Deaicup/Deshab-OS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
301006d611a30db01f88bc7c7e9c59f8186432fe
Deshab-OS/.learnings/ERRORS.md
T

13 KiB
Raw Blame History

Errors

Command failures and integration errors.

[ERR-20260906-001] linux_compat exec returns -4 after guest park

Logged: 2026-09-06T23:10:00+08:00 Command: DSK lxc->exec / exec_async for KDE preflight and deshab-kde-session Error: rc=0xfffffffffffffffc (-4 超时无 EXEC_EXIT) Context: Linux guest already EXEC_READY + parked; virtio-gpu and extra-rootfs backends bound; chooser sent key 2. linux_resume returned via HLT but linux_to_utsm ring had no EXEC_EXIT. Guest serial interleaved ([utsm-linux] ini[VMEXIT]), switch_root/vdc lines not recovered. WSL host only 7.6G so QEMU is 6G with a 3.9G extra-rootfs module.

Next: confirm daemon recv of EXEC_REQUEST after first park; check initramfs vs switch_root; raise WSL memory before blaming KDE packages.

Root cause (2026-09-07): UTSM_IOCTL_RECV_MSG / SEND_MSG in utsm_hcall.c called copy_to_user(umsg.data) / copy_from_user(..., umsg.data) after umsg was already a kernel stack copy. That is -EFAULT. Daemon recv_msg fails, parks with no EXEC_EXIT → host -4. Empty EXEC_READY send worked because data_len==0 skipped the bad copy. Fix: memcpy within the kernel struct, then a single copy_to_user of the whole ioctl arg.

Resolution

  • Resolved: 2026-09-07T06:41:00+08:00
  • Notes: Rebuilt bzImage + patched ESP. QEMU verify: KDE preflight rc=0, exec_async rc=0, KDE spawned; desktop guest ping rc=0.

[ERR-20260907-001] virtio-blk vdb/vdc sector 0 I/O error

Logged: 2026-09-07T22:50:00+08:00 Priority: high Status: in_progress Area: backend

Summary

Linux 6.6 能 probe virtio-blk（blk_size 修好后），但 vdb/vdc 读 sector 0 曾立即 I/O error。2026-09-08 已能 EXT4-fs (vdb): mounted，串口无 I/O error。

Error

virtio_blk virtio2: [vdb] 1572864 512-byte logical blocks
I/O error, dev vdb, sector 0 op 0x0:(READ)
EXT4-fs (vdb): Could not load journal inode
I/O error, dev vdc, sector 0

Context

e820 已改为实际映射大小后仍失败。host 已 log extra-rootfs bound。virtio_gpa_to_host 对请求缓冲返回 NULL、或 chain 解析漏间接描述符，最可疑。

Suggested Fix

对 rootfs_blk_handle_chain 在 IOERR 路径打 GPA/len 日志；确认 desc.addr 落在 0x05000000 RAM 窗口。

Progress (2026-09-08)

  • e820 0x00400000-0x03FFFFFF 标成 60MB RAM，但 EPT 只映射了 kernel/initrd 等几页。virtio-blk READ 缓冲落在空洞且 vCPU 从未 store → virtio_gpa_to_host 返回 NULL → IOERR。
  • 修复：virtio_gpa_to_host 按需映射；启动时填满该窗口空洞（~34MB）；size_max 改为 65535 字节；描述符链支持 INDIRECT + 多段；rootfs 内存可写。
  • 验证：vdb mount OK，无 I/O error。Arch /sbin/init(systemd) 与 /bin/sh 在 glibc 同一偏移 SIGSEGV（segfault at ffffffff，确定性，非随机坏块）。PID 1 改为 utsm_exec_daemon 后 EXEC_READY 恢复。KDE 用户态仍起不来：host exec 的动态链接 sh/startplasma 同样崩；vdc overlay 本轮未见挂上。

Progress (2026-09-08 later)

