SchOS/docs/history/schfs.md

# SchFS — how it was built

The filesystem as it is now is [../architecture/schfs.md](../architecture/schfs.md).
This is the record of building it: the phase plan, what each phase actually
produced, and the bugs that cost real time.

**All five phases are complete.** The plan is kept because the reasoning in it
is still the reasoning behind the format, and because the ordering argument —
*do not start phase 2 before phase 1 is solid, the B-tree is the load-bearing
piece* — is the transferable part.

---

## The phases, as they were planned and as they went

Each phase leaves a bootable system. Do not start phase 2 before phase 1 is
solid — the B-tree is the load-bearing piece and everything else assumes it.

**Phase 1 — the B-tree, in userland first.** Write the CoW B-tree as a plain C
library with a block-device abstraction, and test it on a *host* file with a
fuzz/torture harness (random insert/delete/lookup, simulated crash at every
write point, verify the tree always mounts). This is the one component where a
subtle bug is unfixable later, and testing it in the kernel is far harder. It is
also portable straight into `tools/mkfs`.

**Phase 2 — read-only mount. DONE.** New magic, uberblock scan, tree walk, inode
and dirent key types. `tools/mkschfs` writes images; the kernel only reads them.
The system boots from a ramdisk root (see [../guides/layout.md](../guides/layout.md) and the ramdisk work)
so a broken disk cannot stop you from booting.

Went further than planned in two places, both cheaply: data blocks carry
checksums too (a `CSUM` key type, as btrfs does — without it "end-to-end
checksums" would cover only metadata, which is the part least likely to rot),
and extents landed early because reserving a file's blocks in one go was
simpler than not doing it.

Not yet done: SchFS files are not reachable through `open(2)`. The mount is
observable through `/proc/schfs` and through the kernel API in `schfs_int.h`,
but wiring it into `namei`/`readi` means the VFS work that phase 3 needs anyway.

**Phase 3 — writes, transaction groups. DONE.** Every `schfs_k*` call that
changes anything is one transaction group, committed by writing a new uberblock
before it returns. The full write API — create, link, unlink, rename, write at
an arbitrary offset, truncate — lives in `objw.c` and is exercised by
`tools/schfs_objtest.c`.

Two things went differently from the plan above, both for reasons worth
recording.

*Free space is not tracked as tree keys.* Allocating from a free map that lives
in the tree mutates the tree, which allocates, which changes the map; btrfs
resolves that with a fixpoint loop and ZFS with log-structured space maps. SchFS
sidesteps it: the committed tree already says exactly which blocks are live, so
the map is **derived** by walking the tree once at mount and held only in RAM.
It cannot disagree with the filesystem, because it is computed from it, and a
crash raises no question at all. The cost is mount time proportional to metadata
size, and a ceiling on volume size set by how much bitmap fits in memory. If
mount time ever hurts, the map can be cached on disk purely as a hint,
validated against the generation and discarded on a mismatch — an addition, not
a format change. See `kernel/fs/schfs/space.c`.

*Data blocks needed a generation stamp too.* The B-tree already rewrites its own
nodes in place within a group, since a block that no committed root can reach is
nobody's fallback. Data blocks had no such marker, so every hundred-byte write
copied a whole 4 KB block and split an extent record around it. `schfs_extent`
gained a 64-bit `gen`, and the same rule now applies to file data.

**Phase 3½ — reachable through `open(2)`. DONE.** SchFS is a mounted filesystem:

```
mkdir /mnt
mount -t schfs /dev/hdb /mnt
ls /mnt ; echo hi > /mnt/tmp/w ; cat /mnt/tmp/w ; mkdir /mnt/tmp/d
```

The VFS grew a real dispatch layer rather than a set of `if (dev == SCHFS_DEV)`
checks. `struct fs_ops` (include/file.h) is one table per filesystem, chosen by
`iget()` from the device number and hung off the inode; v1's implementations
keep their bodies and are simply named `v1_*`. Adding a third filesystem is a
new table and nothing else. `kernel/fs/schfs/vfs.c` is SchFS's, and it is pure
translation — object ids in, inodes out — because all the filesystem logic
already lives below it.

