#include <kmalloc.h>
#include <profiler.h>
#include <stdio.h>
-#include <frontend.h>
-#include <colored_caches.h>
#include <hashtable.h>
#include <bitmask.h>
#include <vfs.h>
#include <kprof.h>
#include <termios.h>
#include <manager.h>
+#include <ros/procinfo.h>
-/* Tracing Globals */
-int systrace_flags = 0;
-struct systrace_record *systrace_buffer = 0;
-uint32_t systrace_bufidx = 0;
-size_t systrace_bufsize = 0;
-spinlock_t systrace_lock = SPINLOCK_INITIALIZER_IRQSAVE;
+static int execargs_stringer(struct proc *p, char *d, size_t slen,
+ char *path, size_t path_l,
+ char *argenv, size_t argenv_l);
-static bool __trace_this_proc(struct proc *p)
-{
- return (systrace_flags & SYSTRACE_ON) &&
- ((systrace_flags & SYSTRACE_ALLPROC) || is_traced_proc(p));
-}
+/* Global, used by the kernel monitor for syscall debugging. */
+bool systrace_loud = FALSE;
-static size_t systrace_fill_pretty_buf(struct systrace_record *trace)
+/* Helper, given the trace record, pretty-print the trace's contents into the
+ * trace's pretty buf. 'entry' says whether we're an entry record or not
+ * (exit). Returns the number of bytes put into the pretty_buf. */
+static size_t systrace_fill_pretty_buf(struct systrace_record *trace,
+ bool entry)
{
size_t len = 0;
- struct timespec ts_start;
- struct timespec ts_end;
- tsc2timespec(trace->start_timestamp, &ts_start);
- tsc2timespec(trace->end_timestamp, &ts_end);
-
- len = snprintf(trace->pretty_buf, SYSTR_PRETTY_BUF_SZ - len,
- "[%7d.%09d]-[%7d.%09d] Syscall %3d (%12s):(0x%llx, 0x%llx, "
- "0x%llx, 0x%llx, 0x%llx, 0x%llx) ret: 0x%llx proc: %d core: %d "
- "vcore: %d data: ",
- ts_start.tv_sec,
- ts_start.tv_nsec,
- ts_end.tv_sec,
- ts_end.tv_nsec,
- trace->syscallno,
- syscall_table[trace->syscallno].name,
- trace->arg0,
- trace->arg1,
- trace->arg2,
- trace->arg3,
- trace->arg4,
- trace->arg5,
- trace->retval,
- trace->pid,
- trace->coreid,
- trace->vcoreid);
- /* if we have extra data, print it out on the next line, lined up nicely.
- * this is only useful for looking at the dump in certain terminals. if we
- * have a tool that processes the info, we shouldn't do this. */
- if (trace->datalen)
- len += snprintf(trace->pretty_buf + len, SYSTR_PRETTY_BUF_SZ - len,
- "\n%67s", "");
- len += printdump(trace->pretty_buf + len,
- MIN(trace->datalen, SYSTR_PRETTY_BUF_SZ - len - 1),
+ struct timespec ts_start = tsc2timespec(trace->start_timestamp);
+ struct timespec ts_end = tsc2timespec(trace->end_timestamp);
+
+ /* Slightly different formats between entry and exit. Entry has retval set
+ * to ---, and begins with E. Exit begins with X. */
+ if (entry) {
+ len = snprintf(trace->pretty_buf, SYSTR_PRETTY_BUF_SZ - len,
+ "E [%7d.%09d]-[%7d.%09d] Syscall %3d (%12s):(0x%llx, 0x%llx, "
+ "0x%llx, 0x%llx, 0x%llx, 0x%llx) ret: --- proc: %d core: %d "
+ "vcore: %d data: ",
+ ts_start.tv_sec,
+ ts_start.tv_nsec,
+ ts_end.tv_sec,
+ ts_end.tv_nsec,
+ trace->syscallno,
+ syscall_table[trace->syscallno].name,
+ trace->arg0,
+ trace->arg1,
+ trace->arg2,
+ trace->arg3,
+ trace->arg4,
+ trace->arg5,
+ trace->pid,
+ trace->coreid,
+ trace->vcoreid);
+ } else {
+ len = snprintf(trace->pretty_buf, SYSTR_PRETTY_BUF_SZ - len,
+ "X [%7d.%09d]-[%7d.%09d] Syscall %3d (%12s):(0x%llx, 0x%llx, "
+ "0x%llx, 0x%llx, 0x%llx, 0x%llx) ret: 0x%llx proc: %d core: %d "
+ "vcore: %d data: ",
+ ts_start.tv_sec,
+ ts_start.tv_nsec,
+ ts_end.tv_sec,
+ ts_end.tv_nsec,
+ trace->syscallno,
+ syscall_table[trace->syscallno].name,
+ trace->arg0,
+ trace->arg1,
+ trace->arg2,
+ trace->arg3,
+ trace->arg4,
+ trace->arg5,
+ trace->retval,
+ trace->pid,
+ trace->coreid,
+ trace->vcoreid);
+ }
+ len += printdump(trace->pretty_buf + len, trace->datalen,
+ SYSTR_PRETTY_BUF_SZ - len - 1,
trace->data);
len += snprintf(trace->pretty_buf + len, SYSTR_PRETTY_BUF_SZ - len, "\n");
return len;
}
+/* If some syscalls block, then they can really hurt the user and the
+ * kernel. For instance, if you blocked another call because the trace queue is
+ * full, the 2LS will want to yield the vcore, but then *that* call would block
+ * too. Since that caller was in vcore context, the core will just spin
+ * forever.
+ *
+ * Even worse, some syscalls operate on the calling core or current context,
+ * thus accessing pcpui. If we block, then that old context is gone. Worse, we
+ * could migrate and then be operating on a different core. Imagine
+ * SYS_halt_core. Doh! */
+static bool sysc_can_block(unsigned int sysc_num)
+{
+ switch (sysc_num) {
+ case SYS_proc_yield:
+ case SYS_fork:
+ case SYS_exec:
+ case SYS_pop_ctx:
+ case SYS_getvcoreid:
+ case SYS_halt_core:
+ case SYS_vc_entry:
+ case SYS_change_vcore:
+ case SYS_change_to_m:
+ return FALSE;
+ }
+ return TRUE;
+}
+
+/* Helper: spits out our trace to the various sinks. */
+static void systrace_output(struct systrace_record *trace,
+ struct strace *strace, bool entry)
+{
+ ERRSTACK(1);
+ size_t pretty_len;
+
+ /* qio ops can throw, especially the blocking qwrite. I had it block on the
+ * outbound path of sys_proc_destroy(). The rendez immediately throws. */
+ if (waserror()) {
+ poperror();
+ return;
+ }
+ pretty_len = systrace_fill_pretty_buf(trace, entry);
+ if (strace) {
+ /* At this point, we're going to emit the exit trace. It's just a
+ * question of whether or not we block while doing it. */
+ if (strace->drop_overflow || !sysc_can_block(trace->syscallno))
+ qiwrite(strace->q, trace->pretty_buf, pretty_len);
+ else
+ qwrite(strace->q, trace->pretty_buf, pretty_len);
+ }
+ if (systrace_loud)
+ printk("%s", trace->pretty_buf);
+ poperror();
+}
+
+static bool should_strace(struct proc *p, struct syscall *sysc)
+{
+ unsigned int sysc_num;
+
+ if (systrace_loud)
+ return TRUE;
+ if (!p->strace || !p->strace->tracing)
+ return FALSE;
+ /* TOCTTOU concerns - sysc is __user. */
+ sysc_num = ACCESS_ONCE(sysc->num);
+ if (qfull(p->strace->q)) {
+ if (p->strace->drop_overflow || !sysc_can_block(sysc_num)) {
+ atomic_inc(&p->strace->nr_drops);
+ return FALSE;
+ }
+ }
+ if (sysc_num > MAX_SYSCALL_NR)
+ return FALSE;
+ return test_bit(sysc_num, p->strace->trace_set);
+}
+
+/* Helper, copies len bytes from u_data to the trace->data, if there's room. */
+static void copy_tracedata_from_user(struct systrace_record *trace,
+ long u_data, size_t len)
+{
+ size_t copy_amt;
+
+ copy_amt = MIN(sizeof(trace->data) - trace->datalen, len);
+ copy_from_user(trace->data + trace->datalen, (void*)u_data, copy_amt);
+ trace->datalen += copy_amt;
+}
+
+/* Helper, snprintfs to the trace, if there's room. */
+static void snprintf_to_trace(struct systrace_record *trace, const char *fmt,
+ ...)
