Refuse local-use NAT64 addresses, and IPv4-compatible ones in blocked_networks (closes #299)

The check that keeps pixa from fetching internal addresses let through
64:ff9b:1::/48, the local-use NAT64 range, and matched an address in the
IPv4-compatible form (::/96) against blocked_networks only as IPv6, so
::203.0.113.5 passed with 203.0.113.0/24 listed. The local-use NAT64
range is now blocked next to 64:ff9b::/96, and an address in ::/96 is
refused when blocked_networks holds it as written or the IPv4 address it
holds. The mapped form stays matched as IPv4 only, as Go keeps plain
IPv4 addresses in that form; README.md says so.

Model: opus-5-5

README.md

@@ -323,25 +323,34 @@ first `/`, the size and format after the last `/`, and between them the source
 path up to the first `?` and the source query after it. So the `?` before the
 query, any `?` in the query, and every `%` in the source path or query must be
 percent-encoded, as `%3F` and `%25`. Any other character may be written as
-itself or percent-encoded: `/`, `;`, `,`, `=` and `&` work either way. For
+itself or percent-encoded: `/`, `;`, `,`, `=` and `&` work either way. Neither
-example,
+the source path nor the source query may hold a `#`, and an image URL that
+writes one as `%23` is refused with 400. For example,
 `https://cdn.example.com/img?url=https://origin.example.com/my%20cat.jpg` at
 800x600 WebP is
 `/v1/image/cdn.example.com/img%3Furl=https://origin.example.com/my%2520cat.jpg/800x600.webp`,
 signed as shown under Signature Specification.
 
+The source path is the path with its percent-escapes decoded, and pixa fetches
+it percent-encoded again: every character other than letters, digits, `-`, `.`,
+`_`, `~`, `/`, `$`, `&`, `+`, `,`, `:`, `;`, `=` and `@` is sent as a
+percent-escape. So `my%20cat.jpg` in the image URL is the source path
+`/my cat.jpg`, fetched as `/my%20cat.jpg`; `my%2520cat.jpg` is the source path
+`/my%20cat.jpg`, fetched as `/my%2520cat.jpg`; and `100%25.png` is fetched as
+`/100%25.png`. The source query is the source URL's query as that URL writes it,
+its percent-escapes included, and is sent as it is, except that the characters a
+browser percent-encodes in a query (a space, `"`, `'`, `<`, `>`, control
+characters and characters outside ASCII) and a `%` that does not begin a
+percent-escape are percent-encoded. pixa signs and caches the query in that
+form, so `img%3Fname=my%20cat` and `img%3Fname=my%2520cat` in the image URL are
+the same source query, `name=my%20cat`.
+
 A size with one side `0` has that side scaled to keep the source's proportions,
 and is refused with 400 when that side would come out above 8192, the most
 either side may be; so is a size with `fit=outside`, which scales the image to
 cover the size and so past it on one side, when that side would come out
 above 8192.
 
-An AVIF image of more than 16,777,216 pixels, the number 4096x4096 and 8192x2048
-have, is refused with 400 before anything is decoded, as its encode could take
-more than 40 seconds (see below). That includes `orig` of an AVIF source, and
-`auto` when it chooses AVIF, so a client whose `Accept` names `image/avif` but
-not `image/jxl` is refused such a size.
-
 The source image may be JPEG, PNG, GIF, WebP, AVIF or JPEG XL. The upstream's
 `Content-Type` must name its format, and the image's first bytes must match it.
 An upstream answering with any other `Content-Type`, SVG included, is refused
@@ -391,18 +400,14 @@ transcode that stops this way is answered 503 with the error
 ended. govips gives pixa no way to stop libvips once an encode has begun, so an
 encode keeps its slot until it finishes, which can be after
 `downstream_timeout`. The limit of 8192 on each side of the source and of the
-output, and for AVIF the limit of 16,777,216 pixels, bound how long that takes.
+output bounds how long that takes. With one libvips thread, as pixad runs it,
-With one libvips thread, as pixad runs it, encoding 8192x8192 random pixels
+encoding 8192x8192 random pixels takes up to about 1 second as JPEG, 15 as PNG,
-takes up to about 1 second as JPEG, 15 as PNG and 16 as GIF, the slowest of 8 or
+16 as GIF, 63 as WebP, 110 as AVIF and 96 as JPEG XL, the slowest of 8 or 16
-16 bits per sample, with or without an alpha channel. The WebP and JPEG XL
+bits per sample, with or without an alpha channel. WebP, AVIF and JPEG XL are
-encoders run below libvips' default effort, and AVIF at the lowest effort govips
+that slow only with an alpha channel: 8-bit RGB takes about 44, 53 and 28
-can set, so that no encode takes more than 40 seconds of the default
+seconds. Decoding such a source adds up to about 11 seconds. A WebP, AVIF or
-`downstream_timeout` of 60: at any quality and 8 or 16 bits per sample,
+JPEG XL encode can therefore hold its slot past the default `downstream_timeout`
-8192x8192 random pixels with an alpha channel take up to about 23 seconds as
+of 60 seconds.
-WebP and 27 as JPEG XL, and 4096x4096, 8192x2048 or 2048x8192 up to about 34 as
-AVIF, which at quality 85 would take about 96 at 8192x8192. Without an alpha
-channel they take up to about 16, 13 and 22 seconds. Decoding such a source adds
-up to about 11 seconds.
 
 The login form (`POST /`) is limited to 5 attempts per minute per client
 address, counting an IPv6 client by its /64; an attempt over the limit is
@@ -454,8 +459,13 @@ nor change what it asks for.
    sent with an empty format, or none, also makes a JPEG XL URL. Width and
    height both empty or `0` keep the original size; if only one of them is empty
    or `0`, that side is scaled to keep the image's proportions. A URL whose
-   scaled side would come out above 8192 answers 400, and so does an AVIF URL of
+   scaled side would come out above 8192 answers 400, and so does a source URL
-   more than 16,777,216 pixels (see Routes).
+   whose path holds a `..` segment. So does a source URL that the image URL
+   would fetch written another way, and the page then shows the URL pixa would
+   fetch: one with a user name and password, which pixa does not send, or one
+   whose path holds `%2F`, fetched as `/`, or `(`, fetched as `%28`, as pixa
+   keeps the source path decoded and fetches it percent-encoded again (see
+   Routes).
 3. The page shows the URL, `https://<host>/v1/e/<token>/img.<format>`, and when
    it expires. `<host>` is the host the page was opened on, and the URL starts
    with `http` instead while `debug` is on. The name after the token is ignored
@@ -463,9 +473,11 @@ nor change what it asks for.
    `jxl` for a form with no format.
 
 The token holds the source's host, path and query and the size, format, quality,
-fit and expiry, encrypted with a key derived from `signing_key`. The source
+fit and expiry, encrypted with a key derived from `signing_key`. The host is
-URL's scheme is not kept: the image is fetched like any other (see Routes), and
+kept in lower case and an empty path as `/`, as for `/v1/image/`, so
-the blocked networks still apply.
+`https://CDN.Example.com` and `https://cdn.example.com/` are one source image,
+fetched and cached once. The source URL's scheme is not kept: the image is
+fetched like any other (see Routes), and the blocked networks still apply.
 
 How long the URL lasts is chosen on the page, from 1 minute to 1 year, or never.
 The expiry is fixed in the token when the URL is made and cannot be changed or
@@ -517,13 +529,20 @@ Where:
 - `host` — source host as the URL writes it up to the first `/`, in lower case,
   keeping an IPv6 address's brackets and any `:` and port after it (e.g.
   `cdn.example.com`, `cdn.example.com:8443`, `[2001:db8::1]:8443`)
-- `path` — source path (e.g. `/photos/cat.jpg`), with each `%` written as `%25`
+- `path` — source path, with its percent-escapes decoded as under Routes (e.g.
-  and each `:` as `%3A`, so `/photos/a:b.jpg` is signed as `/photos/a%3Ab.jpg`
+  `/photos/cat.jpg`, or `/photos/my cat.jpg` for `/photos/my%20cat.jpg`), with
-- `query` — source query string, empty string if none, with `%` and `:` written
+  each `%` written as `%25` and each `:` as `%3A`, so `/photos/a:b.jpg` is
-  as in `path`; neither then holds a `:`, so path `/a:b` with no query and path
+  signed as `/photos/a%3Ab.jpg`
-  `/a` with query `b:` are signed differently
+- `query` — source query string as under Routes, empty string if none, with `%`
-- `width` — requested width in pixels, `0` for original
+  and `:` written as in `path`; neither then holds a `:`, so path `/a:b` with no
-- `height` — requested height in pixels, `0` for original
+  query and path `/a` with query `b:` are signed differently
+- `width` — requested width in pixels, or `0`: the width is then worked out from
+  the source's proportions and the requested height, rounded down but at least
+  1, or is the source's own width when the height is `0` too (the `orig` size)
+- `height` — requested height in pixels, or `0`: the height is then worked out
+  from the source's proportions and the requested width, rounded down but at
+  least 1, or is the source's own height when the width is `0` too (the `orig`
+  size)
 - `format` — output format, one of those listed under Routes, with `original`
   signed as `orig`, `jpg` as `jpeg`, and no format as `jxl`, so a URL with no
   format has the signature of the same URL ending in `.jxl`; `auto` is signed as
@@ -536,7 +555,10 @@ Where:
   with 400
 - `fit` — the URL's `fit` query parameter (cover, contain, fill, inside,
   outside), or `cover` when the URL has no `fit`; a request whose `fit` is
-  anything else, an empty `fit=` included, is refused with 400
+  anything else, an empty `fit=` included, is refused with 400. `outside` scales
+  the image to cover the requested size, keeping its proportions, so it can come
+  out larger than that size on one side: an 8192x1 source at `100x0`, which is
+  100x1 by its proportions, comes out at 8192x1
 
 The URL's `sig` is the HMAC-SHA256 result in base64url (the URL-safe alphabet of
 RFC 4648) with the trailing `=` padding kept, 44 characters in all. pixa
@@ -586,7 +608,11 @@ pixa lower-cases the host of a `/v1/image/` URL before anything else uses it, so
 `CDN.Example.com` is matched, fetched, cached and counted against
 `upstream_connections_per_host` as `cdn.example.com`. Such a URL whose host has
 a port, as in `cdn.example.com:8443`, or a `:` with no port, needs a signature
-even when the host is allowlisted. A host holding `@` is refused with 400.
+even when the host is allowlisted. The host must be a host name or an IPv4
+address (letters, digits, hyphens, underscores and dots) or an IPv6 address in
+brackets, either one optionally followed by `:` and a port. Any other host, such
+as one holding `@`, or a `?`, `#`, `%`, `\` or space written percent-encoded, is
+refused with 400.
 
 ### Configuration
 
@@ -668,7 +694,10 @@ Key settings in more detail:
   and generator pages. Default: empty
 - `blocked_networks` — list of CIDR ranges to refuse for SSRF protection, added
   to the always-enforced built-in ranges (loopback, private, link-local, CGNAT,
-  benchmark, NAT64, and the like); an invalid CIDR aborts startup
+  benchmark, NAT64, and the like); an invalid CIDR aborts startup. An IPv6
+  address in the IPv4-mapped form (`::ffff:a.b.c.d`) is matched only as the IPv4
+  address it holds, so an IPv6 range such as `::/0` refuses neither it nor a
+  plain IPv4 address
 - `trusted_proxies` — list of CIDR ranges of the reverse proxies in front of
   pixa. `X-Forwarded-For` is believed only when the direct peer falls inside one
   of these ranges; the logged and login-recorded client address is then the
@@ -699,11 +728,14 @@ Key settings in more detail:
 - `signing_key` — HMAC secret for URL signatures
 - `db_url` — the SQLite database to open; omitted, it is
   `file:<state_dir>/state.sqlite3?_pragma=journal_mode(WAL)`, which keeps the
-  database in WAL mode. pixa opens one connection to it, so its own reads and
+  database in WAL mode. `state_dir` is percent-encoded in it, so a `%`, `?` or
-  writes run one at a time. pixa adds `_pragma=busy_timeout(5000)` to any
+  `#` in the name is written as `%25`, `%3F` or `%23`. For a `state_dir` of `/`
-  `db_url`, so a write that finds another program writing to the file waits up
+  or one starting with `//` it starts `file:////`, so SQLite does not read the
-  to five seconds for it instead of failing. WAL mode comes only from the URL:
+  first name in the path as a host. pixa opens one connection to it, so its own
-  keep `_pragma=journal_mode(WAL)` in one you set
+  reads and writes run one at a time. pixa adds `_pragma=busy_timeout(5000)` to
+  any `db_url`, so a write that finds another program writing to the file waits
+  up to five seconds for it instead of failing. WAL mode comes only from the
+  URL: keep `_pragma=journal_mode(WAL)` in one you set
 - `cache_max_bytes` — disk cache size limit in bytes; `0` disables the disk
   cache entirely; omitted defaults to 75% of the sum of the free space on the
   filesystem containing `<state_dir>/cache/` and the bytes of source and

