Status: Many chapters missed and not posted here yet. Formatting may contain errors and missed entries.
UPD-01-03-10_14-32:
+ Title changed.
+ Metadata circular buffer description (draft) added.
+ LVM extent and stripe difference (in russian yet)
UPD:[2010.02.10]: небольшие улучшения в форматировании.
FIXME: длина строка >= 80.
FIXME: отступы.
FIXME: <> как обозначения, а не как html-тэги.
FIXME: расстояние между словами (см. вывод `od`).
FIXME: и шрифт дурацкий: нули почему-то меньше остальных букв :D
Text below may look much better in vim with folding enabled ('fdm=marker', 'fmr={{{,}}}'). Though indents are probably remain incorrect. Perhaps, someday i'll fix this :-)
Draft of metadata sectors layout:
1. LVM label sector (0-r1) {{{
Location: {{{
- By default, `pvcreate` places the physical volume label
in the sector 1 (2nd 512-byte block).
- This label can optionally be placed in any of the first
four sectors ('--labelsector' option). And due to LVM
tools scan this first four sector for PV label, zeroing
of them ('-Zy' option) is recommended.
}}}
Format: {{{
Let consider on example:
> `pvcreate -vvv -Zy -M2 --metadatacopies=[012] --uuid=.. /dev/sdb4` {{{
>
> /dev/sdb4: size is 58589055 sectors
> with mcopies=0: 58588927 available sectors
> with mcopies=1: 58588671 available sectors
> metadata area at sector 8 size 376 sectors
> with mcopies=2: 58588416 available sectors
> metadata area at sector 8 size 376 sectors
> metadata area at sector 58588800 size 255 sectors
>
> Area sizes (available sectors) in hex:
> data area size =
> (with mcopies=0) = 58588927 sectors = 0x37dfeff sectors = 0x06fbfdfe00 bytes
> (with mcopies=1) = 58588671 sectors = 0x37dfdff sectors = 0x06fbfbfe00 bytes
> (with mcopies=2) = 58588416 sectors = 0x37dfd00 sectors = 0x06fbfa0000 bytes
> meta area size =
> (1st meta area) = 376 sectors = 192512 bytes = 0x02f000 bytes
> (2nd meta area) = 255 sectors = 130560 bytes = 0x01fe00 bytes
>
> Area offsets in hex:
> data area offset =
> (with mcopies=0) = 128 sectors = 65536 bytes = 0x010000 bytes
> (with mcopies=1,2) = 384 sectors = 196608 bytes = 0x030000 bytes
> meta area offset =
> (1st meta area) = 8 sectors = 4096 bytes = 0x1000 bytes
> (2nd meta area) = 58588800 sectors = 0x37dfe80 sectors = 0x06fbfd0000 bytes
>
> PV UUID = 'pesv0I-D0Ok-cVts-73Pg-vIaN-IRz2-LSldOn'
>
> Sector 1 dump: {{{
>
> Below for each 16 bytes row 1st line is for mcopies=0, 2nd - mcopies=1, 3rd -
> mcopies=2.
>
> 000200 4c 41 42 45 4c 4f 4e 45 01 00 00 00 00 00 00 00
> L A B E L O N E soh nul nul nul nul nul nul nul
> 000200 4c 41 42 45 4c 4f 4e 45 01 00 00 00 00 00 00 00
> L A B E L O N E soh nul nul nul nul nul nul nul
> 000200 4c 41 42 45 4c 4f 4e 45 01 00 00 00 00 00 00 00
> L A B E L O N E soh nul nul nul nul nul nul nul
> --
> 000210 e3 bb 4a cb 20 00 00 00 4c 56 4d 32 20 30 30 31
> c ; J K sp nul nul nul L V M 2 sp 0 0 1
> 000210 3c a9 89 2c 20 00 00 00 4c 56 4d 32 20 30 30 31
> < ) ht , sp nul nul nul L V M 2 sp 0 0 1
> 000210 76 4b 22 d4 20 00 00 00 4c 56 4d 32 20 30 30 31
> v K " T sp nul nul nul L V M 2 sp 0 0 1
> --
> 000220 70 65 73 76 30 49 44 30 4f 6b 63 56 74 73 37 33
> p e s v 0 I D 0 O k c V t s 7 3
> 000220 70 65 73 76 30 49 44 30 4f 6b 63 56 74 73 37 33
> p e s v 0 I D 0 O k c V t s 7 3
> 000220 70 65 73 76 30 49 44 30 4f 6b 63 56 74 73 37 33
> p e s v 0 I D 0 O k c V t s 7 3
> --
> 000230 50 67 76 49 61 4e 49 52 7a 32 4c 53 6c 64 4f 6e
> P g v I a N I R z 2 L S l d O n
> 000230 50 67 76 49 61 4e 49 52 7a 32 4c 53 6c 64 4f 6e
> P g v I a N I R z 2 L S l d O n
> 000230 50 67 76 49 61 4e 49 52 7a 32 4c 53 6c 64 4f 6e
> P g v I a N I R z 2 L S l d O n
> --
> 000240 00 fe fd fb 06 00 00 00 00 00 01 00 00 00 00 00
> nul ~ } { ack nul nul nul nul nul soh nul nul nul nul nul
> 000240 00 fe fb fb 06 00 00 00 00 00 03 00 00 00 00 00
> nul ~ { { ack nul nul nul nul nul etx nul nul nul nul nul
> 000240 00 00 fa fb 06 00 00 00 00 00 03 00 00 00 00 00
> nul nul z { ack nul nul nul nul nul etx nul nul nul nul nul
> --
> 000250 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000250 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000250 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> --
