Merge branch 'feat/p3-m5-swmr-reader' into feat/p3-wasm-swmr-python

CHANGELOG.md

@@ -2,6 +2,76 @@
 
 ## Unreleased
 
+### Range reads, milestone M5: reading files a SWMR writer is appending to (2026-09-27)
+Design: `docs/design/swmr.md`.
+- **Fix: files with the SWMR-write flag were bounded by a stale end of
+  file** (since the end-of-file check of 2026-09-26, unreleased). A
+  libhdf5 SWMR writer (h5py `f.swmr_mode = True`) does not keep
+  the superblock's end-of-file address up to date; a copy h5py made of its
+  own file mid-write records 715 in a 6 030-byte file.
+  `Superblock::data_end` only ignored the recorded end when it lay past the
+  end of the file, so such a file listed, but every chunked read failed
+  ("unexpected EOF: need 787 bytes, have 715") and `h5rs check` reported
+  its chunk indexes past the end of the file. For a v3 superblock with the
+  SWMR-write flag the data now ends at the end of the file, as libhdf5's
+  SWMR reader reads it (it skips the end-of-allocation check). Every open
+  path (`File::open`, `open_buffered`, `from_bytes`, `open_storage`,
+  `MmapFile`, `LazyFile`, `h5rs`) reads such a copy now.
+- **`File::open_swmr(path)` / `File::open_storage_swmr(storage)`**: open a
+  file a SWMR writer may still be appending to, as libhdf5's SWMR reader
+  (h5py `File(path, "r", swmr=True)`) does. The file is read with
+  positioned reads (the new `clawhdf5::FileStorage`: `pread`/`seek_read`,
+  never mapped, `len()` the file's current length), reads are bounded by
+  the file's length at the time of each read, and the chunk cache is not
+  used (a cached chunk index would hide new chunks; a cached edge chunk
+  would read as fill where the writer has since written). A file open for
+  writing without SWMR is refused with `Error::Locked`, as libhdf5 refuses
+  it.
+  Only a file whose superblock has the SWMR-write flag when it is opened is
+  read this way; any other file (its writer has closed it) reads exactly as
+  `File::open` reads it, bounded by its recorded end of file, and
+  `is_swmr_read()` is `false`.
+- **`Dataset::refresh()`** reads the dataset's object header again
+  (`H5Drefresh`, h5py `Dataset.refresh()`), so `shape()` and later reads
+  see the writer's appends; a handle keeps its extent until refreshed.
+- **Bounded retries.** On a SWMR-read file, an operation (open, lookups
+  and listings, refresh, every dataset read including strings and
+  variable-length data, attributes, `File::decode_*`) that fails with an error a concurrent write can cause —
+  the failures libhdf5's SWMR reader retries: a checksum mismatch, a read
+  past the file's current end, an object header prefix that does not
+  decode — is run again from the start, up to
+  `File::swmr_read_attempts()` times (`SWMR_READ_ATTEMPTS` = 100,
+  libhdf5's default metadata read attempts for SWMR readers,
+  `H5Pset_metadata_read_attempts`; `set_swmr_read_attempts` changes it),
+  pausing 1 µs doubling to 10 ms between attempts. Results come only from
+  an attempt in which every structure verified, so a torn read is at worst
+  an error, never data. Any other error is returned at once.
+  `File::swmr_retries()` counts the retries.
+  `File::swmr_writer_active()` reads the superblock flags again to tell
+  when the writer has closed the file.
+- Tests (`crates/clawhdf5/tests/swmr_interop.rs`): the mid-write copy
+  (fixture `tests/fixtures/swmr_mid_write.h5`) through every open path and
+  against h5py's SWMR reader; garbled reads (a storage that corrupts the
+  next reads) retried, never returned, and given up after the attempts;
+  every read path (listings, attributes compact and dense, strings,
+  variable-length data; fixture `tests/fixtures/swmr_strings_attrs.h5`)
+  with each of its reads failing once in turn returns the same result;
+  and a live test: an h5py writer appends to a 1-D and a 2-D dataset with
+  one unlimited dimension (Extensible Array index, one of them gzip) and a
+  2-D dataset with two (v2 B-tree index) for 2 500 steps
+  (`CLAWHDF5_SWMR_STEPS`), flushing after each, while two Rust reader
+  threads refresh and read all three in a loop — every value must be the
+  one the writer wrote at its position, extents never shrink — beside
+  h5py's own SWMR reader doing the same checks; after the writer closes,
+  the live handle and a new `File::open` read exactly what h5py reads.
+  Run on tank, 2026-09-27, with h5py 3.16 / HDF5 2.0:
+  `CLAWHDF5_REQUIRE_INTEROP=1 CLAWHDF5_PYTHON=.venv/bin/python cargo test
+  -p clawhdf5 --test swmr_interop`; also at 20 000 steps in a release
+  build. A test with the chunk cache left on in live mode fails it.
+- Not covered: SWMR writing, remote SWMR (`BlockCache` caches blocks and
+  `HttpStorage` pins the length), `MmapFile`/`LazyFile`, and refreshing
+  groups or attributes (a SWMR writer cannot add objects or attributes).
+
 ### Range reads, milestone M3: remote files (2026-09-26)
 - **New crate `clawhdf5-remote`.** `open_url("http://host/file.h5")` gives
   a `clawhdf5::File` (through `File::open_storage`) that reads the file by

README.md

@@ -101,6 +101,10 @@ breaking change, are in [CHANGELOG.md](CHANGELOG.md).
   HTTP server (or in S3/GCS/Azure, opt-in) by range requests through a
   block cache, without downloading it; `h5rs` takes URLs with its `remote`
   feature. See [Reading remote files](#reading-remote-files).
+- `File::open_swmr` follows a file an h5py/libhdf5 SWMR writer is still
+  appending to (`Dataset::refresh`, bounded retries); copies of such files
+  taken mid-write read with every open path. See
+  [Following a file a SWMR writer is appending to](#following-a-file-a-swmr-writer-is-appending-to).
 
 **Tooling**
 - CI now runs the h5py/netCDF4 interop suites for real (they had been skipping
@@ -504,6 +508,43 @@ built with `--features remote` takes the same URLs:
 `https://` is the `https` feature (rustls with ring, which compiles C).
 Limits are in [known issues](docs/known-issues.md).
 
+### Following a file a SWMR writer is appending to
+
+`File::open_swmr` reads a file that a libhdf5 writer in SWMR mode (h5py
+`f.swmr_mode = True`) is still appending to, as h5py's
+`File(path, "r", swmr=True)` does: `Dataset::refresh()` picks up the new
+extent, every read reads the chunk index as it is now, and a read that
+races the writer (a checksum that fails mid-flush) is retried, up to 100
+attempts as in libhdf5, and never returned torn.
+
+```rust
+use std::time::{Duration, Instant};
+
+let file = clawhdf5::File::open_swmr("live.h5")?;
+let mut ds = file.dataset("samples")?;
+let (mut seen, mut last_growth) = (0, Instant::now());
+// Stop when the writer closes the file, or when the dataset has not grown
+// for a minute: a writer that crashed or was killed never clears the
+// SWMR-write flag, so `swmr_writer_active()` alone can stay true forever.
+while file.swmr_writer_active()? && last_growth.elapsed() < Duration::from_secs(60) {
+    ds.refresh()?;
+    let n = ds.shape()?[0];
+    if n > seen {
+        // read rows seen..n ...
+        (seen, last_growth) = (n, Instant::now());
+    }
+    std::thread::sleep(Duration::from_millis(100));
+}
+ds.refresh()?; // the final extent
+```
+
+`swmr_writer_active()` reads the superblock's SWMR-write flag, which libhdf5
+clears only when the writer closes the file; a file whose writer died keeps
+it set (as the mid-write copy in `tests/fixtures/swmr_mid_write.h5` does),
+so a follower needs its own stop condition, like the idle timeout above.
+
+Design and limits: [docs/design/swmr.md](docs/design/swmr.md).
+
 ### Python
 
 `crates/clawhdf5-py` is a Python package (PyO3 + numpy) that reads HDF5 with

crates/clawhdf5-format/src/attribute.rs

@@ -576,7 +576,14 @@ pub fn find_attribute_in_file(
     offset_size: u8,
     length_size: u8,
 ) -> Result<Option<AttributeMessage>, FormatError> {
-    find_attribute_core(file_data, header, name, offset_size, length_size)
+    find_attribute_core(
+        file_data,
+        header,
+        name,
+        offset_size,
+        length_size,
+        &mut Vec::new(),
+    )
 }
 
 /// [`find_attribute_in_file`] over any [`Storage`] (see
@@ -594,16 +601,48 @@ pub fn find_attribute_in<S: Storage + ?Sized>(
 ) -> Result<Option<AttributeMessage>, FormatError> {
     match file_data.as_contiguous() {
         Some(all) => find_attribute_in_file(all, header, name, offset_size, length_size),
-        None => find_attribute_core(file_data, header, name, offset_size, length_size),
+        None => find_attribute_core(
+            file_data,
+            header,
+            name,
+            offset_size,
+            length_size,
+            &mut Vec::new(),
+        ),
     }
 }
 
+/// [`find_attribute_in`], also returning the errors of the attributes it
+/// could not read on the way (which it leaves out rather than failing
+/// the call): the attribute asked for may be one of them. A reader of a
+/// file that is being written uses them to tell a read that raced the
+/// writer from an absent attribute.
+pub fn find_attribute_reporting_in<S: Storage + ?Sized>(
+    file_data: &S,
+    header: &ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(Option<AttributeMessage>, Vec<FormatError>), FormatError> {
+    let mut errors = Vec::new();
+    let found = find_attribute_core(
+        file_data,
+        header,
+        name,
+        offset_size,
+        length_size,
+        &mut errors,
+    )?;
+    Ok((found, errors))
+}
+
 fn find_attribute_core<S: Storage + ?Sized>(
     file_data: &S,
     header: &ObjectHeader,
     name: &str,
     offset_size: u8,
     length_size: u8,
+    errors: &mut Vec<FormatError>,
 ) -> Result<Option<AttributeMessage>, FormatError> {
     let attr_info = find_attribute_info(header, offset_size)?;
     let dense = attr_info
@@ -612,12 +651,10 @@ fn find_attribute_core<S: Storage + ?Sized>(
     let Some((fh_addr, btree_addr)) = dense else {
         // Compact only (or dense storage without a name index, which a
         // listing reports): as a listing finds it.
-        return Ok(
-            extract_attributes_tolerant_in(file_data, header, offset_size, length_size)?
-                .0
-                .into_iter()
-                .find(|a| a.name == name),
-        );
+        let (attrs, errs) =
+            extract_attributes_tolerant_in(file_data, header, offset_size, length_size)?;
+        errors.extend(errs);
+        return Ok(attrs.into_iter().find(|a| a.name == name));
     };
     let btree_hdr = BTreeV2Header::parse_in(
         file_data,
@@ -627,12 +664,10 @@ fn find_attribute_core<S: Storage + ?Sized>(
     )?;
     let fh = FractalHeapHeader::parse_in(file_data, fh_addr, offset_size, length_size)?;
     if btree_hdr.tree_type != ATTRIBUTE_NAME_INDEX || btree_hdr.record_size < 4 {
-        return Ok(
-            extract_attributes_tolerant_in(file_data, header, offset_size, length_size)?
-                .0
-                .into_iter()
-                .find(|a| a.name == name),
-        );
+        let (attrs, errs) =
+            extract_attributes_tolerant_in(file_data, header, offset_size, length_size)?;
+        errors.extend(errs);
+        return Ok(attrs.into_iter().find(|a| a.name == name));
     }
 
     // A listing has the compact attributes first.
@@ -671,10 +706,10 @@ fn find_attribute_core<S: Storage + ?Sized>(
                 AttributeMessage::parse_in_storage(&d, file_data, offset_size, length_size)
             });
         // One that cannot be read is left out, as from a listing.
-        if let Ok(attr) = attr
-            && attr.name == name
-        {
-            return Ok(Some(attr));
+        match attr {
+            Ok(attr) if attr.name == name => return Ok(Some(attr)),
+            Ok(_) => {}
+            Err(e) => errors.push(e),
         }
     }
     Ok(None)

crates/clawhdf5-format/src/superblock.rs

@@ -116,22 +116,27 @@ impl Superblock {
     /// [`FormatError::TruncatedFile`]. Bytes past that address are not part
     /// of the file: libhdf5 fails any read of them ("addr overflow" /
     /// "address plus size exceeds file eoa"), so a reader should parse only
-    /// the data up to the returned end. As libhdf5 does for a SWMR reader,
-    /// the check is skipped for a version-3 superblock whose writer is still
-    /// writing it in SWMR mode (it extends the file as it goes); the data
-    /// then ends at the end of the file.
+    /// the data up to the returned end.
+    ///
+    /// A version-3 superblock with the SWMR-write flag set belongs to a file
+    /// a SWMR writer has open (or had, and did not close). That writer does
+    /// not keep the recorded end of file up to date — a copy taken mid-write
+    /// can record an end of a few hundred bytes in a file of tens of
+    /// kilobytes — and libhdf5's SWMR reader skips its end-of-allocation
+    /// check for every read (`H5FD_read`). For such a superblock the data
+    /// ends at the end of the file, whatever end it records.
     ///
     /// When the superblock's recorded base address differs from where the
     /// superblock actually is (a user block added or removed after the file
     /// was written), libhdf5 moves the recorded end of file by the same
     /// amount, and so does this.
     pub fn data_end(&self, user_block: u64, file_len: u64) -> Result<u64, FormatError> {
-        let eof =
-            i128::from(self.eof_address) - i128::from(self.base_address) + i128::from(user_block);
-        if eof < 0 || eof > i128::from(file_len) {
         if self.version >= 3 && self.is_swmr_write() {
             return Ok(file_len.saturating_sub(user_block));
         }
+        let eof =
+            i128::from(self.eof_address) - i128::from(self.base_address) + i128::from(user_block);
+        if eof < 0 || eof > i128::from(file_len) {
             return Err(FormatError::TruncatedFile {
                 stored_eof: u64::try_from(eof).unwrap_or(self.eof_address),
                 actual_len: file_len,
@@ -617,6 +622,13 @@ mod tests {
         let mut swmr = Superblock::parse(&build_v2_bytes(8, 3), 0).unwrap();
         swmr.consistency_flags = swmr_flags::WRITE_ACCESS | swmr_flags::SWMR_WRITE;
         assert_eq!(swmr.data_end(0, 1000), Ok(1000));
+        // ... nor bounded by its recorded end, which the writer does not
+        // keep up to date (2048 here).
+        assert_eq!(swmr.data_end(0, 17_857), Ok(17_857));
+        assert_eq!(swmr.data_end(512, 17_857), Ok(17_345));
+        // Without the SWMR-write flag the recorded end bounds the data.
+        swmr.consistency_flags = swmr_flags::WRITE_ACCESS;
+        assert_eq!(swmr.data_end(0, 17_857), Ok(2048));
     }
 
