Add optional per-chain byte cap in chained buffers
Previously, njs_chb_t accumulated bytes unbounded through the memory
pool until the final string boundary (njs_string_create_chb()) enforced
the NJS_STRING_MAX_LENGTH limit. For string-producing chains this let
the pool grow to tens of GiB before the final alloc was refused, which
on adversarial inputs (such as String.prototype.replace with a regex
that matches the full string and a replacement containing thousands of
$N references) led to process OOM instead of a catchable RangeError.
The fix is to add two fields to njs_chb_t: total_size (O(1) running byte
count) and max_size (optional upper bound, 0 means unlimited).
Exceeding max_size puts the chain into a sticky NJS_CHB_ERR_OVERFLOW
state. As part of the change njs_chb_drain() now returns early on a
failed chain, in line with njs_chb_drop().
QuickJS property setter and define APIs consume the JSValue argument on
both success and failure. This applies to JS_SetProperty(),
JS_SetPropertyStr(), JS_SetPropertyUint32(), JS_DefinePropertyValue(),
JS_DefinePropertyValueStr(), and JS_DefinePropertyValueUint32().
Do not free values after passing them to these APIs, including failure
paths in shared cleanup labels. Target objects and containers are not
consumed by these calls and still need normal cleanup unless ownership
has been transferred by inserting them into another object.
Also fix adjacent cleanup paths that leaked target objects after a
setter consumed the value being attached.
Align common helper failures with the exception policy used by both engines:
invalid numeric conversion is TypeError, and missing external receiver/context
for common log helpers is TypeError rather than an internal host failure.
The numeric conversion message changes from "is not a number" to
"invalid number" to match the QuickJS helper text.
Report XML API misuse as TypeError and XML parse failures as SyntaxError,
aligning njs and QuickJS behavior.
The QuickJS message "'this' is not XMLNode or XMLDoc" is changed to
"value is not XMLNode or XMLDoc" because qjs_xml_node() is not always
validating this.
Report stream session API misuse as TypeError, including wrong receiver,
unknown or duplicate event handlers, wrong handler phase, and invalid variable
access. Report status bounds violations as RangeError and keep stream output
pipeline failures as InternalError.
Stream: fixed variable value state on allocation failure
Reset the variable value validity flag when allocating storage for a stream
variable fails. This matches the HTTP variable path and prevents a failed
assignment from leaving a half-initialized value marked as valid.
Report HTTP request API misuse as TypeError and status bounds violations as
RangeError. Keep nginx output and request body collection failures as
InternalError, since they are host/runtime failures rather than invalid
JavaScript arguments.
Set pending exceptions before returning NJS_ERROR from njs response output
paths. Previously sendHeader(), send(), and finish() could fail without an
exception being available to the VM.
Report API misuse in Fetch, Request, Response, and Headers as TypeError, and
report status bounds violations as RangeError. Keep internal host failures as
InternalError and preserve conversion helper exceptions where they already
provide the real cause.
Preserve exceptions raised by QuickJS conversions and synthesize an out of
memory exception only for silent pool allocation failures in ngx_qjs_string()
callers.
Use JS_GetOpaque() where a class mismatch is a normal miss, avoiding a stale
pending TypeError while continuing with non-Headers and non-Request input.
Also release the headers init value before throwing for a non-object Headers
option.
Wrong receiver errors are caused by JS API misuse, not by internal host
failures. Report TypeError for TextEncoder, TextDecoder, console, and
fs Stats receiver checks.
Dmitry Volyntsev [Fri, 29 May 2026 22:58:34 +0000 (15:58 -0700)]
Fetch: reject unsafe request methods
Reject empty methods, C0 control bytes, space, and DEL before a fetch
request can be serialized. Preserve existing forbidden-method checks,
common-method normalization, and valid server-side extension methods.
Dmitry Volyntsev [Thu, 28 May 2026 05:06:55 +0000 (22:06 -0700)]
Fetch: fix keepalive with disabled TLS verification
HTTPS connections established with certificate verification disabled are
now excluded from the Fetch keepalive cache. This prevents a later
verified request to the same destination from reusing a TLS connection
that was created with verification disabled.
