Found by running Alpine's apk on SchOS, one failure at a time:
- chroot(2) was a stub returning EPERM. Each process now has a root:
where an absolute path starts and where .. stops. Mounts remember
whose root they were made under, and a walk leaves a mount by the way
it came in, so one filesystem can be mounted twice.
- devtmpfs and proc mounted only on /dev and /proc, whatever the target;
a chroot got "already mounted".
- mmap never reused freed address space, so musl's malloc used up the
256 mapping records while 780 MB were free. A mergeable mapping now
goes straight after an anonymous one whose end has room.
- unlinkat returned -1 (EPERM) where apk expects ENOENT, and renameat
was an old copy of rename: no replacing the target, and v1 dirents
written into SchFS. Both now share rename's implementation.
- execve runs #! scripts, and says ENOEXEC for what it cannot run.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
There were none: fcntl answered F_SETLK with EINVAL and flock was not a
system call. Alpine's apk, which runs unmodified on SchOS through
Alpine's own musl, locks its database before reading it and stopped
there.
kernel/fs/lock.c keeps one table of advisory locks keyed by (dev, inum):
flock() for the whole file, owned by the open file description;
F_SETLK/F_SETLKW/F_GETLK for byte ranges, owned by the process and
released when it closes any descriptor on the file or exits; and the
F_OFD_* forms, ranges owned by the description. flock and record locks
do not see each other, as on Linux. Waiters sleep until any release and
are interrupted by a signal.
cmd/locktest checks each between a parent and a child, including range
splitting, F_GETLK's report, and a blocked waiter being woken.
shell.nix gains pkg-config, meson, ninja, m4, bison, flex and xz, for
the X ports.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Xorg on a framebuffer -- the fbdev video driver, the evdev input driver,
libpciaccess -- asks the kernel a fixed set of questions before it draws
anything, and gives up on the first wrong answer. These are those
answers.
evdev: EVIOCGVERSION, GID, GNAME, GPHYS, GUNIQ, GPROP, GBIT, GKEY, GABS,
GREP and the accepted setters. Xorg and libinput decide what a device
*is* from the capability bitmaps; /dev/input/eventN answered only
EVIOCGRAB, which reads as no device at all. Each device's bitmaps are
its defaults plus whatever it has reported, so a USB tablet makes the
pointer absolute.
fbdev: FBIOPUT_VSCREENINFO succeeds for the mode already set, as on
Linux -- the fbdev driver sets it every time it starts -- and still
refuses any other. FBIOPAN_DISPLAY to the origin; FBIOBLANK.
The console: KDSETMODE(KD_GRAPHICS) stops it drawing on a terminal,
KDSKBMODE(K_OFF) stops keys reaching its tty, and VT_PROCESS now does
the switch handshake: leaving the terminal sends relsig and waits for
VT_RELDISP, arriving sends acqsig. A terminal whose owner has died goes
back to the kernel whole, so a crashed X cannot leave the console dead.
/sys: boot_vga, a writable enable, and graphics/fb0 under the display
device, which is how the fbdev driver pairs a card with /dev/fb0.
ports/linux-headers: <linux/*.h> into musl's prefix, from sabotage-
linux's kernel-headers 4.19.88-2. Its checksum is recorded from the
first download; upstream publishes none.
cmd/xready makes each call and checks the answer, including the switch
handshake on tty2 and an owner that dies holding it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
An ASUS TUF A15 booted with its internal keyboard dead and a USB one
fine. These AMD laptops' firmware describes IRQ 1's polarity wrongly --
Linux keeps a quirk list -- and an edge line programmed the wrong way
up never fires: the controller holds its byte, and only the interrupt
that never comes would read it.
irq_watch() lets a driver say how to ask its device whether something
is waiting. The poller asks every tick; 50 ms of waiting with nothing
delivered turns the line's polarity over and drains the device, which
re-arms the edge, and if it still says nothing the line is polled.
Both are announced. The keyboard and the PS/2 mouse are watched.
ISA lines now also take their mode from the MADT's override entries,
which were parsed and never read.
Tested in QEMU, which ignores polarity on edge lines, with the keyboard
line masked: the first key turns it over and typing works; masked again
after the flip, it is polled and typing still works.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The set libwayland's event loop is made of, in kernel/sys/sys_event.c:
four kinds of descriptor with no inode, FD_EVENT, that read() drains
and poll() reports on. They were table entries returning ENOSYS that
read as done.
epoll registers (descriptor, file) pairs, so a closed descriptor
leaves the set; level- and edge-triggered, EPOLLONESHOT. Waiting is
poll()'s, a tick at a time. The old eventfd and signalfd numbers (284,
282) were aliased to the new calls and read flags from a garbage
register; they are their own entries now, with epoll_create and
epoll_pwait.
cmd/evtest checks each, then all four in one set. The plans are
brought up to date: everything Wayland was waiting on is built except
setsockopt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
sigsend() ended a process for a default-fatal signal (SIGUSR1, SIGTERM,
SIGINT...) on arrival, whatever its signal mask said. Blocking is
exactly the promise that it will not: the signal waits until it is
unblocked, or until the program takes it some other way -- which is
what signalfd is, and how this was found.
A blocked signal is now only made pending. If it is later unblocked
while still pending, deliver_signals() applies the default action
then. SIGKILL cannot be blocked, as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
From 1 MB the 34 MB kernel ran across 8 MB, where OVMF keeps ACPI NVS
memory that outlives the firmware -- so it was overwritten on every
boot, after ExitBootServices. Firmware leaves its leftovers low; above
16 MB is where a machine is most reliably empty.
KERNPHYS in memlayout.h, and kernel.ld agreeing with it. Nothing else
named the old address. Under OVMF the range is now only boot-services
memory: staged, copied home after the firmware exits, no warning. Boots
as a CD and as a USB disk, 1 and 4 CPUs.
The 15 MB below the kernel is not given to the page allocator yet.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
- schinit, schinitd and vgpud build every message from pieces, one
write() each, so a job starting on another CPU landed mid-line:
"schinitd: starting vgpud: os.schos.resolvconfno virtio-gpu on this
machine". They now collect a line and write it at the newline;
resolvconf's verbose line is one write with its prefix.
- The ACPI SCI registers IRQ_QUIET. A desktop's SCI never fires, and the
poller took two seconds of silence for a misrouted line ("irq 9:
nothing has arrived on it") and called the handler every tick. A line
is only exempt while every handler on it is quiet.
- The loader claims each segment from where the last one's claim ended,
so two segments sharing a page no longer ask for it twice. When a
claim is still refused, it names the memory-map entries in the way.
- Stale divesymlink() declaration gone; the kernel.md struct table
re-measured (inode, buf and file pools grow from slabs now).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Sixty-five headers sat side by side in include/, so `#include "irq.h"`
said nothing about where irq.h was or what it belonged to, and errno.h,
stat.h and fcntl.h at the top shadowed the system's for any host tool
built with -I./include. They are in folders now -- arch/ core/ fs/
dev/ gfx/ net/ uapi/ (and ldls/ for the Linux driver layer) -- and
every include names the folder, as Linux's do: "core/proc.h".
Done mechanically and checked the hard way: the kernel built from this
tree disassembles identically to the one built from the tree before
it. The full image boots, and userland is the same set of programs.
Paths in docs and comments follow; docs/guides/layout.md has a table
of what is where, and tools/fix-includes.py rewrites old-style
includes from the folders in the tree, for code written before this.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Nothing ever saved the FPU and SSE registers on a context switch.
musl's memcpy, the compiler's struct copies and any arithmetic on
doubles keep live values in %xmm, so two processes sharing a CPU wrote
over each other's: two shells counting in a loop on one processor took
a NULL+8 page fault within seconds, at a different rip every time. It
reproduces on kernels from before this week too -- it had always been
there, and every crash nobody could explain was a candidate.
sched() now fxsaves the process leaving the CPU and scheduler()
fxrstors the one arriving. A new process starts from the ABI's state
(FCW 0x37F, MXCSR 0x1F80), fork copies the parent's live registers,
exec resets them, and a signal handler's are put back by rt_sigreturn
so the interrupted code finds what it had.
The kernel was free to use %xmm too -- GCC vectorised struct copies
and memset into 153 instructions of it -- so an interrupt could
clobber the registers of the program it interrupted with no switch at
all. Kernel, lwIP and uACPI objects are built -mgeneral-regs-only now
and the image has none; cmd/ keeps SSE.
Checked under QEMU: three concurrent loops three times over, on 1 CPU
and on 4, no faults; before, every run faulted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The scheduler was xv6's round-robin: run everything RUNNABLE in slot
order and yield on every timer tick. Slot order is not fairness, a
process alone on a CPU paid a full context switch a hundred times a
second to be handed straight back to itself, and there was no way to
say that one thing mattered more than another.
Every process now carries vruntime, the CPU time it has used divided
by its weight, and the scheduler runs the runnable process with the
smallest. Fairness and interactivity both fall out of that one rule;
nice is a weight, 1.25x per level as UNIX has always had it. A slice
is 20 ms shared among the runnable processes, never below 2 ms, and
the timer tick preempts only once it is spent. A process waking from
a long sleep is brought up to min_vruntime less one period, so it is
scheduled promptly without being owed everything it slept through;
fork inherits vruntime, so forking does not reset a debt.
getpriority/setpriority are real, with Linux's offset of 20 on the
return; /proc/<pid>/schedstat and the ctxt_switches lines of status
report what the scheduler did. docs/architecture/scheduler.md has the
reasoning, including why the run queue is still a scan of ptable.
This was written in an earlier session and left uncommitted. Finished
here: a signal waking a sleeper goes through the same vruntime floor
as wakeup() does, and sched_yield() yields -- it returned 0 doing
nothing, harmless while every tick was a yield and a spin of a whole
slice once ticks stopped being one.
Checked under QEMU on 1 and 4 CPUs: a nice 10 loop beside a nice 0
one gets about a tenth of a shared CPU, and a command typed meanwhile
answers in 26 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Every page of RAM went onto the free list at boot, which means writing
each one -- a link, and kfree()'s 4 KB of junk -- before the first
process ran: 64 GB of memset on a 64 GB machine to say it was free.
Memory nobody has used yet is kept as a cursor and an end per region
now, and kalloc() takes from those once the list of given-back pages
is empty. A page is first touched by whoever asked for it.