     #[test]

crates/clawhdf5/src/lib.rs

@@ -47,6 +47,7 @@ pub mod lazy;
 #[cfg(feature = "mmap")]
 pub mod mmap_file;
 pub mod reader;
+mod swmr;
 pub mod types;
 pub mod vlen;
 pub mod writer;
@@ -57,6 +58,7 @@ pub use lazy::{LazyDataset, LazyFile, LazyGroup};
 #[cfg(feature = "mmap")]
 pub use mmap_file::{MmapDataset, MmapFile, MmapGroup};
 pub use reader::{Dataset, File, Group, SharedStorage, VdsResolver};
+pub use swmr::{FileStorage, SWMR_READ_ATTEMPTS};
 pub use types::{AttrValue, DType};
 pub use vlen::VlenValue;
 pub use writer::FileBuilder;

crates/clawhdf5/src/reader.rs

@@ -28,7 +28,9 @@ use clawhdf5_format::superblock_ext::{self, CacheImageState};
 
 use crate::cache_image::{self, ImageView};
 use crate::error::Error;
-use crate::types::{AttrValue, DType, classify_datatype, read_attr, read_attrs};
+use crate::types::{
+    AttrValue, DType, classify_datatype, read_attr_reporting, read_attrs_reporting,
+};
 
 // ---------------------------------------------------------------------------
 // FileData — internal storage for owned bytes, an mmap, or any Storage
@@ -115,6 +117,10 @@ struct FileData {
     /// parser reads asks for it, and the patched/overlay checks and range
     /// conversions behind it cost a local metadata walk a few percent.
     contiguous: Option<WholeView>,
+    /// A file a SWMR writer may still be appending to
+    /// ([`File::open_swmr`]): reads are bounded by the storage's current
+    /// length, not by `end`, and nothing is ever read as one slice.
+    live: bool,
 }
 
 /// A borrow of the HDF5 data held by a [`FileData`]'s own `backing` or
@@ -138,7 +144,7 @@ impl FileData {
     /// bytes past the recorded end of file are not read, as in libhdf5.
     fn new(mut backing: Backing) -> Result<(Self, Superblock), Error> {
         if let Backing::Storage(storage) = backing {
-            return Self::new_storage(storage);
+            return Self::new_storage(storage, false, &mut None);
         }
         let whole = backing.whole_file().unwrap_or_default();
         let (user_block, hdf5) = signature::split_user_block(whole)?;
@@ -173,6 +179,7 @@ impl FileData {
             overlay: Vec::new(),
             image_error,
             contiguous: None,
+            live: false,
         };
         data.contiguous = data.find_contiguous();
         Ok((data, superblock))
@@ -180,7 +187,18 @@ impl FileData {
 
     /// [`Self::new`] for a [`Storage`] backend: the same checks, through
     /// reads of the storage.
-    fn new_storage(storage: SharedStorage) -> Result<(Self, Superblock), Error> {
+    ///
+    /// With `swmr` ([`File::open_storage_swmr`]) the file is read live when
+    /// its superblock has the SWMR-write flag (version 3), as libhdf5's SWMR
+    /// reader reads it: reads end where the file ends at the time of each
+    /// read, and nothing is read as one slice. Any other file is read as
+    /// without `swmr`, bounded by its recorded end of file. Once the
+    /// superblock has been read, `flagged` says which it was.
+    fn new_storage(
+        storage: SharedStorage,
+        swmr: bool,
+        flagged: &mut Option<bool>,
+    ) -> Result<(Self, Superblock), Error> {
         let file_len = storage.len();
         let base = signature::find_signature_in(&*storage)?;
         let mut data = Self {
@@ -191,11 +209,16 @@ impl FileData {
             overlay: Vec::new(),
             image_error: None,
             contiguous: None,
+            // Until the superblock says otherwise: its own reads are not
+            // bounded by an end of file it has not read yet.
+            live: swmr,
         };
         // Worked out again below, once the end of file and any cache image
         // are known.
         data.contiguous = data.find_contiguous();
         let superblock = Superblock::parse_in(&data, 0)?;
+        data.live = swmr && superblock.version >= 3 && superblock.is_swmr_write();
+        *flagged = Some(data.live);
         data.end = base + superblock.data_end(base, file_len)?;
         match superblock_ext::cache_image_state_in(&data, &superblock)? {
             CacheImageState::Absent => {}
@@ -230,6 +253,9 @@ impl FileData {
 
     /// [`Self::contiguous`], worked out from `backing` and `patched`.
     fn find_contiguous(&self) -> Option<WholeView> {
+        if self.live {
+            return None;
+        }
         let bytes = self.compute_contiguous()?;
         Some(WholeView {
             ptr: bytes.as_ptr(),
@@ -295,6 +321,12 @@ impl Storage for FileData {
         if let Some(all) = self.contiguous() {
             return all.read_at(offset, len);
         }
+        if self.live {
+            // No end fixed at open: the storage's reads end where the file
+            // ends now.
+            let bytes = cut_to(self.remote()?.read_at(self.base + offset, len)?, len);
+            return Ok(self.with_overlay(offset, bytes));
+        }
         let size = self.end - self.base;
         let len = usize::try_from(size.saturating_sub(offset)).map_or(len, |avail| avail.min(len));
         if len == 0 {
@@ -305,6 +337,11 @@ impl Storage for FileData {
     }
 
     fn len(&self) -> u64 {
+        if self.live
+            && let Backing::Storage(s) = &self.backing
+        {
+            return s.len().saturating_sub(self.base);
+        }
         self.end - self.base
     }
 
@@ -312,7 +349,7 @@ impl Storage for FileData {
         if let Some(all) = self.contiguous() {
             return all.read_ranges(ranges);
         }
-        let size = self.end - self.base;
+        let size = Storage::len(self);
         let shifted: Vec<Range<u64>> = ranges
             .iter()
             .map(|r| {
@@ -375,6 +412,9 @@ pub struct File {
     /// Resolves external Virtual Dataset source files instead of
     /// `base_dir` (see [`File::set_vds_resolver`]).
     vds_resolver: Option<VdsResolver>,
+    /// Attempts per operation on a live file, and the retries made (see
+    /// [`File::open_swmr`]).
+    swmr: crate::swmr::Retries,
 }
 
 impl File {
@@ -394,6 +434,7 @@ impl File {
                 chunk_cache: ChunkCache::new(),
                 base_dir,
                 vds_resolver: None,
+                swmr: crate::swmr::Retries::default(),
             })
         }
         #[cfg(not(feature = "mmap"))]
@@ -428,6 +469,7 @@ impl File {
             chunk_cache: ChunkCache::new(),
             base_dir: None,
             vds_resolver: None,
+            swmr: crate::swmr::Retries::default(),
         })
     }
 
@@ -463,9 +505,205 @@ impl File {
             chunk_cache: ChunkCache::new(),
             base_dir: None,
             vds_resolver: None,
+            swmr: crate::swmr::Retries::default(),
         })
     }
 
+    /// Open a file that a libhdf5 SWMR writer (h5py `f.swmr_mode = True`)
+    /// may still be appending to, as libhdf5's SWMR reader does
+    /// (`H5F_ACC_SWMR_READ`, h5py `File(path, "r", swmr=True)`).
+    ///
+    /// The file is read with positioned reads ([`FileStorage`](crate::FileStorage)),
+    /// never mapped, and reads are bounded by the file's length at the
+    /// time of each read rather than a length fixed at open. The chunk cache
+    /// is not used, so every read reads the chunk index and chunks as they
+    /// are now. A [`Dataset`] handle keeps the extent it was opened (or
+    /// last [refreshed](Dataset::refresh)) with, as in libhdf5: call
+    /// [`Dataset::refresh`] to see the writer's appends.
+    ///
+    /// An operation that fails with an error a concurrent write can cause
+    /// — the failures libhdf5's SWMR reader retries: a checksum mismatch, a
+    /// read past the file's current end, an object header prefix (signature,
+    /// version) that does not decode — is run again from the start, up to
+    /// [`swmr_read_attempts`](Self::swmr_read_attempts) times (100 by
+    /// default, libhdf5's default for SWMR readers), with a pause of 1 µs
+    /// doubling up to 10 ms between attempts. Data is only returned from an
+    /// attempt in which every structure read verified, so a torn read is
+    /// an error (after the last attempt), never data. Every other error is
+    /// returned at once.
+    ///
+    /// All of this applies only to a file whose superblock (version 3) has
+    /// the SWMR-write flag set when it is opened: a file a SWMR writer has
+    /// open, or had and did not close. libhdf5's SWMR writer does not keep
+    /// the superblock's recorded end of file current, so for such a file it
+    /// is ignored, as libhdf5's SWMR reader ignores it. Any other file (one
+    /// whose writer has closed it, or that was never written in SWMR mode)
+    /// is read exactly as [`File::open`] reads it — bounded by its recorded
+    /// end of file, through the chunk cache, each operation tried once —
+    /// and [`is_swmr_read`](Self::is_swmr_read) is `false`. Which of the two
+    /// a handle is does not change after it is opened.
+    ///
+    /// A file whose superblock says it is open for writing without SWMR is
+    /// refused with [`Error::Locked`], as libhdf5 refuses it: such a writer
+    /// does not order its writes for readers.
+    pub fn open_swmr<P: AsRef<std::path::Path>>(path: P) -> Result<Self, Error> {
+        let storage = crate::swmr::FileStorage::open(path.as_ref()).map_err(Error::Io)?;
+        let mut f = Self::open_storage_swmr(Arc::new(storage))?;
+        f.base_dir = path.as_ref().parent().map(|p| p.to_path_buf());
+        Ok(f)
+    }
+
+    /// [`File::open_swmr`] over any [`Storage`] whose [`Storage::len`]
+    /// follows the file as it grows. A storage that caches blocks, or pins
+    /// the file's length at open (`clawhdf5-remote`'s `BlockCache` and
+    /// `HttpStorage`), does not show the writer's appends.
+    pub fn open_storage_swmr(storage: SharedStorage) -> Result<Self, Error> {
+        let swmr = crate::swmr::Retries::default();
+        // Retried only while the superblock cannot be read, or says a SWMR
+        // writer has the file: an error opening any other file is final.
+        let (data, superblock) = swmr.retry(|| {
+            let mut flagged = None;
+            match FileData::new_storage(storage.clone(), true, &mut flagged) {
+                Ok(opened) => Ok(Ok(opened)),
+                Err(e) if flagged == Some(false) => Ok(Err(e)),
+                Err(e) => Err(e),
+            }
+        })??;
+        if superblock.version >= 3 && superblock.is_write_access() && !superblock.is_swmr_write() {
+            return Err(Error::Locked(
+                "the file is open for writing without SWMR (libhdf5: \"file is already open \
+                 for write\"); a SWMR reader needs a SWMR writer"
+                    .into(),
+            ));
+        }
+        Ok(Self {
+            data,
+            superblock,
+            chunk_cache: ChunkCache::new(),
+            base_dir: None,
+            vds_resolver: None,
+            swmr,
+        })
+    }
+
+    /// Whether the file is read live: opened with [`File::open_swmr`] or
+    /// [`File::open_storage_swmr`] while its superblock had the SWMR-write
+    /// flag set. `false` for every other file, including one opened with
+    /// `open_swmr` whose writer had already closed it (it reads as
+    /// [`File::open`] reads it).
+    pub fn is_swmr_read(&self) -> bool {
+        self.data.live
+    }
+
+    /// How many times an operation on a SWMR-read file is tried (see
+    /// [`File::open_swmr`]); 100 unless set. Files not read live
+    /// ([`is_swmr_read`](Self::is_swmr_read) `false`) try every operation
+    /// once.
+    pub fn swmr_read_attempts(&self) -> u32 {
+        self.swmr.attempts
+    }
+
+    /// Set how many times an operation on a SWMR-read file is tried before
+    /// its error is returned (libhdf5's `H5Pset_metadata_read_attempts`); at
+    /// least 1.
+    pub fn set_swmr_read_attempts(&mut self, attempts: u32) {
+        self.swmr.attempts = attempts.max(1);
+    }
+
+    /// How many times an operation on this SWMR-read file has been run
+    /// again because a concurrent write made it fail (the counterpart of
+    /// libhdf5's `H5Fget_metadata_read_retry_info`), counting the attempts
+    /// made while opening it. Always 0 for other files.
+    pub fn swmr_retries(&self) -> u64 {
+        self.swmr.retries()
+    }
+
+    /// Whether a SWMR writer has the file open now, as far as the file
+    /// says: the superblock is read again and its SWMR-write flag returned.
+    /// libhdf5 clears the flag when the writer closes the file, so a reader
+    /// can stop following it then (and one last [`Dataset::refresh`] sees
+    /// the final extents). A writer that crashed or was killed never clears
+    /// it, so the flag alone is not a stop condition: give the loop another
+    /// one, such as a time without growth.
+    ///
+    /// ```no_run
+    /// # fn main() -> Result<(), clawhdf5::Error> {
+    /// use std::time::{Duration, Instant};
+    ///
+    /// let file = clawhdf5::File::open_swmr("live.h5")?;
+    /// let mut ds = file.dataset("samples")?;
+    /// let (mut seen, mut last_growth) = (0, Instant::now());
+    /// while file.swmr_writer_active()? && last_growth.elapsed() < Duration::from_secs(60) {
+    ///     ds.refresh()?;
+    ///     let n = ds.shape()?[0];
+    ///     if n > seen {
+    ///         // read rows seen..n ...
+    ///         (seen, last_growth) = (n, Instant::now());
+    ///     }
+    ///     std::thread::sleep(Duration::from_millis(100));
+    /// }
+    /// ds.refresh()?; // the final extent
+    /// # Ok(())
+    /// # }
+    /// ```
+    pub fn swmr_writer_active(&self) -> Result<bool, Error> {
+        self.retry(|| {
+            let sb = Superblock::parse_in(&self.data, 0)?;
+            Ok(sb.version >= 3 && sb.is_swmr_write())
+        })
+    }
+
+    /// Run `op`, again on errors a SWMR writer can cause when the file is
+    /// live (see [`File::open_swmr`]); once otherwise.
+    fn retry<T>(&self, op: impl FnMut() -> Result<T, Error>) -> Result<T, Error> {
+        if self.data.live {
+            self.swmr.retry(op)
+        } else {
+            let mut op = op;
+            op()
+        }
+    }
+
+    /// An attribute read under [`Self::retry`]. Attribute reads leave out
+    /// (or return as [`AttrValue::Raw`]) an attribute they cannot read
+    /// rather than fail, so `op` also returns the errors behind those: on a
+    /// live file one a concurrent write can cause runs the read again too.
+    /// If every attempt meets one, the last attempt's result is returned,
+    /// as it is for a file that is not live.
+    fn retry_attrs<T>(
+        &self,
+        mut op: impl FnMut() -> Result<(T, Vec<FormatError>), Error>,
+    ) -> Result<T, Error> {
+        let mut last = None;
+        let result = self.retry(|| {
+            last = None;
+            let (value, errors) = op()?;
+            match errors.into_iter().find(crate::swmr::is_transient_format) {
+                Some(e) if self.data.live => {
+                    last = Some(value);
+                    Err(Error::Format(e))
+                }
+                _ => Ok(value),
+            }
+        });
+        match (result, last) {
+            (Err(_), Some(value)) => Ok(value),
+            (result, _) => result,
+        }
+    }
+
+    /// The chunk cache reads of this file go through: the file's own, or
+    /// for a live file a fresh one per read, since a cached chunk index
+    /// would hide the writer's new chunks and a cached edge chunk would
+    /// read as fill where the writer has written since.
+    fn with_chunk_cache<T>(&self, op: impl FnOnce(&ChunkCache) -> T) -> T {
+        if self.data.live {
+            op(&ChunkCache::new())
+        } else {
+            op(&self.chunk_cache)
+        }
+    }
+
     /// Resolve external Virtual Dataset source files (their names as the
     /// mappings store them) with `resolver`, instead of reading them from
     /// the directory of the file (for [`File::open`]) or refusing them (for
@@ -488,6 +726,7 @@ impl File {
     ///
     /// The path uses `/` separators (e.g., `"group1/values"`).
     pub fn dataset(&self, path: &str) -> Result<Dataset<'_>, Error> {
+        self.retry(|| {
             let addr = with_bytes!(self.data.meta()?, |d| group_v2::resolve_path_any_in(
                 d,
                 &self.superblock,
@@ -499,9 +738,11 @@ impl File {
             }
             Dataset {
                 file: self,
+                address: addr,
                 header: hdr,
             }
             .check_open()
+        })
     }
 