+{
+ va_list ap;
+ int rc;
+
+ va_start(ap, fmt);
+ rc = vsnprintf((char*)trace->data + trace->datalen,
+ sizeof(trace->data) - trace->datalen, fmt, ap);
+ va_end(ap);
+ if (!snprintf_error(rc, sizeof(trace->data) - trace->datalen))
+ trace->datalen += rc;
+}
+
+/* Starts a trace for p running sysc, attaching it to kthread. Pairs with
+ * systrace_finish_trace(). */
static void systrace_start_trace(struct kthread *kthread, struct syscall *sysc)
{
- struct systrace_record *trace;
- int coreid, vcoreid;
struct proc *p = current;
+ struct systrace_record *trace;
- if (!__trace_this_proc(p))
+ kthread->strace = 0;
+ if (!should_strace(p, sysc))
return;
- assert(!kthread->trace); /* catch memory leaks */
- coreid = core_id();
- vcoreid = proc_get_vcoreid(p);
- if (systrace_flags & SYSTRACE_LOUD) {
- printk("ENTER [%16llu] Syscall %3d (%12s):(0x%llx, 0x%llx, 0x%llx, "
- "0x%llx, 0x%llx, 0x%llx) proc: %d core: %d vcore: %d\n",
- read_tsc(),
- sysc->num, syscall_table[sysc->num].name,
- sysc->arg0, sysc->arg1, sysc->arg2, sysc->arg3, sysc->arg4,
- sysc->arg5, p->pid, coreid, vcoreid);
- }
- trace = kmalloc(SYSTR_BUF_SZ, 0);
- if (!trace)
- return;
- kthread->trace = trace;
+ /* TODO: consider a block_alloc and qpass, though note that we actually
+ * write the same trace in twice (entry and exit). */
+ trace = kpages_alloc(SYSTR_BUF_SZ, MEM_ATOMIC);
+ if (p->strace) {
+ if (!trace) {
+ atomic_inc(&p->strace->nr_drops);
+ return;
+ }
+ /* Avoiding the atomic op. We sacrifice accuracy for less overhead. */
+ p->strace->appx_nr_sysc++;
+ } else {
+ if (!trace)
+ return;
+ }
+ /* if you ever need to debug just one strace function, this is
+ * handy way to do it: just bail out if it's not the one you
+ * want.
+ * if (sysc->num != SYS_exec)
+ * return; */
trace->start_timestamp = read_tsc();
+ trace->end_timestamp = 0;
trace->syscallno = sysc->num;
trace->arg0 = sysc->arg0;
trace->arg1 = sysc->arg1;
trace->arg3 = sysc->arg3;
trace->arg4 = sysc->arg4;
trace->arg5 = sysc->arg5;
+ trace->retval = 0;
trace->pid = p->pid;
- trace->coreid = coreid;
- trace->vcoreid = vcoreid;
+ trace->coreid = core_id();
+ trace->vcoreid = proc_get_vcoreid(p);
trace->pretty_buf = (char*)trace + sizeof(struct systrace_record);
trace->datalen = 0;
trace->data[0] = 0;
-}
+ switch (sysc->num) {
+ case SYS_write:
+ copy_tracedata_from_user(trace, sysc->arg1, sysc->arg2);
+ break;
+ case SYS_openat:
+ case SYS_chdir:
+ case SYS_rmdir:
+ case SYS_nmount:
+ copy_tracedata_from_user(trace, sysc->arg1, sysc->arg2);
+ break;
+ case SYS_stat:
+ case SYS_lstat:
+ case SYS_access:
+ case SYS_unlink:
+ case SYS_mkdir:
+ case SYS_wstat:
+ copy_tracedata_from_user(trace, sysc->arg0, sysc->arg1);
+ break;
+ case SYS_link:
+ case SYS_symlink:
+ case SYS_rename:
+ case SYS_nbind:
+ copy_tracedata_from_user(trace, sysc->arg0, sysc->arg1);
+ snprintf_to_trace(trace, " -> ");
+ copy_tracedata_from_user(trace, sysc->arg2, sysc->arg3);
+ break;
+ case SYS_nunmount:
+ copy_tracedata_from_user(trace, sysc->arg2, sysc->arg3);
+ break;
+ case SYS_exec:
+ trace->datalen = execargs_stringer(current,
+ (char *)trace->data,
+ sizeof(trace->data),
+ (char *)sysc->arg0,
+ sysc->arg1,
+ (char *)sysc->arg2,
+ sysc->arg3);
+ break;
+ case SYS_proc_create:
+ trace->datalen = execargs_stringer(current,
+ (char *)trace->data,
+ sizeof(trace->data),
+ (char *)sysc->arg0,
+ sysc->arg1,
+ (char *)sysc->arg2,
+ sysc->arg3);
+ break;
+ }
+ systrace_output(trace, p->strace, TRUE);
+
+ kthread->strace = trace;
+}
+
+/* Finishes the trace on kthread for p, with retval being the return from the
+ * syscall we're tracing. Pairs with systrace_start_trace(). */
static void systrace_finish_trace(struct kthread *kthread, long retval)
{
- struct systrace_record *trace = kthread->trace;
- size_t pretty_len;
- if (trace) {
- trace->end_timestamp = read_tsc();
- trace->retval = retval;
- kthread->trace = 0;
- pretty_len = systrace_fill_pretty_buf(trace);
- kprof_tracedata_write(trace->pretty_buf, pretty_len);
- if (systrace_flags & SYSTRACE_LOUD)
- printk("EXIT %s", trace->pretty_buf);
- kfree(trace);
+ struct proc *p = current;
+ struct systrace_record *trace;
+
+ if (!kthread->strace)
+ return;
+ trace = kthread->strace;
+ trace->end_timestamp = read_tsc();
+ trace->retval = retval;
+
+ /* Only try to do the trace data if we didn't do it on entry */
+ if (!trace->datalen) {
+ switch (trace->syscallno) {
+ case SYS_read:
+ if (retval <= 0)
+ break;
+ copy_tracedata_from_user(trace, trace->arg1, retval);
+ break;
+ case SYS_readlink:
+ if (retval <= 0)
+ break;
+ copy_tracedata_from_user(trace, trace->arg0, trace->arg1);
+ snprintf_to_trace(trace, " -> ");
+ copy_tracedata_from_user(trace, trace->arg2, trace->arg3);
+ break;
+ }
}
+
+ systrace_output(trace, p->strace, FALSE);
+ kpages_free(kthread->strace, SYSTR_BUF_SZ);
+ kthread->strace = 0;
}
#ifdef CONFIG_SYSCALL_STRING_SAVING
static void alloc_sysc_str(struct kthread *kth)
{
- kth->name = kmalloc(SYSCALL_STRLEN, KMALLOC_WAIT);
+ kth->name = kmalloc(SYSCALL_STRLEN, MEM_ATOMIC);
+ if (!kth->name)
+ return;
kth->name[0] = 0;
}
static void free_sysc_str(struct kthread *kth)
{
char *str = kth->name;
+
kth->name = 0;
kfree(str);
}
#define sysc_save_str(...) \
{ \
struct per_cpu_info *pcpui = &per_cpu_info[core_id()]; \
- snprintf(pcpui->cur_kthread->name, SYSCALL_STRLEN, __VA_ARGS__); \
+ \
+ if (pcpui->cur_kthread->name) \
+ snprintf(pcpui->cur_kthread->name, SYSCALL_STRLEN, __VA_ARGS__); \
