What shipped, when.

fix: TagFS/ATA bug fixes — braid disk mapping, meta_pool correctness, ata_flush_cache

TagFS / Braid: - braid.c: BraidRawRead/Write were passing disk_id directly as is_master; disk_id=0 mapped to ATA slave (is_master=0) causing all "Device does not exist" errors. Fixed: disk_id 0 → is_master=1 (primary master), disk_id 1 → is_master=0. - tests.c: test_braid_add_disk left disks 0 and 1 in global Braid state; BraidIsHealthy() stayed true → all subsequent TagFS I/O routed through non-existent slave. Fixed: added BraidRemoveDisk(0)/BraidRemoveDisk(1) cleanup at end of test. MetaPool: - meta_pool.c: When chaining to a new block, new block was only initialized in-memory but not written to disk. A crash/Ctrl+C before meta_pool_flush caused the CoW manifest to overwrite the same physical block, triggering "bad magic in chain" on next boot. Fixed: write new block header to disk immediately after setup. - meta_pool.c: pack_record() return value was ignored in meta_pool_write. On invalid extent, pack_record wrote only a partial record (header+tags, no extents, no CRC) then returned 0, but used_bytes was still advanced by the full record_size, silently corrupting the block. Fixed: check pack_record return and propagate error. - meta_pool.c: mirror_deep_copy() silently left tag_ids=NULL with tag_count>0 on kmalloc failure, causing NULL deref in callers that iterate tag_ids[i]. Changed return type to bool; on any allocation failure, already-allocated fields are freed and false is returned. Seqlock write site invalidates the mirror entry on failure; read site returns -1. ATA driver: - ata.c: ata_flush_cache() did not hold g_ata_lock, allowing a race with concurrent PIO operations on the same bus. Fixed: acquire g_ata_lock before issuing flush command. - ata.c: ata_read_sectors_retry / ata_write_sectors_retry were retrying pointlessly when the device does not exist. Fixed: early-exit on ATA_ERR_NO_DEVICE. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: PMT correctness — PHYS_TAG_SHARED, span-find bug, PocketRing/ResultRing tagging

Bug fix: - phys_tag_find_span_from: start_bit underflow when a span crosses word boundaries. When in_span=true and span_start_page belongs to a previous word, the old code computed (span_start_page - w*64) which underflows for w > span_word. Fixed: start_bit = 0 when span started in a prior word, only use page offset when span_start_page is in the current word. New tag: - PHYS_TAG_SHARED (bit 5) — marks physical pages shared between kernel and userspace. Not a zone hint; a semantic annotation. PocketRing/ResultRing tagging: - process_create: PhysTagSet(pocket_phys, ..., PHYS_TAG_SHARED) and PhysTagSet(result_phys, ..., PHYS_TAG_SHARED) immediately after vmm_create_cabin. These ARE shared memory (kernel guide loop writes ResultRing, userspace reads; userspace writes PocketRing, kernel reads). PMT now reflects this. - process_destroy: PhysTagClear on both rings before ref_dec — tags are removed when the cabin is about to be unmapped. PhysTagDump: added SHARED to known_tags list (now 6 entries). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat: Physical Memory Tag (PMT) — replace friend.c with dynamic tag overlay

PMT is a 64-bit tag-per-page overlay on all physical memory. Any physical address range can carry an arbitrary uint64_t tag bitmask. Multiple tags per page (OR semantics). Lookup is O(1) by address (tag_map[page_idx]). Query "all pages with tag X" uses 64-band bitmaps: O(page_count/64) with word-level AND reduction and __builtin_ctzll. Built-in tags: PHYS_TAG_DMA32 / USER / HIGH / KERNEL / MMIO. Bits 16-63 available for drivers, processes, shared-memory annotations. friend.c/h deleted. PhysAllocTagged(pages, required_tags) replaces FriendAlloc. PhysTagInit() called from main.c after vmm_init(); allocates tag_map and 64 band bitmaps via kmalloc, then classifies all E820 regions. pmm_activate_pull_map() calls PhysTagSet(..., PHYS_TAG_HIGH) on deferred regions once buddy activates them. Key properties vs. old friend.c: - No hardcoded zones — boundaries derived from E820 at runtime - Any number of tags on any region (not just 3 zones) - Shared memory: multiple processes tag the same physical region, union of their tags makes it discoverable via PhysTagFindSpan(TAG_A | TAG_B) - PhysAllocTagged retries up to 16 spans on buddy exhaustion - PhysTagGet is lock-free (single 64-bit load after init) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: bugs, hardcode removal, optimizations, complete unfinished items