     /// A `Dataset` handle for the object header at `address` (an address
@@ -511,15 +752,18 @@ impl File {
     /// are scanned); keep the address instead to open the same dataset
     /// repeatedly.
     pub fn dataset_at(&self, address: u64) -> Result<Dataset<'_>, Error> {
+        self.retry(|| {
             let hdr = self.parse_header(address)?;
             if !has_message(&hdr, MessageType::DataLayout) {
                 return Err(Error::NotADataset(format!("object at address {address}")));
             }
             Dataset {
                 file: self,
+                address,
                 header: hdr,
             }
             .check_open()
+        })
     }
 
     /// A `Group` handle for the object header at `address` (from
@@ -538,11 +782,11 @@ impl File {
     /// The path uses `/` separators (e.g., `"sensors"`).
     /// Use `"/"` or `""` for the root group.
     pub fn group(&self, path: &str) -> Result<Group<'_>, Error> {
-        let addr = with_bytes!(self.data.meta()?, |d| group_v2::resolve_path_any_in(
-            d,
-            &self.superblock,
-            path
-        ))?;
+        let addr = self.retry(|| {
+            Ok(with_bytes!(self.data.meta()?, |d| {
+                group_v2::resolve_path_any_in(d, &self.superblock, path)
+            })?)
+        })?;
         Ok(Group {
             file: self,
             address: addr,
@@ -665,6 +909,7 @@ impl File {
     /// [`AttrValue::Raw`] attribute. Variable-length strings are resolved in
     /// this file's global heap; see [`Dataset::read_string`] for the values.
     pub fn decode_strings(&self, datatype: &Datatype, raw: &[u8]) -> Result<Vec<String>, Error> {
+        self.retry(|| {
             with_bytes!(&self.data, |d| crate::vlen::decode_strings(
                 d,
                 datatype,
@@ -672,6 +917,7 @@ impl File {
                 self.offset_size(),
                 self.length_size(),
             ))
+        })
     }
 
     /// Like [`decode_strings`](Self::decode_strings) for variable-length
@@ -682,6 +928,7 @@ impl File {
         datatype: &Datatype,
         raw: &[u8],
     ) -> Result<Vec<Vec<u8>>, Error> {
+        self.retry(|| {
             with_bytes!(self.data.meta()?, |d| crate::vlen::decode_string_bytes(
                 d,
                 datatype,
@@ -689,6 +936,7 @@ impl File {
                 self.offset_size(),
                 self.length_size(),
             ))
+        })
     }
 
     /// Decode the variable-length sequences in `raw`, a buffer of elements
@@ -700,6 +948,7 @@ impl File {
         datatype: &Datatype,
         raw: &[u8],
     ) -> Result<Vec<Vec<T>>, Error> {
+        self.retry(|| {
             with_bytes!(self.data.meta()?, |d| crate::vlen::decode_vlen(
                 d,
                 datatype,
@@ -707,6 +956,7 @@ impl File {
                 self.offset_size(),
                 self.length_size(),
             ))
+        })
     }
 
     fn parse_header(&self, address: u64) -> Result<ObjectHeader, FormatError> {
@@ -750,6 +1000,7 @@ pub struct Group<'f> {
 impl<'f> Group<'f> {
     /// List the names of datasets in this group.
     pub fn datasets(&self) -> Result<Vec<String>, Error> {
+        self.file.retry(|| {
             let entries = self.children()?;
             let mut names = Vec::new();
             for entry in &entries {
@@ -759,10 +1010,12 @@ impl<'f> Group<'f> {
                 }
             }
             Ok(names)
+        })
     }
 
     /// List the names of subgroups in this group.
     pub fn groups(&self) -> Result<Vec<String>, Error> {
+        self.file.retry(|| {
             let entries = self.children()?;
             let mut names = Vec::new();
             for entry in &entries {
@@ -772,6 +1025,7 @@ impl<'f> Group<'f> {
                 }
             }
             Ok(names)
+        })
     }
 
     /// Read all attributes of this group.
@@ -792,26 +1046,31 @@ impl<'f> Group<'f> {
     pub fn attrs_with_errors(
         &self,
     ) -> Result<(HashMap<String, AttrValue>, Vec<FormatError>), Error> {
+        self.file.retry_attrs(|| {
             let hdr = self.file.parse_header(self.address)?;
-        with_bytes!(&self.file.data, |d| read_attrs(
-            d,
-            &hdr,
-            self.file.offset_size(),
-            self.file.length_size()
-        ))
+            let (attrs, errors, read_errors) = with_bytes!(&self.file.data, |d| {
+                read_attrs_reporting(d, &hdr, self.file.offset_size(), self.file.length_size())
+            })?;
+            let transient = errors.iter().chain(&read_errors).cloned().collect();
+            Ok(((attrs, errors), transient))
+        })
     }
 
     /// Get a dataset within this group by name.
     pub fn dataset(&self, name: &str) -> Result<Dataset<'f>, Error> {
-        let hdr = self.file.parse_header(self.child_address(name)?)?;
+        self.file.retry(|| {
+            let address = self.child_address(name)?;
+            let hdr = self.file.parse_header(address)?;
             if !has_message(&hdr, MessageType::DataLayout) {
                 return Err(Error::NotADataset(name.to_string()));
             }
             Dataset {
                 file: self.file,
+                address,
                 header: hdr,
             }
             .check_open()
+        })
     }
 
     /// Get a subgroup within this group by name.
@@ -827,14 +1086,16 @@ impl<'f> Group<'f> {
     /// that name, found without reading the other attributes when they are
     /// stored densely.
     pub fn attr(&self, name: &str) -> Result<Option<AttrValue>, Error> {
+        self.file.retry_attrs(|| {
             let hdr = self.file.parse_header(self.address)?;
-        with_bytes!(&self.file.data, |d| read_attr(
+            with_bytes!(&self.file.data, |d| read_attr_reporting(
                 d,
                 &hdr,
                 name,
                 self.file.offset_size(),
                 self.file.length_size(),
             ))
+        })
     }
 
     /// The object header address of the child called `name`: the entry of
@@ -842,6 +1103,7 @@ impl<'f> Group<'f> {
     /// name index rather than by listing the group (see
     /// [`group_v2::resolve_child`]).
     fn child_address(&self, name: &str) -> Result<u64, Error> {
+        self.file.retry(|| {
             with_bytes!(self.file.data.meta()?, |d| group_v2::resolve_child_in(
                 d,
                 &self.file.superblock,
@@ -849,6 +1111,7 @@ impl<'f> Group<'f> {
                 name
             ))
             .map_err(Error::Format)
+        })
     }
 
     /// This group's children that can be opened, as `(name, object header
@@ -869,10 +1132,12 @@ impl<'f> Group<'f> {
     /// [`group_v2::resolve_group_children`]); dangling, external and
     /// user-defined links are left out.
     fn children(&self) -> Result<Vec<GroupEntry>, Error> {
+        self.file.retry(|| {
             with_bytes!(self.file.data.meta()?, |d| {
                 group_v2::resolve_group_children_in(d, &self.file.superblock, self.address)
             })
             .map_err(Error::Format)
+        })
     }
 }
 
@@ -884,6 +1149,8 @@ impl<'f> Group<'f> {
 #[derive(Debug)]
 pub struct Dataset<'f> {
     file: &'f File,
+    /// Address of the object header, for [`Dataset::refresh`].
+    address: u64,
     header: ObjectHeader,
 }
 
@@ -906,6 +1173,31 @@ impl<'f> Dataset<'f> {
         Ok(self)
     }
 
+    /// Read the dataset's object header again, as libhdf5's `H5Drefresh`
+    /// (h5py `Dataset.refresh()`) does: afterwards [`shape`](Self::shape)
+    /// and every read use the dataset's extent as it is in the file now.
+    /// For a file a SWMR writer is appending to ([`File::open_swmr`]) this
+    /// is how a reader sees the appends; a transient failure is retried
+    /// (see there). On error the handle keeps its old header.
+    pub fn refresh(&mut self) -> Result<(), Error> {
+        let file = self.file;
+        let address = self.address;
+        let fresh = file.retry(|| {
+            let hdr = file.parse_header(address)?;
+            if !has_message(&hdr, MessageType::DataLayout) {
+                return Err(Error::NotADataset(format!("object at address {address}")));
+            }
+            Dataset {
+                file,
+                address,
+                header: hdr,
+            }
+            .check_open()
+        })?;
+        self.header = fresh.header;
+        Ok(())
+    }
+
     /// Returns the shape (dimensions) of the dataset.
     pub fn shape(&self) -> Result<Vec<u64>, Error> {
         let ds = self.dataspace()?;
@@ -938,9 +1230,11 @@ impl<'f> Dataset<'f> {
 
     /// Read all data as `f64` values.
     pub fn read_f64(&self) -> Result<Vec<f64>, Error> {
+        self.file.retry(|| {
             let dt = self.datatype()?;
-        // A contiguous dataset is converted straight from the file bytes; going
-        // through `read_raw` first copied the whole dataset an extra time.
+            // A contiguous dataset is converted straight from the file bytes;
+            // going through `read_raw` first copied the whole dataset an
+            // extra time.
             if let Ok(Some(bytes)) = self.contiguous_raw() {
                 return Ok(data_read::read_as_f64(&bytes, &dt)?);
             }
@@ -949,6 +1243,7 @@ impl<'f> Dataset<'f> {
             }
             let raw = self.read_raw()?;
             Ok(data_read::read_as_f64(&raw, &dt)?)
+        })
     }
 
     /// Zero-copy read of contiguous native-endian `f64` data.
@@ -958,9 +1253,11 @@ impl<'f> Dataset<'f> {
     ///
     /// Read all data as `f32` values.
     pub fn read_f32(&self) -> Result<Vec<f32>, Error> {
+        self.file.retry(|| {
             let dt = self.datatype()?;
-        // A contiguous dataset is converted straight from the file bytes; going
-        // through `read_raw` first copied the whole dataset an extra time.
+            // A contiguous dataset is converted straight from the file bytes;
+            // going through `read_raw` first copied the whole dataset an
+            // extra time.
             if let Ok(Some(bytes)) = self.contiguous_raw() {
                 return Ok(data_read::read_as_f32(&bytes, &dt)?);
             }
@@ -969,13 +1266,16 @@ impl<'f> Dataset<'f> {
             }
             let raw = self.read_raw()?;
             Ok(data_read::read_as_f32(&raw, &dt)?)
+        })
     }
 
     /// Read all data as `i32` values.
     pub fn read_i32(&self) -> Result<Vec<i32>, Error> {
+        self.file.retry(|| {
             let dt = self.datatype()?;
-        // A contiguous dataset is converted straight from the file bytes; going
-        // through `read_raw` first copied the whole dataset an extra time.
+            // A contiguous dataset is converted straight from the file bytes;
+            // going through `read_raw` first copied the whole dataset an
+            // extra time.
             if let Ok(Some(bytes)) = self.contiguous_raw() {
                 return Ok(data_read::read_as_i32(&bytes, &dt)?);
             }
@@ -984,13 +1284,16 @@ impl<'f> Dataset<'f> {
             }
             let raw = self.read_raw()?;
             Ok(data_read::read_as_i32(&raw, &dt)?)
+        })
     }
 
     /// Read all data as `i64` values.
     pub fn read_i64(&self) -> Result<Vec<i64>, Error> {
+        self.file.retry(|| {
             let dt = self.datatype()?;
-        // A contiguous dataset is converted straight from the file bytes; going
-        // through `read_raw` first copied the whole dataset an extra time.
+            // A contiguous dataset is converted straight from the file bytes;
+            // going through `read_raw` first copied the whole dataset an
+            // extra time.
             if let Ok(Some(bytes)) = self.contiguous_raw() {
                 return Ok(data_read::read_as_i64(&bytes, &dt)?);
             }
@@ -999,13 +1302,16 @@ impl<'f> Dataset<'f> {
             }
             let raw = self.read_raw()?;
             Ok(data_read::read_as_i64(&raw, &dt)?)
+        })
     }
 
     /// Read all data as `u64` values.
     pub fn read_u64(&self) -> Result<Vec<u64>, Error> {
+        self.file.retry(|| {
             let dt = self.datatype()?;
-        // A contiguous dataset is converted straight from the file bytes; going
-        // through `read_raw` first copied the whole dataset an extra time.
+            // A contiguous dataset is converted straight from the file bytes;
+            // going through `read_raw` first copied the whole dataset an
+            // extra time.
             if let Ok(Some(bytes)) = self.contiguous_raw() {
                 return Ok(data_read::read_as_u64(&bytes, &dt)?);
             }
@@ -1014,6 +1320,7 @@ impl<'f> Dataset<'f> {
             }
             let raw = self.read_raw()?;
             Ok(data_read::read_as_u64(&raw, &dt)?)
+        })
     }
 
     /// Read all data as `String` values, in row-major order.
@@ -1024,17 +1331,21 @@ impl<'f> Dataset<'f> {
     /// them; bytes that are not valid UTF-8 are replaced with U+FFFD — use
     /// [`read_string_bytes`](Self::read_string_bytes) for the exact bytes.
     pub fn read_string(&self) -> Result<Vec<String>, Error> {
+        self.file.retry(|| {
             let raw = self.read_raw()?;
             let dt = self.datatype()?;
             self.file.decode_strings(&dt, &raw)
+        })
     }
 
     /// Read a variable-length string dataset as the exact bytes of each
     /// string (what h5py's `Dataset[()]` returns), in row-major order.
     pub fn read_string_bytes(&self) -> Result<Vec<Vec<u8>>, Error> {
+        self.file.retry(|| {
             let raw = self.read_raw()?;
             let dt = self.datatype()?;
             self.file.decode_string_bytes(&dt, &raw)
+        })
     }
 
     /// Read the selected elements of a fixed- or variable-length string
@@ -1043,9 +1354,11 @@ impl<'f> Dataset<'f> {
         &self,
         selection: &clawhdf5_format::selection::Selection,
     ) -> Result<Vec<String>, Error> {
+        self.file.retry(|| {
             let raw = self.read_selection(selection)?;
             let dt = self.datatype()?;
             self.file.decode_strings(&dt, &raw)
+        })
     }
 
     /// Read a variable-length sequence dataset (h5py
@@ -1054,9 +1367,11 @@ impl<'f> Dataset<'f> {
     /// is converted to `T` as [`read_f64`](Self::read_f64) and the other
     /// typed readers convert. A null element is an empty sequence.
     pub fn read_vlen<T: crate::vlen::VlenValue>(&self) -> Result<Vec<Vec<T>>, Error> {
+        self.file.retry(|| {
             let raw = self.read_raw()?;
             let dt = self.datatype()?;
             self.file.decode_vlen(&dt, &raw)
+        })
     }
 
