#include <pmap.h>
#include <smp.h>
#include <schedule.h>
+#include <kstack.h>
+#include <kmalloc.h>
+#include <arch/uaccess.h>
+
+#define KSTACK_NR_GUARD_PGS 1
+#define KSTACK_GUARD_SZ (KSTACK_NR_GUARD_PGS * PGSIZE)
+static struct kmem_cache *kstack_cache;
+
+/* We allocate KSTKSIZE + PGSIZE vaddrs. So for one-page stacks, we get two
+ * pages. blob points to the bottom of this space. Our job is to allocate the
+ * physical pages for the stack and set up the virtual-to-physical mappings. */
+int kstack_ctor(void *blob, void *priv, int flags)
+{
+ void *stackbot;
+
+ stackbot = kpages_alloc(KSTKSIZE, flags);
+ if (!stackbot)
+ return -1;
+ if (map_vmap_segment((uintptr_t)blob, 0x123456000, KSTACK_NR_GUARD_PGS,
+ PTE_NONE))
+ goto error;
+ if (map_vmap_segment((uintptr_t)blob + KSTACK_GUARD_SZ, PADDR(stackbot),
+ KSTKSIZE / PGSIZE, PTE_KERN_RW))
+ goto error;
+ return 0;
+error:
+ /* On failure, we only need to undo what our dtor would do. The unmaps
+ * happen in the vmap_arena ffunc. */
+ kpages_free(stackbot, KSTKSIZE);
+ return -1;
+}
+
+/* The vmap_arena free will unmap the vaddrs on its own. We just need to free
+ * the physical memory we allocated in ctor. Although we still have mappings
+ * and TLB entries pointing to the memory after we free it (and thus it can be
+ * reused), this is no more dangerous than just freeing the stack. Errant
+ * pointers into an old kstack are still dangerous. */
+void kstack_dtor(void *blob, void *priv)
+{
+ void *stackbot;
+ pte_t pte;
+
+ pte = pgdir_walk(boot_pgdir, blob + KSTACK_GUARD_SZ, 0);
+ assert(pte_walk_okay(pte));
+ stackbot = KADDR(pte_get_paddr(pte));
+ kpages_free(stackbot, KSTKSIZE);
+}
uintptr_t get_kstack(void)
{
- uintptr_t stackbot;
- if (KSTKSIZE == PGSIZE)
- stackbot = (uintptr_t)kpage_alloc_addr();
- else
- stackbot = (uintptr_t)get_cont_pages(KSTKSHIFT - PGSHIFT, 0);
- assert(stackbot);
- return stackbot + KSTKSIZE;
+ void *blob;
+
+ blob = kmem_cache_alloc(kstack_cache, MEM_ATOMIC);
+ /* TODO: think about MEM_WAIT within kthread/blocking code. */
+ assert(blob);
+ return (uintptr_t)blob + KSTKSIZE + KSTACK_GUARD_SZ;
}
void put_kstack(uintptr_t stacktop)
{
- uintptr_t stackbot = stacktop - KSTKSIZE;
- if (KSTKSIZE == PGSIZE)
- page_decref(kva2page((void*)stackbot));
- else
- free_cont_pages((void*)stackbot, KSTKSHIFT - PGSHIFT);
+ kmem_cache_free(kstack_cache, (void*)(stacktop - KSTKSIZE
+ - KSTACK_GUARD_SZ));
}
uintptr_t *kstack_bottom_addr(uintptr_t stacktop)
void kthread_init(void)
{
kthread_kcache = kmem_cache_create("kthread", sizeof(struct kthread),
- __alignof__(struct kthread), 0, 0, 0);
+ __alignof__(struct kthread), 0,
+ NULL, 0, 0, NULL);
+ kstack_cache = kmem_cache_create("kstack", KSTKSIZE + KSTACK_GUARD_SZ,
+ PGSIZE, 0, vmap_arena, kstack_ctor,
+ kstack_dtor, NULL);
}
/* Used by early init routines (smp_boot, etc) */
return kthread;
}
+/* Helper during early boot, where we jump from the bootstack to a real kthread
+ * stack, then run f(). Note that we don't have a kthread yet (done in smp.c).
