* Barret Rhoden <brho@cs.berkeley.edu>
* See LICENSE for details. */
-#ifdef __SHARC__
-#pragma nosharc
-#endif
-
-#include <ros/bcq.h>
#include <event.h>
#include <arch/arch.h>
#include <bitmask.h>
#include <smp.h>
#include <pmap.h>
#include <trap.h>
+#include <umem.h>
#include <schedule.h>
#include <manager.h>
#include <stdio.h>
#include <hashtable.h>
#include <slab.h>
#include <sys/queue.h>
-#include <frontend.h>
#include <monitor.h>
#include <elf.h>
#include <arsc_server.h>
-#include <devfs.h>
#include <kmalloc.h>
+#include <ros/procinfo.h>
+#include <init.h>
+#include <rcu.h>
struct kmem_cache *proc_cache;
spin_unlock(&pid_bmask_lock);
}
+/* 'resume' is the time int ticks of the most recent onlining. 'total' is the
+ * amount of time in ticks consumed up to and including the current offlining.
+ *
+ * We could move these to the map and unmap of vcores, though not every place
+ * uses that (SCPs, in particular). However, maps/unmaps happen remotely;
+ * something to consider. If we do it remotely, we can batch them up and do one
+ * rdtsc() for all of them. For now, I want to do them on the core, around when
+ * we do the context change. It'll also parallelize the accounting a bit. */
+void vcore_account_online(struct proc *p, uint32_t vcoreid)
+{
+ struct vcore *vc = &p->procinfo->vcoremap[vcoreid];
+ vc->resume_ticks = read_tsc();
+}
+
+void vcore_account_offline(struct proc *p, uint32_t vcoreid)
+{
+ struct vcore *vc = &p->procinfo->vcoremap[vcoreid];
+ vc->total_ticks += read_tsc() - vc->resume_ticks;
+}
+
+uint64_t vcore_account_gettotal(struct proc *p, uint32_t vcoreid)
+{
+ struct vcore *vc = &p->procinfo->vcoremap[vcoreid];
+ return vc->total_ticks;
+}
+
/* While this could be done with just an assignment, this gives us the
* opportunity to check for bad transitions. Might compile these out later, so
* we shouldn't rely on them for sanity checking from userspace. */
* RGM -> RBS
* RGS -> D
* RGM -> D
+ * D -> DA
*
* These ought to be implemented later (allowed, not thought through yet).
* RBS -> D
panic("Invalid State Transition! PROC_WAITING to %02x", state);
break;
case PROC_DYING:
- if (state != PROC_CREATED) // when it is reused (TODO)
+ if (state != PROC_DYING_ABORT)
panic("Invalid State Transition! PROC_DYING to %02x", state);
break;
+ case PROC_DYING_ABORT:
+ panic("Invalid State Transition! PROC_DYING to %02x", state);
+ break;
case PROC_RUNNABLE_M:
if (!(state & (PROC_RUNNING_M | PROC_DYING)))
panic("Invalid State Transition! PROC_RUNNABLE_M to %02x", state);
* so continue
*/
- if (kref_get_not_zero(&p->p_kref, 1)){
+ if (kref_get_not_zero(&p->p_kref, 1)) {
/* this one counts */
if (! n){
printd("pid_nth: at end, p %p\n", p);
kref_put(&p->p_kref);
n--;
}
- if (!hashtable_iterator_advance(iter)){
+ if (!hashtable_iterator_advance(iter)) {
p = NULL;
break;
}
/* Catch issues with the vcoremap and TAILQ_ENTRY sizes */
static_assert(sizeof(TAILQ_ENTRY(vcore)) == sizeof(void*) * 2);
proc_cache = kmem_cache_create("proc", sizeof(struct proc),
- MAX(ARCH_CL_SIZE, __alignof__(struct proc)), 0, 0, 0);
+ MAX(ARCH_CL_SIZE,
+ __alignof__(struct proc)), 0, NULL, 0,
+ 0, NULL);
/* Init PID mask and hash. pid 0 is reserved. */
SET_BITMASK_BIT(pid_bmask, 0);
spinlock_init(&pid_hash_lock);
atomic_init(&num_envs, 0);
}
+void proc_set_username(struct proc *p, char *name)
+{
+ set_username(&p->user, name);
+}
+
+/*
+ * Copies username from the parent process. This is the only case where a
+ * reader blocks writing, just to be extra safe during process initialization.
+ *
+ * Note that since this is intended to be called during initialization, the
+ * child's name lock is NOT used for writing. Nothing else should be able to
+ * read or write yet, so this can be a simple memcpy once the parent is locked.
+ */
+void proc_inherit_parent_username(struct proc *child, struct proc *parent)
+{
+ spin_lock(&parent->user.name_lock);
+
+ // copy entire parent buffer for constant runtime
+ memcpy(child->user.name, parent->user.name, sizeof(child->user.name));
+
+ spin_unlock(&parent->user.name_lock);
+}
+
+void proc_set_progname(struct proc *p, char *name)
+{
+ if (name == NULL)
+ name = DEFAULT_PROGNAME;
+
+ /* might have an issue if a dentry name isn't null terminated, and we'd get
+ * extra junk up to progname_sz. Or crash. */
+ strlcpy(p->progname, name, PROC_PROGNAME_SZ);
+}
+
+void proc_replace_binary_path(struct proc *p, char *path)
+{
+ if (p->binary_path)
+ free_path(p, p->binary_path);
+ p->binary_path = path;
+}
+
/* Be sure you init'd the vcore lists before calling this. */
-static void proc_init_procinfo(struct proc* p)
+void proc_init_procinfo(struct proc* p)
{
p->procinfo->pid = p->pid;
p->procinfo->ppid = p->ppid;
p->procinfo->max_vcores = max_vcores(p);
- p->procinfo->tsc_freq = system_timing.tsc_freq;
- p->procinfo->timing_overhead = system_timing.timing_overhead;
- p->procinfo->heap_bottom = 0;
+ p->procinfo->tsc_freq = __proc_global_info.tsc_freq;
+ p->procinfo->timing_overhead = __proc_global_info.tsc_overhead;
+ p->procinfo->program_end = 0;
/* 0'ing the arguments. Some higher function will need to set them */
- memset(p->procinfo->argp, 0, sizeof(p->procinfo->argp));
- memset(p->procinfo->argbuf, 0, sizeof(p->procinfo->argbuf));
memset(p->procinfo->res_grant, 0, sizeof(p->procinfo->res_grant));
/* 0'ing the vcore/pcore map. Will link the vcores later. */
memset(&p->procinfo->vcoremap, 0, sizeof(p->procinfo->vcoremap));
p->procinfo->num_vcores = 0;
p->procinfo->is_mcp = FALSE;
p->procinfo->coremap_seqctr = SEQCTR_INITIALIZER;
- /* For now, we'll go up to the max num_cpus (at runtime). In the future,
- * there may be cases where we can have more vcores than num_cpus, but for
- * now we'll leave it like this. */
- for (int i = 0; i < num_cpus; i++) {
+ /* It's a bug in the kernel if we let them ask for more than max */
+ for (int i = 0; i < p->procinfo->max_vcores; i++) {
TAILQ_INSERT_TAIL(&p->inactive_vcs, &p->procinfo->vcoremap[i], list);
}
}
-static void proc_init_procdata(struct proc *p)
+void proc_init_procdata(struct proc *p)
{
memset(p->procdata, 0, sizeof(struct procdata));
/* processes can't go into vc context on vc 0 til they unset this. This is
atomic_set(&p->procdata->vcore_preempt_data[0].flags, VC_SCP_NOVCCTX);
}
+static void proc_open_stdfds(struct proc *p)
+{
+ int fd;
+ struct proc *old_current = current;
+
+ /* Due to the way the syscall helpers assume the target process is current,
+ * we need to set current temporarily. We don't use switch_to, since that
+ * actually loads the process's address space, which might be empty or
+ * incomplete. These syscalls shouldn't access user memory, especially
+ * considering how we're probably in the boot pgdir. */
+ current = p;
+ fd = sysopenat(AT_FDCWD, "#cons/stdin", O_READ);
+ assert(fd == 0);
+ fd = sysopenat(AT_FDCWD, "#cons/stdout", O_WRITE);
+ assert(fd == 1);
+ fd = sysopenat(AT_FDCWD, "#cons/stderr", O_WRITE);
+ assert(fd == 2);
+ current = old_current;
+}
+
/* Allocates and initializes a process, with the given parent. Currently
* writes the *p into **pp, and returns 0 on success, < 0 for an error.