     /// Read the selected elements of a variable-length sequence dataset
@@ -1065,9 +1380,11 @@ impl<'f> Dataset<'f> {
         &self,
         selection: &clawhdf5_format::selection::Selection,
     ) -> Result<Vec<Vec<T>>, Error> {
+        self.file.retry(|| {
             let raw = self.read_selection(selection)?;
             let dt = self.datatype()?;
             self.file.decode_vlen(&dt, &raw)
+        })
     }
 
     // ----- Selection-based read methods -----
@@ -1096,6 +1413,15 @@ impl<'f> Dataset<'f> {
         if matches!(selection, clawhdf5_format::selection::Selection::All) {
             return self.read_raw();
         }
+        self.file.retry(|| self.read_selection_once(selection))
+    }
+
+    /// [`read_selection`](Self::read_selection) of a selection other than
+    /// `All`, once.
+    fn read_selection_once(
+        &self,
+        selection: &clawhdf5_format::selection::Selection,
+    ) -> Result<Vec<u8>, Error> {
         let dt = self.datatype()?;
         let ds = self.dataspace()?;
         let dl = self.data_layout()?;
@@ -1182,6 +1508,17 @@ impl<'f> Dataset<'f> {
         if matches!(selection, clawhdf5_format::selection::Selection::All) {
             return full();
         }
+        self.file
+            .retry(|| self.read_typed_selection_once(selection, convert))
+    }
+
+    /// [`read_typed_selection`](Self::read_typed_selection) of a selection
+    /// other than `All`, once.
+    fn read_typed_selection_once<T: data_read::NativeElement>(
+        &self,
+        selection: &clawhdf5_format::selection::Selection,
+        convert: fn(&[u8], &Datatype) -> Result<Vec<T>, FormatError>,
+    ) -> Result<Vec<T>, Error> {
         let dt = self.datatype()?;
         if T::is_native(&dt) && self.file.data.contiguous().is_some() {
             if let Ok(Some(raw)) = self.read_raw_ref() {
@@ -1445,12 +1782,18 @@ impl<'f> Dataset<'f> {
     pub fn attrs_with_errors(
         &self,
     ) -> Result<(HashMap<String, AttrValue>, Vec<FormatError>), Error> {
-        with_bytes!(&self.file.data, |d| read_attrs(
+        self.file.retry_attrs(|| {
+            let (attrs, errors, read_errors) = with_bytes!(&self.file.data, |d| {
+                read_attrs_reporting(
                     d,
                     &self.header,
                     self.file.offset_size(),
                     self.file.length_size(),
-        ))
+                )
+            })?;
+            let transient = errors.iter().chain(&read_errors).cloned().collect();
+            Ok(((attrs, errors), transient))
+        })
     }
 
     /// The attribute called `name`, or `None` if it has none by that name
@@ -1458,13 +1801,15 @@ impl<'f> Dataset<'f> {
     /// that name, found without reading the other attributes when they are
     /// stored densely.
     pub fn attr(&self, name: &str) -> Result<Option<AttrValue>, Error> {
-        with_bytes!(&self.file.data, |d| read_attr(
+        self.file.retry_attrs(|| {
+            with_bytes!(&self.file.data, |d| read_attr_reporting(
                 d,
                 &self.header,
                 name,
                 self.file.offset_size(),
                 self.file.length_size(),
             ))
+        })
     }
 
     /// Verify this dataset's content against its stored provenance hash
@@ -1484,12 +1829,14 @@ impl<'f> Dataset<'f> {
     /// result is not a tamper-evidence or authenticity guarantee.
     #[cfg(feature = "provenance")]
     pub fn verify_provenance(&self) -> Result<clawhdf5_format::provenance::VerifyResult, Error> {
+        self.file.retry(|| {
             Ok(clawhdf5_format::provenance::verify_dataset_in(
                 &self.file.data,
                 &self.header,
                 self.file.offset_size(),
                 self.file.length_size(),
             )?)
+        })
     }
 
     /// A header message's payload, resolved through the shared-message
@@ -1499,11 +1846,10 @@ impl<'f> Dataset<'f> {
         &self,
         msg_type: MessageType,
     ) -> Result<Option<std::borrow::Cow<'_, [u8]>>, Error> {
-        self.header
-            .messages
-            .iter()
-            .find(|m| m.msg_type == msg_type)
-            .map(|msg| {
+        let msg = self.header.messages.iter().find(|m| m.msg_type == msg_type);
+        // A shared message is read from the header it lives in.
+        self.file.retry(|| {
+            msg.map(|msg| {
                 with_bytes!(&self.file.data, |d| {
                     clawhdf5_format::shared_message::message_data_in(
                         d,
@@ -1515,6 +1861,7 @@ impl<'f> Dataset<'f> {
                 .map_err(Error::Format)
             })
             .transpose()
+        })
     }
 
     fn required_payload(&self, msg_type: MessageType) -> Result<std::borrow::Cow<'_, [u8]>, Error> {
@@ -1587,6 +1934,8 @@ impl<'f> Dataset<'f> {
         }
         let ds = self.dataspace()?;
         let pipeline = self.filter_pipeline()?;
+        self.file.retry(|| {
+            self.file.with_chunk_cache(|cache| {
                 Ok(data_read::read_chunked_native_in::<T, _>(
                     &self.header.messages,
                     &self.file.data,
@@ -1596,11 +1945,17 @@ impl<'f> Dataset<'f> {
                     pipeline.as_ref(),
                     self.file.offset_size(),
                     self.file.length_size(),
-            Some(&self.file.chunk_cache),
+                    Some(cache),
                 )?)
+            })
+        })
     }
 
     fn read_raw(&self) -> Result<Vec<u8>, Error> {
+        self.file.retry(|| self.read_raw_once())
+    }
+
+    fn read_raw_once(&self) -> Result<Vec<u8>, Error> {
         let dt = self.datatype()?;
         let ds = self.dataspace()?;
         let dl = self.data_layout()?;
@@ -1622,6 +1977,7 @@ impl<'f> Dataset<'f> {
             self.file.offset_size(),
             self.file.length_size(),
             || {
+                self.file.with_chunk_cache(|cache| {
                     Ok(data_read::read_raw_data_cached_in(
                         &self.file.data,
                         &dl,
@@ -1630,8 +1986,9 @@ impl<'f> Dataset<'f> {
                         pipeline.as_ref(),
                         self.file.offset_size(),
                         self.file.length_size(),
-                    &self.file.chunk_cache,
+                        cache,
                     )?)
+                })
             },
         )
     }

crates/clawhdf5/src/swmr.rs

@@ -0,0 +1,362 @@
+//! Reading files a SWMR writer is still appending to (see
+//! `docs/design/swmr.md` and [`File::open_swmr`](crate::File::open_swmr)):
+//! a [`FileStorage`] whose length grows with the file, and the bounded
+//! retries of operations a concurrent write can make fail.
+
+use std::borrow::Cow;
+use std::cell::Cell;
+use std::path::Path;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::Duration;
+
+use clawhdf5_format::error::FormatError;
+use clawhdf5_format::storage::Storage;
+
+use crate::error::Error;
+
+/// How many times a live file ([`File::open_swmr`](crate::File::open_swmr))
+/// tries an operation that fails with an error a concurrent write can cause
+/// before returning the error: 100, libhdf5's default number of metadata
+/// read attempts for SWMR access (`H5Pset_metadata_read_attempts`).
+pub const SWMR_READ_ATTEMPTS: u32 = 100;
+
+/// Longest pause between two attempts.
+const MAX_PAUSE: Duration = Duration::from_millis(10);
+
+/// A local file read with positioned reads (`pread` on Unix, `seek_read` on
+/// Windows), never mapped, whose [`Storage::len`] is the file's length at
+/// the time of the call: a [`Storage`] for a file that another process is
+/// appending to. A read past the end is short, as the trait allows.
+#[derive(Debug)]
+pub struct FileStorage {
+    file: std::fs::File,
+    #[cfg(not(any(unix, windows)))]
+    lock: std::sync::Mutex<()>,
+}
+
+impl FileStorage {
+    /// Open the file at `path` for reading.
+    pub fn open<P: AsRef<Path>>(path: P) -> std::io::Result<Self> {
+        Ok(Self::new(std::fs::File::open(path)?))
+    }
+
+    /// A storage over an open file.
+    pub fn new(file: std::fs::File) -> Self {
+        Self {
+            file,
+            #[cfg(not(any(unix, windows)))]
+            lock: std::sync::Mutex::new(()),
+        }
+    }
+
+    /// Up to `buf.len()` bytes at `offset`; fewer only at the end of file.
+    fn read_into(&self, offset: u64, buf: &mut [u8]) -> std::io::Result<usize> {
+        let mut got = 0;
+        while got < buf.len() {
+            match self.read_once(offset + got as u64, &mut buf[got..]) {
+                Ok(0) => break,
+                Ok(n) => got += n,
+                Err(e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) => return Err(e),
+            }
+        }
+        Ok(got)
+    }
+
+    #[cfg(unix)]
+    fn read_once(&self, offset: u64, buf: &mut [u8]) -> std::io::Result<usize> {
+        std::os::unix::fs::FileExt::read_at(&self.file, buf, offset)
+    }
+
+    #[cfg(windows)]
+    fn read_once(&self, offset: u64, buf: &mut [u8]) -> std::io::Result<usize> {
+        std::os::windows::fs::FileExt::seek_read(&self.file, buf, offset)
+    }
+
+    #[cfg(not(any(unix, windows)))]
+    fn read_once(&self, offset: u64, buf: &mut [u8]) -> std::io::Result<usize> {
+        use std::io::{Read, Seek, SeekFrom};
+        let _guard = self.lock.lock().unwrap_or_else(|p| p.into_inner());
+        let mut f = &self.file;
+        f.seek(SeekFrom::Start(offset))?;
+        f.read(buf)
+    }
+}
+
+impl Storage for FileStorage {
+    fn read_at(&self, offset: u64, len: usize) -> Result<Cow<'_, [u8]>, FormatError> {
+        // Never allocate more than the file holds, whatever a (possibly
+        // hostile) size field asked for.
+        let avail = self.len().saturating_sub(offset);
+        let len = usize::try_from(avail).map_or(len, |a| a.min(len));
+        let mut buf = vec![0u8; len];
+        let got = self
+            .read_into(offset, &mut buf)
+            .map_err(|e| FormatError::Storage(format!("read of {len} bytes at {offset}: {e}")))?;
+        buf.truncate(got);
+        Ok(Cow::Owned(buf))
+    }
+
+    fn len(&self) -> u64 {
+        self.file.metadata().map_or(0, |m| m.len())
+    }
+}
+
+/// Whether `e` can be caused by reading a structure while a SWMR writer
+/// rewrites it or has not yet written it, so that the operation is worth
+/// running again (see [`is_transient_format`]).
+pub(crate) fn is_transient(e: &Error) -> bool {
+    match e {
+        Error::Format(f) => is_transient_format(f),
+        Error::Io(io) => io.kind() == std::io::ErrorKind::UnexpectedEof,
+        _ => false,
+    }
+}
+
+/// The failures libhdf5's SWMR reader retries, and nothing else. libhdf5
+/// (`H5C__load_entry`) reads a metadata structure again when its checksum
+/// fails, and when the prefix it decodes before the checksum to learn the
+/// structure's size does not decode (for an object header, its signature
+/// and version: a header whose every byte is garbled fails there). A read
+/// past the file's current end is short here; libhdf5 reads zeros there,
+/// which then fail the checksum. So:
+///
+/// - [`FormatError::ChecksumMismatch`], of any checksummed structure;
+/// - [`FormatError::UnexpectedEof`], a read past the current end;
+/// - [`FormatError::InvalidObjectHeaderSignature`] and
+///   [`FormatError::InvalidObjectHeaderVersion`], the object header prefix.
+///
+/// Every other error — an unsupported version or message, a file that is not
+/// HDF5, a structure that is corrupt behind a valid checksum — is returned
+/// at once: a concurrent write does not cause it, and retrying it only
+/// costs up to a second of pauses.
+pub(crate) fn is_transient_format(e: &FormatError) -> bool {
+    use FormatError as F;
+    matches!(
+        e,
+        F::ChecksumMismatch { .. }
+            | F::UnexpectedEof { .. }
+            | F::InvalidObjectHeaderSignature
+            | F::InvalidObjectHeaderVersion(_)
+    )
+}
+
+thread_local! {
+    /// Set while this thread runs an operation under [`retry`], so an
+    /// operation made of retried operations retries as a whole, not each
+    /// part up to the limit.
+    static RETRYING: Cell<bool> = const { Cell::new(false) };
+}
+
+/// A live file's retry policy: attempts per operation, and a count of the
+/// retries made.
+#[derive(Debug)]
+pub(crate) struct Retries {
+    pub(crate) attempts: u32,
+    retried: AtomicU64,
+}
+
+impl Default for Retries {
+    fn default() -> Self {
+        Self {
+            attempts: SWMR_READ_ATTEMPTS,
+            retried: AtomicU64::new(0),
+        }
+    }
+}
+
+impl Retries {
+    /// [`retry`] with this policy, counting the retries.
+    pub(crate) fn retry<T>(&self, op: impl FnMut() -> Result<T, Error>) -> Result<T, Error> {
+        retry(self.attempts, &self.retried, op)
+    }
+
+    /// Retries made so far.
+    pub(crate) fn retries(&self) -> u64 {
+        self.retried.load(Ordering::Relaxed)
+    }
+}
+
+/// Run `op` up to `attempts` times while it fails with a transient error
+/// ([`is_transient`]), pausing 1 µs, 2 µs, 4 µs, … up to 10 ms between
+/// attempts, and adding each retry to `retried`; return its first success
+/// or last error. Inside another `retry` on the same thread, `op` runs
+/// once.
+fn retry<T>(
+    attempts: u32,
+    retried: &AtomicU64,
+    mut op: impl FnMut() -> Result<T, Error>,
+) -> Result<T, Error> {
+    if RETRYING.with(Cell::get) {
+        return op();
+    }
+    struct Reset;
+    impl Drop for Reset {
+        fn drop(&mut self) {
+            RETRYING.with(|r| r.set(false));
+        }
+    }
+    RETRYING.with(|r| r.set(true));
+    let _reset = Reset;
+    let mut pause = Duration::from_micros(1);
+    let mut attempt = 1;
+    loop {
+        match op() {
+            Err(e) if attempt < attempts && is_transient(&e) => {
+                retried.fetch_add(1, Ordering::Relaxed);
+                std::thread::sleep(pause);
+                pause = (pause * 2).min(MAX_PAUSE);
+                attempt += 1;
+            }
+            result => return result,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn file_storage_reads_what_the_file_holds_now() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("grow.bin");
+        std::fs::write(&path, b"hello").unwrap();
+        let s = FileStorage::open(&path).unwrap();
+        assert_eq!(s.len(), 5);
+        assert_eq!(&*s.read_at(1, 3).unwrap(), b"ell");
+        assert_eq!(&*s.read_at(3, 10).unwrap(), b"lo");
+        assert!(s.read_at(9, 4).unwrap().is_empty());
+        // The file grows after the storage was opened.
+        use std::io::Write;
+        std::fs::OpenOptions::new()
+            .append(true)
+            .open(&path)
+            .unwrap()
+            .write_all(b", world")
+            .unwrap();
+        assert_eq!(s.len(), 12);
+        assert_eq!(&*s.read_at(3, 100).unwrap(), b"lo, world");
+    }
+
+    #[test]
+    fn permanent_errors_are_returned_at_once() {
+        // Errors a concurrent write does not cause: none is retried, so
+        // none costs the pauses (about 0.9 s for 100 attempts).
+        let permanent = [
+            FormatError::SignatureNotFound,
+            FormatError::UnsupportedVersion(9),
+            FormatError::UnsupportedMessage(0x99),
+            FormatError::TruncatedFile {
+                stored_eof: 10,
+                actual_len: 5,
+            },
+            FormatError::InvalidDatatypeClass(15),
+            FormatError::InvalidLayoutVersion(9),
+            FormatError::InvalidBTreeSignature,
+            FormatError::ChunkedReadError("x".into()),
+            FormatError::DecompressionError("x".into()),
+            FormatError::Fletcher32Mismatch {
+                expected: 1,
+                computed: 2,
+            },
+        ];
+        let n = AtomicU64::new(0);
+        let started = std::time::Instant::now();
+        for e in permanent {
+            assert!(!is_transient_format(&e), "{e:?}");
+            let mut calls = 0;
+            let r: Result<(), Error> = retry(SWMR_READ_ATTEMPTS, &n, || {
+                calls += 1;
+                Err(Error::Format(e.clone()))
+            });
+            assert!(r.is_err());
+            assert_eq!(calls, 1, "{e:?}");
+        }
+        assert_eq!(n.load(Ordering::Relaxed), 0);
+        assert!(started.elapsed() < Duration::from_millis(50));
+        assert!(!is_transient(&Error::Io(std::io::Error::other("x"))));
+
+        // The transient ones are retried to the limit.
+        for e in [
+            FormatError::ChecksumMismatch {
+                expected: 1,
+                computed: 2,
+            },
+            FormatError::UnexpectedEof {
+                expected: 8,
+                available: 0,
+            },
+            FormatError::InvalidObjectHeaderSignature,
+            FormatError::InvalidObjectHeaderVersion(0x4f),
+        ] {
+            let mut calls = 0;
+            let _: Result<(), Error> = retry(3, &n, || {
+                calls += 1;
+                Err(Error::Format(e.clone()))
+            });
+            assert_eq!(calls, 3, "{e:?}");
+        }
+        assert!(is_transient(&Error::Io(
+            std::io::ErrorKind::UnexpectedEof.into()
+        )));
+    }
+
+    #[test]
+    fn retry_runs_again_only_for_transient_errors() {
+        let n = AtomicU64::new(0);
+        let mut calls = 0;
+        let r: Result<u32, Error> = retry(5, &n, || {
+            calls += 1;
+            if calls < 3 {
+                Err(Error::Format(FormatError::ChecksumMismatch {
+                    expected: 1,
+                    computed: 2,
+                }))
+            } else {
+                Ok(7)
+            }
+        });
+        assert_eq!(r.unwrap(), 7);
+        assert_eq!(calls, 3);
+        assert_eq!(n.load(Ordering::Relaxed), 2);
+
+        // Gives up after `attempts`.
+        let mut calls = 0;
+        let r: Result<(), Error> = retry(4, &n, || {
+            calls += 1;
+            Err(Error::Format(FormatError::UnexpectedEof {
+                expected: 8,
+                available: 0,
+            }))
+        });
+        assert!(r.is_err());
+        assert_eq!(calls, 4);
+        assert_eq!(n.load(Ordering::Relaxed), 5);
+
+        // A permanent error is returned at once.
+        let mut calls = 0;
+        let r: Result<(), Error> = retry(4, &n, || {
+            calls += 1;
+            Err(Error::Format(FormatError::UnsupportedFilter(999)))
+        });
+        assert!(r.is_err());
+        assert_eq!(calls, 1);
+        assert_eq!(n.load(Ordering::Relaxed), 5);
+
+        // Nested: the inner operation runs once per outer attempt.
+        let mut inner = 0;
+        let mut outer = 0;
+        let r: Result<(), Error> = retry(3, &n, || {
+            outer += 1;
+            retry(3, &n, || {
+                inner += 1;
+                Err(Error::Format(FormatError::InvalidObjectHeaderVersion(0x76)))
+            })
+        });
+        assert!(r.is_err());
+        assert_eq!((outer, inner), (3, 3));
+        // The flag is reset afterwards.
+        assert!(!RETRYING.with(Cell::get));
+    }
+}