Three things had to change on the way, none of them SchFS's fault:

- `struct inode.inum` and `.size` were 32-bit, capping objects at 2³² and files
  at 4 GiB on a filesystem that promises far more. Both are `uint64` now, as are
  `readi`/`writei` offsets and `struct file.off`.
- `struct dirent` carries a 16-bit inum and a 14-byte name, so it cannot
  represent a SchFS entry at all. `getdents64` is now an fs_op: each filesystem
  emits the record stream itself, and `f->off` is whatever position it needs to
  resume. v1 walks its hash buckets; SchFS walks a key range and synthesises
  "." and "..", which it does not store.
- Mounting is a substitution made during path resolution (`mount_add` /
  `mount_cross` in `fs.c`). Nothing on either disk records it. The reverse
  direction matters as much as the forward one: `..` at a mounted root has to
  leave the mount, and the filesystem underneath has no idea it is mounted
  anywhere.

Three long-standing bugs surfaced the moment real commands ran against it, all
in the shared VFS rather than in SchFS:

- `include/fcntl.h` used xv6's open flags while every program in the userland
  sends Linux's. `O_APPEND` (0x400) landed on `O_TRUNC`, so `cmd >> file`
  *truncated* it; `O_TRUNC` (0x200) landed on `O_CREATE`, so `cmd > file` left
  whatever was past the new end in place. Nothing about either pointed at a
  header.
- `sys_openat` silently dropped `O_TRUNC` and `O_APPEND` while `sys_open`
  honoured both, so the same redirection behaved differently depending on which
  entry point libc happened to call.
- `rmdir` had no Linux syscall number, so every Linux-ABI program got ENOSYS.

**Phase 3¾ — durability, and the cost of it. DONE.** Three things that the
earlier phases had assumed rather than arranged.

*The commit barrier.* Copy-on-write promises that a crash lands on either the
old tree or the new one, but only if the blocks a root points at reach the
medium *before* the block naming that root. Nothing enforced that: `bwrite`
order is not platter order on any drive with a write cache. `struct schfs_bdev`
gained a `bflush` hook, `schfs_vol_commit` calls it either side of the uberblock
write, and the kernel implements it as "drain this device's dirty buffers, then
ATA FLUSH CACHE" (`ideflush`, and SCSI SYNCHRONIZE CACHE for a USB stick).

*Falling back through the ring.* Mount took the highest-generation uberblock
that validated, checked the tree it named, and gave up if that failed — with up
to 127 older, entirely intact roots sitting on the disk. It now walks the ring
newest-first and adopts the first root it can actually read, counting what it
skipped into `vol->recovered` so the rollback is reported rather than silent.

*Deferred writes, and brackets.* The B-tree seals and writes a node on every
record it inserts, so a leaf collecting one file's data checksums went to the
disk once per 4 KB of that file. The kernel's block hook marks buffers dirty
instead (`bdirty`), and the commit barrier is what makes that safe. On its own
that changes little, because a commit *contains* a flush and every `schfs_k*`
call was one commit — so `schfs_ktx_open`/`schfs_ktx_close` bracket a run of
calls into one group, and `schfs_vfs_write` wraps a whole `write(2)` in one.

Measured over 1 MB of file data, in block writes reaching the device:

| | before | after |
|---|---|---|
| committing every 4 KB | 1947 | 1410 |
| one commit for the megabyte | 1798 | **267** |

267 device writes for 256 blocks of data. Note that the two changes are worth
almost nothing separately — deferral cannot help while every commit flushes,
and batching cannot help while every write goes straight to the disk — which is
why the first attempt at this measured an 8% difference and concluded batching
was not the problem.

**Phase 3⅞ — where the time actually went. DONE.** The filesystem was correct
and durable and the system still felt slow: `ls /bin` took the better part of a
second. Two causes, neither of them the disk.

Measured first, through `/proc/fsstat` (counters added for exactly this). The
decisive number was **398,999 block reads against 445 actual disk misses** —
99.9% of all block reads were re-reading a block already in the buffer cache. So
the disk was never the problem, and every plan that started with "cache more" or
"read ahead" would have been aimed at 1% of the cost.

*The cursor re-read its own buffer.* `cur_load` called `node_read`
unconditionally, and `schfs_cursor_val` calls `cur_load` on **every access to
the current record** — deliberately, so that a callback disturbing the shared
block buffer cannot corrupt a directory walk. But `node_read` copies 4096 bytes
out of the buffer cache and then runs CRC32c over all 4096. Listing a 243-entry
directory paid both once per entry. The cursor now remembers which block its
buffer holds and skips the reload, cross-checked against the block header (so a
second cursor sharing the page cannot fool it) and against a global epoch that
every mutation bumps (so a block rewritten in place cannot be served stale).