* Errors include:
* - ENOFREEPID if it can't get a PID
* - ENOMEM on memory exhaustion */
-error_t proc_alloc(struct proc **pp, struct proc *parent)
+error_t proc_alloc(struct proc **pp, struct proc *parent, int flags)
{
error_t r;
struct proc *p;
if (!(p = kmem_cache_alloc(proc_cache, 0)))
return -ENOMEM;
/* zero everything by default, other specific items are set below */
- memset(p, 0, sizeof(struct proc));
-
- { INITSTRUCT(*p)
+ memset(p, 0, sizeof(*p));
/* only one ref, which we pass back. the old 'existence' ref is managed by
* the ksched */
kref_init(&p->p_kref, __proc_free, 1);
- // Setup the default map of where to get cache colors from
- p->cache_colors_map = global_cache_colors_map;
- p->next_cache_color = 0;
/* Initialize the address space */
if ((r = env_setup_vm(p)) < 0) {
kmem_cache_free(proc_cache, p);
kmem_cache_free(proc_cache, p);
return -ENOFREEPID;
}
+ if (parent && parent->binary_path)
+ kstrdup(&p->binary_path, parent->binary_path);
/* Set the basic status variables. */
spinlock_init(&p->proc_lock);
+ spinlock_init(&p->user.name_lock);
p->exitcode = 1337; /* so we can see processes killed by the kernel */
if (parent) {
p->ppid = parent->pid;
+ proc_inherit_parent_username(p, parent);
+ proc_incref(p, 1); /* storing a ref in the parent */
/* using the CV's lock to protect anything related to child waiting */
cv_lock(&parent->child_wait);
TAILQ_INSERT_TAIL(&parent->children, p, sibling_link);
cv_unlock(&parent->child_wait);
} else {
p->ppid = 0;
+ strlcpy(p->user.name, eve.name, sizeof(p->user.name));
+ printk("Parentless process assigned username '%s'\n", p->user.name);
}
TAILQ_INIT(&p->children);
cv_init(&p->child_wait);
p->state = PROC_CREATED; /* shouldn't go through state machine for init */
p->env_flags = 0;
- p->env_entry = 0; // cheating. this really gets set later
- p->heap_top = 0;
spinlock_init(&p->vmr_lock);
spinlock_init(&p->pte_lock);
TAILQ_INIT(&p->vm_regions); /* could init this in the slab */
+ p->vmr_history = 0;
/* Initialize the vcore lists, we'll build the inactive list so that it
* includes all vcores when we initialize procinfo. Do this before initing
* procinfo. */
SYSEVENTRINGSIZE);
/* Init FS structures TODO: cleanup (might pull this out) */
- kref_get(&default_ns.kref, 1);
- p->ns = &default_ns;
- spinlock_init(&p->fs_env.lock);
- p->fs_env.umask = parent ? parent->fs_env.umask : S_IWGRP | S_IWOTH;
- p->fs_env.root = p->ns->root->mnt_root;
- kref_get(&p->fs_env.root->d_kref, 1);
- p->fs_env.pwd = parent ? parent->fs_env.pwd : p->fs_env.root;
- kref_get(&p->fs_env.pwd->d_kref, 1);
+ p->umask = parent ? parent->umask : S_IWGRP | S_IWOTH;
memset(&p->open_files, 0, sizeof(p->open_files)); /* slightly ghetto */
spinlock_init(&p->open_files.lock);
p->open_files.max_files = NR_OPEN_FILES_DEFAULT;
p->open_files.max_fdset = NR_FILE_DESC_DEFAULT;
p->open_files.fd = p->open_files.fd_array;
p->open_files.open_fds = (struct fd_set*)&p->open_files.open_fds_init;
+ if (parent) {
+ if (flags & PROC_DUP_FGRP)
+ clone_fdt(&parent->open_files, &p->open_files);
+ } else {
+ /* no parent, we're created from the kernel */
+ proc_open_stdfds(p);
+ }
/* Init the ucq hash lock */
p->ucq_hashlock = (struct hashlock*)&p->ucq_hl_noref;
hashlock_init_irqsave(p->ucq_hashlock, HASHLOCK_DEFAULT_SZ);
atomic_inc(&num_envs);
- frontend_proc_init(p);
- plan9setup(p, parent);
+ plan9setup(p, parent, flags);
devalarm_init(p);
TAILQ_INIT(&p->abortable_sleepers);
spinlock_init_irqsave(&p->abort_list_lock);
+ memset(&p->vmm, 0, sizeof(struct vmm));
+ spinlock_init(&p->vmm.lock);
+ qlock_init(&p->vmm.qlock);
printd("[%08x] new process %08x\n", current ? current->pid : 0, p->pid);
- } // INIT_STRUCT
*pp = p;
return 0;
}
spin_unlock(&pid_hash_lock);
}
-/* Creates a process from the specified file, argvs, and envps. Tempted to get
- * rid of proc_alloc's style, but it is so quaint... */
-struct proc *proc_create(struct file *prog, char **argv, char **envp)
+/* Creates a process from the specified file, argvs, and envps. */
+struct proc *proc_create(struct file_or_chan *prog, char **argv, char **envp)
{
struct proc *p;
error_t r;
- if ((r = proc_alloc(&p, current)) < 0)
- panic("proc_create: %e", r); /* one of 3 quaint usages of %e */
- procinfo_pack_args(p->procinfo, argv, envp);
- assert(load_elf(p, prog) == 0);
- /* Connect to stdin, stdout, stderr */
- assert(insert_file(&p->open_files, dev_stdin, 0) == 0);
- assert(insert_file(&p->open_files, dev_stdout, 0) == 1);
- assert(insert_file(&p->open_files, dev_stderr, 0) == 2);
+ if ((r = proc_alloc(&p, current, 0 /* flags */)) < 0)
+ panic("proc_create: %d", r);
+ int argc = 0, envc = 0;
+ if(argv) while(argv[argc]) argc++;
+ if(envp) while(envp[envc]) envc++;
+ proc_set_progname(p, argc ? argv[0] : NULL);
+ assert(load_elf(p, prog, argc, argv, envc, envp) == 0);
__proc_ready(p);
return p;
}
-static int __cb_assert_no_pg(struct proc *p, pte_t *pte, void *va, void *arg)
+static int __cb_assert_no_pg(struct proc *p, pte_t pte, void *va, void *arg)
{
- assert(!*pte);
+ assert(pte_is_unmapped(pte));
return 0;
}
static void __proc_free(struct kref *kref)
{
struct proc *p = container_of(kref, struct proc, p_kref);
+ void *hash_ret;
physaddr_t pa;
printd("[PID %d] freeing proc: %d\n", current ? current->pid : 0, p->pid);
assert(kref_refcnt(&p->p_kref) == 0);
assert(TAILQ_EMPTY(&p->alarmset.list));
+ if (p->strace) {
+ kref_put(&p->strace->procs);
+ kref_put(&p->strace->users);
+ }
+ __vmm_struct_cleanup(p);
+ p->progname[0] = 0;
+ free_path(p, p->binary_path);
cclose(p->dot);
