The AF_INET blocking paths -- connect, accept, read and write -- woke on
every timer tick and looked to see whether anything had changed. That is
a wakeup per blocked socket per 10 ms whether or not the network did
anything, and it answers late even when the data was already sitting in
the buffer.
What stood in the way was a lock mismatch, and it is written down in the
file now because it is the reason this is not simply sleep on the socket:
sleep() wants a spinlock, and the lwIP lock is a sleeplock -- it must be,
since a call into lwIP can reach the driver and the driver can sleep.
Nor can a waiter hold the lwIP lock while sleeping, or the network thread
could never deliver the thing it is waiting for.
So the wait is on a counter with a spinlock of its own. A waiter reads
the counter while it still holds the lwIP lock, right after checking its
condition, then releases and sleeps only if the counter has not moved.
The lwIP callbacks bump it and wake the channel. No wakeup can be lost:
anything that changes a condition does so after the reader's snapshot,
and therefore bumps the counter after it too, and the sleeper re-checks
under the spinlock before committing.
One counter and one channel for all sixteen sockets rather than one
each. Waking a socket with nothing to do costs a few instructions, and a
per-socket channel would have to be torn down in close() against
callbacks that can still arrive -- which is a use-after-free waiting to
be written.
tcp_sent is registered now, on both connected and accepted pcbs. It was
not needed while writes woke on the clock; with this change a write
blocked on a full send buffer would otherwise sleep with nobody left to
wake it, because an acknowledgement is the only thing that frees space.
Verified: bootcheck -Net passes, 38 [ok], no failures. It exercises the
path that matters -- nettest blocks in read() until the host answers, and
the host answers only after reading what the guest sent, so the wakeup
comes from a callback and not from a tick that would have fired anyway.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The TCP/IP stack landed with its end-to-end test performed by hand: start
a listener, boot, type nettest, read the output. That is evidence once
and nothing afterwards, and it is exactly the kind of check that quietly
stops being run.
bootcheck -Net does the whole exchange. A listener runs on this host on a
port picked by binding one and letting go, the guest is told to connect
out to 10.0.2.2 at that port, and both ends are asserted: the guest
reports what it received, the listener reports what arrived. One side
alone proves much less than it looks -- a stack that completes a
handshake and then drops every byte satisfies any check made from inside
the guest.
The listener is a background job, a separate process, because this
harness has one hard rule: nothing may block the writer feeding QEMU's
stdin. A socket accept is precisely the wait that deadlocked two earlier
attempts at reading QEMU's output. It answers the protocol nettest
speaks and then closes -- the close is load-bearing, since nettest reads
until end of file and a listener that stayed open would hang the guest
rather than fail it.
The NIC is spelled out (-netdev user + -device e1000) rather than left to
QEMU's default. Other runs can take the default and still answer "did a
driver bind"; a run that asserts on an address should be reading its own
configuration.
/proc/netif is new, and the address is why. Until now the only report was
a line on the console at boot, which userland cannot read and a test can
only find by scraping the transcript. It gives interface, MAC, state,
address, netmask, gateway, MTU and frame counts. Named netif rather than
net on purpose: /proc/net is a *directory* on Linux, and a plain file
there would give ENOTDIR to anything reaching for /proc/net/dev, which is
worse than not having it. The counters are kept at the lwIP seam rather
than in the driver, because a frame the driver received and the netif
dropped never reached the stack.
Verified: 38 [ok], no failures. Eleven of them are new -- lwIP up, DHCP
obtaining 10.0.2.15 from SLIRP (a number nothing in the guest knows in
advance, so it is the exchange completing rather than a guess), the four
/proc/netif fields, frames counted in both directions, the handshake, the
reply read back, and the listener confirming HELLO-FROM-SCHOS arrived.
One bug found by writing it: $netPort was chosen after the command list
was built, so the guest ran `nettest 10.0.2.2` with no port and printed
its usage. An unset PowerShell variable interpolates to nothing rather
than failing, which is worth remembering next time a command in this
file looks right and behaves oddly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
lwIP 2.2.0, fetched by ports/lwip and compiled straight into the kernel.
Not written from scratch, and the reason is not the packet formats: it is
congestion control, retransmission, window management and twenty years of
corner cases that only show up against a real peer. Not FreeBSD's
netinet either, though the licence would have allowed it -- that stack is
not a library, it reaches into mbufs, epoch(9), netisr, VNET and uma(9),
so adopting it means porting FreeBSD's kernel infrastructure first. lwIP
was written for exactly this situation: a small kernel with a raw
Ethernet driver and a port layer of a few hundred lines.