That makes the cap on RAM cheap to raise, and PHYSTOP_MAX goes from 64
to 256 GB, which still leaves 248 GB of the kernel PDPT for device
windows. MEMREGION_MAX goes from 32 to 64 for the firmware maps that
cut RAM into many pieces.
kalloc_free_bytes() walked the whole free list under the allocator's
lock -- two million links on 8 GB, every second top or fastfetch was
running, with every CPU that wanted a page waiting behind it. It reads
a counter now. And the buffer cache, which keeps what it grows to, is
reported as Buffers in /proc/meminfo and sysinfo(2) rather than as
memory used by nothing.
Checked under QEMU at 1, 2 and 6 GB: MemFree right, zsh splitting
200000 words, fork loops, a 74 MB write.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The fixed tables that a bigger machine than the one they were written
on reaches first, each raised to where it costs little:
NPCI 64 -> 256 a desktop with cards and Thunderbolt passes 64,
and the devices past it -- often the disk or
NIC last in bus order -- were ignored; 44 KB
NSOCK 32 -> 128 Unix sockets, which a desktop session uses by
the dozen; 4.9 KB each, 470 KB in all
NDEVICE 32 -> 128 120 bytes each
NIFDEV 16 -> 64 one dock is a hub, Ethernet, audio and a card
reader at two or three interfaces apiece
NETIF_MAX 4 -> 8
Device names took one digit and clamped there, so the tenth device of
a kind and every one after it was named "...9". Two digits now.
Checked under QEMU with two bridges of test devices: 68 functions,
all in lspci, /proc/pci and /sys/bus/pci.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
NCPU was 16, under what a desktop has today. 64 costs 48 more TSS
I/O bitmaps and scheduler stacks, about 800 KB of BSS. Past 254 the
APIC would need x2APIC mode, which is a separate change.
With more than 16 there were two other ceilings. sched_getaffinity
wrote "CPU 0" and returned 0, and musl clears every byte past the
count returned -- so the mask came back empty and nproc, make -j and
sysconf(_SC_NPROCESSORS_ONLN) all said 1, on any machine. It returns
the mask of processors that started, and its size, now. And /proc
rendered every file into 2 KB, which /proc/cpuinfo ran out of at the
eighteenth processor without a word; /proc/pci, /proc/interrupts and
/proc/mounts were each one bigger machine from the same cut. There is
one 64 KB buffer for it now behind a sleeplock, since there is no
allocator for contiguous pages, and the generators that write raw
are given the real capacity instead of a constant.
Checked under QEMU with -smp 24: nproc 24, /proc/cpuinfo whole.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The inode table was NINODE (200) entries, searched end to end on every
iget(), and a panic -- "iget: no inodes" -- when all of them were held.
Held means open, a cwd, a running binary or a directory mid-walk, so a
few busy processes reached it. iget() now finds an inode through a hash
on (dev, inum), takes a free one from a list, and makes more from a
slab cache once the NINODE in the image are all in use.
The open-file table had the same shape, NFILE (512) and then ENFILE,
and gets the same treatment: a free list, and the slab past NFILE.
/proc/fsstat's peaks now read against how many exist.
getrlimit(RLIMIT_NOFILE) said unlimited, which musl turns into
sysconf(_SC_OPEN_MAX) == -1, after which zsh believed it had no
descriptors and `exec {fd}<file` failed on the first one. It reports
NOFILE now. And a comment in sys_fs.c held a "*/" that ended it early.
Checked under QEMU: six processes holding 115 files each, 713 open
descriptions and 341 inodes at once, all back to the baseline after.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The cache was NBUF buffers, 1 MB, on every machine: enough for one with
tens of megabytes and nothing on one with gigabytes, which re-read the
same directory and binary blocks off the disk for every command. It
still starts as those 256 in the kernel image, and a miss with every
buffer busy now makes a new one rather than evicting, up to 1/32 of RAM
or 256 MB. Growing on demand means nothing is spent that is not used,
and binit() needs nothing from the page allocator.
Lookups go through a hash on (dev, blockno) instead of walking the LRU
list, and bflush() walks a creation-order list with an unlocked filter
first, so neither costs more at 65536 buffers than it did at 256.
struct buf's data is a pointer to its own page now; the drivers copy
through their own DMA buffers, so where it sits does not matter.
On a 2 GB QEMU guest: a 40 MB file read back in 0.89 s instead of 1.83,
the big binaries in 0.17 s instead of 0.26. /proc/fsstat shows the
cache as buf_cached now / most.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
exec took 32 arguments and 64 environment variables, in fixed arrays,
so a glob over more than 31 files failed with E2BIG. It now reads the
caller's argv and envp where they are, through the uaccess helpers and
a bounce page -- a second thread rewriting the array after a check
cannot slip a kernel address in -- and sizes the new stack to fit
them. The limits are Linux's: ARG_MAX, 2 MB of strings and pointers,
and 128 KB for any one string (USER_STR_MAX follows it).
Every mmap took one of a process's 256 VMA slots, and zsh takes one per
16 KB heap block, so `x=($(<file))` over 30000 words ran out after 4 MB
of heap and exited 1 without a word. Neighbouring anonymous mappings
now share a record, as Linux merges them -- the bump allocator puts
each new one where the last ended -- and munmap trims or splits it.
Partial unmaps of device or segment mappings are refused with EINVAL
instead of kfree()ing pages that are not the process's, and a failed
allocation half way through a mapping no longer leaks what it had.
Checked under QEMU: touch over 3000 names, /bin/echo with 50000
arguments, E2BIG past 2 MB, zsh splitting 100000 words, fork loops.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
hwreport writes one folder to the first read-write FAT mount, or mounts
the first USB stick that takes vfat itself and unmounts it afterwards:
summary.txt (machine, firmware, CPU, memory, every PCI and USB device
with its driver or "-- no driver"), dmesg.txt, lspci.txt in the shape of
lspci -nnk -xxx, lsusb.txt, dmi.txt, memmap.txt, the ACPI tables as raw
.dat files iasl -d reads, and copies of /proc. Serials, the UUID and the
MSDM/SLIC tables (a Windows licence key) are left out.
It reads /sys, which is new: sysfs at Linux's paths and in Linux's
formats -- /sys/bus/pci/devices with uevent, config and resource,
/sys/bus/usb/devices, /sys/firmware/acpi/tables, dmi/tables, memmap
and efi, /sys/class/dmi/id. toybox's lspci and lsusb work unmodified
against it. The USB core now keeps a table of attached devices for it.
The machine's name comes from SMBIOS (kernel/dev/dmi.c). The UEFI loader
hands over the entry point from the configuration table, the kernel
falls back to the 0xF0000 scan, and the log gets Linux's "DMI:" line.
Two bugs found on the way:
- The loader handed over at most 40 memory-map entries, in firmware
order. A laptop's UEFI map is 60-150 descriptors with RAM above 4 GB
listed last, so that is what was dropped. The entries now have their
own area at 0x31000 with room for 256, are sorted before merging, and
the kernel's 32 usable regions keep the largest when there are more.
- MemTotal and sysinfo's totalram were PHYSTOP, an address, which counts
the PCI hole: a 6 GB machine reported 8. Now the usable RAM, summed.
usb.ids moves to /usr/share/hwdata, where lsusb looks for it, and
pci.ids joins it (`make pci-ids`, BSD-3-Clause like usb.ids).
Checked under QEMU q35: the report written to a FAT32 stick, fsck.fat
clean, DSDT.dat disassembled by iasl on the host, lspci -k and lsusb
showing names, 6 GB of RAM seen as 6137 MB.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
NBLKDEV was 16 for every disk and partition together, which a laptop's
NVMe drive with its eight partitions and a USB stick come close to. The
mount table held 8, SchFS 4 volumes and 4 snapshots, FAT 4 volumes.
Now 32, 32, 16, 8 and 8; the device numbers still fit between the ones
already fixed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Every path component was copied into a DIRSIZ buffer and every path
into a MAXPATH one, so names were cut at 29 bytes and paths at 127
without an error. Two names sharing their first 29 bytes opened the
same file, on SchFS and FAT alike, though both hold 255.
VFS_NAME_MAX (255) and VFS_PATH_MAX (4096) are the limits now, and
anything longer is ENAMETOOLONG. namex_from() takes its path buffers
from kalloc(), a page each, rather than putting two of them on an 8 KB
kernel stack. The v1 format still stores DIRSIZ bytes, and refuses a
longer name instead of matching on its first 30.
Checked: 255-byte names on SchFS and FAT (fsck.fat clean, and the FAT
name answers to its uppercase form), 256 refused, a 200-byte path and a
link to it, and no pages lost over 700 lookups through links.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
kd_write() went through bread(), so every block a write allocated was
read from the disk first and then replaced whole. That was a quarter of
the disk operations of a SchFS write. bgetblk() is bread() without the
read, as Linux's getblk() is.
1 MB on AHCI under QEMU: 2.0 -> 2.8 MB/s in 4 KB writes, 4.5 -> 5.6
MB/s in 64 KB ones. 3 MB of random data written and copied through the
cache's evictions reads back with the same md5.
The 63 KB/s seen earlier is the legacy IDE driver on QEMU's default
machine, at about 10 ms per emulated PIO write; the same disk on AHCI
(q35) is 40 times faster.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Symlinks are followed in namex_from() now, in the middle of a path as
well as at the end, and a relative target resolves from the directory
that holds the link rather than from the caller's cwd. `ln -s f d1/s;
cat d1/s` said "Symbolic link loop" because it looked for ./f. A loop is
ELOOP after 40 links, a dangling link ENOENT, and open(O_CREAT) through
a link to a file opens the file. divesymlink() is gone.
mknod from a Linux program passes (mode, dev) where the native call
passes (major, minor), and was read as the native one: `mknod n c 1 2`
made node 432,258, and mkfifo made a character device. Both are decoded
now; a FIFO is EPERM, and an existing name is EEXIST.
/proc/mounts listed devtmpfs twice: once from a fixed line and once
from the mount table. The fixed line is gone, and mount_add() refuses
a filesystem mounted on its own root, which is where a second devtmpfs
mount lands. A read-only FAT mount says ro.
/proc/partitions exists, in Linux's format, and /proc/filesystems lists
vfat.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The filesystem every other system can read: a USB stick, an SD card, the
EFI system partition. It is here first for the hardware report, which has
to leave a machine on something that can be read on another one.
mount -t vfat /dev/usb0p1 /mnt (msdos and fat work too; -o ro)
Written from Microsoft's specification, and matched to Linux vfat where
the specification leaves a choice: 8.3 names with the NT lowercase bits,
long-name entries otherwise, Linux's alias rule (REPORTS, then BIGNUM~1),
case-insensitive lookup, UTC times, the boot-sector dirty bit, every FAT
copy and FSInfo kept up to date. It uses its own sector cache over the
new blkdev_xfer, since a partition need not be a whole number of 4 KB
blocks, and every operation that changes something reaches the disk
before it returns. Inode numbers come from the entry position, or the
first cluster for a directory, plus a small table for entries that moved.
