charsets(7), iso_8859-1(7), cp1252(7), koi8-r(7) and six more: Linux's map of everything before Unicode¶
Level: reference · for anyone who has typed man 7 cp1252 or man 7 charsets and wanted every row on it explained
One line: Ten Linux pages, none of them on the Mac, print the upper half of each single-byte code page as a table you can grep, and every row checked against iconv today holds: Latin-9 is Latin-1 with exactly eight bytes changed and € first among them, Windows-1252 assigns 27 of the 32 positions Latin-1 gives to controls and leaves five holes, and clearing bit 7 of a KOI8-R byte gives a Latin letter that spells the Russian one.
The pages: charsets(7) (Linux man-pages 6.7, 2024-01-28; copyright Eric S. Raymond 1996 and Andries Brouwer) · iso_8859-1(7) (2024-01-28; Daniel Quinlan 1993–1995) · iso_8859-2(7), iso_8859-15(7), iso_8859-16(7) (2024-01-28) · cp1251(7), cp1252(7) (2024-01-28; Marko Myllynen 2014) · koi8-r(7) (2022-12-15; Alexey Mahotkin 2001) · koi8-u(7) (2022-12-15) · armscii-8(7) (2022-12-15; Lefteris Dimitroulakis 2009). macOS ships none of them: man -w 7 charsets answers No manual entry. Dumped 2026-09-13 by dump_linux.sh; the machines are in raw/PROVENANCE-linux.txt and raw/PROVENANCE-macos.txt.
What the pages are for¶
charsets(7) says what it is in its first sentence: an overview on different character set standards and how they were used on Linux before Unicode became ubiquitous. It is a map, in section 7 because it documents no file, no function and no command, and it walks from ASCII through the fifteen parts of ISO 8859, KOI8, the East Asian sets, the ISO 2022 switching scheme, TIS-620 and finally Unicode, one paragraph each. Its copyright line is from 1996, when the map was of the present; the page still describes the sets as used by locale character sets, and glibc still ships a converter for every set named below.
The other nine are tables. Each prints the upper half of one single-byte code page as one row per byte with the byte in three bases, the glyph and the Unicode character name, and each says of itself that it shows the characters ... that are printable and unlisted in the ascii(7) manual page, which is the whole idea of a code page in one clause: the low half is ASCII and the page is about the disagreement above it, Code pages. ascii(7)'s SEE ALSO there lists a page for every part of ISO 8859 that exists, and beside them are two Windows code pages, two KOI8 variants and ArmSCII; this family takes the ten that the library's cast of characters lands on.
macOS has none of these pages, though it has the tables: iconv -l on the Mac lists every name used below, locale -a lists fifty locales whose encoding is one of Latin-2, Latin-9, KOI8-R, KOI8-U, CP1251 or ArmSCII-8, and every experiment below that could run on both machines gave the same bytes on both but one. The pages are Linux's; the facts are everybody's.
The page, with notes¶
charsets(7): the map¶
ISO/IEC 8859 is a series of 15 8-bit character sets, all of which
have ASCII in their low (7-bit) half, invisible control characters in
positions 128 to 159, and 96 fixed-width graphics in positions
160–255.
...
ISO/IEC 8859-12
This character set does not exist.
Three numbers describe the whole family: 128 ASCII, 32 controls, 96 graphics. The 32 are the C1 controls, 0x80–0x9F, which is why the ISO tables below begin at 0xA0 and have 96 rows, and why Windows, which had no use for C1 controls, put printable characters there instead. Part 12 does not exist is the page being precise about a standard that skipped a number. The Unicode section at the end says the codespace permits 20.1 bits, which is log₂ of 1,114,112, and still prints the six-byte UTF-8 progression that utf8(5) and utf-8(7) dates to 1998.
KOI8-R is a non-ISO character set popular in Russia before Unicode.
The lower half is ASCII; the upper is a Cyrillic character set some‐
what better designed than ISO/IEC 8859-5.
...
Switching between character sets is done using the shift functions ^N
(SO or LS1), ^O (SI or LS0), ESC n (LS2), ESC o (LS3), ESC N (SS2),
ESC O (SS3), ESC ~ (LS1R), ESC } (LS2R), ESC | (LS3R).
Somewhat better designed is explained by the measurement below, not by the page. The ISO 2022 paragraph is the other model of a multi-script world, the one the code pages replaced and UTF-8 replaced in turn: escape sequences that switch which table the following bytes belong to, with four slots G0–G3 and the same single shift idea that euc(5) uses for its third and fourth code sets. A reader who has met ESC ( B in a terminal log has met this paragraph.
