nebulon

nebulon запушил(а) master в nebulon/nebos

11 часов назад

nebulon запушил(а) master в nebulon/nebos

  • 99417e8e77 settings: real windowed app with WM preferences Replace the old control-center popover with a titled Settings window: a section sidebar and a scrollable page for Network, Power, Date & Time, Appearance (desktop colour and wallpaper picker), Windows and About. The top-bar netmenu is now a compact quick-settings popover whose six pills open the matching section through SYS_WM_OPEN_SETTINGS. A running Settings window is reused, and the WM hands it the requested pane via /share/settings.req. Add SYS_WM_PREFS so the Settings app can apply and persist window shadows, edge snapping, drag resistance and the desktop colour. Preferences live in /data/SETTINGS.CFG (or /share/settings.cfg) and are replayed by the WM after boot_ready; the WM and sys_wm objects build -Os to stay under 0x90000.
  • 1123caf1a4 wm: soft drop shadow for normal windows
  • b70c54bedc browser: auto table layout for overflowing columns, <br> empty lines Hacker News comment pages laid out 5000 px wide: - A one-column fallback table row only ever scaled its cell up to the table width; a cell whose one-line text is wider is now scaled down. - When the columns' one-line widths do not fit, both table paths (layout_table and the per-row table_columns) give each column its min-content width plus a share of the rest in proportion to max - min, like CSS auto table layout, so long text cells wrap. intrinsic_width() gained a min-content mode (iw_min). - A <br> on an empty line is one empty line; HN's comment header (margin-bottom: -10px followed by <br>) no longer overlaps the text, and <br><br> after a block leaves its blank lines. The compat ladder, news.ycombinator.com, docs.cloudron.io, google.com, getbootstrap 3.4, lobste.rs and heise.de render pixel-identical; cloudron.io gains the two blank lines its hero <br><br> asks for.
  • 06acbc67d3 browser: cache fallback-table column widths per layout pass Rows of a table laid out through the per-row fallback (a table nested in another table's cell) each recomputed the whole table's column widths, measuring the text of every cell of every row: O(rows^2) text measurement. A Hacker News comment page (one row per comment) took ~1.75 s per measure and per render pass on the host, many times that on the ION, and froze the browser. table_columns() now keeps the result for the current pass in an 8-slot cache keyed by table and width (one slot per nesting level, since each comment row holds its own nested table); inline_box_pass_id changes every pass, so DOM or style changes are never served stale widths. The item page's passes drop to ~0.25-0.3 s.
  • 1e391e1572 wm: pad the title and use a hover-bright close icon
  • Просмотр сравнение для этих 10 коммитов »

11 часов назад

nebulon запушил(а) master в nebulon/nebos

  • 6b7600b448 init: PID-1 service manager and parent-death process tree Add a ring-3 init as PID 1. The kernel starts /ramfs/bin/init after wm_server_init(), records init_pid before the process is made runnable, protects it from SYS_KILL and restarts it from the kworker. Init is the only WM creator: the kernel launcher/fallback and wm_expected_pid are gone, and wm_server_attach accepts only an init child named "wm". Every PCB now carries parent_pid; process_exit(), process_kill() and the reaper kill the descendant subtree, and a child born while its parent dies is aborted. SYS_SPAWN gains a flags word: the default parents the child to the caller, SYS_SPAWN_DETACH attaches it to the caller's parent (the terminal/WM session). Terminal commands, WM apps and kernel commands stay parented; files/disks/chat and the WiFi helpers use detached spawns. Browser and gallery launchers spawn the real app parented and wait, so terminal launches stay foreground and Ctrl-C or closing the terminal kills the app. Init owns one private tty per service, drains it to /share/svc-<name>.log, writes /share/svc-<name>.state and serves the NBI1 control protocol on its own tty. The service CLI/GUI starts, stops, restarts, kills and tails websrv (the first unit) and the WM; logs and state are RAM-only, and an init crash resets the ring-3 tree and restarts the desired services. New syscalls are SYS_TTY_ALLOC, SYS_TTY_WRITE, SYS_TTY_CLOSE and SYS_IS_INIT; TTY_MAX is 16. SYS_PROCINFO exposes parent_pid, so the process monitor now draws an indented parent/child tree while still sorting sibling groups. It protects init, and the WM start menu gains a Services entry (/ramfs/bin/service --gui). The spawn/tty ABI is deliberately not backward compatible: kernel, command pack, root.img, UEFI payload and host harness must be rebuilt together, and the new kernel needs init in the rootfs for a WM. Tested in QEMU on BIOS and UEFI: init PID 1 starts the WM and terminal; service list/status/start/stop/restart/kill/logs and service --gui; WM restart; an init crash reset; gallery foreground/Ctrl-C parent-death; a file-manager launched viewer surviving the file manager; the process tree; and the start-menu Services entry.
  • 15a0c3e30c sched: account the BSP tick with the period it covered sched_tick_bsp() charged the elapsed tick with timer_tick_ms() read after wake_sleepers_locked(). When the idle process is current, waking a sleeper calls smp_resched_wake(), which reprograms the PIT to 1 kHz and sets tick_ms to 1 for the next interval, so every 10 ms idle tick was charged as 1 ms of idle. On a 2-CPU machine the BSP therefore looked ~90% busy at rest: the process monitor's CPU card showed ~58% while its process rows summed to ~8%. Read the period at entry, before any wake-up can shorten it, and use it for power_tick() and tick_decide_locked(). The AP path captures c->tick_ms the same way. Tested in QEMU: before, BSP idle_ticks advanced ~10% of the wall clock and most tick time was charged to neither idle nor a process; after, idle_ticks + process ticks equal the elapsed ticks, and the process app with 1 CPU shows 14.4% busy against 13.8% summed over the rows.
  • 2fa060e0ee acpi: reset via the FADT RESET_REG on UEFI reboot The reboot command only pulsed the 8042 controller, which does nothing on the Chromebook and left the BSP halted with the BKL held. Cache the ACPI 2.0 RESET_REG/RESET_VALUE pair (RESET_REG_SUP) and write it first; fall back to the Intel PCH 0xCF9 sequence, the 8042 pulse and a triple fault so a reboot cannot end in a frozen hlt loop. terminal.o, process.o and reboot.o build with -Oz to keep _end below the 0x90000 wall after the S5/reset code crossed the 0x69000 data/.bss boundary. QEMU BIOS and UEFI both boot twice through the reboot command.
  • ccb4fdc5ba rootfs: raise the compressed pack cap to 4 MB and reloads to 8 MB PACK_MAX (vesa_stage.S), CMDS_GZ_MAX (Makefile) and the embedded-pack check in the UEFI loader now allow a 4 MB gzip pack; the legacy copy ends at 0xD00000, inside the 16 MB no-E820 fallback. KEXEC_MAX, PAYLOAD_MAX and netkexec_send take up to 8 MB, the ramfs cap, so the 4.35 MB raw pack of cmds-dev/net-cmds is accepted again. Tested in QEMU: a 4.15 MB pack boots on the legacy and UEFI paths and over net-dev, the raw pack goes through net-cmds, and make test-install passes.
  • 0c0473ea82 browser: reliable tooltips, opacity, transitions and modal stacking - Forced reflows are capped at 8 per main-loop pass instead of 40 ms per 100 ms: on slow boards every reflow was refused and Bootstrap placed its tooltip from stale boxes. - opacity is a real alpha; translucent absolute/fixed boxes are blended over what lies beneath them (group opacity). - transition / transition-* animate opacity and translate() (percentages of the own box) from the value on screen, with easing, and dispatch transitionend; the start state comes from the last paint, so a forced reflow before adding a class works as in browsers. - A fixed box pinned by top and bottom spans the viewport (modal backdrops). - Painted absolute/fixed boxes with a background or z-index occlude the content beneath them for clicks and hover; pointer-events: none is parsed. - Animations of elements no longer painted are dropped. Bootstrap's modal on cloudron.io now fades its backdrop in, slides the dialog down and closes on a backdrop click.
  • Просмотр сравнение для этих 5 коммитов »