Around it:
- /dev lists every disk and partition the block layer found (major 8,
minor = block device), and raw reads and writes on them go by sector.
- The ESP and MBR FAT partitions are registered, typed so the root probe
passes over them. MBR entries with a bad status byte are skipped, so a
superfloppy's boot code is not read as a partition table.
- statfs, sync, fsync, umount (EBUSY while in use), shutdown, reboot and
the power button all know about FAT.
- create undoes and reports a failed ialloc or dirlink (ENOSPC, EINVAL,
EPERM) instead of panicking; fs_ops gains refuse_types and no_hardlinks.
- filewrite returns a short write instead of panicking on it.
- utimensat is implemented. As a stub returning 0 it made toybox touch
believe the file existed, so `touch newfile` created nothing on any
filesystem.
Tested in QEMU against mkfs.vfat images (FAT12 floppy, FAT16 on MBR,
FAT32 superfloppy and GPT ESP) by writing, renaming, deleting, filling to
ENOSPC and truncating there, then checking with fsck.fat -n and mtools on
the host: all clean, contents match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
On the first real machine with a discrete card (an RTX 3060, 800x600 from
the UEFI GOP) the console was slow enough to photograph mid-scroll, and the
boot died on CPU 1's "cpu1: starting": CPU 0 spun 200M times on the console
lock while CPU 1 held it, scrolling, and called it a deadlock.
The framebuffer is device memory in the uncached 3-4 GB page of the direct
map. QEMU's stdvga is RAM underneath, so nothing showed there; on a card,
gfx_scroll_up's memmove read the whole screen back across the bus one byte
at a time, and every anti-aliased glyph read back each pixel it blended.
So the console now draws into a copy of the screen in RAM (vmap_alloc: kalloc
pages behind a page table of their own, since nothing needs it contiguous),
and gfx_flush() copies the changed rectangle onto the card through an
ioremap_wc() mapping -- PAT entry 4 set to write-combining on every CPU.
Flushes happen where drawing ends: a message, a terminal write, an echo, a
terminal switch. Scrolling moves eight bytes at a time instead of one.
/dev/fb0 reads and writes the card itself, as before. The early console's
fb_info now carries the physical address it always should have.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
The firmware's AML is now interpreted. uACPI 6.1.1 (MIT) is fetched by
ports/uacpi, checksum-verified, and compiled into the kernel unmodified,
the same way lwIP is. kernel/acpi/host.c is its host layer: memory,
mapping, port and PCI config I/O, time, mutexes and events that spin
before the scheduler exists, the SCI, and an "acpi" thread for the
work the SCI defers.
PCI interrupts go where the board says. pci_init() now routes every
device's INTx through the root bridge's _PRT before drivers bind. It
follows link devices' _CRS, uses a bridge's own _PRT where the namespace
has one, and swizzles through bridges that don't. Before this, INTx
waited on the 8259's line number, which is the wrong pin in APIC mode.
irq_register() takes a trigger mode rather than a level/low flag,
because q35's links and most SCIs are level, active-high. The MADT's
interrupt source overrides are now parsed, and are what places the SCI.
The power button syncs the filesystems and powers off. shutdown() and
reboot() go through the firmware's S5 and reset register first, and
only then fall back to the emulators' ports, which a real board does
not answer.
Tested under QEMU: q35 with MSI off routes xHCI to GSI 22,
level/active-high, and it takes real interrupts with nothing polled.
system_powerdown powers off q35 and i440fx (QEMU exits 0), and q35
with 4 CPUs boots from NVMe with USB hot-plug working alongside.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1giNvrHTrPMPw7V6x1kQ4
Devices can now arrive and leave after boot. A "usbhotplug" kernel thread
handles root-port changes, which it is told about by Port Status Change
events and also checks for every quarter second. It also handles changes
behind hubs, which each hub reports on its status-change endpoint. An
unplugged device has its endpoints closed and any transfer in flight
failed. Then its drivers' disconnect hooks run, the slot is disabled and
its memory is freed. Everything behind a hub that goes goes with it. This
needed the interrupt handler to file command and endpoint-0 completions
for their waiters, so that enumeration can run while keyboards interrupt.
On real hardware: the controller is taken from the firmware through USB
Legacy Support before reset, with SMIs off. On Intel 7/8/9 series parts
the ports are switched over from EHCI, which is where firmware leaves
them.
Every interface of a device is offered to the class drivers, not just
the first, so a receiver's mouse on interface 1 gets a driver. Configure
Endpoint now copies the slot context from the controller. Before, it
rebuilt it and lost the route string and TT fields of devices behind
hubs.
HID: pointers are read in report protocol from their report descriptor.
That covers buttons, X/Y and wheel, relative or absolute, and report IDs.
Boot protocol is kept as the fallback for mice and for keyboards.
Non-pointer HID interfaces are declined rather than treated as tablets.
USB keyboards get software key repeat. Mass storage handles up to four
sticks, each unregistered from the block layer when it is pulled out,
along with its cached blocks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HH7xukt4aaiGj1nQEAoQf1
q35 stopped with "no root filesystem" because its disk is SATA, like
every PC since 2008, and SchOS only had IDE. And SchFS could only sit
on a whole raw device, so nothing could boot from a disk that also
carried an ESP.
- include/blkdev.h, kernel/dev/blkdev.c: a registry of block devices
found at probe time, numbered from 4. A disk's GPT (or MBR) is read
when it registers, and each partition becomes a device whose block 0
is the partition's first sector. SchFS needs no format change to live
in a partition; a Linux SchFS driver will see the same bytes on
/dev/sdX2. SchFS has its own GPT type:
47138FA9-B79C-47A6-9F7D-AC1F7F572D5E.
- AHCI driver, written from the AHCI 1.3.1 spec: BIOS handoff,
COMRESET when the link is down, IDENTIFY, 48-bit DMA read/write,
FLUSH. Completion is polled, so the root disk does not depend on
interrupt routing.
- USB sticks also register as usb0, so their partitions are found.
- The root probe tries registered devices between USB and IDE,
SchFS-typed partitions first.
- make disk / tools/posix/mkdisk.sh: a GPT image with a 64 MB ESP and
a SchFS partition, ready to dd to a stick or a disk.
Tested in QEMU q35: the old two-disk setup roots on sata1; disk.img as
the only SATA disk roots on sata0p2; disk.img as a USB stick (firmware
boots from it) roots on usb0p2. Writes survive sync in each case.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HH7xukt4aaiGj1nQEAoQf1
PCI drivers took their interrupt from INTERRUPT_LINE. On real hardware
that is the 8259's number, not the IOAPIC pin the device is wired to
(that is in ACPI _PRT, which needs AML), so nothing ever arrived. xHCI
is entirely interrupt-driven, so the USB keyboard and a USB root would
have hung even on a clean boot. Under QEMU the two numbers agree,
which is why it never showed.
- pci_irq_attach()/pci_irq_detach(): MSI, then MSI-X if the driver
allows it, then INTx. The handler is registered before MSI is
enabled, so no edge is lost. xHCI, e1000, r8169 and rtl8139 use it.
- Lines 24-29 are MSI vectors. The LAPIC error vector moves from 19,
which is a PCI pin on Intel chipsets, to 30. Nothing is placed at or
above vector 64 (T_SYSCALL).
- A storming line is masked and its handlers are polled instead of
being left to eat the CPU. A level line that has had a handler for 2 s
and never fired is polled every tick by a kernel thread.
Tested in QEMU q35 with the root on a USB stick: xHCI and e1000e on
MSI; then MSI disabled and INTx deliberately misrouted, where xHCI got
zero interrupts and the root still mounted through the poller. i440fx
INTx and the rtl8139 hand-over are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HH7xukt4aaiGj1nQEAoQf1
Every one of these works under QEMU and stops a real machine:
- MADT processors that are "online capable" but not enabled were
counted as CPUs. AMD firmware lists absent cores that way (so does
QEMU with maxcpus), and startothers() then waited for ever for an
AP that does not exist. Only enabled entries count now, x2APIC
(type 9) entries are read too, and the wait is bounded to a second;
an AP that misses it is parked with INIT and the boot goes on.
- The PIT calibration loop had no bound. Recent boards clock-gate the
8254: the output never rises (a hang) or floats high (TSC "0 MHz"
and every clock thousands of times fast). The loop is bounded and
checked, and falls back to the ACPI PM timer, then CPUID 0x15/0x16.
Under QEMU pit=off the PM timer gives the same figures the PIT did.
- Firmware that hands over in x2APIC mode left lapic[] reading
garbage: cpunum() panicked or the timer never started. Each CPU is
switched back to xAPIC first (x2APIC -> disabled -> xAPIC).
- The IOAPIC was always at 0xFEC00000 and interrupts were aimed at
APIC ID "0". Both now come from the firmware's tables.
- INIT-to-STARTUP was 100 us ("Bochs is slow"); it is the 10 ms the
MP specification asks for.
- mouseinit() drained the 8042 with an unbounded loop. With no
controller the port reads 0xFF and the loop never ends. It now
notices there is no controller and skips the mouse.
Tested in QEMU: i440fx and q35, -smp 4, -smp 2,maxcpus=4, pit=off.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HH7xukt4aaiGj1nQEAoQf1
Nexoniarz is the sole author and SchOS was never published under
Apache-2.0, so this is a straight relicence. LICENSE is the FSF GPLv2 text,
every first-party SPDX header moves to GPL-2.0-or-later, and NOTICE,
README.md, docs/licensing.md and the site say so. projects/schfs-linux
follows the rest of the tree.
The GPL gives copyleft: a fork has to ship its source. It also lifts the
old "Linux drivers, never" rule, since GPL-2.0-only code combines with
GPL-2.0-or-later. "Or later" leaves room for GPLv3, but one binary still
has to pick one version, so docs/licensing.md now keeps Apache-2.0 and
GPLv3 code out of the kernel to keep Linux code possible there. BSD
4-clause code goes from "avoid" to "no", because the advertising clause
is GPL-incompatible.