  • vdc 已挂上：init 必须在 mount --move /dev 之前 mount vdc。现串口有 persist moved to new root /mnt/persist + overlay /opt /home /root。
  • WRMSR IA32_FS_BASE 必须写入 VMCS_GUEST_FS_BASE（忽略写会留下 TLS=0）。
  • musl busybox（无 SHSTK note）在 push 上 segfault at 0；改回 Arch /bin/sh→bash 后变成 glibc #GP，IP 正好是 endbr64（libc RX 映射 +0x175140）。
  • 在 host 上 wrmsr IA32_U_CET=0 或临时打开 CR4.CET 会让 UTSM 在 Linux park 前就死掉，不要再试。
  • 当前 blocker：动态链接 glibc 的 IBT landing pad（endbr64）在 nested VMX 下 #GP。Plasma 仍起不来。

Progress (2026-09-10)

  • 盘上 libc 0x199140 已是 90 90 90 90；guest 仍在同一 RIP #GP。
  • 用户态 #GP 会 VM-exit 到 L1；guest_fetch 读到全 0，ENDBR skip 从未命中。
  • ibt=off 内核认；user_shstk=off 在 Linux 6.6 未知。
  • nested KVM 不提供 VMCS CET save/load。
  • /bin/sh→busybox.static 后：segfault at 0 error 6（PUSH/SSP=0）。
  • 不要用 0F 1F 40 00 替换 ENDBR。

Progress (2026-09-10 afternoon)

  • 限制 RIP 的盲 skip：glibc 0x7f…140 连续 skip 8×4 字节后 segfault at 7f…d000；busybox.static push r13 仍写地址 0（RSP 正常，是 SHSTK/SSP=0）。
  • L1 CR4=0x20a0（无 CET）；L1 CPUID 也无 CET_SS/IBT。不能在 L1 WRMSR CET。
  • VM-entry MSR-load IA32_U_CET=0：VMLAUNCH 指令成功，随后 exit reason 34（MSR loading failure），Linux 未 park。已停用该 list。
  • QEMU 进程 arch_prctl(ARCH_SHSTK_DISABLE)：WSL 返回 EINVAL，/proc/cpuinfo 无 shstk/ibt。LD_PRELOAD 无效。
  • Guest rdmsrl_safe/wrmsrl_safe(IA32_U_CET)：rd=wr=-EIO（#GP）。L2 也碰不到 CET MSR，但硬件 IBT/SHSTK 仍作用于用户态。
  • CET 在 L0 生效且对 L1/L2 都不可编程。在禁止 host wrmsr 的前提下 Plasma 起不来。

Progress (2026-09-10 evening)

  • 去掉 #PF 拦截后 park + desktop ready 恢复。
  • KVM_SET_MSRS CET 返回 r=0（0 个 MSR 写成功），KVM 不虚拟化 CET。
  • patch_boot_on_img.sh 把 /bin/sh 改成 busybox.static 会让所有脚本在 0x4bede0 PUSH 写地址 0 上死掉。已改回 /bin/sh -> bash。
  • bash/glibc 用户态 #GP：RIP 0x7f…140，ilen=2 或 3（不是 ENDBR 的 4），页表走到 hole 零页（mem=0）。盲 skip 4 会到 …144 继续 #GP。
  • CR3_LOAD_EXITING 能拦住 MOV CR3（见 0x514c000 等高位 PGD），但故障时 tracked CR3=0x2fa2000 仍把 libc GVA 走到零页。

Progress (2026-09-10 EPT misconfig)

  • persist 能挂上 /dev/vdc；desktop ready 后 EPT misconfig gpa=0x3e20000 把 vCPU 停掉。
  • guest clear_page_erms #GP，RDI=0x0005088000005000（非规范），随后 ext4lazyinit 在已 tainted 内核里 oops。
  • 根因更像 EPT leaf 未 Ignore PAT + misconfig 杀 vCPU，而不是这一次的 EXT4-fs error。

Metadata

  • Reproducible: yes (glibc SIGSEGV)
  • Related Files: CODE/UTSM/vmm/virtio_blk.c, CODE/UTSM/vmm/virtio_mmio.c, CODE/UTSM/vmm/linux_loader.c, CODE/linux/initramfs/init