TODO.md

@@ -30,18 +30,100 @@ P2: security: per-origin rate limiting
 
 # Completed Steps
 
-- 2026-10-15 a WebP, AVIF or JPEG XL encode at the limits takes at most about 34
+- 2026-10-16 P1 security bug: the check that keeps pixa from fetching internal
-  seconds (closes #272): with one libvips thread, 8192x8192 random pixels with
+  addresses refuses the local-use NAT64 range (`64:ff9b:1::/48`, RFC 8215) as it
-  an alpha channel took about 61 seconds to encode as WebP, 96 as AVIF and 81 as
+  refuses `64:ff9b::/96`, and refuses an IPv6 address in the IPv4-compatible
-  JPEG XL, so with the default `downstream_timeout` of 60 seconds such an image
+  form (`::/96`) when `blocked_networks` contains either the address as written
-  was never delivered, yet its encode held a processing slot to the end. WebP is
+  or the IPv4 address it holds (closes #299). A source host such as
-  now saved at effort 2 and JPEG XL at effort 4, which take up to about 23 and
+  `[64:ff9b:1::a00:1]` was fetched, which on a network that runs NAT64 on that
-  27 seconds. AVIF was already at the lowest effort govips can set, so an AVIF
+  range reaches the IPv4 address it holds, here `10.0.0.1`; and with
-  output of more than 16,777,216 pixels is refused with 400 before anything is
+  `203.0.113.0/24` in `blocked_networks`, `[::ffff:203.0.113.5]` was refused but
-  decoded; 4096x4096, 8192x2048 or 2048x8192 takes up to about 34 seconds. Each
+  `[::203.0.113.5]` was fetched. Both now answer 403. The mapped form
-  time is the slowest at any quality. Resizing with `fit=cover` already decodes
+  (`::ffff:a.b.c.d`) is still matched only as the IPv4 address it holds: Go
-  the input, so the output's dimensions are worked out before resizing.
+  keeps a plain IPv4 address in that form, and an IPv6 entry such as `::/0` must
-  `README.md` gives the limit and the new times.
+  not refuse every IPv4 address.
+- 2026-10-16 P1 security bug: the check that keeps pixa from fetching internal
+  addresses judges an IPv6 address in the IPv4-compatible form (`::/96`), such
+  as `::127.0.0.1` or `::10.0.0.1`, by the IPv4 address in its last four bytes
+  (closes #297). It read only the mapped form (`::ffff:a.b.c.d`) as IPv4, so a
+  source host such as `[::127.0.0.1]` was checked as a public IPv6 address and
+  fetched, which on a network that routes it reaches the internal IPv4 host; it
+  is now refused with 403. The NAT64 form (`64:ff9b::/96`) was already refused
+  as a whole, as one of the built-in blocked ranges, and still is.
+- 2026-10-16 a source host that is an IPv6 address in brackets with no port,
+  such as `[2001:db8::a]`, is fetched like one with a port, through `/v1/image/`
+  and through an encrypted URL (closes #295). The fetcher's check of the source
+  URL kept the brackets when no port followed them, could not look the host up,
+  and the request answered 500. It now checks the address without brackets, so a
+  private one is refused with 403, as with a port.
+- 2026-10-16 a `state_dir` whose name holds `%`, `?` or `#` keeps its database
+  at `state.sqlite3` in that directory (closes #290): the `db_url` derived from
+  it put the name into a `file:` URL as it was, so SQLite read `%41` as `A`, a
+  `?` or `#` ended the path early, and the journal mode parameter was lost. The
+  derived `db_url` is now built with Go's `net/url`, which writes those
+  characters as `%25`, `%3F` and `%23`. For a `state_dir` of `/` or one starting
+  with `//` it starts `file:////`, so SQLite no longer reads the first name in
+  the path as a host.
+- 2026-10-16 a size with one side `0` keeps the side computed from the source's
+  proportions at 1 or more (closes #283): it was rounded down, so for a thin
+  source it came out at 0, and libvips then made the size square. An 8192x1
+  source at `100x0` came out at 100x100, and at `1300x0` with `fit=cover` made
+  libvips ask for about 39 GB at once; an 8192x2 source at `100x0` answered 500.
+  `fit=contain` and `fit=inside` work both sides out again from the smaller of
+  the two scales and `fit=outside` from the larger, and each keeps both sides at
+  1 or more too, so a 1x8192 source at `0x100` with `fit=contain`, and a 49x8192
+  source at `0x100` or a 49x49 source at `1x1` with `fit=outside`, no longer
+  answer 500. An 8192x1 source at `1300x0` now comes out at 1300x1, and the
+  process running it alone peaked at about 34 MiB resident. With `fit=outside`
+  the image is scaled to cover the size, so that source at `100x0` comes out at
+  8192x1. `README.md` now says this under Signature Specification, and says how
+  a `0` side is worked out where it said `0` meant the original.
+- 2026-10-16 the generator page lower-cases the source host and writes an empty
+  source path as `/` before it makes the encrypted URL, as `/v1/image/` does
+  (closes #293). A host written in upper and in lower case, or a source URL with
+  no path and with the path `/`, made two cache keys and two upstream requests
+  for one image.
+- 2026-10-16 the source URL a request fetches is built with `net/url` from the
+  source host, path and query, not joined as text (closes #288). The source
+  path, decoded, is fetched percent-encoded again, so `100%25.png` fetches
+  `/100%25.png` where it answered 401, and `my%20cat.png` and `my%2520cat.png`
+  fetch different paths where both fetched `/my%20cat.png` under two cache keys.
+  The source query stays as the source URL writes it; the characters a browser
+  percent-encodes in a query, such as `'` or a space, which was sent raw and
+  answered 502, are percent-encoded when the image URL is read, so the query is
+  signed and cached as it is sent. The generator page answers 400, showing the
+  URL pixa would fetch, for a source URL its image URL would fetch written
+  another way, such as with `%2F` in the path, whatever else the path holds, or
+  with a user name and password. It also answers 400 for a source path holding a
+  `..` segment, as `/v1/image/` refuses one.
+- 2026-10-15 a thin source requested with `fit=cover` no longer makes libvips
+  build a huge image (closes #282): libvips' thumbnail scaled the whole source
+  to cover the requested size and copied that into memory before it cut out the
+  centre, so an 8192x64 source at 8192x8192 made it allocate about 24 GiB, and
+  an 8192x1 source answered 500. The image processor now first cuts the centre
+  of the source, widened to whole source pixels, and thumbnail scales and cuts
+  only that, by the same amount on both sides as before. What it copies into
+  memory is then larger than the output by less than one scaled source pixel on
+  each side. With the runtime image's libvips and JPEG output at 8192x8192,
+  resident memory peaked at about 242 MiB for the 8192x64 source and 431 MiB for
+  the 8192x1 one. The output can sit a fraction of an output pixel from where it
+  was.
+- 2026-10-15 a `/v1/image/` URL whose source path or source query holds a `#`,
+  written as `%23`, is refused with 400 (closes #286): the source URL built from
+  it ended at the `#` when fetched, so each text after it fetched the same
+  upstream image, without a signature for an allowlisted host, and cached it
+  under keys of its own. `README.md` says neither may hold a `#`.
+- 2026-10-15 the host of a `/v1/image/` URL must be a host name or an IPv4
+  address (letters, digits, hyphens, underscores and dots) or an IPv6 address in
+  brackets, either one optionally followed by `:` and a port, and any other host
+  is refused with 400 (closes #275): only `@` was refused, so a `?` written as
+  `%3F` gave the host `cdn.example.com?x=1`, which Go's URL parser reads as
+  `cdn.example.com` with a query. Such a URL passed the allowlist, and each text
+  after the `?` fetched the host's root path and cached the answer under keys of
+  its own. A `#`, `%`, `\` or space written percent-encoded is refused the same
+  way, and so is an IPv6 address without brackets, which needed a signature
+  before. Underscores are kept because `allowlist_hosts` accepts host names
+  holding them.
 - 2026-10-15 the README says how a source URL with a query is written into an
   image URL, and the image route accepts it (closes #255): the query follows the
   source path after `%3F`. pixa now reads the image URL's path decoded once, so

internal/config/config.go

@@ -338,8 +338,7 @@ func newFromSmartConfig(sc *smartconfig.Config) (*Config, error) {
 				settingName(keyDBURL), errValueEmpty)
 		}
 
-		// The driver sets the journal mode only through a _pragma parameter.
+		c.DBURL = defaultDBURL(c.StateDir)
-		c.DBURL = fmt.Sprintf("file:%s/state.sqlite3?_pragma=journal_mode(WAL)", c.StateDir)
 	}
 
 	if loader.err != nil {
@@ -354,6 +353,30 @@ func newFromSmartConfig(sc *smartconfig.Config) (*Config, error) {
 	return c, nil
 }
 
+// defaultDBURL returns the db_url used when db_url is omitted: the file
+// state.sqlite3 in stateDir, in WAL mode, which the driver sets only
+// through a _pragma parameter. String() percent-encodes any %, ? or # in
+// the directory name, which SQLite decodes back, so the URL names exactly
+// this file. OmitHost writes file:/var/lib/pixa/... instead of
+// file:///var/lib/pixa/..., and keeps a relative stateDir from being read
+// as a host name. The path written into the URL is stateDir followed by
+// /state.sqlite3; when it starts with //, as it does for a stateDir of /
+// or one starting with //, the empty host is kept, so the URL starts
+// file://// and SQLite does not read the first name in the path as a
+// host.
+func defaultDBURL(stateDir string) string {
+	dbPath := stateDir + "/state.sqlite3"
+
+	dbURL := url.URL{
+		Scheme:   "file",
+		OmitHost: !strings.HasPrefix(dbPath, "//"),
+		Path:     dbPath,
+		RawQuery: "_pragma=journal_mode(WAL)",
+	}
+
+	return dbURL.String()
+}
+
 // validateKnownKeys rejects configuration files containing keys the
 // application does not understand, so typos fail at startup instead of
 // being silently ignored, and rejects keys that are explicitly set to

internal/config/config_validation_internal_test.go

@@ -135,6 +135,104 @@ func TestDefaultDBURLOpensTheDatabaseInWALMode(t *testing.T) {
 	}
 }
 
+// TestDerivedDBURLOpensStateDirHoldingURLCharacters opens the db_url
+// derived from a state_dir whose name holds %41, ? and #, and checks that
+// the database is state.sqlite3 in that directory, in WAL mode. Put into
+// a file: URL as it is, the name would lose its meaning: %41 is read as
+// A, and the path ends at the ?, which also loses the journal mode.
+func TestDerivedDBURLOpensStateDirHoldingURLCharacters(t *testing.T) {
+	t.Parallel()
+
+	stateDir := filepath.Join(t.TempDir(), "a%41b?c#d")
+
+	err := os.Mkdir(stateDir, 0o750)
+	if err != nil {
+		t.Fatalf("failed to create %q: %v", stateDir, err)
+	}
+
+	// Quoted so YAML takes the whole name as it is.
+	c, err := configFromYAML(t, signingKeyLine+"state_dir: '"+stateDir+"'\n")
+	if err != nil {
+		t.Fatalf("config with state_dir %q should be valid, got: %v", stateDir, err)
+	}
+
+	t.Logf("state_dir %q gives db_url %q", stateDir, c.DBURL)
+
+	checkDBURLOpensFileInWALMode(t, c.DBURL, filepath.Join(stateDir, "state.sqlite3"))
+}
+
+// TestDerivedDBURLOpensStateDirStartingWithTwoSlashes opens the db_url
+// derived from a state_dir that starts with //, and checks that the
+// database is state.sqlite3 in that directory, in WAL mode. Written right
+// after file:, such a name gives a URL starting file://, and SQLite reads
+// the first directory name as a host and refuses to open it.
+func TestDerivedDBURLOpensStateDirStartingWithTwoSlashes(t *testing.T) {
+	t.Parallel()
+
+	dir := t.TempDir()
+	stateDir := "/" + dir
+
+	c, err := configFromYAML(t, signingKeyLine+"state_dir: "+stateDir+"\n")
+	if err != nil {
+		t.Fatalf("config with state_dir %q should be valid, got: %v", stateDir, err)
+	}
+
+	t.Logf("state_dir %q gives db_url %q", stateDir, c.DBURL)
+
+	checkDBURLOpensFileInWALMode(t, c.DBURL, filepath.Join(dir, "state.sqlite3"))
+}
+
+// TestDerivedDBURLForRootStateDir checks the db_url derived from a
+// state_dir of /. The path written into the URL is then
+// //state.sqlite3, and written right after file: it gives a URL starting
+// file://, where SQLite reads state.sqlite3 as a host and refuses to open
+// it. The test only checks the URL, since it cannot write to /.
+func TestDerivedDBURLForRootStateDir(t *testing.T) {
+	t.Parallel()
+
+	c, err := configFromYAML(t, signingKeyLine+"state_dir: /\n")
+	if err != nil {
+		t.Fatalf("config with state_dir / should be valid, got: %v", err)
+	}
+
+	t.Logf("state_dir / gives db_url %q", c.DBURL)
+
+	wantDBURL := "file:////state.sqlite3?_pragma=journal_mode(WAL)"
+	if c.DBURL != wantDBURL {
+		t.Errorf("DBURL = %q, want %q", c.DBURL, wantDBURL)
+	}
+}
+
+// checkDBURLOpensFileInWALMode opens dbURL with the SQLite driver pixad
+// uses, and checks that the database is in WAL mode and is the file
+// dbFile.
+func checkDBURLOpensFileInWALMode(t *testing.T, dbURL, dbFile string) {
+	t.Helper()
+
+	db, err := sql.Open("sqlite", dbURL)
+	if err != nil {
+		t.Fatalf("failed to open %q: %v", dbURL, err)
+	}
+
+	t.Cleanup(func() { _ = db.Close() })
+
+	var journalMode string
+
+	err = db.QueryRowContext(t.Context(), "PRAGMA journal_mode").Scan(&journalMode)
+	if err != nil {
+		t.Fatalf("failed to read the journal mode of %q: %v", dbURL, err)
+	}
+
+	if journalMode != "wal" {
+		t.Errorf("journal mode of %q = %q, want wal", dbURL, journalMode)
+	}
+
+	_, err = os.Stat(dbFile)
+	if err != nil {
+		t.Errorf("opening %q did not create %q: %v", dbURL, dbFile, err)
+	}
+}
+
 func TestExplicitValidValuesAreUsed(t *testing.T) {
 	t.Parallel()
 

internal/handlers/auth.go

@@ -9,7 +9,9 @@ import (
 	"math"
 	"net/http"
 	"net/url"
+	"slices"
 	"strconv"
+	"strings"
 	"time"
 
 	"sneak.berlin/go/pixa/internal/clientip"
@@ -125,7 +127,7 @@ func (s *Handlers) HandleGenerateURL() http.HandlerFunc {
 		sourceURL := r.FormValue("url")
 
 		// Validate source URL
-		parsed, err := url.Parse(sourceURL)
+		parsed, err := parseSourceURL(sourceURL)
 		if err != nil || parsed.Host == "" {
 			s.renderGeneratorWithForm(w, r, "Invalid source URL", r.Form,
 				http.StatusBadRequest)
@@ -133,6 +135,16 @@ func (s *Handlers) HandleGenerateURL() http.HandlerFunc {
 			return
 		}
 
+		// /v1/image/ refuses a ".." segment in the source path, and so does
+		// the generator. The path is decoded, so %2e%2e counts too
+		if slices.Contains(strings.Split(parsed.Path, "/"), "..") {
+			s.renderGeneratorWithForm(w, r,
+				"Source URL path cannot hold a .. segment", r.Form,
+				http.StatusBadRequest)
+
+			return
+		}
+
 		payload, expiresAt, ttl, err := buildGeneratePayload(parsed, r.Form)
 		if err != nil {
 			s.renderGeneratorWithForm(w, r, err.Error(), r.Form,
@@ -141,6 +153,19 @@ func (s *Handlers) HandleGenerateURL() http.HandlerFunc {
 			return
 		}
 