LDLS still ships no .ko. That boundary is now project policy rather than
a legal requirement, and ldls.md, devices.md and licensing.md section 6
say so. refmod declares license=GPL, and ldls-check passes (27 checks).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three things, and the first one is the freeze.
**The loader jumped out from under itself.** A machine with a lot of RAM
stopped at "Starting the kernel" with the splash still up: no panic, no reset,
nothing. The firmware's own exception handler had the answer the whole time,
once a machine was found that printed it:
X64 Exception Type - 0E(#PF - Page-Fault)
ExceptionData - 0x10 I:1 P:0
RIP - 0000000140004DEE <- the loader's own code, at 5.4 GB
RAX - 0000000000001000 <- the page tables it had just installed
RCX - 00000000001000F0 <- the kernel it was about to jump to
I:1 P:0 is an instruction fetch from a page that is not present. The loader
built page tables covering four gigabytes, loaded cr3 -- and the firmware had
placed the loader image *above* four gigabytes, so the next instruction fetch
was unmapped. It reproduces exactly on `-m 8G` and never on `-m 1G`, which is
why every test here passed.
The PDPT is one page whether four entries are filled or all 512, so all 512
are: 512 GB of identity map, replaced by kvmalloc() a few hundred instructions
later.
**And 3 GB was never the limit, it was a circle.** PHYSTOP_MAX was 0xC0000000
because kvmalloc() mapped exactly four 1 GB pages, and kvmalloc mapped four
because that was the cap -- each holding the other up, throwing away thirteen
gigabytes of a sixteen gigabyte machine. kvmalloc now maps as many gigabytes as
the machine has RAM and tells ioremap where the direct map stopped, instead of
ioremap assuming it stopped at four.
Freeing memory becomes two passes, and the ordering is forced rather than
chosen: kinit1() runs *before* kvmalloc because kvmalloc calls kalloc for its
own page tables, and freeing a page means writing into it -- so kinit1 can only
touch what the loader mapped. kinit_high() adds the rest afterwards.
`phystop` and the new `lowtop` are now two different numbers: the DMA arena
stays below 4 GB because a device is not obliged to reach the top of memory,
and an RTL8139's buffer address register is thirty-two bits wide.
**And the 0 MB.** The System Information screen reported "0 MB usable" on a
machine with sixteen gigabytes in it, because memory_mb() had its own 16 KB
guess at how big a memory map gets, and returned zero when GetMemoryMap said
EFI_BUFFER_TOO_SMALL. That is the same bug as the previous commit's, in a
second copy, and this one failed silently. There is one buffer and one
map_read() now.
Verified at 1 GB, 8 GB and 16 GB: all three boot, no traps, no panics, socktest
and memfdtest green in each. 8 GB and 16 GB did not boot at all before -- and
16 GB now reports 17408 MB with "mem: 13312 MB above 4 GB added to the
allocator", where the old kernel would have used three. Both display
configurations still clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
schdbgd ships Disabled, and its manifest explains why: a debugger that can read
the process table should not run on a machine whose owner never asked for one.
That default is right and has one bad consequence -- **the machines that most
need debugging are the ones you cannot get a shell on to enable it.** A laptop
with no serial port that stops before it can draw has nowhere to say why, and
`schinitd enable` needs a prompt that does not exist.
So `make DEBUG=1` adds a sixth entry to the F8 menu. Picking it toggles;
booting carries the choice through SCHBOOT_DEBUG_FLAG; the kernel publishes
/proc/bootdebug; schinitd starts the one job that flag may un-disable, with
the manifest's Args now naming 0.0.0.0 -- a debugger reachable only from the
machine being debugged is no use when that machine is the one that will not
talk to you.
Four decisions, since this is a network-reachable debugger:
compiled out, not hidden A release build cannot set the flag because the
code that would set it is not in the binary. A
menu item that is present and refuses is an
invitation to make it not refuse.
consent at the machine Somebody standing at the keyboard during boot is
the strongest consent this system can ask for and
the weakest one that is any use on hardware that
will not boot.
one job, named in schinitd Not a manifest key like EnableIfDebug: the general
version of "the boot menu starts things that ship
off" is worse than the specific one, and this way
what the flag can start is one readable line.
the kernel only publishes Nothing in the kernel behaves differently, so a
debug kernel and a release kernel are the same
kernel and a bug cannot be the flag's fault.
/proc/bootdebug is its own file rather than a second line in /proc/bootmode,
because bootmode's whole promise is that reading it is a string compare and
not a parser.
**Also: `tone` stopped lying.** On a machine with no sound card it reported
"the device will not take 16-bit samples", which sends you looking at the
format code instead of at the empty PCI slot -- /dev/dsp exists whether or not
hardware is behind it, and the ioctl is where the absence shows up. ENODEV is
now told apart from a format refusal.
Verified end to end: F8 -> item 6 -> boot -> schinitd says it is starting
schdbgd -> a client outside the VM connected to the forwarded port and ran
`dma`, getting the real arena dump back. The release build has no such entry
in the binary at all, and boots clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A program that wants to make a sound opens /dev/dsp, writes PCM, and closes
it. That is what UNIX audio has always been and it now exists here:
tone 440 2
Three pieces, in the order a sample passes through them. **/dev/dsp** is a
character device where write() is audio and a handful of OSS ioctls say what
the samples are. **AudioEngine** is everything above a driver that is not
specific to a chip -- the format, whether there is room, and what a write does
when there is not. **A driver** is bound by class, exactly as a display driver
is: device_find_class(DEVCLASS_AUDIO).
OSS numbering rather than something new, because every UNIX audio program ever
written already knows it and the only thing a different set would achieve is
making ports harder. They are read/write ioctls, which is the part that
matters: a program asks for 44100, the device sets 48000 and writes 48000
back, the program resamples. One that ignores the answer plays 9% fast, and
that is its bug rather than the driver's.
**The engine is deliberately one stream.** Mixing several programs is a real
feature with a real design behind it, and building it before anything had
played a single sound would be exactly what docs/plan/1.0.md keeps a register
for: work that exists and is connected to nothing. The second program that
wants to play at the same time is what should cause it to be written.
**audio_wait() sleeps on the timer tick, not on the device.** That looks like
the lazy choice and is not: a ring that drains is a ring with room, so asking
again in ten milliseconds always makes progress, whereas sleeping on the card
makes audio depend on an interrupt arriving. On a machine whose line is
misrouted that is not a stutter, it is a writer that never wakes. The handler
still accounts for finished buffers -- deleting it would cost efficiency, not
correctness -- and it retires descriptors through the same `tail` pointer
queue() uses, so the two cannot disagree about what is in flight.
**AC'97 first because HDA is two unproven things at once.** It is the simplest
audio device that exists on real machines and in QEMU both. Intel HDA is what
the laptop actually has and is a much larger driver; writing it against an
interface nothing had ever played through would have meant debugging the
driver and the interface simultaneously. HDA arrives as a second
struct audio_ops and nothing above it changes -- which is the entire point of
the class.
Two things the wiring needed: PCI class 0x04 now maps to DEVCLASS_AUDIO (the
driver bound and /proc/devices showed it, but audio_get() found nothing,
because the class byte had never been translated), and include/x86.h grew
inw() -- AC'97's mixer is all 16-bit registers and reading one a byte at a
time returns the low half twice on some codecs.
**Tested by measuring what came out, not by not crashing.** QEMU's wav backend
writes the guest's audio to a file; `tone 440 2` produced 1.97 seconds of
44.1 kHz stereo with peak amplitude 14177 against the generator's table
maximum of 14178, and 880 zero crossings per second -- 440 Hz exactly. wav is
on every QEMU target rather than a host backend, because a build machine may
have no sound card and it is the only backend whose output can be checked.
Both supported configurations: one boot each, no traps, no panics, socktest
and memfdtest green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first piece of real-hardware work, and it is three bugs that QEMU cannot
show you. The report was an ASUS laptop -- Ryzen 7000-series APU, Radeon
integrated, RTX discrete -- booting to a white underscore on a black screen,
and unable to be offered 1920x1080.
**The framebuffer address was truncated.** The loader handed over
`(UINT32)gop->Mode->FrameBufferBase` and the kernel read a 32-bit field out of
a VBE mode-info block -- a *BIOS* structure the UEFI path had been filling in
because the kernel already knew how to read one. A machine that puts its
aperture above 4 GB, which is ordinary on a laptop with a discrete GPU, got a
plausible-looking address pointing at nothing. The splash went into it, the
kernel was handed it, and the firmware's own cursor kept blinking on an
otherwise black screen. Every step in that sequence succeeded on a number that
was wrong, which is why there is no error message anywhere in it.
SCHBOOT_FB_BASE is the loader's own display block: 64-bit address, 64-bit
size, 32-bit width, height and stride, and the pixel format. The VBE block is
still written for the BIOS loader, but anything that does not fit is left at
zero rather than wrapped -- a truncated address is this bug, and a zero is at
least honest.
The kernel then has a problem the loader does not: its direct map covers the
low 4 GB, so a framebuffer above it needs ioremap(), which needs kpdpt, which
kvmalloc() has not allocated when the early console starts. A low framebuffer
is attached exactly as before; a high one is attached immediately after
kvmalloc(), still thousands of lines before anything else.
**There was more than one adapter, and the first one was taken.**
LocateProtocol answers with *a* GOP and the specification says nothing about
which. QEMU with a stdvga and a virtio-gpu has **three** handles and two of
them are Blt-only with a framebuffer address of zero -- so this is not a
hardware-only hazard, it was a one-in-three chance of drawing into address 0
that happened to keep coming up heads. Every handle is scored now: no
aperture, no candidacy; a readable current mode beats one that needs changing;
size breaks the tie.
**mode_usable() rejected modes real machines have.** A padded scanline (every
GPU that aligns its rows -- and very likely the reason 1920x1080 could not be
offered), a width not divisible by eight (1366x768, which is half the laptops
in the world), and anything below 800x600 (the only mode some firmware gives
over a KVM). All three passed on QEMU, which is why they survived. The mode
list now offers 1366x768 and 640x480 where it did not.
The loader also prints what it found, to the firmware's text console, which is
cleared before the splash: invisible on a boot that works, and the only thing
on the screen on a boot that dies before the splash.
One process note, recorded because it cost a wrong test: `make UEFI_KEEP=1`
ships the loader that is already built. Its own comment says "a build that
quietly ships a stale bootloader is how you spend an evening debugging code
that is not running", and that is exactly what happened -- a regression run
that looked clean had tested the old loader. Loader work gets a full build.
