* 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 <elf.h>
#include <arsc_server.h>
#include <devfs.h>
+#include <kmalloc.h>
struct kmem_cache *proc_cache;
/* Other helpers, implemented later. */
-static void __proc_startcore(struct proc *p, trapframe_t *tf);
static bool is_mapped_vcore(struct proc *p, uint32_t pcoreid);
static uint32_t get_vcoreid(struct proc *p, uint32_t pcoreid);
static uint32_t try_get_pcoreid(struct proc *p, uint32_t vcoreid);
static uint32_t get_pcoreid(struct proc *p, uint32_t vcoreid);
static void __proc_free(struct kref *kref);
static bool scp_is_vcctx_ready(struct preempt_data *vcpd);
+static void save_vc_fp_state(struct preempt_data *vcpd);
+static void restore_vc_fp_state(struct preempt_data *vcpd);
/* PID management. */
#define PID_MAX 32767 // goes from 0 to 32767, with 0 reserved
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. */
* RGS -> W
* RGM -> W
* W -> RBS
+ * W -> RGS
* W -> RBM
+ * W -> D
* RGS -> RBM
* RBM -> RGM
* RGM -> RBM
panic("Invalid State Transition! PROC_RUNNING_S to %02x", state);
break;
case PROC_WAITING:
- if (!(state & (PROC_RUNNABLE_S | PROC_RUNNABLE_M)))
+ if (!(state & (PROC_RUNNABLE_S | PROC_RUNNING_S | PROC_RUNNABLE_M |
+ PROC_DYING)))
panic("Invalid State Transition! PROC_WAITING to %02x", state);
break;
case PROC_DYING:
return p;
}
+/* Used by devproc for successive reads of the proc table.
+ * Returns a pointer to the nth proc, or 0 if there is none.
+ * This uses get_not_zero, since it is possible the refcnt is 0, which means the
+ * process is dying and we should not have the ref (and thus return 0). We need
+ * to lock to protect us from getting p, (someone else removes and frees p),
+ * then get_not_zero() on p.
+ * Don't push the locking into the hashtable without dealing with this. */
+struct proc *pid_nth(unsigned int n)
+{
+ struct proc *p;
+ spin_lock(&pid_hash_lock);
+ if (!hashtable_count(pid_hash)) {
+ spin_unlock(&pid_hash_lock);
+ return NULL;
+ }
+ struct hashtable_itr *iter = hashtable_iterator(pid_hash);
+ p = hashtable_iterator_value(iter);
+
+ while (p) {
+ /* if this process is not valid, it doesn't count,
+ * so continue
+ */
+
+ if (kref_get_not_zero(&p->p_kref, 1)){
+ /* this one counts */
+ if (! n){
+ printd("pid_nth: at end, p %p\n", p);
+ break;
+ }
+ kref_put(&p->p_kref);
+ n--;
+ }
+ if (!hashtable_iterator_advance(iter)){
+ p = NULL;
+ break;
+ }
+ p = hashtable_iterator_value(iter);
+ }
+
+ spin_unlock(&pid_hash_lock);
+ kfree(iter);
+ return p;
+}
+
/* Performs any initialization related to processes, such as create the proc
* cache, prep the scheduler, etc. When this returns, we should be ready to use
* any process related function. */
/* 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(HW_CACHE_ALIGN, __alignof__(struct proc)), 0, 0, 0);
+ MAX(ARCH_CL_SIZE, __alignof__(struct proc)), 0, 0, 0);
/* 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_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->heap_bottom = (void*)UTEXT;
+ p->procinfo->timing_overhead = system_timing.timing_overhead;
+ p->procinfo->heap_bottom = 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
* 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;
-
- { INITSTRUCT(*p)
+ /* zero everything by default, other specific items are set below */
+ memset(p, 0, sizeof(*p));
/* only one ref, which we pass back. the old 'existence' ref is managed by
* the ksched */
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);
p->exitcode = 1337; /* so we can see processes killed by the kernel */
- init_sem(&p->state_change, 0);
- p->ppid = parent ? parent->pid : 0;
+ if (parent) {
+ p->ppid = parent->pid;
+ 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;
+ }
+ 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 = (void*)UTEXT; /* heap_bottom set in proc_init_procinfo */
- memset(&p->env_ancillary_state, 0, sizeof(p->env_ancillary_state));
- memset(&p->env_tf, 0, sizeof(p->env_tf));
- spinlock_init(&p->mm_lock);
+ 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 */
- /* 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. */
+ 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. */
TAILQ_INIT(&p->online_vcs);
TAILQ_INIT(&p->bulk_preempted_vcs);
TAILQ_INIT(&p->inactive_vcs);
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 */
+ int fd;
+ fd = insert_file(&p->open_files, dev_stdin, 0, TRUE, FALSE);
+ assert(fd == 0);
+ fd = insert_file(&p->open_files, dev_stdout, 1, TRUE, FALSE);
+ assert(fd == 1);
+ fd = insert_file(&p->open_files, dev_stderr, 2, TRUE, FALSE);
+ assert(fd == 2);
+ }
/* Init the ucq hash lock */
p->ucq_hashlock = (struct hashlock*)&p->ucq_hl_noref;
- hashlock_init(p->ucq_hashlock, HASHLOCK_DEFAULT_SZ);
+ hashlock_init_irqsave(p->ucq_hashlock, HASHLOCK_DEFAULT_SZ);
atomic_inc(&num_envs);
frontend_proc_init(p);
+ 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));
+ qlock_init(&p->vmm.qlock);
printd("[%08x] new process %08x\n", current ? current->pid : 0, p->pid);
- } // INIT_STRUCT
*pp = p;
return 0;
}
* push setting the state to CREATED into here. */
void __proc_ready(struct proc *p)
{
- /* Tell the ksched about us */
- register_proc(p);
+ /* Tell the ksched about us. TODO: do we need to worry about the ksched
+ * doing stuff to us before we're added to the pid_hash? */
+ __sched_proc_register(p);
spin_lock(&pid_hash_lock);
hashtable_insert(pid_hash, (void*)(long)p->pid, p);
spin_unlock(&pid_hash_lock);
{
struct proc *p;
error_t r;
- if ((r = proc_alloc(&p, current)) < 0)
+ if ((r = proc_alloc(&p, current, 0 /* flags */)) < 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);
+ 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)
+{
+ assert(pte_is_unmapped(pte));
+ return 0;
+}
+
/* This is called by kref_put(), once the last reference to the process is
* gone. Don't call this otherwise (it will panic). It will clean up the
* address space and deallocate any other used memory. */
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);
// All parts of the kernel should have decref'd before __proc_free is called
assert(kref_refcnt(&p->p_kref) == 0);
-
+ assert(TAILQ_EMPTY(&p->alarmset.list));
+
+ __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 */
kref_put(&p->fs_env.root->d_kref);
kref_put(&p->fs_env.pwd->d_kref);
- destroy_vmrs(p);
+ /* 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) {
}
/* 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);
- /* Flush all mapped pages in the user portion of the address space */
- env_user_mem_free(p, 0, UVPT);
- /* These need to be free again, since they were allocated with a refcnt. */
+ /* 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 page and procinfo/procdata */
+ env_user_mem_free(p, (void*)UMAPTOP, UVPT - UMAPTOP); /* 3rd arg = len... */
+ 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));
env_pagetable_free(p);
- p->env_pgdir = 0;
+ arch_pgdir_clear(&p->env_pgdir);
p->env_cr3 = 0;
atomic_dec(&num_envs);
kmem_cache_free(proc_cache, p);
}
-/* Whether or not actor can control target. Note we currently don't need
- * locking for this. TODO: think about that, esp wrt proc's dying. */
+/* Whether or not actor can control target. TODO: do something reasonable here.