The table, and how to read a row¶
The following table displays the characters in ISO/IEC 8859-1 that
are printable and unlisted in the ascii(7) manual page.
Oct Dec Hex Char Description
────────────────────────────────────────────────────────────────────
240 160 A0 NO-BREAK SPACE
...
351 233 E9 é LATIN SMALL LETTER E WITH ACUTE
...
377 255 FF ÿ LATIN SMALL LETTER Y WITH DIAERESIS
One byte per row, written three ways, then the glyph, then a name. The three bases are the same number and the reason all three are printed is that od speaks octal, xxd hex and Python's bytes decimal: Hex is a shorthand. The Char column is empty for the no-break space and the soft hyphen, which print as nothing. The Description column is the Unicode character name, so every row is a byte-to-code-point mapping with the number left out: é at E9 is U+00E9, and for Latin-1 alone byte and code point are always equal, the fact A character is a number rests on. On every other page the name is the only key, and the way back to a number is unicodedata.lookup or uni.
iso_8859-1(7) and iso_8859-15(7): eight bytes apart¶
< 244 164 A4 ¤ CURRENCY SIGN
> 244 164 A4 € EURO SIGN
< 246 166 A6 ¦ BROKEN BAR
> 246 166 A6 Š LATIN CAPITAL LETTER S WITH CARON
< 250 168 A8 ¨ DIAERESIS
> 250 168 A8 š LATIN SMALL LETTER S WITH CARON
< 264 180 B4 ´ ACUTE ACCENT
> 264 180 B4 Ž LATIN CAPITAL LETTER Z WITH CARON
< 270 184 B8 ¸ CEDILLA
> 270 184 B8 ž LATIN SMALL LETTER Z WITH CARON
< 274 188 BC ¼ VULGAR FRACTION ONE QUARTER
< 275 189 BD ½ VULGAR FRACTION ONE HALF
< 276 190 BE ¾ VULGAR FRACTION THREE QUARTERS
> 274 188 BC ΠLATIN CAPITAL LIGATURE OE
> 275 189 BD œ LATIN SMALL LIGATURE OE
> 276 190 BE Ÿ LATIN CAPITAL LETTER Y WITH DIAERESIS
The two pages are 96 rows each and differ in exactly eight. Latin-9, which the NOTES also calls Latin-0, was Latin-1 revised for the euro, and the eight victims were the eight characters judged least missed: the currency placeholder ¤, the broken bar, three spacing accents and the three fractions. In their places, €, the Finnish Š š Ž ž and the French Œ œ Ÿ that Latin-1 had left out, the omission charsets(7) records as considered tolerable. The consequence for data is that a Latin-9 file holding € is a Latin-1 file holding ¤, byte for byte, and nothing in the file says which: Mojibake.
iso_8859-2(7) and iso_8859-16(7): where ż and Ł live¶
243 163 A3 Ł LATIN CAPITAL LETTER L WITH STROKE
...
277 191 BF ż LATIN SMALL LETTER Z WITH DOT ABOVE
...
363 243 F3 ó LATIN SMALL LETTER O WITH ACUTE
Latin-2 is the Central European page and the one Polish text was written in before UTF-8. Two of the library's cast are on it, ż at BF and Ł at A3, and Łódź encodes as a3 f3 64 bc. Notice ó at F3: the letters Latin-2 shares with Latin-1 sit at the same bytes, so a Polish word misread as Latin-1 keeps its vowels and loses its consonants, £ód¼, which is how Mojibake teaches you to recognise the culprit from the damage. Latin-10, iso_8859-16(7), is Romanian in its NAME line and moves Ł ł Ż ż to other bytes while adding € at A4 and the comma-below Ș ș Ț ț that charsets(7) says Latin-2 had settled for cedillas instead of.
cp1252(7) and cp1251(7): the C1 block filled in¶
The Windows Code Pages include several 8-bit extensions to the ASCII
character set (also known as ISO/IEC 646-IRV). CP 1252 encodes the
characters used in many West European languages.
...