Dmitry Volyntsev [Sat, 23 May 2026 01:18:39 +0000 (18:18 -0700)]
Fixed Buffer allocation length checks
On 32-bit platforms, where size_t is 32 bits, callers passing
int64_t lengths >= 2^32 to njs_buffer_alloc() had the value
silently truncated before reaching the UINT32_MAX check in
njs_array_buffer_alloc(), so Buffer.from({length: 0x100000000})
returned a zero-sized buffer instead of raising a RangeError.
Widened the size parameter to uint64_t so the value reaches
njs_typed_array_alloc() intact and the "invalid index" range
error fires consistently on 32-bit and 64-bit builds.
Buffer.concat() is protected by the same change: it likewise
reads an int64_t length from JS and forwards it to
njs_buffer_alloc() without an upper bound check of its own.
Dmitry Volyntsev [Fri, 22 May 2026 04:05:59 +0000 (21:05 -0700)]
QuickJS: fixed fetch error-handling typos
Returned the JS_ThrowOutOfMemory() result from
ngx_qjs_fetch_response_ctor() on ngx_list_init() failure: previously
the call result was discarded and execution fell through.
Replaced a stale "ret" check with "rc" after ngx_qjs_headers_fill(),
so a non-NGX_OK return is no longer ignored.
Added a NULL check after JS_ToCStringLen() on the header key in
ngx_qjs_headers_ext_keys(), and replaced a stale "value" check with
"item" when testing JS_NewStringLen() for failure.
Dmitry Volyntsev [Thu, 21 May 2026 23:01:56 +0000 (16:01 -0700)]
QuickJS: fixed Buffer.from() and encoding error paths
Freed the typed array constructor object after reading constructor.name
while detecting Float32Array in Buffer.from(). This keeps the exception
path from leaking the constructor object.
Handled JS_ToCString() failure for constructor.name before comparing the
name, avoiding a NULL dereference when the property converts to an
exception, such as a Symbol value.
Divided the source offset by the element size in qjs_buffer_from()
typed-array path so the offset addresses the right element for 2-, 4-,
and 8-byte element types (previously the offset was left in byte units
while size was already in element units).
Added a NULL check on JS_ToCStringLen() in qjs_buffer_encoding(), and
moved the JS_FreeCString() call after the JS_ThrowTypeError() so the
encoding name remains valid while the error message is formatted.
Routed array source errors in qjs_buffer_from_object() through a single
fail label so the destination buffer is freed once on every failure path
(previously leaked on three of them).
Dmitry Volyntsev [Wed, 20 May 2026 23:27:46 +0000 (16:27 -0700)]
Added agent developer guide.
The new top-level AGENTS.md (with CLAUDE.md symlinked to it) is the
canonical index for agent and contributor instructions. Detailed
guides live under docs/agent/:
docs/agent/engine-dev.md building, testing, debugging the engine
and the nginx modules.
docs/agent/js-dev.md writing JavaScript for either engine,
with the common nginx API surface and
engine differences.
docs/agent/js-dev-njs.md specifics of the deprecated njs engine
and migration to QuickJS.
Dmitry Volyntsev [Fri, 15 May 2026 23:28:47 +0000 (16:28 -0700)]
Fixed call argument value snapshotting.
Since fd5e523f (0.9.7), njs has evaluated all call argument expressions
before emitting the frame and PUT_ARG instructions. This fixed await
expressions in call arguments and preserved argument side effects before
callee validation, but left PUT_ARG reading argument values only after later
argument expressions had already run.
As a result, an earlier argument backed by a mutable variable slot could
observe a later mutation of the same slot. For example, f(a, a = 2) passed
2 as both arguments instead of preserving the first value as 1. The same
issue applied to later argument effects such as getters.
The fix is to preserve affected argument values in temporary registers where
appropriate.
njs_chb_destroy() frees chain nodes through chain->free() and guards
against a NULL free callback. njs_chb_drain() and the tail-freeing
path of njs_chb_drop() called njs_mp_free() directly, which is wrong
for chains initialized with NJS_CHB_CTX_INIT() where chain->free is
js_free(), and unsafe for NGX_CHB_CTX_INIT() chains where chain->free
is NULL.
Both paths now route through chain->free() with the same NULL guard
as njs_chb_destroy().
Dmitry Volyntsev [Thu, 14 May 2026 23:10:33 +0000 (16:10 -0700)]
HTTP: fixed internalRedirect() in js_access.