+ *
+ * After this, our callee (f) can free the bootstack, if we care, by adding it
+ * to the base arena (use the KERNBASE addr, not the KERN_LOAD_ADDR). */
+void __use_real_kstack(void (*f)(void *arg))
+{
+ struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ uintptr_t new_stacktop;
+
+ new_stacktop = get_kstack();
+ set_stack_top(new_stacktop);
+ __reset_stack_pointer(0, new_stacktop, f);
+}
+
/* Starts kthread on the calling core. This does not return, and will handle
* the details of cleaning up whatever is currently running (freeing its stack,
* etc). Pairs with sem_down(). */
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
uintptr_t current_stacktop;
- struct kthread *current_kthread;
+ struct kthread *cur_kth;
/* Avoid messy complications. The kthread will enable_irqsave() when it
* comes back up. */
disable_irq();
put_kstack(pcpui->spare->stacktop);
kmem_cache_free(kthread_kcache, pcpui->spare);
}
- current_kthread = pcpui->cur_kthread;
- current_stacktop = current_kthread->stacktop;
- assert(!current_kthread->sysc); /* catch bugs, prev user should clear */
+ cur_kth = pcpui->cur_kthread;
+ current_stacktop = cur_kth->stacktop;
+ assert(!cur_kth->sysc); /* catch bugs, prev user should clear */
/* Set the spare stuff (current kthread, which includes its stacktop) */
- pcpui->spare = current_kthread;
+ pcpui->spare = cur_kth;
/* When a kthread runs, its stack is the default kernel stack */
set_stack_top(kthread->stacktop);
pcpui->cur_kthread = kthread;
-#ifdef CONFIG_KTHREAD_POISON
- /* Assert and switch to cur stack not in use, kthr stack in use */
- uintptr_t *cur_stack_poison, *kth_stack_poison;
- cur_stack_poison = kstack_bottom_addr(current_stacktop);
- assert(*cur_stack_poison == 0xdeadbeef);
- *cur_stack_poison = 0;
- kth_stack_poison = kstack_bottom_addr(kthread->stacktop);
- assert(!*kth_stack_poison);
- *kth_stack_poison = 0xdeadbeef;
-#endif /* CONFIG_KTHREAD_POISON */
/* Only change current if we need to (the kthread was in process context) */
if (kthread->proc) {
- /* Load our page tables before potentially decreffing cur_proc */
- lcr3(kthread->proc->env_cr3);
- /* Might have to clear out an existing current. If they need to be set
- * later (like in restartcore), it'll be done on demand. */
- if (pcpui->cur_proc)
- proc_decref(pcpui->cur_proc);
- /* We also transfer our counted ref from kthread->proc to cur_proc */
- pcpui->cur_proc = kthread->proc;
+ if (kthread->proc == pcpui->cur_proc) {
+ /* We're already loaded, but we do need to drop the extra ref stored
+ * in kthread->proc. */
+ proc_decref(kthread->proc);
+ kthread->proc = 0;
+ } else {
+ /* Load our page tables before potentially decreffing cur_proc.
+ *
+ * We don't need to do an EPT flush here. The EPT is flushed and
+ * managed in sync with the VMCS. We won't run a different VM (and
+ * thus *need* a different EPT) without first removing the old GPC,
+ * which ultimately will result in a flushed EPT (on x86, this
+ * actually happens when we clear_owning_proc()). */
+ lcr3(kthread->proc->env_cr3);
+ /* Might have to clear out an existing current. If they need to be
+ * set later (like in restartcore), it'll be done on demand. */
+ if (pcpui->cur_proc)
+ proc_decref(pcpui->cur_proc);
+ /* Transfer our counted ref from kthread->proc to cur_proc. */
+ pcpui->cur_proc = kthread->proc;
+ kthread->proc = 0;
+ }
}
/* Finally, restart our thread */
- pop_kernel_ctx(&kthread->context);
+ longjmp(&kthread->context, 1);
}
/* Kmsg handler to launch/run a kthread. This must be a routine message, since
struct kthread *kthread = (struct kthread*)a0;
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
struct proc *cur_proc = pcpui->cur_proc;
-
+
/* Make sure we are a routine kmsg */
assert(in_early_rkmsg_ctx(pcpui));
if (pcpui->owning_proc && pcpui->owning_proc != kthread->proc) {
sem_down(sem);
}
+void kthread_usleep(uint64_t usec)
+{
+ ERRSTACK(1);