cclose(p->slash);
p->dot = p->slash = 0; /* catch bugs */
- /* can safely free the fgrp, now that no one is accessing it */
- kfree(p->fgrp->fd);
- kfree(p->fgrp);
- kref_put(&p->fs_env.root->d_kref);
- kref_put(&p->fs_env.pwd->d_kref);
/* now we'll finally decref files for the file-backed vmrs */
unmap_and_destroy_vmrs(p);
- frontend_proc_free(p); /* TODO: please remove me one day */
- /* Free any colors allocated to this process */
- if (p->cache_colors_map != global_cache_colors_map) {
- for(int i = 0; i < llc_cache->num_colors; i++)
- cache_color_free(llc_cache, p->cache_colors_map);
- cache_colors_map_free(p->cache_colors_map);
- }
/* Remove us from the pid_hash and give our PID back (in that order). */
spin_lock(&pid_hash_lock);
- if (!hashtable_remove(pid_hash, (void*)(long)p->pid))
- panic("Proc not in the pid table in %s", __FUNCTION__);
+ hash_ret = hashtable_remove(pid_hash, (void*)(long)p->pid);
spin_unlock(&pid_hash_lock);
- put_free_pid(p->pid);
- /* all memory below UMAPTOP should have been freed via the VMRs. the stuff
- * above is the global page and procinfo/procdata */
- env_user_mem_free(p, (void*)UMAPTOP, UVPT - UMAPTOP); /* 3rd arg = len... */
+ /* might not be in the hash/ready, if we failed during proc creation */
+ if (hash_ret)
+ put_free_pid(p->pid);
+ else
+ printd("[kernel] pid %d not in the PID hash in %s\n", p->pid,
+ __FUNCTION__);
+ /* All memory below UMAPTOP should have been freed via the VMRs. The stuff
+ * above is the global info/page and procinfo/procdata. We free procinfo
+ * and procdata, but not the global memory - that's system wide. We could
+ * clear the PTEs of the upper stuff (UMAPTOP to UVPT), but we shouldn't
+ * need to. */
env_user_mem_walk(p, 0, UMAPTOP, __cb_assert_no_pg, 0);
- /* These need to be freed again, since they were allocated with a refcnt. */
- free_cont_pages(p->procinfo, LOG2_UP(PROCINFO_NUM_PAGES));
- free_cont_pages(p->procdata, LOG2_UP(PROCDATA_NUM_PAGES));
+ kpages_free(p->procinfo, PROCINFO_NUM_PAGES * PGSIZE);
+ kpages_free(p->procdata, PROCDATA_NUM_PAGES * PGSIZE);
env_pagetable_free(p);
- p->env_pgdir = 0;
+ arch_pgdir_clear(&p->env_pgdir);
p->env_cr3 = 0;
atomic_dec(&num_envs);
/* We use the pcpui to access 'current' to cut down on the core_id() calls,
* though who know how expensive/painful they are. */
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ struct proc *old_proc;
+
/* If the process wasn't here, then we need to load its address space. */
if (p != pcpui->cur_proc) {
proc_incref(p, 1);
* previous lcr3 unloaded the previous proc's context. This should
* rarely happen, since we usually proactively leave process context,
* but this is the fallback. */
- if (pcpui->cur_proc)
- proc_decref(pcpui->cur_proc);
+ old_proc = pcpui->cur_proc;
pcpui->cur_proc = p;
+ if (old_proc)
+ proc_decref(old_proc);
}
}
}
/* Dispatches a _S process to run on the current core. This should never be
- * called to "restart" a core.
+ * called to "restart" a core.
*
* This will always return, regardless of whether or not the calling core is
* being given to a process. (it used to pop the tf directly, before we had
spin_lock(&p->proc_lock);
switch (p->state) {
case (PROC_DYING):
+ case (PROC_DYING_ABORT):
spin_unlock(&p->proc_lock);
printk("[kernel] _S %d not starting due to async death\n", p->pid);
return;
case (PROC_RUNNABLE_S):
__proc_set_state(p, PROC_RUNNING_S);
- /* We will want to know where this process is running, even if it is
- * only in RUNNING_S. can use the vcoremap, which makes death easy.
- * Also, this is the signal used in trap.c to know to save the tf in
- * scp_ctx. */
+ /* SCPs don't have full vcores, but they act like they have vcore 0.
+ * We map the vcore, since we will want to know where this process
+ * is running, even if it is only in RUNNING_S. We can use the
+ * vcoremap, which makes death easy. num_vcores is still 0, and we
+ * do account the time online and offline. */
__seq_start_write(&p->procinfo->coremap_seqctr);
- p->procinfo->num_vcores = 0; /* TODO (VC#) */
- /* TODO: For now, we won't count this as an active vcore (on the
- * lists). This gets unmapped in resource.c and yield_s, and needs
- * work. */
- __map_vcore(p, 0, coreid); /* not treated like a true vcore */
+ p->procinfo->num_vcores = 0;
+ __map_vcore(p, 0, coreid);
+ vcore_account_online(p, 0);
__seq_end_write(&p->procinfo->coremap_seqctr);
/* incref, since we're saving a reference in owning proc later */
proc_incref(p, 1);
* for now. can simply clear_owning if we want to. */
assert(!pcpui->owning_proc);
pcpui->owning_proc = p;
- pcpui->owning_vcoreid = 0; /* TODO (VC#) */
+ pcpui->owning_vcoreid = 0;
restore_vc_fp_state(vcpd);
/* similar to the old __startcore, start them in vcore context if
* they have notifs and aren't already in vcore context. o/w, start
vcpd->uthread_ctx = p->scp_ctx;
pcpui->cur_ctx = &pcpui->actual_ctx;
memset(pcpui->cur_ctx, 0, sizeof(struct user_context));
- proc_init_ctx(pcpui->cur_ctx, 0, p->env_entry,
- vcpd->transition_stack, vcpd->vcore_tls_desc);
+ proc_init_ctx(pcpui->cur_ctx, 0, vcpd->vcore_entry,
+ vcpd->vcore_stack, vcpd->vcore_tls_desc);
} else {
/* If they have no transition stack, then they can't receive
* events. The most they are getting is a wakeup from the
switch (p->state) {
case (PROC_WAITING):
case (PROC_DYING):
+ case (PROC_DYING_ABORT):
warn("ksched tried to run proc %d in state %s\n", p->pid,
procstate2str(p->state));
return;
* bypass the routine_kmsg check. Interrupts should be off when you call this.