Both supported configurations verified after a clean build: one boot each, no
traps, no panics, socktest and memfdtest green, and the adapter report shows
the right one being chosen out of three.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gate 0's last item, "delete the cell grid" -- done as far as it should be, and
the rest of it retired, which is the part worth explaining.
include/text.h made the case against the grid in four charges: it cannot do
variable-width glyphs, cannot position text off a multiple of eight, cannot
anti-alias, and cannot hold a codepoint above 255. Every one of those was
about the *drawing layer*, and every one is now false -- glyphs carry metrics,
PSOGM positions text at arbitrary pixels, the TrueType path blends 8-bit
coverage, and vtcell.ch is a uint32.
What is left is a grid in vt.c, and deleting that would not modernise the
console; it would stop it being a terminal. VT100 escapes address columns,
TIOCGWINSZ reports rows and columns, and every program that draws a box
assumes both. So the item is checked with its remainder rewritten to say what
was actually wrong, rather than ticked as if the original wording had been
carried out.
**What was actually wrong was the duplicate.** gfx.c walked font_data and
blitted glyphs its own way; text.c did the same job into a struct surface, for
no caller but a sysrq self-test. Two implementations of bearings, coverage and
clipping, which had to agree about a cell size only one of them was told had
changed -- and they did not agree, which is why every line of PSOGM once drew
the next character's bitmap underneath itself.
There is one now, and the mechanism is the pleasing part: the console hands
text_draw_glyph() **the cell as a surface of its own** -- same pixels, same
stride, origin moved and bounds set to that cell. Clipping a tall ascender to
its line is then the same bounds check that clips to the edge of a screen, and
there is no second notion of clipping in the tree. `surface_sub()` is six
lines and replaced fourteen of hand-written comparisons.
The bitmap is reached through font_builtin() like any other provider, so the
early console and the real one are one path with different metrics, and the
`if (!cfont)` branch that walked the array is gone. gfx.c no longer includes
font.h; text.c does, being the only file left that reads it -- which also
retires the note in text.c explaining why it could not.
struct surface gained a bpp, because gfx_draw_pixel handles 16, 24 and 32 and
whatever replaces it has to as well. Zero means 32: a UEFI GOP has no other
format and the loader publishes 32 unconditionally, so callers that do not
care say nothing.
One behaviour change, and it is the documented one: a codepoint the font
cannot draw is now a substituted glyph rather than a blank cell. text.h argues
for that directly -- "a missing glyph that draws nothing makes text silently
shorter and the fault impossible to see".
Verified by screendump: the console renders correctly, PSOGM renders
correctly, and `echo t > /proc/sysrq` still draws its specimen through the
surface path with the back buffer at 32bpp. Both supported configurations: one
boot each, no traps, no panics, socktest and memfdtest green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gate 0's last-but-one open item. Drop a TrueType file at \schos\console.ttf on
the ESP and the console is drawn with it; with no file, the face linked into
the kernel image, which is what an early panic needs and is not a degraded
mode.
**The loader reads it, and that is the decision worth writing down.** The
kernel opening the file once the root filesystem is mounted is the obvious
design. It is also wrong: the console has been drawing since the kernel's
first line, so a font arriving later changes the cell size, which changes the
column count, which re-lays out every terminal -- and blanks them. The kernel
log would vanish off the screen at the moment the root filesystem mounted,
which is exactly the log a machine being brought up for the first time is
being read for. Since the next thing on the list is real hardware, losing that
was not a trade worth making.
So the choice happens before anything is drawn. Three details make it cheap:
- **EfiReservedMemoryType**, not EfiLoaderData. The E820 handoff turns
loader data into usable memory, on the argument that the loader's own
pages are the kernel's the moment it runs -- true of the loader, not true
of a buffer it wants the kernel to keep reading from.
- **Below 4 GB**, because the kernel reaches this through its direct map and
that is how far the direct map goes. A font above it would need an ioremap
the console cannot do before there is a console.
- **Used in place.** No copy, which matters because the kernel has no
contiguous allocator to copy 112 KB into -- the only one is the DMA arena,
and that belongs to devices.
Sized without EFI_FILE_INFO and its GUID: seek past the end and ask where that
landed, which the specification requires to clamp.
Every failure falls back and keeps going -- no file, implausible size, no
memory, short read, unparseable file. The last says so in the log, because a
font that silently did not load looks exactly like a font that did.
Verified by booting with JetBrainsMono-ExtraBold.ttf on the ESP and
screendumping the console: visibly the bold face, against the Regular of a
boot with no file. Three boots each way, no traps and no panics, socktest and
memfdtest green in all six.
Nothing ships a console.ttf: the file would be the file already in the kernel,
and two copies of a font are two things that can disagree.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two things, both asked for.
**The crash report no longer says which monitors it reached.** "The report is
on the virtio-gpu display too" was a line about the crash screen printed on the
crash screen. The report says what happened and where to read it again; that is
not part of it.
**And the second head now shows the F8 menu and the boot.** It used to be black
until vgpud started several seconds into userspace -- through the splash,
through the menu, through every boot message -- which from the outside is
indistinguishable from a display that does not work. Three stages now cover the
whole boot and hand over to each other without coordinating:
*The loader.* LocateProtocol() answers with the first GOP and offers no way to
ask for the rest, so the rest are found by handle and fed the primary's
framebuffer as a Blt source at the points where a screen is finished: the menu,
the resolution chooser, system information, the splash, each segment of the
kernel load. Blt rather than a second framebuffer to draw into, because a
virtio GOP does not have one -- the same fact that made a virtio-only machine
reboot-loop two commits ago.
*The kernel.* gfx remembers which rows were drawn, as one span rather than a
list of rectangles, because console output is lines and a span is what lines
make. display_mirror_console() copies that span across: a line of boot text is
one cell-row transfer, not a screen. Primed once after pci_init(), because a
display that has just appeared has a black scanout and the console it should be
showing was drawn by the loader into a framebuffer it knows nothing about.
*Then vgpud*, from the same console framebuffer, unchanged.
**Two bugs found building it, both worth the retelling.**
The mirror ran per character from inside vt_feed -- which is inside the console
lock -- and vga_disp_acquire() announces its back buffer with a cprintf. A CPU
taking a lock it already holds, with a call chain that said exactly that. It
runs at the end of consolewrite() and cprintf() now, outside the lock, which is
also the better batching: one transfer per write instead of one per character.
And "is this the console's own screen" was asked of the pointer acquire()
returned. vga_fb hands back a back buffer it copies onto the console's
framebuffer afterwards, so on a one-screen machine the answer was "a different
screen" -- and the console was mirrored onto itself through two megabytes of
intermediate, allocated for the purpose. display_ops grew scanout_phys, which
is a fact about the hardware rather than a policy a driver can get subtly
wrong, and it is asked first: it has no side effects and acquire() has two.
Verified with two heads under QEMU, screendumping the virtio head every 400 ms
from power-on: splash, splash, splash, then the console, and it stays. The F8
menu appears on it; choosing 1024x768 there brings both heads up at 1024x768
byte-identical. A panic still reaches both. Both supported configurations: one
boot each, no traps, no panics, socktest and memfdtest green.
One gap remains and is documented rather than hidden: while the kernel's driver
resets the device during probe, QEMU shows "Display output is not active" for a
fraction of a second. A device being taken over has no scanout until the new
driver sets one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"it also doesn't show crash screen on virtio gpu or just other gpu than default
vga" -- true, and on the virtio-gpu it was worse than nothing. The head kept
showing the console from a second before the panic: one screen saying the
kernel is dead, the other showing a shell prompt.
panic() painted into the console framebuffer and stopped there. Two cases were
missing, and they are not the same problem.
**The kernel still drives it.** display_acquire() hands back a buffer; the
report is copied in and presented. Ordinary, and a no-op when the display
class's buffer *is* the console's, so a one-screen machine does not do the work
twice.
**A process drives it.** The hard one, and the one hybrid.md owed an answer
for. The kernel let go of the device deliberately, and the process holding it
will never run again -- there is no scheduling after a panic -- so there is
nobody left to ask. It is not a race that can be won by being careful.
So the userspace driver says where it is scanning out *in advance*:
PCIIOC_PANICFB, registered by vgpud after bring-up and after every modeset. On
a panic the kernel paints the report into that memory and asks the kernel
driver the device was taken from to put it on screen one last time --
`panic_present` in struct pci_driver, which virtio-gpu implements by resetting
the device out from under a driver that is still nominally running. That is
safe here and nowhere else, because nothing is going to run again.
The panic path allocates nothing. It cannot: a corrupted heap is one of the
ways a machine gets here, and an allocator that deadlocks takes the crash
screen *and* the serial log's last chance to be read. So the driver keeps its
BAR mappings across the handover, rebuilds one queue from 20 KB of static
memory, and bounds every spin -- including the reset acknowledge, which is an
unbounded `while` in probe() and correct there.
**The registration is checked like the promise it is.** The range must lie
inside a block that process got from PCIIOC_DMA on that device. Without that,
a claimed device would be a way to nominate any physical address in the machine
-- page tables, kernel text -- to be overwritten with four megabytes of crash
report at the one moment nothing is left to notice. That is a worse bug than
the one this fixes. It goes away with the claim and with the process.
What it buys is worth stating plainly: a device can be evicted to userspace
without the kernel losing its ability to tell you the machine died. The
alternative was a driver boundary that is also a blind spot.
Verified both ways under QEMU with two heads, `echo c > /proc/sysrq`, and a
screendump of each over QMP:
vgpud holding the device both heads byte-identical, serial says
"The report is on the virtio-gpu display too."
kernel holding it same, via display_acquire/display_present
(/proc/pci shows 1af4:1050 [virtio-gpu])
Both supported configurations: one boot each, no traps, no panics, socktest and
memfdtest green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gate 1, item 3, and the last of foundations.md.
bind() creates a T_SOCK node at the path; connect() resolves the path and looks
for the socket bound to **that inode**. Matching on the inode rather than on
the path *string* is the part that matters: the name used to live only in the
socket table, so two spellings of one path missed each other, a relative path
never matched an absolute one, and there was nothing to stat() or unlink().
An inode is the thing a path resolves *to*, so connect() can walk the path
however it likes and arrive at the same socket.
# ls -l /tmp
srwxr-xr-x 1 root root 0 2026-09-22 08:04 socktest
The node is a marker, not the binding. It holds no data and the live
association stays in the socket table, so a socket file left by a crash is
what Linux leaves behind -- a dead file that connect() refuses until something
binds again -- and unlink()ing a live socket's path does not disturb anyone
already connected.