+ * Just checking for the parent is a bit limiting. Could walk the parent-child
+ * tree, check user ids, or some combination. Make sure actors can always
+ * control themselves. */
bool proc_controls(struct proc *actor, struct proc *target)
{
+ return TRUE;
+ #if 0 /* Example: */
return ((actor == target) || (target->ppid == actor->pid));
+ #endif
}
/* Helper to incref by val. Using the helper to help debug/interpose on proc
*
* 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
- * cur_tf).
+ * cur_ctx).
*
* Since it always returns, it will never "eat" your reference (old
* documentation talks about this a bit). */
void proc_run_s(struct proc *p)
{
- int8_t state = 0;
uint32_t coreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0];
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
- * env_tf. */
+ /* 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);
- /* disable interrupts to protect cur_tf, owning_proc, and current */
- disable_irqsave(&state);
- /* wait til ints are disabled before unlocking, in case someone else
- * grabs the lock and IPIs us before we get set up in cur_tf */
+ /* lock was protecting the state and VC mapping, not pcpui stuff */
spin_unlock(&p->proc_lock);
/* redundant with proc_startcore, might be able to remove that one*/
__set_proc_current(p);
* for now. can simply clear_owning if we want to. */
assert(!pcpui->owning_proc);
pcpui->owning_proc = p;
- pcpui->owning_vcoreid = 0; /* TODO (VC#) */
- /* TODO: (HSS) set silly state here (__startcore does it instantly) */
+ 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
* them wherever they were before (could be either vc ctx or not) */
if (!vcpd->notif_disabled && vcpd->notif_pending
&& scp_is_vcctx_ready(vcpd)) {
vcpd->notif_disabled = TRUE;
- /* save the _S's tf in the notify slot, build and pop a new one
- * in actual/cur_tf. */
- vcpd->notif_tf = p->env_tf;
- pcpui->cur_tf = &pcpui->actual_tf;
- memset(pcpui->cur_tf, 0, sizeof(struct trapframe));
- proc_init_trapframe(pcpui->cur_tf, 0, p->env_entry,
- vcpd->transition_stack);
+ /* save the _S's ctx in the uthread slot, build and pop a new
+ * one in actual/cur_ctx. */
+ 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, 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
* that for them. */
if (!scp_is_vcctx_ready(vcpd))
vcpd->notif_pending = FALSE;
- /* this is one of the few times cur_tf != &actual_tf */
- pcpui->cur_tf = &p->env_tf;
+ /* this is one of the few times cur_ctx != &actual_ctx */
+ pcpui->cur_ctx = &p->scp_ctx;
}
- enable_irqsave(&state);
/* When the calling core idles, it'll call restartcore and run the
* _S process's context. */
return;
return;
case (PROC_RUNNABLE_M):
/* vcoremap[i] holds the coreid of the physical core allocated to
- * this process. It is set outside proc_run. For the kernel
- * message, a0 = struct proc*, a1 = struct trapframe*. */
+ * this process. It is set outside proc_run. */
if (p->procinfo->num_vcores) {
__send_bulkp_events(p);
__proc_set_state(p, PROC_RUNNING_M);
* turn online */
TAILQ_FOREACH(vc_i, &p->online_vcs, list) {
send_kernel_message(vc_i->pcoreid, __startcore, (long)p,
- (long)vcore2vcoreid(p, vc_i), 0,
- KMSG_IMMEDIATE);
+ (long)vcore2vcoreid(p, vc_i),
+ (long)vc_i->nr_preempts_sent,
+ KMSG_ROUTINE);
}
} else {
warn("Tried to proc_run() an _M with no vcores!");
}
}
-/* Actually runs the given context (trapframe) of process p on the core this
+/* You must disable IRQs and PRKM before calling this.
+ *
+ * Actually runs the given context (trapframe) of process p on the core this
* code executes on. This is called directly by __startcore, which needs to
* bypass the routine_kmsg check. Interrupts should be off when you call this.
*
* 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. */
-static void __proc_startcore(struct proc *p, trapframe_t *tf)
+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(!is_ktask(pcpui->cur_kthread));
__set_proc_current(p);
- /* need to load our silly state, preferably somewhere other than here so we
- * can avoid the case where the context was just running here. it's not
- * sufficient to do it in the "new process" if-block above (could be things
- * like page faults that cause us to keep the same process, but want a
- * different context.
- * for now, we load this silly state here. (TODO) (HSS)
- * We also need this to be per trapframe, and not per process...
- * For now / OSDI, only load it when in _S mode. _M mode was handled in
- * __startcore. */
- if (p->state == PROC_RUNNING_S)
- env_pop_ancillary_state(p);
- /* Clear the current_tf, since it is no longer used */
- current_tf = 0; /* TODO: might not need this... */
- env_pop_tf(tf);
-}
-
-/* Restarts/runs the current_tf, which must be for the current process, on the
+ /* 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.
*
* 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.
- *
- * Refcnting: this will not return, and it assumes that you've accounted for
- * your reference as if it was the ref for "current" (which is what happens when
- * returning from local traps and such. */
+ * but that would have crappy overhead. */
void proc_restartcore(void)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- assert(!pcpui->cur_sysc);
+ 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 processing (race on missing
+ /* Need ints disabled when we return from PRKM (race on missing
* messages/IPIs) */
disable_irq();
- process_routine_kmsg(pcpui->cur_tf);
+ 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. */
abandon_core();
smp_idle();
}
- assert(pcpui->cur_tf);
- __proc_startcore(pcpui->owning_proc, pcpui->cur_tf);
+ assert(pcpui->cur_ctx);
+ __proc_startcore(pcpui->owning_proc, pcpui->cur_ctx);
}
-/* Destroys the process. This should be called by the ksched, which needs to
- * hold the lock. It will destroy the process and return any cores allocated to
- * the proc via pc_arr and nr_revoked. It's up to the caller to have enough
- * space for pc_arr. This will return TRUE if we successfully killed it, FALSE
- * otherwise. Failure isn't a big deal either - it can happen due to concurrent
- * calls to proc_destroy.
+/* Destroys the process. It will destroy the process and return any cores
+ * to the ksched via the __sched_proc_destroy() CB.
*
* Here's the way process death works:
* 0. grab the lock (protects state transition and core map)
* come in, making you abandon_core, as if you weren't running. It may be that
* the only reference to p is the one you passed in, and when you decref, it'll
* get __proc_free()d. */
-bool __proc_destroy(struct proc *p, uint32_t *pc_arr, uint32_t *nr_revoked)
+void proc_destroy(struct proc *p)
{
+ 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.