18 часов назад

nebulon запушил(а) master в nebulon/nebos

  • 75e0dc9ead acpi: real S5 poweroff on BIOS and UEFI The kernel only poked the Bochs/QEMU shutdown ports, so poweroff froze the BSP with the BKL held on real hardware. Parse the FADT PM1 blocks and the DSDT \_S5 package, enable ACPI through SMI_CMD when SCI_EN is clear, and write SLP_TYP first and SLP_EN in a second write while preserving SCI_EN. Intel-style southbridges get SMI_EN.SLP_SMI_EN cleared first, since that bit turns the SLP_EN write into an SMI. NVIDIA nForce/MCP boards (the ION) run S5 through the firmware SMM: replicate the DSDT's _PTS(5) by writing ASSB/AOTB/AAXB/OSTP in the BIOS NVS and arming PS1S/PS1E in the SMI block before SLP_EN. A failed poweroff now returns instead of halting; the WM shows a notice and the netlog carries the FADT/S5/SCI_EN state. A new acpidump command writes the mapped DSDT to /share/DSDT.BIN for board diagnostics. acpi.o and poweroff.o build with -Oz to stay under the 0x90000 kernel wall.
  • 56aba2f2e6 browser: en-US Intl, form constraint validation, stripped prelude QuickJS has no Intl, so the Cloudron dashboard's Domains view threw "'Intl' is not defined" in prettyLongDate. The prelude now provides an en-US-only subset: DateTimeFormat (UTC), NumberFormat, RelativeTimeFormat, DurationFormat, DisplayNames, Collator, PluralRules, ListFormat and a stub Locale, and routes toLocale*String and localeCompare through it (checked against node's ICU Intl on 525 cases). Behind it the domain dialog called form.checkValidity(); controls and forms now have checkValidity, reportValidity, validity and setCustomValidity. To pay for it in the rootfs pack, the new host tool tools/jsstrip.c drops the prelude's comments, indentation and non-separating blanks before js_prelude.S embeds it: prelude gzip 18,467 -> 20,252 bytes in total.
  • 29b7a1c185 wm: fix wallpaper blits in partial damage rects
  • dac20d37c3 wm: persisted desktop wallpaper from the image viewer
  • f268094347 usb: arm EHCI bulk-IN qTDs with the exact transfer length A bulk-IN qTD was always armed with the 2048-byte bounce buffer, even when usbmsc asked for one 512-byte sector. A device that ends the data stage without a zero-length packet (the ION's SanDisk) left the qTD ACTIVE forever, so install could not read its USB source. Arm each qTD with exactly the length of the call that starts it and do not pre-arm the next transfer: a BOT data stage's final chunk can be shorter than the previous one. This also stops a completed larger chunk from being clipped to a later, smaller length.
  • Просмотр сравнение для этих 6 коммитов »

1 день назад

nebulon запушил(а) master в nebulon/nebos

  • 8373a22461 fs: skip raw reads of a failed USB slot The boot background probe can leave a mass-storage device that never answers its media read. The Disks app then issued SYS_BLK_READ on it again and waited out another full timeout before the window appeared. Mark the MSC slot failed when a read command finally fails, and return EIO immediately from SYS_BLK_READ for such a BLK_USB slot. A storage rescan clears the mark and retries the device with the full recovery path.
  • 2ee880a4fc fs: keep raw block reads preemptible SYS_BLK_READ (the Disks app's partition view) ran in the non-preemptible filesystem section and busy-waited on a stuck USB disk while holding the BKL, so the whole desktop stopped until the device timeout expired. - Publish the raw read as a scan-owned section without preempt_disable(): disk file operations still wait, but the USB HCD waits can sleep/yield and drop the BKL. - Mark BLK_USB raw reads as a best-effort probe (one 5 s attempt) instead of the full reset/retry chain.
  • 7f6dbd12ae wm: remap content canvases after the frame begins The WM mapped window contents before u_wm_begin(), so a client self-resize in that gap could unmap the front canvas under the draw. usurf_canvas() then read released pages and the WM died with a #PF (cr2=0x59420000, content slot 1). Re-read each window's content generation after u_wm_begin(), when the kernel refuses further self-resizes, and remap changed canvases in all draw paths: full and partial redraw, geometry changes and drag.
  • 17fcbc512e usb: fix boot-time mass-storage stalls The ION's SanDisk NAKs its first READ(10) data stage after enumeration, and a timed-out EHCI bulk-IN qTD stayed armed, so the boot scan sat in the reset/retry chain for 9-21 s. - EHCI: ring the async-advance doorbell after an idle gap and before cancelling a stuck bulk transfer; cancel a timed-out IN qTD and reset its data toggle; xHCI stops the endpoint before dropping its ring. - usbmsc: clear both host-side toggles after the BOT reset's CLEAR_FEATURE requests, re-apply SET_CONFIGURATION once, request sense only after a CSW status error, and give the first media read a longer deadline (5 s best-effort boot probe, 15 s full scan). - VFS: defer USB mounts until the bootscan window has closed; recognize the NebOS MBR layout without reading P1; do not probe a valid non-FAT MBR as a superfloppy; skip a partition whose boot sector is unreadable. - The boot background USB pass can no longer delay the desktop, /ramfs, /share, the bootscan window or the WiFi autoconnect.
  • 35d464536c browser: hover events and Bootstrap tooltips - Dispatch mouseover/mouseout (bubbling, with relatedTarget) and mouseenter/mouseleave on the ancestors entered/left when the element under the pointer changes; jQuery's mouseenter and Bootstrap tooltips build on them. - Geometry reads after a DOM change run a synchronous layout first (forced reflow, budgeted to 40 ms per 100 ms); documentElement clientWidth/clientHeight report the viewport. - Absolute/fixed boxes with opacity:0 or visibility:hidden keep a measurable box (not painted, not hit-tested); an auto width with one horizontal offset is shrink-to-fit; absolute/fixed children no longer widen a shrink-to-fit parent or end a float row. - <a href=""> is a link (hand cursor, click target). - browser-host: {move:X,Y} key moves the headless pointer for hover tests.
  • Просмотр сравнение для этих 27 коммитов »