Previously, r.internalRedirect() only stored the redirect URI and set
ctx->status to NGX_DONE. The stored URI was handled only by the js_content
finalization path.
As a result, when internalRedirect() was called from js_access, the access
handler returned NGX_DONE to the phase engine without starting the redirect,
leaving the request unfinished.
Now the internal redirect handling is shared by js_access and js_content, so
internalRedirect() works consistently in both handlers.
Dmitry Volyntsev [Thu, 14 May 2026 02:06:24 +0000 (19:06 -0700)]
HTTP: fixed r.return() with body in js_access.
Previously, when a js_access handler called r.return(status, body),
ngx_http_send_response() sent the response but returned NGX_OK. The
access handler then returned NGX_OK to the phase engine, which treated
the access check as allowed and continued to the content phase.
As a result, a denied request could still reach proxy_pass or another
content handler after the response had already been sent.
Now the access handler finalizes the request if a response was already
sent during js_access execution. This keeps r.return() behavior in
js_content unchanged, while making r.return(status, body) terminal in
js_access.
Dmitry Volyntsev [Thu, 14 May 2026 00:41:05 +0000 (17:41 -0700)]
Improved error stack frame handling.
Native frames always have an associated function. Keep the function
lookup in the native-frame branch to make the invariant explicit and to
avoid confusing static analysis.
Dmitry Volyntsev [Tue, 12 May 2026 23:52:56 +0000 (16:52 -0700)]
Fetch: fixed heap buffer overflow in proxy URL credentials.
The destination buffers for the decoded user and password in
ngx_js_parse_proxy_url() were a fixed 128 bytes, while the encoded
input length was bounded only by the URL length. Since
ngx_unescape_uri() writes at most one byte per input byte, raw
credentials longer than 128 bytes overflowed the buffer; the
length check ran only after the decode.
The fix is to size the destination buffer based on the encoded
input length.
The async method parses the client request body as an HTML form
and returns a Promise resolving to a form object with get(),
getAll(), has(), forEach(), hasFiles() accessors.
Supports "application/x-www-form-urlencoded" and "multipart/form-data"
content types. File parts are detected but their contents are not
exposed. An optional maxKeys option caps the number of fields.
File parts are detected but their contents are not exposed. A
proper File API with streaming Blob semantics is a significant amount
of work and is out of scope.
Dmitry Volyntsev [Sat, 28 Mar 2026 00:33:37 +0000 (17:33 -0700)]
HTTP: added js_access directive.
The directive registers a JavaScript handler in the access phase,
running after built-in access checkers (allow/deny, auth_basic,
auth_request). r.subrequest(), ngx.fetch() and other async operations
are supported.
The handler defaults to NGX_OK (access granted) on normal completion,
matching the behavior of other access phase modules. The r.decline()
method allows the handler to return NGX_DECLINED (no opinion), deferring
the decision to other access checkers under "satisfy any".
The r.return() method can send any HTTP response from the access phase,
including 3xx redirects for authentication flows.
Modules: fixed loading of the built-in "crypto" module.
Since 3185ce81 (0.9.7), njs_crypto_module has been registered
conditionally in auto/modules under NJS_HAVE_OPENSSL. libnjs.a is
built for the nginx module via nginx/config.make with
"./configure --no-openssl ...", so libnjs.a's njs_modules[] no
longer contains the crypto module. The nginx addon lists
(njs_http_js_addon_modules[] and njs_stream_js_addon_modules[],
plus the qjs variants) were not updated accordingly, making
"import cr from 'crypto'" fail at nginx -t with ENOENT.
Added &njs_crypto_module and &qjs_crypto_module to the shared addon
list under the existing NJS_HAVE_OPENSSL guard, next to the webcrypto
entries.
While here, the near-identical addon arrays were factored into
shared macros in a new header nginx/ngx_js_modules.h, so adding a
future conditional addon needs a single edit instead of four.
Modules: removed "js vm init" during configuration parsing.
Previously, a notice-level log message "js vm init %s: %p" was emitted
for each JS engine created during configuration parsing. This message
leaked to the compiled-in default error log even when the user
explicitly configured logging elsewhere, because config-time logging
uses the initial cycle log whose file descriptor is opened before the
user's error_log directive takes effect.
The log message was used for VM deduplication tests. The tests are
updated to verify unique engine identifiers via HTTP and stream
responses instead of grepping for log messages.