+		// The image URL must fetch the source URL as it was written
+		sourceRequest := payload.ToImageRequest()
+		sourceRequest.AllowHTTP = s.config.AllowHTTP
+		fetchURL := sourceRequest.SourceURL()
+
+		if !fetchedAsWritten(parsed, fetchURL) {
+			s.renderGeneratorWithForm(w, r, "Source URL cannot be fetched as "+
+				"written: pixa would fetch "+fetchURL,
+				r.Form, http.StatusBadRequest)
+
+			return
+		}
+
 		// Generate encrypted token
 		token, err := s.encGen.Generate(payload)
 		if err != nil {
@@ -173,6 +198,48 @@ func (s *Handlers) HandleGenerateURL() http.HandlerFunc {
 	}
 }
 
+// parseSourceURL parses the source URL entered on the generator page with its
+// host lower-cased and an empty path written as "/", as /v1/image/ reads a
+// source URL, so that each spelling of one source URL makes one cache key and
+// one upstream request.
+func parseSourceURL(sourceURL string) (*url.URL, error) {
+	parsed, err := url.Parse(sourceURL)
+	if err != nil {
+		return nil, err
+	}
+
+	parsed.Host = strings.ToLower(parsed.Host)
+
+	if parsed.Path == "" {
+		parsed.Path = "/"
+	}
+
+	return parsed, nil
+}
+
+// fetchedAsWritten reports whether fetchURL, the URL pixa fetches for an
+// image URL made from the source URL source, is source as it was written,
+// apart from the scheme, which the token does not keep. It is not when source
+// holds a user name and password, which the token does not keep either, or a
+// "?" with no query after it. The token keeps the source path decoded and
+// pixa fetches it percent-encoded again, so a path written another way, such
+// as with %2F, is fetched as another path. url.URL keeps the path as written
+// in RawPath when it is not the decoded path percent-encoded again.
+// EscapedPath ignores RawPath when it also holds a character net/url encodes,
+// such as a space, so RawPath is compared itself.
+func fetchedAsWritten(source *url.URL, fetchURL string) bool {
+	writtenPath := source.RawPath
+	if writtenPath == "" {
+		writtenPath = source.EscapedPath()
+	}
+
+	fetched, err := url.Parse(fetchURL)
+
+	return err == nil && source.User == nil && !source.ForceQuery &&
+		fetched.EscapedPath() == writtenPath &&
+		fetched.RawQuery == source.RawQuery
+}
+
 // buildGeneratePayload parses the numeric form fields and assembles the
 // encrypted URL payload. ttl=0 means never expires (ExpiresAt stays 0). A
 // non-numeric or out-of-range width, height, quality or ttl, or an

internal/handlers/avif_too_large_internal_test.go

@@ -1,49 +0,0 @@
-package handlers
-
-import (
-	"fmt"
-	"log/slog"
-	"net/http"
-	"net/http/httptest"
-	"strings"
-	"testing"
-
-	"sneak.berlin/go/pixa/internal/imageprocessor"
-	"sneak.berlin/go/pixa/internal/imgcache"
-)
-
-// TestAVIFOutputTooLargeAnswers400 checks that respondImageError and
-// handleImageError, which answer the errors of the two image routes, answer
-// 400 naming the AVIF pixel limit when the image processor refuses an AVIF
-// output above it, wrapped as the service wraps it. The request's source only
-// goes into a log line, so it is left empty.
-func TestAVIFOutputTooLargeAnswers400(t *testing.T) {
-	t.Parallel()
-
-	h := &Handlers{log: slog.New(slog.DiscardHandler)}
-	err := fmt.Errorf("image processing failed: %w: 4097x4096",
-		imageprocessor.ErrAVIFOutputTooLarge)
-	wantBody := fmt.Sprintf("above %d pixels", imageprocessor.MaxAVIFOutputPixels)
-
-	plain := httptest.NewRecorder()
-	h.respondImageError(plain, &imgcache.ImageRequest{}, err)
-
-	encrypted := httptest.NewRecorder()
-	h.handleImageError(encrypted, err)
-
-	for name, rec := range map[string]*httptest.ResponseRecorder{
-		"respondImageError": plain, "handleImageError": encrypted,
-	} {
-		t.Logf("%s: status = %d, body = %q", name, rec.Code, rec.Body.String())
-
-		if rec.Code != http.StatusBadRequest {
-			t.Errorf("%s: status = %d, want %d",
-				name, rec.Code, http.StatusBadRequest)
-		}
-
-		if !strings.Contains(rec.Body.String(), wantBody) {
-			t.Errorf("%s: body = %q, want it to say %q",
-				name, rec.Body.String(), wantBody)
-		}
-	}
-}

internal/handlers/image.go

@@ -264,14 +264,6 @@ func (s *Handlers) respondImageError(
 		return
 	}
 
-	if errors.Is(err, imageprocessor.ErrAVIFOutputTooLarge) {
-		s.respondError(w, fmt.Sprintf("size too large for AVIF: the output "+
-			"would be above %d pixels", imageprocessor.MaxAVIFOutputPixels),
-			http.StatusBadRequest)
-
-		return
-	}
-
 	s.respondError(w, "internal error", http.StatusInternalServerError)
 }
 

internal/handlers/image_errors_internal_test.go

@@ -264,3 +264,127 @@ func TestHandleImage_AllowlistOrSignature(t *testing.T) {
 		})
 	}
 }
+
+// TestHandleImage_InvalidHost_Returns400 sends image URLs whose host no URL
+// host can be, and checks that each is refused with 400 naming the host as
+// decoded: allowlistedHost with a character after it that a URL host cannot
+// hold, percent-encoded, an IPv6 address without brackets, and brackets
+// around something that is not an IPv6 address. A "?" or "#" after
+// allowlistedHost would otherwise end the host when the source URL is read,
+// so the request would pass the allowlist and be fetched and cached under a
+// host that is not the one fetched.
+func TestHandleImage_InvalidHost_Returns400(t *testing.T) {
+	t.Parallel()
+
+	photos := newPhotoFetcher(t, allowlistedHost)
+
+	tests := []struct {
+		name        string
+		writtenHost string
+		decodedHost string
+	}{
+		{"? written as %3F", allowlistedHost + "%3Fx=1", allowlistedHost + "?x=1"},
+		{"# written as %23", allowlistedHost + "%23x", allowlistedHost + "#x"},
+		{"% written as %25", allowlistedHost + "%253F", allowlistedHost + "%3F"},
+		{`\ written as %5C`, allowlistedHost + "%5Cx", allowlistedHost + `\x`},
+		{"space written as %20", allowlistedHost + "%20", allowlistedHost + " "},
+		{"IPv6 address without brackets", "2001:db8::1", "2001:db8::1"},
+		{"IPv4 address in brackets", "[203.0.113.5]", "[203.0.113.5]"},
+		{"no address in brackets", "[abc]", "[abc]"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			rec := sendGet(t, newImageRoute(t, photos), photoURL(tt.writtenHost))
+			checkErrorBody(t, rec, http.StatusBadRequest,
+				fmt.Sprintf("invalid image URL: invalid source host: %q",
+					tt.decodedHost))
+		})
+	}
+}
+
+// TestHandleImage_HostForms_Served checks that the image route serves, with
+// a valid signature, the image at photoPath on a host name, an IPv4 address,
+// an IPv6 address in brackets, and a host followed by ":" with or without a
+// port.
+func TestHandleImage_HostForms_Served(t *testing.T) {
+	t.Parallel()
+
+	hosts := []string{
+		allowlistedHost,
+		"img_1.example.com",
+		"203.0.113.5",
+		"[2001:db8::1]",
+		allowlistedHost + ":8443",
+		"[2001:db8::1]:8443",
+		allowlistedHost + ":",
+	}
+
+	photos := newPhotoFetcher(t, hosts...)
+	exp := time.Now().Add(time.Hour)
+
+	for _, host := range hosts {
+		t.Run(host, func(t *testing.T) {
+			t.Parallel()
+
+			sig := photoSignature(testSigningKey, host, exp)
+
+			rec := sendGet(t, newImageRoute(t, photos),
+				photoURLWithSig(host, sig, exp))
+			if rec.Code != http.StatusOK {
+				t.Errorf("status = %d, want %d: %s",
+					rec.Code, http.StatusOK, rec.Body)
+			}
+		})
+	}
+}
+
+// TestHandleImage_HashInSourcePathOrQuery_Returns400 sends image URLs whose
+// source path or source query holds a "#", written as %23, and checks that
+// each is refused with 400, the one signed for that source path included. The
+// source URL would otherwise end at the "#" when it is fetched, so each text
+// after it would fetch the image at photoPath and cache it under keys of its
+// own.
+func TestHandleImage_HashInSourcePathOrQuery_Returns400(t *testing.T) {
+	t.Parallel()
+
+	photos := newPhotoFetcher(t, allowlistedHost, signedHost)
+	exp := time.Now().Add(time.Hour)
+
+	sig := signature.New(testSigningKey).Sign(&signature.Request{
+		SourceHost: signedHost,
+		SourcePath: photoPath + "#x",
+		Width:      50,
+		Height:     50,
+		Format:     string(imgcache.FormatJPEG),
+		Quality:    85,
+		FitMode:    string(imgcache.FitCover),
+		Expires:    exp,
+	})
+
+	tests := []struct {
+		name   string
+		target string
+	}{
+		{"# in the source path",
+			"/v1/image/" + allowlistedHost + photoPath + "%23x/50x50.jpeg"},
+		{"# in the source query",
+			"/v1/image/" + allowlistedHost + photoPath + "%3Fv=1%23x/50x50.jpeg"},
+		{"# in the source path, signed",
+			fmt.Sprintf("/v1/image/%s%s%%23x/50x50.jpeg?sig=%s&exp=%d",
+				signedHost, photoPath, sig, exp.Unix())},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			rec := sendGet(t, newImageRoute(t, photos), tt.target)
+			checkErrorBody(t, rec, http.StatusBadRequest,
+				`invalid image URL: invalid image path: `+
+					`the source path or query holds a "#"`)
+		})
+	}
+}

internal/handlers/imageenc.go

@@ -170,10 +170,6 @@ func (s *Handlers) handleImageError(w http.ResponseWriter, err error) {
 		s.respondError(w, fmt.Sprintf("size too large for the source: the side "+
 			"computed from its proportions is above %d",
 			imageprocessor.MaxOutputDimension), http.StatusBadRequest)
-	case errors.Is(err, imageprocessor.ErrAVIFOutputTooLarge):
-		s.respondError(w, fmt.Sprintf("size too large for AVIF: the output "+
-			"would be above %d pixels", imageprocessor.MaxAVIFOutputPixels),
-			http.StatusBadRequest)
 	default:
 		s.log.Error("image request failed", "error", err)
 		s.respondError(w, "internal error", http.StatusInternalServerError)

internal/httpfetcher/httpfetcher.go

@@ -59,8 +59,9 @@ const (
 // builtinBlockedPrefixes are internal or special-use ranges that Go's
 // net.IP predicates (IsPrivate, IsLinkLocalUnicast, and the like) do not
 // already cover. They are always blocked, in addition to any
-// operator-supplied networks. IPv4-mapped IPv6 addresses are unmapped
+// operator-supplied networks. IPv4-mapped and IPv4-compatible IPv6
-// before matching, so these IPv4 ranges are caught in both forms.
+// addresses are read as the IPv4 address they hold before matching, so
+// these IPv4 ranges are caught in all three forms.
 //
 //nolint:gochecknoglobals // immutable built-in blocklist
 var builtinBlockedPrefixes = []netip.Prefix{
@@ -68,6 +69,7 @@ var builtinBlockedPrefixes = []netip.Prefix{
 	netip.MustParsePrefix("192.0.0.0/24"),   // RFC 6890 IETF protocol assignments
 	netip.MustParsePrefix("198.18.0.0/15"),  // RFC 2544 benchmarking range
 	netip.MustParsePrefix("64:ff9b::/96"),   // RFC 6052 NAT64 (maps onto IPv4)
+	netip.MustParsePrefix("64:ff9b:1::/48"), // RFC 8215 local-use NAT64 (maps onto IPv4)
 }
 
 // Fetcher errors.
@@ -473,16 +475,16 @@ func validateURL(ctx context.Context, rawURL string, allowHTTP bool) error {
 		return ErrUnsupportedScheme
 	}
 
-	// Parse to extract host
+	parsedURL, err := neturl.Parse(rawURL)
-	host := extractHost(rawURL)
+	if err != nil {
-	if host == "" {
+		return fmt.Errorf("%w: %w", ErrInvalidHost, err)
-		return ErrInvalidHost
 	}
 
-	// Remove port if present
+	// Hostname drops any port, and the brackets of an IPv6 address whether
-	h, _, err := net.SplitHostPort(host)
+	// or not a port follows them
-	if err == nil {
+	host := parsedURL.Hostname()
-		host = h
+	if host == "" {
+		return ErrInvalidHost
 	}
 
 	// Block obvious localhost patterns
@@ -543,6 +545,9 @@ func isPrivateIP(ip net.IP) bool {
 		return true
 	}
 
+	// An IPv4-compatible IPv6 address is judged as the IPv4 address it holds
+	ip = ipv4CompatibleToIPv4(ip)
+
 	// Check for loopback
 	if ip.IsLoopback() {
 		return true
@@ -594,7 +599,12 @@ func isPrivateIP(ip net.IP) bool {
 }
 
 // isBlockedIP reports whether ip is refused, either by the built-in
-// internal-range check or by one of the operator-supplied prefixes.
+// internal-range check or by one of the operator-supplied prefixes. An
+// IPv4-compatible address (in ::/96) is refused when a prefix contains the
+// address as written or the IPv4 address it holds. An IPv4-mapped address (in
+// ::ffff:0:0/96) is matched only as the IPv4 address it holds: Go keeps a
+// plain IPv4 address in that form too, and an IPv6 prefix such as ::/0 must
+// not refuse every IPv4 address.
 func isBlockedIP(ip net.IP, blocked []netip.Prefix) bool {
 	if isPrivateIP(ip) {
 		return true
@@ -607,11 +617,28 @@ func isBlockedIP(ip net.IP, blocked []netip.Prefix) bool {
 
 	addr = addr.Unmap()
 
+	// ip read as IPv4 where it is or holds an IPv4 address, otherwise the
+	// zero netip.Addr, which no prefix contains. Matching asIPv4 as well as
+	// addr refuses more only for an IPv4-compatible address.
+	asIPv4, _ := netip.AddrFromSlice(ipv4CompatibleToIPv4(ip).To4())
+
 	return slices.ContainsFunc(blocked, func(prefix netip.Prefix) bool {
-		return prefix.Contains(addr)
+		return prefix.Contains(addr) || prefix.Contains(asIPv4)
 	})
 }
 