1 день назад

nebulon запушил(а) master в nebulon/nebos

  • 828cba05ea ngui: share one corner radius and fix its rounding Cards, tiles, list rows and dialogs each picked their own radius (6 to 10 px). The corner scanline inset also used a half-resolution distance, so a corner was taller than it was wide. Add NGUI_RADIUS for the shared radius and compute the inset from pixel centres, then use the constant in disks and the control panel.
  • f3def28d01 mm: raise the physical ceiling to 1 GB, move the frame bitmap out of BSS The frame bitmap was a 16 KB BSS array sized for 512 MB, and the kernel image must end below 0x90000. mm_init() now sizes the bitmap to the highest E820 RAM address and places it in the first RAM run at or above 0x900000 that avoids the boot rootfs pack and the LFB (a 512-byte static array remains for the 16 MB no-E820 fallback). _end drops from ~0x8ef20 to ~0x8bf20. The allocators keep handing out the lowest free frames, but no longer scan from page 0 with interrupts off: a first_free word hint skips the full low part, mm_alloc_page() finds the bit with bsf, and mm_alloc_pages() steps over full and empty words 32 pages at a time. MM_MAX_PHYS is now 1 GB, exactly USER_BASE; paging.c statically asserts the identity map never reaches a ring-3 window. The UEFI loader may place the pack anywhere below 1 GB, and QEMU guests default to 1G. Verified in QEMU with a temporary randomized self-test against a naive lowest-fit scan (128M/512M/1G/2G: 0 mismatches, no leak), a post-paging write/read of every free frame up to 0x3FFFF000, and test-uefi/test-amd. Known issue 6 (user pages leaked on reap) was stale: the reaper already frees them. It now records the real limit, RAM above 1 GB.
  • ed2d7f479a Follow the primary plane's gamma in the Intel cursor The Ironlake display palette sits after the plane blender. We forced CURSOR_GAMMA_ENABLE on the cursor plane even when the VBE primary plane bypasses the palette, so the cursor's premultiplied antialiased edge pixels went through the LUT alone and appeared as a white halo on any non-black background (the X201). Read the matching DSPxCNTR gamma-enable bit and mirror it, and drop the pipe-select field: Gen5 has one cursor register per pipe, so bits 29:28 are reserved (Linux skips them for Gen5). Verified on the X201: the halo is gone and the pointer looks correct over both light and dark windows.
  • da99a8231b ngui: hide scrollbars until the pointer enters a pane The bar was always painted over the right edge of a scroll pane, so files, chat, disks and gallery reserved a column for it and the content never used the full width. Draw an 8 px rounded overlay pill instead: it stays hidden while the pointer is outside the pane, slides in from the right edge over ~100 ms on hover, brightens over the hit strip and uses the accent colour while dragged. Presses in the strip start the drag (or jump the thumb) and are consumed so the row or button beneath it does not activate. Remove the per-app reservations so rows, messages and columns use the full pane width, and let the control panel use the overlay bar too; hide_bar remains as an explicit opt-out. The terminal keeps its own cell-rendered scrollbar.
  • a2c9570412 browser: fix unfocusable form fields and the external-link icon Docusaurus' navbar search field showed the I-beam on the board but ignored clicks. doc_reset() cleared every node field except the cached attribute bits, so a node slot kept BR_ATTR_DISABLED (and the open/inert bits) from the previous page, and real hardware does not hand out zeroed memory either. The bits are now cleared on reset, node_append() zeroes the whole node, and removing an attribute clears its cached bit (`input.disabled = false` used to leave the field disabled). Related form fixes found on the way: - A field takes focus before `click` is dispatched; a cancelled click only skips the activation behaviour (as focus follows mousedown). - Full form-state slots are recycled from controls that left the document (client-side routers replace content without navigating). - The status bar no longer labels a hovered form control "Link". The external-link icon after "Cloudron.io"/"Demo" drew only a corner: a <use> without width/height now sizes its symbol to 100% of the viewport instead of the symbol's viewBox units. browser-host --keys gains {click:X,Y} for headless clicks.
  • Просмотр сравнение для этих 5 коммитов »

2 дней назад

nebulon запушил(а) master в nebulon/nebos

  • 4996d78353 browser: draw ::before/::after icon boxes and data: images Docusaurus' sidebar chevrons are empty pseudo-elements with a size and an inline-SVG background (`content:""; background:url("data:image/svg+xml; utf8,<svg ...>") 50%/2rem 2rem; transform:rotate(90deg)`), which the text-only generated content dropped. - `content: ""` is kept (BR_CONTENT_EMPTY); br_pseudo() reports such a box when it has a size and a background. It becomes one RUN_PBOX run: background colour, then the data: image at its background-size and position, clipped to the box and rotated by transform. Flex rows reserve their first/last column for these boxes (margin-left:auto puts one at the right edge), intrinsic_width() counts them, and an empty <button> whose icon is such a box is laid out as an element. - data: URLs (percent-encoded or base64) are decoded for background images, <img> sources and icon boxes (data_url_image() caches by URL and size, failures included). - Declaration values end at the first `;` outside strings and parens, url() honours quotes and unescapes CSS string escapes, background URLs and custom-property values may be up to 8 KB, and the background shorthand reads its keywords/colours with url() bodies blanked (an SVG's own rgba() fill no longer becomes the background colour). - SVG images without a viewBox scale their user units to the up-to-4x raster: a 40x20 icon was drawn at a quarter of its size. The compat ladder renders pixel-identical to the previous commit.
  • 312f63a340 disks: keep scroll pane content clear of the scrollbar ngui paints the scrollbar over the right edge of a scroll pane, so the device list and the details page drew rows and buttons underneath it. Narrow the content while a pane scrolls, and redraw once when the details page starts or stops needing the bar (its height is measured while drawing). Co-Authored-By: Claude <noreply@anthropic.com>
  • 06a7aa59be wm: add a Disks launcher to the top bar The new last launcher spawns /ramfs/bin/disks; WM_BTN_END_X moves to its right edge, so the bar notice and the right-aligned status start 70 px further right. Co-Authored-By: Claude <noreply@anthropic.com>
  • d097c55b1a browser: show the hand cursor over URL-bar suggestions Co-Authored-By: Claude <noreply@anthropic.com>
  • c6eac494e7 ngui: drop the checkmark on selected list rows The selection background already marks the row; the extra check also shifted the selected row's text out of line with the others. Affects every ngui_list_row user (disks, netmenu). Co-Authored-By: Claude <noreply@anthropic.com>
  • Просмотр сравнение для этих 7 коммитов »