crates/clawhdf5/src/types.rs

@@ -170,16 +170,38 @@ pub(crate) fn read_attrs<S: clawhdf5_format::storage::Storage + ?Sized>(
     ),
     crate::Error,
 > {
+    read_attrs_reporting(file_data, header, offset_size, length_size)
+        .map(|(attrs, errors, _)| (attrs, errors))
+}
+
+/// What [`read_attrs_reporting`] returns: the attributes, the errors of
+/// those left out, and the errors behind values returned as
+/// [`AttrValue::Raw`].
+pub(crate) type AttrsReport = (
+    HashMap<String, AttrValue>,
+    Vec<clawhdf5_format::error::FormatError>,
+    Vec<clawhdf5_format::error::FormatError>,
+);
+
+/// [`read_attrs`], also returning the errors of reading the file for a
+/// value (a variable-length string's global heap) that was returned as
+/// [`AttrValue::Raw`] instead: a live file retries on them (see
+/// `File::open_swmr`).
+pub(crate) fn read_attrs_reporting<S: clawhdf5_format::storage::Storage + ?Sized>(
+    file_data: &S,
+    header: &clawhdf5_format::object_header::ObjectHeader,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<AttrsReport, crate::Error> {
     let (msgs, errors) = clawhdf5_format::attribute::extract_attributes_tolerant_in(
         file_data,
         header,
         offset_size,
         length_size,
     )?;
-    Ok((
-        attrs_to_map(&msgs, file_data, offset_size, length_size),
-        errors,
-    ))
+    let mut read_errors = Vec::new();
+    let map = attrs_to_map_reporting(&msgs, file_data, offset_size, length_size, &mut read_errors);
+    Ok((map, errors, read_errors))
 }
 
 /// The attribute called `name` on the object with header `header`, decoded
@@ -192,30 +214,50 @@ pub(crate) fn read_attr<S: clawhdf5_format::storage::Storage + ?Sized>(
     offset_size: u8,
     length_size: u8,
 ) -> Result<Option<AttrValue>, crate::Error> {
-    let Some(msg) = clawhdf5_format::attribute::find_attribute_in(
+    read_attr_reporting(file_data, header, name, offset_size, length_size).map(|(v, _)| v)
+}
+
+/// [`read_attr`], also returning the errors of the attributes it could not
+/// read on the way (the one asked for may be among them) and of reading the
+/// file for its value, if one made it [`AttrValue::Raw`] (see
+/// [`read_attrs_reporting`]).
+pub(crate) fn read_attr_reporting<S: clawhdf5_format::storage::Storage + ?Sized>(
+    file_data: &S,
+    header: &clawhdf5_format::object_header::ObjectHeader,
+    name: &str,
+    offset_size: u8,
+    length_size: u8,
+) -> Result<(Option<AttrValue>, Vec<clawhdf5_format::error::FormatError>), crate::Error> {
+    let (found, mut read_errors) = clawhdf5_format::attribute::find_attribute_reporting_in(
         file_data,
         header,
         name,
         offset_size,
         length_size,
-    )?
-    else {
-        return Ok(None);
+    )?;
+    let Some(msg) = found else {
+        return Ok((None, read_errors));
     };
-    Ok(attrs_to_map(
+    let value = attrs_to_map_reporting(
         std::slice::from_ref(&msg),
         file_data,
         offset_size,
         length_size,
+        &mut read_errors,
     )
-    .remove(name))
+    .remove(name);
+    Ok((value, read_errors))
 }
 
-pub(crate) fn attrs_to_map<S: clawhdf5_format::storage::Storage + ?Sized>(
+/// The attributes `attrs` by name, decoded; an error reading the file for
+/// a value that is returned as [`AttrValue::Raw`] because of it is added
+/// to `read_errors`.
+pub(crate) fn attrs_to_map_reporting<S: clawhdf5_format::storage::Storage + ?Sized>(
     attrs: &[clawhdf5_format::attribute::AttributeMessage],
     file_data: &S,
     offset_size: u8,
     length_size: u8,
+    read_errors: &mut Vec<clawhdf5_format::error::FormatError>,
 ) -> HashMap<String, AttrValue> {
     let mut map = HashMap::new();
     for attr in attrs {
@@ -224,13 +266,11 @@ pub(crate) fn attrs_to_map<S: clawhdf5_format::storage::Storage + ?Sized>(
         // verbatim as `AttrValue::Raw` rather than dropped — a partial
         // attribute list with no indication anything is missing is worse than
         // an undecoded value.
-        let val =
-            decode_attr_value(attr, file_data, offset_size, length_size).unwrap_or_else(|| {
-                AttrValue::Raw {
+        let val = decode_attr_value(attr, file_data, offset_size, length_size, read_errors)
+            .unwrap_or_else(|| AttrValue::Raw {
                 datatype: attr.datatype.clone(),
                 shape: attr.dataspace.dimensions.clone(),
                 data: attr.raw_data.clone(),
-                }
             });
         map.insert(attr.name.clone(), val);
     }
@@ -271,6 +311,7 @@ fn decode_attr_value<S: clawhdf5_format::storage::Storage + ?Sized>(
     file_data: &S,
     offset_size: u8,
     length_size: u8,
+    read_errors: &mut Vec<clawhdf5_format::error::FormatError>,
 ) -> Option<AttrValue> {
     use clawhdf5_format::datatype::Datatype;
 
@@ -310,9 +351,13 @@ fn decode_attr_value<S: clawhdf5_format::storage::Storage + ?Sized>(
         Datatype::VariableLength {
             is_string: true, ..
         } => {
-            let strings = attr
-                .read_vl_strings_in(file_data, offset_size, length_size)
-                .ok()?;
+            let strings = match attr.read_vl_strings_in(file_data, offset_size, length_size) {
+                Ok(strings) => strings,
+                Err(e) => {
+                    read_errors.push(e);
+                    return None;
+                }
+            };
             if strings.len() == 1 {
                 Some(AttrValue::String(strings[0].clone()))
             } else {