- return FALSE;
- case PROC_RUNNABLE_M:
- /* Need to reclaim any cores this proc might have, even though it's
- * not running yet. */
- *nr_revoked = __proc_take_allcores(p, pc_arr, FALSE);
- // fallthrough
+ case PROC_DYING: /* someone else killed this already. */
+ spin_unlock(&p->proc_lock);
+ return;
+ case PROC_CREATED:
case PROC_RUNNABLE_S:
- /* might need to pull from lists, though i'm currently a fan of the
- * model where external refs notice DYING (if it matters to them)
- * and decref when they are done. the ksched will notice the proc
- * is dying and handle it accordingly (which delay the reaping til
- * the next call to schedule()) */
+ case PROC_WAITING:
+ break;
+ case PROC_RUNNABLE_M:
+ case PROC_RUNNING_M:
+ /* Need to reclaim any cores this proc might have, even if it's not
+ * running yet. Those running will receive a __death */
+ nr_cores_revoked = __proc_take_allcores(p, pc_arr, FALSE);
break;
case PROC_RUNNING_S:
#if 0
}
#endif
send_kernel_message(get_pcoreid(p, 0), __death, 0, 0, 0,
- KMSG_IMMEDIATE);
+ 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) */
break;
- case PROC_RUNNING_M:
- /* Send the DEATH message to every core running this process, and
- * deallocate the cores.
- * The rule is that the vcoremap is set before proc_run, and reset
- * within proc_destroy */
- *nr_revoked = __proc_take_allcores(p, pc_arr, FALSE);
- break;
- case PROC_CREATED:
- break;
default:
warn("Weird state(%s) in %s()", procstate2str(p->state),
__FUNCTION__);
- return FALSE;
+ spin_unlock(&p->proc_lock);
+ return;
}
/* At this point, a death IPI should be on its way, either from the
* RUNNING_S one, or from proc_take_cores with a __death. in general,
* 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);
- /* This prevents processes from accessing their old files while dying, and
- * will help if these files (or similar objects in the future) hold
- * references to p (preventing a __proc_free()). */
- close_all_files(&p->open_files, FALSE);
- /* Signal our state change. Assuming we only have one waiter right now. */
- sleeper = __up_sem(&p->state_change, TRUE);
- if (sleeper)
- kthread_runnable(sleeper);
- return TRUE;
+ /* 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);
+ /* 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).
+ *
+ * Also note that any mmap'd files will still be mmapped. You can close the
+ * file after mmapping, with no effect. */
+ close_fdt(&p->open_files, FALSE);
+ /* 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) */
+ proc_signal_parent(p);
+}
+
+/* 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. */
+void proc_signal_parent(struct proc *child)
+{
+ struct kthread *sleeper;
+ struct proc *parent = pid2proc(child->ppid);
+ if (!parent)
+ return;
+ /* there could be multiple kthreads sleeping for various reasons. even an
+ * SCP could have multiple async syscalls. */
+ cv_broadcast(&parent->child_wait);
+ /* if the parent was waiting, there's a __launch kthread KMSG out there */
+ proc_decref(parent);
+}
+
+/* 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. */
+int __proc_disown_child(struct proc *parent, struct proc *child)
+{
+ /* Bail out if the child has already been reaped */
+ if (!child->ppid)
+ return -1;
+ assert(child->ppid == parent->pid);
+ /* lock protects from concurrent inserts / removals from the list */
+ TAILQ_REMOVE(&parent->children, child, sibling_link);
+ /* 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 on the list */
+ return 0;
}
/* Turns *p into an MCP. Needs to be called from a local syscall of a RUNNING_S
- * process. Returns 0 if it succeeded, an error code otherwise. You should
- * hold the lock before calling. */
-int __proc_change_to_m(struct proc *p)
+ * process. Returns 0 if it succeeded, an error code otherwise. */
+int proc_change_to_m(struct proc *p)
{
- int8_t state = 0;
+ int retval = 0;
+ spin_lock(&p->proc_lock);
/* in case userspace erroneously tries to change more than once */
if (__proc_is_mcp(p))
- return -EINVAL;
+ goto error_out;
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.");
- /* save the tf so userspace can restart it. Like in __notify,
- * this assumes a user tf is the same as a kernel tf. We save
- * it in the preempt slot so that we can also save the silly
- * state. */
struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0];
- disable_irqsave(&state); /* protect cur_tf */
- /* Note this won't play well with concurrent proc kmsgs, but
- * since we're _S and locked, we shouldn't have any. */
- assert(current_tf);
- /* Copy uthread0's context to the notif slot */
- vcpd->notif_tf = *current_tf;
+ assert(current_ctx);
+ /* Copy uthread0's context to VC 0's uthread slot */
+ copy_current_ctx_to(&vcpd->uthread_ctx);
clear_owning_proc(core_id()); /* so we don't restart */
- save_fp_state(&vcpd->preempt_anc);
- enable_irqsave(&state);
+ save_vc_fp_state(vcpd);
/* Userspace needs to not fuck with notif_disabled before
* transitioning to _M. */
if (vcpd->notif_disabled) {
/* 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);
p->procinfo->is_mcp = TRUE;
- break;
+ spin_unlock(&p->proc_lock);
+ /* Tell the ksched that we're a real MCP now! */
+ __sched_proc_change_to_m(p);
+ return 0;
case (PROC_RUNNABLE_S):
/* Issues: being on the runnable_list, proc_set_state not liking
* it, and not clearly thinking through how this would happen.