2 дней назад

nebulon запушил(а) master в nebulon/nebos

  • c5f21a2ab9 Use the Adwaita cursor set instead of the single arrow NebOS had one 16x16 two-colour arrow. It now has the 35 CSS cursor shapes from GNOME's Adwaita theme (adwaita-cursors 51, LGPL-3 / CC-BY-SA-3.0): the arrow, hand, I-beam, the resize arrows, move, help, not-allowed, zoom, the animated wait/progress spinners and the rest. This system ships no SVG cursor theme, only Xcursor bitmaps, so the assets are Adwaita's Xcursor files cut down to their 24 px images (assets/cursors/adwaita/, 424 KB instead of 12 MB, by tools/xcursor_trim.py). The host tool mkcursors turns them into one NBCUR1 bundle (include/cursor_abi.h: size, hotspot, frames, premultiplied ARGB), packed as cursors/adwaita24.cur in the full pack only (~18 KB gzip; the floppy has no room and keeps the built-in arrow). The spinners keep every fourth of their 60 x 16 ms frames. kernel/cursor.c copies the bundle out of the ramfs at boot and after a runtime pack replacement. mouse.c treats the position as the hotspot, clips the image at the screen edges and blends it over the saved background (up to 32x32); Intel Ironlake and Radeon load the ARGB image into their 64x64 planes, Cirrus reduces it to its two colours. SYS_WIN_SET_CURSOR (117) sets a window's content cursor and SYS_WM_SET_CURSOR (118) is the WM's override. The window server resolves the shape on every WM poll: the override, else the window holding the pointer capture, else the window under the pointer, else the arrow. The WM uses resize arrows on the grip, the move cross while dragging a title bar and a hand over launchers and the status area. ngui asks for the I-beam over text fields, the terminal uses it for its cells, the browser honours CSS `cursor` (a hand over links, the I-beam over text fields and the progress cursor while loading for `auto`), and the gallery shows every shape. browser-host maps them to SDL system cursors. make test-cursors checks the bundle. Verified in QEMU with the software and Cirrus cursors; the Intel/Radeon paths are untested.
  • ce76a5ac91 Add browser visit history with URL-bar autocomplete Every page that finishes loading is recorded (URL, title, visit count and a stamp counter that orders visits across windows and reboots) in a 128-entry BSS table. Browser windows share /share/bhist, merged on every write, and it is persisted to /data/BHIST.DAT at most every 3 s. While the address bar has the focus a dropdown lists up to six entries: the top sites for an empty bar, otherwise URL prefixes (with or without the scheme and www.) ranked above URL and title substrings, plus visits and recency. Typing at the end completes inline to the next '/' of the best match, and Enter opens the history URL it stands for. Up/Down preview a row, Delete forgets it, Esc closes the list, a click opens it. The browser no longer fits the floppy lite pack: it is now in the full pack only, and floppy boots run BROWSER from root.img through the existing launcher fallback. browser-host gains --keys for scripted headless typing.
  • d00a58bec0 Use the Adwaita symbolic icons for ngui ngui's icons (browser back/forward/reload/home, the files and gallery icons) are now GNOME Adwaita symbolic icons instead of line drawings. The unchanged *-symbolic.svg files live in assets/icons/adwaita/ (LGPL-3 / CC-BY-SA-3.0). ICON_LIST in include/icon_abi.h names them and generates enum ngui_icon and ngui_icon_name(); the order is the bundle order, so append only. STOP, CLOSE, SEARCH and MENU are new. The host tool mkicons rasterizes each SVG with the browser's svg_raster.c into a 16x16 coverage mask and writes one NBICO1 bundle, /ramfs/icons/adwaita16.icn (3.3 KB, ~0.8 KB gzip). Its XML reader only accepts what symbolic icons use and fails the build on transforms, strokes or unknown elements; -v prints an ASCII preview. ngui_icon() reads the bundle once per process into BSS (safe across the browser's heap_reset()), capped at the ids it knows, and blends the mask tinted with the caller's colour. The procedural fallback is gone: it cost ~2 KB of code per app, and dropping it lets both packs carry the bundle while the floppy lite pack still shrinks. browser-host symlinks icons/ into its /ramfs like fonts/.
  • 9eb38faa14 Fix xHCI EP0 max packet size so the VL805 sees the boot stick On the 970A-DS3P's VIA VL805 (1106:3483) a high-speed USB stick never enumerated: Address Device kept the USB core's 8-byte first-read guess for EP0, and the 18-byte device descriptor read failed with Babble Detected, so the booted stick was never usb0 and install had no source. - Address Device uses the speed's fixed EP0 size (8 low, 64 high, 512 SuperSpeed); a full-speed EP0 is corrected with Evaluate Context once the core knows bMaxPacketSize0. - Do not PED-disable a SuperSpeed port before its reset (that leaves the link in SS.Disabled). - Set the Setup TRB transfer type only on xHCI >= 1.0. - Always log the controller id/version and every failed control transfer (completion code, PORTSC, slot/EP0 state) to COM1/netmon. Verified on the board: the SanDisk stick enumerates first try and mounts as usb0; QEMU xHCI USB3 boot + HID unchanged.
  • 7254b400be Render google.com: inert noscript, 4 KB URLs and author-styled controls google.com showed only its "Klicke hier ... nicht automatisch weitergeleitet" fallback. Its <noscript> holds a meta refresh and `table,div,span,p{display:none}`, and the parser treated noscript content as live DOM/CSS although scripting is always on. <noscript> is now a raw-text, display:none element (HTML's scripting-enabled parsing). The page's form controls then rendered in the dark UI theme: - Google's stylesheet URL (~720 B) and main script URL (~3.7 KB) were truncated by 512 B browser URLs, 480 B href/src attributes, the curl helper's 768 B request file and uhttp's 1 KB paths. All are 4 KB now (BR_URL_MAX, BR_URL_ATTR_MAX, UHTTP_PATH_MAX; header lines 4.5 KB). - ~94 KB of inline <style> overflowed the 64 KB BR_CSS_SIZE, dropping the form-control rules; it is 256 KB. - Inputs, textareas and buttons with author CSS (background, border or radius) are painted with that background/border/radius and text colour and sized from padding, height, min-width and margins; unstyled controls keep the ngui theme. `background: transparent|none` is now a cascaded value (PROP_BG_NONE shares PROP_BG's slot). Layout fixes found on the same page: - float shrink-to-fit width includes borders ("Anmelden" wrapped); - a max-width clamp makes the width definite, so margin:auto centres; - a flex row under 160 px stays a row when its items fit; - align-self overrides the row's align-items; - forced-colors/prefers-contrast/hover/pointer media features are answered for a normal-contrast mouse screen (Google's forced-colors border no longer applies); - U+2010..2015 map to '-'. Large leaf buffers in uhttp and the browser are static to spare the 128 KB ring-3 stack. New compat pages 072 (styled controls) and 073 (noscript) fail on the old build and pass now; pages 001-071 render pixel-identical in browser-host.
  • Просмотр сравнение для этих 5 коммитов »

3 дней назад

nebulon запушил(а) master в nebulon/nebos

  • 346fd927bf Split the tree into boot/, kernel/ and user/ directories Move the flat repository root into: boot/ boot sector, MBR, VESA stage, boot/efi/ UEFI loader kernel/ the kernel, drivers, services and kernel.ld common/ inflate and the 8x16 font, linked into kernel and userland include/ ABI headers shared by boot, kernel, userland and tools user/apps/<n>/ one directory per ring-3 program (ucmd_<n>.c + modules) user/lib/<l>/ libc, ngui, ufont, img, tls, uhttp, ujson, uwm, wifi tools/ host tools (mkpack, mkrootfs, kexec_send, flash.sh, ...) The top-level Makefile still builds everything with the same targets. Objects go to build/ with compiler-generated header dependencies; the images and workflow artifacts (kernel.elf, kernel.bin, floppy.img, usb.img, usb-efi.img, BOOTX64.EFI, cmds.pack, browser.elf, ...) stay at the top. Per-object flags are target-specific variables, and the kernel link order is kept explicit in KERNEL_LINK_ORDER. A clean build is byte-identical to the previous tree for kernel.bin, boot.bin, mbr.bin, the floppy/USB images, both packs, the rootfs images, browser.elf, gallery.elf and BOOTX64.EFI. AGENTS.md gains a source-tree section and the docs' source maps use the new paths.

