feat: extend syscall with LoadedImage and loadImageByPath

src/syscall.c

@@ -101,11 +101,202 @@ static int64 doRead(uint64 fd, uint64 buffer, uint64 length)
     return vfsRead((int)fd, (void*)buffer, length);
 }
 
+typedef struct LoadedImage {
+    uint64 entry;
+    uint64 stackTop;
+    uint64 brkStart;
+} LoadedImage;
+
+/* Shared by doExec and doSpawn: resolves `path` through the VFS, loads its
+   LOAD segments via elfLoad (which maps them at whatever fixed virtual
+   address they were linked at -- see elf.c), then derives and maps a fresh
+   stack for that slot from nothing but the ELF's own entry point (see
+   elfSlotStackTop/elf.h). Leaves the caller's thread state untouched;
+   callers decide what to do with a successfully loaded image. */
+static bool loadImageByPath(const char* path, LoadedImage* out)
+{
+    VfsNode* node = vfsLookup(path);
+    if (!node)
+        return false;
+    /* Real EACCES-shaped refusal: a file without the right exec bit for the
+       CALLING thread's identity (vfsPermitted, owner-vs-other-vs-root) can
+       never be exec'd or spawned, exactly like real Unix's own execve()
+       check -- jobctl.c's own jobctlSpawn (the shell's separate, parallel
+       loading path) enforces the identical check the identical way. */
+    if (!vfsPermitted(vfsMode(node), vfsUid(node), vfsGid(node), currentThreadUid(), currentThreadGid(), fsModeOwnerExec))
+        return false;
+
+    const void* image = vfsFileData(node);
+    uint64 size = vfsSize(node);
+    uint64 entry = elfLoad(image, size);
+    if (entry == 0)
+        return false;
+
+    out->entry = entry;
+    out->stackTop = elfSlotStackTop(entry);
+    out->brkStart = elfSlotBrkStart(entry);
+
+    return pagingMapUserRangeFlags(out->stackTop - execStackSize, execStackSize, pageNoExecute);
+}
+
+/* execve()-lite: replaces the calling thread's own program image with a
+   different one, in place -- same pid, same open file descriptors (POSIX's
+   default, no O_CLOEXEC support to opt out of here), new code/stack/brk.
+   Mutates the live interrupt frame's rip/rsp directly (the same trick
+   threadCreateUser uses to bootstrap a brand new thread's very first
+   frame): on success this syscall does not "return" to the caller in any
+   meaningful sense, since regs->rip now points at the new image's entry
+   point instead of back into the old one's code -- exactly real exec()
+   semantics. General-purpose registers are zeroed so the new program never
+   observes leftover state from a syscall it never made. Failure (bad path,
+   not a valid ELF, thread table/frame mapping issue) leaves regs untouched
+   apart from the -1 return value: the caller keeps running its old image,
+   also matching real exec().
+
+   Also reclaims the OLD image's slot: frostbyte's flat single-address-space
+   design (see doSpawn below) means a program that exec()s into a binary
+   linked at a different fixed base -- the common case, since two distinct
+   binaries can never share a slot -- would otherwise leak its entire
+   elfSlotSize worth of physical frames every time, forever. Skipped only
+   when the new image happens to land in the SAME slot as the old one (a
+   binary exec()ing itself): unmapping first would free frames the new
+   image's own loadImageByPath call had already mapped moments earlier.
+
+   argv/argc are already kernel-side copies (copyUserArgv, called from the
+   sysExec dispatch case below, BEFORE this runs) -- critical ordering,
+   since they may point into the OLD image's own memory (a process's argv
+   commonly lives in its own .rodata/stack), which this function is about
+   to unmap out from under it. */
+static bool doExec(const char* path, const char* const* argv, int argc, Registers* regs)
+{
+    uint32 pid = currentThreadId();
+    uint64 oldSlotBase = threadUserSlotBase(pid);
+
+    LoadedImage img;
+    if (!loadImageByPath(path, &img))
+        return false;
+
+    uint64 initialRsp;
+    if (!elfBuildInitialStack(img.stackTop, execStackSize, argv, argc, &initialRsp))
+        return false;
+
+    uint64 newSlotBase = elfSlotBase(img.entry);
+    if (oldSlotBase != 0 && oldSlotBase != newSlotBase) {
+        uint64 stackProbe = oldSlotBase + elfSlotStackOffset - 1;
+        bool wasMapped = pagingResolve(stackProbe) != 0;
+
+        pagingUnmapUserRange(oldSlotBase, elfSlotSize);
+
+        bool nowFreed = pagingResolve(stackProbe) == 0;
+        kprintf("[exec] pid=%u freed old slot %p (was-mapped=%s now-freed=%s) -> %s\n",
+                pid, (void*)oldSlotBase, wasMapped ? "yes" : "no", nowFreed ? "yes" : "no",
+                (wasMapped && nowFreed) ? "exec-old-slot-freed-ok" : "exec-old-slot-freed-FAIL");
+    } else if (oldSlotBase != 0) {
+        /* Self-exec (or the pid==0/no-prior-slot case, unreachable for a
+           real ring-3 caller): the branch above is skipped precisely to