+ /* TODO: classic ksched issue: where do we want the wake up to happen? */
+ struct timer_chain *tchain = &per_cpu_info[core_id()].tchain;
+ struct rendez rv;
+
+ int ret_zero(void *ignored)
+ {
+ return 0;
+ }
+
+ /* "discard the error" style (we run the conditional code) */
+ if (!waserror()) {
+ rendez_init(&rv);
+ rendez_sleep_timeout(&rv, ret_zero, 0, usec);
+ }
+ poperror();
+}
+
static void __ktask_wrapper(uint32_t srcid, long a0, long a1, long a2)
{
+ ERRSTACK(1);
void (*fn)(void*) = (void (*)(void*))a0;
void *arg = (void*)a1;
char *name = (char*)a2;
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- assert(pcpui->cur_kthread->is_ktask);
+ assert(is_ktask(pcpui->cur_kthread));
pcpui->cur_kthread->name = name;
+ /* There are some rendezs out there that aren't wrapped. Though no one can
+ * abort them. Yet. */
+ if (waserror()) {
+ printk("Ktask %s threw error %s\n", name, current_errstr());
+ goto out;
+ }
enable_irq();
fn(arg);
+out:
disable_irq();
pcpui->cur_kthread->name = 0;
+ poperror();
/* if we blocked, when we return, PRKM will smp_idle() */
}
(long)name, KMSG_ROUTINE);
}
-void check_poison(char *msg)
-{
-#ifdef CONFIG_KTHREAD_POISON
- struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- assert(pcpui->cur_kthread && pcpui->cur_kthread->stacktop);
- if (*kstack_bottom_addr(pcpui->cur_kthread->stacktop) != 0xdeadbeef) {
- printk("\nBad kthread canary, msg: %s\n", msg);
- panic("");
- }
-#endif /* CONFIG_KTHREAD_POISON */
-}
-
/* Semaphores, using kthreads directly */
static void debug_downed_sem(struct semaphore *sem);
static void debug_upped_sem(struct semaphore *sem);
+static void debug_lock_semlist(void);
+static void debug_unlock_semlist(void);
static void sem_init_common(struct semaphore *sem, int signals)
{
sem->nr_signals = signals;
#ifdef CONFIG_SEMAPHORE_DEBUG
sem->is_on_list = FALSE;
- sem->bt_pc = 0;
- sem->bt_fp = 0;
- sem->calling_core = 0;
#endif
}
sem->irq_okay = TRUE;
}
-bool sem_trydown(struct semaphore *sem)
+bool sem_trydown_bulk(struct semaphore *sem, int nr_signals)
{
bool ret = FALSE;
+
+ /* lockless peek */
+ if (sem->nr_signals - nr_signals < 0)
+ return ret;
+ debug_lock_semlist();
spin_lock(&sem->lock);
- if (sem->nr_signals > 0) {
+ if (sem->nr_signals - nr_signals >= 0) {
sem->nr_signals--;
ret = TRUE;
debug_downed_sem(sem);
}
spin_unlock(&sem->lock);
+ debug_unlock_semlist();
return ret;
}
+bool sem_trydown(struct semaphore *sem)
+{
+ return sem_trydown_bulk(sem, 1);
+}
+
+/* Bottom-half of sem_down. This is called after we jumped to the new stack. */
+static void __attribute__((noreturn)) __unlock_and_idle(void *arg)
+{
+ struct semaphore *sem = (struct semaphore*)arg;
+
+ spin_unlock(&sem->lock);
+ debug_unlock_semlist();
+ smp_idle();
+}
+
/* This downs the semaphore and suspends the current kernel context on its
* waitqueue if there are no pending signals. Note that the case where the
* signal is already there is not optimized. */
void sem_down(struct semaphore *sem)
{
- volatile bool blocking = TRUE; /* signal to short circuit when restarting*/
struct kthread *kthread, *new_kthread;
register uintptr_t new_stacktop;
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ bool irqs_were_on = irq_is_enabled();
assert(can_block(pcpui));
/* Make sure we aren't holding any locks (only works if SPINLOCK_DEBUG) */
- assert(!pcpui->lock_depth);
- assert(pcpui->cur_kthread);
+ if (pcpui->lock_depth)
+ panic("Kthread tried to sleep, with lockdepth %d\n", pcpui->lock_depth);
/* Try to down the semaphore. If there is a signal there, we can skip all
* of the sleep prep and just return. */
+#ifdef CONFIG_SEM_SPINWAIT
+ for (int i = 0; i < CONFIG_SEM_SPINWAIT_NR_LOOPS; i++) {
+ if (sem_trydown(sem))
+ goto block_return_path;
+ cpu_relax();
+ }
+#else
if (sem_trydown(sem))
goto block_return_path;
+#endif
+ assert(pcpui->cur_kthread);
/* We're probably going to sleep, so get ready. We'll check again later. */
kthread = pcpui->cur_kthread;