4 дней назад

nebulon запушил(а) master в nebulon/nebos

  • ee904a6a3c Decode images at display size and fetch chat media asynchronously Remote <img> jobs now run the JPEG/scale work in the curl __serve helper: the request carries the display bounds, the JPEG decoder can emit a 1/2, 1/4 or 1/8 IDCT-scaled bitmap, and the UI only copies the published RGB565+alpha bytes. The browser keeps a 6 M-pixel LRU of display-size bitmaps so repaints are 1:1 blits, coalesces image-driven repaints, bounds decode time per main-loop pass and records the net/io/dec/scale/paint breakdown in /share/IMGSTAT. img_blit adds the 1:1 opaque copy and integer-alpha paths; the image tests cover the reduced IDCT output. Chat message images and avatars use the same helper (480x320 and 32x32) instead of blocking on chat_http(); uhttp_read consumes a buffered body incrementally with a 16 KB reader, and ngui scroll panes can hide their scrollbar for the control panel.
  • 8d63339700 Bring up the Gigabyte 970A-DS3P and fix SB950 OHCI input AMD/board: parse both MADT IOAPICs and mask their normal redirection entries after ACPI init (ioapic.c), sample family 15h Tctl through D18F3xA4 with the kworker under the BKL, and add a separate AMD WC MTRR path that uses only a free variable entry with a zero high mask and never writes MTRRdefType (the 970A reset during paging init otherwise). Cap the VBE/GOP mode at 1920x1080 so the full-screen RGB565 canvas still fits the 4 MB window slot, and make vga_fb_lfb_wc() report the MTRR-backed LFB path as well as the ION alias. USB: the SB950's legacy OHCI DMA does not snoop dirty CPU cache lines, so the EDs, TDs, setup packet and interrupt buffers move into a dedicated page that paging_init maps uncacheable before paging is enabled. Without it the controller sometimes read an empty ED, skipped the transfer and never re-armed the endpoint, freezing the cursor with no error. Control data buffers stay write-back and are flushed around the transfer. Enumeration now starts on the USB default 8-byte EP0 and adopts the descriptor's real bMaxPacketSize0, settles 20 ms after a port reset, and retries a failed port three times with a fresh reset. The boot scan gets a second OHCI pass 60 ms after the EHCI-first walk, because the companion handback can leave a low-speed port briefly disconnected. HID detach is debounced (three disconnected reads, 50 TD errors) and forces a clean full rescan, since a false port-loss leaves no CSC to wake the per-port path. DoneHead/WDH is acknowledged on every poll as well. Docs: AGENTS.md and the architecture pages cover the 970A target, IOAPIC masking, AMD thermal/MTRR, the 1080p limits and the SB950 USB path.
  • e762867916 Add a Linux SDL3 host harness for the browser The browser only needs the small SYS_* ABI, so host/ now implements it for Linux: the same ucmd_lib.h wrappers dispatch to host files, BSD sockets behind uhttp/TLS, fork/exec plus a pipe tty for the curl __serve helpers, and an SDL3 window (native Wayland here). The ring-3 allocator is reused over an mmap arena, keeping heap_reset() behaviour. make browser-host builds browser-host and host-curl without touching the target images. --headless --wait N --shot out.ppm writes the RGB565 canvas as a PPM for the compatibility ladder. Host sockets replace the NebOS TCP stack, so the WM and kernel drivers are not exercised.
  • 4148a04a2b Load and lay out arstechnica.com: TLS 1.2, jQuery/Alpine DOM, Tailwind CSS TLS: the same ClientHello now also offers TLS 1.2 (ECDHE over X25519 or P-256 with ChaCha20-Poly1305 or AES-128-GCM, extended master secret), for servers without 1.3. New tls_gcm.c (AES-128-GCM on wpa_crypto's AES) and tls_p256.c (P-256 ECDH), tls12_prf() in tls_sha256.c, all with host vectors in make test-crypto. The 1.3 middlebox CCS moves before the client Finished, since a 1.2 server rejects it after the hello. DOM: createHTMLDocument returned the live document, so jQuery's support probe wrote two <form>s over the page body; it and a new DOMParser now use an inert detached document. parentNode of children of node 0 (the <html> root) is no longer null. Add document.currentScript, document.scripts/ images/forms/links/embeds and inert <template> elements with a content fragment (Alpine x-for/x-if/x-teleport). CSS/layout: class/style attributes keep 2 KB instead of 120/240 bytes (Tailwind class lists), rules are indexed by their unescaped class (.md\:flex), flex-wrap: nowrap no longer wraps, aspect-ratio and repeat(N, 1fr) grids are supported, inline SVG icons and text-only inline-blocks stay in the text line, block/absolute images keep their aspect ratio under max-width, sticky z-index boxes paint above the flow, form controls and SVGs have an intrinsic width, and a failed stylesheet fetch is retried once.
  • 70a4a22f0f Keep the WM running while lsusb and storage wait on hardware The USB host-controller completion and port-reset loops, the ASIX/RNDIS/ usbmsc probe delays and the SDHCI command/CMD1 polls busy-waited with the BKL held, so every other syscall (the WM included) stalled for the whole scan. The HCD poll loops now call usb_poll_relax(), which yields once a wait passes 2 ms when the caller owns the bus for a sleepable scan, and the probe delays go through usb_delay_ms(). A nestable usb_bus_claim() replaces the scan-only flag; usbmsc claims the bus per command, so the storage rescan can drop the BKL mid-transfer safely. SDHCI waits use vfs_scan_poll()/vfs_scan_relax().
  • Просмотр сравнение для этих 5 коммитов »

6 дней назад

nebulon запушил(а) master в nebulon/nebos

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • c746da5f08 Fix more browser polyfils
  • 40be0da0eb Lay out the Cloudron dashboard: CSS variable scoping, % heights, alignment The logged-in dashboard rendered its sidebar full-width and centred, with the app grid pushed below it. Engine fixes, all general: - Custom properties: :root/html/body/*/:host definitions win; a component-level definition applies only while no other rule gives the name a different value, else var() uses its fallback. Last-wins made Font Awesome's --fa-width 100% for every icon, so each menu entry measured ~1100 px and the sidebar took the whole row. - Percentage heights resolve against the containing block's definite height (viewport for html/body), as a minimum since inner boxes do not scroll; they used to resolve against the width. - Inline-block and flex rows restore the line alignment, so a centred .fa-fw icon no longer centres the label after it. - Flex-item form controls get a default width; placeholders do not scroll; align-items:center centres a flex column's lines. - offsetWidth/clientWidth fall back to the viewport before layout, and window gets a resize event after the first layout and on width changes (the app grid measured 0 on mount and chose the phone layout).
  • Просмотр сравнение для этих 2 коммитов »