EADDRINUSE comes from the *directory* rather than from a scan of the table,
deliberately: a stale socket file has to refuse a rebind for the same reason a
running server's does, or two servers both believe they own it. The creation
lives in sys_bind() and not in unix.c, because that is the layer with both --
kernel/net/unix.c knows about sockets and has no business calling create().
T_SOCK is a compat addition to SchFS by the rule in schfs.h: a driver that does
not know SCHFS_T_SOCK maps it to a zero-length regular file, so nothing goes
stale and nothing becomes unreadable.
cmd/socktest section 15 asserts the plan's own wording -- bind, stat, see an
`s`, unlink -- plus the two a name table could never fail: a second bind to a
live path is refused, and connecting to a path that is a *regular file* is
refused rather than matched by name.
**And the semantics being real broke socktest on its second run**, correctly:
the socket files persisted and every bind() hit EADDRINUSE. Every real server
unlinks before binding; socktest now does too, which is the change the new
behaviour demands rather than a workaround for it.
Chasing that turned up something worse. `socktest` reported rc=1 with every
check passing, and so did every other first-party program:
- **lib/start.S called main and threw %eax away.** sys_exit() reads the
status from %rdi and nothing put it there, so every program exited with
whatever that register happened to hold. Every test suite ends in
`return failures ? 1 : 0` and none of it reached the shell.
- **exit(void) passed no status either**, for the same reason: the
declaration takes no argument, so %rdi was the caller's litter.
`if socktest; then` works now, and a failing command reports 1 -- checked both
ways, because a fix that makes everything return 0 would look identical from
the success side.
Also factored: unlink_path() out of sys_unlink, so a failed bind can take its
own node back out without going through the syscall entry, and create_sock()
as a named door rather than exporting create() with an arbitrary type.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gate 1, item 2, and the thing a Wayland buffer crosses on.
memfd_create(2) is the same segment machinery shmget() already had -- pages,
refcounts, no_free VMAs, the fork and exit accounting -- reached by
**descriptor** rather than by System V key. That difference is the entire
point and the reason it exists on Linux at all: two processes share memory by
passing the fd over a socket, so neither has to agree on a path or race to
invent a key. It is also why SCM_RIGHTS had to be item 1.
docs/plan/foundations.md proposed exactly this and it was right: what memfd
adds on top of shmget is a segment that starts empty and is sized by
ftruncate(), and a file that owns one reference to it. FD_MEMFD, and
read/write/lseek/fstat all learn about it -- a descriptor that refuses read()
is a surprise nobody needs.
Three things are refused rather than faked, each the honest version of
something this kernel cannot do:
- MAP_PRIVATE on a memfd. Every caller is sharing on purpose; a private
mapping of an anonymous object is an expensive MAP_ANONYMOUS.
- Mapping past the end. Linux faults those pages in; there is no handler
here that could.
- Shrinking while mapped. Linux truncates the pages out from under the
mapper; here a freed page would stay mapped to memory somebody else now
owns.
Writable MAP_SHARED of a *real file* is still ENODEV. That needs a page cache,
and an anonymous object has no such contract to break -- which is precisely
why memfd is the narrow answer.
**cmd/memfdtest**, seventeen checks, written so a permissive implementation
cannot pass: an mmap that quietly returned private zeroed pages satisfies every
return code and fails the one assertion that matters, because the reader is a
different process. It ends with the combination the plan singled out -- a memfd
sent over a socketpair to a receiver that closes every inherited copy first,
maps only the descriptor that arrived, reads what the sender wrote and writes
back where the sender can see it.
Two things were found by writing that test, both of which had been refusing
everything that was not FD_INODE:
- sys_fstat() never reached filestat(), so the memfd branch there was
unreachable from the syscall that needs it. It also refused pipes and
sockets, and returned -1 (which musl reads as EPERM) rather than an errno.
- sys_lseek() the same, and SEEK_END needs a length that is not in an inode.
Verified: 40 consecutive runs leave MemFree at exactly the value it started
at -- 936656 kB before and after -- so nothing leaks across creation, dual
mapping, descriptor passing and exit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Gate 0's last structural row. kernel/core/kmem.c: slabs are pages from
kalloc() with a header at the front, and free objects hold the free list in
their own first eight bytes.
That last detail is the whole design. **An object carries no header and
freeing performs no search**: the page base of any object *is* its slab, and
the slab names its cache. It is also exactly why an object cannot exceed a
page, which matters below.
Two interfaces. kmem_cache_create/alloc/free for a subsystem with one kind of
thing -- it gets its own line in /proc/slab, which is most of why naming it is
worth the argument -- and kmalloc/kmfree over power-of-two classes from 16 to
2048 for code with no single type to build a cache around. Each cache keeps
partial and full lists; a slab that empties goes straight back to kalloc()
rather than being kept, so /proc/slab never overstates what is held.
`make kmem-test` asserts the bookkeeping on the build machine. Same arrangement
as schfs-objtest, and for the same reason: slab lists, free-list threading, the
page-to-cache lookup and whether pages actually come back are all things that
can be wrong without a kernel to be wrong in. 26 checks, including one that a
permissive allocator cannot pass -- write a distinct pattern through 512
objects and read them all back, because two objects sharing storage shows up
there and nowhere else.
One bug the test found immediately: kmem_max_obj() returned the raw page
remainder, which kmem_cache_create() then rounded *up* past its own limit. The
largest legal size was reported as illegal.
**It does not do what the register said it would.** The claim was that every
fixed table exists to work around kalloc. Measured:
struct proc 10656 NPROC 128 larger than a page
struct socket 4928 NSOCK 32 larger than a page
struct buf 4272 NBUF 256 larger than a page
struct inode 280 NINODE 200 a cache, bounded on purpose
struct device 120 NDEVICE 32 3.8 KB, never hit
struct file 48 NFILE 512 24 KB, never hit
The three largest cannot go in a page-backed slab whatever it does. NINODE is
a cache and is meant to be bounded. NFILE and NDEVICE are small enough that
their tables are rounding errors and their ceilings have never been reached.
Freeing the big ones needs a contiguous multi-page allocator, which is a
different thing and is not written. That is now in the register and in
docs/architecture/kernel.md rather than left as a claim nobody checked.
What it does fix is real, and was measured rather than assumed:
- **A pipe is 640 bytes and came from a whole 4096-byte page.** Every `|` in
a shell command burned three and a half kilobytes it never touched. Six
pipes per page now, and /proc/slab shows the page going back when they
close.
- **procread() built its answer in a 2048-byte array on the kernel stack**,
under a comment that had warned about exactly that since it was written --
it stayed a local because a whole page for 2 KB felt worse, and kmalloc is
the answer that was missing. Its frame is 64 bytes now instead of ~2100.
- procfs's other generator took a page from kalloc for the same 2 KB.
Verified on a boot: /proc/slab tracks pipe and kmalloc-2048 through real
pipelines, allocs and frees balance, and $(), ps, time, readlink, sigtest and
rename are all still green with no traps.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Four Gate 0 rows, and a fifth found while checking one of them.
**getrusage() did not distinguish RUSAGE_CHILDREN from RUSAGE_SELF**, which
undersells it: it ignored `who` entirely and reported the machine's *uptime*
as the caller's user time. Every process of every age answered with the same
number. Per-process CPU time is sampled at the timer tick now and split on the
interrupted privilege level; children's time is rolled up in wait() rather than
exit(), because RUSAGE_CHILDREN means *reaped* children and a running one has
not contributed yet.
**And wait4()'s fourth argument is a `struct rusage *` that was read off the
stack and thrown away.** That is the one shells actually use -- wait3() is the
same syscall with pid -1 -- so zsh's `time` reported
16777216.01s user 1073.78s system 927983249% cpu
for every command. 2^24 seconds is uninitialised memory, and it did not vary
with the workload because it was never written. Now:
busy loop 2.66s user 97% cpu 2.720 total
sleep 2 0.02s user 0% cpu 2.248 total
ls -lR / 0.70s user 65% cpu 1.070 total
**ru_stime is a structural zero and says so.** MSR_SFMASK includes FL_IF, so
SYSCALL clears interrupts and no tick can land while a process is inside a
syscall. Tick sampling cannot see kernel time at all here. Charging it needs
TSC deltas closed and reopened around every sched(), which is scheduler
accounting rather than sampling; the limitation is written down in
docs/architecture/kernel.md rather than papered over.
**/proc/<pid>/exe is a symlink.** It also needed a real st_size, and every
other procfs file deliberately reports zero -- readlink(2) bounds its copy by
it, so a symlink of size 0 resolves to the empty string, which looks like it
worked and is worse than the regular file was. Generating that one path in
ilock() costs one small render per stat() of one file.
**OVMF memory reporting is closed**: 1 GB reports 1017 MB usable, 2 GB reports
2041 MB. The firmware keeps about 7 MB either way, which is the right shape.
**And `ps` printed a header and nothing else**, which is not in any list
because nobody had run it. `ps` with no arguments shows only processes on
*this* terminal; /proc/<pid>/stat hardcoded tty_nr to 0, so nothing could ever
match. `ps -e` always worked, which is exactly what hid it. p->ctty was already
tracked -- procfs simply was not asked. Fields 14..17 were hardcoded zero too,
so TIME read 00:00:00 however long a process had run; they carry the new
counters now, and a shell that has burned 653 ticks reports 00:00:06.
Verified: utime climbs 18 -> 653 across a loop in the shell itself, `ps` lists
its own terminal, `ps -e` still lists everything, `readlink /proc/1/exe` gives
/bin/schinit, and sigtest/$()/rename/dmesg are all still green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`#define ROOTDEV 1` -- IDE disk 1 -- so a kernel could only ever root on the
disk it was *built* expecting. Booting the same image from a USB stick was
impossible by construction, and that was the one thing standing between this
kernel and its own acceptance test, which begins "boot on real hardware from a
USB stick".
rootdev_mount() probes. schfs_kmount() was already written for exactly this --
its own comment says "the uberblock scan doubles as the probe: no valid
uberblock means this is simply not a SchFS volume, which is not an error worth
reporting loudly" -- and a failed probe leaves the slot unmounted and reusable,
which is what makes a loop over candidates legal at all. So the change is one
function and one call site; the only real use of ROOTDEV was `schfs_kmount()`.