crates/clawhdf5/tests/swmr_interop.rs

@@ -0,0 +1,847 @@
+//! Files a libhdf5 SWMR writer (h5py `f.swmr_mode = True`) has open: a copy
+//! taken mid-write (fixture), and a live file appended to by an h5py writer
+//! process while clawhdf5 and h5py's own SWMR reader read it (see
+//! `docs/design/swmr.md`).
+//!
+//! The live tests need python3 with h5py; they are skipped without it,
+//! unless `CLAWHDF5_REQUIRE_INTEROP=1`.
+
+use std::io::{BufRead, BufReader, Read};
+use std::path::{Path, PathBuf};
+use std::process::{Command, Stdio};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicU32, Ordering};
+
+use clawhdf5::{File, Selection};
+
+fn python() -> String {
+    std::env::var("CLAWHDF5_PYTHON").unwrap_or_else(|_| "python3".to_string())
+}
+
+fn interop_required() -> bool {
+    std::env::var("CLAWHDF5_REQUIRE_INTEROP").is_ok_and(|v| v == "1")
+}
+
+fn python_available() -> bool {
+    Command::new(python())
+        .args(["-c", "import h5py, numpy"])
+        .output()
+        .map(|o| o.status.success())
+        .unwrap_or(false)
+}
+
+macro_rules! skip_if_no_python {
+    () => {
+        if !python_available() {
+            assert!(
+                !interop_required(),
+                "CLAWHDF5_REQUIRE_INTEROP=1 but python3 with h5py is not available"
+            );
+            eprintln!("SKIP: python3 with h5py not available");
+            return;
+        }
+    };
+}
+
+fn fixture(name: &str) -> PathBuf {
+    Path::new(env!("CARGO_MANIFEST_DIR"))
+        .join("tests/fixtures")
+        .join(name)
+}
+
+/// `swmr_mid_write.h5`: a copy h5py 3.16 (HDF5 2.0) made of its own file
+/// while writing it in SWMR mode, after 4 appends of 37 rows to
+/// `/a` (int64, chunks of 100, no filter: `a[i] = i + 1`) and `/b` (float64
+/// `(n, 4)`, chunks of 16 x 4, gzip: `b[i, j] = 10 i + j + 1`), each
+/// followed by a flush. Its superblock (v3) still has the SWMR-write flag
+/// set and records an end of file of 715 in a 6 030-byte file.
+fn check_mid_write_copy(f: &File) {
+    let sb = f.superblock();
+    assert_eq!(sb.version, 3);
+    assert!(sb.is_swmr_write());
+    let a = f.dataset("a").unwrap();
+    assert_eq!(a.shape().unwrap(), vec![148]);
+    let want_a: Vec<i64> = (1..=148).collect();
+    assert_eq!(a.read_i64().unwrap(), want_a);
+    let b = f.dataset("b").unwrap();
+    assert_eq!(b.shape().unwrap(), vec![148, 4]);
+    let want_b: Vec<f64> = (0..148)
+        .flat_map(|i| (0..4).map(move |j| (10 * i + j + 1) as f64))
+        .collect();
+    assert_eq!(b.read_f64().unwrap(), want_b);
+}
+
+#[test]
+fn a_copy_taken_mid_write_reads_past_its_recorded_end_of_file() {
+    let path = fixture("swmr_mid_write.h5");
+    // The recorded end of file (715) is far below the file's length; the
+    // chunk indexes and chunks lie past it. libhdf5's SWMR reader does not
+    // bound reads by it, and neither does any open path here.
+    check_mid_write_copy(&File::open(&path).unwrap());
+    check_mid_write_copy(&File::open_buffered(&path).unwrap());
+    check_mid_write_copy(&File::from_bytes(std::fs::read(&path).unwrap()).unwrap());
+    let want_a: Vec<i64> = (1..=148).collect();
+    let want_b: Vec<f64> = (0..148)
+        .flat_map(|i| (0..4).map(move |j| (10 * i + j + 1) as f64))
+        .collect();
+    let mm = clawhdf5::MmapFile::open(&path).unwrap();
+    assert_eq!(mm.dataset("a").unwrap().read_i64().unwrap(), want_a);
+    assert_eq!(mm.dataset("b").unwrap().read_f64().unwrap(), want_b);
+    let lazy = clawhdf5::LazyFile::open_mmap(&path).unwrap();
+    assert_eq!(lazy.dataset("a").unwrap().read_i64().unwrap(), want_a);
+    assert_eq!(lazy.dataset("b").unwrap().read_f64().unwrap(), want_b);
+    let storage = File::open_storage(Arc::new(std::fs::read(&path).unwrap())).unwrap();
+    check_mid_write_copy(&storage);
+}
+
+#[test]
+fn a_copy_taken_mid_write_reads_as_h5py_swmr_reader_reads_it() {
+    skip_if_no_python!();
+    let path = fixture("swmr_mid_write.h5");
+    // libhdf5 refuses a non-SWMR open of this file ("file is already open
+    // for write"); its SWMR reader reads the values checked above.
+    let script = format!(
+        r#"
+import h5py, numpy as np
+with h5py.File("{p}", "r", swmr=True, locking=False) as f:
+    a = f["a"][()]
+    b = f["b"][()]
+assert np.array_equal(a, np.arange(148) + 1), a
+assert np.array_equal(b, (np.arange(148)[:, None] * 10 + np.arange(4) + 1).astype("f8")), b
+print("ok")
+"#,
+        p = path.display()
+    );
+    let out = Command::new(python())
+        .args(["-c", &script])
+        .output()
+        .unwrap();
+    assert!(
+        out.status.success(),
+        "h5py failed:\n{}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+    check_mid_write_copy(&File::open(&path).unwrap());
+}
+
+#[test]
+fn open_swmr_reads_the_mid_write_copy() {
+    let f = File::open_swmr(fixture("swmr_mid_write.h5")).unwrap();
+    assert!(f.is_swmr_read());
+    assert_eq!(f.swmr_read_attempts(), clawhdf5::SWMR_READ_ATTEMPTS);
+    // The copy still has the SWMR-write flag: as far as it says, its writer
+    // is still writing.
+    assert!(f.swmr_writer_active().unwrap());
+    check_mid_write_copy(&f);
+    let mut a = f.dataset("a").unwrap();
+    a.refresh().unwrap();
+    assert_eq!(a.shape().unwrap(), vec![148]);
+    assert_eq!(f.swmr_retries(), 0);
+    assert!(
+        !File::open(fixture("swmr_mid_write.h5"))
+            .unwrap()
+            .is_swmr_read()
+    );
+}
+
+/// The mid-write copy with its superblock's flags set to `flags` (and the
+/// superblock checksum updated).
+fn mid_write_copy_with_flags(flags: u8) -> Vec<u8> {
+    let mut bytes = std::fs::read(fixture("swmr_mid_write.h5")).unwrap();
+    // Superblock v3, 8-byte offsets: 12 bytes, 4 addresses, checksum.
+    bytes[11] = flags;
+    let sum = clawhdf5_format::checksum::jenkins_lookup3(&bytes[..44]);
+    bytes[44..48].copy_from_slice(&sum.to_le_bytes());
+    bytes
+}
+
+#[test]
+fn open_swmr_refuses_a_file_open_for_writing_without_swmr() {
+    // libhdf5 refuses it too: "file is already open for write".
+    let bytes = mid_write_copy_with_flags(0x01);
+    let err = File::open_storage_swmr(Arc::new(bytes)).unwrap_err();
+    assert!(matches!(err, clawhdf5::Error::Locked(_)), "{err}");
+    // Closed (flags clear): opens, and no writer is active.
+    let f = File::open_storage_swmr(Arc::new(mid_write_copy_with_flags(0))).unwrap();
+    assert!(!f.swmr_writer_active().unwrap());
+}
+
+/// The outcome of reading every dataset of `f` in full, as text: the
+/// values, or the error.
+fn read_all(f: &File) -> Vec<String> {
+    ["a", "b"]
+        .iter()
+        .map(|n| match f.dataset(n).and_then(|d| d.read_f64()) {
+            Ok(v) => format!("{n}: {} values", v.len()),
+            Err(e) => format!("{n}: error {e}"),
+        })
+        .collect()
+}
+
+#[test]
+fn open_swmr_reads_a_file_without_the_swmr_write_flag_as_file_open_does() {
+    // The mid-write copy with its flags cleared: a closed file that
+    // records an end of file of 715 in 6 030 bytes, its chunk indexes past
+    // that end. File::open (and libhdf5's plain reader) refuse what lies
+    // past the recorded end; open_swmr must too, and must not retry (the
+    // file is not live). See the h5py test below for libhdf5's SWMR reader.
+    let bytes = mid_write_copy_with_flags(0);
+    let dir = tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
+    let path = dir.path().join("closed_short_eof.h5");
+    std::fs::write(&path, &bytes).unwrap();
+
+    let plain = File::open(&path).unwrap();
+    let swmr = File::open_swmr(&path).unwrap();
+    let swmr_storage = File::open_storage_swmr(Arc::new(bytes)).unwrap();
+    let want = read_all(&plain);
+    assert!(
+        want.iter().all(|r| r.contains("error")),
+        "File::open reads past the recorded end: {want:?}"
+    );
+    for f in [&swmr, &swmr_storage] {
+        assert!(!f.is_swmr_read());
+        assert!(!f.swmr_writer_active().unwrap());
+        let started = std::time::Instant::now();
+        assert_eq!(read_all(f), want);
+        assert!(started.elapsed() < std::time::Duration::from_millis(100));
+        assert_eq!(f.swmr_retries(), 0);
+    }
+
+    // The same bytes with the SWMR-write flag (and write access) set are
+    // read live, past the recorded end.
+    let live = File::open_storage_swmr(Arc::new(mid_write_copy_with_flags(0x05))).unwrap();
+    assert!(live.is_swmr_read());
+    check_mid_write_copy(&live);
+}
+
+#[test]
+fn how_h5py_reads_the_closed_copy_past_its_recorded_end_of_file() {
+    skip_if_no_python!();
+    let dir = tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
+    let path = dir.path().join("closed_short_eof.h5");
+    std::fs::write(&path, mid_write_copy_with_flags(0)).unwrap();
+    // libhdf5's plain reader refuses `a` (its chunk index is past the
+    // recorded end of file), as File::open and open_swmr do above.
+    // libhdf5's SWMR reader skips its end-of-allocation check for every
+    // file it opens (H5FD_read), and reads it; it still refuses an object
+    // *header* past the end ("address of object past end of allocation",
+    // H5O_protect). open_swmr does not copy that half-way rule for files
+    // without the SWMR-write flag: they read as File::open reads them.
+    let script = format!(
+        r#"
+import h5py
+for swmr in (False, True):
+    try:
+        with h5py.File("{p}", "r", swmr=swmr, locking=False) as f:
+            f["a"][()]
+    except Exception as e:
+        print("refused", swmr)
+    else:
+        print("read", swmr)
+"#,
+        p = path.display()
+    );
+    let out = Command::new(python())
+        .args(["-c", &script])
+        .output()
+        .unwrap();
+    let stdout = String::from_utf8_lossy(&out.stdout);
+    assert_eq!(
+        stdout.split_whitespace().collect::<Vec<_>>(),
+        ["refused", "False", "read", "True"],
+        "{}",
+        String::from_utf8_lossy(&out.stderr)
+    );
+}
+
+#[test]
+fn open_swmr_returns_a_permanent_error_at_once() {
+    // Not HDF5 at all: SignatureNotFound is not something a writer
+    // causes, so it is not retried (100 attempts would pause about 0.9 s).
+    let started = std::time::Instant::now();
+    let err = File::open_storage_swmr(Arc::new(vec![7u8; 4096])).unwrap_err();
+    assert!(
+        matches!(
+            err,
+            clawhdf5::Error::Format(clawhdf5_format::error::FormatError::SignatureNotFound)
+        ),
+        "{err}"
+    );
+    assert!(started.elapsed() < std::time::Duration::from_millis(100));
+
+    // A live file: a lookup of a name it does not have fails at once, and
+    // a torn object header read (below) is retried.
+    let f = File::open_storage_swmr(Arc::new(mid_write_copy_with_flags(0x05))).unwrap();
+    assert!(f.is_swmr_read());
+    let started = std::time::Instant::now();
+    assert!(f.dataset("no_such").is_err());
+    assert!(started.elapsed() < std::time::Duration::from_millis(100));
+    assert_eq!(f.swmr_retries(), 0);
+}
+
+/// A storage whose next `torn` reads come back garbled, as a read racing a
+/// rewrite of the structure can see them: the middle byte changed, or with
+/// `invert` every byte (signatures included).
+struct Torn {
+    bytes: Vec<u8>,
+    torn: AtomicU32,
+    invert: AtomicBool,
+}
+
+impl clawhdf5::Storage for Torn {
+    fn read_at(
+        &self,
+        offset: u64,
+        len: usize,
+    ) -> Result<std::borrow::Cow<'_, [u8]>, clawhdf5_format::error::FormatError> {
+        let got = self.bytes.as_slice().read_at(offset, len)?;
+        let tear = self
+            .torn
+            .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |n| n.checked_sub(1))
+            .is_ok();
+        if tear && !got.is_empty() {
+            let mut v = got.into_owned();
+            if self.invert.load(Ordering::SeqCst) {
+                v.iter_mut().for_each(|b| *b = !*b);
+            } else {
+                let mid = v.len() / 2;
+                v[mid] ^= 0x5a;
+            }
+            return Ok(v.into());
+        }
+        Ok(got)
+    }
+
+    fn len(&self) -> u64 {
+        self.bytes.len() as u64
+    }
+}
+
+#[test]
+fn a_torn_metadata_read_is_retried_and_never_returned() {
+    use clawhdf5::Storage as _;
+    let torn = Arc::new(Torn {
+        bytes: std::fs::read(fixture("swmr_mid_write.h5")).unwrap(),
+        torn: AtomicU32::new(0),
+        invert: AtomicBool::new(false),
+    });
+    let f = File::open_storage_swmr(torn.clone()).unwrap();
+    assert_eq!(f.swmr_retries(), 0);
+    assert!(torn.len() > 0);
+
+    // The next 3 reads (object headers) come back garbled: the lookup
+    // fails its checks, is run again, and returns the right dataset. The
+    // same with every byte garbled, signatures included.
+    for invert in [false, true] {
+        torn.invert.store(invert, Ordering::SeqCst);
+        let before = f.swmr_retries();
+        torn.torn.store(3, Ordering::SeqCst);
+        let a = f.dataset("a").unwrap();
+        assert!(f.swmr_retries() > before, "invert {invert}");
+        assert_eq!(a.read_i64().unwrap(), (1..=148).collect::<Vec<i64>>());
+        // Two garbled reads: the prefix of the root group's header (whose
+        // garbled byte may go unused, as it is read again whole) and the
+        // whole header, which fails its checksum.
+        let before = f.swmr_retries();
+        torn.torn.store(2, Ordering::SeqCst);
+        let b = f.dataset("b").unwrap();
+        assert_eq!(b.read_f64().unwrap().len(), 148 * 4);
+        assert!(f.swmr_retries() > before, "invert {invert}");
+    }
+    torn.invert.store(false, Ordering::SeqCst);
+
+    // With one attempt the error is returned instead (both reads of the
+    // header garbled, as above).
+    let mut f = File::open_storage_swmr(torn.clone()).unwrap();
+    f.set_swmr_read_attempts(1);
+    torn.torn.store(2, Ordering::SeqCst);
+    assert!(f.dataset("a").is_err());
+    assert_eq!(f.swmr_retries(), 0);
+    torn.torn.store(0, Ordering::SeqCst);
+
+    // A refresh that keeps failing gives up after the attempts, and the
+    // handle keeps its extent.
+    f.set_swmr_read_attempts(5);
+    let mut b = f.dataset("b").unwrap();
+    torn.torn.store(1000, Ordering::SeqCst);
+    assert!(b.refresh().is_err());
+    assert_eq!(f.swmr_retries(), 4);
+    torn.torn.store(0, Ordering::SeqCst);
+    assert_eq!(b.shape().unwrap(), vec![148, 4]);
+    b.refresh().unwrap();
+    assert_eq!(b.read_f64().unwrap().len(), 148 * 4);
+
+    // A file not opened for SWMR reading does not retry.
+    let plain = File::open_storage(torn.clone()).unwrap();
+    torn.torn.store(2, Ordering::SeqCst);
+    assert!(plain.dataset("a").is_err());
+    assert_eq!(plain.swmr_retries(), 0);
+}
+
+// ---------------------------------------------------------------------------
+// A live file: an h5py SWMR writer appends while clawhdf5 and h5py read.
+// ---------------------------------------------------------------------------
+
+/// The h5py SWMR writer. It creates three chunked datasets whose values are
+/// a function of their position, so a reader can check every value it
+/// reads without knowing when it was written:
+///
+/// - `a`: int64 `(n,)`, chunks of 100, no filter — `a[i] = i + 1`
+///   (Extensible Array index);
+/// - `b`: float64 `(n, 4)`, chunks of 16 x 4, gzip — `b[i, j] = 10 i + j + 1`
+///   (Extensible Array index);
+/// - `c`: int32 `(r, k)`, both unlimited, chunks of 8 x 8, no filter —
+///   `c[i, j] = 1000 i + j + 1` (version-2 B-tree index).
+///
+/// It switches to SWMR mode, prints `ready`, then for `steps` steps appends a
+/// random number of rows to each dataset (and every 7th step a column to
+/// `c`), resizing before writing as SWMR writers must, flushing each
+/// dataset and pausing 2 ms, and finally closes the file and writes the
+/// final contents to `<path>.<name>.bin`. `CLAWHDF5_SWMR_STEPS` sets the
+/// number of steps (default 2500).
+const WRITER: &str = r#"
+import sys, time, random, h5py, numpy as np
+path, steps = sys.argv[1], int(sys.argv[2])
+rng = random.Random(1234)