}
#else
* to not muck with the flags while we're signalling. */
atomic_or(&sysc->flags, SC_K_LOCK | SC_DONE);
__signal_syscall(sysc, p);
- atomic_and(&sysc->flags, ~SC_K_LOCK);
+ atomic_and(&sysc->flags, ~SC_K_LOCK);
}
/* Helper that "finishes" the current async syscall. This should be used with
{
va_list ap;
+ assert(fmt);
va_start(ap, fmt);
vset_errstr(fmt, ap);
va_end(ap);
set_errno(error);
+ assert(fmt);
va_start(ap, fmt);
- vset_errstr(fmt != NULL ? fmt: errno_to_string(error), ap);
+ vset_errstr(fmt, ap);
va_end(ap);
}
/* Diagnostic function: blocks the kthread/syscall, to help userspace test its
* async I/O handling. */
-static int sys_block(struct proc *p, unsigned int usec)
+static int sys_block(struct proc *p, unsigned long usec)
{
- /* Note printing takes a few ms, so your printds won't be perfect. */
- printd("[kernel] sys_block(), sleeping at %llu\n", read_tsc());
+ sysc_save_str("block for %lu usec", usec);
kthread_usleep(usec);
- printd("[kernel] sys_block(), waking up at %llu\n", read_tsc());
return 0;
}
-// Writes 'val' to 'num_writes' entries of the well-known array in the kernel
-// address space. It's just #defined to be some random 4MB chunk (which ought
-// to be boot_alloced or something). Meant to grab exclusive access to cache
-// lines, to simulate doing something useful.
-static int sys_cache_buster(struct proc *p, uint32_t num_writes,
- uint32_t num_pages, uint32_t flags)
-{
- #define BUSTER_ADDR 0xd0000000L // around 512 MB deep
- #define MAX_WRITES 1048576*8
- #define MAX_PAGES 32
- #define INSERT_ADDR (UINFO + 2*PGSIZE) // should be free for these tests
- uint32_t* buster = (uint32_t*)BUSTER_ADDR;
- static spinlock_t buster_lock = SPINLOCK_INITIALIZER;
- uint64_t ticks = -1;
- page_t* a_page[MAX_PAGES];
-
- /* Strided Accesses or Not (adjust to step by cachelines) */
- uint32_t stride = 1;
- if (flags & BUSTER_STRIDED) {
- stride = 16;
- num_writes *= 16;
- }
-
- /* Shared Accesses or Not (adjust to use per-core regions)
- * Careful, since this gives 8MB to each core, starting around 512MB.
- * Also, doesn't separate memory for core 0 if it's an async call.
- */
- if (!(flags & BUSTER_SHARED))
- buster = (uint32_t*)(BUSTER_ADDR + core_id() * 0x00800000);
-
- /* Start the timer, if we're asked to print this info*/
- if (flags & BUSTER_PRINT_TICKS)
- ticks = start_timing();
+/* Pause execution for a number of nanoseconds.
+ * The current implementation rounds up to the nearest microsecond. If the
+ * syscall is aborted, we return the remaining time the call would have ran
+ * in the 'rem' parameter. */
+static int sys_nanosleep(struct proc *p,
+ const struct timespec *req,
+ struct timespec *rem)
+{
+ ERRSTACK(1);
+ uint64_t usec;
+ struct timespec kreq, krem = {0, 0};
+ uint64_t tsc = read_tsc();
- /* Allocate num_pages (up to MAX_PAGES), to simulate doing some more
- * realistic work. Note we don't write to these pages, even if we pick
- * unshared. Mostly due to the inconvenience of having to match up the
- * number of pages with the number of writes. And it's unnecessary.
- */
- if (num_pages) {
- spin_lock(&buster_lock);
- for (int i = 0; i < MIN(num_pages, MAX_PAGES); i++) {
- upage_alloc(p, &a_page[i],1);
- page_insert(p->env_pgdir, a_page[i], (void*)INSERT_ADDR + PGSIZE*i,
- PTE_USER_RW);
- page_decref(a_page[i]);
- }
- spin_unlock(&buster_lock);
+ /* Check the input arguments. */
+ if (memcpy_from_user(p, &kreq, req, sizeof(struct timespec))) {
+ set_errno(EFAULT);
+ return -1;
}
-
- if (flags & BUSTER_LOCKED)
- spin_lock(&buster_lock);
- for (int i = 0; i < MIN(num_writes, MAX_WRITES); i=i+stride)
- buster[i] = 0xdeadbeef;
- if (flags & BUSTER_LOCKED)
- spin_unlock(&buster_lock);
-
- if (num_pages) {
- spin_lock(&buster_lock);
- for (int i = 0; i < MIN(num_pages, MAX_PAGES); i++) {
- page_remove(p->env_pgdir, (void*)(INSERT_ADDR + PGSIZE * i));
- page_decref(a_page[i]);
- }
- spin_unlock(&buster_lock);
+ if (rem && memcpy_to_user(p, rem, &krem, sizeof(struct timespec))) {
+ set_errno(EFAULT);
+ return -1;
+ }
+ if (kreq.tv_sec < 0) {
+ set_errno(EINVAL);
+ return -1;
+ }
+ if ((kreq.tv_nsec < 0) || (kreq.tv_nsec > 999999999)) {
+ set_errno(EINVAL);
+ return -1;
}
- /* Print info */
- if (flags & BUSTER_PRINT_TICKS) {
- ticks = stop_timing(ticks);
- printk("%llu,", ticks);
+ /* Convert timespec to usec. Ignore overflow on the tv_sec field. */
+ usec = kreq.tv_sec * 1000000;
+ usec += DIV_ROUND_UP(kreq.tv_nsec, 1000);
+
+ /* Attempt to sleep. If we get aborted, copy the remaining time into
+ * 'rem' and return. We assume the tsc is sufficient to tell how much
+ * time is remaining (i.e. it only overflows on the order of hundreds of
+ * years, which should be sufficiently long enough to ensure we don't
+ * overflow). */
+ if (waserror()) {
+ krem = tsc2timespec(read_tsc() - tsc);
+ if (rem && memcpy_to_user(p, rem, &krem, sizeof(struct timespec)))
+ set_errno(EFAULT);
+ poperror();
+ return -1;
}
+ sysc_save_str("nanosleep for %d usec", usec);
+ kthread_usleep(usec);
+ poperror();
return 0;
}
/* sys_reboot(): called directly from dispatch table. */
-/* Print a string to the system console. */
-static ssize_t sys_cputs(struct proc *p, const char *string,
- size_t strlen)
-{
- char *t_string;
- t_string = user_strdup_errno(p, string, strlen);
- if (!t_string)
- return -1;
- printk("%.*s", strlen, t_string);
- user_memdup_free(p, t_string);
- return (ssize_t)strlen;
-}
-
-// Read a character from the system console.