WebCrypto: fixed RSA JWK test data for Node.js compatibility.
Removed the incorrect "alg":"RS256" field from RSA JWK test data
files. The field is optional per RFC 7517, and the value RS256
(RSASSA-PKCS1-v1_5) was wrong for tests that use RSA-OAEP and
RSA-PSS algorithms. Node.js correctly rejects JWK imports when
the "alg" field does not match the requested algorithm.
The hash mismatch negative test was updated to use an inline JWK
with an explicit "alg" field instead of the shared file.
This ensures: `test/test262 --binary=node test/webcrypto` pass.
Dmitry Volyntsev [Tue, 24 Mar 2026 01:46:02 +0000 (18:46 -0700)]
Crypto: switched to OpenSSL EVP for hashing.
Previously, the crypto module used built-in software implementations
for a limited set of hash algorithms (md5, sha1, sha256). This
prevented users from using algorithms like sha384, sha512, and sha3
family, even when the underlying OpenSSL library supported them.
The change replaces built-in hash implementations with OpenSSL
EVP_MD_CTX for createHash() and HMAC_CTX for createHmac(), following
the webcrypto module. Algorithm lookup now uses EVP_get_digestbyname(),
making any digest supported by the linked OpenSSL available to
JavaScript code.
The module now requires OpenSSL and is conditionally compiled, same as
the webcrypto module. Builds without OpenSSL (--no-openssl) will no
longer have the crypto module available.
Tested with OpenSSL 3.0, OpenSSL 1.1.1w, LibreSSL 3.9.2, and
BoringSSL. SHA-3 tests are skipped when the SSL library does not
support them (e.g. BoringSSL).
Dmitry Volyntsev [Fri, 27 Mar 2026 01:03:41 +0000 (18:03 -0700)]
Modules: removed shared dict expiration from read-locked paths.
Previously, keys(), items(), and size() called ngx_js_dict_expire()
under a read lock. Since ngx_js_dict_expire() deletes nodes from
both rbtrees and frees slab memory, concurrent readers on different
worker processes could corrupt shared memory by freeing the same
expired nodes simultaneously.
The fix removes ngx_js_dict_expire() calls from all read-locked
paths and instead skips expired entries during iteration, consistent
with how get() and has() already handle expiry. Actual cleanup of
expired entries is deferred to write-side operations (set, add,
delete, clear).
Dmitry Volyntsev [Fri, 27 Mar 2026 00:16:25 +0000 (17:16 -0700)]
Modules: improved shared dict eviction strategy.
Previously, when a slab allocation failed in evict mode, only 16
entries were evicted with a single retry. This could still result
in SharedMemoryError when the freed slab slots did not match the
requested allocation size class, even though the zone had plenty
of evictable entries.
In practice, it might happen when the following conditions are met:
- The shared zone is full
- evict flag is enabled
- key/value entries differ in size
The allocation now retries in a loop, evicting 16 entries at a time,
until the allocation succeeds or no more entries remain in the expire
tree.
After this change, allocation with evict enabled can only fail when:
- the value is larger than the zone's usable space
- the expire tree has no entries left to evict
- zone metadata overhead leaves insufficient room
Dmitry Volyntsev [Fri, 27 Mar 2026 00:16:11 +0000 (17:16 -0700)]
Modules: fixed double-free in shared dict update with eviction.
Previously, when updating an existing key's string value in a shared
dictionary with timeout and evict enabled, ngx_js_dict_alloc() could
trigger ngx_js_dict_evict() if the zone was full. Since the node being
updated was still in the expire tree, eviction could free it. The
subsequent ngx_slab_free_locked() call in the update path then freed the
already-freed string data, causing the "chunk is already free" alert
followed by a segfault.
The fix removes the node from the expire tree before allocating
memory for the new value, preventing eviction from reaching it.
On allocation failure the node is re-inserted with its original
expiry time.
Dmitry Volyntsev [Thu, 12 Mar 2026 23:50:34 +0000 (16:50 -0700)]
Modules: preserved per-entry TTL on shared dict incr() calls.
Previously, incr() without an explicit timeout argument always
reset the entry expiry to the directive default, discarding any
per-entry timeout set by a prior add(), set(), or incr() call.
This aligns the behavior with Redis INCR and OpenResty
ngx.shared.DICT:incr() where value mutation does not touch the
existing TTL. An explicit timeout argument still updates it.