+f = h5py.File(path, "w", libver="latest")
+a = f.create_dataset("a", shape=(0,), maxshape=(None,), chunks=(100,), dtype="i8")
+b = f.create_dataset("b", shape=(0, 4), maxshape=(None, 4), chunks=(16, 4), dtype="f8",
+                     compression="gzip")
+c = f.create_dataset("c", shape=(0, 1), maxshape=(None, None), chunks=(8, 8), dtype="i4")
+f.swmr_mode = True
+print("ready", flush=True)
+na = nb = rc = 0
+kc = 1
+for step in range(steps):
+    k = rng.randint(1, 60)
+    a.resize((na + k,)); a[na:na + k] = np.arange(na, na + k) + 1; na += k
+    a.flush()
+    k = rng.randint(1, 20)
+    rows = np.arange(nb, nb + k)[:, None] * 10 + np.arange(4) + 1
+    b.resize((nb + k, 4)); b[nb:nb + k] = rows; nb += k
+    b.flush()
+    if step % 7 == 6:
+        c.resize((rc, kc + 1))
+        if rc:
+            c[:, kc] = np.arange(rc) * 1000 + kc + 1
+        kc += 1
+    k = rng.randint(1, 3)
+    c.resize((rc + k, kc))
+    c[rc:rc + k] = np.arange(rc, rc + k)[:, None] * 1000 + np.arange(kc) + 1
+    rc += k
+    c.flush()
+    time.sleep(0.002)
+f.close()
+with h5py.File(path, "r") as f:
+    for name in "abc":
+        open(f"{path}.{name}.bin", "wb").write(f[name][()].tobytes())
+        print(name, *f[name].shape, flush=True)
+"#;
+
+/// h5py's own SWMR reader, run beside ours as the reference: the same
+/// checks, until `<path>.stop` exists. Prints its iteration count.
+const H5PY_READER: &str = r#"
+import os, sys, h5py, numpy as np
+path = sys.argv[1]
+f = h5py.File(path, "r", swmr=True)
+ds = {n: f[n] for n in "abc"}
+last = {n: (0,) * ds[n].ndim for n in "abc"}
+its = 0
+while True:
+    stop = os.path.exists(path + ".stop")
+    for n, d in ds.items():
+        d.refresh()
+        shape = d.shape
+        assert all(s >= l for s, l in zip(shape, last[n])), (n, shape, last[n])
+        last[n] = shape
+        v = d[()]
+        if n == "a":
+            want = np.arange(shape[0]) + 1
+        elif n == "b":
+            want = np.arange(shape[0])[:, None] * 10 + np.arange(4) + 1
+        else:
+            want = np.arange(shape[0])[:, None] * 1000 + np.arange(shape[1]) + 1
+        bad = np.argwhere(v != want)
+        assert len(bad) == 0, (n, shape, bad[:5], v[tuple(bad[0])], want[tuple(bad[0])])
+    its += 1
+    if stop:
+        break
+print("iterations", its, *last["a"], *last["b"], *last["c"], flush=True)
+"#;
+
+fn want_a(n: u64) -> Vec<i64> {
+    (1..=n as i64).collect()
+}
+
+fn want_b(rows: std::ops::Range<u64>) -> Vec<f64> {
+    rows.flat_map(|i| (0..4).map(move |j| (10 * i + j + 1) as f64))
+        .collect()
+}
+
+fn want_c(rows: std::ops::Range<u64>, cols: u64) -> Vec<i32> {
+    rows.flat_map(|i| (0..cols).map(move |j| (1000 * i + j + 1) as i32))
+        .collect()
+}
+
+/// One pass of the Rust reader: refresh every dataset, check its extent
+/// did not shrink, and check what it reads (the last rows of each, and all
+/// of them every `full_every` passes).
+fn check_pass(
+    datasets: &mut [clawhdf5::Dataset<'_>; 3],
+    last: &mut [Vec<u64>; 3],
+    pass: u64,
+    full_every: u64,
+) {
+    for (k, d) in datasets.iter_mut().enumerate() {
+        d.refresh().unwrap_or_else(|e| panic!("refresh {k}: {e}"));
+        let shape = d.shape().unwrap();
+        assert!(
+            shape.iter().zip(&last[k]).all(|(s, l)| s >= l),
+            "dataset {k} shrank: {shape:?} after {:?}",
+            last[k]
+        );
+        last[k] = shape;
+    }
+    let full = pass.is_multiple_of(full_every);
+    let na = last[0][0];
+    let from = if full { 0 } else { na.saturating_sub(200) };
+    let got = if full {
+        datasets[0].read_i64()
+    } else {
+        datasets[0].read_i64_selection(&Selection::slice(std::slice::from_ref(&(from..na))))
+    }
+    .unwrap_or_else(|e| panic!("read a: {e}"));
+    let want: Vec<i64> = want_a(na).split_off(from as usize);
+    assert!(got == want, "a {from}..{na}: {:?}", first_diff(&got, &want));
+
+    let nb = last[1][0];
+    let from = if full { 0 } else { nb.saturating_sub(50) };
+    let sel = Selection::slice(&[from..nb, 0..4]);
+    let got = datasets[1]
+        .read_f64_selection(&sel)
+        .unwrap_or_else(|e| panic!("read b: {e}"));
+    let want = want_b(from..nb);
+    assert!(got == want, "b {from}..{nb}: {:?}", first_diff(&got, &want));
+
+    let (rc, kc) = (last[2][0], last[2][1]);
+    let from = if full { 0 } else { rc.saturating_sub(9) };
+    let got = if full {
+        datasets[2].read_i32()
+    } else {
+        datasets[2].read_i32_selection(&Selection::slice(&[from..rc, 0..kc]))
+    }
+    .unwrap_or_else(|e| panic!("read c: {e}"));
+    let want = want_c(from..rc, kc);
+    assert!(
+        got == want,
+        "c {from}..{rc} x {kc}: {:?}",
+        first_diff(&got, &want)
+    );
+}
+
+fn first_diff<T: PartialEq + std::fmt::Debug>(got: &[T], want: &[T]) -> String {
+    if got.len() != want.len() {
+        return format!("{} values, want {}", got.len(), want.len());
+    }
+    let i = got.iter().zip(want).position(|(g, w)| g != w).unwrap();
+    format!("first difference at {i}: {:?}, want {:?}", got[i], want[i])
+}
+
+#[test]
+fn a_live_file_reads_consistently_while_an_h5py_swmr_writer_appends() {
+    skip_if_no_python!();
+    let steps: u32 = std::env::var("CLAWHDF5_SWMR_STEPS")
+        .ok()
+        .and_then(|s| s.parse().ok())
+        .unwrap_or(2500);
+    let dir = tempfile::tempdir_in(env!("CARGO_TARGET_TMPDIR")).unwrap();
+    let path = dir.path().join("live.h5");
+
+    let mut writer = Command::new(python())
+        .args(["-c", WRITER, path.to_str().unwrap(), &steps.to_string()])
+        .stdout(Stdio::piped())
+        .spawn()
+        .unwrap();
+    let mut out = BufReader::new(writer.stdout.take().unwrap());
+    let mut line = String::new();
+    out.read_line(&mut line).unwrap();
+    assert_eq!(line.trim(), "ready", "writer did not start");
+
+    // h5py's SWMR reader on the same file, as the reference.
+    let h5py_reader = Command::new(python())
+        .args(["-c", H5PY_READER, path.to_str().unwrap()])
+        .stdout(Stdio::piped())
+        .stderr(Stdio::piped())
+        .spawn()
+        .unwrap();
+
+    let file = File::open_swmr(&path).unwrap();
+    assert!(file.is_swmr_read());
+    assert!(file.swmr_writer_active().unwrap());
+    // Two reader threads share the file, each with its own handles; each
+    // makes passes until it has seen the writer exit, then one more.
+    let writer_done = AtomicBool::new(false);
+    let reader = || {
+        let mut datasets = ["a", "b", "c"].map(|n| file.dataset(n).unwrap());
+        let mut last = [vec![0], vec![0, 4], vec![0, 1]];
+        let (mut passes, mut grew) = (0u64, 0u64);
+        loop {
+            let done = writer_done.load(Ordering::Acquire);
+            let before = last[0][0];
+            check_pass(&mut datasets, &mut last, passes, 50);
+            passes += 1;
+            grew += u64::from(last[0][0] > before);
+            if done {
+                break;
+            }
+        }
+        // The writer has closed the file: one more pass reads the final
+        // extents, all of them.
+        check_pass(&mut datasets, &mut last, 0, 50);
+        (passes, grew, last)
+    };
+    let (status, results) = std::thread::scope(|s| {
+        let readers = [s.spawn(reader), s.spawn(reader)];
+        let status = writer.wait().unwrap();
+        writer_done.store(true, Ordering::Release);
+        (status, readers.map(|r| r.join().unwrap()))
+    });
+    assert!(status.success(), "writer failed");
+    assert!(!file.swmr_writer_active().unwrap());
+    let last = results[0].2.clone();
+    assert_eq!(results[1].2, last);
+    let passes: Vec<u64> = results.iter().map(|r| r.0).collect();
+    let grew: u64 = results.iter().map(|r| r.1).min().unwrap();
+
+    std::fs::write(path.with_extension("h5.stop"), b"").unwrap();
+    let reference = h5py_reader.wait_with_output().unwrap();
+    assert!(
+        reference.status.success(),
+        "h5py's SWMR reader failed:\n{}",
+        String::from_utf8_lossy(&reference.stderr)
+    );
+    let reference = String::from_utf8_lossy(&reference.stdout).into_owned();
+
+    // The writer's final contents, as h5py reads the closed file.
+    let mut finals = String::new();
+    out.read_to_string(&mut finals).unwrap();
+    let shapes: std::collections::HashMap<&str, Vec<u64>> = finals
+        .lines()
+        .map(|l| {
+            let mut w = l.split_whitespace();
+            let name = w.next().unwrap();
+            (name, w.map(|v| v.parse().unwrap()).collect())
+        })
+        .collect();
+    assert_eq!(shapes["a"], last[0]);
+    assert_eq!(shapes["b"], last[1]);
+    assert_eq!(shapes["c"], last[2]);
+    let bin = |name: &str| std::fs::read(format!("{}.{name}.bin", path.display())).unwrap();
+    let le = |b: &[u8], n: usize| -> Vec<[u8; 8]> {
+        b.chunks(n)
+            .map(|c| {
+                let mut v = [0u8; 8];
+                v[..n].copy_from_slice(c);
+                v
+            })
+            .collect()
+    };
+    // The live handle and a fresh non-SWMR open of the closed file both
+    // read exactly what h5py reads.
+    for f in [&file, &File::open(&path).unwrap()] {
+        let a: Vec<[u8; 8]> = f
+            .dataset("a")
+            .unwrap()
+            .read_i64()
+            .unwrap()
+            .iter()
+            .map(|v| v.to_le_bytes())
+            .collect();
+        assert_eq!(a, le(&bin("a"), 8));
+        let b: Vec<[u8; 8]> = f
+            .dataset("b")
+            .unwrap()
+            .read_f64()
+            .unwrap()
+            .iter()
+            .map(|v| v.to_le_bytes())
+            .collect();
+        assert_eq!(b, le(&bin("b"), 8));
+        let c: Vec<[u8; 8]> = f
+            .dataset("c")
+            .unwrap()
+            .read_i32()
+            .unwrap()
+            .iter()
+            .map(|v| {
+                let mut x = [0u8; 8];
+                x[..4].copy_from_slice(&v.to_le_bytes());
+                x
+            })
+            .collect();
+        assert_eq!(c, le(&bin("c"), 4));
+    }
+    eprintln!(
+        "SWMR: {steps} writer steps; clawhdf5 readers made {passes:?} passes (each saw `a` \
+         grow at least {grew} times), {} retries; h5py reader: {}",
+        file.swmr_retries(),
+        reference.trim()
+    );
+    assert!(
+        grew >= 2,
+        "a reader never saw the file grow ({passes:?} passes)"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// Every read path retries: strings, variable-length data, attributes.
+// ---------------------------------------------------------------------------
+
+/// A storage over `bytes` whose read number `fail_at` (counted from the
+/// last [`Flaky::arm`]) fails with a checksum mismatch, as a read racing a
+/// SWMR writer's flush can; it counts the reads.
+struct Flaky {
+    bytes: Vec<u8>,
+    reads: AtomicU32,
+    fail_at: AtomicU32,
+}
+
+impl Flaky {
+    fn arm(&self, fail_at: u32) {
+        self.reads.store(0, Ordering::SeqCst);
+        self.fail_at.store(fail_at, Ordering::SeqCst);
+    }
+}
+
+impl clawhdf5::Storage for Flaky {
+    fn read_at(
+        &self,
+        offset: u64,
+        len: usize,
+    ) -> Result<std::borrow::Cow<'_, [u8]>, clawhdf5_format::error::FormatError> {
+        let n = self.reads.fetch_add(1, Ordering::SeqCst);
+        if n == self.fail_at.load(Ordering::SeqCst) {
+            return Err(clawhdf5_format::error::FormatError::ChecksumMismatch {
+                expected: 0,
+                computed: 1,
+            });
+        }
+        self.bytes.as_slice().read_at(offset, len)
+    }
+
+    fn len(&self) -> u64 {
+        self.bytes.len() as u64
+    }
+}
+
+/// `swmr_strings_attrs.h5`: a copy h5py 3.16 (HDF5 2.0) made of its own
+/// file right after switching it to SWMR mode (the SWMR-write flag is set),
+/// holding a variable-length string dataset `s` (`alpha beta gamma
+/// delta`), a variable-length int32 dataset `v` (`[1] [2 3] [4 5 6]`), an
+/// int64 dataset `d` (`0..10`) with 12 string attributes `k00`..`k11`
+/// (`value 0`..; dense storage: a fractal heap and a v2 B-tree), a group
+/// `g` with attribute `note`, and root attributes `title` (a
+/// variable-length string) and `n`.
+#[test]
+fn every_read_path_of_a_live_file_retries_a_failed_read() {
+    let flaky = Arc::new(Flaky {
+        bytes: std::fs::read(fixture("swmr_strings_attrs.h5")).unwrap(),
+        reads: AtomicU32::new(0),
+        fail_at: AtomicU32::new(u32::MAX),
+    });
+    let f = File::open_storage_swmr(flaky.clone()).unwrap();
+    assert!(f.is_swmr_read());
+    let root = f.root();
+    let g = f.group("g").unwrap();
+    let s = f.dataset("s").unwrap();
+    let v = f.dataset("v").unwrap();
+    let d = f.dataset("d").unwrap();
+    let sorted = |m: std::collections::HashMap<String, clawhdf5::AttrValue>| {
+        format!(
+            "{:?}",
+            m.into_iter().collect::<std::collections::BTreeMap<_, _>>()
+        )
+    };
+    type Op<'a> = Box<dyn Fn() -> Result<String, clawhdf5::Error> + 'a>;
+    let ops: Vec<(&str, Op<'_>)> = vec![
+        ("root.attrs", Box::new(|| root.attrs().map(sorted))),
+        (
+            "root.attr",
+            Box::new(|| root.attr("title").map(|a| format!("{a:?}"))),
+        ),
+        ("g.attrs", Box::new(|| g.attrs().map(sorted))),
+        ("d.attrs", Box::new(|| d.attrs().map(sorted))),
+        (
+            "d.attr",
+            Box::new(|| d.attr("k05").map(|a| format!("{a:?}"))),
+        ),
+        (
+            "root.datasets",
+            Box::new(|| root.datasets().map(|n| format!("{n:?}"))),
+        ),
+        (
+            "root.groups",
+            Box::new(|| root.groups().map(|n| format!("{n:?}"))),
+        ),
+        ("d.shape", Box::new(|| d.shape().map(|n| format!("{n:?}")))),
+        (
+            "d.read_i64",
+            Box::new(|| d.read_i64().map(|n| format!("{n:?}"))),
+        ),
+        (
+            "s.read_string",
+            Box::new(|| s.read_string().map(|n| format!("{n:?}"))),
+        ),
+        (
+            "s.read_string_bytes",
+            Box::new(|| s.read_string_bytes().map(|n| format!("{n:?}"))),
+        ),
+        (
+            "s.read_string_selection",
+            Box::new(|| {
+                s.read_string_selection(&Selection::slice(std::slice::from_ref(&(1..3))))
+                    .map(|n| format!("{n:?}"))
+            }),
+        ),
+        (
+            "v.read_vlen",
+            Box::new(|| v.read_vlen::<i32>().map(|n| format!("{n:?}"))),
+        ),
+        (
+            "v.read_vlen_selection",
+            Box::new(|| {
+                v.read_vlen_selection::<i32>(&Selection::slice(std::slice::from_ref(&(1..3))))
+                    .map(|n| format!("{n:?}"))
+            }),
+        ),
+    ];
+    for (name, op) in &ops {
+        flaky.arm(u32::MAX);
+        let want = op().unwrap_or_else(|e| panic!("{name}: {e}"));
+        let reads = flaky.reads.load(Ordering::SeqCst);
+        // Each read the operation makes fails once in turn: the operation
+        // is run again and returns the same result.
+        for k in 0..reads {
+            flaky.arm(k);
+            let before = f.swmr_retries();
+            let got = op().unwrap_or_else(|e| panic!("{name}, read {k} of {reads} failing: {e}"));
+            assert_eq!(got, want, "{name}, read {k} of {reads} failing");
+            assert_eq!(f.swmr_retries(), before + 1, "{name}, read {k} of {reads}");
+        }
+    }
+    flaky.arm(u32::MAX);
+    assert!(ops[0].1().unwrap().contains("live strings"));
+    assert!(ops[3].1().unwrap().contains("value 11"));
+    assert_eq!(
+        ops[9].1().unwrap(),
+        r#"["alpha", "beta", "gamma", "delta"]"#
+    );
+    assert_eq!(ops[12].1().unwrap(), "[[1], [2, 3], [4, 5, 6]]");
+
+    // With one attempt the failure reaches the caller.
+    let mut f = File::open_storage_swmr(flaky.clone()).unwrap();
+    f.set_swmr_read_attempts(1);
+    let s = f.dataset("s").unwrap();
+    flaky.arm(0);
+    assert!(s.read_string().is_err());
+}