/* We need to have a spare slot for restart, so we also use it when
new_kthread = pcpui->spare;
new_stacktop = new_kthread->stacktop;
pcpui->spare = 0;
- /* Based on how we set is_ktask (in PRKM), we'll usually have a spare
- * with is_ktask set, even though the default setting is off. The
- * reason is that the launching of blocked kthreads also uses PRKM, and
- * that KMSG (__launch_kthread) doesn't return. Thus the soon-to-be
- * spare kthread, that is launching another, has is_ktask set. */
- new_kthread->is_ktask = FALSE;
+ /* The old flags could have KTH_IS_KTASK set. The reason is that the
+ * launching of blocked kthreads also uses PRKM, and that KMSG
+ * (__launch_kthread) doesn't return. Thus the soon-to-be spare
+ * kthread, that is launching another, has flags & KTH_IS_KTASK set. */
+ new_kthread->flags = KTH_DEFAULT_FLAGS;
new_kthread->proc = 0;
new_kthread->name = 0;
} else {
new_kthread = __kthread_zalloc();
+ new_kthread->flags = KTH_DEFAULT_FLAGS;
new_stacktop = get_kstack();
new_kthread->stacktop = new_stacktop;
-#ifdef CONFIG_KTHREAD_POISON
- *kstack_bottom_addr(new_stacktop) = 0;
-#endif /* CONFIG_KTHREAD_POISON */
}
/* Set the core's new default stack and kthread */
set_stack_top(new_stacktop);
pcpui->cur_kthread = new_kthread;
-#ifdef CONFIG_KTHREAD_POISON
- /* Mark the new stack as in-use, and unmark the current kthread */
- uintptr_t *new_stack_poison, *kth_stack_poison;
- new_stack_poison = kstack_bottom_addr(new_stacktop);
- assert(!*new_stack_poison);
- *new_stack_poison = 0xdeadbeef;
- kth_stack_poison = kstack_bottom_addr(kthread->stacktop);
- assert(*kth_stack_poison == 0xdeadbeef);
- *kth_stack_poison = 0;
-#endif /* CONFIG_KTHREAD_POISON */
/* Kthreads that are ktasks are not related to any process, and do not need
* to work in a process's address space. They can operate in any address
- * space that has the kernel mapped (like boot_pgdir, or any pgdir).
+ * space that has the kernel mapped (like boot_pgdir, or any pgdir). Some
+ * ktasks may switch_to, at which point they do care about the address
+ * space and must maintain a reference.
*
- * Other kthreads need to stay in the process context (if there is one), but
- * we want the core (which could be a vcore) to stay in the context too. In
- * the future, we could check owning_proc. If it isn't set, we could leave
- * the process context and transfer the refcnt to kthread->proc. */
- if (!kthread->is_ktask) {
+ * Normal kthreads need to stay in the process context, but we want the core
+ * (which could be a vcore) to stay in the context too. */
+ if (kthread->flags & KTH_SAVE_ADDR_SPACE) {
kthread->proc = current;
+ assert(kthread->proc);
+ /* In the future, we could check owning_proc. If it isn't set, we could
+ * clear current and transfer the refcnt to kthread->proc. If so, we'll
+ * need to reset the cr3 to something (boot_cr3 or owning_proc's cr3),
+ * which might not be worth the potentially excessive TLB flush. */
proc_incref(kthread->proc, 1);
} else {
- kthread->proc = 0;
- }
- /* Save the context, toggle blocking for the reactivation */
- save_kernel_ctx(&kthread->context);
- if (!blocking)
+ assert(kthread->proc == 0);
+ }
+ if (setjmp(&kthread->context))
goto block_return_path;
- blocking = FALSE; /* for when it starts back up */
- /* Down the semaphore. We need this to be inline. If we're sleeping, once
- * we unlock the kthread could be started up again and can return and start
- * trashing this function's stack, hence the weird control flow. */
+ debug_lock_semlist();
spin_lock(&sem->lock);
- if (sem->nr_signals-- <= 0) {
+ sem->nr_signals -= 1;
+ if (sem->nr_signals < 0) {
TAILQ_INSERT_TAIL(&sem->waiters, kthread, link);
- debug_downed_sem(sem);
- /* At this point, we know we'll sleep and change stacks later. Once we
- * unlock, we could have the kthread restarted (possibly on another
- * core), so we need to disable irqs until we are on our new stack.