+// ipv4CompatibleToIPv4 returns the IPv4 address that an IPv4-compatible IPv6
+// address (in ::/96, such as ::127.0.0.1) holds in its last four bytes, and
+// any other address unchanged.
+func ipv4CompatibleToIPv4(ip net.IP) net.IP {
+	addr, ok := netip.AddrFromSlice(ip)
+	if ok && netip.MustParsePrefix("::/96").Contains(addr) {
+		return ip[12:]
+	}
+
+	return ip
+}
+
 // ssrfSafeDialer validates IP addresses against the built-in blocked ranges
 // before connecting. New wraps dialSSRFSafe with the operator-supplied
 // blocklist; this entry point enforces the built-in ranges alone.

internal/httpfetcher/httpfetcher_internal_test.go

@@ -186,6 +186,44 @@ func TestIsPrivateIP(t *testing.T) {
 	}
 }
 
+// TestIsPrivateIPJudgesEmbeddedIPv4Address checks IPv6 addresses that hold an
+// IPv4 address in their last four bytes: the IPv4-compatible form (::/96) and
+// the NAT64 form (64:ff9b::/96). One holding a loopback, private or
+// link-local IPv4 address is refused as that address is, and one holding a
+// public IPv4 address in the IPv4-compatible form passes.
+func TestIsPrivateIPJudgesEmbeddedIPv4Address(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		ip   string
+		want bool
+	}{
+		{"::127.0.0.1", true},       // loopback
+		{"::10.0.0.1", true},        // private
+		{"::169.254.169.254", true}, // link-local
+		{"64:ff9b::7f00:1", true},   // NAT64 form of 127.0.0.1
+		{"::8.8.8.8", false},        // public
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.ip, func(t *testing.T) {
+			t.Parallel()
+
+			ip := net.ParseIP(tc.ip)
+			if ip == nil {
+				t.Fatalf("failed to parse IP %q", tc.ip)
+			}
+
+			got := isPrivateIP(ip)
+			t.Logf("isPrivateIP(%s) = %v", ip, got)
+
+			if got != tc.want {
+				t.Errorf("isPrivateIP(%q) = %v, want %v", tc.ip, got, tc.want)
+			}
+		})
+	}
+}
+
 func TestValidateURL_RejectsNonHTTPS(t *testing.T) {
 	t.Parallel()
 

internal/imageprocessor/avif_too_large_internal_test.go

@@ -1,64 +0,0 @@
-package imageprocessor
-
-import (
-	"bytes"
-	"errors"
-	"testing"
-)
-
-// TestImageProcessor_RefusesAVIFAboveMaxPixels verifies that an AVIF output
-// of more than MaxAVIFOutputPixels is refused with ErrAVIFOutputTooLarge
-// before anything is decoded, whatever size and fit give it, and that an AVIF
-// output of at most that many pixels, or one in another format, is not. The
-// input is a 64x32 PNG cut off in its pixel data: libvips reads its size from
-// its header but fails when it decodes it, so a request that gets that failure
-// was not refused, and a refusal came before any decoding.
-func TestImageProcessor_RefusesAVIFAboveMaxPixels(t *testing.T) {
-	t.Parallel()
-
-	whole := createTestPNG(t, 64, 32)
-	cutOff := whole[:len(whole)/2]
-
-	tests := []struct {
-		name          string
-		format        Format
-		width, height int
-		fit           FitMode
-		refused       bool
-	}{
-		{"AVIF at the limit", FormatAVIF, 4096, 4096, FitCover, false},
-		{"AVIF above the limit", FormatAVIF, 4096, 4097, FitCover, true},
-		{"WebP above the AVIF limit", FormatWebP, 4096, 4097, FitCover, false},
-		// FitInside does not enlarge, so the output stays 64x32
-		{"AVIF inside a larger size", FormatAVIF, 8192, 8192, FitInside, false},
-		// FitOutside scales 64x32 by 128 to cover 8192x1: 8192x4096
-		{"AVIF outside a smaller size", FormatAVIF, 8192, 1, FitOutside, true},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			t.Parallel()
-
-			_, err := New(Params{}).Process(t.Context(), bytes.NewReader(cutOff),
-				&Request{
-					Size:    Size{Width: tt.width, Height: tt.height},
-					Format:  tt.format,
-					FitMode: tt.fit,
-				})
-			t.Logf("Process() error = %v", err)
-
-			if tt.refused {
-				if !errors.Is(err, ErrAVIFOutputTooLarge) {
-					t.Errorf("Process() error = %v, want %v", err, ErrAVIFOutputTooLarge)
-				}
-
-				return
-			}
-
-			if err == nil || errors.Is(err, ErrAVIFOutputTooLarge) {
-				t.Errorf("Process() error = %v, want the failure to decode "+
-					"the cut-off input", err)
-			}
-		})
-	}
-}

internal/imageprocessor/computed_side_internal_test.go

@@ -50,3 +50,69 @@ func TestImageProcessor_RefusesComputedSideAboveMaximum(t *testing.T) {
 		})
 	}
 }
+
+// TestImageProcessor_ComputedSideIsAtLeastOne verifies that, for a size with
+// one side 0, the side computed from the input's proportions is 1 where it
+// would round down to 0, and that Process makes that size in every fit mode
+// but FitOutside, which scales the input to cover the size and so can make
+// it longer. Each case checks the size targetDimensions gives before it
+// calls Process: with that side at 0, an 8192x1 input at 1300x0 makes
+// libvips ask for about 39 GB with FitCover.
+func TestImageProcessor_ComputedSideIsAtLeastOne(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name                    string
+		inputWidth, inputHeight int
+		size                    Size
+		wantWidth, wantHeight   int
+	}{
+		{"8192x1 at 100x0", 8192, 1, Size{Width: 100}, 100, 1},
+		{"8192x1 at 1300x0", 8192, 1, Size{Width: 1300}, 1300, 1},
+		{"8192x2 at 100x0", 8192, 2, Size{Width: 100}, 100, 1},
+		{"1x8192 at 0x100", 1, 8192, Size{Height: 100}, 1, 100},
+		{"1x8192 at 0x1300", 1, 8192, Size{Height: 1300}, 1, 1300},
+		{"2x8192 at 0x100", 2, 8192, Size{Height: 100}, 1, 100},
+	}
+
+	fits := []FitMode{FitCover, FitContain, FitFill, FitInside, FitOutside}
+
+	for _, tt := range tests {
+		for _, fit := range fits {
+			t.Run(tt.name+" "+string(fit), func(t *testing.T) {
+				t.Parallel()
+
+				width, height, err := targetDimensions(tt.size,
+					tt.inputWidth, tt.inputHeight)
+				if err != nil || width != tt.wantWidth || height != tt.wantHeight {
+					t.Fatalf("targetDimensions() = %dx%d, %v, want %dx%d",
+						width, height, err, tt.wantWidth, tt.wantHeight)
+				}
+
+				input := createTestJPEG(t, tt.inputWidth, tt.inputHeight)
+
+				result, err := New(Params{}).Process(t.Context(),
+					bytes.NewReader(input),
+					&Request{Size: tt.size, Format: FormatJPEG, FitMode: fit})
+				if err != nil {
+					t.Fatalf("Process() error = %v", err)
+				}
+
+				defer func() { _ = result.Content.Close() }()
+
+				t.Logf("Process() made %dx%d", result.Width, result.Height)
+
+				ok := result.Width == tt.wantWidth && result.Height == tt.wantHeight
+				if fit == FitOutside {
+					ok = result.Width >= tt.wantWidth && result.Height >= tt.wantHeight
+				}
+
+				if !ok {
+					t.Errorf("Process() made %dx%d, want %dx%d (at least, for %s)",
+						result.Width, result.Height, tt.wantWidth, tt.wantHeight,
+						FitOutside)
+				}
+			})
+		}
+	}
+}

internal/imageprocessor/encoder_settings_internal_test.go

@@ -77,10 +77,10 @@ func TestImageProcessor_PNGIsCompressed(t *testing.T) {
 	}
 }
 
-// TestImageProcessor_WebPAtEffort2 verifies that WebP is saved at effort 2,
+// TestImageProcessor_WebPAtDefaultEffort verifies that WebP is saved at
-// not libvips' default of 4: the output is the size of the same image saved
+// libvips' default effort, 4: the output is the size of the same image saved
-// at effort 2.
+// at that effort.
-func TestImageProcessor_WebPAtEffort2(t *testing.T) {
+func TestImageProcessor_WebPAtDefaultEffort(t *testing.T) {
 	t.Parallel()
 
 	input := createTestJPEG(t, 200, 150)
@@ -90,14 +90,14 @@ func TestImageProcessor_WebPAtEffort2(t *testing.T) {
 	want, _, err := decode(t, input).ExportWebp(&vips.WebpExportParams{
 		StripMetadata:   true,
 		Quality:         85,
-		ReductionEffort: 2,
+		ReductionEffort: 4,
 	})
 	if err != nil {
 		t.Fatalf("ExportWebp() error = %v", err)
 	}
 
 	if size != int64(len(want)) {
-		t.Errorf("WebP output is %d bytes, want %d, the size at effort 2",
+		t.Errorf("WebP output is %d bytes, want %d, the size at effort 4",
 			size, len(want))
 	}
 }
@@ -143,29 +143,3 @@ func TestImageProcessor_AVIFAtEffort1(t *testing.T) {
 			"and 8 bits", size, len(want))
 	}
 }
-
-// TestImageProcessor_JPEGXLAtEffort4 verifies that JPEG XL is saved at
-// effort 4, not libvips' default of 7: the output is the size of the same
-// image saved at effort 4, with the distance exportJXL gives quality 85.
-func TestImageProcessor_JPEGXLAtEffort4(t *testing.T) {
-	t.Parallel()
-
-	input := createTestJPEG(t, 200, 150)
-
-	size := processedSize(t, input, &Request{Format: FormatJXL, Quality: 85})
-
-	params := vips.NewJxlExportParams()
-	params.Quality = 85
-	params.Distance = jxlDistance(85)
-	params.Effort = 4
-
-	want, _, err := decode(t, input).ExportJxl(params)
-	if err != nil {
-		t.Fatalf("ExportJxl() error = %v", err)
-	}
-
-	if size != int64(len(want)) {
-		t.Errorf("JPEG XL output is %d bytes, want %d, the size at effort 4",
-			size, len(want))
-	}
-}

internal/imageprocessor/fit_cover_internal_test.go

@@ -0,0 +1,268 @@
+package imageprocessor
+
+import (
+	"bytes"
+	"image"
+	"image/color"
+	"image/png"
+	"math"
+	"testing"
+)
+
+// TestImageProcessor_FitCoverServesThinSource verifies that FitCover serves
+// a thin source at the requested size. An 8192x1 source scaled to cover
+// 8192x8192 is 8192 times as long as the output on its long side. The centre
+// of an 8192x1 source with the proportions of 1x8192 is 1/8192 of a pixel
+// wide. Each case is also turned on its side.
+func TestImageProcessor_FitCoverServesThinSource(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name                    string
+		inputWidth, inputHeight int
+		size                    Size
+	}{
+		{"8192x1 at 8192x8192", 8192, 1, Size{Width: 8192, Height: 8192}},
+		{"8192x1 at 1x8192", 8192, 1, Size{Width: 1, Height: 8192}},
+		{"1x8192 at 8192x8192", 1, 8192, Size{Width: 8192, Height: 8192}},
+		{"1x8192 at 8192x1", 1, 8192, Size{Width: 8192, Height: 1}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			input := createTestJPEG(t, tt.inputWidth, tt.inputHeight)
+
+			result, err := New(Params{}).Process(t.Context(),
+				bytes.NewReader(input),
+				&Request{Size: tt.size, Format: FormatJPEG, FitMode: FitCover})
+			if err != nil {
+				t.Fatalf("Process() error = %v", err)
+			}
+
+			defer func() { _ = result.Content.Close() }()
+
+			t.Logf("Process() made %dx%d", result.Width, result.Height)
+
+			if result.Width != tt.size.Width || result.Height != tt.size.Height {
+				t.Errorf("Process() made %dx%d, want %dx%d",
+					result.Width, result.Height, tt.size.Width, tt.size.Height)
+			}
+		})
+	}
+}
+
+// createBandsPNG creates a PNG of three equal bands, red, green and blue,
+// side by side when it is wider than tall and top to bottom otherwise.
+func createBandsPNG(t *testing.T, width, height int) []byte {
+	t.Helper()
+
+	bands := []color.RGBA{
+		{R: 255, A: 255},
+		{G: 255, A: 255},
+		{B: 255, A: 255},
+	}
+
+	img := image.NewRGBA(image.Rect(0, 0, width, height))
+	for y := range height {
+		for x := range width {
+			band := 3 * y / height
+			if width > height {
+				band = 3 * x / width
+			}
+
+			img.Set(x, y, bands[band])
+		}
+	}
+
+	return encodePNG(t, img)
+}
+
+// TestImageProcessor_FitCoverCropsTheCentre verifies that FitCover keeps the
+// centre of the source: a 300x100 source of red, green and blue bands side
+// by side, or a 100x300 one with them top to bottom, comes out at 50x50 as
+// its green band.
+func TestImageProcessor_FitCoverCropsTheCentre(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name                    string
+		inputWidth, inputHeight int
+	}{
+		{"bands side by side", 300, 100},
+		{"bands top to bottom", 100, 300},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			input := createBandsPNG(t, tt.inputWidth, tt.inputHeight)
+
+			result, err := New(Params{}).Process(t.Context(),
+				bytes.NewReader(input),
+				&Request{
+					Size:    Size{Width: 50, Height: 50},
+					Format:  FormatPNG,
+					FitMode: FitCover,
+				})
+			if err != nil {
+				t.Fatalf("Process() error = %v", err)
+			}
+
+			defer func() { _ = result.Content.Close() }()
+
+			output, err := png.Decode(result.Content)
+			if err != nil {
+				t.Fatalf("failed to decode output: %v", err)
+			}
+
+			bounds := output.Bounds()
+			t.Logf("Process() made %dx%d", bounds.Dx(), bounds.Dy())
+
+			if bounds.Dx() != 50 || bounds.Dy() != 50 {
+				t.Fatalf("Process() made %dx%d, want 50x50",
+					bounds.Dx(), bounds.Dy())
+			}
+
+			// Resampling may blend a little of the bands beside the
+			// green one into the output's edges.
+			for y := bounds.Min.Y; y < bounds.Max.Y; y++ {
+				for x := bounds.Min.X; x < bounds.Max.X; x++ {
+					r, g, b, _ := output.At(x, y).RGBA()
+					if r>>8 > 64 || g>>8 < 192 || b>>8 > 64 {
+						t.Fatalf("output pixel %d,%d is %d,%d,%d, want green",
+							x, y, r>>8, g>>8, b>>8)
+					}
+				}
+			}
+		})
+	}
+}
+
+// The greys of createSquarePNG, and the grey halfway between them.
+const (
+	backgroundGrey = 192
+	squareGrey     = 64
+	halfwayGrey    = (backgroundGrey + squareGrey) / 2
+)
+
+// createSquarePNG creates a grey PNG side pixels square with a darker grey
+// square squareSide pixels wide in its centre.
+func createSquarePNG(t *testing.T, side, squareSide int) []byte {
+	t.Helper()
+
+	first := (side - squareSide) / 2
+	last := first + squareSide - 1
+
+	img := image.NewGray(image.Rect(0, 0, side, side))
+	for y := range side {
+		for x := range side {
+			grey := uint8(backgroundGrey)
+			if x >= first && x <= last && y >= first && y <= last {
+				grey = squareGrey
+			}
+
+			img.SetGray(x, y, color.Gray{Y: grey})
+		}
+	}
+
+	return encodePNG(t, img)
+}
+
+// squareLength returns how long the square is along greys, the greys of a
+// row or column of the output that crosses it: the distance between the
+// first and the last point where greys passes halfwayGrey, or 0 when there
+// is no such point. Resampling blurs the square's edges over several
+// pixels, and each point is placed between the two pixels on either side
+// of it in proportion to their greys, so the length is measured to a
+// fraction of a pixel.
+func squareLength(greys []float64) float64 {
+	first, last := -1.0, -1.0
+
+	for i := 1; i < len(greys); i++ {
+		before, after := greys[i-1], greys[i]
+		if (before > halfwayGrey) == (after > halfwayGrey) {
+			continue
+		}
+
+		point := float64(i-1) + (before-halfwayGrey)/(before-after)
+		if first < 0 {
+			first = point
+		}
+
+		last = point
+	}
+
+	return last - first
+}
+
+// TestImageProcessor_FitCoverKeepsProportions verifies that FitCover scales
+// both sides of the source by the same amount: a square in the centre of
+// the source comes out square. Its width and height are measured along the
+// output's middle row and column.
+func TestImageProcessor_FitCoverKeepsProportions(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name             string
+		side, squareSide int
+		size             Size
+	}{
+		{"16x16 at 100x60", 16, 8, Size{Width: 100, Height: 60}},
+		{"16x16 at 60x100", 16, 8, Size{Width: 60, Height: 100}},
+		{"10x10 at 100x35", 10, 2, Size{Width: 100, Height: 35}},
+		{"10x10 at 35x100", 10, 2, Size{Width: 35, Height: 100}},
+		{"64x64 at 300x100", 64, 16, Size{Width: 300, Height: 100}},
+		{"64x64 at 100x300", 64, 16, Size{Width: 100, Height: 300}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			input := createSquarePNG(t, tt.side, tt.squareSide)
+
+			result, err := New(Params{}).Process(t.Context(),
+				bytes.NewReader(input),
+				&Request{Size: tt.size, Format: FormatPNG, FitMode: FitCover})
+			if err != nil {
+				t.Fatalf("Process() error = %v", err)
+			}
+
+			defer func() { _ = result.Content.Close() }()
+
+			output, err := png.Decode(result.Content)
+			if err != nil {
+				t.Fatalf("failed to decode output: %v", err)
+			}
+
+			bounds := output.Bounds()
+			if bounds.Dx() != tt.size.Width || bounds.Dy() != tt.size.Height {
+				t.Fatalf("Process() made %dx%d, want %dx%d",
+					bounds.Dx(), bounds.Dy(), tt.size.Width, tt.size.Height)
+			}
+
+			row := make([]float64, bounds.Dx())
+			for x := range row {
+				r, _, _, _ := output.At(x, bounds.Dy()/2).RGBA()
+				row[x] = float64(r >> 8)
+			}
+
+			column := make([]float64, bounds.Dy())
+			for y := range column {
+				r, _, _, _ := output.At(bounds.Dx()/2, y).RGBA()
+				column[y] = float64(r >> 8)
+			}
+
+			width, height := squareLength(row), squareLength(column)
+			t.Logf("square is %.2f wide and %.2f high", width, height)
+
+			if width == 0 || math.Abs(width-height) > 0.5 {
+				t.Errorf("square is %.2f wide and %.2f high, want it square",
+					width, height)
+			}
+		})
+	}
+}