*The checksum was bit-at-a-time.* `schfs_crc32c` did eight shifts per byte —
32,768 iterations to checksum one block — to stay usable before any
initialisation. It builds a 1 KB table on first use now, which keeps that
property and is eight times faster. This only became the top cost *after* the
cursor fix, which is the usual shape: removing the largest item promotes
whatever was beneath it.

| cycles, one session | before | after |
|---|---|---|
| `getdents` | 26.8 G | **0.14 G** (196×) |
| `read` | 26.4 G | **5.5 G** (4.8×) |
| `exec` | 18.7 G | **4.0 G** (4.7×) |
| block reads | 398,999 | **127,713** |

`ls /bin` went from ~400 ms to ~20 ms. Verified by reading an 802 KB binary
through the whole path and checking its md5 against the host's copy.

There was also a third of the problem that was not the kernel at all: `make
qemu` ran QEMU without `-enable-kvm`, so every instruction was emulated. The run
targets probe `/dev/kvm` now. Measure against the KVM number or you are
measuring TCG.

**Phase 4 — extents with refcounts. DONE.** `cp` that costs nothing.

A reference count per run of blocks, in its own tree (`SCHFS_KEY_EXTREF` under
`SCHFS_EXTREF_OBJECTID`) rather than in the file's own extent record — because a
count kept per file cannot be decremented by the *other* file sharing the run.
`schfs_file_reflink` copies one extent record per run and bumps the count; both
files stay ordinary and writable, and the first write to either copies only the
run it touches. Reachable from userland as `FICLONE`, Linux's ioctl, so
`cp --reflink` works unmodified; `cmd/reflink` is a two-line caller of it.

Measured on the running system: cloning a 411 KB file moved `df` by 24 KB.

Three things had to be right, and each was wrong first:

- **The count is "how many extent records point into this run", not "how many
  files".** Splitting a shared extent around a written block leaves a head *and*
  a tail, which is one more reference than before. Missing that leaked four
  blocks per split.
- **A file's extent may point into the middle of a run.** Records are keyed by
  the block a run was *allocated* at, so every lookup is a range query
  (`extref_find`), not an exact one. An exact lookup finds nothing for a split
  tail, "nothing" means unshared, and the tail is then freed while another file
  still reaches it.
- **The space scan has to walk the reference records too.** A file truncated
  inside a shared run describes fewer blocks than exist; the rest would be
  derived as free while still referenced, and the allocator would hand out a
  block that a later release frees a second time.

None of those is visible until every holder of a run is deleted, which is why
the test that matters is the least glamorous one: after deleting every file that
shared a run, free space must be back to exactly what it was.

**Phase 5 — attribute index and query. DONE.** The reason the design was one
B-tree with many key types from the start: the "database filesystem" feature
turned out to be `kernel/fs/schfs/attr.c` and nothing else.

An attribute is a named value on an object, stored under `(ino, ATTR,
hash-of-name)` and packed exactly as directory entries are — same collision
argument, same shape. An attribute may be *indexed*, and an indexed attribute
gets a second set of records keyed by value: `(hash-of-name, INDEX,
hash-of-value)` holding the object ids that carry it. So "every object whose
`tag` is `holiday`" is one tree descent instead of a walk of every object.

There is no query language, no parser and no planner, and §6 above explains why:
the useful half of "query the filesystem" is the index, and WinFS spent three
years and a large team on the other interpretation.

```
schattr index tag                 # declare the index
schattr set /tmp/a tag holiday
schattr find tag holiday          # one descent, whatever the volume holds
```

Four decisions worth keeping:

- **Which attributes are indexed is cached in the volume**, read once at mount.
  Every `setxattr` has to ask, and the answer is almost always no; a tree
  descent to say so would make every attribute write pay for a feature it is
  not using.
- **The value is stored in the index entry, not just its hash.** Two values can
  hash alike, and a query that returned the wrong file would be far worse than a
  slow one.
- **An index covers values written after it is declared.** There is no
  reindexing pass. Saying so is better than a half-truth that makes a query
  answer confidently and wrongly.
- **A query on an unindexed attribute is refused, not answered by scanning.** A
  query that silently degrades into a full walk teaches you to use it where it
  will not scale.

Reachable as `setxattr`/`getxattr`/`removexattr` at Linux's syscall numbers, so
anything that already knows how to attach metadata needs no changes, plus two
native calls (`schfs_index`, `schfs_query`) that have no Linux equivalent
because asking a filesystem a question does not.