- * Otherwise, if we take an IRQ, we'll be using our stack while another
- * core is using it (restarted kthread). Basically, disabling irqs
- * allows us to atomically unlock and 'yield'. */
- disable_irq();
- } else { /* we didn't sleep */
- debug_downed_sem(sem);
- goto unwind_sleep_prep;
+ debug_downed_sem(sem); /* need to debug after inserting */
+ /* At this point, we know we'll sleep and change stacks. Once we unlock
+ * the sem, we could have the kthread restarted (possibly on another
+ * core), so we need to leave the old stack before unlocking. If we
+ * don't and we stay on the stack, then if we take an IRQ or NMI (NMI
+ * that doesn't change stacks, unlike x86_64), we'll be using the stack
+ * at the same time as the kthread. We could just disable IRQs, but
+ * that wouldn't protect us from NMIs that don't change stacks. */
+ __reset_stack_pointer(sem, new_stacktop, __unlock_and_idle);
+ assert(0);
}
- spin_unlock(&sem->lock);
- /* Switch to the core's default stack. After this, don't use local
- * variables. TODO: we shouldn't be using new_stacktop either, can't always
- * trust the register keyword (AFAIK). */
- set_stack_pointer(new_stacktop);
- smp_idle(); /* reenables irqs eventually */
- /* smp_idle never returns */
- assert(0);
-unwind_sleep_prep:
/* We get here if we should not sleep on sem (the signal beat the sleep).
- * Note we are not optimizing for cases where the signal won. */
+ * We debug_downed_sem since we actually downed it - just didn't sleep. */
+ debug_downed_sem(sem);
spin_unlock(&sem->lock);
+ debug_unlock_semlist();
printd("[kernel] Didn't sleep, unwinding...\n");
/* Restore the core's current and default stacktop */
- current = kthread->proc; /* arguably unnecessary */
- if (kthread->proc)
+ if (kthread->flags & KTH_SAVE_ADDR_SPACE) {
proc_decref(kthread->proc);
+ kthread->proc = 0;
+ }
set_stack_top(kthread->stacktop);
pcpui->cur_kthread = kthread;
/* Save the allocs as the spare */
assert(!pcpui->spare);
pcpui->spare = new_kthread;
-#ifdef CONFIG_KTHREAD_POISON
- /* switch back to old stack in use, new one not */
- *new_stack_poison = 0;
- *kth_stack_poison = 0xdeadbeef;
-#endif /* CONFIG_KTHREAD_POISON */
block_return_path:
printd("[kernel] Returning from being 'blocked'! at %llu\n", read_tsc());
+ /* restart_kthread and longjmp did not reenable IRQs. We need to make sure
+ * irqs are on if they were on when we started to block. If they were
+ * already on and we short-circuited the block, it's harmless to reenable
+ * them. */
+ if (irqs_were_on)
+ enable_irq();
return;
}
+void sem_down_bulk(struct semaphore *sem, int nr_signals)
+{
+ /* This is far from ideal. Our current sem code expects a 1:1 pairing of
+ * signals to waiters. For instance, if we have 10 waiters of -1 each or 1
+ * waiter of -10, we can't tell from looking at the overall structure. We'd
+ * need to track the desired number of signals per waiter.
+ *
+ * Note that if there are a bunch of signals available, sem_down will
+ * quickly do a try_down and return, so we won't block repeatedly. But if
+ * we do block, we could wake up N times. */
+ for (int i = 0; i < nr_signals; i++)
+ sem_down(sem);
+}
+
/* Ups the semaphore. If it was < 0, we need to wake up someone, which we do.
* Returns TRUE if we woke someone, FALSE o/w (used for debugging in some
* places). If we need more control, we can implement a version of the old
bool sem_up(struct semaphore *sem)
{
struct kthread *kthread = 0;
+
+ debug_lock_semlist();
spin_lock(&sem->lock);
if (sem->nr_signals++ < 0) {
assert(!TAILQ_EMPTY(&sem->waiters));
}
debug_upped_sem(sem);
spin_unlock(&sem->lock);
+ debug_unlock_semlist();
/* Note that once we call kthread_runnable(), we cannot touch the sem again.