*
* A note on refcnting: this function will not return, and your proc reference
- * will end up stored in current. This will make no changes to p's refcnt, so
- * do your accounting such that there is only the +1 for current. This means if
- * it is already in current (like in the trap return path), don't up it. If
- * it's already in current and you have another reference (like pid2proc or from
- * an IPI), then down it (which is what happens in __startcore()). If it's not
- * in current and you have one reference, like proc_run(non_current_p), then
- * also do nothing. The refcnt for your *p will count for the reference stored
- * in current. */
+ * will be ignored (not decreffed). It may be incref'd, if cur_proc was not
+ * set. Pass in an already-accounted-for ref, such as owning_proc. */
void __proc_startcore(struct proc *p, struct user_context *ctx)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
assert(!irq_is_enabled());
/* Should never have ktask still set. If we do, future syscalls could try
* to block later and lose track of our address space. */
- assert(!pcpui->cur_kthread->is_ktask);
+ assert(!is_ktask(pcpui->cur_kthread));
__set_proc_current(p);
- /* Clear the current_ctx, since it is no longer used */
- current_ctx = 0; /* TODO: might not need this... */
+ __set_cpu_state(pcpui, CPU_STATE_USER);
proc_pop_ctx(ctx);
}
/* Restarts/runs the current_ctx, which must be for the current process, on the
- * core this code executes on. Calls an internal function to do the work.
+ * core this code executes on.
+ *
+ * For now, we just smp_idle. We used to do something similar, but customized
+ * for expecting to return to the process. But it was a source of bugs. If we
+ * want to optimize for the case where we know we had a process current, then we
+ * can do so here.
*
- * In case there are pending routine messages, like __death, __preempt, or
- * __notify, we need to run them. Alternatively, if there are any, we could
- * self_ipi, and run the messages immediately after popping back to userspace,
- * but that would have crappy overhead. */
+ * Note that PRKM currently calls smp_idle() if it ever has a message, so the
+ * value of optimizing may depend on the semantics of PRKM. */
void proc_restartcore(void)
{
- struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- assert(!pcpui->cur_kthread->sysc);
- /* TODO: can probably remove this enable_irq. it was an optimization for
- * RKMs */
- /* Try and get any interrupts before we pop back to userspace. If we didn't
- * do this, we'd just get them in userspace, but this might save us some
- * effort/overhead. */
- enable_irq();
- /* Need ints disabled when we return from PRKM (race on missing
- * messages/IPIs) */
- disable_irq();
- process_routine_kmsg();
- /* If there is no owning process, just idle, since we don't know what to do.
- * This could be because the process had been restarted a long time ago and
- * has since left the core, or due to a KMSG like __preempt or __death. */
- if (!pcpui->owning_proc) {
- abandon_core();
- smp_idle();
+ smp_idle();
+}
+
+/* Helper for proc_destroy. Disowns any children. */
+static void proc_disown_children(struct proc *parent)
+{
+ struct proc *child_i, *temp;
+ struct proc_list todo = TAILQ_HEAD_INITIALIZER(todo);
+ int ret;
+
+ cv_lock(&parent->child_wait);
+ TAILQ_FOREACH_SAFE(child_i, &parent->children, sibling_link, temp) {
+ ret = __proc_disown_child(parent, child_i);
+ /* should never fail, lock should cover the race. invariant: any child
+ * on the list should have us as a parent */
+ assert(!ret);
+ TAILQ_INSERT_TAIL(&todo, child_i, sibling_link);
}
- assert(pcpui->cur_ctx);
- __proc_startcore(pcpui->owning_proc, pcpui->cur_ctx);
+ cv_unlock(&parent->child_wait);
+
+ TAILQ_FOREACH_SAFE(child_i, &todo, sibling_link, temp)
+ proc_decref(child_i);
}
/* Destroys the process. It will destroy the process and return any cores
uint32_t nr_cores_revoked = 0;
struct kthread *sleeper;
struct proc *child_i, *temp;
- /* Can't spin on the proc lock with irq disabled. This is a problem for all
- * places where we grab the lock, but it is particularly bad for destroy,
- * since we tend to call this from trap and irq handlers */
- assert(irq_is_enabled());
+
spin_lock(&p->proc_lock);
/* storage for pc_arr is alloced at decl, which is after grabbing the lock*/
uint32_t pc_arr[p->procinfo->num_vcores];
switch (p->state) {
case PROC_DYING: /* someone else killed this already. */
+ case (PROC_DYING_ABORT):
spin_unlock(&p->proc_lock);
return;
case PROC_CREATED:
// here's how to do it manually
if (current == p) {
lcr3(boot_cr3);
- proc_decref(p); /* this decref is for the cr3 */
current = NULL;
+ proc_decref(p); /* this decref is for the cr3 */
}
#endif
- send_kernel_message(get_pcoreid(p, 0), __death, 0, 0, 0,
+ send_kernel_message(get_pcoreid(p, 0), __death, (long)p, 0, 0,
KMSG_ROUTINE);
__seq_start_write(&p->procinfo->coremap_seqctr);
- // TODO: might need to sort num_vcores too later (VC#)
- /* vcore is unmapped on the receive side */
+ __unmap_vcore(p, 0);
__seq_end_write(&p->procinfo->coremap_seqctr);
/* If we ever have RUNNING_S run on non-mgmt cores, we'll need to
* tell the ksched about this now-idle core (after unlocking) */
* interrupts should be on when you call proc_destroy locally, but currently
* aren't for all things (like traphandlers). */
__proc_set_state(p, PROC_DYING);
- /* Disown any children. If we want to have init inherit or something,
- * change __disown to set the ppid accordingly and concat this with init's
- * list (instead of emptying it like disown does). Careful of lock ordering
- * between procs (need to lock to protect lists) */
- TAILQ_FOREACH_SAFE(child_i, &p->children, sibling_link, temp) {
- int ret = __proc_disown_child(p, child_i);
- /* should never fail, lock should cover the race. invariant: any child
- * on the list should have us as a parent */
- assert(!ret);
- }
spin_unlock(&p->proc_lock);
+ proc_disown_children(p);
/* Wake any of our kthreads waiting on children, so they can abort */
cv_broadcast(&p->child_wait);
- /* Abort any abortable syscalls. This won't catch every sleeper, but future
- * abortable sleepers are already prevented via the DYING state. (signalled
- * DYING, no new sleepers will block, and now we wake all old sleepers). */
- abort_all_sysc(p);
/* we need to close files here, and not in free, since we could have a
* refcnt indirectly related to one of our files. specifically, if we have
* a parent sleeping on our pipe, that parent won't wake up to decref until
* the pipe closes. And if the parent doesnt decref, we don't free.
- * alternatively, we could send a SIGCHILD to the parent, but that would
- * require parent's to never ignore that signal (or risk never reaping).
+ * Even if we send a SIGCHLD to the parent, that would require that the
+ * parent to never ignores that signal (or we risk never reaping).