Order is preference, not confidence: a RAMDISK_ROOT build wins outright, then
USB, then IDE 1, then IDE 0. USB before the fixed disks because a stick that is
plugged in was plugged in on purpose. iderw() panics on an absent drive rather
than failing, so each candidate is checked with ide_disk_present() /
usb_storage_present() first -- which is what those predicates are for.
Verified with **no IDE disk attached at all**: boots, mounts the stick, writes
a file, and the file is still there on the next boot.
[ 0.000021] root: USB disk (device 3)
And when nothing matches, it names what it looked at, because "no root
filesystem" with no list is the least useful thing a kernel can say at the one
moment somebody needs to know which disk it expected:
root: no SchFS volume on any candidate device. Looked at:
device 3, USB disk: not present
device 1, IDE disk 1: not present
device 0, IDE disk 0: present, but no SchFS on it
**That message is also how dmesg turned out to be broken**, because it did not
appear in `dmesg` and should have. `dmesg` printed *nothing at all*, on a
system whose boot banner says "or dmesg once you are in" and whose entire
quiet-boot design is "route the detail into the kernel log and say how to read
it". The log was there the whole time -- `cat /proc/kmsg` has always worked --
but the one command anybody would reach for could not see it.
Two separate faults, each alone enough:
- **sys_syslog() was `return 0`.** dmesg does not read /proc/kmsg; it calls
klogctl(10) for the buffer size and klogctl(3) to read. 0 to the first one
means "the log is zero bytes", and a stub that answers 0 is
indistinguishable from a machine that logged nothing. Implemented for the
actions that mean something here; the rest return -EINVAL rather than 0,
because a console log level this kernel does not have is not a request that
succeeded.
- **With that fixed, dmesg was still empty.** It parses every line with
`sscanf(msg, "<%u>[%llu.%llu] %n", ...)` and silently drops anything that
does not match. The ring's `[ 0.00] text` already matches everything
after the priority -- scanf skips the leading spaces -- so the only missing
thing was the `<6>`. klog_render_syslog() adds it on the way out rather
than storing it: the ring is also what a person reads with `cat
/proc/kmsg` and what PSOGM points at after a crash, and an ABI detail
belongs at the boundary that has the ABI.
Also: the first version of the no-root diagnostic printed
%-11s device 1596400
because cprintf() has no field-width specifier -- it prints "%-11s" literally
and then reads the next argument for the %d. Plain %s now, with the note.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Went looking for why `X=$(echo hi)` hangs every shell. Did not find it. Found
two other things, and built the instrument that rules out the largest suspect.
**The signal return trampoline was ABI-blind.** deliver_signals() manufactures
a return address for the handler -- seven bytes on the user stack reading
`mov eax, 15 ; syscall`. Fifteen is *Linux's* rt_sigreturn, which is right for
anything musl-linked, and that is every port and every shell. For a first-party
binary it is native syscall 15, which is `open`. So a handler in any cmd/
program ran, returned into the trampoline, called open() with whatever the
handler had left in its registers, got -ENOENT, and carried on executing the
stack bytes after the trampoline.
pid 14 sigtest: trap 14 err 4 rip 0x46ca7 addr 0xfffffffffffffffe
rax=fffffffffffffffe ... rsp=46ca0
ustack: 46e30 46ca0 50f0000000fb8 0 0 0
rip is rsp+7 -- one past a seven-byte trampoline -- and the third word of the
stack is the trampoline's own bytes. It emits the number for the process's ABI
now; syscall_map is the thing that says which numbering this process speaks.
**SIG_BLOCK, SIG_UNBLOCK and SIG_SETMASK were not defined anywhere.**
sys_sigprocmask branched on bare 0, 1 and 2 with the names in comments, so
anything outside the kernel had to spell the number and hope it had the right
one. Defined in posix.h, used in the kernel.
**cmd/sigtest** asserts the SIGCHLD/sigsuspend contract every shell is built
on, and it is written so a permissive stub cannot pass: a blocked SIGCHLD stays
pending, sigsuspend delivers an *already pending* signal without spinning, the
handler runs exactly once, the mask is restored, the child is still reapable,
and a signal arriving mid-wait also wakes it. Eight checks, all green. The
kernel's half of that contract is now proven rather than assumed, which is what
lets the remaining bug be pursued somewhere else.
Along with it, `rt_sigsuspend` in lib/usys.S -- the syscall existed and the
first-party library simply never exposed it.
On the hang itself: it is not job control, not SMP, not zsh (mksh does it too),
and not any value the kernel returns -- a full syscall trace is in
docs/history/kernel.md section 7, which also records what the contradiction
narrows to. That section had said "root cause not yet isolated" with three
suspects; all three are now eliminated, and the note says what to do next
instead of what to guess.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
One to four processes died per boot with a fault they could not have caused:
`rip 0x0` -- an instruction fetch at zero, which a process does not branch to
-- or a dereference of `0xffffffffffffffea`, which is `-EINVAL`. It hit shells
and schinitd alike and had been carried for a long time as "zsh sometimes jumps
into .data".
`syscall_entry` has to swap FS_BASE from the user's TLS to the kernel's before
it can touch any `%fs:` data, and `wrmsr` destroys rax, rcx and rdx. Those have
to be parked somewhere first -- not on the user stack, which is untrusted and
has not been switched away from yet. They were parked in three globals:
// Single-CPU scratch for syscall_entry FS swap
// (safe while IF cleared by SYSCALL/SFMASK)
uint64 _sc_rax, _sc_rcx, _sc_rdx;
IF being clear answers for an interrupt arriving on *this* CPU. It says nothing
whatever about a second CPU executing the same three stores at the same
instant, and with more than one processor entering syscalls that is not a rare
interleaving -- it is the ordinary one.
**%rcx is the user return address.** SYSCALL puts the user's rip there and
SYSRET takes it back from there. So the CPU that lost the race returned its
process into *another process's* instruction stream, with %rax -- the syscall
number on the way in, the return value on the way out -- swapped alongside it.
Both symptoms fall straight out of that: a zero handed over and followed, and
an errno used as a pointer.
Per-CPU words at %gs:32..48 now. %gs is usable there, before the FS swap, for
exactly the reason RESTORE_KERNEL_FS can already read %gs:0: GS base is per-CPU
in hardware, seginit() sets it once, and userspace cannot change it --
sys_arch_prctl() implements ARCH_SET_FS and ARCH_GET_FS and rejects the rest.
Measured rather than argued. Both variants built once, `build/esp` saved for
each, the same pristine schfs.img restored before every boot, -smp 4:
three globals 10 boots 10 faults, in five of the ten
three per-CPU 18 boots 0 faults
A _Static_assert pins all seven per-CPU offsets, because trapasm.S spells them
as literals and cannot include a header -- a slot that moves would leave a
register quietly holding somebody else's value, which is the hardest possible
thing to find.
**This is the second bug of this exact shape in this file.** The first was
`kernel_fs_base`, a single global every CPU wrote during its own seginit(), so
the last writer won and every CPU restored one CPU's per-CPU block;
RESTORE_KERNEL_FS reads %gs:0 for that reason. The lesson generalises and is
now written down in both places: **on the entry path, a global is a bug.**
Closes the row added to Gate 0 four commits ago. It was the right row to put
first -- a compositor built on this would have failed in ways indistinguishable
from its own bugs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Painting every kernel stack at allocation and scanning it on each trap puts the
high-water mark at **3736 of 4096 bytes** -- 91%, in netd, inside lwIP's input
path. 360 bytes of headroom on a kernel that takes interrupts on the current
stack is not headroom: a timer landing at that depth pushes a trapframe and
runs trap() straight off the bottom of the page and into whatever kalloc handed
out next, which is very often another process's kernel stack.
The canary at kstack[0] did not catch this and could not have. It is checked on
trap entry and exit, and a frame allocated in one `sub $N, %rsp` steps clean
over the bottom eight bytes and writes below them without ever touching it. It
catches a slow walk past the end, not one deep call.
8192 now, and the stacks are static -- one per ptable slot, NPROC * KSTACKSIZE
= 1 MB of BSS. Static rather than a bigger kalloc because kalloc() hands out
exactly one page and there is no contiguous multi-page allocator to ask. A
process's slot is fixed for its lifetime, so the stack belonging to that slot is
the natural answer: no allocation to fail, no free to forget, and the kalloc
failure path in allocproc() goes away with it. The page was being paid for out
of kalloc before, one at a time, and undercounted.
**And the AP scheduler stack was already wrong.** startothers() did
`stack = kalloc()` for one page and then handed the AP `stack + KSTACKSIZE` as
its stack top. Exactly right for as long as KSTACKSIZE happened to equal
PGSIZE, and a whole page past the allocation the moment it did not -- which
this commit would have made true. That is not a throwaway boot stack either:
the AP runs mpmain() and then scheduler() on it for the rest of the machine's
life, and takes interrupts on it whenever no process is current. Static too,
NCPU of them.
**This does not fix the intermittent zsh crash**, and the measurement above was
taken while looking for it. It is worth saying plainly: 1-3 shells per boot
still die, so the hypothesis was wrong and the stack was a separate latent bug
that the search turned up.
What the search did establish is worth more than the guess was:
- **It is SMP-only.** Three boots at -smp 1: zero faults. Three at -smp 2:
one, three, one.
- **It is not only zsh.** schinitd took one too.
- **The faults name their cause if you print enough.** `rip 0x0` is a process
handed a zero and following it, not a process that branched anywhere. And
schinitd's `addr 0xffffffffffffffea` is -22, which is -EINVAL: a call
returned an error and the caller used it as a pointer.
So trap.c dumps registers, segments, and the top of the user stack for a fault
at rip below a page or a faulting address near zero, the way it already did for
a GP fault. Neither of the two facts above was visible from the one-line
report, and both came out of the first dump that included them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`∞`, `→`, `≤` and `'` could not be typed. `ą` and `ę` could. That pairing is
the whole diagnosis: KBD_UP was 0xE2 and its neighbours filled 0xE0..0xE9,
which are exactly the UTF-8 lead bytes for three-byte characters. Every
character in U+0800..U+FFFF -- most punctuation, every arrow, every maths
symbol, the whole of CJK -- arrived as a value the console read as an arrow key
and helpfully turned into `ESC [ A`. Two-byte sequences lead with 0xC4 and
nothing claimed 0xC4, so Polish came through and the rest did not.