* Perhaps an async call that gets serviced after you're
* descheduled? */
warn("Not supporting RUNNABLE_S -> RUNNABLE_M yet.\n");
- return -EINVAL;
+ goto error_out;
case (PROC_DYING):
warn("Dying, core request coming from %d\n", core_id());
- return -EINVAL;
+ goto error_out;
default:
- return -EINVAL;
+ goto error_out;
}
- return 0;
+error_out:
+ spin_unlock(&p->proc_lock);
+ return -EINVAL;
}
/* Old code to turn a RUNNING_M to a RUNNING_S, with the calling context
* by the proc. */
uint32_t __proc_change_to_s(struct proc *p, uint32_t *pc_arr)
{
- int8_t state = 0;
+ 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 */
- disable_irqsave(&state); /* protect cur_tf */
- assert(current_tf);
- p->env_tf = *current_tf;
+ assert(current_ctx);
+ copy_current_ctx_to(&p->scp_ctx);
clear_owning_proc(core_id()); /* so we don't restart */
- enable_irqsave(&state);
- env_push_ancillary_state(p); // TODO: (HSS)
+ save_vc_fp_state(vcpd);
/* sending death, since it's not our job to save contexts or anything in
* this case. */
num_revoked = __proc_take_allcores(p, pc_arr, FALSE);
return try_get_pcoreid(p, vcoreid);
}
-/* Helper: saves the SCP's tf state and unmaps vcore 0. In the future, we'll
- * probably use vc0's space for env_tf and the silly state. */
-void __proc_save_context_s(struct proc *p, struct trapframe *tf)
+/* Saves the FP state of the calling core into VCPD. Pairs with
+ * restore_vc_fp_state(). On x86, the best case overhead of the flags:
+ * FNINIT: 36 ns
+ * FXSAVE: 46 ns
+ * FXRSTR: 42 ns
+ * Flagged FXSAVE: 50 ns
+ * Flagged FXRSTR: 66 ns
+ * Excess flagged FXRSTR: 42 ns
+ * If we don't do it, we'll need to initialize every VCPD at process creation
+ * time with a good FPU state (x86 control words are initialized as 0s, like the
+ * rest of VCPD). */
+static void save_vc_fp_state(struct preempt_data *vcpd)
+{
+ save_fp_state(&vcpd->preempt_anc);
+ vcpd->rflags |= VC_FPU_SAVED;
+}
+
+/* Conditionally restores the FP state from VCPD. If the state was not valid,
+ * we don't bother restoring and just initialize the FPU. */
+static void restore_vc_fp_state(struct preempt_data *vcpd)
+{
+ if (vcpd->rflags & VC_FPU_SAVED) {
+ restore_fp_state(&vcpd->preempt_anc);
+ vcpd->rflags &= ~VC_FPU_SAVED;
+ } else {
+ init_fp_state();
+ }
+}
+
+/* Helper for SCPs, saves the core's FPU state into the VCPD vc0 slot */
+void __proc_save_fpu_s(struct proc *p)
{
- p->env_tf= *tf;
- env_push_ancillary_state(p); /* TODO: (HSS) */
- __unmap_vcore(p, 0); /* VC# keep in sync with proc_run_s */
+ struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0];
+ save_vc_fp_state(vcpd);
+}
+
+/* Helper: saves the SCP's GP tf state and unmaps vcore 0. This does *not* save
+ * the FPU state.
+ *
+ * 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)
+{
+ 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.
* Also note that it needs a non-current/edible reference, since it will abandon
* and continue to use the *p (current == 0, no cr3, etc).
*
- * We disable interrupts for most of it too, since we need to protect current_tf
- * and not race with __notify (which doesn't play well with concurrent
- * yielders). */
-void proc_yield(struct proc *SAFE p, bool being_nice)
+ * 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 *p, bool being_nice)
{
uint32_t vcoreid, pcoreid = core_id();
+ struct per_cpu_info *pcpui = &per_cpu_info[pcoreid];
struct vcore *vc;
struct preempt_data *vcpd;
- int8_t state = 0;
- /* Need to disable before even reading vcoreid, since we could be unmapped
- * by a __preempt or __death. _S also needs ints disabled, so we'll just do
- * it immediately. */
- disable_irqsave(&state);
- /* Need to lock before checking the vcoremap to find out who we are, in case
- * we're getting __preempted and __startcored, from a remote core (in which
- * case we might have come in thinking we were vcore X, but had X preempted
- * and Y restarted on this pcore, and we suddenly are the wrong vcore
- * yielding). Arguably, this is incredibly rare, since you'd need to
- * preempt the core, then decide to give it back with another grant in
- * between. */
+ /* Need to lock to prevent concurrent vcore changes (online, inactive, the
+ * mapping, etc). This plus checking the nr_preempts is enough to tell if
+ * our vcoreid and cur_ctx ought to be here still or if we should abort */
spin_lock(&p->proc_lock); /* horrible scalability. =( */
switch (p->state) {
case (PROC_RUNNING_S):
if (!being_nice) {
/* waiting for an event to unblock us */
vcpd = &p->procdata->vcore_preempt_data[0];
- /* this check is an early optimization (check, signal, check
- * again pattern). We could also lock before spamming the
- * vcore in event.c */
- if (vcpd->notif_pending) {
- /* they can't handle events, just need to prevent a yield.
- * (note the notif_pendings are collapsed). */
- if (!scp_is_vcctx_ready(vcpd))
- vcpd->notif_pending = FALSE;
- goto out_failed;
- }
/* syncing with event's SCP code. we set waiting, then check
* pending. they set pending, then check waiting. it's not
* possible for us to miss the notif *and* for them to miss
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.
+ * (note the notif_pendings are collapsed). */
if (!scp_is_vcctx_ready(vcpd))
vcpd->notif_pending = FALSE;
goto out_failed;
/* 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_tf);
- spin_unlock(&p->proc_lock); /* note irqs are not enabled yet */
+ __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_tf);
- spin_unlock(&p->proc_lock); /* note irqs are not enabled yet */
+ __proc_save_context_s(p);
+ spin_unlock(&p->proc_lock);
/* immediately wake up the proc (makes it runnable) */
proc_wakeup(p);
}
panic("Weird state(%s) in %s()", procstate2str(p->state),
__FUNCTION__);
}
- /* If we're already unmapped (__preempt or a __death was sent and the caller
- * unmapped us), bail out. Note that if a __death hit us, we should have
- * bailed when we saw PROC_DYING. Also note we might not have received the
- * __preempt or __death kmsg yet. */
- if (!is_mapped_vcore(p, pcoreid))
- goto out_failed;
- vcoreid = get_vcoreid(p, pcoreid);
+ /* This is which vcore this pcore thinks it is, regardless of any unmappings
+ * that may have happened remotely (with __PRs waiting to run) */
+ vcoreid = pcpui->owning_vcoreid;
vc = vcoreid2vcore(p, vcoreid);
vcpd = &p->procdata->vcore_preempt_data[vcoreid];
+ /* This is how we detect whether or not a __PR happened. */
+ if (vc->nr_preempts_sent != vc->nr_preempts_done)
+ goto out_failed;
+ /* Sanity checks. If we were preempted or are dying, we should have noticed
+ * by now. */
+ assert(is_mapped_vcore(p, pcoreid));
+ assert(vcoreid == get_vcoreid(p, pcoreid));
/* no reason to be nice, return */
if (being_nice && !vc->preempt_pending)
goto out_failed;
- /* Sanity check, can remove after a while (we should have been unmapped) */
- assert(!vc->preempt_served);
/* At this point, AFAIK there should be no preempt/death messages on the
* way, and we're on the online list. So we'll go ahead and do the yielding
* business. */
goto out_failed;
}
/* Don't let them yield if they are missing a notification. Userspace must
- * not leave vcore context without dealing with notif_pending. pop_ros_tf()
- * handles leaving via uthread context. This handles leaving via a yield.
+ * not leave vcore context without dealing with notif_pending.
+ * pop_user_ctx() handles leaving via uthread context. This handles leaving
+ * via a yield.