> 000260 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000260 00 00 00 00 00 00 00 00 00 10 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul dle nul nul nul nul nul nul
> +0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +a +b +c +d +e +f
> 000260 00 00 00 00 00 00 00 00 00 10 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul dle nul nul nul nul nul nul
> --
> 000270 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000270 00 f0 02 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul p stx nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000270 00 f0 02 00 00 00 00 00 00 00 fd fb 06 00 00 00
> nul p stx nul nul nul nul nul nul nul } { ack nul nul nul
> --
> 000280 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000280 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> +0 +1 +2 +3 +4 +5 +6 +7 +8 +9 +a +b +c +d +e +f
> 000280 00 fe 01 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul ~ soh nul nul nul nul nul nul nul nul nul nul nul nul nul
> --
> 000290 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000290 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> 000290 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul nul
> --
> (last NULL line repeated up to sector end)
>
> }}}
>
> }}}
Notes: {{{
- Size and offset below is in bytes.
- All ranges below includes their boundaries.
- Numbers (like size and offset) written on disk with least
significant byte first. I.e least significant byte of
number on disk will has less offset from disk beginning,
than most significant.
- All areas location rounded by sector boundary (512 bytes
block), hence all size and offset values will be
divisible by 512 (0x200), hence in hex them will always
has NULL least significant byte (0x00).
- I don't know for sure, is this NULL least significant
byte actually written on disk or not, but rather is, than
not. The most confusing place with this byte is beginning
of data area size after UUID: NULL-byte at 0x240+0 may be
NULL-termination byte after UUID as well as least
significant byte of data area size value. I assume, that
this is NULL least significant byte (see below).
}}}
Assumption: {{{
- NULL least significant byte written on disk as is.
- size value always occupy +0 -> +7 range of row.
- offset value always occupy +8 -> +f range of row.
- so, size and offset can be up to 16-digit hex number and
occupy exactly one row.
}}}
Some sort of prove for assumptions: {{{
- Throwing out part of number (NULL least significant byte)
and not writing it to disk will not save space allocated
for metadata, but may cause problems in future, when this
least significant byte may become not NULL. So i don't
see any sane reason for this.
- Not writing NULL-termination byte after UUID looks
possible, because UUID size is fixed (i think).
- If offset and size regions on disk occupy different
number of bytes, we either can not provide a method to
address all allocated size (through offset) or will have
offset values, which never be used - both is only
introduces future incompatibilities. Hence, offset and
size maximum allowed value should be the same and them
should occupy the same number of bytes on disk.
- Splitting one 16 bytes row on two equal parts by 8 bytes
for size (fisrt) and offset (second) seems to be very
reasonable: we have size and offset disk regions occupy
the same number of bytes and also, as example shows,
entire values (including NULL least significant byte)
written into this regions.
- Eventually, i do not find any conflict with any of my
assumptions :-)
}}}
Short Physical Volume label sector format: {{{
0x200 <LVM_magic_string>
0x210 <smth_unknown_and_LVM_version>
0x220 <PV_UUID>
0x230 <PV_UUID(continue)>
0x240 <data_size(8b)> <data_offset(8b)>
0x250 <NULL>
0x260 <NULL(8b)> <1st_meta_offset(8b)>
0x270 <1st_meta_size(8b)> <2nd_meta_offset(8b)>
0x280 <2nd_meta_size(8b)> <NULL(8b)>
0x290 <NULL_(up_to_0x400)>
}}}
Detailed Physical Volume label sector format: {{{
0x200+0 -> 0x200+f: LVM magic string. Identical for all
three PVs.