* Some sems are on stacks. The caller can touch sem, if it knows about the
* memory/usage of the sem. Likewise, we can't touch the kthread either. */
return FALSE;
}
-bool sem_trydown_irqsave(struct semaphore *sem, int8_t *irq_state)
+bool sem_trydown_bulk_irqsave(struct semaphore *sem, int nr_signals,
+ int8_t *irq_state)
{
bool ret;
+
disable_irqsave(irq_state);
- ret = sem_trydown(sem);
+ ret = sem_trydown_bulk(sem, nr_signals);
enable_irqsave(irq_state);
return ret;
}
-void sem_down_irqsave(struct semaphore *sem, int8_t *irq_state)
+bool sem_trydown_irqsave(struct semaphore *sem, int8_t *irq_state)
+{
+ return sem_trydown_bulk_irqsave(sem, 1, irq_state);
+}
+
+void sem_down_bulk_irqsave(struct semaphore *sem, int nr_signals,
+ int8_t *irq_state)
{
disable_irqsave(irq_state);
- sem_down(sem);
+ sem_down_bulk(sem, nr_signals);
enable_irqsave(irq_state);
}
+void sem_down_irqsave(struct semaphore *sem, int8_t *irq_state)
+{
+ sem_down_bulk_irqsave(sem, 1, irq_state);
+}
+
bool sem_up_irqsave(struct semaphore *sem, int8_t *irq_state)
{
bool retval;
#ifdef CONFIG_SEMAPHORE_DEBUG
struct semaphore_tailq sems_with_waiters =
TAILQ_HEAD_INITIALIZER(sems_with_waiters);
+/* The lock ordering is sems_with_waiters_lock -> any_sem_lock */
spinlock_t sems_with_waiters_lock = SPINLOCK_INITIALIZER_IRQSAVE;
+static void debug_lock_semlist(void)
+{
+ spin_lock_irqsave(&sems_with_waiters_lock);
+}
+
+static void debug_unlock_semlist(void)
+{
+ spin_unlock_irqsave(&sems_with_waiters_lock);
+}
+
/* this gets called any time we downed the sem, regardless of whether or not we
* waited */
static void debug_downed_sem(struct semaphore *sem)
{
- sem->bt_pc = read_pc();
- sem->bt_fp = read_bp();
- sem->calling_core = core_id();
if (TAILQ_EMPTY(&sem->waiters) || sem->is_on_list)
return;
- spin_lock_irqsave(&sems_with_waiters_lock);
TAILQ_INSERT_HEAD(&sems_with_waiters, sem, link);
- spin_unlock_irqsave(&sems_with_waiters_lock);
sem->is_on_list = TRUE;
}
static void debug_upped_sem(struct semaphore *sem)
{
if (TAILQ_EMPTY(&sem->waiters) && sem->is_on_list) {
- spin_lock_irqsave(&sems_with_waiters_lock);
TAILQ_REMOVE(&sems_with_waiters, sem, link);
- spin_unlock_irqsave(&sems_with_waiters_lock);
sem->is_on_list = FALSE;
}
}
#else
+static void debug_lock_semlist(void)
+{
+ /* no debugging */
+}
+
+static void debug_unlock_semlist(void)
+{
+ /* no debugging */
+}
+
static void debug_downed_sem(struct semaphore *sem)
{
/* no debugging */
#endif /* CONFIG_SEMAPHORE_DEBUG */
-void print_sem_info(struct semaphore *sem)
+static bool __sem_has_pid(struct semaphore *sem, pid_t pid)
{
struct kthread *kth_i;
+
+ if (pid == -1)
+ return TRUE;
+ TAILQ_FOREACH(kth_i, &sem->waiters, link) {
+ if (kth_i->proc) {
+ if (kth_i->proc->pid == pid)
+ return TRUE;
+ } else {
+ if (pid == 0)
+ return TRUE;
+ }
+ }
+ return FALSE;
+}
+
+static void print_sem_info(struct semaphore *sem, pid_t pid)
+{
+ struct kthread *kth_i;
+
/* Always safe to irqsave */
spin_lock_irqsave(&sem->lock);
- printk("Semaphore %p has %d signals (neg = waiters)", sem, sem->nr_signals);
-#ifdef CONFIG_SEMAPHORE_DEBUG
- printk(", recently downed on core %d with pc/frame %p %p\n",
- sem->calling_core, sem->bt_pc, sem->bt_fp);
-#else
- printk("\n");
-#endif /* CONFIG_SEMAPHORE_DEBUG */
+ if (!__sem_has_pid(sem, pid)) {