*
* This early check is an optimization. The real check is below when it
* works with the online_vcs list (syncing with event.c and INDIR/IPI
goto out_failed;
/* Now we'll actually try to yield */
printd("[K] Process %d (%p) is yielding on vcore %d\n", p->pid, p,
- get_vcoreid(p, coreid));
+ get_vcoreid(p, pcoreid));
/* Remove from the online list, add to the yielded list, and unmap
* the vcore, which gives up the core. */
TAILQ_REMOVE(&p->online_vcs, vc, list);
/* Note we need interrupts disabled, since a __notify can come in
* and set pending to FALSE */
if (vcpd->notif_pending) {
- /* We lost, put it back on the list and abort the yield */
+ /* We lost, put it back on the list and abort the yield. If we ever
+ * build an myield, we'll need a way to deal with this for all vcores */
TAILQ_INSERT_TAIL(&p->online_vcs, vc, list); /* could go HEAD */
goto out_failed;
}
+ /* Not really a kmsg, but it acts like one w.r.t. proc mgmt */
+ pcpui_trace_kmsg(pcpui, (uintptr_t)proc_yield);
/* We won the race with event sending, we can safely yield */
TAILQ_INSERT_HEAD(&p->inactive_vcs, vc, list);
/* Note this protects stuff userspace should look at, which doesn't
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 */
- put_idle_core(p, pcoreid);
+ __sched_put_idle_core(p, pcoreid);
goto out_yield_core;
out_failed:
/* for some reason we just want to return, either to take a KMSG that cleans
* us up, or because we shouldn't yield (ex: notif_pending). */
spin_unlock(&p->proc_lock);
- enable_irqsave(&state);
return;
out_yield_core: /* successfully yielded the core */
proc_decref(p); /* need to eat the ref passed in */
- /* Clean up the core and idle. Need to do this before enabling interrupts,
- * since once we put_idle_core() and unlock, we could get a startcore. */
+ /* Clean up the core and idle. */
clear_owning_proc(pcoreid); /* so we don't restart */
abandon_core();
- smp_idle(); /* will reenable interrupts */
+ smp_idle();
}
/* Sends a notification (aka active notification, aka IPI) to p's vcore. We
printd("[kernel] sending notif to vcore %d\n", vcoreid);
/* This use of try_get_pcoreid is racy, might be unmapped */
send_kernel_message(try_get_pcoreid(p, vcoreid), __notify, (long)p,
- 0, 0, KMSG_IMMEDIATE);
+ 0, 0, KMSG_ROUTINE);
}
}
}
-/* Makes sure p is runnable. May be spammed, via the ksched. Called only by
- * the ksched when it holds the ksched lock (or whatever). We need to lock both
- * the ksched and the proc at some point, so we need to start this call in the
- * ksched (lock ordering).
- *
- * Will call back to the ksched via one of the __sched_.cp_wakeup() calls. */
-void __proc_wakeup(struct proc *p)
+/* Makes sure p is runnable. Callers may spam this, so it needs to handle
+ * repeated calls for the same event. Callers include event delivery, SCP
+ * yield, and new SCPs. Will trigger __sched_.cp_wakeup() CBs. Will only
+ * trigger the CB once, regardless of how many times we are called, *until* the
+ * proc becomes WAITING again, presumably because of something the ksched did.*/
+void proc_wakeup(struct proc *p)
{
spin_lock(&p->proc_lock);
if (__proc_is_mcp(p)) {
/* we only wake up WAITING mcps */
- if (p->state != PROC_WAITING)
- goto out_unlock;
- if (!p->procdata->res_req[RES_CORES].amt_wanted)
- p->procdata->res_req[RES_CORES].amt_wanted = 1;
+ if (p->state != PROC_WAITING) {
+ spin_unlock(&p->proc_lock);
+ return;
+ }
__proc_set_state(p, PROC_RUNNABLE_M);
spin_unlock(&p->proc_lock);
__sched_mcp_wakeup(p);
- goto out;
+ return;
} else {
/* SCPs can wake up for a variety of reasons. the only times we need
* to do something is if it was waiting or just created. other cases
case (PROC_RUNNABLE_S):
case (PROC_RUNNING_S):
case (PROC_DYING):
- goto out_unlock;
+ spin_unlock(&p->proc_lock);
+ return;
case (PROC_RUNNABLE_M):
case (PROC_RUNNING_M):
warn("Weird state(%s) in %s()", procstate2str(p->state),
__FUNCTION__);
- goto out_unlock;
+ spin_unlock(&p->proc_lock);
+ return;
}
printd("[kernel] FYI, waking up an _S proc\n"); /* thanks, past brho! */
spin_unlock(&p->proc_lock);
__sched_scp_wakeup(p);
- goto out;
}
-out_unlock:
- spin_unlock(&p->proc_lock);
-out:
- return;
}
/* Is the process in multi_mode / is an MCP or not? */
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.
*
{
uint32_t vcoreid = get_vcoreid(p, pcoreid);
struct event_msg preempt_msg = {0};
- p->procinfo->vcoremap[vcoreid].preempt_served = TRUE;
+ /* works with nr_preempts_done to signal completion of a preemption */
+ p->procinfo->vcoremap[vcoreid].nr_preempts_sent++;
// expects a pcorelist. assumes pcore is mapped and running_m
__proc_take_corelist(p, &pcoreid, 1, TRUE);
/* Only send the message if we have an online core. o/w, it would fuck
* calling. */
uint32_t __proc_preempt_all(struct proc *p, uint32_t *pc_arr)
{
- /* instead of doing this, we could just preempt_served all possible vcores,
- * and not just the active ones. We would need to sort out a way to deal
- * with stale preempt_serveds first. This might be just as fast anyways. */
struct vcore *vc_i;
/* TODO:(BULK) PREEMPT - don't bother with this, set a proc wide flag, or
* just make us RUNNABLE_M. Note this is also used by __map_vcore. */
TAILQ_FOREACH(vc_i, &p->online_vcs, list)
- vc_i->preempt_served = TRUE;
+ vc_i->nr_preempts_sent++;
return __proc_take_allcores(p, pc_arr, TRUE);
}
/* TODO: when we revise this func, look at __put_idle */
/* Return the cores to the ksched */
if (num_revoked)
- put_idle_cores(p, pc_arr, num_revoked);
+ __sched_put_idle_cores(p, pc_arr, num_revoked);
}
/* Give the specific pcore to proc p. Lots of assumptions, so don't really use
* out). */
uint32_t proc_get_vcoreid(struct proc *p)
{
- return per_cpu_info[core_id()].owning_vcoreid;
+ struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
+ if (pcpui->owning_proc == p) {
+ return pcpui->owning_vcoreid;
+ } else {
+ warn("Asked for vcoreid for %p, but %p is pwns", p, pcpui->owning_proc);
+ return (uint32_t)-1;
+ }
}
/* TODO: make all of these static inlines when we gut the env crap */
if (!new_vc)
return FALSE;
printd("setting vcore %d to pcore %d\n", vcore2vcoreid(p, new_vc),
- pcorelist[i]);
+ pcore);
TAILQ_REMOVE(vc_list, new_vc, list);
TAILQ_INSERT_TAIL(&p->online_vcs, new_vc, list);
__map_vcore(p, vcore2vcoreid(p, new_vc), pcore);
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), 0, KMSG_IMMEDIATE);
+ (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):
/* Lock the vcore's state (necessary for preemption recovery) */
vcpd = &p->procdata->vcore_preempt_data[vcoreid];
atomic_or(&vcpd->flags, VC_K_LOCK);
- send_kernel_message(pcoreid, __preempt, (long)p, 0, 0, KMSG_IMMEDIATE);
+ send_kernel_message(pcoreid, __preempt, (long)p, 0, 0, KMSG_ROUTINE);
} else {
- send_kernel_message(pcoreid, __death, 0, 0, 0, KMSG_IMMEDIATE);
+ send_kernel_message(pcoreid, __death, 0, 0, 0, KMSG_ROUTINE);
}
}
* calling. */
void __map_vcore(struct proc *p, uint32_t vcoreid, uint32_t pcoreid)
{
- /* Need to spin until __preempt is done saving state and whatnot before we
- * give the core back out. Note that __preempt doesn't need the mapping: we
- * just need to not give out the same vcore (via a __startcore) until the
- * state is saved so __startcore has something to start. (and spinning in
- * startcore won't work, since startcore has no versioning). */
- while (p->procinfo->vcoremap[vcoreid].preempt_served)
- cpu_relax();
p->procinfo->vcoremap[vcoreid].pcoreid = pcoreid;
p->procinfo->vcoremap[vcoreid].valid = TRUE;
p->procinfo->pcoremap[pcoreid].vcoreid = vcoreid;
/* Stop running whatever context is on this core and load a known-good cr3.