*
* Also note that any mmap'd files will still be mmapped. You can close the
* file after mmapping, with no effect. */
- close_9ns_files(p, FALSE);
- close_all_files(&p->open_files, FALSE);
+ close_fdt(&p->open_files, FALSE);
+ /* Abort any abortable syscalls. This won't catch every sleeper, but future
+ * abortable sleepers are already prevented via the DYING_ABORT state.
+ * (signalled DYING_ABORT, no new sleepers will block, and now we wake all
+ * old sleepers). */
+ __proc_set_state(p, PROC_DYING_ABORT);
+ abort_all_sysc(p);
/* Tell the ksched about our death, and which cores we freed up */
__sched_proc_destroy(p, pc_arr, nr_cores_revoked);
/* Tell our parent about our state change (to DYING) */
}
/* Can use this to signal anything that might cause a parent to wait on the
- * child, such as termination, or (in the future) signals. Change the state or
- * whatever before calling. */
+ * child, such as termination, or signals. Change the state or whatever before
+ * calling. */
void proc_signal_parent(struct proc *child)
{
struct kthread *sleeper;
struct proc *parent = pid2proc(child->ppid);
if (!parent)
return;
+ send_posix_signal(parent, SIGCHLD);
/* there could be multiple kthreads sleeping for various reasons. even an
* SCP could have multiple async syscalls. */
cv_broadcast(&parent->child_wait);
/* Called when a parent is done with its child, and no longer wants to track the
* child, nor to allow the child to track it. Call with a lock (cv) held.
- * Returns 0 if we disowned, -1 on failure. */
+ * Returns 0 if we disowned, -1 on failure.
+ *
+ * If we disowned, (ret == 0), the caller must decref the child. */
int __proc_disown_child(struct proc *parent, struct proc *child)
{
/* Bail out if the child has already been reaped */
/* After this, the child won't be able to get more refs to us, but it may
* still have some references in running code. */
child->ppid = 0;
- proc_decref(child); /* ref that was keeping the child alive after dying */
return 0;
}
switch (p->state) {
case (PROC_RUNNING_S):
/* issue with if we're async or not (need to preempt it)
- * either of these should trip it. TODO: (ACR) async core req
- * TODO: relies on vcore0 being the caller (VC#) */
+ * either of these should trip it. TODO: (ACR) async core req */
if ((current != p) || (get_pcoreid(p, 0) != core_id()))
panic("We don't handle async RUNNING_S core requests yet.");
struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0];
assert(current_ctx);
/* Copy uthread0's context to VC 0's uthread slot */
- vcpd->uthread_ctx = *current_ctx;
+ copy_current_ctx_to(&vcpd->uthread_ctx);
clear_owning_proc(core_id()); /* so we don't restart */
save_vc_fp_state(vcpd);
/* Userspace needs to not fuck with notif_disabled before
/* this process no longer runs on its old location (which is
* this core, for now, since we don't handle async calls) */
__seq_start_write(&p->procinfo->coremap_seqctr);
- // TODO: (VC#) might need to adjust num_vcores
// TODO: (ACR) will need to unmap remotely (receive-side)
- __unmap_vcore(p, 0); /* VC# keep in sync with proc_run_s */
+ __unmap_vcore(p, 0);
+ vcore_account_offline(p, 0);
__seq_end_write(&p->procinfo->coremap_seqctr);
/* change to runnable_m (it's TF is already saved) */
__proc_set_state(p, PROC_RUNNABLE_M);
warn("Not supporting RUNNABLE_S -> RUNNABLE_M yet.\n");
goto error_out;
case (PROC_DYING):
+ case (PROC_DYING_ABORT):
warn("Dying, core request coming from %d\n", core_id());
goto error_out;
default:
{
struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0];
uint32_t num_revoked;
+ /* Not handling vcore accounting. Do so if we ever use this */
printk("[kernel] trying to transition _M -> _S (deprecated)!\n");
assert(p->state == PROC_RUNNING_M); // TODO: (ACR) async core req
/* save the context, to be restarted in _S mode */
assert(current_ctx);
- p->scp_ctx = *current_ctx;
+ copy_current_ctx_to(&p->scp_ctx);
clear_owning_proc(core_id()); /* so we don't restart */
save_vc_fp_state(vcpd);
/* sending death, since it's not our job to save contexts or anything in
* In the future, we'll probably use vc0's space for scp_ctx and the silly
* state. If we ever do that, we'll need to stop using scp_ctx (soon to be in
* VCPD) as a location for pcpui->cur_ctx to point (dangerous) */
-void __proc_save_context_s(struct proc *p, struct user_context *ctx)
+void __proc_save_context_s(struct proc *p)
{
- p->scp_ctx = *ctx;
- __unmap_vcore(p, 0); /* VC# keep in sync with proc_run_s */
+ copy_current_ctx_to(&p->scp_ctx);
+ __seq_start_write(&p->procinfo->coremap_seqctr);
+ __unmap_vcore(p, 0);
+ __seq_end_write(&p->procinfo->coremap_seqctr);
+ vcore_account_offline(p, 0);
}
/* Yields the calling core. Must be called locally (not async) for now.
* We disable interrupts for most of it too, since we need to protect
* current_ctx and not race with __notify (which doesn't play well with
* concurrent yielders). */
-void proc_yield(struct proc *SAFE p, bool being_nice)
+void proc_yield(struct proc *p, bool being_nice)
{
uint32_t vcoreid, pcoreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[pcoreid];
* WAITING. one (or both) of us will see and make sure the proc
* wakes up. */
__proc_set_state(p, PROC_WAITING);
- wrmb(); /* don't let the state write pass the notif read */
+ wrmb(); /* don't let the state write pass the notif read */
if (vcpd->notif_pending) {
__proc_set_state(p, PROC_RUNNING_S);
/* they can't handle events, just need to prevent a yield.
/* if we're here, we want to sleep. a concurrent event that
* hasn't already written notif_pending will have seen WAITING,
* and will be spinning while we do this. */
- __proc_save_context_s(p, current_ctx);
+ __proc_save_context_s(p);
spin_unlock(&p->proc_lock);
} else {
/* yielding to allow other processes to run. we're briefly
* WAITING, til we are woken up */
__proc_set_state(p, PROC_WAITING);
- __proc_save_context_s(p, current_ctx);
+ __proc_save_context_s(p);
spin_unlock(&p->proc_lock);
/* immediately wake up the proc (makes it runnable) */
proc_wakeup(p);
case (PROC_RUNNING_M):
break; /* will handle this stuff below */
case (PROC_DYING): /* incoming __death */
+ case (PROC_DYING_ABORT):
case (PROC_RUNNABLE_M): /* incoming (bulk) preempt/myield TODO:(BULK) */
goto out_failed;
default:
p->procinfo->num_vcores--;
p->procinfo->res_grant[RES_CORES] = p->procinfo->num_vcores;
__seq_end_write(&p->procinfo->coremap_seqctr);
+ vcore_account_offline(p, vcoreid);
/* No more vcores? Then we wait on an event */
if (p->procinfo->num_vcores == 0) {
/* consider a ksched op to tell it about us WAITING */
__proc_set_state(p, PROC_WAITING);
}
spin_unlock(&p->proc_lock);
+ /* We discard the current context, but we still need to restore the core */
+ arch_finalize_ctx(pcpui->cur_ctx);
/* Hand the now-idle core to the ksched */
__sched_put_idle_core(p, pcoreid);
goto out_yield_core;
void proc_notify(struct proc *p, uint32_t vcoreid)
{
struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[vcoreid];
+
+ /* If you're thinking about checking notif_pending and then returning if it
+ * is already set, note that some callers (e.g. the event system) set
+ * notif_pending when they deliver a message, regardless of whether there is
+ * an IPI or not. Those callers assume that we don't care about
+ * notif_pending, only notif_disabled. So don't change this without
+ * changing them (probably can't without a lot of thought - that
+ * notif_pending is about missing messages. It might be possible to say "no
+ * IPI, but don't let me miss messages that were delivered." */
vcpd->notif_pending = TRUE;
wrmb(); /* must write notif_pending before reading notif_disabled */
if (!vcpd->notif_disabled) {
case (PROC_RUNNABLE_S):
case (PROC_RUNNING_S):
case (PROC_DYING):
+ case (PROC_DYING_ABORT):
spin_unlock(&p->proc_lock);
return;
case (PROC_RUNNABLE_M):
return p->procinfo->is_mcp;
}
+bool proc_is_vcctx_ready(struct proc *p)
+{
+ struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0];
+ return scp_is_vcctx_ready(vcpd);
+}
+
/************************ Preemption Functions ******************************
* Don't rely on these much - I'll be sure to change them up a bit.