200 128 80 € EURO SIGN
202 130 82 ‚ SINGLE LOW-9 QUOTATION MARK
214 140 8C ΠLATIN CAPITAL LIGATURE OE
216 142 8E Ž LATIN CAPITAL LETTER Z WITH CARON
221 145 91 ‘ LEFT SINGLE QUOTATION MARK
Read the octal column and watch it skip: 200, 202, then 214, 216, 221. The page has 123 rows where Latin-1 has 96, because Windows-1252 puts printable characters in the 32 positions ISO reserves for C1 controls, and it has 27 rows in that block rather than 32, because five bytes, 81 8D 8F 90 9D, are assigned nothing at all. That is the count Windows-1252 vs Latin-1 makes, and the page's own table agrees with it: Windows-1252 is not a superset of Latin-1, since over these 32 bytes it keeps none of Latin-1's mappings and drops five. The experiment below shows both iconvs refusing the five and Latin-1 accepting everything, which is why one table can always round-trip and the other cannot: The mojibake round trip. cp1251(7), Windows Cyrillic, fills the same block with Serbian and Macedonian letters and the same typographic quotes, puts € at 88, and lays out the Russian alphabet in dictionary order from А at C0 to я at FF, which is exactly what KOI8-R does not do.
koi8-r(7) and koi8-u(7): the ordering trick¶
300 192 C0 ю CYRILLIC SMALL LETTER YU
301 193 C1 а CYRILLIC SMALL LETTER A
302 194 C2 б CYRILLIC SMALL LETTER BE
303 195 C3 ц CYRILLIC SMALL LETTER TSE
304 196 C4 д CYRILLIC SMALL LETTER DE
305 197 C5 е CYRILLIC SMALL LETTER IE
...
NOTES
The differences with KOI8-U are in the hex positions A4, A6, A7, AD,
B4, B6, B7, and BD.
ю а б ц д е is not any alphabet's order. Subtract 0x80 from each byte and read the ASCII: @ A B C D E F G, and а is A, б is B, ц is C, д is D. KOI8-R orders its Cyrillic by the Latin letter that transliterates it, so that a byte stream with its high bit stripped, which is what a seven-bit channel does to it, comes out as readable Russian in Latin letters rather than as noise. The lowercase letters sit against ASCII's uppercase, C0–DF against 40–5F, so the case is inverted on the way, and the experiment below shows it. This is the design charsets(7) calls somewhat better designed than ISO/IEC 8859-5 without saying why, and it is also why sorting KOI8-R bytes does not sort Russian: A character is a number is a number chosen for a reason, and the reason here was transmission, not collation. The NOTES on each KOI8 page name eight positions; a diff of the two dumps finds exactly those eight, box-drawing pieces in KOI8-R replaced by the Ukrainian є і ї ґ and their capitals in KOI8-U.
armscii-8(7): a second copy of the punctuation¶
244 164 A4 ) RIGHT PARENTHESIS
245 165 A5 ( LEFT PARENTHESIS
...
251 169 A9 . FULL STOP
...
253 171 AB , COMMA
254 172 AC - HYPHEN-MINUS
...
262 178 B2 Ա ARMENIAN CAPITAL LETTER AYB
263 179 B3 ա ARMENIAN SMALL LETTER AYB
The Armenian page has 94 rows and two things on it break rules the other pages keep. Bytes A4, A5, A9, AB and AC are a parenthesis pair, a full stop, a comma and a hyphen, characters ASCII already has at 29, 28, 2E, 2C and 2D, so under this table one character has two byte spellings and the 1:1 mapping utf8(5) treats as a security property does not hold. (Note the pair is stored closing-first, ) before (.) And the letters alternate capital, small, capital, small from B2, so the case bit of an Armenian letter is bit 0, not the 0x20 of ascii(7). A byte-oriented tolower written for Latin gets both wrong.
Terms on the page¶
| On the page | What it means | Read more |
|---|---|---|
| ISO/IEC 646-IRV | The International Reference Version of ISO 646, which is ASCII; the pages' formal name for the low half every code page shares | ascii(7) |
| invisible control characters in positions 128 to 159 | The C1 controls, 0x80–0x9F; why the ISO tables start at A0 and why Windows put printable characters there instead |
Control characters |
| 96 fixed-width graphics | 0xA0–0xFF: the one printable block an ISO 8859 part defines; 96 rows on each ISO page |
Code pages |
| Latin-1 … Latin-10 | The alphabet numbering, which does not follow the part numbering: Latin-5 is 8859-9, Latin-9 is 8859-15, Latin-10 is 8859-16 | Code pages |
Oct Dec Hex Char Description |
One byte in three bases, its glyph, and its Unicode character name; the name is the only key to the code point on nine of the ten pages | Hex is a shorthand |
| NO-BREAK SPACE, SOFT HYPHEN | A0 and AD on every ISO page, printed with an empty Char cell; both invisible, both a common source of "the two strings look equal" bugs |
A code point is not a character |
CURRENCY SIGN ¤ |
Latin-1's A4: a placeholder for whichever currency you use, which is why the euro took its byte |
Windows-1252 vs Latin-1 |
| Windows Code Pages, CP 1252, Windows-1252, Windows Cyrillic | Microsoft's numbered single-byte tables; the same tables under SAP's and IBM's numbers are on the SAP page | SAP code pages |
| KOI8-R, KOI8-U, RFC 1489, RFC 2310 | The Russian and Ukrainian eight-bit sets, each defined by an informational RFC rather than an ISO part; KOI is a Russian abbreviation for code for information interchange | A character is a number |
G0, G1, G2, G3; SO, SI; SS2, SS3; ESC ( B |
ISO 2022's four table slots, the shift bytes that select one, and the escape that designates ASCII into G0 | The CJK pages |
| ArmSCII-8 | Armenian Standard Code for Information Interchange, eight-bit; the page's expansion of its own name | |
| 94×94, EUC-CN, EUC-JP, JIS X 0208, not ISO/IEC 2022 compatible | The two-byte national sets and the encodings built on them, one paragraph each here and six pages on the Mac; KOI8 and Big5 are not compatible because their upper halves are not a set the escape scheme could switch to | The CJK pages |
Try it on your machine¶
Where the pages are. Ten on Ubuntu, none on the Mac; every table on both.