* Note this leaves no trace of what was running. This "leaves the process's
- * context. Also, we want interrupts disabled, to not conflict with kmsgs
- * (__launch_kthread, proc mgmt, etc).
+ * context.
*
* This does not clear the owning proc. Use the other helper for that. */
void abandon_core(void)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- assert(!irq_is_enabled());
/* 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_sysc = 0;
+ pcpui->cur_kthread->sysc = 0;
+ pcpui->cur_kthread->errbuf = 0; /* just in case */
if (pcpui->cur_proc)
__abandon_core();
}
{
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
struct proc *p = pcpui->owning_proc;
- assert(!irq_is_enabled());
pcpui->owning_proc = 0;
pcpui->owning_vcoreid = 0xdeadbeef;
- pcpui->cur_tf = 0; /* catch bugs for now (will go away soon) */
- if (p);
+ pcpui->cur_ctx = 0; /* catch bugs for now (may go away) */
+ if (p)
proc_decref(p);
}
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
struct proc *old_proc;
- int8_t irq_state = 0;
- disable_irqsave(&irq_state);
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);
}
- enable_irqsave(&irq_state);
return old_proc;
}
void switch_back(struct proc *new_p, struct proc *old_proc)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
- int8_t irq_state = 0;
if (old_proc != new_p) {
- disable_irqsave(&irq_state);
pcpui->cur_proc = old_proc;
if (old_proc)
lcr3(old_proc->env_cr3);
else
lcr3(boot_cr3);
- enable_irqsave(&irq_state);
}
}
* 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. */
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);
}
-/* Helper, used by __startcore and change_to_vcore, which sets up cur_tf to run
- * a given process's vcore. Caller needs to set up things like owning_proc and
+/* Helper, used by __startcore and __set_curctx, which sets up cur_ctx to run a
+ * given process's vcore. Caller needs to set up things like owning_proc and
* whatnot. Note that we might not have p loaded as current. */
-static void __set_curtf_to_vcoreid(struct proc *p, uint32_t vcoreid)
+static void __set_curctx_to_vcoreid(struct proc *p, uint32_t vcoreid,
+ uint32_t old_nr_preempts_sent)
{
struct per_cpu_info *pcpui = &per_cpu_info[core_id()];
struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[vcoreid];
-
- /* We could let userspace do this, though they come into vcore entry many
- * times, and we just need this to happen when the cores comes online the
- * first time. That, and they want this turned on as soon as we know a
- * vcore *WILL* be online. We could also do this earlier, when we map the
- * vcore to its pcore, though we don't always have current loaded or
- * otherwise mess with the VCPD in those code paths. */
- vcpd->can_rcv_msg = TRUE;
+ struct vcore *vc = vcoreid2vcore(p, vcoreid);
+ /* Spin until our vcore's old preemption is done. When __SC was sent, we
+ * were told what the nr_preempts_sent was at that time. Once that many are
+ * done, it is time for us to run. This forces a 'happens-before' ordering
+ * on a __PR of our VC before this __SC of the VC. Note the nr_done should
+ * not exceed old_nr_sent, since further __PR are behind this __SC in the
+ * KMSG queue. */
+ while (old_nr_preempts_sent != vc->nr_preempts_done)
+ cpu_relax();
+ cmb(); /* read nr_done before any other rd or wr. CPU mb in the atomic. */
/* Mark that this vcore as no longer preempted. No danger of clobbering
* other writes, since this would get turned on in __preempt (which can't be
* concurrent with this function on this core), and the atomic is just
* toggling the one bit (a concurrent VC_K_LOCK will work) */
atomic_and(&vcpd->flags, ~VC_PREEMPTED);
+ /* Once the VC is no longer preempted, we allow it to receive msgs. We
+ * could let userspace do it, but handling it here makes it easier for them
+ * to handle_indirs (when they turn this flag off). Note the atomics
+ * provide the needed barriers (cmb and mb on flags). */
+ atomic_or(&vcpd->flags, VC_CAN_RCV_MSG);
printd("[kernel] startcore on physical core %d for process %d's vcore %d\n",
core_id(), p->pid, vcoreid);
/* If notifs are disabled, the vcore was in vcore context and we need to
- * restart the preempt_tf. o/w, we give them a fresh vcore (which is also
+ * restart the vcore_ctx. o/w, we give them a fresh vcore (which is also
* what happens the first time a vcore comes online). No matter what,
* they'll restart in vcore context. It's just a matter of whether or not
* it is the old, interrupted vcore context. */
if (vcpd->notif_disabled) {
- restore_fp_state(&vcpd->preempt_anc);
/* copy-in the tf we'll pop, then set all security-related fields */
- pcpui->actual_tf = vcpd->preempt_tf;
- proc_secure_trapframe(&pcpui->actual_tf);
+ 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);
- /* TODO: consider 0'ing the FP state. We're probably leaking. */
- proc_init_trapframe(&pcpui->actual_tf, vcoreid, p->env_entry,
- vcpd->transition_stack);
+ 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_tf was built above (in actual_tf), now use it */
- pcpui->cur_tf = &pcpui->actual_tf;
- /* this cur_tf will get run when the kernel returns / idles */
+ /* Regardless of whether or not we have a 'fresh' VC, we need to restore the
+ * FPU state for the VC according to VCPD (which means either a saved FPU
+ * state or a brand new init). Starting a fresh VC is just referring to the
+ * GP context we run. The vcore itself needs to have the FPU state loaded
+ * from when it previously ran and was saved (or a fresh FPU if it wasn't
+ * saved). For fresh FPUs, the main purpose is for limiting info leakage.
+ * I think VCs that don't need FPU state for some reason (like having a
+ * current_uthread) can handle any sort of FPU state, since it gets sorted
+ * when they pop their next uthread.
+ *
+ * Note this can cause a GP fault on x86 if the state is corrupt. In lieu
+ * of reading in the huge FP state and mucking with mxcsr_mask, we should
+ * handle this like a KPF on user code. */
+ restore_vc_fp_state(vcpd);
+ /* 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
* state calling vcore wants to be left in. It will look like caller_vcoreid
- * was preempted. Note we don't care about notif_pending. */
-void proc_change_to_vcore(struct proc *p, uint32_t new_vcoreid,
- bool enable_my_notif)
+ * was preempted. Note we don't care about notif_pending.