+ spin_unlock_irqsave(&sem->lock);
+ return;
+ }
+ printk("Semaphore %p has %d signals (neg = waiters)\n", sem,
+ sem->nr_signals);
TAILQ_FOREACH(kth_i, &sem->waiters, link)
- printk("\tKthread %p (%s), proc %d (%p), sysc %p\n", kth_i, kth_i->name,
- kth_i->proc ? kth_i->proc->pid : 0, kth_i->proc, kth_i->sysc);
+ printk("\tKthread %p (%s), proc %d, sysc %p, pc/frame %p %p\n",
+ kth_i, kth_i->name, kth_i->proc ? kth_i->proc->pid : 0,
+ kth_i->sysc, jmpbuf_get_pc(&kth_i->context),
+ jmpbuf_get_fp(&kth_i->context));
+ printk("\n");
spin_unlock_irqsave(&sem->lock);
}
-void print_all_sem_info(void)
+void print_all_sem_info(pid_t pid)
{
#ifdef CONFIG_SEMAPHORE_DEBUG
struct semaphore *sem_i;
printk("All sems with waiters:\n");
spin_lock_irqsave(&sems_with_waiters_lock);
TAILQ_FOREACH(sem_i, &sems_with_waiters, link)
- print_sem_info(sem_i);
+ print_sem_info(sem_i, pid);
spin_unlock_irqsave(&sems_with_waiters_lock);
#else
printk("Failed to print all sems: build with CONFIG_SEMAPHORE_DEBUG\n");
static void sem_wake_one(struct semaphore *sem)
{
struct kthread *kthread;
+
/* these locks will be irqsaved if the CV is irqsave (only need the one) */
+ debug_lock_semlist();
spin_lock(&sem->lock);
assert(sem->nr_signals < 0);
sem->nr_signals++;
TAILQ_REMOVE(&sem->waiters, kthread, link);
debug_upped_sem(sem);
spin_unlock(&sem->lock);
+ debug_unlock_semlist();
kthread_runnable(kthread);
}
enable_irqsave(irq_state);
}
-/* Helper, aborts and releases a CLE. dereg_ spinwaits on abort_in_progress. */
+/* Helper, aborts and releases a CLE. dereg_ spinwaits on abort_in_progress.
+ * This can throw a PF */
static void __abort_and_release_cle(struct cv_lookup_elm *cle)
{
int8_t irq_state = 0;
* all the memory for CLE is safe */
bool abort_sysc(struct proc *p, struct syscall *sysc)
{
+ ERRSTACK(1);
struct cv_lookup_elm *cle;
int8_t irq_state = 0;
+
spin_lock_irqsave(&p->abort_list_lock);
TAILQ_FOREACH(cle, &p->abortable_sleepers, link) {
if (cle->sysc == sysc) {
spin_unlock_irqsave(&p->abort_list_lock);
if (!cle)
return FALSE;
- __abort_and_release_cle(cle);
+ if (!waserror()) /* discard error */
+ __abort_and_release_cle(cle);
+ poperror();
return TRUE;
}
-/* This will abort any abortabls at the time the call was started. New
- * abortables could be registered concurrently. The main caller I see for this
- * is proc_destroy(), so DYING will be set, and new abortables will quickly
- * abort and dereg when they see their proc is DYING. */
-void abort_all_sysc(struct proc *p)
+/* This will abort any abortables at the time the call was started for which
+ * should_abort(cle, arg) returns true. New abortables could be registered
+ * concurrently.
+ *
+ * One caller for this is proc_destroy(), in which case DYING_ABORT will be set,
+ * and new abortables will quickly abort and dereg when they see their proc is
+ * DYING_ABORT. */
+static int __abort_all_sysc(struct proc *p,
+ bool (*should_abort)(struct cv_lookup_elm*, void*),
+ void *arg)
{
+ ERRSTACK(1);
struct cv_lookup_elm *cle;
int8_t irq_state = 0;
struct cv_lookup_tailq abortall_list;
- struct proc *old_proc = switch_to(p);
+ uintptr_t old_proc = switch_to(p);
+ int ret = 0;
+
/* Concerns: we need to not remove them from their original list, since
* concurrent wake ups will cause a dereg, which will remove from the list.