*
for (int i = 0; i < num; i++) {
assert(__proc_give_a_pcore(p, pc_arr[i], &p->inactive_vcs, &vc_i));
send_kernel_message(pc_arr[i], __startcore, (long)p,
- (long)vcore2vcoreid(p, vc_i),
+ (long)vcore2vcoreid(p, vc_i),
(long)vc_i->nr_preempts_sent, KMSG_ROUTINE);
}
__seq_end_write(&p->procinfo->coremap_seqctr);
int __proc_give_cores(struct proc *p, uint32_t *pc_arr, uint32_t num)
{
/* should never happen: */
- assert(num + p->procinfo->num_vcores <= MAX_NUM_CPUS);
+ assert(num + p->procinfo->num_vcores <= MAX_NUM_CORES);
switch (p->state) {
case (PROC_RUNNABLE_S):
case (PROC_RUNNING_S):
warn("Don't give cores to a process in a *_S state!\n");
return -1;
case (PROC_DYING):
+ case (PROC_DYING_ABORT):
case (PROC_WAITING):
/* can't accept, just fail */
return -1;
atomic_or(&vcpd->flags, VC_K_LOCK);
send_kernel_message(pcoreid, __preempt, (long)p, 0, 0, KMSG_ROUTINE);
} else {
- send_kernel_message(pcoreid, __death, 0, 0, 0, KMSG_ROUTINE);
+ send_kernel_message(pcoreid, __death, (long)p, 0, 0, KMSG_ROUTINE);
}
}
* Note this leaves no trace of what was running. This "leaves the process's
* context.
*
- * This does not clear the owning proc. Use the other helper for that. */
-void abandon_core(void)
+ * This does not clear the owning proc. Use the other helper for that.
+ *
+ * Returns whether or not there was a process present. */
+bool abandon_core(void)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
/* Syscalls that don't return will ultimately call abadon_core(), so we need
* to make sure we don't think we are still working on a syscall. */
pcpui->cur_kthread->sysc = 0;
pcpui->cur_kthread->errbuf = 0; /* just in case */
- if (pcpui->cur_proc)
+ if (pcpui->cur_proc) {
__abandon_core();
+ return true;
+ }
+ return false;
}
/* Helper to clear the core's owning processor and manage refcnting. Pass in
{
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
struct proc *p = pcpui->owning_proc;
+
+ __clear_owning_proc(coreid);
pcpui->owning_proc = 0;
pcpui->owning_vcoreid = 0xdeadbeef;
pcpui->cur_ctx = 0; /* catch bugs for now (may go away) */
/* Switches to the address space/context of new_p, doing nothing if we are
* already in new_p. This won't add extra refcnts or anything, and needs to be
- * paired with switch_back() at the end of whatever function you are in. Don't
- * migrate cores in the middle of a pair. Specifically, the uncounted refs are
- * one for the old_proc, which is passed back to the caller, and new_p is
- * getting placed in cur_proc. */
-struct proc *switch_to(struct proc *new_p)
+ * paired with switch_back() at the end of whatever function you are in.
+ * Specifically, the uncounted refs are one for the old_proc, which is passed
+ * back to the caller, and new_p is getting placed in cur_proc. */
+uintptr_t switch_to(struct proc *new_p)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ struct kthread *kth = pcpui->cur_kthread;
struct proc *old_proc;
+ uintptr_t ret;
+
old_proc = pcpui->cur_proc; /* uncounted ref */
/* If we aren't the proc already, then switch to it */
if (old_proc != new_p) {
pcpui->cur_proc = new_p; /* uncounted ref */
- lcr3(new_p->env_cr3);
+ if (new_p)
+ lcr3(new_p->env_cr3);
+ else
+ lcr3(boot_cr3);
+ }
+ ret = (uintptr_t)old_proc;
+ if (is_ktask(kth)) {
+ if (!(kth->flags & KTH_SAVE_ADDR_SPACE)) {
+ kth->flags |= KTH_SAVE_ADDR_SPACE;
+ /* proc pointers are aligned; we can use the lower bit as a signal
+ * to turn off SAVE_ADDR_SPACE. */
+ ret |= 0x1;
+ }
}
- return old_proc;
+ return ret;
}
-/* This switches back to old_proc from new_p. Pair it with switch_to(), and
- * pass in its return value for old_proc. */
-void switch_back(struct proc *new_p, struct proc *old_proc)
+/* This switches back from new_p to the original process. Pair it with
+ * switch_to(), and pass in its return value for old_ret. */
+void switch_back(struct proc *new_p, uintptr_t old_ret)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ struct kthread *kth = pcpui->cur_kthread;
+ struct proc *old_proc;
+
+ if (is_ktask(kth)) {
+ if (old_ret & 0x1) {
+ kth->flags &= ~KTH_SAVE_ADDR_SPACE;
+ old_ret &= ~0x1;
+ }
+ }
+ old_proc = (struct proc*)old_ret;
if (old_proc != new_p) {
pcpui->cur_proc = old_proc;
if (old_proc)
* immediate message. */
void proc_tlbshootdown(struct proc *p, uintptr_t start, uintptr_t end)
{
+ /* TODO: need a better way to find cores running our address space. we can
+ * have kthreads running syscalls, async calls, processes being created. */
struct vcore *vc_i;
/* TODO: we might be able to avoid locking here in the future (we must hit
* all online, and we can check __mapped). it'll be complicated. */
tlbflush();
break;
case (PROC_RUNNING_M):
- /* TODO: (TLB) sanity checks and rounding on the ranges */
+ /* TODO: (TLB) sanity checks and rounding on the ranges.