-// Returns the character.
-/* TODO: remove me */
-static uint16_t sys_cgetc(struct proc *p)
-{
- uint16_t c;
-
- // The cons_get_any_char() primitive doesn't wait for a character,
- // but the sys_cgetc() system call does.
- while ((c = cons_get_any_char()) == 0)
- cpu_relax();
-
- return c;
-}
-
/* Returns the id of the physical core this syscall is executed on. */
static uint32_t sys_getpcoreid(void)
{
/************** Process management syscalls **************/
-/* Returns the calling process's pid */
-static pid_t sys_getpid(struct proc *p)
+/* Helper for proc_create and fork */
+static void inherit_strace(struct proc *parent, struct proc *child)
{
- return p->pid;
+ if (parent->strace && parent->strace->inherit) {
+ /* Refcnt on both, put in the child's ->strace. */
+ kref_get(&parent->strace->users, 1);
+ kref_get(&parent->strace->procs, 1);
+ child->strace = parent->strace;
+ }
}
/* Creates a process from the file 'path'. The process is not runnable by
/* TODO: 9ns support */
program = do_file_open(t_path, O_READ, 0);
if (!program)
- goto error_user_memdup;
-
+ goto error_with_path;
+ if (!is_valid_elf(program)) {
+ set_errno(ENOEXEC);
+ goto error_with_file;
+ }
/* Check the size of the argenv array, error out if too large. */
if ((argenv_l < sizeof(struct argenv)) || (argenv_l > ARG_MAX)) {
set_error(EINVAL, "The argenv array has an invalid size: %lu\n",
argenv_l);
- goto error_user_memdup;
+ goto error_with_file;
}
/* Copy the argenv array into a kernel buffer. Delay processing of the
* array to load_elf(). */
kargenv = user_memdup_errno(p, argenv, argenv_l);
if (!kargenv) {
- set_errstr("Failed to copy in the args");
- goto error_user_memdup;
+ set_error(EINVAL, "Failed to copy in the args");
+ goto error_with_file;
}
/* Unpack the argenv array into more usable variables. Integrity checking
* done along side this as well. */
if (unpack_argenv(kargenv, argenv_l, &argc, &argv, &envc, &envp)) {
- set_errstr("Failed to unpack the args");
- goto error_unpack;
+ set_error(EINVAL, "Failed to unpack the args");
+ goto error_with_kargenv;
}
-
/* TODO: need to split the proc creation, since you must load after setting
* args/env, since auxp gets set up there. */
//new_p = proc_create(program, 0, 0);
if (proc_alloc(&new_p, current, flags)) {
- set_errstr("Failed to alloc new proc");
- goto error_proc_alloc;
+ set_error(ENOMEM, "Failed to alloc new proc");
+ goto error_with_kargenv;
}
+ inherit_strace(p, new_p);
/* close the CLOEXEC ones, even though this isn't really an exec */
close_fdt(&new_p->open_files, TRUE);
/* Load the elf. */
if (load_elf(new_p, program, argc, argv, envc, envp)) {
- set_errstr("Failed to load elf");
- goto error_load_elf;
+ set_error(EINVAL, "Failed to load elf");
+ goto error_with_proc;
}
/* progname is argv0, which accounts for symlinks */
proc_set_progname(new_p, argc ? argv[0] : NULL);
profiler_notify_new_process(new_p);
proc_decref(new_p); /* give up the reference created in proc_create() */
return pid;
-error_load_elf:
- set_errno(EINVAL);
+error_with_proc:
/* proc_destroy will decref once, which is for the ref created in
* proc_create(). We don't decref again (the usual "+1 for existing"),
* since the scheduler, which usually handles that, hasn't heard about the
* process (via __proc_ready()). */
proc_destroy(new_p);
-error_proc_alloc:
- kref_put(&program->f_kref);
-error_unpack:
+error_with_kargenv:
user_memdup_free(p, kargenv);
-error_user_memdup:
+error_with_file:
+ kref_put(&program->f_kref);
+error_with_path:
free_path(p, t_path);
return -1;
}
printd("[%d] destroying proc %d\n", p->pid, p_to_die->pid);
}
proc_destroy(p_to_die);
- /* we only get here if we weren't the one to die */
proc_decref(p_to_die);
return 0;
}
static ssize_t sys_fork(env_t* e)
{
- struct proc *temp;
- int8_t state = 0;
+ uintptr_t temp;
int ret;
+ struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
// TODO: right now we only support fork for single-core processes
if (e->state != PROC_RUNNING_S) {
assert(env != NULL);
proc_set_progname(env, e->progname);
- disable_irqsave(&state); /* protect cur_ctx */
/* Can't really fork if we don't have a current_ctx to fork */
if (!current_ctx) {
- enable_irqsave(&state);
proc_destroy(env);
proc_decref(env);
set_errno(EINVAL);
return -1;
}
- env->scp_ctx = *current_ctx;
- enable_irqsave(&state);
-
- env->cache_colors_map = cache_colors_map_alloc();
- for (int i = 0; i < llc_cache->num_colors; i++)
- if (GET_BITMASK_BIT(e->cache_colors_map,i))
- cache_color_alloc(llc_cache, env->cache_colors_map);
+ assert(pcpui->cur_proc == pcpui->owning_proc);
+ copy_current_ctx_to(&env->scp_ctx);
/* Make the new process have the same VMRs as the older. This will copy the
* contents of non MAP_SHARED pages to the new VMRs. */
switch_back(env, temp);
/* Copy some state from the original proc into the new proc. */
- env->heap_top = e->heap_top;
env->env_flags = e->env_flags;
+ inherit_strace(e, env);
+
/* In general, a forked process should be a fresh process, and we copy over
* whatever stuff is needed between procinfo/procdata. */
*env->procdata = *e->procdata;
- env->procinfo->heap_bottom = e->procinfo->heap_bottom;
+ env->procinfo->program_end = e->procinfo->program_end;
/* FYI: once we call ready, the proc is open for concurrent usage */
__proc_ready(env);
return ret;
}
+/* string for sys_exec arguments. Assumes that d is pointing to zero'd
+ * storage or storage that does not require null termination or
+ * provides the null. */
+static int execargs_stringer(struct proc *p, char *d, size_t slen,
+ char *path, size_t path_l,
+ char *argenv, size_t argenv_l)
+{
+ int argc, envc, i;
+ char **argv, **envp;
+ struct argenv *kargenv;
+ int amt;
+ char *s = d;
+ char *e = d + slen;
+
+ if (path_l > slen)
+ path_l = slen;
+ if (memcpy_from_user(p, d, path, path_l)) {
+ s = seprintf(s, e, "Invalid exec path");
+ return s - d;
+ }
+ s += path_l;
+
+ /* yes, this code is cloned from below. I wrote a helper but
+ * Barret and I concluded after talking about it that the
+ * helper was not really helper-ful, as it has almost 10
+ * arguments. Please, don't suggest a cpp macro. Thank you. */
+ /* Check the size of the argenv array, error out if too large. */
+ if ((argenv_l < sizeof(struct argenv)) || (argenv_l > ARG_MAX)) {
+ s = seprintf(s, e, "The argenv array has an invalid size: %lu\n",
+ argenv_l);
+ return s - d;
+ }
+ /* Copy the argenv array into a kernel buffer. */
+ kargenv = user_memdup_errno(p, argenv, argenv_l);
+ if (!kargenv) {
+ s = seprintf(s, e, "Failed to copy in the args and environment");
+ return s - d;
+ }
+ /* Unpack the argenv array into more usable variables. Integrity checking
+ * done along side this as well. */
+ if (unpack_argenv(kargenv, argenv_l, &argc, &argv, &envc, &envp)) {
+ s = seprintf(s, e, "Failed to unpack the args");
+ user_memdup_free(p, kargenv);
+ return s - d;
+ }
+ s = seprintf(s, e, "[%d]{", argc);
+ for (i = 0; i < argc; i++)
+ s = seprintf(s, e, "%s, ", argv[i]);
+ s = seprintf(s, e, "}");
+
+ user_memdup_free(p, kargenv);
+ return s - d;
+}
+
/* Load the binary "path" into the current process, and start executing it.