internal/imageprocessor/fit_outside_internal_test.go

@@ -60,3 +60,45 @@ func TestImageProcessor_RefusesFitOutsideAboveMaximum(t *testing.T) {
 		})
 	}
 }
+
+// TestImageProcessor_FitOutsideSideIsAtLeastOne verifies that FitOutside
+// keeps each side at 1 or more when it works both sides out again from the
+// larger scale. A side of 49 scaled to 1 is 49 times 1/49, which in floating
+// point is just under 1 and so rounded down to 0, a width libvips refuses.
+func TestImageProcessor_FitOutsideSideIsAtLeastOne(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name                    string
+		inputWidth, inputHeight int
+		size                    Size
+		wantWidth, wantHeight   int
+	}{
+		{"49x8192 at 0x100", 49, 8192, Size{Height: 100}, 1, 167},
+		{"49x49 at 1x1", 49, 49, Size{Width: 1, Height: 1}, 1, 1},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			input := createTestJPEG(t, tt.inputWidth, tt.inputHeight)
+
+			result, err := New(Params{}).Process(t.Context(),
+				bytes.NewReader(input),
+				&Request{Size: tt.size, Format: FormatJPEG, FitMode: FitOutside})
+			if err != nil {
+				t.Fatalf("Process() error = %v", err)
+			}
+
+			defer func() { _ = result.Content.Close() }()
+
+			t.Logf("Process() made %dx%d", result.Width, result.Height)
+
+			if result.Width != tt.wantWidth || result.Height != tt.wantHeight {
+				t.Errorf("Process() made %dx%d, want %dx%d",
+					result.Width, result.Height, tt.wantWidth, tt.wantHeight)
+			}
+		})
+	}
+}

internal/imageprocessor/imageprocessor.go

@@ -142,19 +142,6 @@ const MaxOutputDimension = 8192
 // requested size.
 var ErrOutputTooLarge = errors.New("output image dimensions exceed maximum")
 
-// MaxAVIFOutputPixels is the most pixels an AVIF output may have, so that an
-// encode takes at most 40 seconds of the default downstream_timeout of 60
-// seconds; AVIF is already at the lowest effort govips can set (avifEffort).
-// With one thread, as pixad runs libvips, saving 4096x4096, 8192x2048 or
-// 2048x8192 random pixels with an alpha channel takes up to about 34
-// seconds, the slowest at any quality and at 8 or 16 bits per sample, and
-// 5120x4096 at 8 bits already takes about 36.
-const MaxAVIFOutputPixels = 4096 * 4096
-
-// ErrAVIFOutputTooLarge is returned, before anything is decoded, when an AVIF
-// output would have more than MaxAVIFOutputPixels pixels.
-var ErrAVIFOutputTooLarge = errors.New("AVIF output has too many pixels")
-
 // ErrUnsupportedOutputFormat is returned when the requested output format is
 // not supported.
 var ErrUnsupportedOutputFormat = errors.New("unsupported output format")
@@ -223,8 +210,7 @@ func New(params Params) *ImageProcessor {
 // ErrTooManyImages. When ctx has ended once Process has a slot, or before it
 // encodes, it returns ctx's error and gives the slot back. govips gives pixa
 // no way to stop libvips once an encode has begun, so an encode keeps its
-// slot until it finishes. An AVIF output of more than MaxAVIFOutputPixels is
+// slot until it finishes.
-// refused with ErrAVIFOutputTooLarge before anything is decoded.
 func (p *ImageProcessor) Process(
 	ctx context.Context,
 	input io.Reader,
@@ -237,9 +223,18 @@ func (p *ImageProcessor) Process(
 
 	defer release()
 
-	data, err := p.readInput(input)
+	// Read input with a size limit to prevent unbounded memory consumption.
+	// We read at most maxInputBytes+1 so we can detect if the input exceeds
+	// the limit without consuming additional memory.
+	limited := io.LimitReader(input, p.maxInputBytes+1)
+
+	data, err := io.ReadAll(limited)
 	if err != nil {
-		return nil, err
+		return nil, fmt.Errorf("failed to read input: %w", err)
+	}
+
+	if int64(len(data)) > p.maxInputBytes {
+		return nil, ErrInputDataTooLarge
 	}
 
 	// govips picks the libvips loader from the input's first bytes with
@@ -282,28 +277,20 @@ func (p *ImageProcessor) Process(
 		return nil, err
 	}
 
+	// Resize if needed
+	if targetWidth != origWidth || targetHeight != origHeight {
+		err := p.resize(img, targetWidth, targetHeight, req.FitMode)
+		if err != nil {
+			return nil, fmt.Errorf("failed to resize: %w", err)
+		}
+	}
+
 	// orig is the source's own format; encode refuses an empty format
 	outputFormat := req.Format
 	if outputFormat == FormatOriginal {
 		outputFormat = sourceFormat
 	}
 
-	// Determine output dimensions; too large ones are refused before
-	// resizing, which with FitCover already decodes the input
-	outputWidth, outputHeight, err := outputDimensions(origWidth, origHeight,
-		targetWidth, targetHeight, req.FitMode, outputFormat)
-	if err != nil {
-		return nil, err
-	}
-
-	// Resize if needed
-	if targetWidth != origWidth || targetHeight != origHeight {
-		err := p.resize(img, outputWidth, outputHeight, req.FitMode)
-		if err != nil {
-			return nil, fmt.Errorf("failed to resize: %w", err)
-		}
-	}
-
 	// Encode to target format
 	output, err := p.encode(ctx, img, outputFormat, req.Quality)
 	if err != nil {
@@ -324,8 +311,10 @@ func (p *ImageProcessor) Process(
 
 // targetDimensions calculates the output dimensions for a requested size,
 // scaling proportionally when only one dimension is given and keeping the
-// original dimensions when both are zero. A side scaled proportionally can
+// original dimensions when both are zero. A side scaled proportionally is
-// come out above MaxOutputDimension, which is ErrOutputTooLarge.
+// rounded down but kept at 1 or more, as for a thin image it would round
+// down to 0, a side of no pixels. It can come out above MaxOutputDimension,
+// which is ErrOutputTooLarge.
 func targetDimensions(size Size, origWidth, origHeight int) (int, int, error) {
 	width, height := size.Width, size.Height
 
@@ -335,10 +324,10 @@ func targetDimensions(size Size, origWidth, origHeight int) (int, int, error) {
 		return origWidth, origHeight, nil
 	case width == 0:
 		// Only height specified: calculate width proportionally
-		width = origWidth * height / origHeight
+		width = max(1, origWidth*height/origHeight)
 	case height == 0:
 		// Only width specified: calculate height proportionally
-		height = origHeight * width / origWidth
+		height = max(1, origHeight*width/origWidth)
 	}
 
 	if width > MaxOutputDimension || height > MaxOutputDimension {
@@ -456,23 +445,6 @@ func (p *ImageProcessor) acquireSlot(ctx context.Context) (func(), error) {
 	return release, nil
 }
 
-// readInput reads input with a size limit to prevent unbounded memory
-// consumption, refusing more than maxInputBytes with ErrInputDataTooLarge. It
-// reads at most maxInputBytes+1 so it can detect if the input exceeds the
-// limit without consuming additional memory.
-func (p *ImageProcessor) readInput(input io.Reader) ([]byte, error) {
-	data, err := io.ReadAll(io.LimitReader(input, p.maxInputBytes+1))
-	if err != nil {
-		return nil, fmt.Errorf("failed to read input: %w", err)
-	}
-
-	if int64(len(data)) > p.maxInputBytes {
-		return nil, ErrInputDataTooLarge
-	}
-
-	return data, nil
-}
-
 // detectFormat returns the format string for a govips image type.
 func (p *ImageProcessor) detectFormat(imageType vips.ImageType) string {
 	switch imageType {
@@ -497,89 +469,107 @@ func (p *ImageProcessor) detectFormat(imageType vips.ImageType) string {
 	}
 }
 
-// outputDimensions returns the dimensions resize gives an image of imgWidth x
+// resize resizes the image according to the fit mode. With FitOutside one
-// imgHeight for the target dimensions and fit mode, without resizing it, and
+// side can come out larger than requested; when it would be above
-// refuses an output too large for format. With FitOutside one side can come
+// MaxOutputDimension, resize returns ErrOutputTooLarge without resizing.
-// out larger than the target; when it would be above MaxOutputDimension,
-// outputDimensions returns ErrOutputTooLarge. An AVIF output of more than
-// MaxAVIFOutputPixels is ErrAVIFOutputTooLarge.
-func outputDimensions(
-	imgWidth, imgHeight, width, height int, fit FitMode, format Format,
-) (int, int, error) {
-	scaleW := float64(width) / float64(imgWidth)
-	scaleH := float64(height) / float64(imgHeight)
-	outputWidth, outputHeight := width, height
-
-	switch fit {
-	case FitCover, FitFill, "":
-		// Cover crops, and fill stretches, to exactly the target dimensions
-
-	case FitContain:
-		// Fit within the target dimensions, maintaining aspect ratio
-		scale := min(scaleW, scaleH)
-		outputWidth = int(float64(imgWidth) * scale)
-		outputHeight = int(float64(imgHeight) * scale)
-
-	case FitInside:
-		// Same as contain, but only shrink
-		scale := min(scaleW, scaleH, 1)
-		outputWidth = int(float64(imgWidth) * scale)
-		outputHeight = int(float64(imgHeight) * scale)
-
-	case FitOutside:
-		// The smallest dimension fits, the other may exceed the target
-		scale := max(scaleW, scaleH)
-		outputWidth = int(float64(imgWidth) * scale)
-		outputHeight = int(float64(imgHeight) * scale)
-
-	default:
-		return 0, 0, fmt.Errorf("%w: %s", ErrInvalidFitMode, fit)
-	}
-
-	// The side FitOutside scales past the target is held to the maximum too
-	if max(outputWidth, outputHeight) > MaxOutputDimension {
-		return 0, 0, fmt.Errorf("%w: %dx%d is above %d", ErrOutputTooLarge,
-			outputWidth, outputHeight, MaxOutputDimension)
-	}
-
-	if format == FormatAVIF && outputWidth*outputHeight > MaxAVIFOutputPixels {
-		return 0, 0, fmt.Errorf("%w: %dx%d is above %d pixels",
-			ErrAVIFOutputTooLarge, outputWidth, outputHeight, MaxAVIFOutputPixels)
-	}
-
-	return outputWidth, outputHeight, nil
-}
-
-// resize resizes the image to width x height, the dimensions
-// outputDimensions gives for the fit mode.
 func (p *ImageProcessor) resize(
 	img *vips.ImageRef, width, height int, fit FitMode,
 ) error {
 	switch fit {
 	case FitCover, "":
-		// Resize and crop to fill exact dimensions (default)
+		// Crop the centre and resize to fill exact dimensions (default)
-		return img.Thumbnail(width, height, vips.InterestingCentre)
+		return cropCentreAndResize(img, width, height)
+
+	case FitContain:
+		// Resize to fit within dimensions, maintaining aspect ratio. Scaled
+		// by the smaller scale, the short side of a thin image can round
+		// down to 0, which libvips refuses as a width, so each side is at
+		// least 1.
+		imgW, imgH := img.Width(), img.Height()
+		scaleW := float64(width) / float64(imgW)
+		scaleH := float64(height) / float64(imgH)
+		scale := min(scaleW, scaleH)
+		newW := max(1, int(float64(imgW)*scale))
+		newH := max(1, int(float64(imgH)*scale))
+
+		return img.Thumbnail(newW, newH, vips.InterestingNone)
 
 	case FitFill:
 		// Resize to exact dimensions (may distort)
 		return img.ThumbnailWithSize(width, height, vips.InterestingNone, vips.SizeForce)
 
 	case FitInside:
-		// outputDimensions keeps the dimensions of an image that already fits
+		// Same as contain, but only shrink
-		if img.Width() == width && img.Height() == height {
+		if img.Width() <= width && img.Height() <= height {
-			return nil
+			return nil // Already fits
 		}
 
