diff --git a/lib/fs.js b/lib/fs.js
index 9f6cfbddb657..51c574cb3ede 100644
--- a/lib/fs.js
+++ b/lib/fs.js
@@ -451,10 +451,11 @@ function readFile(path, options, callback) {
   const flagsNumber = stringToFlags(options.flag, 'options.flag');
   path = getValidatedPath(path);
   if (options.buffer === undefined) {
-    // Open + fstat + read + close in one thread pool round trip for files of
-    // up to one chunk; larger files come back as an open fd + size and take
-    // the chunked reader below (readFileAfterOneShot). `true`: a handed-back
-    // fd will be closed through fs.close(), so track it as unmanaged.
+    // Open + fstat + read + close in one thread pool round trip for regular
+    // files of up to one chunk; larger and non-regular files (size 0) come
+    // back as an open fd + size and take the chunked reader below
+    // (readFileAfterOneShot). `true`: a handed-back fd will be closed through
+    // fs.close(), so track it as unmanaged.
     const job = new ReadFileJob(path, flagsNumber, kReadFileBufferLength, true);
     job.context = context;
     job.ondone = readFileAfterOneShot;
@@ -479,7 +480,8 @@ function readFileAfterOneShot(err, buffer, fd, size, closeErr) {
   }
   if (fd !== -1) {
     // (context.read() below performs the abort check for this case.)
-    // Larger than one chunk: continue exactly like after open + fstat.
+    // Larger than one chunk or not a regular file: continue exactly like
+    // after open + fstat.
     context.fd = fd;
     context.size = size;
     if (size > kIoMaxLength) {
diff --git a/lib/internal/fs/promises.js b/lib/internal/fs/promises.js
index 47574a5a4711..35268b33496f 100644
--- a/lib/internal/fs/promises.js
+++ b/lib/internal/fs/promises.js
@@ -2305,9 +2305,9 @@ async function readFile(path, options) {
   checkAborted(options.signal);
 
   if (options.buffer === undefined && vfsState.handlers === null) {
-    // Open + fstat + read + close in one thread pool round trip for files of
-    // up to one chunk; larger files come back as an open fd + size and are
-    // read by readFileHandle() as before.
+    // Open + fstat + read + close in one thread pool round trip for regular
+    // files of up to one chunk; larger and non-regular files (size 0) come
+    // back as an open fd + size and are read by readFileHandle() as before.
     path = getValidatedPath(path);
     const { 0: buffer, 1: fd, 2: size } = await readFileInOneRoundTrip(path, stringToFlags(flag));
     if (fd === -1) {
diff --git a/src/node_file.cc b/src/node_file.cc
index f5e7cc80e724..85327608bd28 100644
--- a/src/node_file.cc
+++ b/src/node_file.cc
@@ -3613,18 +3613,19 @@ static void ReadFileUtf8(const FunctionCallbackInfo<Value>& args) {
 // read + close -- instead of one round trip per step, which is what dominates
 // fs.readFile() for the typical small file and multiplies thread pool
 // contention when many files are read at once. Files whose size exceeds
-// `limit` are not read here: the job hands the open fd and the size back so
-// that the caller continues with the chunked reader (which stays fair and
-// abortable for large files); it then only saved the fstat() round trip.
+// `limit`, and files that are not regular files (a FIFO or a device may never
+// reach EOF), are not read here: the job hands the open fd and the size back
+// so that the caller continues with the chunked reader (which stays fair and
+// abortable between chunks); it then only saved the fstat() round trip.
 //
 // JS: const job = new ReadFileJob(path, flags, limit, trackFd);
 //     job.ondone = (err, buffer, fd, size, closeErr) => {...}; job.run(path);
 // Exactly one of these outcomes is reported:
 //   err            -- open/fstat/read failed (any fd opened here was closed);
-//   fd >= 0, size  -- file is larger than `limit`: caller owns fd now (it is
-//                     registered as an unmanaged fd iff trackFd, i.e. when the
-//                     caller will close it through fs.close() rather than a
-//                     FileHandle);
+//   fd >= 0, size  -- file is larger than `limit`, or not a regular file and
+//                     size is 0: caller owns fd now (it is registered as an
+//                     unmanaged fd iff trackFd, i.e. when the caller will
+//                     close it through fs.close() rather than a FileHandle);
 //   buffer         -- the whole content; closeErr set if only close() failed.
 class ReadFileJob final : public AsyncWrap, public ThreadPoolWork {
  public:
@@ -3689,8 +3690,9 @@ class ReadFileJob final : public AsyncWrap, public ThreadPoolWork {
     const uint64_t size = regular ? static_cast<uint64_t>(st->st_size) : 0;
     uv_fs_req_cleanup(&req);
 
-    if (size > limit_) {
-      // Too large to read in one go here: hand the fd back.
+    if (size > limit_ || !regular) {
+      // Too large to read in one go here, or possibly endless: hand the fd
+      // back.
       fd_ = fd;
       size_ = size;
       return;
diff --git a/test/parallel/test-fs-readfile-fifo-abort.js b/test/parallel/test-fs-readfile-fifo-abort.js
new file mode 100644
index 000000000000..d22f457a34c0
--- /dev/null
+++ b/test/parallel/test-fs-readfile-fifo-abort.js
@@ -0,0 +1,48 @@
+'use strict';
+// Env: UV_THREADPOOL_SIZE=2
+// fs.readFile() and fs.promises.readFile() stay abortable on a FIFO whose
+// writer stays open, so the read never reaches EOF.
+//
+// The reader's open() holds a thread pool thread until a writer opens the
+// FIFO, and the writer's open() needs a second one.
+const common = require('../common');
+if (common.isWindows)
+  common.skip('no mkfifo on Windows');
+const tmpdir = require('../common/tmpdir');
+const assert = require('assert');
+const { execFileSync } = require('child_process');
+const fs = require('fs');
+const { promisify } = require('util');
+
+tmpdir.refresh();
+
+async function testAbort(name, read) {
+  const fifo = tmpdir.resolve(`${name}.fifo`);
+  execFileSync('mkfifo', [fifo]);
+  const ac = new AbortController();
+  const settled = assert.rejects(read(fifo, { signal: ac.signal }), { name: 'AbortError' });
+  const writer = await fs.promises.open(fifo, 'w');
+  ac.abort();
+  // A reader that honors the abort closes the FIFO after at most one more
+  // read, and the next write fails with EPIPE. One that ignores it keeps
+  // draining the FIFO until the writer closes.
+  let closedByReader = false;
+  const chunk = Buffer.alloc(1024);
+  for (let i = 0; i < 1000; i++) {
+    try {
+      await writer.write(chunk);
+    } catch (err) {
+      assert.strictEqual(err.code, 'EPIPE');
+      closedByReader = true;
+      break;
+    }
+  }
+  await writer.close();
+  assert.ok(closedByReader, `${name} kept reading after the abort`);
+  await settled;
+}
+
+(async () => {
+  await testAbort('callback', promisify(fs.readFile));
+  await testAbort('promises', fs.promises.readFile);
+})().then(common.mustCall());