+ *
+ * Will return:
+ * 0 if we successfully changed to the target vcore.
+ * -EBUSY if the target vcore is already mapped (a good kind of failure)
+ * -EAGAIN if we failed for some other reason and need to try again. For
+ * example, the caller could be preempted, and we never even attempted to
+ * change.
+ * -EINVAL some userspace bug */
+int proc_change_to_vcore(struct proc *p, uint32_t new_vcoreid,
+ bool enable_my_notif)
{
uint32_t caller_vcoreid, pcoreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[pcoreid];
struct preempt_data *caller_vcpd;
struct vcore *caller_vc, *new_vc;
struct event_msg preempt_msg = {0};
- int8_t state = 0;
- /* Need to disable before even reading caller_vcoreid, since we could be
- * unmapped by a __preempt or __death, like in yield. */
- disable_irqsave(&state);
- /* Need to lock before reading the vcoremap, like in yield */
+ int retval = -EAGAIN; /* by default, try again */
+ /* Need to not reach outside the vcoremap, which might be smaller in the
+ * future, but should always be as big as max_vcores */
+ if (new_vcoreid >= p->procinfo->max_vcores)
+ return -EINVAL;
+ /* Need to lock to prevent concurrent vcore changes, like in yield. */
spin_lock(&p->proc_lock);
/* new_vcoreid is already runing, abort */
- if (vcore_is_mapped(p, new_vcoreid))
- goto out_failed;
+ if (vcore_is_mapped(p, new_vcoreid)) {
+ retval = -EBUSY;
+ goto out_locked;
+ }
/* Need to make sure our vcore is allowed to switch. We might have a
* __preempt, __death, etc, coming in. Similar to yield. */
switch (p->state) {
case (PROC_RUNNING_S): /* user bug, just return */
case (PROC_DYING): /* incoming __death */
case (PROC_RUNNABLE_M): /* incoming (bulk) preempt/myield TODO:(BULK) */
- goto out_failed;
+ goto out_locked;
default:
panic("Weird state(%s) in %s()", procstate2str(p->state),
__FUNCTION__);
}
- /* Make sure we're still mapped in the proc. */
- if (!is_mapped_vcore(p, pcoreid))
- goto out_failed;
- /* Get all our info */
- caller_vcoreid = get_vcoreid(p, pcoreid);
- assert(caller_vcoreid == pcpui->owning_vcoreid);
- caller_vcpd = &p->procdata->vcore_preempt_data[caller_vcoreid];
+ /* This is which vcore this pcore thinks it is, regardless of any unmappings
+ * that may have happened remotely (with __PRs waiting to run) */
+ caller_vcoreid = pcpui->owning_vcoreid;
caller_vc = vcoreid2vcore(p, caller_vcoreid);
+ caller_vcpd = &p->procdata->vcore_preempt_data[caller_vcoreid];
+ /* This is how we detect whether or not a __PR happened. If it did, just
+ * abort and handle the kmsg. No new __PRs are coming since we hold the
+ * lock. This also detects a __PR followed by a __SC for the same VC. */
+ if (caller_vc->nr_preempts_sent != caller_vc->nr_preempts_done)
+ goto out_locked;
+ /* Sanity checks. If we were preempted or are dying, we should have noticed
+ * by now. */
+ assert(is_mapped_vcore(p, pcoreid));
+ assert(caller_vcoreid == get_vcoreid(p, pcoreid));
/* Should only call from vcore context */
if (!caller_vcpd->notif_disabled) {
+ retval = -EINVAL;
printk("[kernel] You tried to change vcores from uthread ctx\n");
- goto out_failed;
+ goto out_locked;
}
- /* Sanity check, can remove after a while (we should have been unmapped) */
- assert(!caller_vc->preempt_served);
/* Ok, we're clear to do the switch. Lets figure out who the new one is */
new_vc = vcoreid2vcore(p, new_vcoreid);
printd("[kernel] changing vcore %d to vcore %d\n", caller_vcoreid,
/* enable_my_notif signals how we'll be restarted */
if (enable_my_notif) {
/* if they set this flag, then the vcore can just restart from scratch,
- * and we don't care about either the notif_tf or the preempt_tf. */
+ * 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. 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 tf so that it'll get restarted by
+ /* 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->preempt_tf = *current_tf;
- save_fp_state(&caller_vcpd->preempt_anc);
- /* Mark our core as preempted (for userspace recovery). */
- atomic_or(&caller_vcpd->flags, VC_PREEMPTED);
+ copy_current_ctx_to(&caller_vcpd->vcore_ctx);
+ save_vc_fp_state(caller_vcpd);
}
+ /* 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);
/* Move the new one from inactive to online */
TAILQ_REMOVE(&p->inactive_vcs, new_vc, list);
TAILQ_INSERT_TAIL(&p->online_vcs, new_vc, list);
- /* Change the vcore map (TODO: might get rid of this seqctr) */
+ /* Change the vcore map */
__seq_start_write(&p->procinfo->coremap_seqctr);
__unmap_vcore(p, caller_vcoreid);
__map_vcore(p, new_vcoreid, pcoreid);
__seq_end_write(&p->procinfo->coremap_seqctr);
- /* So this core knows which vcore is here: */
- pcpui->owning_vcoreid = new_vcoreid;
+ 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. */
* In this case, it's the one we just changed to. */
assert(!TAILQ_EMPTY(&p->online_vcs));
send_kernel_event(p, &preempt_msg, new_vcoreid);
- /* Change cur_tf so we'll be the new vcoreid */
- __set_curtf_to_vcoreid(p, new_vcoreid);
- /* Fall through to exit (we didn't fail) */
-out_failed:
+ /* So this core knows which vcore is here. (cur_proc and owning_proc are
+ * already correct): */
+ pcpui->owning_vcoreid = new_vcoreid;
+ /* Until we set_curctx, we don't really have a valid current tf. The stuff
+ * in that old one is from our previous vcore, not the current
+ * owning_vcoreid. This matters for other KMSGS that will run before
+ * __set_curctx (like __notify). */
+ pcpui->cur_ctx = 0;
+ /* Need to send a kmsg to finish. We can't set_curctx til the __PR is done,
+ * but we can't spin right here while holding the lock (can't spin while
+ * waiting on a message, roughly) */
+ send_kernel_message(pcoreid, __set_curctx, (long)p, (long)new_vcoreid,
+ (long)new_vc->nr_preempts_sent, KMSG_ROUTINE);
+ retval = 0;
+ /* Fall through to exit */
+out_locked:
spin_unlock(&p->proc_lock);
- enable_irqsave(&state);
+ return retval;
}
/* Kernel message handler to start a process's context on this core, when the
* core next considers running a process. Tightly coupled with __proc_run_m().
