *eb, void *ptr, \ unsigned long off, u##bits val) \ { \fs/btrfs/accessors.h +0
-37@@ -16,14 +16,6 @@struct extent_buffer ; -struct btrfs_map_token { -struct extent_buffer *eb; -char *kaddr; -unsigned long offset; -}; - -void btrfs_init_map_token (struct btrfs_map_token *token,struct extent_buffer *eb); -/* * Some macros to generate set/get functions for the struct fields. This * assumes there is a lefoo_to_cpu for every type, so lets make a simple one @@ -56,11 +48,6 @@ static inline void put_unaligned_le8(u8 val, void *p) sizeof_field(type, member)))#define DECLARE_BTRFS_SETGET_BITS (bits) \ -u##bits btrfs_get_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off); \ -void btrfs_set_token_##bits(struct btrfs_map_token *token, \ - const void *ptr, unsigned long off, \ - u##bits val); \ u##bits btrfs_get_##bits(const struct extent_buffer *eb, \ const void *ptr, unsigned long off); \ void btrfs_set_##bits(const struct extent_buffer *eb, void *ptr, \@@ -83,18 +70,6 @@ static inline void btrfs_set_##name(const struct extent_buffer *eb, type *s, \ { \ static_assert(sizeof(u##bits) == sizeof_field(type, member)); \ btrfs_set_##bits(eb, s, offsetof(type, member), val); \ -} \ -static inline u##bits btrfs_token_##name(struct btrfs_map_token *token, \ - const type *s) \ -{ \ - static_assert(sizeof(u##bits) == sizeof_field(type, member)); \ - return btrfs_get_token_##bits(token, s, offsetof(type, member));\ -} \ -static inline void btrfs_set_token_##name(struct btrfs_map_token *token,\ - type *s, u##bits val) \ -{ \ - static_assert(sizeof(u##bits) == sizeof_field(type, member)); \ - btrfs_set_token_##bits(token, s, offsetof(type, member), val); \ }#define BTRFS_SETGET_HEADER_FUNCS (name, type, member, bits) \@@ -479,18 +454,6 @@ static inline void btrfs_set_item_##member(const struct extent_buffer *eb, \ int slot, u32 val) \ { \ btrfs_set_raw_item_##member(eb, btrfs_item_nr(eb, slot), val); \ -} \ -static inline u32 btrfs_token_item_##member(struct btrfs_map_token *token, \ - int slot) \ -{ \ - struct btrfs_item *item = btrfs_item_nr(token->eb, slot); \ - return btrfs_token_raw_item_##member(token, item); \ -} \ -static inline void btrfs_set_token_item_##member(struct btrfs_map_token *token, \ - int slot, u32 val) \ -{ \ - struct btrfs_item *item = btrfs_item_nr(token->eb, slot); \ - btrfs_set_token_raw_item_##member(token, item, val); \ }BTRFS_ITEM_SETGET_FUNCS (offset)
Quoted 37 lines · David Sterba
This patchset is first in a series to cleanup and optimize the accessors, now removing code that's not needed anymore, explained with references in the last patch.
Overall effects on .ko:
text data bss dec hex filename
1463615 115665 16088 1595368 1857e8 pre/btrfs.ko 1456601 115665 16088 1588354 183c82 post/btrfs.koDELTA: -7014
And stack consumption:
__push_leaf_left -32 (176 -> 144) copy_for_split -32 (144 -> 112) fill_inode_item -32 (80 -> 48) btrfs_truncate_item -24 (152 -> 128) btrfs_extend_item -16 (104 -> 88) btrfs_del_items -16 (144 -> 128) setup_items_for_insert -32 (144 -> 112) __push_leaf_right -24 (168 -> 144)REMOVED (744):
btrfs_get_token_32 88 btrfs_set_token_64 96 btrfs_get_token_64 88 btrfs_set_token_32 96 btrfs_get_token_16 88 btrfs_get_token_8 88 btrfs_set_token_16 96 btrfs_set_token_8 96
btrfs_init_map_token8REMOVED/NEW DELTA: -744 PRE/POST DELTA: -952
Quoted 6 lines
David Sterba (4):
btrfs: don't use token set/get accessors forbtrfs_itemmembers
btrfs: don't use token set/get accessors in inode.c:fill_inode_item()
btrfs: tree-log: don't use token set/get accessors in
fill_inode_item()
btrfs: accessors: delete token versions of set/get helpers
Looks good to me.
Although it also exposed that we have two different fill_inode_item() functions.
I know the tree log code needs some special handling, the fill_inode_item() in inode.c looks like can be implemented by the one in tree-log.
It may be a good time to merge them into one in another patch.
Thanks,
Qu
Quoted 8 lines
fs/btrfs/accessors.c | 78 -------------------------------------------- fs/btrfs/accessors.h | 37 --------------------- fs/btrfs/ctree.c | 51 ++++++++++------------------- fs/btrfs/inode.c | 50 ++++++++++++---------------- fs/btrfs/tree-log.c | 48 +++++++++++---------------- 5 files changed, 57 insertions(+), 207 deletions(-)
Nice work,
Quoted 11 lines · Qu Wenruo
Quoted 6 lines
David Sterba (4):
btrfs: don't use token set/get accessors forbtrfs_itemmembers
btrfs: don't use token set/get accessors in inode.c:fill_inode_item()
btrfs: tree-log: don't use token set/get accessors in
fill_inode_item()
btrfs: accessors: delete token versions of set/get helpersLooks good to me.
Reviewed-by Qu Wenruo <[email protected]>
Quoted 7 lines
Although it also exposed that we have two different
fill_inode_item()functions.
I know the tree log code needs some special handling, thefill_inode_item()ininode.clooks like can be implemented by the one in tree-log.It may be a good time to merge them into one in another patch.
Yes, the difference is not that big so it makes sense to merge them.
Re: [PATCH 4/4] btrfs: accessors: delete token versions of set/get helpers
In reply to Qu WenruoQuoted 15 lines · Qu Wenruo
Quoted 11 lines · David Sterba
Once upon a time there was a need to cache address of extent buffer pages, as it was a costly operation (
map_private_extent_buffer(),cfed81a04eb555("Btrfs: add the ability to cache a pointer into the eb")). This was not even due to use of HIGHMEM, this had been removed before that due to possible locking issues (a65917156e3459("Btrfs:
stop using highmem forextent_buffers")).Over the time the amount of work in the set/get helpers got reduced and became quite straightforward bounds checking with an unaligned read/write, commit
db3756c879773c("btrfs: remove unusedmap_private_extent_buffer").
Quoted 6 lines
Thanks a lot for the history of the token accessors.
And this is also a great chance to us to sync the removal of token eb accessors to the progs.
If you haven't yet started that progs removal, I'm totally fine to do that.
I usually do the kernel->progs sync before a release, inspired by the kernel git log where it applies. I haven't started with the accessors updates as I have a few more changes to those files pending but I wouldn't mind if you do it now.