+ *
+ * We need to make sure that once a core that was online has been
+ * removed from the online list, then it must receive a TLB flush
+ * (abandon_core()) before running the process again. Either that,
+ * or make other decisions about who to TLB-shootdown. */
TAILQ_FOREACH(vc_i, &p->online_vcs, list) {
send_kernel_message(vc_i->pcoreid, __tlbshootdown, start, end,
0, KMSG_IMMEDIATE);
}
break;
- case (PROC_DYING):
- /* if it is dying, death messages are already on the way to all
- * cores, including ours, which will clear the TLB. */
- break;
default:
- /* will probably get this when we have the short handlers */
- warn("Unexpected case %s in %s", procstate2str(p->state),
- __FUNCTION__);
+ /* TODO: til we fix shootdowns, there are some odd cases where we
+ * have the address space loaded, but the state is in transition. */
+ if (p == current)
+ tlbflush();
}
spin_unlock(&p->proc_lock);
}
pcpui->actual_ctx = vcpd->vcore_ctx;
proc_secure_ctx(&pcpui->actual_ctx);
} else { /* not restarting from a preemption, use a fresh vcore */
- assert(vcpd->transition_stack);
- proc_init_ctx(&pcpui->actual_ctx, vcoreid, p->env_entry,
- vcpd->transition_stack, vcpd->vcore_tls_desc);
+ assert(vcpd->vcore_stack);
+ proc_init_ctx(&pcpui->actual_ctx, vcoreid, vcpd->vcore_entry,
+ vcpd->vcore_stack, vcpd->vcore_tls_desc);
/* Disable/mask active notifications for fresh vcores */
vcpd->notif_disabled = TRUE;
}
/* cur_ctx was built above (in actual_ctx), now use it */
pcpui->cur_ctx = &pcpui->actual_ctx;
/* this cur_ctx will get run when the kernel returns / idles */
+ vcore_account_online(p, vcoreid);
}
/* Changes calling vcore to be vcoreid. enable_my_notif tells us about how the
break; /* the only case we can proceed */
case (PROC_RUNNING_S): /* user bug, just return */
case (PROC_DYING): /* incoming __death */
+ case (PROC_DYING_ABORT):
case (PROC_RUNNABLE_M): /* incoming (bulk) preempt/myield TODO:(BULK) */
goto out_locked;
default:
* and we don't care about either the uthread_ctx or the vcore_ctx. */
caller_vcpd->notif_disabled = FALSE;
/* Don't need to save the FPU. There should be no uthread or other
- * reason to return to the FPU state. */
+ * reason to return to the FPU state. But we do need to finalize the
+ * context, even though we are throwing it away. We need to return the
+ * pcore to a state where it can run any context and not be bound to
+ * the old context. */
+ arch_finalize_ctx(pcpui->cur_ctx);
} else {
/* need to set up the calling vcore's ctx so that it'll get restarted by
* __startcore, to make the caller look like it was preempted. */
- caller_vcpd->vcore_ctx = *current_ctx;
+ copy_current_ctx_to(&caller_vcpd->vcore_ctx);
save_vc_fp_state(caller_vcpd);
- /* Mark our core as preempted (for userspace recovery). */
- atomic_or(&caller_vcpd->flags, VC_PREEMPTED);
}
+ /* Mark our core as preempted (for userspace recovery). Userspace checks
+ * this in handle_indirs, and it needs to check the mbox regardless of
+ * enable_my_notif. This does mean cores that change-to with no intent to
+ * return will be tracked as PREEMPTED until they start back up (maybe
+ * forever). */
+ atomic_or(&caller_vcpd->flags, VC_PREEMPTED);
/* Either way, unmap and offline our current vcore */
/* Move the caller from online to inactive */
TAILQ_REMOVE(&p->online_vcs, caller_vc, list);
__unmap_vcore(p, caller_vcoreid);
__map_vcore(p, new_vcoreid, pcoreid);
__seq_end_write(&p->procinfo->coremap_seqctr);
+ vcore_account_offline(p, caller_vcoreid);
/* Send either a PREEMPT msg or a CHECK_MSGS msg. If they said to
* enable_my_notif, then all userspace needs is to check messages, not a
* full preemption recovery. */
uint32_t vcoreid = (uint32_t)a1;
uint32_t coreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
- struct proc *p_to_run = (struct proc *CT(1))a0;
+ struct proc *p_to_run = (struct proc *)a0;
uint32_t old_nr_preempts_sent = (uint32_t)a2;
assert(p_to_run);
vcpd->notif_disabled = TRUE;
/* save the old ctx in the uthread slot, build and pop a new one. Note that
* silly state isn't our business for a notification. */
- vcpd->uthread_ctx = *pcpui->cur_ctx;
+ copy_current_ctx_to(&vcpd->uthread_ctx);
memset(pcpui->cur_ctx, 0, sizeof(struct user_context));
- proc_init_ctx(pcpui->cur_ctx, vcoreid, p->env_entry,
- vcpd->transition_stack, vcpd->vcore_tls_desc);
+ proc_init_ctx(pcpui->cur_ctx, vcoreid, vcpd->vcore_entry,
+ vcpd->vcore_stack, vcpd->vcore_tls_desc);
/* this cur_ctx will get run when the kernel returns / idles */
}
* cur_ctx in the uthread slot, and it'll appear to the vcore when it comes
* back up the uthread just took a notification. */
if (vcpd->notif_disabled)
- vcpd->vcore_ctx = *pcpui->cur_ctx;
+ copy_current_ctx_to(&vcpd->vcore_ctx);
else
- vcpd->uthread_ctx = *pcpui->cur_ctx;
+ copy_current_ctx_to(&vcpd->uthread_ctx);
/* Userspace in a preemption handler on another core might be copying FP
* state from memory (VCPD) at the moment, and if so we don't want to
* clobber it. In this rare case, our current core's FPU state should be
atomic_and(&vcpd->flags, ~VC_K_LOCK);
/* either __preempt or proc_yield() ends the preempt phase. */
p->procinfo->vcoremap[vcoreid].preempt_pending = 0;
+ vcore_account_offline(p, vcoreid);
wmb(); /* make sure everything else hits before we finish the preempt */
/* up the nr_done, which signals the next __startcore for this vc */
p->procinfo->vcoremap[vcoreid].nr_preempts_done++;
{
uint32_t vcoreid, coreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
- struct proc *p = pcpui->owning_proc;
- if (p) {
- vcoreid = pcpui->owning_vcoreid;
- printd("[kernel] death on physical core %d for process %d's vcore %d\n",
- coreid, p->pid, vcoreid);
- /* We won't restart the process later. current gets cleared later when
- * we notice there is no owning_proc and we have nothing to do
- * (smp_idle, restartcore, etc) */
- clear_owning_proc(coreid);
+ struct proc *p = (struct proc*)a0;
+
+ assert(p);
+ if (p != pcpui->owning_proc) {
+ /* Older versions of Akaros thought it was OK to have a __death hit a
+ * core that no longer had a process. I think it's a bug now. */