It is kernel code, so nothing is installed into the sysroot and there is
no second build system: the top-level Makefile compiles the files it
wants against our port headers. The file list is explicit rather than a
wildcard, because lwIP ships IPv6, 6LoWPAN, PPP, SLIP, DNS-SD and a dozen
apps in the same tree, and building what is not configured yields either
dead code or a link error naming a port function nobody meant to write.
Its sources are unmodified and its warnings are silenced for its files
only -- patching them would mean maintaining a fork.
The port layer is ours: a lock, a clock, the netif that hands frames to
e1000, lwipopts.h, and a shim/ directory of tiny headers so lwIP's
<string.h> and friends resolve to the kernel's functions instead of the
host's, whose prototypes disagree and would otherwise be found on the
default include path.
NO_SYS=1, so lwIP has no threads of its own and one kernel thread feeds
it. A thread rather than the interrupt handler, because a call into lwIP
can reach the driver, the driver can sleep, and an interrupt handler
cannot -- so the interrupt still only clears the cause register and the
thread polls. kthread_create is new and generally useful: a process with
a kernel-only page directory, no files, no cwd, no parent, dispatched
from forkret, which can therefore sleep and be preempted like anything
else. DHCP obtains an address at boot and the console says which one.
AF_INET went in exactly where phase 2 said it would. socknew dispatches
on the domain, kernel/net/inet.c implements the internet side on lwIP's
raw API, and not one syscall changed shape: bind, connect, listen,
accept, read and write reach the same descriptors they always did. TCP
and UDP both work. cmd/nettest opens a connection, writes, and reads the
answer back.
Two things are owed and named here so they are not discovered later:
recvfrom does not yet report the source address, and the socket paths
poll on the timer tick instead of sleeping on the socket. Both are
follow-ups, not redesigns.
NOTICE carries lwIP's BSD-3 text. Clause 3 is no-endorsement, not the
old advertising clause, so it sits with Apache-2.0 without difficulty.
Verified: the tree builds clean, and the headless boot check passes with
27 [ok] and no failures with the stack compiled in and the thread
running -- which is the part that would break first if netinit deadlocked
or the thread starved anything. An end-to-end connection test belongs in
that suite and is the next commit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The last three phases of the filesystem, which were written together and
touch the same files throughout -- kmount.c, objw.c and vfs.c are each
rewritten by all three -- so they land as one commit rather than three
that could not be built separately.
Durability. A transaction that reaches the platter out of order is worse
than one that never reached it: the uberblock can name a tree whose
blocks are still in the drive's write cache. bdirty/bflush give the
buffer layer an explicit barrier, ideflush issues ATA FLUSH CACHE and
usb_storage_flush the SCSI SYNCHRONIZE CACHE that means the same thing,
and a commit now waits for the data before it writes the root that
points at it. The uberblock groups are read newest-first with fallback,
so a volume whose newest roots are torn mounts at an older one and says
which generations were lost.
Reflink. Extents carry a refcount, so two objects can name the same
blocks and neither frees them while the other holds them. Exposed as
FICLONE -- Linux's ioctl value, so `cp --reflink` works unmodified --
and as cmd/reflink for the case where there is no cp to hand.
Attributes, and the one genuinely new idea here. attr.c stores named
values on an object, reachable through setxattr/getxattr/removexattr at
Linux's syscall numbers. On top of that, schfs_index declares an index
over an attribute name and schfs_query answers "which objects carry this
value" from the B-tree rather than by walking. A query against an
unindexed attribute is refused with an error, never silently answered by
a full scan: a query that quietly becomes a scan is worse than one that
admits it cannot help. cmd/schattr drives both.
Three bugs fixed on the way, each found by the numbers being wrong
rather than by inspection:
- statfs's f_fsid was declared int64[2]. Linux's fsid_t is two *ints*,
so every field after it landed one slot late: userland read f_frsize
out of f_namelen, got 255, and df reported a 32 MB root filesystem
as 1992 KB. It hid for as long as it did because the fields used to
be invented constants, so wrong numbers looked like the usual ones.