0x210+0 -> 0x210+7: Unknown.
0x210+8 -> 0x210+b: LVM version (i suppose).
0x210+c -> 0x210+c: Simply a separator (i suppose).
0x210+d -> 0x210+f: LVM PV label sector (this sector)
format version (i suppose).
0x220+0 -> 0x230+f: PV UUID.
0x240+0 -> 0x240+7: Data area size (number of available
bytes). Always present.
0x240+8 -> 0x240+f: Data area offset. Always present.
0x250+0 -> 0x250+f: NULL. Why?
0x260+0 -> 0x260+7: NULL. Why?
0x260+8 -> 0x260+f: 1st metadata circular buffer offset.
Only for mcopies=1,2.
0x270+0 -> 0x270+7: 1st metadata circular buffer size.
Only for mcopies=1,2.
0x270+8 -> 0x270+f: 2nd metadata circular buffer offset.
Only for mcopies=2.
0x280+0 -> 0x280+7: 2nd metadata circular buffer size
Only for mcopies=2.
0x280+8 -> 0x3f0+f: NULL and seems not used.
}}}
Last notes: {{{
- metadata area size and location (sector number, where
metadata circular buffer begins) can be obtained from
`pvcreate -vvv` output. If no metadata areas (circular
buffer) selected during PV creation (`pvcreate
--metadatacopies=0`), than metadata buffer area size will
be set to 120 sectors (it seems, that this is the lowest
size), though will not be used (no pointers to metadata
buffer will be in label sector and sector 8 remains
unchanged as well). This can be determined by
substraction count of available sectors from entire count
of sectors on disk (you get 128 = 120 + 8).
}}}
}}}
}}}
2. LVM circular buffer (0-r1) {{{
Short sector 8 format: {{{
0x1000 <Unknown>
0x1010 <Unknown(8b)> <1st_meta_offset(8b)>
0x1020 <1st_meta_size(8b)> <latest_entry_offset(8b)>
0x1030 <latest_entry_size(8b)> <Unknown(8b)>
0x1040 <NULLs_(up_to_0x1200)>
}}}
Detailed sector 8 format: {{{
0x1000+0 -> 0x1000+f: Unknown (2).
0x1010+0 -> 0x1010+7: Unknown (2).
0x1010+8 -> 0x1010+f: 1st metadata circular buffer (this
buffer) offset.
0x1020+0 -> 0x1020+7: 1st metadata circular buffer (this
buffer) size.
0x1020+8 -> 0x1020+f: Latest metadata entry in 1st circular
buffer (this buffer) offset. Offset
from beginning of the buffer, but
NOT from beginning of the PV.
0x1030+0 -> 0x1030+7: Latest metadata entry in 1st
circular buffer (this buffer) size,
including null-terminator.
0x1030+8 -> 0x1030+f: Unknown (1).
0x1040+0 -> 0x11f0+f: NULLs and seems not used.
(1): This number rather is not:
- first unallocated PE (checked by value on example).
- PE size (checked by value on example).
- PV size (checked by value on example).
- 2nd metadata buffer offset (it presents even in PVs
with single metadata buffer).
- latest metadata entry timestamp or smth else related
to latest metadata entry (it differs for different
PVs in the same VG, but latest metadata entry are
the same for all PVs of the same VG).
Also:
- this value does not divisible without remainder by
1024 or 512.
(2): Notes:
- range 0x1000+4 -> 0x1010+7 seems to be the same for
all PVs (even from different VGs).
}}}
Metadata entries location: {{{
- lvm metadata entry on disk location aligned roughly by
sector boundary (512 bytes block). I.e words such
vg_mp3 {
id = "YWvzHx-M1X5-TWtl-vCD1-w2zn-y0da-qui6PK"
seqno = 50
status = ["RESIZEABLE", "READ", "WRITE"]
will be placed only at sector's boundary (beginning).
- lvm metadata entry can occupy several sectors, though, if
last occupied sector not fully filled, all trailing
sector's part will not be cleared and, hence, it can
contain some garbage (exactly, some part of data from
previous record, occuping this sector).
}}}
Metadata entries format: {{{
- Each metadata entry on disk ends with null-terminator.
- On disk metadata timestamp (information about how and
when metadata entry was created) written after VG
description (information about volume group structure,
see below) to which it relates (in contrast with metadata
backup file, produced by `vgcfgbackup`, where timestamps
written first).
- On disk in metadata timestamp 'description' field is
empty (but in metadata backup file, produced by
`vgcfgbackup`, is not). (why?)