* argv and envp are magically bundled in procinfo for now. Keep in sync with
* glibc's sysdeps/ros/execve.c. Once past a certain point, this function won't
* return. It assumes (and checks) that it is current. Don't give it an extra
- * refcnt'd *p (syscall won't do that).
+ * refcnt'd *p (syscall won't do that).
* Note: if someone batched syscalls with this call, they could clobber their
* old memory (and will likely PF and die). Don't do it... */
static int sys_exec(struct proc *p, char *path, size_t path_l,
char *t_path = NULL;
struct file *program;
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- int8_t state = 0;
int argc, envc;
char **argv, **envp;
struct argenv *kargenv;
return -1;
}
- disable_irqsave(&state); /* protect cur_ctx */
/* Can't exec if we don't have a current_ctx to restart (if we fail). This
* isn't 100% true, but I'm okay with it. */
if (!pcpui->cur_ctx) {
- enable_irqsave(&state);
set_errno(EINVAL);
return -1;
}
/* Preemptively copy out the cur_ctx, in case we fail later (easier on
* cur_ctx if we do this now) */
- p->scp_ctx = *pcpui->cur_ctx;
- /* Clear the current_ctx. We won't be returning the 'normal' way. Even if
- * we want to return with an error, we need to go back differently in case
- * we succeed. This needs to be done before we could possibly block, but
- * unfortunately happens before the point of no return.
- *
- * Note that we will 'hard block' if we block at all. We can't return to
- * userspace and then asynchronously finish the exec later. */
- clear_owning_proc(core_id());
- enable_irqsave(&state);
-
+ assert(pcpui->cur_proc == pcpui->owning_proc);
+ copy_current_ctx_to(&p->scp_ctx);
/* Check the size of the argenv array, error out if too large. */
if ((argenv_l < sizeof(struct argenv)) || (argenv_l > ARG_MAX)) {
set_error(EINVAL, "The argenv array has an invalid size: %lu\n",
/* This could block: */
/* TODO: 9ns support */
program = do_file_open(t_path, O_READ, 0);
+ /* Clear the current_ctx. We won't be returning the 'normal' way. Even if
+ * we want to return with an error, we need to go back differently in case
+ * we succeed. This needs to be done before we could possibly block, but
+ * unfortunately happens before the point of no return.
+ *
+ * Note that we will 'hard block' if we block at all. We can't return to
+ * userspace and then asynchronously finish the exec later. */
+ clear_owning_proc(core_id());
if (!program)
goto early_error;
if (!is_valid_elf(program)) {
proc_replace_binary_path(p, t_path);
proc_set_progname(p, argc ? argv[0] : NULL);
proc_init_procdata(p);
- p->procinfo->heap_bottom = 0;
+ p->procinfo->program_end = 0;
/* When we destroy our memory regions, accessing cur_sysc would PF */
pcpui->cur_kthread->sysc = 0;
unmap_and_destroy_vmrs(p);
static pid_t try_wait(struct proc *parent, struct proc *child, int *ret_status,
int options)
{
- if (child->state == PROC_DYING) {
+ if (proc_is_dying(child)) {
/* Disown returns -1 if it's already been disowned or we should o/w
* abort. This can happen if we have concurrent waiters, both with
* pointers to the child (only one should reap). Note that if we don't
/* If we're dying, then we don't need to worry about waiting. We don't
* do this yet, but we'll need this outlet when we deal with orphaned
* children and having init inherit them. */
- if (parent->state == PROC_DYING)
+ if (proc_is_dying(parent))
goto out_unlock;
/* Any child can wake us up, but we check for the particular child we
* care about */
while (!retval) {
cpu_relax();
cv_wait(&parent->child_wait);
- if (parent->state == PROC_DYING)
+ if (proc_is_dying(parent))
goto out_unlock;
/* Any child can wake us up from the CV. This is a linear try_wait
* scan. If we have a lot of children, we could optimize this. */
pid_t retval = 0;
int ret_status = 0;
+ sysc_save_str("waitpid on %d", pid);
/* -1 is the signal for 'any child' */
if (pid == -1) {
retval = wait_any(parent, &ret_status, options);
return 0;
}
+static int sys_send_event(struct proc *p, struct event_queue *ev_q,
+ struct event_msg *u_msg, uint32_t vcoreid)
+{
+ struct event_msg local_msg = {0};
+
+ if (memcpy_from_user(p, &local_msg, u_msg, sizeof(struct event_msg))) {
+ set_errno(EINVAL);
+ return -1;
+ }
+ send_event(p, ev_q, &local_msg, vcoreid);
+ return 0;
+}
+
/* Puts the calling core into vcore context, if it wasn't already, via a
* self-IPI / active notification. Barring any weird unmappings, we just send
* ourselves a __notify. */
* is trying to halt. The core need not abort the halt for notif_pending for
* the vcore, only for a __notify or other RKM. Anyone setting notif_pending
* should then attempt to __notify (o/w it's probably a bug). */
-static int sys_halt_core(struct proc *p, unsigned int usec)
+static int sys_halt_core(struct proc *p, unsigned long usec)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
struct preempt_data *vcpd;
return retval;
}
-/* Initializes a process to run virtual machine contexts, returning the number
- * initialized, optionally setting errno */
-static int sys_setup_vmm(struct proc *p, unsigned int nr_guest_pcores,
- int flags)
+/* Assists the user/2LS by atomically running *ctx and leaving vcore context.
+ * Normally, the user can do this themselves, but x86 VM contexts need kernel
+ * support. The caller ought to be in vcore context, and if a notif is pending,
+ * then the calling vcore will restart in a fresh VC ctx (as if it was notified
+ * or did a sys_vc_entry).