-		return img.Thumbnail(width, height, vips.InterestingNone)
+		imgW, imgH := img.Width(), img.Height()
+		scaleW := float64(width) / float64(imgW)
+		scaleH := float64(height) / float64(imgH)
+		scale := min(scaleW, scaleH)
+		newW := max(1, int(float64(imgW)*scale))
+		newH := max(1, int(float64(imgH)*scale))
 
-	case FitContain, FitOutside:
+		return img.Thumbnail(newW, newH, vips.InterestingNone)
-		return img.Thumbnail(width, height, vips.InterestingNone)
+
+	case FitOutside:
+		// Resize so smallest dimension fits, may exceed target on other
+		// dimension. A side scaled to 1 can come out just under 1 in
+		// floating point, as 49 times 1/49 does, and round down to 0, which
+		// libvips refuses as a width, so each side is at least 1.
+		imgW, imgH := img.Width(), img.Height()
+		scaleW := float64(width) / float64(imgW)
+		scaleH := float64(height) / float64(imgH)
+		scale := max(scaleW, scaleH)
+		newW := max(1, int(float64(imgW)*scale))
+		newH := max(1, int(float64(imgH)*scale))
+
+		// The dimension that exceeds the target is held to the maximum too
+		if newW > MaxOutputDimension || newH > MaxOutputDimension {
+			return fmt.Errorf("%w: %dx%d is above %d", ErrOutputTooLarge,
+				newW, newH, MaxOutputDimension)
+		}
+
+		return img.Thumbnail(newW, newH, vips.InterestingNone)
 
 	default:
 		return fmt.Errorf("%w: %s", ErrInvalidFitMode, fit)
 	}
 }
 
+// cropCentreAndResize scales img by the same amount on both sides, the
+// least that makes it cover width x height, and cuts the centre width x
+// height out of it, as libvips' thumbnail does with InterestingCentre.
+// thumbnail alone scales the whole image and copies the result into memory
+// before it cuts, and for a thin image that is far larger than the image or
+// the output: 8192x64 at 8192x8192 is 1048576x8192, about 24 GiB. So the
+// centre is cut from img first, widened to whole pixels of img, which adds
+// less than one pixel of img on each side, and thumbnail scales and cuts
+// only that.
+func cropCentreAndResize(img *vips.ImageRef, width, height int) error {
+	imgW, imgH := img.Width(), img.Height()
+
+	// An image wider than the requested proportions keeps its height and
+	// loses width down to imgH*width/height, any other keeps its width and
+	// loses height down to imgW*height/width. Half of what is lost comes
+	// off each side, rounded down to whole pixels, so the part kept holds
+	// all of the centre.
+	left, top := 0, 0
+	if imgW*height > imgH*width {
+		left = (imgW*height - imgH*width) / (2 * height) //nolint:mnd // each side
+	} else {
+		top = (imgH*width - imgW*height) / (2 * width) //nolint:mnd // each side
+	}
+
+	err := img.ExtractArea(left, top, imgW-2*left, imgH-2*top)
+	if err != nil {
+		return err
+	}
+
+	return img.Thumbnail(width, height, vips.InterestingCentre)
+}
+
 const defaultQuality = 85
 
 // encode encodes an image to the specified format. govips gives pixa no way
@@ -677,16 +667,11 @@ func exportGIF(img *vips.ImageRef) ([]byte, error) {
 	return output, err
 }
 
-// webpEffort is the WebP effort, from 0 (fastest) to 6. It is below libvips'
+// webpEffort is libvips' default WebP effort, from 0 (fastest) to 6.
-// default of 4 so that an encode takes at most 40 seconds of the default
+const webpEffort = 4
-// downstream_timeout of 60 seconds. With one thread, as pixad runs libvips,
-// saving an 8192x8192 image of random pixels with an alpha channel takes up
-// to about 23 seconds at 2, and up to about 16 without one, at any quality;
-// with an alpha channel, 3 takes about 50 seconds and 4 about 61.
-const webpEffort = 2
 
-// exportWebP encodes img as lossy WebP at quality and webpEffort, without
+// exportWebP encodes img as lossy WebP at quality and libvips' default
-// metadata.
+// effort, without metadata.
 func exportWebP(img *vips.ImageRef, quality int) ([]byte, error) {
 	output, _, err := img.ExportWebp(&vips.WebpExportParams{
 		StripMetadata:   true,
@@ -698,13 +683,12 @@ func exportWebP(img *vips.ImageRef, quality int) ([]byte, error) {
 }
 
 // avifEffort is the AVIF effort, from 0 (fastest) to 9; 1 is the lowest
-// govips can set. With one thread, as pixad runs libvips, 1 takes about 52
+// govips can set. With one thread, as pixad runs libvips, 1 takes about 51
-// seconds to save an 8192x8192 RGB image of random pixels at quality 85,
+// seconds to save an 8192x8192 RGB image of random pixels, and about 110
-// and about 96 with an alpha channel, against the default
+// with an alpha channel, the worst case, against the default
-// downstream_timeout of 60 seconds, so an AVIF output is held to
+// downstream_timeout of 60 seconds. On an image of milder noise, which 1
-// MaxAVIFOutputPixels. On an image of milder noise, which 1 saves in about
+// saves in about 12 seconds, 2 takes nearly a minute and libvips' default,
-// 12 seconds, 2 takes nearly a minute and libvips' default, 4, takes
+// 4, takes minutes.
-// minutes.
 const avifEffort = 1
 
 // avifBitdepth is the AVIF bit depth, 8 bits per sample for every image.
@@ -731,18 +715,10 @@ func exportAVIF(img *vips.ImageRef, quality int) ([]byte, error) {
 // JPEG that names none.
 const jxlResolution = 72 / 25.4
 
-// jxlEffort is the JPEG XL effort, from 1 (fastest) to 9. It is below
+// exportJXL encodes img as JPEG XL at quality, with libvips' default effort
-// libvips' default of 7 so that an encode takes at most 40 seconds of the
+// and without metadata. govips sends libvips a distance, the JPEG XL
-// default downstream_timeout of 60 seconds. With one thread, as pixad runs
+// encoder's own measure of quality, along with the quality, and libvips then
-// libvips, saving an 8192x8192 image of random pixels with an alpha channel
+// uses the distance alone, so the quality is also given as a distance.
-// takes up to about 27 seconds at 4, and up to about 13 without one, at any
-// quality; with an alpha channel, 5 takes about 43 seconds and 7 about 81.
-const jxlEffort = 4
-
-// exportJXL encodes img as JPEG XL at quality and jxlEffort, without
-// metadata. govips sends libvips a distance, the JPEG XL encoder's own
-// measure of quality, along with the quality, and libvips then uses the
-// distance alone, so the quality is also given as a distance.
 func exportJXL(img *vips.ImageRef, quality int) ([]byte, error) {
 	// libvips converts a CMYK image to sRGB before it saves WebP, AVIF or
 	// PNG, but cannot save one as JPEG XL. encode has already converted
@@ -781,7 +757,6 @@ func exportJXL(img *vips.ImageRef, quality int) ([]byte, error) {
 	params := vips.NewJxlExportParams()
 	params.Quality = quality
 	params.Distance = jxlDistance(quality)
-	params.Effort = jxlEffort
 
 	output, _, err := toSave.ExportJxl(params)
 	if err != nil {

internal/imgcache/imgcache.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"errors"
 	"io"
+	"net/url"
 	"time"
 )
 
@@ -86,9 +87,11 @@ func ValidateImageRequest(req *ImageRequest) error {
 type ImageRequest struct {
 	// SourceHost is the origin host (e.g., "cdn.example.com")
 	SourceHost string
-	// SourcePath is the path on the origin (e.g., "/photos/cat.jpg")
+	// SourcePath is the path on the origin with its percent-escapes decoded,
+	// as url.URL's Path holds it (e.g., "/photos/my cat.jpg")
 	SourcePath string
-	// SourceQuery is the optional query string for the origin URL
+	// SourceQuery is the optional query string for the origin URL as the URL
+	// writes it, percent-escapes included, as url.URL's RawQuery holds it
 	SourceQuery string
 	// Size is the requested output dimensions
 	Size Size
@@ -107,20 +110,25 @@ type ImageRequest struct {
 	AllowHTTP bool
 }
 
-// SourceURL returns the full upstream URL to fetch.
+// SourceURL returns the full upstream URL to fetch: the source path
-// Uses http:// scheme when AllowHTTP is true, otherwise https://.
+// percent-encoded again, as net/url writes a path, so each source path is one
+// upstream path, and the source query as it is, with the characters
+// escapeQuery encodes percent-encoded. Uses http:// scheme when AllowHTTP is
+// true, otherwise https://.
 func (r *ImageRequest) SourceURL() string {
 	scheme := "https"
 	if r.AllowHTTP {
 		scheme = "http"
 	}
 
-	url := scheme + "://" + r.SourceHost + r.SourcePath
+	sourceURL := url.URL{
-	if r.SourceQuery != "" {
+		Scheme:   scheme,
-		url += "?" + r.SourceQuery
+		Host:     r.SourceHost,
+		Path:     r.SourcePath,
+		RawQuery: escapeQuery(r.SourceQuery),
 	}
 
-	return url
+	return sourceURL.String()
 }
 
 // ImageResponse represents a processed image ready to serve

internal/imgcache/urlparser.go

@@ -3,6 +3,7 @@ package imgcache
 import (
 	"errors"
 	"fmt"
+	"net/netip"
 	"net/url"
 	"regexp"
 	"strconv"
@@ -28,7 +29,9 @@ var (
 const MaxDimension = imageprocessor.MaxOutputDimension
 
 // ValidateDimension checks one requested width or height; name ("width" or
-// "height") appears in the error. 0 means "original size" and is valid.
+// "height") appears in the error. 0 is valid: that side is then worked out from
+// the source's proportions and the other requested side, rounded down but at
+// least 1, or is the source's own side when the other side is 0 too.
 func ValidateDimension(name string, value int) error {
 	if value < 0 {
 		return fmt.Errorf("%w: %s is negative", ErrInvalidSize, name)
@@ -50,6 +53,21 @@ var sizeFormatRegex = regexp.MustCompile(
 // percentEscapeRegex matches one percent-escape, such as "%2e"
 var percentEscapeRegex = regexp.MustCompile(`%[0-9A-Fa-f]{2}`)
 
+// queryEscapeRegex matches a percent-escape, or one character that a browser
+// percent-encodes in the query of an http or https URL: one outside the
+// printable ASCII characters "!" to "~", such as a space, a control character
+// or a character outside ASCII, or one of `"#'<>`. A "%" that does not begin a
+// percent-escape matches too.
+var queryEscapeRegex = regexp.MustCompile(`%[0-9A-Fa-f]{2}|[^!-~]|["#'<>%]`)
+
+// hostRegex matches the forms the host of an image URL may take: a host name
+// or an IPv4 address (letters, digits, "-", "_" and "."), or an IPv6 address
+// in brackets (hex digits, ":" and "."), either one optionally followed by ":"
+// and a port, which may be empty. Its one group holds the text in brackets,
+// which parseImageComponents also checks is an IPv6 address.
+var hostRegex = regexp.MustCompile(
+	`^(?:[A-Za-z0-9_.-]+|\[([0-9A-Fa-f:.]+)\])(?::[0-9]*)?$`)
+
 // ParsedURL contains the parsed components of an image proxy URL.
 type ParsedURL struct {
 	// Host is the source origin host (e.g., "cdn.example.com")
@@ -141,10 +159,20 @@ func parseImageComponents(remainder string) (*ParsedURL, error) {
 		return nil, ErrMissingHost
 	}
 
-	// A URL reads what comes before "@" as a user name and password, and the
+	// Only the forms hostRegex describes are accepted. A "?", "#" or "@" in
-	// rest as the host
+	// the host would make Go's URL parser, which the allowlist check and the
-	if strings.Contains(host, "@") {
+	// fetcher use, read a different host from the one allowlisted and cached
-		return nil, fmt.Errorf("%w: %q holds @", ErrInvalidHost, host)
+	hostMatch := hostRegex.FindStringSubmatch(host)
+	if hostMatch == nil {
+		return nil, fmt.Errorf("%w: %q", ErrInvalidHost, host)
+	}
+
+	// Go's URL parser refuses brackets around anything but an IPv6 address
+	if inBrackets := hostMatch[1]; inBrackets != "" {
+		addr, err := netip.ParseAddr(inBrackets)
+		if err != nil || !addr.Is6() {
+			return nil, fmt.Errorf("%w: %q", ErrInvalidHost, host)
+		}
 	}
 
 	// Host names do not depend on letter case, so the allowlist, the upstream
@@ -152,9 +180,20 @@ func parseImageComponents(remainder string) (*ParsedURL, error) {
 	// host
 	host = strings.ToLower(host)
 