macOS ubuntu:24.04
$ man -w 7 charsets No manual entry for charsets /usr/share/man/man7/charsets.7.gz
$ locale -a | grep -cE 'ISO8859-2$|ISO8859-15$|KOI8|CP1251|ARMSCII'
50 0 (C, C.utf8, POSIX only)
$ LC_ALL=pl_PL.ISO8859-2 locale charmap
ISO8859-2 ANSI_X3.4-1968, after three 'Cannot set ... to default locale' warnings
One row from each page, and what iconv says about it. printf '\xNN' | iconv -f NAME -t UTF-8, shown as the character and its UTF-8 bytes. Byte-identical on both machines.
byte ISO-8859-1 ISO-8859-15 CP1252 KOI8-R ISO-8859-2 ARMSCII-8
80 (C1) c280 (C1) c280 € e282ac ─ e29480 (C1) c280 (C1) c280
a3 £ c2a3 £ c2a3 £ c2a3 ё d191 Ł c581 ։ d689
a4 ¤ c2a4 € e282ac ¤ c2a4 ╓ e29593 ¤ c2a4 ) 29
bf ¿ c2bf ¿ c2bf ¿ c2bf © c2a9 ż c5bc է d5a7
c1 Á c381 Á c381 Á c381 а d0b0 Á c381 ը d5a8
$ man 7 iso_8859-15 | grep -E '^ +244 ' 244 164 A4 € EURO SIGN
$ man 7 iso_8859-1 | grep -E '^ +244 ' 244 164 A4 ¤ CURRENCY SIGN
One byte, six meanings, and the rows are the pages' own rows: grep the octal on Ubuntu and iconv on either machine agree on every cell. a4 is the page's whole story in one line: the currency placeholder, the euro, the placeholder again, a box-drawing corner, and in ArmSCII a closing parenthesis. (C1) is the byte decoding to the control character U+0080, invisible and legal under every table but Windows-1252 and KOI8-R, refused by none of them.
The five holes in Windows-1252, and the eight changes in Latin-9. The counts, made from the dumps and checked against two iconvs and Python.
$ diff <(grep -E '^ {7}[0-7]{3} ' raw/linux/iso_8859-1.7.txt) <(grep -E '^ {7}[0-7]{3} ' raw/linux/iso_8859-15.7.txt) | grep -c '^<'
8
$ grep -cE '^ {7}2[0-3][0-7] ' raw/linux/cp1252.7.txt # rows in octal 200-237, i.e. 0x80-0x9F
27
missing from cp1252.7.txt: 201(=81) 215(=8D) 217(=8F) 220(=90) 235(=9D)
$ printf '\x81' | iconv -f CP1252 -t UTF-8 iconv: illegal input sequence at position 0 (Apple: Illegal byte sequence)
$ printf '\x81' | iconv -f ISO-8859-1 -t UTF-8 c2 81 (U+0081, a C1 control)
$ python3 -c "..." cp1252 refuses: 81 8D 8F 90 9D
$ python3 -c "print(bytes(range(0x80,0xa0)).decode('cp1252','replace'))"
€�‚ƒ„…†‡ˆ‰Š‹Œ�Ž��‘’“”•–—˜™š›œ�žŸ
Twenty-seven assigned, five refused, and the five are the same five on the page, in both iconvs and in Python's codec. The last line is the 32-byte block decoded with errors='replace', and the five replacement characters sit exactly where the octal column skipped. Latin-1 never refuses, because it has a character for all 256 bytes; that asymmetry is the whole of The mojibake round trip.