+ *
+ * Note that this will set the TLS too, which is part of the context. Parlib's
+ * pop_user_ctx currently does *not* do this, since the TLS is managed
+ * separately. If you want to use this syscall for testing, you'll need to 0
+ * out fsbase and conditionally write_msr in proc_pop_ctx(). */
+static int sys_pop_ctx(struct proc *p, struct user_context *ctx)
+{
+ int pcoreid = core_id();
+ struct per_cpu_info *pcpui = &per_cpu_info[pcoreid];
+ int vcoreid = pcpui->owning_vcoreid;
+ struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[vcoreid];
+
+ /* With change_to, there's a bunch of concerns about changing the vcore map,
+ * since the kernel may have already locked and sent preempts, deaths, etc.
+ *
+ * In this case, we don't care as much. Other than notif_pending and
+ * notif_disabled, it's more like we're just changing a few registers in
+ * cur_ctx. We can safely order-after any kernel messages or other changes,
+ * as if the user had done all of the changes we'll make and then did a
+ * no-op syscall.
+ *
+ * Since we are mucking with current_ctx, it is important that we don't
+ * block before or during this syscall. */
+ arch_finalize_ctx(pcpui->cur_ctx);
+ if (copy_from_user(pcpui->cur_ctx, ctx, sizeof(struct user_context))) {
+ /* The 2LS isn't really in a position to handle errors. At the very
+ * least, we can print something and give them a fresh vc ctx. */
+ printk("[kernel] unable to copy user_ctx, 2LS bug\n");
+ memset(pcpui->cur_ctx, 0, sizeof(struct user_context));
+ proc_init_ctx(pcpui->cur_ctx, vcoreid, vcpd->vcore_entry,
+ vcpd->vcore_stack, vcpd->vcore_tls_desc);
+ return -1;
+ }
+ proc_secure_ctx(pcpui->cur_ctx);
+ /* The caller leaves vcore context no matter what. We'll put them back in
+ * if they missed a message. */
+ vcpd->notif_disabled = FALSE;
+ wrmb(); /* order disabled write before pending read */
+ if (vcpd->notif_pending)
+ send_kernel_message(pcoreid, __notify, (long)p, 0, 0, KMSG_ROUTINE);
+ return 0;
+}
+
+static int sys_vmm_add_gpcs(struct proc *p, unsigned int nr_more_gpcs,
+ struct vmm_gpcore_init *gpcis)
{
- return vmm_struct_init(p, nr_guest_pcores, flags);
+ ERRSTACK(1);
+ struct vmm *vmm = &p->vmm;
+
+ qlock(&vmm->qlock);
+ if (waserror()) {
+ qunlock(&vmm->qlock);
+ poperror();
+ return -1;
+ }
+ __vmm_struct_init(p);
+ __vmm_add_gpcs(p, nr_more_gpcs, gpcis);
+ qunlock(&vmm->qlock);
+ poperror();
+ return nr_more_gpcs;
+}
+
+static int sys_vmm_poke_guest(struct proc *p, int guest_pcoreid)
+{
+ return vmm_poke_guest(p, guest_pcoreid);
+}
+
+static int sys_vmm_ctl(struct proc *p, int cmd, unsigned long arg1,
+ unsigned long arg2, unsigned long arg3,
+ unsigned long arg4)
+{
+ ERRSTACK(1);
+ int ret;
+ struct vmm *vmm = &p->vmm;
+
+ /* Protects against concurrent setters and for gets that are not atomic
+ * reads (say, multiple exec ctls). */
+ qlock(&vmm->qlock);
+ if (waserror()) {
+ qunlock(&vmm->qlock);
+ poperror();
+ return -1;
+ }
+ __vmm_struct_init(p);
+ switch (cmd) {
+ case VMM_CTL_GET_EXITS:
+ if (vmm->amd)
+ error(ENOTSUP, "AMD VMMs unsupported");
+ ret = vmx_ctl_get_exits(&vmm->vmx);
+ break;
+ case VMM_CTL_SET_EXITS:
+ if (arg1 & ~VMM_CTL_ALL_EXITS)
+ error(EINVAL, "Bad vmm_ctl_exits %x (%x)", arg1,
+ VMM_CTL_ALL_EXITS);
+ if (vmm->amd)
+ error(ENOTSUP, "AMD VMMs unsupported");
+ ret = vmx_ctl_set_exits(&vmm->vmx, arg1);
+ break;
+ case VMM_CTL_GET_FLAGS:
+ ret = vmm->flags;
+ break;
+ case VMM_CTL_SET_FLAGS:
+ if (arg1 & ~VMM_CTL_ALL_FLAGS)
+ error(EINVAL, "Bad vmm_ctl flags. Got 0x%lx, allowed 0x%lx\n",
+ arg1, VMM_CTL_ALL_FLAGS);
+ vmm->flags = arg1;
+ ret = 0;
+ break;
+ default:
+ error(EINVAL, "Bad vmm_ctl cmd %d", cmd);
+ }
+ qunlock(&vmm->qlock);
+ poperror();
+ return ret;
}
/* Pokes the ksched for the given resource for target_pid. If the target pid
* == 0, we just poke for the calling process. The common case is poking for
- * self, so we avoid the lookup.
+ * self, so we avoid the lookup.
*
* Not sure if you could harm someone via asking the kernel to look at them, so
* we'll do a 'controls' check for now. In the future, we might have something
static intreg_t sys_read(struct proc *p, int fd, void *buf, size_t len)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- struct systrace_record *t = pcpui->cur_kthread->trace;
ssize_t ret;
struct file *file = get_file_from_fd(&p->open_files, fd);
sysc_save_str("read on fd %d", fd);
/* plan9, should also handle errors (EBADF) */
ret = sysread(fd, buf, len);
}
-
- if ((ret > 0) && t) {
- t->datalen = MIN(sizeof(t->data), ret);
- memcpy(t->data, buf, t->datalen);
- }
-
return ret;
}
static intreg_t sys_write(struct proc *p, int fd, const void *buf, size_t len)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- struct systrace_record *t = pcpui->cur_kthread->trace;
ssize_t ret;
struct file *file = get_file_from_fd(&p->open_files, fd);
+
sysc_save_str("write on fd %d", fd);
/* VFS */
if (file) {
/* plan9, should also handle errors */
ret = syswrite(fd, (void*)buf, len);
}
-
- if (t) {
- t->datalen = MIN(sizeof(t->data), len);
- memcpy(t->data, buf, t->datalen);
- }
return ret;
-
}
/* Checks args/reads in the path, opens the file (relative to fromfd if the path
}
/* 9ns, should also handle errors (bad FD, etc) */
retval = sysclose(fd);
- if (retval < 0) {
- /* no one checks their retvals. a double close will cause problems. */
- printk("[kernel] sys_close failed: proc %d fd %d. Check your rets.\n",
- p->pid, fd);
- }
return retval;
}
-/* kept around til we remove the last ufe */
-#define ufe(which,a0,a1,a2,a3) \
- frontend_syscall_errno(p,APPSERVER_SYSCALL_##which,\
- (int)(a0),(int)(a1),(int)(a2),(int)(a3))
-
static intreg_t sys_fstat(struct proc *p, int fd, struct kstat *u_stat)
{
struct kstat *kbuf;
ret = file->f_op->llseek(file, tempoff, &retoff, whence);
kref_put(&file->f_kref);
} else {
- /* won't return here if error ... */
- ret = sysseek(fd, tempoff, whence);
- retoff = ret;
- ret = 0;
+ retoff = sysseek(fd, tempoff, whence);
+ ret = (retoff < 0);
}
if (ret)
path_d = lookup_dentry(t_path, 0);
if (!path_d){
int n = 2048;
- buf = kmalloc(n*2, KMALLOC_WAIT);
+ buf = kmalloc(n*2, MEM_WAIT);
struct dir *d = (void *)&buf[n];
/* try 9ns. */
if (sysstat(t_path, buf, n) > 0) {
if (symname){
copy_amt = strnlen(symname, buf_l - 1) + 1;
- if (! memcpy_to_user_errno(p, u_buf, symname, copy_amt))
+ if (!memcpy_to_user_errno(p, u_buf, symname, copy_amt))
ret = copy_amt - 1;
}
if (path_d)
return retval;
}
-intreg_t sys_pipe(struct proc *p, int *u_pipefd, int flags)
-{
- int pipefd[2] = {0};
- int retval = syspipe(pipefd);
-
- if (retval)
- return -1;
- if (memcpy_to_user_errno(p, u_pipefd, pipefd, sizeof(pipefd))) {
- sysclose(pipefd[0]);
- sysclose(pipefd[1]);
- set_errno(EFAULT);
- return -1;
- }
- return 0;
-}
-
-intreg_t sys_gettimeofday(struct proc *p, int *buf)
-{
- static spinlock_t gtod_lock = SPINLOCK_INITIALIZER;
- static int t0 = 0;
-
- spin_lock(>od_lock);
- if(t0 == 0)
-
-#if (defined CONFIG_APPSERVER)
- t0 = ufe(time,0,0,0,0);
-#else
- // Nanwan's birthday, bitches!!