Bugs fixed: - async_io: BGND lane starvation — weighted round-robin (1 BGND per 8 DATA) - buddy: out-of-bounds in buddy_free() — range validation before page_index() - ahci: ci_snapshot race — atomic_load_n(ACQUIRE) on port->ci and hba_mem->is - pmm: MAXPHYADDR assertion at init so post-alloc check is unreachable - kcore: spin-wait deadlock guard — KCORE_POP_SPIN_LIMIT resets stale slot Hardcode removed: - kernel_config.h: all tunables centralized (AHCI timeout, lane caps, spin limits, zone boundaries, listen table max, cache limits) - friend.c: zone boundaries now derived from pmm_get_mem_end() + E820 at FriendInit() — correct on machines with < 1GB or > 4GB RAM - ahci.c: CONFIG_AHCI_CMD_TIMEOUT_MS replaces magic 2000 Optimizations: - listen_table: O(1) duplicate detection via 256-bucket hash index on (pid × 2654435761 XOR required_tags) - async_io_cancel_by_pid: flag-based cancel (sets req->cancelled=1, no kmalloc, no queue rebuild); dequeue_lane() skips cancelled entries - buddy_alloc_range: fast-reject continue for blocks outside [min,max) Unfinished completed: - FriendFree: evicts to buddy when zone cache >= CONFIG_FRIEND_ZONE_CACHE_MAX_PAGES - listen_table_add/remove: return error_t (ERR_LISTEN_ALREADY, ERR_NO_MEMORY, ERR_LISTEN_TABLE_FULL, ERR_PROCESS_NOT_FOUND) instead of magic int - listen_table: LISTEN_TABLE_MAX_ENTRIES enforced at add time Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

feat: Friend Allocator, KResult, Tag-Priority Lanes, Listen Table tags

- Friend Allocator (friend.h/c): zone-aware PMM layer over buddy allocator. PhysZones DMA32/USER/HIGH with per-order free-list caches. O(1) hot path, buddy_alloc_range() fallback for cache misses. FriendInit() called from pmm_activate_pull_map() after high-memory is activated. - KResult (kresult.h): typed result struct {status, context, value}. KRESULT_OK/KRESULT_ERR macros with KResultContext subsystem IDs. Result.context field replaces _reserved — userspace now sees which subsystem produced the result (KCTX_GUIDE, KCTX_IPC, etc.). - Tag-Priority Lanes (async_io): replaced single FIFO with 3 semantic lanes (LANE_META/DATA/BGND, capacities 64/256/64). K-Core dequeues META first, guaranteeing TagFS metadata ordering before data during crash window. - Listen Table tag namespace: replaced hardcoded LISTEN_KEYBOARD=0 enum with uint64_t event_tags bitmask. Listeners register required_tags mask; match: (event_tags & required_tags) == required_tags. Eliminates enum limit — new devices need only a new tag bit. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

added check.sh

fix: bootloader A20/E820, PMM deferred regions, async_io cross-core safety, scheduler return type

Bootloader (stage2.asm): - Fix test_a20: xor ax,ax → mov ax,0 to preserve ZF from cmp (xor silently set ZF=1 unconditionally, making test_a20 always return "disabled") - Add fatal halt after all 3 A20 methods if A20 remains disabled - Replace INT 15h 0x88 fallback (64MB cap) with INT 15h 0xE801 (up to 4GB); 0x88 kept as last resort PMM (pmm.c): - Remove static 8-slot PMM_DEFERRED_MAX array; allocate deferred region array from bootstrap physical memory sized to E820 entry count — no high-memory regions silently dropped async_io (async_io.h/c, ata_dma.c, storage_deck.c): - Add uint8_t port_num to async_io_request_t for AHCI multi-port routing - submit_time (rdtsc) → submit_tick (kernel_tick_get) — PIT-driven global tick avoids cross-core TSC skew - Add src/include/kernel_clock.h: single declaration of g_global_tick + kernel_tick_get() Process (process.h/c): - volatile uint32_t ref_count → atomic_u32_t for type consistency with atomics.h conventions - plain assignment at creation → atomic_store_u32 Scheduler (scheduler.h/c, main.c): - scheduler_init() void → error_t; returns ERR_INVALID_ARGUMENT / ERR_NO_MEMORY on failure - main.c panics on scheduler_init failure instead of continuing silently broken Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