[ERR-20260910-001] verify kde_ok=2 after persist check stall

Logged: 2026-09-10T20:22:00+08:00 Command: patch_boot_on_img.sh + tmp_reverify2.sh (landing-pad rebuild) Error: exit 1; guest_ok=1 session_ok=1 kde_ok=2 sent_kde=1 desktop_ready=1; last session kmsg checking persist mount; no persist check done / entering KDE Context: WSL NAT warning on stdout; QEMU 480s; serial 2 spawned deshab-kde-session. Host KDE loop still exec sleep/test -f every frame. drm KDE scanout w=1024 is fbdev, not Plasma.

Next: run_slice pump + persist remount kmsg; rebuild UTSM/desktop/DSK; inject session via patch_boot_on_img.sh (do not loop-mount extra-rootfs from drvfs).

Follow-up (2026-09-11)

run_slice 打进镜像后仍 kde_ok=2：checking persist mount 后无 vdc present/persist check done。同步 test -f 仍 EXEC_EXIT -4。async 时 daemon sched_yield 空转占满 vCPU；init mount --move 失败后未 umount，二次 mount 因 vdc busy 失败。

[ERR-20260911-001] verify6 froze at Limine, no SUMMARY

Logged: 2026-09-11T21:35:00+08:00 Command: tmp_run_verify6.sh (initrd+patch+qemu_verify_loop) Error: serial log 20KB Limine cursor text only; no UTSM/DSK markers. Wrapper exec >file hid later output. QEMU gone on resume. Next: re-run verify only with stdbuf -oL + timeout --kill-after=15 540.

[ERR-20260911-002] Cursor-attached wsl.exe kills the VM in ~45s

Logged: 2026-09-11T22:02:00+08:00 Command: long-running wsl -d Ubuntu from Cursor Shell (even sleep 620) Error: wsl.exe exits 1 at ~45s; dmesg systemd-shutdow + journal "uncleanly shut down"; QEMU dies while Limine is still Loading module linux-extra-rootfs.img. Serial stays ~20KB cursor text. STATUS_STACK_BUFFER_OVERRUN / UTF-16 NAT warning on the Windows pipe. Next: run tmp_run_verify9.sh from Windows Terminal, not Cursor. Image already patched (initrd+daemon+session+desktop).

Follow-up (2026-09-11 22:07)

User ran it in Windows Terminal: no tty output (script >>file), QEMU gone at ~1 min, serial still 20KB Limine extra-rootfs, no timed_exit. Same kill during 3.9G module load from drvfs, not only Cursor. Next: tmp_run_visible.sh (tee + heartbeat + copy image to /home/deshab).

[ERR-20260911-003] persist remount failed, no startplasma

Logged: 2026-09-11T23:04:00+08:00 Command: tmp_run_visible.sh (native /home/deshab/deshab-dev.img) Error: kde_ok=2. Session: persist remount failed, ERROR: no startplasma at /mnt/persist. Init: overlays on persist then mount --move fail; umount+remount also fail because overlay upperdirs keep vdc busy. Next: bind-mount persist into the new root; session bind from existing /dev/vdc mount if remount EBUSY.

[ERR-20260912-001] verify dies at log=21013 in Windows Terminal

Logged: 2026-09-12T12:40:00+08:00 Priority: critical Status: in_progress Area: infra

Summary

tmp_run_visible.sh 在 WT 里 t=26s、log=21013 整段退出，没有 SUMMARY。Limine 仍在 Loading linux-extra-rootfs.img。

Error

[verify] t=26s log=21013 guest=0 session=0 desktop=0 kde=0
PS C:\Windows\system32>

Context

Native /home/deshab/deshab-dev.img 已完整。第一次成功启动曾在 21013 停约 80s。之后多次 26–45s 被掐。不是用户提前 Ctrl+C。

Suggested Fix

停止 Limine 预加载 extra-rootfs；vdc 改为 ESP FAT32+AHCI 映射。

Follow-up (2026-09-12 13:21)