* We also can't touch freed memory, so we need a refcnt to keep cles
TAILQ_INIT(&abortall_list);
spin_lock_irqsave(&p->abort_list_lock);
TAILQ_FOREACH(cle, &p->abortable_sleepers, link) {
+ if (!should_abort(cle, arg))
+ continue;
atomic_inc(&cle->abort_in_progress);
TAILQ_INSERT_HEAD(&abortall_list, cle, abortall_link);
+ ret++;
}
spin_unlock_irqsave(&p->abort_list_lock);
- TAILQ_FOREACH(cle, &abortall_list, abortall_link)
- __abort_and_release_cle(cle);
+ if (!waserror()) { /* discard error */
+ TAILQ_FOREACH(cle, &abortall_list, abortall_link)
+ __abort_and_release_cle(cle);
+ }
+ poperror();
switch_back(p, old_proc);
+ return ret;
+}
+
+static bool always_abort(struct cv_lookup_elm *cle, void *arg)
+{
+ return TRUE;
+}
+
+void abort_all_sysc(struct proc *p)
+{
+ __abort_all_sysc(p, always_abort, 0);
+}
+
+/* cle->sysc could be a bad pointer. we can either use copy_from_user (btw,
+ * we're already in their addr space) or we can use a waserror in
+ * __abort_all_sysc(). Both options are fine. I went with it here for a couple
+ * reasons. It is only this abort function pointer that accesses sysc, though
+ * that could change. Our syscall aborting isn't plugged into a broader error()
+ * handler yet, which means we'd want to poperror instead of nexterror in
+ * __abort_all_sysc, and that would required int ret getting a volatile flag. */
+static bool sysc_uses_fd(struct cv_lookup_elm *cle, void *fd)
+{
+ struct syscall local_sysc;
+ int err;
+
+ err = copy_from_user(&local_sysc, cle->sysc, sizeof(struct syscall));
+ /* Trigger an abort on error */
+ if (err)
+ return TRUE;
+ return syscall_uses_fd(&local_sysc, (int)(long)fd);
+}
+
+int abort_all_sysc_fd(struct proc *p, int fd)
+{
+ return __abort_all_sysc(p, sysc_uses_fd, (void*)(long)fd);
}
/* Being on the abortable list means that the CLE, KTH, SYSC, and CV are valid
cle->cv = cv;
cle->kthread = pcpui->cur_kthread;
/* Could be a ktask. Can build in support for aborting these later */
- if (cle->kthread->is_ktask) {
+ if (is_ktask(cle->kthread)) {
cle->sysc = 0;
return;
}
cle->sysc = cle->kthread->sysc;
- assert(cle->sysc);
cle->proc = pcpui->cur_proc;
atomic_init(&cle->abort_in_progress, 0);
spin_lock_irqsave(&cle->proc->abort_list_lock);
* CV lock. So if we hold the CV lock, we can deadlock (circular dependency).*/
void dereg_abortable_cv(struct cv_lookup_elm *cle)
{
- if (cle->kthread->is_ktask)
+ if (is_ktask(cle->kthread))
return;
assert(cle->proc);
spin_lock_irqsave(&cle->proc->abort_list_lock);
* this with things for ktasks in the future. */
bool should_abort(struct cv_lookup_elm *cle)
{
- if (cle->kthread->is_ktask)
+ struct syscall local_sysc;
+ int err;
+
+ if (is_ktask(cle->kthread))
return FALSE;
- if (cle->proc && (cle->proc->state == PROC_DYING))
- return TRUE;
- if (cle->sysc && (atomic_read(&cle->sysc->flags) & SC_ABORT))
+ if (cle->proc && (cle->proc->state == PROC_DYING_ABORT))
return TRUE;
+ if (cle->sysc) {
+ assert(cle->proc && (cle->proc == current));
+ err = copy_from_user(&local_sysc, cle->sysc,
+ offsetof(struct syscall, flags) +
+ sizeof(cle->sysc->flags));
+ /* just go ahead and abort if there was an error */
+ if (err || (atomic_read(&local_sysc.flags) & SC_ABORT))
+ return TRUE;
+ }
return FALSE;
}
+
+/* Sometimes the kernel needs to switch out of process context and into a
+ * 'process-less' kernel thread. This is basically a ktask. We use this mostly
+ * when performing file ops as the kernel. It's nasty, and all uses of this
+ * probably should be removed. (TODO: KFOP). */
+uintptr_t switch_to_ktask(void)
+{
+ struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ struct kthread *kth = pcpui->cur_kthread;
+
+ if (is_ktask(kth))
+ return 0;
+ /* We leave the SAVE_ADDR_SPACE flag on. Now we're basically a ktask that
+ * cares about its addr space, since we need to return to it (not that we're
+ * leaving). */
+ kth->flags |= KTH_IS_KTASK;
+ return 1;
+}
+
+void switch_back_from_ktask(uintptr_t old_ret)
+{
+ struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ struct kthread *kth = pcpui->cur_kthread;
+
+ if (old_ret)
+ kth->flags &= ~KTH_IS_KTASK;
+}
projects / akaros.git / blobdiff