- struct inode's nlink was a short against SchFS's uint32, so a file
past 32767 names came back negative -- and iput frees an object
whose count reaches zero. It is an int now, signed on purpose, so an
underflow lands somewhere a check can see rather than wrapping.
- filewrite split every write to fit v1's log. That limit belongs to
the journal, not to writing, and SchFS makes one transaction of
whatever it is handed: a 1 MB write was becoming 170 commits and
170 flushes for nothing. fs_ops carries max_write now, and SchFS
sets no limit.
Also here: a real mount table, so `mount -t schfs /dev/hdb /mnt` shows
up in /proc/mounts -- df looks a path up there before it calls statfs at
all, so a mount missing from that list was invisible to it -- several
volumes mounted at once, snapshots mounted by name, and /proc/fsstat
with an rdtsc-based counter, which is the only instrument fine enough to
see inside a single syscall.
Verified: `make schfs-check` passes in full, including the sweep that
crashes the volume at each of 119 points mid-transaction and requires
every one to mount with exactly its committed data, and the case where
both newest uberblocks are rubbish and the volume still comes back. The
headless boot check passes with 27 [ok] and no failures, on a SchFS root.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
struct file already carried a uint64 offset and bread already took a
uint64 block number. The narrowing was one boundary in the middle:
devsw's read and write took a 32-bit uint, so fileread passed a 64-bit
f->off into a 32-bit parameter and the top half was discarded.
The effect was not a refusal at 4 GB but silent wraparound past it. A
read at 5 GB returned the bytes at 1 GB, reporting success. disk.c had a
DISK_MAX_BYTES constant documenting the limit, and nothing anywhere
referenced it -- the cap was described but never enforced, which is the
worst of the three options.
The signature is now uint64 across all nine files that carry it, and
diskread/diskwrite no longer narrow the block number on the way to
bread. Widening this is a prerequisite for NVMe and for any real disk:
namespaces start well above 4 GB and the failure mode reads valid data
from the wrong place.
Verified end to end rather than by inspection. A 6 GB sparse image with
"PAST4GB-MARKER-OK" written at byte 5368709120, attached as /dev/hdb to a
ramdisk-root kernel so device 1 was free:
# dd if=/dev/hdb bs=4096 skip=1310720 count=1 | od -c | head -2
0000000 P A S T 4 G B - M A R K E R - O
Under the old code that offset truncates to 1 GB, where the image is
zeroes. The whole path is exercised: f->off, devsw, diskread, bread and
ide.c's LBA48.
Worth recording for next time: `make ramroot` stages the ESP itself, and
running `make esp` afterwards rebuilds the kernel without RAMROOT and
silently undoes it. The tell is a 855 KB kernel where a 25 MB one is
expected.
110 [ok], no failures, on the default build.
Two loose ends from the mount-by-name commit, both closed.
A slot is a device number, so supporting more than one snapshot is a
small array rather than a redesign: SCHFS_SNAP_DEV through
SCHFS_SNAP_DEV + SCHFS_SNAP_SLOTS - 1, each with its own volume and its
own scratch pages, allocated on first use. Four of them -- "one at a
time" is an odd thing to have to explain, and the cost is two pages per
slot. Not many, because each one is a device number and those are a
small fixed space.
The leak mattered more. umount removed the mount table entry and left the
slot marked in use, so a snapshot could be unmounted and then never
mounted again until reboot. sys_umount2 releases it now, keyed off the
device it already has in hand.
Verified in the running system: two snapshots mounted at once, both
readable, both unmounted, and a third mount afterwards succeeds and reads
correctly -- which is the case that failed before.
schfs-check passes; bootcheck 108 [ok].
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`mount -t schfs @golden /mnt` mounts a snapshot. The source argument
finally means something: sys_mount ignored it outright, because there was
one volume and nothing to choose between. `@name` rather than a path
because a snapshot is not a device -- there is no node in /dev for it and
there should not be, since it lives inside the volume already mounted.
A second volume rather than a parameter threaded through all 23 schfs_k*
entry points. Only the read side can ever address it, so the write path
keeps talking to the live volume and needed no change; the four read
entry points grew _dev variants and pick the volume from the device the
inode already carries. One fs_ops table serves both, because every entry
takes an inode and dispatches on it -- a second table would be the same
pointers under a different name.
It needs its own device number: a snapshot's object ids are the *same
numbers* as the live tree's, so one device would make the two views alias
in the inode cache and hand a reader whichever was cached first.