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • abbc1fbd1b Answer ARP/ICMP while receiving a netkexec upload The board owns the NIC for the whole transfer, so the idle loop's ARP/ICMP responder could not run. A multi-megabyte upload outlives the host's neighbour entry: after three unanswered probes the host marked the board FAILED and dropped every datagram while sendto() still returned success, which stalled make net-dev at a random offset. Answer ARP/ICMP in netkexec_recv() for the receiving NIC, using that NIC's MAC and link-local default address, so probes and pings keep working during the transfer. On the host side, require the offset to advance: a 10 s no-progress guard now reports a dead neighbour instead of retrying forever on stale ACKs, and an ACK_BADSUM is reported as a checksum failure instead of being treated as completion. Update AGENTS.md, docs/network.html and docs/development.html.
  • aca4c5c3c5 Turn the control panel into collapsible quick settings The whole right top-bar status area is now one button. It opens a control popover flush under the bar and the right screen edge. The panel starts compact with the battery/clock summary, the brightness slider and Quick settings pills for Network, Power mode and Date & Time; clicking a pill expands that section in place. The popover resizes itself to the content (SYS_WIN_RESIZE_SELF plus SYS_SCREEN_INFO) and scrolls when a section is taller than the screen. - Network expands to the existing status card, device/WiFi lists and connection details; it is hidden by default. - Power mode expands to Balanced/Performance/Power saver rows through the new SYS_CPUFREQ. - Date & Time expands to an inline editor (date/time/NTP fields, Set, Now and Sync NTP) instead of spawning settime. SYS_BACKLIGHT drives the brightness slider. - ngui gains ngui_tile() and ngui_slider(). - The WM gives every window creation a globally unique content_gen, so a new window in a just-closed slot cannot reuse a stale front-canvas mapping. - AGENTS.md and docs (graphics, network, userland, cpu, drivers) updated. make all passes; a headless QEMU run shows the compact panel, the Network expansion, inline Date & Time editing and the right status area toggling the panel.
  • 0e15afe99c Keep browser sessions across login POSTs and route #fragments in place - Cookies: the curl helpers re-attach the shared jar before every request (uhttp_init() after a POST cleared h.jar, so the request after the OIDC login POST went out cookie-less: "authorization request has expired"), and save by merging their own changes into /share/bcookies instead of overwriting it. The jar holds 48 cookies and evicts the one set or updated longest ago, so a refreshed session cookie outlives the one-shot OIDC interaction cookies. - Fragment navigation: location.hash is live; a change of only the #fragment (location.hash/href/assign/replace, #... links, back/forward) updates the URL, address bar and history and fires hashchange instead of reloading, which the hash-routed Cloudron dashboard depends on. window events no longer run their handlers twice. - <select> exposes options/multiple, and an <option> without a value reports its text (Vue's v-model on selects).
  • a6d2afec15 Persist browser cookies and localStorage in /data When the installed disk's /data volume is mounted, the browser keeps the helpers' cookie jar in /data/BCOOKIES.BIN and every origin's localStorage in one /data/BSTORE.DAT, so logins survive a reboot. /share stays the working copy: the jar is restored into /share at start, a store missing from /share is read from disk on first use, and the main loop writes back what changed at most every 3 s. A cleared store is kept as "{}" so the disk copy cannot resurrect it. sessionStorage stays in RAM.
  • Просмотр сравнение для этих 4 коммитов »

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • d57c67b1e4 Snap windows to screen edges and resist adjacent windows Dropping a title-bar drag on the left or right screen edge snaps the window to that half of the work area at full height; dropping it at the top edge maximizes it. The restore geometry is saved only when leaving the normal state, so snap -> maximize -> restore returns to the original, and dragging a snapped or maximized window restores it first while keeping the grab point at the same fraction of the title bar. While dragging, the screen edges and the facing edges of windows that overlap on the other axis resist the window: it stays flush until the pointer overshoots by WM_RESIST (24 px), which makes exact side-by-side placement easy. AGENTS.md and docs/graphics.html updated. Ignore the generated mkrtwtbl host tool.

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • 8513cec844 Run ES modules, fetch and Web Storage in the browser Gets the Cloudron dashboard (a Vite/Vue 3 app) through its OIDC login: - <script type="module"> runs through QuickJS's module loader; imports load synchronously via the helpers, modulepreload chunks are fetched in parallel first, and unhandled promise rejections are reported. - js_prelude.js adds fetch/XHR, URL/URLSearchParams, TextEncoder, btoa, Web Storage (persisted per origin in /share), crypto.subtle SHA-256, observers, DOM classes and common Element methods. - The curl helpers share one cookie jar (/share/bcookies) and accept any method/content type; empty-body form POSTs stay POSTs. - uhttp percent-encodes unsafe path bytes; URL normalization no longer collapses "//" inside query strings (broke redirect_uri). - location.replace/assign/reload navigate; the script budget excludes network waits; data-block <script> text is kept for textContent. - Script-built inputs get their type/placeholder, flex-item controls are drawn, fields stretch in flex columns and clip/scroll their text, and clicks reach plain elements with listeners.
  • 47aeb06a63 Use one font size in the network popover Render every heading, section caption, row and footer line at the toolkit's 14 px default (bold for headings and captions) instead of mixing 11, 14 and 16 px. Give the panel a consistent 12 px outer inset and 10 px gaps, use 28 px section headers, grow two-line rows to 42 px and single-line detail rows to 30 px, and drop the "Join again without the passphrase" subtitle from the Saved networks row. make all passes; the popover layout was checked in a headless QEMU run.
  • 484756be4d Turn the network menu into a flat popover Open the top-bar network settings as a chromeless WIN_F_POPOVER canvas window anchored under the status icon instead of a titled window. The WM keeps it topmost, dismisses it on an outside click or Esc, and toggles it from the icon; it draws no title bar, shadow or rounded outer corners, so the panel reads as a flat menu. - wm_abi.h, wm.c: carry the popover flag, use the window origin for content-relative pointer/scroll coordinates and clamp the panel fully on screen. - ucmd_lib.h: add u_win_create_popover(). - ucmd_wm.c: skip window chrome for popovers, keep them above every window, anchor them under the network icon, fix the damage origin and dismiss them on an outside press. - ngui.c/h: add the popover body/footer component, section headers, two-line list rows and signal bars. - ucmd_netmenu.c: rebuild the menu around those components with a connection status card, scrollable Devices/Wi-Fi/Connection sections, passphrase, progress and saved-network views and a footer status line. - AGENTS.md and docs (graphics, network, userland) updated. make all passes. A headless QEMU run opened the popover from the status icon, checked the flat panel and the rich content, and verified outside click, icon toggle and Esc dismissal.
  • 3527e786f9 Clip chat avatars to circles Avatar bitmaps now get a 4x4 supersampled circular alpha mask when they are cached, so message avatars and the New DM picker show round pictures instead of square images over the round fallback. The initials fallback is unchanged, and the cache keeps the alpha plane for opaque pictures as well. AGENTS.md and docs (network, graphics) updated. The chat e2e run passes; its screenshots show the circular avatars.
  • cce9edb987 Add Cherryview panel backlight and a brightness OSD The HP Chromebook 11 G5 (google/setzer, Braswell) drives its internal panel PWM from the display engine, not the Chrome EC. backlight.c/h matches the Intel 8086:22b0/22b1 graphics device, maps the graphics MMIO BAR and drives the VLV/CHV BLC_PWM_CTL2/BLC_PWM_CTL registers (pipe A at 0x1E1250/0x1E1254, pipe B +0x100), following the pipe the firmware enabled. The 10% steps clamp at 5..100% and the level is read back from the register at boot. - backlight.c/h, gpu.c: probe the CHV PWM, expose present/percent/step/seq and a status snapshot for the command. - io.c/h: consume F6/F7 (the ChromeOS EC action row, guarded by cros_ec_present()) and E0 14/E0 15 (standard brightness keys); add a 64-entry raw scancode tap for the keytest helper. - terminal.c: add the backlight command (probe/up/down/0-100) and keytest, which prints every raw byte with its E0/make/break state and decoded key. - wm_abi.h, wm.c: carry the backlight level and change counter in wm_status. - ucmd_wm.c: show a centered 3 s Brightness bubble with a sun icon, the percentage and a level bar, composited above every window (including the drag/resize passes). - Makefile, AGENTS.md and docs (graphics, drivers, development, userland) updated. make kernel.bin and ucmd_wm.elf pass; QEMU boot and keytest were checked. The CHV PWM path was validated on the target and the OSD layout was checked with a forced level in QEMU.
  • Просмотр сравнение для этих 10 коммитов »