Dmitry Volyntsev [Thu, 12 Mar 2026 23:48:14 +0000 (16:48 -0700)]
Modules: added ttl() method to shared dictionaries.
The method returns the remaining time-to-live in milliseconds
for a given key, or undefined if the key does not exist or has
expired. Throws TypeError if the dictionary was declared without
the timeout parameter.
Previously, call lowering created FUNCTION_FRAME and METHOD_FRAME before
argument evaluation. This made call ordering observably wrong: for
non-callable callees, the error was thrown before arguments with side
effects were evaluated, violating the ECMAScript specification. It also
prevented await expressions in call arguments, which were rejected at
parse time because suspending inside a half-created frame was not
supported.
The fix evaluates arguments first, then emits the frame, PUT_ARG, and
FUNCTION_CALL. Callee and receiver values are captured into temporaries
before argument evaluation to guard against argument side effects.
Method properties are resolved via PROPERTY_GET before arguments.
METHOD_FRAME is redefined from a composite opcode (property lookup +
callability check + frame creation) to a pure frame-creation opcode
that takes an already-resolved function and explicit "this" value.
The parser always wraps call expressions in a NJS_TOKEN_FUNCTION_CALL
node, removing the NJS_TOKEN_NAME special case.
Previously, the "in" operator swap flag was relayed through generator
context. This broke when call-end handlers switched to
njs_generator_stack_pop(NULL), which released the context before the
swap was read in njs_generate_3addr_operation_end().
The fix derives the swap directly from the node token type
(NJS_TOKEN_IN), eliminating the context relay.
Introduce NJS_TOKEN_OPTIONAL_PRESERVE for optional-chain preserve nodes
instead of reusing OBJECT_VALUE, so OBJECT_VALUE remains strictly an
object/array literal structure token.
Route optional-preserve, object-value, and optional method-preserve
access through dedicated helper functions with narrow assertions,
removing direct u.object/left/right access from general parser and
generator paths.
Previously, grouped optional calls like (o?.m)() resolved the callee
through the optional chain but dispatched via plain FUNCTION_CALL,
losing the original receiver.
The fix stores the receiver on the call node so the upcoming
call-argument reorder can emit METHOD_FRAME with an explicit "this".
Call-expression setup and optional-chain preserve lookup are routed
through named helpers with generator-side validation.
Previously, the generator inferred reference intent from raw AST shape.
Now the parser marks the relevant property node as PROPERTY_REF before
building METHOD_CALL, assignment, or update nodes, and the generator
accepts both PROPERTY and PROPERTY_REF via
njs_generate_is_property_lvalue().
Introduce NJS_TOKEN_PROPERTY_REF as an explicit parser-side marker for
property accesses that carry reference semantics (assignment targets,
delete operands, increment/decrement, method-call receivers).
WebCrypto: validate JWK key type against algorithm in importKey().
Previously, importKey() did not verify that the JWK "kty" field
matched the requested algorithm. For example, importing a JWK with
kty "oct" (symmetric) while specifying an asymmetric algorithm like
ECDH caused a SEGV in EVP_PKEY_free() during cleanup. This happened
because the symmetric key data written into the union's "raw" member
overlapped with the "pkey" pointer, corrupting it.
The fix validates kty before calling any JWK import function:
- "RSA" is only accepted for RSA-OAEP, RSA-PSS, RSASSA-PKCS1-v1_5
- "EC" is only accepted for ECDSA, ECDH
- "oct" is only accepted for HMAC, AES-GCM, AES-CTR, AES-CBC
Dmitry Volyntsev [Thu, 26 Feb 2026 15:45:46 +0000 (07:45 -0800)]
Fixed string offset map corruption in scope values hash.
The issue was introduced in e7caa46d (0.9.5). When compile-time
UTF-8 string constants were copied into the values hash in
njs_scope_value_index(), the string data layout was calculated
incorrectly: the map offset did not account for the null terminator
added in e7caa46d, the "size" variable was overwritten corrupting
the subsequent memcpy, and the offset map was never initialized
to zero.
This caused SEGV/SIGBUS crashes for any multi-byte UTF-8 string
constant with more than 32 characters when accessing a character
at index >= 32 (e.g. via .replace() or bracket notation). The bug
only manifested when the string byte size was 4-byte aligned, as
otherwise alignment padding absorbed the missing byte.