KOI8-R with its high bit stripped. Encode a Russian phrase, clear bit 7 of every byte, decode the result as ASCII.
$ python3 -c "
s='Привет, мир'.encode('koi8-r')
print('koi8-r bytes :', s.hex(' '))
print('bit 7 cleared:', bytes(b & 0x7f for b in s).decode('ascii'))
"
koi8-r bytes : f0 d2 c9 d7 c5 d4 2c 20 cd c9 d2
bit 7 cleared: pRIWET, MIR
$ printf '\xa4' | iconv -f KOI8-R -t UTF-8 ╓ $ printf '\xa4' | iconv -f KOI8-U -t UTF-8 є
Privet, mir with the case swapped, because KOI8-R's lowercase sits against ASCII's uppercase. ц at C3 came out as C and в at D7 as W, the transliteration the table was built around; the comma and the space, being ASCII, came through untouched. The last line is one of the eight positions the NOTES name, a box-drawing piece in Russian and a Ukrainian letter in the neighbour.
Latin-2 has the Polish letters and Latin-1 does not, and the one place the machines differ. Łódź encoded into three tables, then its Latin-2 bytes read back under the wrong one.
macOS ubuntu:24.04
$ printf 'Łódź' | iconv -f UTF-8 -t ISO-8859-2 | xxd -p a3f364bc a3f364bc
$ printf 'Łódź' | iconv -f UTF-8 -t ISO-8859-1 | xxd -p 4cf364b47a exit 0 illegal input sequence at position 0 exit 1
$ printf 'Łódź' | iconv -f UTF-8 -t CP1252 | xxd -p 4cf364b47a exit 0 illegal input sequence at position 0 exit 1
$ printf '\xa3\xf3\x64\xbc' | iconv -f ISO-8859-1 -t UTF-8 £ód¼ £ód¼
Row one is the page: Ł at A3, ó at F3, ż's neighbour ź at BC. Rows two and three are the difference. GNU iconv refuses to write Ł into a table that has no Ł, and says so with exit 1; Apple's writes 4c f3 64 b4 7a, which is Lód´z, with the stroke dropped and the acute turned into a spacing accent, and exits 0 as if nothing had happened. GNU will do something similar only when asked with //TRANSLIT, and does it differently: 4c f3 64 7a, Lódz, measured on Ubuntu only. Which behaviour you get is a fact about the iconv on the machine, and neither is on any page here: iconv. The last row is the mojibake the page predicts: the bytes are Polish, the reader is Latin-1, the vowels survive.
Where the page is dated, and what it does not say¶
charsets(7) carries a 1996 copyright and a 2024 date, and reads as both: the code-page paragraphs describe a present that is now a past, and its Unicode paragraph still prints the six-byte UTF-8 progression and a 512-glyph Linux console. Its map is nonetheless the only one on either machine. Nor does koi8-r(7) explain its own ordering: the transliteration design is visible in the table and named nowhere, and charsets(7) calls it better designed and stops.
The table pages give names, not numbers. Every row's Description is a Unicode character name, and no row says U+. For Latin-1 the number is the byte; for the other nine you look the name up.
No page says what happens to a byte it does not list. The C1 range on the ISO pages and the five holes on cp1252(7) are simply absent; that iconv refuses the holes and passes the C1 bytes through as controls is measured above and written nowhere.
cp1252(7) does not say that browsers read ISO-8859-1 as Windows-1252, the rule that makes the superset claim feel true on the web and false in a data pipeline: Windows-1252 vs Latin-1.
None of the ten exists on the Mac, and the Mac's iconv(1) page does not point at anything that replaces them. The nearest thing on the Mac to man 7 cp1252 is iconv -l, which gives the names and none of the rows.
See also¶
ascii(7)— the low half every one of these pages says it leaves oututf8(5)andutf-8(7)— the encoding that made all ten tables legacy, and the 1:1-mapping rule ArmSCII breaksiconv(1)andiconv(3)— the tool that has every table above and whose Apple build substitutes silentlylocale(5)andlocaledef(1)— the charmap files these tables are compiled from on Linux- The CJK pages — the two-byte sets
charsets(7)gives a paragraph each and the Mac gives a page each - Code pages — ASCII plus a second opinion, the idea behind every table here
- Windows-1252 vs Latin-1 — the count the
cp1252(7)table confirms - Mojibake —
£ód¼andé, and how to name the culprit from the damage - What the page does not say — the general rule for the gaps listed above