**Phase 5 — attribute index and query.** See below.


---

## What was reused from v1

Not starting from zero:

- **`struct schfs_extent`** is already defined in `include/fs.h` and unused.
  The extent concept is already sketched; v2 gives it refcounts and 64-bit
  fields.
- **CRC32c checksum fields** already exist in `dinode` and `superblock`, and
  mkfs already computes them. Same algorithm, relocated into parent pointers.
- **`kernel/fs/log.c`** is a working transaction log with `begin_op`/`end_op`
  call sites throughout the FS. Under CoW the log is not needed for consistency,
  but `begin_op`/`end_op` are exactly the right places to bracket a transaction
  group, so the *call sites* survive even though the implementation is replaced.
- **The buffer cache** (`bio.c`) is format-agnostic and needs no change beyond
  the block-size constant.


---

## What it was estimated to cost

Phases 1-3 are roughly 3,000-4,000 lines and are the bulk of the work. Phase 1
alone (B-tree + torture harness) is worth doing carefully and unhurriedly; every
later phase compounds on it.

This is the largest single piece of work in SchOS. It is also entirely
ordinary filesystem engineering with excellent published references
(Giampaolo for BeFS; the btrfs design papers; the ZFS uberblock/checksum-in-
parent scheme) — there is no research risk here, only volume.

**Do this before the installer.** The installer's entire job is to create and
populate this format. Building it against v1 means building it twice.


---

## Things learned the hard way

Recorded because each one cost real time and none is obvious from the code.

- **The root's object id must be reserved.** `next_objectid` started at
  `SCHFS_FIRST_OBJECTID`, which *is* the root's id, so the first directory
  created silently became the root. It then contained itself, and every tree
  walk recursed until the stack ran out. `schfs_fsck --check` now detects
  directory loops explicitly rather than crashing.
- **Cursor buffers alias.** A range callback gets a pointer into the cursor's
  block buffer, and almost every readdir caller immediately looks the entry up,
  which overwrites it. Directory iteration copies each entry out before calling
  back. Extracting a tree with directories reparented under their own children
  is what this looks like when it goes wrong.
- **A 4096-byte block does not fit on a 4096-byte kernel stack.** The cursor
  exists precisely to avoid one block buffer per tree level, but
  `schfs_vol_open` still kept a whole uberblock as a local and blew the stack on
  the first boot. The kstack canary caught it cleanly.
- **Extents only merge if you allocate the run first.** Writing an extent record
  allocates blocks for its own copy-on-write path, so allocating data one block
  at a time guarantees the blocks are never adjacent and every block needs its
  own extent. `schfs_file_reserve` claims the whole run up front; an 822 KB
  binary is one extent instead of 201.
- **Freed blocks are reusable during a build.** Copy-on-write forbids reusing a
  released block because the previously committed tree still points at it — but
  a build has committed nothing. Recycling them cut the test image from 3596
  blocks to 374.
- **Reproducibility dies from atime.** mkschfs reads every source file, which
  updates its atime, so copying atime into the image made two builds of the same
  tree differ. All three timestamps come from mtime.
- **An optional hook must still be initialised.** `struct schfs_bdev` gained
  `bflush`, and every host tool built one on the stack field by field. The tools
  that did not set the new field handed `schfs_vol_commit` a garbage pointer to
  test and call, which segfaulted at the first commit. Each of those helpers
  memsets the struct first now — the fix is not "remember to set it".
- **Deferring writes makes an all-dirty buffer cache normal.** `bget` panicked
  with "no buffers" when it could find nothing clean to recycle, which was
  correct while a dirty buffer meant "the log has not committed". It now writes
  the least recently used victim out and takes it; a transaction touching more
  than `NBUF` blocks reaches that state long before it commits.
- **The two performance fixes are worthless apart.** Deferring tree writes buys
  nothing while every commit flushes, and batching commits buys nothing while
  every write goes straight to the disk. Measuring either alone says "this is
  not where the time goes", and both measurements are correct.
- **`fsid_t` is eight bytes.** `struct statfs` declared `f_fsid` as `int64[2]`,
  which is one slot too long, so everything after it landed in the wrong place:
  userland read `f_frsize` out of `f_namelen` and got 255. df then computed in
  units of 255 bytes and called a 32 MB root filesystem 1992 KB. It survived
  because every field in that struct used to be an invented constant, so no
  number in it had ever been checked against reality.

---

## The one about the tests themselves

`tools/posix/mutate.sh` injects six deliberate bugs and fails if any goes
undetected.

It exists because a suite that passes is not evidence until something has
checked that it can fail. The most valuable single assertion in the whole SchFS
suite is also the least glamorous: **after deleting everything, free space must
be back to exactly what it was on an empty volume.** Every other test still
passes on a filesystem that leaks a block on every delete.