-	// Extract query string if present in path
+	// Go's URL parser, which the fetcher uses, reads a "#" as the start of a
+	// fragment and drops the rest, so each text after it would be another
+	// spelling, cached under its own keys, of one upstream request
+	if strings.Contains(path, "#") {
+		return nil, fmt.Errorf(
+			`%w: the source path or query holds a "#"`, ErrInvalidPath)
+	}
+
+	// Extract query string if present in path. It is the source URL's query
+	// as that URL writes it. A character a URL's query cannot hold as it is,
+	// such as a space written %20 in the image URL, is percent-encoded here,
+	// so the query is signed and cached as it is sent upstream
 	if qIndex := strings.Index(path, "?"); qIndex != -1 {
-		query = path[qIndex+1:]
+		query = escapeQuery(path[qIndex+1:])
 		path = path[:qIndex]
 	}
 
@@ -243,6 +282,20 @@ func decodePercentEscapes(s string) string {
 	})
 }
 
+// escapeQuery returns a source query as a URL's query holds it: each
+// character that queryEscapeRegex names, such as a space, percent-encoded,
+// and the rest, the query's percent-escapes included, kept as they are. The
+// result is unchanged by a second escapeQuery.
+func escapeQuery(query string) string {
+	return queryEscapeRegex.ReplaceAllStringFunc(query, func(match string) string {
+		if percentEscapeRegex.MatchString(match) {
+			return match
+		}
+
+		return url.PathEscape(match)
+	})
+}
+
 // parseSizeFormat parses strings like "800x600.webp", "orig.png" or "800x600"
 func parseSizeFormat(s string) (Size, ImageFormat, error) {
 	matches := sizeFormatRegex.FindStringSubmatch(s)

internal/server/image_proxy_internal_test.go

@@ -14,10 +14,12 @@ import (
 	"net"
 	"net/http"
 	"net/http/httptest"
+	"net/netip"
 	"os"
 	"path/filepath"
 	"sync/atomic"
 	"testing"
+	"time"
 
 	"go.uber.org/fx"
 	"go.uber.org/fx/fxtest"
@@ -183,6 +185,57 @@ func startImageProxy(
 	return s, db.DB(), stateDir
 }
 
+// startImageProxyWithBlockedNetworks starts the components pixad's fx app
+// builds, from a config with a fresh state directory and blockedNetworks as
+// blocked_networks. Unlike startImageProxy it provides no upstream fetcher, so
+// the handlers build the one pixad uses, which checks every address it
+// connects to against the built-in blocked ranges and blocked_networks. It
+// returns the server with its routes.
+func startImageProxyWithBlockedNetworks(
+	t *testing.T, blockedNetworks []netip.Prefix,
+) *Server {
+	t.Helper()
+
+	stateDir := t.TempDir()
+	cfg := &config.Config{
+		SigningKey:      testSigningKey,
+		StateDir:        stateDir,
+		DBURL:           "file:" + filepath.Join(stateDir, "state.sqlite3"),
+		BlockedNetworks: blockedNetworks,
+		// A fetch that gets past the checks and finds nothing at the address
+		// fails within this time at the latest.
+		UpstreamFetchTimeout: 2 * time.Second,
+		// With no connection slots, the fetch would fail before connecting.
+		UpstreamConnections: config.DefaultUpstreamConnections,
+		DownstreamTimeout:   config.DefaultDownstreamTimeout,
+	}
+
+	var (
+		h  *handlers.Handlers
+		mw *middleware.Middleware
+	)
+
+	app := fxtest.New(t,
+		fx.Supply(cfg),
+		fx.Provide(
+			globals.New,
+			logger.New,
+			database.New,
+			healthcheck.New,
+			handlers.New,
+			middleware.New,
+		),
+		fx.Populate(&h, &mw),
+	)
+	app.RequireStart()
+	t.Cleanup(app.RequireStop)
+
+	s := &Server{config: cfg, mw: mw, h: h}
+	s.SetupRoutes()
+
+	return s
+}
+
 // getImage sends a GET for target to s and fails unless it answers 200.
 func getImage(t *testing.T, s *Server, target string) *httptest.ResponseRecorder {
 	t.Helper()