}}}
Last notes: {{{
- When PV contain no metadata circular buffer areas
('--metadatacopies=0' by `pvcreate`), than restoring VG
does not change anything in the PV metadata.
- To obtain offset from beginning of PV to latest metadata
entry in circular buffer, sum up '0x1010+8 -> 0x1010+f'
value with '0x1020+8 -> 0x1020+f' value.
- In order to locate latest metadata entry in raw 'on disk'
metadata copy, you should look up metadata circular
buffer (mostly, starting from sector 9, but if not, exact
offsets you can obtained from PV label sector) dump
splitted be sectors (512 bytes block) for sectors,
beginning like
vg_mp3 {
id = "YWvzHx-M1X5-TWtl-vCD1-w2zn-y0da-qui6PK"
seqno = 50
status = ["RESIZEABLE", "READ", "WRITE"]
This is the beginning of correct metadata entry.
Afterwards, you should choose one with latest 'seqno'
field. As explained above, simply search by word 'seqno'
may match with some garbage data after end of correct
metadata entry. Though, because we look up for entry with
latest 'seqno', anyway we'll select correct one. Also,
note, that 'seqno' word may appear as garbage only in few
first bytes of sector.
}}}
}}}
LVM extent and stripe:(копия письма :-)
<...>
А под stripe-ами я имел в виду логический том LVM с stripe mapping, те,
например, вот такая команда
`lvcreate -vvv -i3 -I32 -l32000 -n striped_lv_3x32k test_vg`
И непонятно мне было почему LVM использует два вида блоков - extent и stripe,
и почему нельзя было реализовать все виды отображения (mapping) - и линейное,
и stripe mapping, - используя только один вид блоков. Но я тут еще посмотрел
логи `lvcreate -vvv` и в части, относящейся к активации логического раздела,
мне кажется, я нашел ответ. Хотя не уверен, что полностью правильный :-) Вот,
например:
--- Volume group ---
VG Name vg_4k
PE Size 4.00 KB
VG UUID M7TWGx-Rggp-Okru-nWGY-c7Mh-MaLi-LhYxFm
теперь если создать логический том (размером 32000 extent-ов) с stripe
mapping, разделенный на 3 физических тома и с размером stripe-а в 4КБайта
получится:
`lvcreate -vvv -i3 -I4 -l32000 -n s_lv_3x16k vg_4k`
<..>
Creating vg_4k-s_lv_3x16k
dm create vg_4k-s_lv_3x16k
LVM-M7TWGxRggpOkrunWGYc7MhMaLiLhYxFm0YvE0mFyjZS7Q103PbM4efKUb2lUXnR8
NF [16384]
Loading vg_4k-s_lv_3x16k table
Adding target: 0 256008 striped 3 8 7:1 384 7:2 384 7:0 384
dm table (253:0) OF [16384]
dm reload (253:0) NF [16384]
Resuming vg_4k-s_lv_3x16k (253:0)
dm resume (253:0) NF [16384]
<..>
Хотя это, конечно, и без лога известно было, но все же. Т.е вся
функциональность LVM реализована через device-mapper, но device-mapper ничего
не знает об extent-ах и не использует их. В своих таблицах он (`dmsetup
table`) для всех размеров и смещений использует дисковые блоки (512байт). А
для striped таблиц он также использует stripe - блок данных, который будет
записан на одно физическое устройство (т.е стандартное определение stripe-а).
Т.е получается, что device-mapper как раз и использует всего один тип блоков -
только stripe. И получается, что extent - блок, используемый только для
удобства управления LVM томами, а при работе LVM (I/O) он не используется. Т.е
блок, используемый только LVM тулсетом. Тогда становится понятно вот это
замечание из описания опции '-s' в `man vgcreate`:
-s, --physicalextentsize PhysicalExtentSize[kKmMgGtT]
<..>
If the volume group metadata uses lvm2 format those restrictions
do not apply, but having a large number of extents will slow
down the tools but have no impact on I/O performance to the log-
ical volume. The smallest PE is 1KB.
А ограничения на размеры stripe и extent, видимо, сделаны для того, чтобы все
они друг в друге помещались: 512байт - дисковый блок - степень двойки,
поэтому, наверно, stripe и extent тоже должны быть степенью двойки, чтобы
содержали целое число дисковых блоков. Кроме того, т.к и stripe, и extent
степень двойки, extent всегда будет содержать целое количество stripe-ов.
Правда, не очень понятно, почему device-mapper не позволяет устанавливать
размер stripe-а меньше 4Кбайт.
<...>