- t0 = 1242129600;
-#endif
- spin_unlock(>od_lock);
-
- long long dt = read_tsc();
- /* TODO: This probably wants its own function, using a struct timeval */
- long kbuf[2] = {t0+dt/system_timing.tsc_freq,
- (dt%system_timing.tsc_freq)*1000000/system_timing.tsc_freq};
-
- return memcpy_to_user_errno(p,buf,kbuf,sizeof(kbuf));
-}
-
intreg_t sys_tcgetattr(struct proc *p, int fd, void *termios_p)
{
int retval = 0;
int m_sz;
int retval = 0;
- dir = kzmalloc(sizeof(struct dir) + stat_sz, KMALLOC_WAIT);
+ dir = kzmalloc(sizeof(struct dir) + stat_sz, MEM_WAIT);
m_sz = convM2D(stat_m, stat_sz, &dir[0], (char*)&dir[1]);
if (m_sz != stat_sz) {
- set_error(EINVAL, NULL);
+ set_error(EINVAL, ERROR_FIXME);
kfree(dir);
return -1;
}
char *new_path, size_t new_path_l)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- struct systrace_record *t = pcpui->cur_kthread->trace;
ERRSTACK(1);
int mountpointlen = 0;
char *from_path = copy_in_path(p, old_path, old_path_l);
if ((!from_path) || (!to_path))
return -1;
printd("sys_rename :%s: to :%s: : ", from_path, to_path);
- if (t) {
- t->datalen = snprintf((char *)t->data, sizeof(t->data), "Rename :%s: to :%s:", from_path, to_path);
- }
/* we need a fid for the wstat. */
/* TODO: maybe wrap the 9ns stuff better. sysrename maybe? */
printd("Newchan: %C\n", newchan);
printd("Newchan: mchan %C\n", newchan->mchan);
- if ((newchan->dev != oldchan->dev) ||
+ if ((newchan->dev != oldchan->dev) ||
(newchan->type != oldchan->type)) {
printd("Old chan and new chan do not match\n");
set_errno(ENODEV);
}
mlen = convD2M(&dir, mbuf, sizeof(mbuf));
- if (! mlen) {
+ if (!mlen) {
printk("convD2M failed\n");
set_errno(EINVAL);
goto done;
};
printk("syswstat returns %d\n", retval);
-done:
+done:
free_path(p, from_path);
free_path(p, to_path);
cclose(oldchan);
const struct sys_table_entry syscall_table[] = {
[SYS_null] = {(syscall_t)sys_null, "null"},
[SYS_block] = {(syscall_t)sys_block, "block"},
- [SYS_cache_buster] = {(syscall_t)sys_cache_buster, "buster"},
[SYS_cache_invalidate] = {(syscall_t)sys_cache_invalidate, "wbinv"},
[SYS_reboot] = {(syscall_t)reboot, "reboot!"},
- [SYS_cputs] = {(syscall_t)sys_cputs, "cputs"},
- [SYS_cgetc] = {(syscall_t)sys_cgetc, "cgetc"},
[SYS_getpcoreid] = {(syscall_t)sys_getpcoreid, "getpcoreid"},
[SYS_getvcoreid] = {(syscall_t)sys_getvcoreid, "getvcoreid"},
- [SYS_getpid] = {(syscall_t)sys_getpid, "getpid"},
[SYS_proc_create] = {(syscall_t)sys_proc_create, "proc_create"},
[SYS_proc_run] = {(syscall_t)sys_proc_run, "proc_run"},
[SYS_proc_destroy] = {(syscall_t)sys_proc_destroy, "proc_destroy"},
- [SYS_yield] = {(syscall_t)sys_proc_yield, "proc_yield"},
+ [SYS_proc_yield] = {(syscall_t)sys_proc_yield, "proc_yield"},
[SYS_change_vcore] = {(syscall_t)sys_change_vcore, "change_vcore"},
[SYS_fork] = {(syscall_t)sys_fork, "fork"},
[SYS_exec] = {(syscall_t)sys_exec, "exec"},
[SYS_provision] = {(syscall_t)sys_provision, "provision"},
[SYS_notify] = {(syscall_t)sys_notify, "notify"},
[SYS_self_notify] = {(syscall_t)sys_self_notify, "self_notify"},
+ [SYS_send_event] = {(syscall_t)sys_send_event, "send_event"},
[SYS_vc_entry] = {(syscall_t)sys_vc_entry, "vc_entry"},
[SYS_halt_core] = {(syscall_t)sys_halt_core, "halt_core"},
#ifdef CONFIG_ARSC_SERVER
[SYS_init_arsc] = {(syscall_t)sys_init_arsc, "init_arsc"},
#endif
[SYS_change_to_m] = {(syscall_t)sys_change_to_m, "change_to_m"},
- [SYS_setup_vmm] = {(syscall_t)sys_setup_vmm, "setup_vmm"},
+ [SYS_vmm_add_gpcs] = {(syscall_t)sys_vmm_add_gpcs, "vmm_add_gpcs"},
+ [SYS_vmm_poke_guest] = {(syscall_t)sys_vmm_poke_guest, "vmm_poke_guest"},
+ [SYS_vmm_ctl] = {(syscall_t)sys_vmm_ctl, "vmm_ctl"},
[SYS_poke_ksched] = {(syscall_t)sys_poke_ksched, "poke_ksched"},
[SYS_abort_sysc] = {(syscall_t)sys_abort_sysc, "abort_sysc"},
[SYS_abort_sysc_fd] = {(syscall_t)sys_abort_sysc_fd, "abort_sysc_fd"},
[SYS_populate_va] = {(syscall_t)sys_populate_va, "populate_va"},
+ [SYS_nanosleep] = {(syscall_t)sys_nanosleep, "nanosleep"},
+ [SYS_pop_ctx] = {(syscall_t)sys_pop_ctx, "pop_ctx"},
[SYS_read] = {(syscall_t)sys_read, "read"},
[SYS_write] = {(syscall_t)sys_write, "write"},
[SYS_getcwd] = {(syscall_t)sys_getcwd, "getcwd"},
[SYS_mkdir] = {(syscall_t)sys_mkdir, "mkdir"},
[SYS_rmdir] = {(syscall_t)sys_rmdir, "rmdir"},
- [SYS_pipe] = {(syscall_t)sys_pipe, "pipe"},
- [SYS_gettimeofday] = {(syscall_t)sys_gettimeofday, "gettime"},
[SYS_tcgetattr] = {(syscall_t)sys_tcgetattr, "tcgetattr"},
[SYS_tcsetattr] = {(syscall_t)sys_tcsetattr, "tcsetattr"},
[SYS_setuid] = {(syscall_t)sys_setuid, "setuid"},
[SYS_tap_fds] = {(syscall_t)sys_tap_fds, "tap_fds"},
};
const int max_syscall = sizeof(syscall_table)/sizeof(syscall_table[0]);
+
/* Executes the given syscall.