The fix factors out njs_string_data_size() and njs_string_data_init()
helpers shared by njs_string_alloc() and njs_scope_value_index(),
eliminating the duplicated layout logic that caused the divergence.
Fixed logical assignment short-circuit with non-writable properties.
When the logical condition was already satisfied (e.g., falsy for &&=,
truthy for ||=, non-nullish for ??=), the short-circuit jump landed on
the property set instruction instead of past it. This caused spurious
TypeErrors for non-writable, getter-only, and non-extensible property
targets even though no assignment should occur.
This fixes logical assignment introduced in 1a64ba68.
This change fixes 9 more tests in test262.
Dmitry Volyntsev [Wed, 11 Feb 2026 15:13:06 +0000 (07:13 -0800)]
Added support for ||= and &&= logical assignment operators.
Unlike regular compound assignments (+=, -=), these operators
short-circuit: the RHS is not evaluated and no assignment occurs
if the logical condition is already satisfied.
Dmitry Volyntsev [Fri, 13 Feb 2026 05:31:23 +0000 (21:31 -0800)]
Shell: fixed interactive mode detection for piped stdin.
Previously, libedit callback API (rl_callback_read_char) was used for
interactive input processing. Unlike the blocking readline() API, the
callback interface does not work correctly when stdin is a pipe: input
characters are silently dropped and the line handler is never invoked.
Dmitry Volyntsev [Wed, 18 Feb 2026 01:56:22 +0000 (17:56 -0800)]
Replace per-VM indexed array for modules with hash-based lookup.
Previously, evaluated module values were cached in a separate
vm->modules array indexed by module->index assigned at compile time.
This required keeping the array size in sync with shared->module_items,
which was error-prone in interactive mode where new modules could be
compiled across commands.
Instead, store evaluated module values directly in the per-VM module
copy already maintained by vm->modules_hash. This unifies the import
and require() caching paths and eliminates the index-based array along
with module->index and shared->module_items fields.
This fixes `make shell_test` when configured with --debug=YES.
Dmitry Volyntsev [Wed, 25 Feb 2026 07:13:54 +0000 (23:13 -0800)]
Modules: suppressed slab log_nomem for evict shared dict zones.
When evict is enabled, memory allocation failures are expected
and handled by evicting old entries. The slab allocator's
"no memory" log messages are now suppressed for such zones.
Dmitry Volyntsev [Fri, 20 Feb 2026 23:45:25 +0000 (15:45 -0800)]
Modules: fixed expire field truncation in shared dict state files.
The njs_sprintf buffer for the expire field was sized for 10-digit
numbers, but current millisecond timestamps are 13 digits. This caused
silent truncation, making entries appear expired on a full restart.
The issue has been present since eca03622 (0.9.1), which introduced
the shared dictionary state file support.
Dmitry Volyntsev [Tue, 17 Feb 2026 00:20:20 +0000 (16:20 -0800)]
Fixed heap-buffer-overflow in atom hash caused by Symbol().
Previously, there was a key_hash collision between strings and
symbols in the atom hash table that led to use of uninitialized memory
and desynchronization of vm->atom_id_generator.
In the atom hash, string entries used djb_hash(content) as key_hash
while symbol entries used raw atom_id. Since djb_hash produces arbitrary
uint32_t values and atom_ids grow monotonically from NJS_ATOM_SIZE,
these spaces overlapped. When a symbol's atom_id matched an existing
string's djb_hash, njs_flathsh_insert() invoked the test function which
read uninitialized fhq.key fields (the function expected string data,
but the caller never initialized it for symbols). Additionally,
atom_id_generator was incremented before the insert, so on failure the
counter diverged from the actual element count, causing subsequent
njs_atom_to_value() calls to read out of bounds.
The fix is to partition the key_hash space using bit 31.
Dmitry Volyntsev [Tue, 17 Feb 2026 00:16:12 +0000 (16:16 -0800)]
Refactored atom hash into orthogonal find and add operations.
Previously, njs_atom_find_or_add() and njs_atom_find_or_add_string()
duplicated the lookup logic and coupled find with insert. Splitting
them into njs_atom_find() and njs_atom_add() separates concerns:
find does a pure lookup with raw bytes, add inserts a pre-built value.