Limine 已过：DSK 20s、guest+desktop。init 报 /dev/vdc missing：linux_launch 在 dsm_load（AHCI）之前，盘上映射没有块设备。已改为先加载 manifest stage 0–1。

Metadata

  • Reproducible: yes
  • Related Files: SYSTEM/limine/limine.conf, CODE/UTSM/vmm/virtio_blk.c
  • See Also: ERR-20260911-002

[ERR-20260912-002] entering KDE then bind /dev failed

Logged: 2026-09-12T14:10:18+08:00 Priority: high Status: in_progress Area: backend

Summary

Guest parked, persist has startplasma, session reached entering KDE root, then mount --rbind /dev failed. Verify printed VERIFY_KDE_PLASMA_OK because entering KDE was treated as Plasma.

Error

[deshab-kde] persist already has startplasma
[deshab-kde] entering KDE root at /mnt/persist
[deshab-kde] ERROR: could not bind /dev into the KDE root

Context

e2fsck of Arch vdb was clean. extra-rootfs /dev /proc /sys /run are 0555 empty placeholders. Session stderr is /dev/null (async daemon). drm SET_SCANOUT at 7.2s is virtio-gpu probe, not Plasma.

Suggested Fix

chmod 0755 placeholders, mount --bind (not rbind), log mount stderr to kmsg, seed /dev/dri on failure. Verify kde_ok=1 only on starting Plasma.

Follow-up (2026-09-12 14:16)

Real bind error: persist /usr/bin/mount + vdb ld.so → libmount.so.1: Error 74 (EBADMSG). vdb ext4_lookup checksum invalid on inodes 1673/1810/1826 (/usr/bin/mount is 1826). Use persist ld.so --library-path to run persist mount/chroot.

Follow-up (2026-09-12 14:45)

Static utsm_bindmount binds /dev /proc /sys /run /tmp. persist mount/ld.so are unusable (SHSTK or XSAVE). e2fsck of vdb is clean; checksums are runtime virtio reads (grep=1673 mkdir=1810 mount=1826). chroot persist bash exits 127 (new glibc wants XSAVE/YMM; CPUID hides them). Overlay vdb ld.so+libc → chroot 139 (SIGSEGV). Plasma still not painted.

[ERR-20260912-003] idle=poll still on guest cmdline after header edit

Logged: 2026-09-12T19:50:00+08:00 Priority: critical Status: in_progress Area: backend

Summary

PARK/HLT 分流已进树，但 guest 仍带着 nohlt idle=poll，Xorg 等待期间 cpu_idle_poll 空指针，idle task panic。

Error

Command line: ... no_timer_check nohlt idle=poll noapic nolapic ...
RIP: cpu_idle_poll.isra.0+0x2b
RDX: 0000000000000000
Kernel panic - not syncing: Attempted to kill the idle task!

Context

linux_loader.o 未随头文件重编（缺 .d）。SYSTEM/boot/utsm.elf 19:34 链接仍含旧字符串。验证镜像是 /home/deshab/deshab-dev.img。

Suggested Fix

删除 vmm/linux_loader.o 后 make；确认二进制无 nohlt/idle=poll；同时 patch ISO 与 native GPT。

Metadata

  • Reproducible: yes
  • Related Files: CODE/UTSM/vmm/linux_loader.c, CODE/UTSM/include/utsm/linux_loader.h
  • See Also: LRN-20260912-003

Follow-up (2026-09-12 20:33)

删掉 linux_loader.o 后 cmdline 已干净，idle panic 消失。vdb /tmp//run bind 导致只读，改为 tmpfs 后 seatd socket 出现。persist /usr/lib/Xorg ELF 头损坏；换官方包后 Xorg 不再 139，改为 libXdmcp.so.6: invalid ELF header。debugfs 确认 libXdmcp.so.6.0.0 在 extra-rootfs 上就不是 ELF。

Reference in New Issue View Git Blame Copy Permalink