The chord codes already knew this rule and said so in a comment. The navigation
keys predate them and were never brought in line. Stated once, in kbd.h: a key
produces either a Unicode codepoint or a value above 0x10FFFF, and nothing in
between means anything. KEY_* -> KBD_*, which also settles a silent collision
-- include/input.h defines its own KEY_HOME, KEY_UP and KEY_END as evdev
numbers, and kbd.c includes both headers, so three macros were being quietly
redefined in the one file that used them.
Then the layouts, which is what the codepoint-sized keycodes were for.
kernel/dev/keymap.c: us, pl, de. A layout is four tables -- plain, Shift,
AltGr, Shift+AltGr -- of uint32 codepoints. uint32 and not uchar is the reason
the file exists: `ó` fits in a byte and `ą` does not, and half a Polish
alphabet is not a Polish layout.
Two things deliberately are *not* tables:
- **Ctrl is derived from the plain codepoint.** The table it replaces was
indexed by scancode, which quietly assumed US -- on QWERTZ the Y and Z keys
swap, so Ctrl+Z would have suspended a job only on some layouts.
- **Caps Lock is a case toggle on the result**, so AltGr+a with Caps Lock is
`Ą`. No second table could say that, because Caps Lock combines with AltGr
rather than replacing it. Latin Extended-A alternates upper/lower by
codepoint but not all the same way round -- Ł is U+0141 (odd) and Ą is
U+0104 (even) -- so Polish lands in both halves and the parity matters.
kbdgetc() encodes to UTF-8 and returns the bytes one call at a time.
consoleintr()'s contract is bytes: the serial line hands it UTF-8 already and
it could not tell an encoded codepoint from a byte of the same value.
The layouts live in the kernel. Linux's loadkeys compiles a text keymap and
uploads it a key at a time through KDSKBENT, which puts the data in userspace
-- and a machine whose /usr has not mounted then has no layout at all. Here the
kernel always has a working keyboard and `loadkeys` only names which one.
KB_GETLAYOUT/KB_SETLAYOUT/KB_LISTLAYOUT; /etc/vconsole.conf and an
os.schos.keymap job for the boot-time choice. No dead keys, so French and
Spanish are absent rather than wrong.
/etc/locale.conf and LANG=C.UTF-8 in profile and zshenv. Only the charset half
of that is real -- musl decodes UTF-8 whatever the locale says and ships no
locale data -- but an unset LANG reads as "ASCII only" to everything that asks,
which is its own way of producing `?`.
And two symlink bugs found while chasing
zsh: locking failed for /root/.zsh_history: operation not permitted
on every boot. symlink() is the classic atomic lock -- create the link, and if
it exists someone else holds it -- and zsh distinguishes EEXIST ("contended,
retry") from anything else ("this cannot work, give up"). sys_symlink returned
a bare -1, which is -EPERM, so every zsh on the machine took the second branch
on a filesystem that supports symlinks perfectly well. Same errno class as
commit 75ccf98.
The second one the first was hiding: **divesymlink() has been in fs.c since the
fork and nothing ever called it.** A symlink behaved as a small file containing
its own target, so `cat link` printed the target's *name*. open() and openat()
follow the final component now (O_NOFOLLOW gives ELOOP), and stat() and lstat()
stop being the same function, since which one follows is the only difference
between them. Mid-path symlinks are still not resolved -- that belongs in
namex(), which would re-enter lookup holding a directory lock.
Verified on the emulated keyboard, not just over serial: AltGr+a/e/s gives
ąęś, Shift+AltGr gives ĄĘŚ, Caps Lock+AltGr gives ĄĘŚ, de types zy for the
physical y/z keys and @€µ on AltGr. `∞ 'x' → ≤ • — …` typed over serial now
survives. Three fresh boots: 0 lock failures, 0 faults.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"Ą Ę Ś Ł Ż" drew as "?????" and the obvious diagnosis -- the font has no
glyphs for them -- was wrong twice over. `struct vtcell.ch` was **one byte**
holding a CP437 index, and the console mapped UTF-8 down to it on the way in.
U+0104 does not fit in a byte. No font could have fixed that; the cell had to
get wider first.
So: `ch` is a uint32 codepoint, the mapping on the way in is gone, and the cell
draws through kernel/dev/ttf.c against JetBrains Mono linked into the image.
Cell size comes from the font's measured advance instead of a hardcoded 8x16.
The grid costs 864 KB of BSS instead of 432 KB, which is the honest price of
holding a codepoint per cell for six terminals at 240x75.
The rasteriser needed three fixes before it could be trusted with this.
**21 KB of locals on a 4 KB kernel stack.** `ttf_render` held an outline (512
points as three arrays) and a raster (1024 edges) as automatics. KSTACKSIZE is
4096. They walked off the bottom of the stack into whatever was underneath,
which surfaced as user processes faulting on addresses they never touched and
then a panic inside freevm() on an overwritten page directory. Nothing pointed
anywhere near a font.
**And then again, recursively.** load_outline() held a 4736-byte outline as a
local *and recursed* for composite glyphs. Composite glyphs are exactly Ą, Ę,
Ś, Ł, Ż -- so the second crash arrived the moment the first fix let Unicode
reach the screen, looking like a fresh bug. Now a static outline per recursion
level with a depth cap of 4 (the font's deepest real nesting is 2), and the
leaf parse split into load_simple() so its 512-byte flag array is not
multiplied by the depth. Worst chain measured from the object file: ~2.3 KB.
Neither showed in tools/ttftest, which has an 8 MB stack. That is the standing
hazard of one source file built for host and kernel: the host is most forgiving
about exactly the resource the kernel is strictest about.
**Component transforms were being skipped**, on the stated theory that a wrong
transform is worse than none. It is not. `‘` in this font is the comma glyph
with a scale of -1 -- a 180-degree rotation -- and the offset that ships with
it only lands correctly after the rotation. Without it the mark went twelve
pixels into a ten-pixel cell and was clipped away entirely, so the character
read as missing from the font rather than as misplaced. F2Dot14 2x2 now, with
SCALED_COMPONENT_OFFSET honoured.
The glyph cache was ASCII-only, so every accented letter re-rasterised on every
repaint: a screenful of Polish did hundreds of bezier flattenings per scroll
where the same screenful of English did none. 95 reserved ASCII slots plus 224
shared slots keyed by codepoint, 64 KB pool. Verified 13680 cached renders
byte-identical to direct ones across collisions and a pool flush.
tools/ttftest speaks UTF-8 now. Iterating bytes meant it could only ever ask
for ASCII, and ASCII is entirely simple glyphs -- it could not reach the code
path that mattered.
Also: resolvconf's success line moves behind /proc/bootmode, same rule as
vgpud. A quiet boot opened with "resolvconf: nameserver 10.0.2.3" where the
one line was supposed to be "Welcome to SchOS!". Failures stay loud.
Verified on screen, not just on the host: Ą Ę Ś Ł Ż ć ń ó ź, Ω λ π Œ Ð Æ ß,
∞ ‘x’ → ≤ all render, panic=0 across four boots.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
vgpud: wanted 7365 KB of DMA memory for a 1920x982 framebuffer
vgpud: not enough contiguous DMA memory
schinitd: restarting os.schos.vgpud (and again, and again, until
schinit: supervisor stopped answering schinit replaced schinitd)
Every number in that diagnostic was about the arena, and the arena had 61 MB
free. The refusal came from do_dma()'s per-request cap, which was 4 MB --
exactly one framebuffer at 1280x800, chosen when that was the only thing
anybody asked for.
virtio-gpu's display size is the size of the **host's window**. Drag it out and
the guest is asked for 1920x982x4 in one contiguous piece, and no amount of
free arena helps. 32 MB now, which is 2560x1600 with room over. The cap is not
there to ration memory -- the arena does that and says so properly -- it is
there so a wild size cannot ask for everything by accident.
DMA_ARENA_BYTES 16 -> 64 MB alongside, for the reason its own comment already
gave and then got the arithmetic wrong on: either side of the handover *both*
drivers hold a framebuffer, so the requirement is two at the largest mode, not
one. 16 MB is two at 1280x800 and nothing bigger. Verified booting at 256M and
512M as well as 1G.
And vgpud no longer dies when it cannot have what it asked for. It claims the
framebuffer before creating the resource, and if the full size cannot be had it
steps down by thirds to 640x480 before giving up -- and gives up with
SCHINIT_EXIT_NORESTART rather than die(), because a driver that will fail
identically on every restart should not be restarted. That loop is what made
schinit replace the supervisor mid-boot.
vgpud: the display is 1920x1080 but only 1280x720 of DMA could be had
is what it says when it steps down. With the cap raised it does not need to:
1920x1080 now allocates its full 8100 KB, 0 restarts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VQ7Qx8E1q1M1bbCq8QLHS2
Ctrl-C killed the shell. consoleread() breaks off the read in progress when
SIGINT arrives and returned -1 to do it -- and -1 is -EPERM by the time it
reaches a program, so every program that handles SIGINT by carrying on saw a
permanent error instead of an interrupted call:
zsh: error on TTY read: operation not permitted
schinitd: restarting os.schos.console-shell
and the next Ctrl-C did it to the replacement. EINTR now, which is what a
signal-interrupted read is required to give. Verified: Ctrl-C an interruptible
command, shell survives, zero restarts.
Same mistake, same file family: link, unlink, rename and renameat returned -1
on every error path. Callers branch on errno -- `if (unlink(p) < 0 && errno !=
ENOENT)` is the standard shape, and against EPERM it always complains -- which
is why zsh reported failures for things that had worked. They return ENOENT,
EEXIST, EXDEV, EISDIR, ENOTEMPTY and EFAULT now. About ninety other `return -1`s
remain in sys_fs.c; these four are done because these four were what a program
tripped over. /dev/mem's unknown minors return ENODEV rather than -1 too.
/dev/urandom was a linear congruential generator seeded with a **constant**, so
every boot produced the same bytes in the same order -- not weak, the same. A
key, a nonce, a temporary filename: shared with every SchOS machine that had
ever run. Seeded at boot from RDRAND where the CPU has it, from the cycle
counter and the wall clock where it does not, and re-stirred per byte. Still an
LCG and still not a CSPRNG; what changed is the part that was dangerous rather
than merely poor.
**Not fixed, and recorded as such:** rename() still will not replace an existing
target, so zsh's history save fails on every exit after the first. An attempt to
unlink the target first was reverted -- the directory is an HTree, dirunlink
followed by dirlink of the same name did not behave, and it panicked in
filewrite with the old file already gone. Losing both files is worse than
refusing, so it refuses until the HTree's slot reuse is understood.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VQ7Qx8E1q1M1bbCq8QLHS2
Nexoniarz