docs/design/range-reads.md

@@ -7,7 +7,9 @@ change. Progress: M0 and M1 are done, and so is M2 (branch
 `File::open_storage` gives the facade's read API over any `Storage` (see
 `CHANGELOG.md`, "Range reads, milestone M2"). M3 is done on branch
 `feat/p3-m3-remote`: the `clawhdf5-remote` crate (block cache, HTTP(S),
-object stores) and URLs in `h5rs` (see the M3 status below). M4 (wasm) is
+object stores) and URLs in `h5rs` (see the M3 status below). M5 (SWMR) is
+done on branch `feat/p3-m5-swmr-reader`, with its own design in
+[`swmr.md`](swmr.md) (see the M5 status below). M4 (wasm) is
 next. Every count in §1–§2 was
 
 change. Progress: M1, first part (the `Storage` trait and the metadata
@@ -523,6 +525,19 @@ fast path within benchmark noise.
 **M5 — SWMR and growth (later, separate design).** `Storage::len()` may grow;
 add `File::refresh()` that re-reads the superblock/EOF and invalidates cached
 blocks past the old end. Needs libhdf5 SWMR semantics research first.
+- *Status 2026-09-27:* done on branch `feat/p3-m5-swmr-reader`; design and
+  libhdf5 research in [`swmr.md`](swmr.md). Differences from the sketch
+  above: the refresh is per dataset (`Dataset::refresh`, as libhdf5's
+  `H5Drefresh`), not per file — a SWMR writer only grows datasets, and the
+  superblock's EOF is not kept up to date by it, so there is nothing to
+  re-read there (`File::swmr_writer_active` re-reads its flags). A live
+  file (`File::open_swmr`) reads through `FileStorage` (positioned reads,
+  `len()` the current length) with no block or chunk cache, rather than
+  invalidating cached blocks: `BlockCache`/`HttpStorage` stay snapshot
+  readers. Operations that fail with an error a racing write can cause are
+  retried up to 100 times (libhdf5's default for SWMR readers).
+  Tested against a live h5py writer and h5py's SWMR reader
+  (`crates/clawhdf5/tests/swmr_interop.rs`).
 
 Total: roughly 6–10 engineer-weeks for M0–M4 (estimate, not measured).
 

docs/design/swmr.md

@@ -0,0 +1,178 @@
+# Design: reading files a SWMR writer is still appending to (range-read M5)
+
+Status: design 2026-09-27, implemented on branch `feat/p3-m5-swmr-reader`
+(see "Status" at the end). This is milestone M5 of
+[`range-reads.md`](range-reads.md): "`Storage::len()` may grow; add a
+refresh". It covers the reader only; clawhdf5 does not write SWMR files.
+
+## What libhdf5 does
+
+A SWMR ("single writer, multiple readers") writer is a libhdf5 process that
+opened a file with `libver='latest'` and switched to SWMR mode (h5py
+`f.swmr_mode = True`, `H5Fstart_swmr_write`). Readers open the same file
+with `H5F_ACC_SWMR_READ` (h5py `File(path, 'r', swmr=True)`) while the
+writer keeps appending. What the format and the library guarantee:
+
+- **Superblock v3, flags set.** The writer sets the superblock's
+  file-consistency flags to write access + SWMR write (`0x05`) and clears
+  them on close. The superblock's end-of-file address is *not* kept up to
+  date while writing: a copy of a file taken mid-write records an EOF of a
+  few hundred bytes while the file is tens of kilobytes (checked on tank,
+  2026-09-27, h5py 3.16 / HDF5 2.0: EOF 715 in a 17 857-byte file). A SWMR
+  reader therefore skips libhdf5's end-of-allocation check for every read
+  (`H5FD_read`: "allow access to data past the end of the allocated space
+  … for SWMR read access"), and bounds reads by the file's real length.
+- **A non-SWMR open of such a file fails** in libhdf5: "file is already
+  open for write (may use <h5clear file> to clear file consistency flags)".
+- **The writer only appends.** New objects and attributes cannot be created
+  in SWMR mode; datasets grow with `H5Dset_extent` and are written. Chunked
+  datasets with one unlimited dimension use an Extensible Array index, with
+  more than one a version-2 B-tree; both are updated in a SWMR-safe way.
+  Fixed Array and single-chunk indexes are for datasets that cannot grow.
+- **Flush ordering.** Every metadata structure the writer uses is
+  checksummed, and flush dependencies order the writes: a chunk's data is
+  written before the index entry that points to it, and index blocks before
+  the object header whose dataspace announces the new extent. A reader
+  that reads the object header first and the index after sees an index at
+  least as new as the extent, so every chunk inside the extent it read is
+  either in the index or never written (then it reads as the fill value, as
+  it does for libhdf5's reader).
+- **Refresh.** A reader sees a dataset's new extent only when it refreshes
+  it (`H5Drefresh`, h5py `Dataset.refresh()`), which evicts the dataset's
+  cached metadata and reads the object header again.
+- **Retries.** Reads are not atomic against writes on every system, so a
+  checksum can fail when a structure is read while the writer rewrites it.
+  A SWMR reader reads checksummed metadata up to 100 times before failing
+  (`H5Pset_metadata_read_attempts`; default 100 for SWMR access, 1
+  otherwise — `H5Ppublic.h` of HDF5 1.14.6).
+
+## What clawhdf5 did before
+
+- `File::open` of a SWMR-flagged file bounded every read by the recorded
+  EOF (`Superblock::data_end` only tolerated an EOF *past* the end of the
+  file). A file copied or read mid-write therefore listed, but every
+  chunked read failed ("unexpected EOF: need 787 bytes, have 715"), and
+  `h5rs check` reported chunk indexes "past the end of the file".
+- A `File` is a snapshot: an mmap (or a buffer) of the length at open, and a
+  per-file chunk cache that keeps each dataset's chunk index and decoded
+  chunks for the life of the `File`. A reader could not see growth at all,
+  and a mapping of a file that is being rewritten can change under a read.
+
+## Design
+
+1. **Bound SWMR files by their length.** `Superblock::data_end` returns the
+   file's length for a version-3 superblock with the SWMR-write flag,
+   whatever EOF it records, as libhdf5's SWMR reader does. Every existing
+   open path (`File::open`, `open_storage`, `h5rs`) then reads a finished
+   copy of a live file. We keep opening such files without a SWMR flag
+   (libhdf5 refuses): it is read-only and the alternative is an error.
+2. **A live open: `File::open_swmr(path)` / `File::open_storage_swmr`.**
+   - The file is read with positioned reads (`FileStorage`, `pread` on
+     Unix, `seek_read` on Windows), never mapped, and `len()` is the file's
+     current length, so reads past the length seen at open work.
+   - The facade's view of the file (`FileData`) is *live*: reads are not
+     clamped to an end fixed at open, only by the storage's current length.
+   - The chunk cache is not used: every read reads the chunk index and the
+     chunks it needs again. A cached index would hide new chunks, and a
+     cached partial edge chunk would read as fill where the writer has since
+     written data (unfiltered edge chunks are rewritten in place).
+   - Only a file whose superblock has the SWMR-write flag when it is opened
+     is read live. Any other file is read exactly as `File::open` reads it
+     (bounded by its recorded EOF, through the chunk cache, no retries), and
+     `is_swmr_read()` is `false`. libhdf5's SWMR reader is looser: it skips
+     the end-of-allocation check in `H5FD_read` for every file it opens,
+     flagged or not, yet still refuses an object header past the EOF
+     (`H5O_protect`, "address of object past end of allocation"). On a
+     closed file whose EOF is below its length (checked 2026-09-27, h5py
+     3.16 / HDF5 2.0, the mid-write fixture with its flags cleared) it
+     therefore reads a dataset whose chunk index lies past the EOF, which
+     its plain reader refuses. We follow the plain reader there: the EOF of
+     a file no SWMR writer has open is what the file says it is.
+   - A storage that caches blocks (`clawhdf5-remote`'s `BlockCache`) would
+     serve stale bytes; `HttpStorage` also pins a file by ETag and length and
+     refuses a changed file. Remote SWMR is out of scope.
+3. **`Dataset::refresh()`** reads the dataset's object header again (same
+   address) and replaces the handle's copy, so `shape()` and every later
+   read use the new extent. Like h5py, a handle that is not refreshed keeps
+   its extent; its reads still read the index as it is now, and only return
+   elements inside that extent.
+4. **Bounded retries.** In a live file, an operation (open, lookups and
+   listings, refresh, every dataset read — typed, raw, selections, strings
+   and variable-length data with their global-heap decoding — attributes,
+   `File::decode_*`, `verify_provenance`) that fails with an error a
+   concurrent write can
+   cause is run again from the start, up to `File::swmr_read_attempts()`
+   times (default 100, libhdf5's default; `set_swmr_read_attempts` changes
+   it), sleeping 1 µs, 2 µs, … up to 10 ms between attempts (under a second
+   in all). libhdf5 retries the one structure whose checksum failed; we
+   retry the whole operation, because the parsers are pure functions of the
+   bytes they read. Which errors: those libhdf5 retries (`H5C__load_entry`)
+   — a checksum mismatch, and a failure to decode the prefix it reads
+   before the checksum to size the structure (for an object header its
+   signature and version: a test that garbles every byte of a header gets
+   `InvalidObjectHeaderVersion`) — and a read past the file's current end
+   (short here; libhdf5 reads zeros there, which fail the checksum). Every
+   other error (an unsupported version or message, a file that is not HDF5,
+   a structure corrupt behind a valid checksum) is returned at once: an
+   earlier version retried nearly every format error, so a permanent one
+   cost about 0.9 s of pauses. Non-live files never retry.
+   Results are only returned from a run where every structure verified, so
+   a torn metadata read is an error, never data. `File::swmr_retries()`
+   counts the retries (libhdf5: `H5Fget_metadata_read_retry_info`).
+   Attribute reads leave out an attribute they cannot read (or return a
+   variable-length string one as `AttrValue::Raw`) instead of failing;
+   on a live file such an error of a retried kind runs the read again too,
+   and after the last attempt the last result is returned as before. The
+   zero-copy reads (`read_raw_ref`, `read_as_slice`, `read_*_zerocopy`)
+   need the file in memory, which a live file never is: they report
+   `None` / `ContiguousStorageRequired` without reading data.
+   Global heap collections (variable-length data) have no checksum in
+   HDF5, like raw data, so a torn read of one is only caught when it fails
+   a check (its signature, a bound).
+   Raw data has no checksum in HDF5 (unless Fletcher-32 is on), in libhdf5
+   as here: correctness rests on the writer's ordering (chunk data before
+   the index entry, only appends), as for libhdf5's reader.
+5. **Writer state.** `File::swmr_writer_active()` reads the superblock
+   flags again, so a reader can tell when the writer has closed the file.
+   A writer that crashed or was killed never clears the flag, so a reader
+   that follows a file needs a second stop condition (the README example
+   stops after a minute without growth).
+
+What stays out: SWMR writing, VFD SWMR (HDF5 1.13's page-buffer protocol,
+not in 1.14 or 2.0), remote SWMR, refresh of groups/attributes (a SWMR
+writer cannot add them), and `MmapFile`/`LazyFile`.
+
+## Tests
+
+- `crates/clawhdf5-format`: `data_end` of a SWMR-flagged v3 superblock whose
+  EOF is below the file length.
+- `crates/clawhdf5/tests/swmr_interop.rs`:
+  - a copy of a file taken mid-write (fixture) reads like h5py's SWMR reader;
+  - the same bytes with the flags cleared read as `File::open` reads them
+    (and how h5py's two readers read them);
+  - a permanent error returns at once; every read path (listings,
+    attributes, strings, variable-length data) with each of its reads
+    failing once in turn returns the same result;
+  - a live test: an h5py writer (`swmr_mode = True`) appends to a 1-D and a
+    2-D dataset with one unlimited dimension (Extensible Array, one of them
+    gzip) and a 2-D dataset with two (v2 B-tree), flushing after every step,
+    while a Rust reader refreshes and reads them in a loop, and an h5py SWMR
+    reader does the same as the reference. Every value read must be the value
+    the writer wrote (a deterministic function of its position), extents
+    never shrink, and after the writer closes both readers must read the
+    same data as h5py.
+
+## Status
+
+Implemented 2026-09-27 on branch `feat/p3-m5-swmr-reader` as designed
+above (`CHANGELOG.md`, "Range reads, milestone M5"). Observed on tank the
+same day (h5py 3.16 / HDF5 2.0, `cargo test -p clawhdf5 --test
+swmr_interop`, and once with `CLAWHDF5_SWMR_STEPS=20000` in a release
+build): no read returned a value the writer had not written at that
+position, h5py's reader agreed, and retries were needed but rare (the
+failures seen were checksum mismatches, each cured by one retry). A
+variant of the test with the chunk cache left on in live mode fails it
+(stale chunk index / edge chunk), which is why live files do not use it.
+
+Also found: `File::open` of such a file had been failing since the
+end-of-file check of 2026-09-26 (item 1; `docs/known-issues.md`).

docs/known-issues.md

@@ -7,6 +7,29 @@ deleting it.
 
 ---
 
+## Files a SWMR writer had open could not be read past a stale end of file
+
+**Status:** fixed 2026-09-27 (branch `feat/p3-m5-swmr-reader`), before any
+release: reads have been bounded by the recorded end of file since
+`7d7a7e7` (2026-09-26), which no release contains.
+
+A libhdf5 writer in SWMR mode (h5py `f.swmr_mode = True`) sets the
+superblock's SWMR-write flag and does not keep its end-of-file address up to
+date: a copy h5py made of its own file mid-write records 715 in a
+6 030-byte file (h5py 3.16 / HDF5 2.0, tank). `Superblock::data_end` only
+ignored the recorded end when it lay *past* the end of the file, so every
+reader bounded such a file at 715 bytes: it listed, but every chunked read
+failed ("unexpected EOF: need 787 bytes, have 715") and `h5rs check`
+reported the chunk indexes past the end of the file. Never wrong data.
+
+**Fix:** for a v3 superblock with the SWMR-write flag, the data ends at the
+end of the file, as libhdf5's SWMR reader reads it. **Test:**
+`crates/clawhdf5/tests/swmr_interop.rs` (the mid-write copy,
+`tests/fixtures/swmr_mid_write.h5`, through every open path and against
+h5py's SWMR reader). A file still being written is read with
+`File::open_swmr` (see `docs/design/swmr.md`); `File::open` maps the file
+at its length at open and is not meant for files that change while open.
+
 ## Fletcher-32 checksums disagreed with libhdf5 on about 1 chunk in 32768
 
 **Status:** fixed 2026-09-26, after v2.7.0. **Every release (v2.1.0 to