1 неделя назад

nebulon удалил ветку main в nebulon/nebos

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • 852e70701c Speed up boot to the terminal and keep storage/lsusb from stalling The terminal is on screen ~1.5 s after QEMU starts instead of ~5.3 s (the kernel now starts at ~0.8 s instead of ~2.2 s). Tested on target. - boot.S: EDD reads up to 64 sectors per INT 13h into ES:0 (was one sector per call, ~700 calls for the kernel); a refused multi-sector read falls back to one sector per call. CHS stays single-sector. - vesa_stage.S: one INT 13h per 16 KB rootfs chunk (32 sectors over EDD, up to the end of the track over CHS) instead of ~3600 single-sector calls; same single-sector fallback. - inflate.c: the whole-buffer helpers (ramfs pack, image decoders) use 10-bit Huffman lookup tables, decode into the output buffer and checksum once: the 3.7 MB pack inflates in ~0.25 s instead of ~1.1 s. The streaming decoder is unchanged. - e1000e.c: skip the PCIe GIO-master handshake on the PCI 82540EM; its status bit never clears in QEMU and cost the full 1 s timeout. - disk.c: skip the floppy probe when CMOS register 0x10 says no drive. - kernel.c: the boot storage scan (and the /data WiFi autoconnect after the boot DHCP) runs in a "bootscan" kernel thread; boot_ready is set right after the WM starts. /share is mounted with /ramfs at boot. - sys_fs.c/vfs.c: /ramfs and /share operations and spawns from them no longer wait for a rescan; the storage rescan runs preemptible and its floppy/ATA/AHCI/SDHCI wait loops sleep (vfs_scan_poll/relax) for the scan's owner, so `storage` no longer freezes the WM. - usb.c and the HCDs: usb_delay_ms() sleeps inside a scan owned by a sleepable process (lsusb, the kworker's HID rescan/hotplug); other contexts are refused controller access while it sleeps (usb_scan_foreign), and usbhid_poll skips its tick. - sys_wm.c: COM1 prints "wm: first window frame (term) at N ms" as the boot-time measurement end point. AGENTS.md and docs/ (boot, filesystems, drivers, network, development) updated. test-usb, test-install, test-uefi, test-images, test-libc and test-js pass.
  • 3d3fef3263 Add Realtek RTL8821AU USB WiFi (TP-Link Archer T2U Nano) A port of the Linux rtw88 8821A/USB bring-up as a third WiFi backend, so the wifi command, network menu, WPA2 supplicant and autoconnect work unchanged with a plugged-in stick. Tested through QEMU usb-host passthrough on xHCI: 2.4 + 5 GHz scan, WPA2-PSK connect and DHCP. - rtl8821au.c: USB probe (Linux rtw88_8821au ID list), vendor-request register access, eFuse, power sequences, LLT, 8051 firmware download, MAC/BB/AGC/RF tables with the 8821A condition rules, band/channel switch, eFuse TX power, software channel-hop scan, RX descriptor/RSSI parsing, fixed-rate TX on a single bulk-OUT endpoint (the xHCI driver has one bulk ring). Delays are busy waits: SYS_WIFI runs with interrupts off, so the tick is frozen inside a call. - third_party/rtw88 + mkrtwtbl.c: the Linux register tables (BSD-3-Clause) become firmware/rtl8821a.tbl in the pack, keeping ~11 KB out of the kernel; rtw8821a_fw.bin is packed from linux-firmware. - ccmp.c: ccmp_tx_encrypt() (the chip's crypto engine stays off), with 802.11 and RFC 3610 encrypt vectors in make test-ccmp. - wifi_fw.c (was iwl_fw.c): backend dispatch; a plugged-in stick wins over the built-in Intel card. - netdev_chip(): the network menu names the real WiFi chip. - usb.c: six device-probe slots. - AGENTS.md, docs: RTL8821AU section, known issue 38, size numbers.
  • 02ed133120 Add Intel 6200/6300 WiFi (iwlwifi DVM) for the ThinkPad X201 The X201's Centrino card (a 6300, 8086:4238, on the test machine) runs the older iwlwifi DVM firmware instead of the 7265D's MVM firmware. Share the PCIe transport and wifi.c between both and add a DVM firmware layer. - iwl_trans.c: per-family descriptor (fam_7000/fam_6000): queue count, command queue/FIFO, header size, byte-count units, SCD chain-ext workaround, shadow registers, firmware path and the pre-TLV v2 .ucode format; transport-programmed AC queues (iwl_txq_enable_fifo), a NIC config hook and iwl_fatal(). - iwl_dvm.c: OTP/EEPROM read (MAC, radio config, crystal calibration, regulatory channel tables), INIT image calibration capture and runtime replay, RXON commits with station restore and fixed-rate LQ tables, TX power, per-band scans over the EEPROM's valid channels, ADD_STA keys, TX and the DVM tx_resp layout (status at byte 32). - iwl_fw.c: dispatch between iwl_mvm.c and iwl_dvm.c by probed family. - wifi.c: a firmware assert drops WiFi to DOWN so the next scan or connect reloads the firmware. - Makefile: pack iwlwifi-6000-4.ucode next to the 7265D firmware (WIFI_FW_7265D / WIFI_FW_6000 override the host search). Tested on the X201: firmware boot, 2.4 GHz scan, WPA2 connect, DHCP and ping. The EEPROM channel scan, crash reload and tx status fix are built but not yet run on the board. Docs and AGENTS.md updated.
  • 22a97523c0 Share one selector engine, add ::before/::after and speed up page scripts The cascade, querySelector(All) and matches now use one selector engine: attribute values (= ~= |= ^= $= *=, i flag), the + and ~ combinators with backtracking, the nth-* pseudo-classes, :empty, :checked, :disabled/:enabled, :required/:optional, :not/:is/:where with selector lists and :has. User-state pseudo-classes never match and their rules are dropped at parse time. The rule index no longer files li:not(.x) under class x. ::before/::after generated text (strings, attr(), quotes) is cascaded as an anonymous inline box by br_pseudo() and emitted as literal runs, so scripts never see it (tagesschau's "| mehr"). Text offsets in nodes and layout runs are 32-bit: the text arena is 512 KB, so pages with more than 64 KB of text drew the wrong strings. QuickJS is built -O1: on the Cloudron store it is faster than -Os and -O2. The status bar shows a per-phase load breakdown, and the store loads in 2572 ms under TCG (was 3414). Scripts get a 20 s budget and an interrupted script still reports its error. e1000: pin the 82540EM to legacy RX descriptors. QEMU stores RFCTL, so EXTEN read back set and RX was dead.
  • Просмотр сравнение для этих 4 коммитов »