+           avoid unmapping the NEW image's own LOAD segments/stack, which
+           already landed on these same pages -- but any mmap mapping from
+           the OLD image is a disjoint sub-range (mmap.h's own mmapReserve
+           region, [slot+elfSlotStackOffset, slot+elfSlotSize)) that
+           loadImageByPath/elfBuildInitialStack never touch, so it still
+           needs reclaiming here explicitly or its frames would leak
+           forever every time a binary exec()s itself. */
+        pagingUnmapUserRange(oldSlotBase + elfSlotStackOffset, elfSlotSize - elfSlotStackOffset);
+    }
+    if (oldSlotBase != 0)
+        mmapReleaseAll(threadAt(pid));
+    threadSetUserSlotBase(pid, newSlotBase);
+
+    threadSetName(pid, path);
+    threadSetBrk(pid, img.brkStart);
+
+    regs->rip = img.entry;
+    regs->rsp = initialRsp;
+    regs->rbp = 0;
+    regs->rdi = 0;
+    regs->rsi = 0;
+    regs->rdx = 0;
+    regs->rcx = 0;
+    regs->rbx = 0;
+    regs->rax = 0;
+    regs->r8 = 0;
+    regs->r9 = 0;
+    regs->r10 = 0;
+    regs->r11 = 0;
+    regs->r12 = 0;
+    regs->r13 = 0;
+    regs->r14 = 0;
+    regs->r15 = 0;
+    return true;
+}
+
+/* posix_spawn()-lite: loads `path` into a brand new thread/process rather
+   than replacing the caller, and returns its pid immediately without
+   waiting -- the caller joins it later with sysWait (threadJoin). This is
+   deliberately not a literal fork(): frostbyte has one shared address space
+   system-wide (no per-process page tables/CR3, see paging.c), so there is
+   no address space to duplicate, only a fresh slot to load a binary into.
+   As with any binary in this kernel, spawning two live instances of the
+   same path (or of two paths whose fixed link addresses happen to share a
+   slot) at once would map both onto the same pages and corrupt each other;
+   this is an accepted limitation of the flat single-address-space design,
+   not something doSpawn can detect from a path string alone. */
+static int64 doSpawn(const char* path, const char* const* argv, int argc)
+{
+    LoadedImage img;
+    if (!loadImageByPath(path, &img))
+        return -1;
+
+    uint64 initialRsp;
+    if (!elfBuildInitialStack(img.stackTop, execStackSize, argv, argc, &initialRsp))
+        return -1;
+
+    uint32 pid = threadCreateUser(path, img.entry, initialRsp);
+    if (pid == threadInvalidId)
+        return -1;
+
+    threadSetBrk(pid, img.brkStart);
+    return (int64)pid;
+}
+
 uint64 syscallDispatch(Registers* regs)
 {
+    /* This thread is not (yet, or any more) parked mid a retryable
+       syscall's own block-and-retry pattern the instant ANY dispatch for
+       it begins -- whether this dispatch turns out to BE that exact retry
+       re-executing (sysRead/sysWrite/sysWait's own cases set it true again
+       below if they block once more) or a completely unrelated syscall.
+       See sched.h's Thread.sigInRetryableSyscall and signal.c's
+       sigConsumePending for why this precise a window matters: without it,
+       a signal delivered while this thread is doing perfectly ordinary
+       work could be mistaken for one interrupting a slow syscall. */
+    currentThread()->sigInRetryableSyscall = false;
+
     switch (regs->rax) {
-    case sysWrite:
+    case sysWrite: {
-        regs->rax = (uint64)doWrite(regs->rdi, regs->rsi, regs->rdx);
+        int64 result = doWrite(regs->rdi, regs->rsi, regs->rdx);
+        if (result == -1) {
+            /* A "would block" result on a fdFlagBlocking pipe end (see
+               vfs.h) is retried, not returned: park on the pipe's channel
+               and rewind rip by the 2-byte `int $0x80` this trap came from
+               so the whole syscall re-executes once schedWakeChannel makes
+               this thread Runnable again (pipeWrite, called from doWrite
+               above, already does that whenever it frees or fills space) --
+               the exact restart-syscall pattern sysWait uses for
+               threadJoin, for the same reason: halt()ing here directly
+               would freeze the whole machine, not just this thread, since
+               this handler runs with interrupts disabled. A non-blocking
+               pipe end (vfsPipe, e.g. pipetest.c) or a permanently broken
+               one gets nullptr back and the -1 passes through untouched. */
+            const void* channel = vfsBlockingChannel((int)regs->rdi, true);
+            if (channel) {
+                currentThread()->sigInRetryableSyscall = true;
+                schedBlockCurrent(channel);
+                regs->rip -= 2;
+                return schedule((uint64)regs);
+            }
+        }
+        regs->rax = (uint64)result;
         return (uint64)regs;
     case sysRead:
         regs->rax = (uint64)doRead(regs->rdi, regs->rsi, regs->rdx);