*
* Note tf is passed in, which points to the tf of the context on the kernel
void run_local_syscall(struct syscall *sysc)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ struct proc *p = pcpui->cur_proc;
- assert(irq_is_enabled()); /* in case we proc destroy */
/* In lieu of pinning, we just check the sysc and will PF on the user addr
* later (if the addr was unmapped). Which is the plan for all UMEM. */
if (!is_user_rwaddr(sysc, sizeof(struct syscall))) {
}
pcpui->cur_kthread->sysc = sysc; /* let the core know which sysc it is */
systrace_start_trace(pcpui->cur_kthread, sysc);
+ pcpui = &per_cpu_info[core_id()]; /* reload again */
alloc_sysc_str(pcpui->cur_kthread);
/* syscall() does not return for exec and yield, so put any cleanup in there
* too. */
pcpui = &per_cpu_info[core_id()];
free_sysc_str(pcpui->cur_kthread);
systrace_finish_trace(pcpui->cur_kthread, sysc->retval);
+ pcpui = &per_cpu_info[core_id()]; /* reload again */
/* Some 9ns paths set errstr, but not errno. glibc will ignore errstr.
* this is somewhat hacky, since errno might get set unnecessarily */
if ((current_errstr()[0] != 0) && (!sysc->err))
* at least one, it will run it directly. */
void prep_syscalls(struct proc *p, struct syscall *sysc, unsigned int nr_syscs)
{
- int retval;
/* Careful with pcpui here, we could have migrated */
if (!nr_syscs) {
printk("[kernel] No nr_sysc, probably a bug, user!\n");
}
}
-/* Syscall tracing */
-static void __init_systrace(void)
-{
- systrace_buffer = kmalloc(MAX_SYSTRACES*sizeof(struct systrace_record), 0);
- if (!systrace_buffer)
- panic("Unable to alloc a trace buffer\n");
- systrace_bufidx = 0;
- systrace_bufsize = MAX_SYSTRACES;
- /* Note we never free the buffer - it's around forever. Feel free to change
- * this if you want to change the size or something dynamically. */
-}
-
-/* If you call this while it is running, it will change the mode */
-void systrace_start(bool silent)
-{
- static bool init = FALSE;
- spin_lock_irqsave(&systrace_lock);
- if (!init) {
- __init_systrace();
- init = TRUE;
- }
- systrace_flags = silent ? SYSTRACE_ON : SYSTRACE_ON | SYSTRACE_LOUD;
- spin_unlock_irqsave(&systrace_lock);
-}
-
-int systrace_reg(bool all, struct proc *p)
-{
- spin_lock_irqsave(&systrace_lock);
- if (all) {
- printk("Tracing syscalls for all processes\n");
- systrace_flags |= SYSTRACE_ALLPROC;
- } else {
- set_traced_proc(p, TRUE);
-
- printk("Tracing syscalls for process %d\n", p->pid);
- }
- spin_unlock_irqsave(&systrace_lock);
- return 0;
-}
-
-int systrace_trace_pid(struct proc *p)
-{
- if (systrace_reg(false, p))
- error(EFAIL, "no more processes");
- systrace_start(true);
- return 0;
-}
-
-void systrace_stop(void)
-{
- spin_lock_irqsave(&systrace_lock);
- systrace_flags = 0;
- spin_unlock_irqsave(&systrace_lock);
-}
-
-/* If you registered a process specifically, then you need to dereg it
- * specifically. Or just fully stop, which will do it for all. */
-int systrace_dereg(bool all, struct proc *p)
-{
- spin_lock_irqsave(&systrace_lock);
- if (all) {
- printk("No longer tracing syscalls for all processes.\n");
- systrace_flags &= ~SYSTRACE_ALLPROC;
- } else {
- set_traced_proc(p, FALSE);
-
- printk("No longer tracing syscalls for process %d\n", p->pid);
- }
- spin_unlock_irqsave(&systrace_lock);
- return 0;
-}
-
-/* Regardless of locking, someone could be writing into the buffer */
-void systrace_print(bool all, struct proc *p)
-{
- spin_lock_irqsave(&systrace_lock);
- /* if you want to be clever, you could make this start from the earliest
- * timestamp and loop around. Careful of concurrent writes. */
- for (int i = 0; i < systrace_bufsize; i++)
- if (systrace_buffer[i].start_timestamp)
- printk("[%16llu] Syscall %3d (%12s):(%p, %p, %p, %p, %p,"
- "%p) proc: %d core: %d vcore: %d\n",
- systrace_buffer[i].start_timestamp,
- systrace_buffer[i].syscallno,
- syscall_table[systrace_buffer[i].syscallno].name,
- systrace_buffer[i].arg0,
- systrace_buffer[i].arg1,
- systrace_buffer[i].arg2,
- systrace_buffer[i].arg3,
- systrace_buffer[i].arg4,
- systrace_buffer[i].arg5,
- systrace_buffer[i].pid,
- systrace_buffer[i].coreid,
- systrace_buffer[i].vcoreid);
- spin_unlock_irqsave(&systrace_lock);
-}
-
-void systrace_clear_buffer(void)
-{
- spin_lock_irqsave(&systrace_lock);
- memset(systrace_buffer, 0, sizeof(struct systrace_record) * MAX_SYSTRACES);
- spin_unlock_irqsave(&systrace_lock);
-}
-
bool syscall_uses_fd(struct syscall *sysc, int fd)
{
switch (sysc->num) {
void print_sysc(struct proc *p, struct syscall *sysc)
{
- struct proc *old_p = switch_to(p);
+ uintptr_t old_p = switch_to(p);
printk("SYS_%d, flags %p, a0 %p, a1 %p, a2 %p, a3 %p, a4 %p, a5 %p\n",
sysc->num, atomic_read(&sysc->flags),
sysc->arg0, sysc->arg1, sysc->arg2, sysc->arg3, sysc->arg4,
projects / akaros.git / blobdiff