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • dc5395d1aa Start with a link in qemu
  • ecbc4c04aa Replace MuJS with QuickJS and make the Cloudron app store work MuJS only implements ES5, so modern pages (const, arrow functions, classes, async/await) failed to parse; cloudron.io/store's ES2015 controller never registered its Angular module. Engine: - Vendor QuickJS 2026-06-04 (ES2025) in third_party/quickjs with small __nebos__ patches (no atomics/pthreads, no fenv.h, alloca.h, UTC timezone), built freestanding with -nostdinc. MuJS and mujs_test are removed. - js_mu.c/h keeps the MuJS value-stack API the DOM layer was written against: C functions see `this` at index 0, userdata has/put/delete hooks become QuickJS exotic-class methods with a backing object for expandos, js_try/js_typeerror longjmp to the innermost try or C-function boundary, registry refs. Outermost engine entries run on a private 1 MB stack (mu_stack_call; the ring-3 stack is 128 KB), Promise jobs run after each script/timer/event, C functions are constructible (new Event(...)), and a 10 s wall-clock interrupt replaces the instruction limit. - js_browser.c: wrappers inherit from Object.prototype (Angular calls node.toString()), errors report "Name: message" plus the stack, and an inline-handler catch path no longer pops its try frame twice. - libc: INT*_MIN/SIZE_MAX/INT64_C, memchr, malloc_usable_size and the C99 math QuickJS needs (hypot, cbrt, log2, log1p, expm1, hyperbolic, lrint), with host checks in test_math.c. New make test-js (ASan/UBSan). Browser fixes found on the store page: - br_clone_node wrote the copy's attributes onto the previous node (it passed a 1-based handle to br_set_attr), breaking every ng-repeat. - element.click() now dispatches a bubbling, cancelable click event before the default action. - Images: 256 bitmaps within a 4 M-pixel budget (was 16) and 512 background URLs (was 16); background bitmaps scale to cover their box; one reusable decode arena per page instead of malloc/free per image, which fragmented the heap until 200 icons ran out of memory; lazy background loading against measure-pass block_top[]; in-flight jobs are marked so rescans skip them; style/class/src changes from scripts re-arm nodes the first scan marked empty, and script DOM changes trigger a rescan; image jobs time out after 30 s; rounded boxes paint their background image inside the corner arcs; alpha images are composited over white instead of the dark theme. - The ring-3 heap cap (USER_HEAP_MAX, libc HEAP_MAX_BYTES) is now 64 MB: the DOM arenas plus the QuickJS heap use ~30 MB before images. Compat ladder: new 071-modern-js; 036, 054 and 055 pass on the new engine.
  • Просмотр сравнение для этих 2 коммитов »

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • 094a42f397 Fix the X201 rootfs load with a protected-mode pack copy The X201 BIOS's INT 15h AH=87h block move silently copies zeros to PACK_HIGH_ADDR, so the kernel received an empty pack, took the raw-pack path and wedged before the WM. Copy each 16 KB staging chunk with a temporary flat protected-mode rep movsb instead; leave through a 16-bit protected-mode code segment before clearing PE and returning to real mode. Update the boot-chain documentation in AGENTS.md and docs/boot.html.
  • 13291ce526 Render inline SVG and fix tagesschau.de layout gaps tagesschau.de showed no header logo, wordmark or burger icon, a dark box in the navigation, its hero image below a grey placeholder, doubled teaser line spacing and "mehr" links on their own line. - Inline SVG is now rendered by svg_raster.c: every path command including curves and arcs, rect/circle/ellipse/line/polyline/polygon, <g>/nested <svg>/<use> of a <symbol>, transform lists, viewBox with preserveAspectRatio, nonzero/even-odd fill, strokes, opacity, currentColor and gradients painted as their average stop colour. Coverage is antialiased into an RGB565 + alpha layer cached per node and size; percentages resolve against the viewBox. Host test: make test-svg. - SVG paint is its own inherited style state (fill/stroke modes, colours and stroke width) instead of reusing background/border, so CSS fill on an HTML element (the burger button) no longer paints its background. getComputedStyle exposes fill and stroke. - Absolutely positioned boxes: the deferred paint list held 16 entries, so later posters fell back into the flow below their padding-top aspect box. It now holds 256, and an overflow still paints against the containing block. - line-height in rem is an absolute length, not a multiplier. - An inline-flex box around inline content stays in the text line. Compat ladder: new pages 067-070, and 058 now checks fill instead of the old background shortcut.
  • a1e381e370 Drain the fetch helpers' log tty so cross-origin loads cannot hang The curl __serve helpers write their DNS/TLS log lines to the private tty the browser creates with u_tty_create(), but nothing ever read it. A full 4 KB tty ring blocks its writers while the owner is alive, so after a page that opened a dozen or so TLS connections (tagesschau.de) every helper hung in the next log line of a fresh connection, and the following cross-origin navigation (cloudron.io) sat at "Loading..." until it timed out. The browser now drains and discards that tty from every fetch poll loop and the idle main loop. tty_write also drops output instead of sleeping when the ring's owner writes to its own full ring (the browser's [js] lines go there too), since only the owner can drain it. The docs now describe the helpers staying warm across origins, which the previous commit introduced.
  • c1edf1b343 Fix stalled keep-alive requests and image decode bursts Treat an HTTP/1.0 response as connection-close unless it explicitly opts into keep-alive: reusing the half-closed socket discarded the next request, so after the first worker batch later images waited for a response that could not arrive until the retry timeout. A reused connection now waits 3 s for response headers (a fresh one keeps the 10 s idle timeout) so a silently closed keep-alive is retried promptly. Decode at most one image per main-loop pass so the TCP pump, timers and input run between JPEG decodes on slow CPUs, and keep the four idle fetch helpers across navigations instead of killing and respawning all of them while images are still loading.
  • b2bdb6517e Fix tagesschau layout and large-page loading Let the shared HTTP client survive slow but steady transfers (the idle timeout still catches a stalled peer), and grow the browser attribute arena to 4 MB so the 1.5 MB tagesschau page leaves room for its external CSS selectors. Treat a flex container on an inline tag (link cards) as a block box, ignore width/height on non-replaced inline boxes, accept minified leading-dot lengths and fractional percent font sizes, support structural pseudo-classes, keep fully transparent borders invisible, and raise the layout recursion budget (a DOM level costs about two units) so deeply nested teaser text is no longer skipped. Adds ladder 061-066.
  • Просмотр сравнение для этих 8 коммитов »

1 неделя назад

nebulon запушил(а) master в nebulon/nebos

  • 3b2607df02 Prefer a wired link over WiFi at boot Expose the physical link state from e1000e, r8168 and ASIX through the optional netdev link_up() hook, and skip the boot DHCP when the active NIC has no link. wifi autoconnect now checks that link instead of a DHCP lease, so a connected Ethernet cable wins and WiFi is only probed and joined when the cable is down. Keep an r8168 link that survived the probe reset, with a bounded wait only when the reset drops a link that was already up. Update AGENTS.md and the network page.
  • 63a76a46bc Turn the top-bar power text into status icons Show the battery percentage only in a hover tooltip below the far-right power icon, and drop the CPU prefix from the temperature. The icon is a plug while the adapter is online and a level-filled battery on battery. Draw the tooltip after the windows in every render path, including window moves and drags, so it always stays above them. Keep the bar hover highlight above the rainbow separator. Update AGENTS.md and the graphics/drivers/cpu pages.
  • Просмотр сравнение для этих 2 коммитов »

1 неделя назад