Three things this turned up, none of them new but all of them reachable
for the first time:
- sys_mount refused with `dev == SCHFS_DEV`, meaning "do not mount SchFS
onto SchFS". A snapshot's mount point is necessarily on the live root,
so that refused the entire feature. The real condition is mounting a
device onto *itself*, which is `dev == fsdev`.
- createat allocates an inode and then panics if dirlink fails --
reasonable about a writable filesystem, a crash on a read-only one.
The refusal now happens before anything is allocated. Unlink and
rename already handled a refusal properly.
- a refused create reported ENOENT, so writing to a snapshot said "No
such file or directory" -- the sentence a shell prints for every
create failure. It reports EROFS now.
Verified in the running system against an image built with
--snapshot golden: the mount succeeds, ls lists the snapshot's tree,
/mnt2/etc/motd reads back through it, and a write is refused with
"Read-only file system".
One at a time, and no umount path for it yet: mount_del_by_root removes
the mount table entry, but ksnap_unmount is only called on the failure
paths. schfs-check passes; bootcheck 108 [ok].
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
SchFS is a 64-bit format addressing 64 ZB. The stack under it was not,
and the lowest limit was three orders of magnitude below the one anybody
would have guessed:
ide.c panicked on any block >= FSSIZE -- 150000 blocks, 614 MB
ide.c computed the sector in an int, and addressed it with LBA28
bio.c took a 32-bit block number, capping any device at 16 TB
schfs refused blocks past 0xffffffff, mirroring the cache
The FSSIZE check is the one that mattered. FSSIZE describes the size of a
v1 filesystem *image*; it had no business bounding a physical device, and
it meant a disk larger than half a gigabyte panicked the kernel rather
than being read.
struct buf's blockno is uint64 now, and so is bread's. idestart computes
the sector in 64 bits and picks LBA48 when it does not fit in 28 -- per
request, so everything that fits stays on the LBA28 path every existing
image has been read through. The 48-bit registers are written high half
first, because each is a two-deep FIFO and the other order leaves the
drive addressing whatever the low bytes alone describe.
Verified with a 2 GB disk carrying a marker at 1 GiB -- block 262144,
well past the old panic at 150000. The guest reads it back:
"MARKER-PAST-THE-OLD-LIMIT". schfs-check passes including the round-trip;
bootcheck 108 [ok].
Two limits deliberately left, both noted rather than guessed at: raw
/dev/hdN access still goes through devsw's 32-bit offset, so a raw device
is addressable to 4 GB; and usb_storage computes a 32-bit LBA, which is
what SCSI READ(10) carries -- READ(16) is a different command, not a
wider variable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Groundwork for usb-tablet, which is the fix for the guest cursor and the
host's never being in the same place: a PS/2 mouse reports how far it
moved, two pointers accumulating deltas independently drift apart, and no
amount of care in the compositor can make them agree. An absolute device
reports where it is, so they coincide by construction.
- the HID driver no longer matches the boot subclass only. usb-tablet
declares subclass 0 and was refused at the door; hid_probe decides
what an interface actually is, and still declines what it cannot make
sense of.
- a non-boot HID interface is taken to be an absolute pointer and is
not asked for boot protocol, which it does not have -- that request
failing is what stopped it binding.
- tablet_report decodes the fixed 6-byte layout (buttons, 16-bit X,
16-bit Y, wheel) and scales the logical range to pixels there,
because it is the only place that knows the range.
- input_mouse_abs reports EV_ABS/ABS_X/ABS_Y so a reader can tell an
absolute position from a delta, and mouse_set_abs places the pointer
rather than nudging it.
- Schcom treats EV_ABS as authoritative and EV_REL as a nudge.
The device binds and identifies correctly: "usb-hid: tablet (absolute) on
slot 1, endpoint 0x81 every 4 ms".
It is NOT finished, and QEMU_INPUT is empty by default because of it. No
reports have yet been observed arriving on the tablet's interrupt
endpoint, so making it the default pointer would trade a cursor that
drifts for one that does not move at all. Opt in with
make qemu QEMU_INPUT='-device qemu-xhci,id=xhci -device usb-tablet,bus=xhci.0'
Next step is to find out why nothing arrives: whether the first transfer
is queued but never completes, whether a completion needs re-arming for
this endpoint, or whether QEMU is not routing motion to the tablet at all
(`info mice`).
bootcheck 108 [ok], schcom-test passes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
kernel/dev/mouse.c held the pointer inside literal 800x600 bounds. That
was correct while 800x600 was the only mode the loader could set, which
it was until the resolution became a boot-time choice. At 1280x800 the
pointer simply stopped two thirds of the way across and would not go
further -- which reads as the cursor drifting away from the host's rather
than as a clamp.
vga.c grows vga_get_width/vga_get_height for it. The mode is chosen at
boot now, so nothing above the framebuffer may hardcode its size, and
these are the first two callers that needed to ask.