+ panic("__death arrived for a process (%p) that was not owning (%p)!",
+ p, pcpui->owning_proc);
}
+ vcoreid = pcpui->owning_vcoreid;
+ printd("[kernel] death on physical core %d for process %d's vcore %d\n",
+ coreid, p->pid, vcoreid);
+ vcore_account_offline(p, vcoreid); /* in case anyone is counting */
+ /* We won't restart the process later. current gets cleared later when
+ * we notice there is no owning_proc and we have nothing to do
+ * (smp_idle, restartcore, etc). */
+ arch_finalize_ctx(pcpui->cur_ctx);
+ clear_owning_proc(coreid);
}
/* Kernel message handler, usually sent IMMEDIATE, to shoot down virtual
void print_allpids(void)
{
- void print_proc_state(void *item)
+ void print_proc_state(void *item, void *opaque)
{
struct proc *p = (struct proc*)item;
assert(p);
- printk("%8d %-10s %6d\n", p->pid, procstate2str(p->state), p->ppid);
+ /* this actually adds an extra space, since no progname is ever
+ * PROGNAME_SZ bytes, due to the \0 counted in PROGNAME. */
+ printk("%8d %-*s %-10s %6d\n", p->pid, PROC_PROGNAME_SZ, p->progname,
+ procstate2str(p->state), p->ppid);
}
- printk(" PID STATE Parent \n");
- printk("------------------------------\n");
+ char dashes[PROC_PROGNAME_SZ];
+ memset(dashes, '-', PROC_PROGNAME_SZ);
+ dashes[PROC_PROGNAME_SZ - 1] = '\0';
+ /* -5, for 'Name ' */
+ printk(" PID Name %-*s State Parent \n",
+ PROC_PROGNAME_SZ - 5, "");
+ printk("------------------------------%s\n", dashes);
spin_lock(&pid_hash_lock);
- hash_for_each(pid_hash, print_proc_state);
+ hash_for_each(pid_hash, print_proc_state, NULL);
spin_unlock(&pid_hash_lock);
}
-void print_proc_info(pid_t pid)
+void proc_get_set(struct process_set *pset)
+{
+ void enum_proc(void *item, void *opaque)
+ {
+ struct proc *p = (struct proc*) item;
+ struct process_set *pset = (struct process_set *) opaque;
+
+ if (pset->num_processes < pset->size) {
+ proc_incref(p, 1);
+
+ pset->procs[pset->num_processes] = p;
+ pset->num_processes++;
+ }
+ }
+
+ static const size_t num_extra_alloc = 16;
+
+ pset->procs = NULL;
+ do {
+ if (pset->procs)
+ proc_free_set(pset);
+ pset->size = atomic_read(&num_envs) + num_extra_alloc;
+ pset->num_processes = 0;
+ pset->procs = (struct proc **)
+ kzmalloc(pset->size * sizeof(struct proc *), MEM_WAIT);
+ if (!pset->procs)
+ error(-ENOMEM, ERROR_FIXME);
+
+ spin_lock(&pid_hash_lock);
+ hash_for_each(pid_hash, enum_proc, pset);
+ spin_unlock(&pid_hash_lock);
+
+ } while (pset->num_processes == pset->size);
+}
+
+void proc_free_set(struct process_set *pset)
+{
+ for (size_t i = 0; i < pset->num_processes; i++)
+ proc_decref(pset->procs[i]);
+ kfree(pset->procs);
+}
+
+void print_proc_info(pid_t pid, int verbosity)
{
int j = 0;
+ uint64_t total_time = 0;
struct proc *child, *p = pid2proc(pid);
struct vcore *vc_i;
+ struct preempt_data *vcpd;
+
if (!p) {
printk("Bad PID.\n");
return;
}
+ vcpd = &p->procdata->vcore_preempt_data[0];
+ print_lock();
spinlock_debug(&p->proc_lock);
//spin_lock(&p->proc_lock); // No locking!!
printk("struct proc: %p\n", p);
+ printk("Program name: %s\n", p->progname);
printk("PID: %d\n", p->pid);
printk("PPID: %d\n", p->ppid);
printk("State: %s (%p)\n", procstate2str(p->state), p->state);
+ printk("\tIs %san MCP\n", p->procinfo->is_mcp ? "" : "not ");
+ if (!scp_is_vcctx_ready(vcpd))
+ printk("\tIs NOT vcctx ready\n");
+ if (verbosity > 0 && !p->procinfo->is_mcp) {
+ printk("Last saved SCP context:");
+ backtrace_user_ctx(p, &p->scp_ctx);
+ }
printk("Refcnt: %d\n", atomic_read(&p->p_kref.refcount) - 1);
printk("Flags: 0x%08x\n", p->env_flags);
printk("CR3(phys): %p\n", p->env_cr3);
printk("Inactive / Yielded:\n");
TAILQ_FOREACH(vc_i, &p->inactive_vcs, list)
printk("\tVcore %d\n", vcore2vcoreid(p, vc_i));
+ if (verbosity > 0) {
+ printk("Nsec Online, up to the last offlining:\n");
+ printk("------------------------");
+ for (int i = 0; i < p->procinfo->max_vcores; i++) {
+ uint64_t vc_time = tsc2nsec(vcore_account_gettotal(p, i));
+
+ if (i % 4 == 0)
+ printk("\n");
+ printk(" VC %3d: %14llu", i, vc_time);
+ total_time += vc_time;
+ }
+ printk("\n");
+ printk("Total CPU-NSEC: %llu\n", total_time);
+ }
printk("Resources:\n------------------------\n");
for (int i = 0; i < MAX_NUM_RESOURCES; i++)
printk("\tRes type: %02d, amt wanted: %08d, amt granted: %08d\n", i,
p->procdata->res_req[i].amt_wanted, p->procinfo->res_grant[i]);
printk("Open Files:\n");
- struct files_struct *files = &p->open_files;
+ struct fd_table *files = &p->open_files;
+ if (spin_locked(&files->lock)) {
+ spinlock_debug(&files->lock);
+ printk("FILE LOCK HELD, ABORTING\n");
+ print_unlock();
+ proc_decref(p);
+ return;
+ }
spin_lock(&files->lock);
- for (int i = 0; i < files->max_files; i++)
- if (files->fd[i].fd_file) {
- printk("\tFD: %02d, File: %p, File name: %s\n", i,
- files->fd[i].fd_file, file_name(files->fd[i].fd_file));
+ for (int i = 0; i < files->max_files; i++) {
+ if (GET_BITMASK_BIT(files->open_fds->fds_bits, i)) {
+ printk("\tFD: %02d, ", i);
+ assert(files->fd[i].fd_chan);
+ print_chaninfo(files->fd[i].fd_chan);
}
+ }
spin_unlock(&files->lock);
- print_9ns_files(p);
printk("Children: (PID (struct proc *))\n");
TAILQ_FOREACH(child, &p->children, sibling_link)
printk("\t%d (%p)\n", child->pid, child);
+ print_unlock();
/* no locking / unlocking or refcnting */
// spin_unlock(&p->proc_lock);
proc_decref(p);
void check_my_owner(void)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- void shazbot(void *item)
+ void shazbot(void *item, void *opaque)
{
struct proc *p = (struct proc*)item;
struct vcore *vc_i;
spin_unlock(&p->proc_lock);
}
assert(!irq_is_enabled());
- extern int booting;
if (!booting && !pcpui->owning_proc) {
spin_lock(&pid_hash_lock);
- hash_for_each(pid_hash, shazbot);
+ hash_for_each(pid_hash, shazbot, NULL);
spin_unlock(&pid_hash_lock);
}
}
-
-/* Use this via kfunc */
-void print_9ns(void)
-{
- void print_proc_9ns(void *item)
- {
- struct proc *p = (struct proc*)item;
- print_9ns_files(p);
- }
- spin_lock(&pid_hash_lock);
- hash_for_each(pid_hash, print_proc_9ns);
- spin_unlock(&pid_hash_lock);
-}
projects / akaros.git / blobdiff