internal/server/source_url_internal_test.go

@@ -0,0 +1,665 @@
+package server
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"net/netip"
+	"net/url"
+	"regexp"
+	"slices"
+	"strconv"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"sneak.berlin/go/pixa/internal/encurl"
+	"sneak.berlin/go/pixa/internal/imgcache"
+	"sneak.berlin/go/pixa/internal/session"
+	"sneak.berlin/go/pixa/internal/signature"
+)
+
+// percentSignPath is a source path holding a "%", which a URL writes %25.
+const percentSignPath = "/100%25.png"
+
+// generateFormPattern extracts the CSRF token rendered into the generator
+// page's form.
+var generateFormPattern = regexp.MustCompile(
+	`(?s)action="/generate".*?name="gorilla\.csrf\.Token" value="([^"]+)"`)
+
+// generatedURLPattern extracts the path of the encrypted URL the generator
+// page shows.
+var generatedURLPattern = regexp.MustCompile(`value="[^"]*(/v1/e/[^"]+)"`)
+
+// recordingUpstream is a test upstream that answers every request with a PNG
+// and records the request target, the path and query as sent, of each request
+// it receives.
+type recordingUpstream struct {
+	*httptest.Server
+
+	mu       sync.Mutex
+	received []string
+}
+
+// startRecordingUpstream starts a recordingUpstream, which is closed when the
+// test ends.
+func startRecordingUpstream(t *testing.T) *recordingUpstream {
+	t.Helper()
+
+	source := encodeTestPNG(t, 64, 48)
+	upstream := &recordingUpstream{}
+
+	upstream.Server = httptest.NewServer(http.HandlerFunc(
+		func(w http.ResponseWriter, r *http.Request) {
+			upstream.mu.Lock()
+			upstream.received = append(upstream.received, r.RequestURI)
+			upstream.mu.Unlock()
+
+			w.Header().Set("Content-Type", "image/png")
+			_, _ = w.Write(source)
+		}))
+	t.Cleanup(upstream.Close)
+
+	return upstream
+}
+
+// requestTargets returns the request targets the upstream has received, in
+// the order it received them.
+func (u *recordingUpstream) requestTargets() []string {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+
+	return slices.Clone(u.received)
+}
+
+// pageOrigin is the origin of the login and generator pages below, opened
+// over HTTPS; httptest gives every request the host example.com.
+const pageOrigin = "https://example.com"
+
+// openRoot sends GET / to s over HTTPS, with cookie when it is not nil, and
+// returns the page: the login page, or the generator page for a login session.
+func openRoot(
+	t *testing.T, s *Server, cookie *http.Cookie,
+) *httptest.ResponseRecorder {
+	t.Helper()
+
+	req := httptest.NewRequestWithContext(t.Context(), http.MethodGet,
+		pageOrigin+"/", nil)
+	if cookie != nil {
+		req.AddCookie(cookie)
+	}
+
+	page := httptest.NewRecorder()
+	s.ServeHTTP(page, req)
+
+	return page
+}
+
+// submitForm posts form to path on s as a browser submits it from page over
+// HTTPS: with the CSRF token tokenPattern finds in page, the cookies page set,
+// cookie when it is not nil, and a Referer naming the page, which pixa needs
+// while debug is off, as it is in startImageProxy's config.
+func submitForm(
+	t *testing.T, s *Server, page *httptest.ResponseRecorder,
+	tokenPattern *regexp.Regexp, path string, form url.Values,
+	cookie *http.Cookie,
+) *httptest.ResponseRecorder {
+	t.Helper()
+
+	match := tokenPattern.FindStringSubmatch(page.Body.String())
+	if match == nil {
+		t.Fatalf("the page has no form for %s with a CSRF token", path)
+	}
+
+	form.Set("gorilla.csrf.Token", match[1])
+
+	req := httptest.NewRequestWithContext(t.Context(), http.MethodPost,
+		pageOrigin+path, strings.NewReader(form.Encode()))
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.Header.Set("Referer", pageOrigin+"/")
+
+	for _, c := range page.Result().Cookies() {
+		req.AddCookie(c)
+	}
+
+	if cookie != nil {
+		req.AddCookie(cookie)
+	}
+
+	rec := httptest.NewRecorder()
+	s.ServeHTTP(rec, req)
+
+	return rec
+}
+
+// postGenerate logs in to s with the signing key, submits the generator page's
+// form with sourceURL, and returns the response.
+func postGenerate(
+	t *testing.T, s *Server, sourceURL string,
+) *httptest.ResponseRecorder {
+	t.Helper()
+
+	login := submitForm(t, s, openRoot(t, s, nil), csrfFieldPattern, "/",
+		url.Values{"key": {testSigningKey}}, nil)
+
+	var sessionCookie *http.Cookie
+
+	for _, c := range login.Result().Cookies() {
+		if c.Name == session.CookieName {
+			sessionCookie = c
+		}
+	}
+
+	if sessionCookie == nil {
+		t.Fatalf("logging in answered %d and set no session cookie", login.Code)
+	}
+
+	rec := submitForm(t, s, openRoot(t, s, sessionCookie), generateFormPattern,
+		"/generate", url.Values{"url": {sourceURL}}, sessionCookie)
+	t.Logf("POST /generate with url %s: %d", sourceURL, rec.Code)
+
+	return rec
+}
+
+// TestImageURLFetchesItsSourcePathAndQuery requests image URLs whose source
+// query holds a space, written percent-encoded, or a "'", and whose source
+// path holds a "%", written percent-encoded, and checks the one request the
+// upstream receives: the source path percent-encoded again, and the source
+// query with the space written %20 and the "'" written %27, as a browser
+// writes them in a query.
+func TestImageURLFetchesItsSourcePathAndQuery(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name string
+		// pathAndQuery is the source path and query as the image URL writes
+		// them
+		pathAndQuery string
+		// wantRequest is the path and query the upstream must receive
+		wantRequest string
+	}{
+		{"space in the query", "/photo.png%3Fname=my%20cat",
+			"/photo.png?name=my%20cat"},
+		{"apostrophe in the query", "/photo.png%3Fname=it's",
+			"/photo.png?name=it%27s"},
+		{"percent sign in the path", percentSignPath, percentSignPath},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			upstream := startRecordingUpstream(t)
+			s, _, _ := startImageProxy(t, upstream.Server)
+
+			getImage(t, s, "/v1/image/"+upstreamHost+tt.pathAndQuery+"/orig.png")
+
+			got := upstream.requestTargets()
+			t.Logf("upstream received %q", got)
+
+			if !slices.Equal(got, []string{tt.wantRequest}) {
+				t.Errorf("upstream received %q, want [%q]", got, tt.wantRequest)
+			}
+		})
+	}
+}
+
+// TestSourcePathsHoldingSpaceAndPercent20AreFetchedApart requests twice each
+// of the source paths /my cat.png and /my%20cat.png, which the image URL
+// writes my%20cat.png and my%2520cat.png. Each is fetched once, as a request
+// of its own, and its second request is served from the cache.
+func TestSourcePathsHoldingSpaceAndPercent20AreFetchedApart(t *testing.T) {
+	t.Parallel()
+
+	upstream := startRecordingUpstream(t)
+	s, _, _ := startImageProxy(t, upstream.Server)
+
+	imageURLs := []string{
+		"/v1/image/" + upstreamHost + "/my%20cat.png/orig.png",
+		"/v1/image/" + upstreamHost + "/my%2520cat.png/orig.png",
+	}
+
+	for _, imageURL := range imageURLs {
+		getImage(t, s, imageURL)
+	}
+
+	for _, imageURL := range imageURLs {
+		cacheStatus := getImage(t, s, imageURL).Header().Get("X-Pixa-Cache")
+		if cacheStatus != string(imgcache.CacheHit) {
+			t.Errorf("second GET %s: X-Pixa-Cache = %q, want HIT",
+				imageURL, cacheStatus)
+		}
+	}
+
+	got := upstream.requestTargets()
+	t.Logf("upstream received %q", got)
+
+	want := []string{"/my%20cat.png", "/my%2520cat.png"}
+	if !slices.Equal(got, want) {
+		t.Errorf("upstream received %q, want %q", got, want)
+	}
+}
+
+// TestSourceQueryWrittenTwoWaysIsOneUpstreamRequest requests the source query
+// name=my%20cat written as the README says, name=my%2520cat, and with the
+// space written %20, name=my%20cat. Both are one source query: the upstream
+// receives one request, and the second image URL is served from the cache.
+func TestSourceQueryWrittenTwoWaysIsOneUpstreamRequest(t *testing.T) {
+	t.Parallel()
+
+	upstream := startRecordingUpstream(t)
+	s, _, _ := startImageProxy(t, upstream.Server)
+
+	getImage(t, s, "/v1/image/"+upstreamHost+"/photo.png%3Fname=my%2520cat/orig.png")
+
+	second := getImage(t, s,
+		"/v1/image/"+upstreamHost+"/photo.png%3Fname=my%20cat/orig.png")
+	if got := second.Header().Get("X-Pixa-Cache"); got != string(imgcache.CacheHit) {
+		t.Errorf("second X-Pixa-Cache = %q, want HIT", got)
+	}
+
+	got := upstream.requestTargets()
+	t.Logf("upstream received %q", got)
+
+	want := []string{"/photo.png?name=my%20cat"}
+	if !slices.Equal(got, want) {
+		t.Errorf("upstream received %q, want %q", got, want)
+	}
+}
+
+// TestGeneratedURLFetchesTheSourceURLAsEntered makes an encrypted URL on the
+// generator page for source URLs holding %20, %25 and %23 in the path or the
+// query, requests it, and checks that the upstream receives the path and
+// query exactly as entered.
+func TestGeneratedURLFetchesTheSourceURLAsEntered(t *testing.T) {
+	t.Parallel()
+
+	// The names hold no "%": each test's state directory is named after it
+	// and goes into the database URL, which decodes percent-escapes
+	tests := []struct {
+		name         string
+		pathAndQuery string
+	}{
+		{"space in the path", "/my%20cat.png"},
+		{"percent sign in the path", percentSignPath},
+		{"number sign in the path", "/a%23b.png"},
+		{"all three in the query", "/photo.png?name=my%20cat%25%23"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			upstream := startRecordingUpstream(t)
+			s, _, _ := startImageProxy(t, upstream.Server)
+
+			rec := postGenerate(t, s, "https://"+upstreamHost+tt.pathAndQuery)
+
+			match := generatedURLPattern.FindStringSubmatch(rec.Body.String())
+			if rec.Code != http.StatusOK || match == nil {
+				t.Fatalf("POST /generate: %d with no encrypted URL; body %s",
+					rec.Code, rec.Body.String())
+			}
+
+			getImage(t, s, match[1])
+
+			got := upstream.requestTargets()
+			t.Logf("upstream received %q", got)
+
+			if !slices.Equal(got, []string{tt.pathAndQuery}) {
+				t.Errorf("upstream received %q, want [%q]", got, tt.pathAndQuery)
+			}
+		})
+	}
+}
+
+// TestGeneratorMakesOneImageURLForTwoSpellingsOfASourceURL enters two
+// spellings of one source URL on the generator page, requests the encrypted
+// URL made for each, and checks that the upstream receives one request and
+// that the second encrypted URL is served from the cache. One pair writes the
+// host in upper and in lower case, which /v1/image/ lower-cases. The host is
+// an IPv6 address, whose hex digits have letter case, because the fetcher
+// would look a host name up in DNS. It has a port, as the fetcher refuses an
+// IPv6 address written without one. The other pair is a source URL with no
+// path and with the path "/", both of which fetch "/".
+func TestGeneratorMakesOneImageURLForTwoSpellingsOfASourceURL(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name string
+		// sourceURLs are the two spellings, after "https://"
+		sourceURLs []string
+		// wantRequest is the path the upstream must receive
+		wantRequest string
+	}{
+		{"host in upper and in lower case",
+			[]string{"[2001:DB8::A]:8443/a.png", "[2001:db8::a]:8443/a.png"},
+			"/a.png"},
+		{"no path and the path written as a slash",
+			[]string{upstreamHost, upstreamHost + "/"}, "/"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			upstream := startRecordingUpstream(t)
+			s, _, _ := startImageProxy(t, upstream.Server)
+
+			var imageURLs []string
+
+			for _, sourceURL := range tt.sourceURLs {
+				rec := postGenerate(t, s, "https://"+sourceURL)
+
+				match := generatedURLPattern.FindStringSubmatch(rec.Body.String())
+				if rec.Code != http.StatusOK || match == nil {
+					t.Fatalf("POST /generate: %d with no encrypted URL; body %s",
+						rec.Code, rec.Body.String())
+				}
+
+				imageURLs = append(imageURLs, match[1])
+			}
+
+			getImage(t, s, imageURLs[0])
+
+			second := getImage(t, s, imageURLs[1])
+			if got := second.Header().Get("X-Pixa-Cache"); got != string(imgcache.CacheHit) {
+				t.Errorf("second X-Pixa-Cache = %q, want HIT", got)
+			}
+
+			got := upstream.requestTargets()
+			t.Logf("upstream received %q", got)
+
+			if !slices.Equal(got, []string{tt.wantRequest}) {
+				t.Errorf("upstream received %q, want [%q]", got, tt.wantRequest)
+			}
+		})
+	}
+}
+
+// TestGeneratorRefusesSourceURLItCannotFetchAsWritten submits source URLs the
+// generator page must make no encrypted URL for, and checks that it answers
+// 400 and shows why. pixa keeps the source path decoded and fetches it
+// percent-encoded again, so a path holding %2F would be fetched with a "/"
+// there, whatever else the path holds. pixa sends no user name or password,
+// and no "?" with no query after it. For these the page shows the URL pixa
+// would fetch. A source path holding a ".." segment is refused, as
+// /v1/image/ refuses it.
+func TestGeneratorRefusesSourceURLItCannotFetchAsWritten(t *testing.T) {
+	t.Parallel()
+
+	// The test config sets allow_http, so pixa would fetch over plain HTTP
+	const fetchedFrom = "http://" + upstreamHost
+
+	tests := []struct {
+		name string
+		// sourceURL is the source URL after "https://"
+		sourceURL string
+		// wantShown is what the page must show: the URL pixa would fetch, or
+		// why it makes no encrypted URL
+		wantShown string
+	}{
+		{"encoded slash", upstreamHost + "/a%2Fb.png",
+			fetchedFrom + "/a/b.png"},
+		{"encoded slash and a space", upstreamHost + "/a%2Fb c.png",
+			fetchedFrom + "/a/b%20c.png"},
+		{"encoded slash and a bar", upstreamHost + "/a%2Fb|c.png",
+			fetchedFrom + "/a/b%7Cc.png"},
+		{"encoded slash and a letter outside ASCII", upstreamHost + "/café%2Fx.png",
+			fetchedFrom + "/caf%C3%A9/x.png"},
+		{"user name and password", "user:pass@" + upstreamHost + "/a.png",
+			fetchedFrom + "/a.png"},
+		{"question mark with no query", upstreamHost + "/a.png?",
+			fetchedFrom + "/a.png"},
+		{"two-dot segment", upstreamHost + "/a/../b.png",
+			"cannot hold a .. segment"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+
+			s, _, _ := startImageProxy(t, startRecordingUpstream(t).Server)
+
+			rec := postGenerate(t, s, "https://"+tt.sourceURL)
+			body := rec.Body.String()
+
+			if rec.Code != http.StatusBadRequest {
+				t.Errorf("status = %d, want %d", rec.Code, http.StatusBadRequest)
+			}
+
+			if strings.Contains(body, "/v1/e/") {
+				t.Error("the page shows an encrypted URL")
+			}
+
+			if !strings.Contains(body, tt.wantShown) {
+				t.Errorf("the page does not show %s; body %s", tt.wantShown, body)
+			}
+		})
+	}
+}
+
+// ipv6SourcePath is the source path of the image URLs signedImageURL and
+// encryptedImageURL make.
+const ipv6SourcePath = "/ipv6.png"
+
+// TestImageRoutesFetchIPv6SourceHostWithOrWithoutPort requests ipv6SourcePath
+// on an IPv6 source host, written with and without a port, through a
+// /v1/image/ URL with a signature and through an encrypted URL. The public
+// address 2001:db8::a, a documentation address (RFC 3849) the fetcher accepts
+// as public, answers 200, and the upstream receives the request for
+// ipv6SourcePath. The private address fd00::1 answers 403, and the upstream
+// receives no request: the test's fetcher connects every address to the test
+// upstream, so only the fetcher's check of the source URL can refuse it.
+func TestImageRoutesFetchIPv6SourceHostWithOrWithoutPort(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name       string
+		sourceHost string
+		wantStatus int
+		// wantRequests are the request targets the upstream must receive
+		wantRequests []string
+	}{
+		{"public address without a port", "[2001:db8::a]",
+			http.StatusOK, []string{ipv6SourcePath}},
+		{"public address with a port", "[2001:db8::a]:8443",
+			http.StatusOK, []string{ipv6SourcePath}},
+		{"private address without a port", "[fd00::1]",
+			http.StatusForbidden, nil},
+		{"private address with a port", "[fd00::1]:8443",
+			http.StatusForbidden, nil},
+	}
+
+	for _, tt := range tests {
+		imageURLs := map[string]string{
+			"image URL with a signature": signedImageURL(t, tt.sourceHost),
+			"encrypted URL":              encryptedImageURL(t, tt.sourceHost),
+		}
+
+		for route, imageURL := range imageURLs {
+			t.Run(tt.name+" "+route, func(t *testing.T) {
+				t.Parallel()
+
+				upstream := startRecordingUpstream(t)
+				s, _, _ := startImageProxy(t, upstream.Server)
+
+				rec := httptest.NewRecorder()
+				s.ServeHTTP(rec, httptest.NewRequestWithContext(
+					t.Context(), http.MethodGet, imageURL, nil))
+				t.Logf("GET %s: %d", imageURL, rec.Code)
+
+				if rec.Code != tt.wantStatus {
+					t.Errorf("status = %d, want %d; body %s",
+						rec.Code, tt.wantStatus, rec.Body.String())
+				}
+
+				got := upstream.requestTargets()
+				t.Logf("upstream received %q", got)
+
+				if !slices.Equal(got, tt.wantRequests) {
+					t.Errorf("upstream received %q, want %q", got, tt.wantRequests)
+				}
+			})
+		}
+	}
+}
+
+// signedImageURL returns the /v1/image/ URL of ipv6SourcePath on sourceHost at
+// its original size as PNG, with a signature made with the test signing key,
+// which a host not on the allowlist needs. The URL has no q or fit, so it is
+// signed with the quality 85 and the fit cover.
+func signedImageURL(t *testing.T, sourceHost string) string {
+	t.Helper()
+
+	path, sig, exp := signature.New(testSigningKey).GenerateSignedURL(
+		&signature.Request{
+			SourceHost: sourceHost,
+			SourcePath: ipv6SourcePath,
+			Format:     string(imgcache.FormatPNG),
+			Quality:    85,
+			FitMode:    string(imgcache.FitCover),
+		}, time.Hour)
+
+	return path + "?sig=" + sig + "&exp=" + strconv.FormatInt(exp, 10)
+}
+
+// encryptedImageURL returns the encrypted URL of ipv6SourcePath on sourceHost
+// at its original size as PNG, made with the test signing key.
+func encryptedImageURL(t *testing.T, sourceHost string) string {
+	t.Helper()
+
+	generator, err := encurl.NewGenerator(testSigningKey)
+	if err != nil {
+		t.Fatalf("encurl.NewGenerator: %v", err)
+	}
+
+	token, err := generator.Generate(&encurl.Payload{
+		SourceHost: sourceHost, SourcePath: ipv6SourcePath,
+		Format: imgcache.FormatPNG,
+	})
+	if err != nil {
+		t.Fatalf("encurl Generate: %v", err)
+	}
+
+	return "/v1/e/" + token + "/img.png"
+}
+
+// TestImageRoutesRefuseIPv6SourceHostHoldingInternalIPv4Address requests
+// ipv6SourcePath on IPv6 source hosts that hold a loopback or private IPv4
+// address in their last four bytes, in the IPv4-compatible form (::/96) or the
+// NAT64 form (64:ff9b::/96), through a /v1/image/ URL with a signature and
+// through an encrypted URL. Each answers 403, and the upstream receives no
+// request: the test's fetcher connects every address to the test upstream, so
+// only the fetcher's check of the source URL can refuse it.
+func TestImageRoutesRefuseIPv6SourceHostHoldingInternalIPv4Address(t *testing.T) {
+	t.Parallel()
+
+	upstream := startRecordingUpstream(t)
+	s, _, _ := startImageProxy(t, upstream.Server)
+
+	sourceHosts := []string{
+		"[::127.0.0.1]",     // IPv4-compatible form of a loopback address
+		"[::10.0.0.1]",      // IPv4-compatible form of a private address
+		"[64:ff9b::7f00:1]", // NAT64 form of 127.0.0.1
+	}
+
+	for _, sourceHost := range sourceHosts {
+		imageURLs := []string{
+			signedImageURL(t, sourceHost),
+			encryptedImageURL(t, sourceHost),
+		}
+
+		for _, imageURL := range imageURLs {
+			rec := httptest.NewRecorder()
+			s.ServeHTTP(rec, httptest.NewRequestWithContext(
+				t.Context(), http.MethodGet, imageURL, nil))
+			t.Logf("GET %s (source host %s): %d", imageURL, sourceHost, rec.Code)
+
+			if rec.Code != http.StatusForbidden {
+				t.Errorf("source host %s: status = %d, want %d; body %s",
+					sourceHost, rec.Code, http.StatusForbidden, rec.Body.String())
+			}
+		}
+	}
+
+	if got := upstream.requestTargets(); len(got) != 0 {
+		t.Errorf("upstream received %q, want no request", got)
+	}
+}
+
+// TestImageRoutesCheckIPv4AddressInIPv6SourceHost requests ipv6SourcePath on
+// source hosts that hold an IPv4 address, through a /v1/image/ URL with a
+// signature and through an encrypted URL, with the fetcher pixad uses.
+// [64:ff9b:1::a00:1], in the local-use NAT64 range (64:ff9b:1::/48, RFC 8215),
+// answers 403. [::203.0.113.5] holds the public address 203.0.113.5 in the
+// IPv4-compatible form (::/96) and answers 403 when blocked_networks lists
+// 203.0.113.0/24 or ::/96. [::ffff:203.0.113.5] holds it in the IPv4-mapped
+// form (::ffff:0:0/96) and answers 403 when blocked_networks lists
+// 203.0.113.0/24. Go keeps the plain address 203.0.113.5 in the mapped form
+// too, and an IPv6 range must not refuse it: with ::/0 listed it is not
+// refused. Neither is [::203.0.113.5] with blocked_networks empty. Where a
+// host is not refused, the fetcher tries to connect to it, finds nothing
+// there, and the request fails with another status.
+func TestImageRoutesCheckIPv4AddressInIPv6SourceHost(t *testing.T) {
+	t.Parallel()
+
+	const (
+		ipv4CompatibleHost = "[::203.0.113.5]"
+		ipv4MappedHost     = "[::ffff:203.0.113.5]"
+	)
+
+	tests := []struct {
+		name            string
+		sourceHost      string
+		blockedNetworks []netip.Prefix
+		wantForbidden   bool
+	}{
+		{"local-use NAT64 form of 10.0.0.1", "[64:ff9b:1::a00:1]", nil, true},
+		{"IPv4-compatible form of an address in blocked_networks",
+			ipv4CompatibleHost,
+			[]netip.Prefix{netip.MustParsePrefix("203.0.113.0/24")}, true},
+		{"IPv4-compatible form with ::/96 in blocked_networks",
+			ipv4CompatibleHost,
+			[]netip.Prefix{netip.MustParsePrefix("::/96")}, true},
+		{"IPv4-mapped form of an address in blocked_networks",
+			ipv4MappedHost,
+			[]netip.Prefix{netip.MustParsePrefix("203.0.113.0/24")}, true},
+		{"plain IPv4 address with ::/0 in blocked_networks",
+			"203.0.113.5",
+			[]netip.Prefix{netip.MustParsePrefix("::/0")}, false},
+		{"IPv4-compatible form of an address with blocked_networks empty",
+			ipv4CompatibleHost, nil, false},
+	}
+
+	for _, tt := range tests {
+		imageURLs := map[string]string{
+			"image URL with a signature": signedImageURL(t, tt.sourceHost),
+			"encrypted URL":              encryptedImageURL(t, tt.sourceHost),
+		}
+
+		for route, imageURL := range imageURLs {
+			t.Run(tt.name+" "+route, func(t *testing.T) {
+				t.Parallel()
+
+				s := startImageProxyWithBlockedNetworks(t, tt.blockedNetworks)
+
+				rec := httptest.NewRecorder()
+				s.ServeHTTP(rec, httptest.NewRequestWithContext(
+					t.Context(), http.MethodGet, imageURL, nil))
+				t.Logf("GET %s: %d", imageURL, rec.Code)
+
+				forbidden := rec.Code == http.StatusForbidden
+				if tt.wantForbidden && !forbidden {
+					t.Errorf("status = %d, want %d; body %s",
+						rec.Code, http.StatusForbidden, rec.Body.String())
+				}
+
+				if !tt.wantForbidden && forbidden {
+					t.Errorf("status = %d, want any other status; body %s",
+						rec.Code, rec.Body.String())
+				}
+			})
+		}
+	}
+}