Verified that the system still builds and boots at 1280x800, and that
bootcheck is 108 [ok] and schcom-test passes. I did not manage to
photograph the pointer past the old limit in the time available -- there
is no reader for /dev/mouse in the image and the screendump attempt did
not resolve, so the fix rests on inspection rather than on a picture.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Nothing outside the kernel could touch a disk. sys_mount does not even
look at the device it is handed -- it mounts the one SchFS volume found
at boot and ignores `source` -- so a program had no way to read a
partition table, format a volume or write a boot sector. An installer is
a program that needs all three, so this is the first piece of it.
A character device in devsw rather than a separate block layer, because
that is where the offset already lives: fileread/filewrite pass f->off
and advance it by what comes back, which is everything a seekable device
needs. The minor is the IDE device number, the same numbering iderw()
uses; two schemes for the same disks would be a bug waiting to happen.
Writes bypass every filesystem above them, which is inherent to a raw
device and why an installer must target a disk it has not mounted.
The nodes are created unconditionally, like every other entry in /dev.
That matters more than it looks: with no node, `echo x > /dev/hdb`
silently creates an ordinary file of that name and appears to work. It
did, and it agreed with itself on read-back; only the host's copy of the
disk showed nothing had been written.
st_rdev was never filled, so every node in /dev reported "0, 0" under
ls -l. Harmless while nothing read it, wrong as soon as something has to
tell two disks apart. Filled with the full Linux dev_t encoding rather
than (major << 8 | minor), which only works while both stay small.
Verified against the host as an independent oracle, both directions: the
guest wrote to /dev/hdb and the host saw those bytes at offset 0; the
host wrote at offset 4096 and the guest read them back with dd skip=.
ls -l now shows 1,1 for console, 2,0 for fb0 and 8,1 for hdb.
Known limit, commented in disk.c: devsw passes the offset as a 32-bit
uint, so a raw disk is addressable to 4 GB. That must be widened before
anyone installs onto a real disk.
bootcheck 108 [ok] with -Fb -Tty -Usb -Sock -Evdev -Ps.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every driver reached its registers with P2V(bar.base), which assumes the
BAR sits inside the 0-4 GB the direct map covers. Under SeaBIOS it always
does. Under OVMF it does not: firmware may place a 64-bit BAR anywhere,
and it puts virtio-gpu's at about 768 GB. P2V produced an address with no
translation behind it, so the driver's first register write took a page
fault, then a double fault, then a triple fault and a reboot loop --
which, coming from a driver bind, reads as broken hardware rather than a
missing mapping.
ioremap() carries each region on a 1 GB page in PDPT entries 4-511, which
were unused: the direct map occupies 0-3, and kpml4[0], kpml4[KERNBASE]
and every process's PML4 all share that one table, so a mapping appears
in every address space at once with no page table to allocate. The pages
are cache-disabled and write-through for the same reason the top GB of
the direct map is -- these are registers, and a write sitting in a cache
line is a write the device never sees.
A physical address below 4 GB returns the direct-map address unchanged,
so the BIOS path, where every BAR is low, is byte for byte what it was.
virtio-gpu now initialises under OVMF (1280x800, queue 64) instead of
triple faulting. e1000 and xhci had the identical latent bug; xhci was
already refusing to bind above 4 GB with a comment saying the mapping did
not exist yet, and now it does.
BIOS unchanged: bootcheck 108 [ok] with -Fb -Tty -Usb -Sock -Evdev -Ps,
e1000 and xhci both binding.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An x86-64 operating system built from xv6, now largely rewritten: SMP,
a copy-on-write root filesystem (SchFS), PCI/USB/virtio drivers, a TTF
rasteriser, and a Linux-numbered syscall ABI.
No history before this point -- the tree was never under version control.
This commit exists so the rename that follows has something to undo to.
Nexoniarz