TagFS and other bug fixes

fix: TagFS subsystems fully functional — DiskBook, CoW, Dedup, Braid integrated

- DiskBook: fixed sector addressing (was calling tagfs_read/write_block treating sector as block number); now uses direct disk_read/write_sectors with correct LBA - CoW: real metadata passed to DiskBookLogMetadata (was NULL → silent fail); snapshot_count++ moved inside lock region (TOCTOU fix); CowAllocSnapshotId atomic - Dedup: integrated into tagfs_write — dedup check before alloc, register on write - Braid: integrated into read_block/write_block; BraidVerifyBlock uses cross-disk consensus; BraidAutoHeal verifies before propagating; correct LBA passed (physical sector via block_to_sector, not raw block num × 8) - meta_pool_init: chain traversal capped at block_count to prevent infinite loop - ata.c: fixed lock leak — spin_unlock missing before protected/OOB early returns, causing all subsequent ATA ops to deadlock after any blocked write Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

fix: scheduler races, recalc throttle, tagfs double-release panic

- tagfs/tagfs.c: ofe_release — replace silent debug_printf+return on ref_count==0 with spin_unlock+panic; double-release is a kernel invariant violation and must not be silently ignored in any build. - scheduler.h: is_parked bool → _Atomic bool; the field is written by the BSP (PIT IRQ recalc) and read by AP cores (LAPIC timer IRQ) with no synchronization — C11 UB, eliminated by making the type atomic. - scheduler.h: add SCHED_RECALC_INTERVAL=10 constant. - scheduler.c: throttle scheduler_recalc_parameters() to fire at most once every SCHED_RECALC_INTERVAL global ticks (~25/s at 250 Hz instead of 250/s); reduces O(cores) work and PIT reprogramming rate. - runqueue.h/runqueue.c: add _Atomic uint32_t total to RunQueue; maintained with RELEASE stores in enqueue/dequeue/remove; exposed via RunqueueAtomicTotal() with ACQUIRE load for lock-free reads. - scheduler.c: replace all lock-guarded or bare runqueue_total_count() calls in recalc_parameters, sched_enqueue, and sched_try_steal with RunqueueAtomicTotal(); removes an unnecessary lock/unlock pair from the steal victim scan. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

added claude

added claude skills

DEBUG=on making now

new commit for github

Dataroll rnamed to DiskBook and rewrited! adeed some unique features like Tag-Aware-Journaling. Journal duplicate deleted. Let me what is this now?

1. Система ошибок (error.h, error.c)

- Добавлено 40+ новых кодов ошибок для полного покрытия всех подсистем - Переименовано error_string() → ErrorString() (PascalCase) - Добавлены категории: General, Memory, I/O, Filesystem, Process, Security, Hardware, ACPI, Pocket/K-Core, Scheduler, Boot 2. PMM (pmm.c, pmm.h) - pmm_init() → возвращает error_t - pmm_set_maxphyaddr() → возвращает error_t - Исправлена гонка: проверка переполнения pmm_deferred_count перед записью 3. VMM (vmm.c) - pcid_alloc() → pcid_alloc_safe() с error_t - Исправлена гонка: bounds check для pcid_free_stack[] - PCID exhaustion возвращает ERR_PCID_EXHAUSTED вместо паники 4. K-Core (kcore.c, kcore.h) - Критическая гонка исправлена: удалена очистка kcore_pending при переполнении очереди - kcore_queue_push() → возвращает error_t - kcore_submit() → возвращает error_t - Добавлен mfence() для правильного memory ordering 5. Scheduler (scheduler.c, scheduler.h) - sched_enqueue() → error_t вместо bool - sched_enqueue_on() → error_t вместо void - sched_dequeue() → error_t вместо void - Валидация home_core перед доступом к массиву 6. Process (process.c) - Обработка ошибок от sched_enqueue() в process_set_state() - Использует ErrorString() для консистентных сообщений 7. Guide (guide.c) - Исправлено error_string() → ErrorString() Ключевые исправленные гонки 1. K-Core pocket loss - kcore_pending очищался преждевременно, вызывая потерю pockets 2. PMM deferred region overflow - отсутствовала проверка границ перед записью в массив 3. VMM PCID allocation - потенциальный array overflow 4. Memory ordering - добавлен mfence() в kcore queue операциях

some fixes

some fixes

fixed PID issue

some additions/fixes

added test hello.txt file

USB HID here. so-so maybe?

some fixed races maybe?

added build script

improved tagfs! so-so?

some kill for hardcode!!

some additions.

some additions to keyboards

some fixes

some additions.

some additions.

some additions - refactoring. Deleted BRUH externs...

some additions & fixes MB

some fixes

tagfs some deadlocks fixes

some races fixes

makefile improved

reboot/shutdown some additions 3 )0)0)0))0)0

reboot/shutdown some additions 2

reboot/shutdown additions

pmm some tests added

some fixes

added some fetures. TLB also

async syscall so-so???

async syscall skeleton

compilation warnings fixed

amp phase 3

amp skeleton 2 phase

amp skeleton phase 1