* Interrupts are disabled. */
-void __startcore(struct trapframe *tf, uint32_t srcid, long a0, long a1, long a2)
+void __startcore(uint32_t srcid, long a0, long a1, long a2)
{
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);
/* Can not be any TF from a process here already */
assert(!pcpui->owning_proc);
- /* the sender of the amsg increfed already for this saved ref to p_to_run */
+ /* the sender of the kmsg increfed already for this saved ref to p_to_run */
pcpui->owning_proc = p_to_run;
pcpui->owning_vcoreid = vcoreid;
/* sender increfed again, assuming we'd install to cur_proc. only do this
}
/* Note we are not necessarily in the cr3 of p_to_run */
/* Now that we sorted refcnts and know p / which vcore it should be, set up
- * pcpui->cur_tf so that it will run that particular vcore */
- __set_curtf_to_vcoreid(p_to_run, vcoreid);
+ * pcpui->cur_ctx so that it will run that particular vcore */
+ __set_curctx_to_vcoreid(p_to_run, vcoreid, old_nr_preempts_sent);
+}
+
+/* Kernel message handler to load a proc's vcore context on this core. Similar
+ * to __startcore, except it is used when p already controls the core (e.g.
+ * change_to). Since the core is already controlled, pcpui such as owning proc,
+ * vcoreid, and cur_proc are all already set. */
+void __set_curctx(uint32_t srcid, long a0, long a1, long a2)
+{
+ struct proc *p = (struct proc*)a0;
+ uint32_t vcoreid = (uint32_t)a1;
+ uint32_t old_nr_preempts_sent = (uint32_t)a2;
+ __set_curctx_to_vcoreid(p, vcoreid, old_nr_preempts_sent);
}
-/* Bail out if it's the wrong process, or if they no longer want a notif. Don't
- * use the TF we passed in, we care about cur_tf. Try not to grab locks or
- * write access to anything that isn't per-core in here. */
-void __notify(struct trapframe *tf, uint32_t srcid, long a0, long a1, long a2)
+/* Bail out if it's the wrong process, or if they no longer want a notif. Try
+ * not to grab locks or write access to anything that isn't per-core in here. */
+void __notify(uint32_t srcid, long a0, long a1, long a2)
{
uint32_t vcoreid, coreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
/* Not the right proc */
if (p != pcpui->owning_proc)
return;
- /* Common cur_tf sanity checks. Note cur_tf could be an _S's env_tf */
- assert(pcpui->cur_tf);
- assert(!in_kernel(pcpui->cur_tf));
+ /* the core might be owned, but not have a valid cur_ctx (if we're in the
+ * process of changing */
+ if (!pcpui->cur_ctx)
+ return;
+ /* Common cur_ctx sanity checks. Note cur_ctx could be an _S's scp_ctx */
vcoreid = pcpui->owning_vcoreid;
vcpd = &p->procdata->vcore_preempt_data[vcoreid];
/* for SCPs that haven't (and might never) call vc_event_init, like rtld.
if (vcpd->notif_disabled)
return;
vcpd->notif_disabled = TRUE;
- /* save the old tf in the notify slot, build and pop a new one. Note that
+ /* 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->notif_tf = *pcpui->cur_tf;
- memset(pcpui->cur_tf, 0, sizeof(struct trapframe));
- proc_init_trapframe(pcpui->cur_tf, vcoreid, p->env_entry,
- vcpd->transition_stack);
- /* this cur_tf will get run when the kernel returns / idles */
+ copy_current_ctx_to(&vcpd->uthread_ctx);
+ 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);
+ /* this cur_ctx will get run when the kernel returns / idles */
}
-void __preempt(struct trapframe *tf, uint32_t srcid, long a0, long a1, long a2)
+void __preempt(uint32_t srcid, long a0, long a1, long a2)
{
uint32_t vcoreid, coreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
panic("__preempt arrived for a process (%p) that was not owning (%p)!",
p, pcpui->owning_proc);
}
- /* Common cur_tf sanity checks */
- assert(pcpui->cur_tf);
- assert(pcpui->cur_tf == &pcpui->actual_tf);
- assert(!in_kernel(pcpui->cur_tf));
+ /* Common cur_ctx sanity checks */
+ assert(pcpui->cur_ctx);
+ assert(pcpui->cur_ctx == &pcpui->actual_ctx);
vcoreid = pcpui->owning_vcoreid;
vcpd = &p->procdata->vcore_preempt_data[vcoreid];
printd("[kernel] received __preempt for proc %d's vcore %d on pcore %d\n",
p->procinfo->pid, vcoreid, coreid);
/* if notifs are disabled, the vcore is in vcore context (as far as we're
- * concerned), and we save it in the preempt slot. o/w, we save the
- * process's cur_tf in the notif slot, and it'll appear to the vcore when it
- * comes back up that it just took a notification. */
+ * concerned), and we save it in the vcore slot. o/w, we save the process's
+ * 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->preempt_tf = *pcpui->cur_tf;
+ copy_current_ctx_to(&vcpd->vcore_ctx);
else
- vcpd->notif_tf = *pcpui->cur_tf;
- /* either way, we save the silly state (FP) */
- save_fp_state(&vcpd->preempt_anc);
+ 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
+ * the same as whatever is in VCPD, so this shouldn't be necessary, but the
+ * arch-specific save function might do something other than write out
+ * bit-for-bit the exact same data. Checking STEALING suffices, since we
+ * hold the K_LOCK (preventing userspace from starting a fresh STEALING
+ * phase concurrently). */
+ if (!(atomic_read(&vcpd->flags) & VC_UTHREAD_STEALING))
+ save_vc_fp_state(vcpd);
/* Mark the vcore as preempted and unlock (was locked by the sender). */
atomic_or(&vcpd->flags, VC_PREEMPTED);
atomic_and(&vcpd->flags, ~VC_K_LOCK);
- wmb(); /* make sure everything else hits before we finish the preempt */
- p->procinfo->vcoremap[vcoreid].preempt_served = FALSE;
/* 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++;
/* 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) */
* Note this leaves no trace of what was running.
* It's okay if death comes to a core that's already idling and has no current.
* It could happen if a process decref'd before __proc_startcore could incref. */
-void __death(struct trapframe *tf, uint32_t srcid, long a0, long a1, long a2)
+void __death(uint32_t srcid, long a0, long a1, long a2)
{
uint32_t vcoreid, coreid = core_id();
struct per_cpu_info *pcpui = &per_cpu_info[coreid];
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) */
+ * (smp_idle, restartcore, etc). */
+ arch_finalize_ctx(pcpui->cur_ctx);
clear_owning_proc(coreid);
}
}
/* Kernel message handler, usually sent IMMEDIATE, to shoot down virtual
* addresses from a0 to a1. */
-void __tlbshootdown(struct trapframe *tf, uint32_t srcid, long a0, long a1,
- long a2)
+void __tlbshootdown(uint32_t srcid, long a0, long a1, long a2)
{
/* TODO: (TLB) something more intelligent with the range */
tlbflush();
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 %s\n", p->pid, procstate2str(p->state));
+ /* 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 \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 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 *), KMALLOC_WAIT);
+ if (!pset->procs)
+ error(-ENOMEM, NULL);
+
+ 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 j = 0;
- struct proc *p = pid2proc(pid);
+ uint64_t total_time = 0;
+ struct proc *child, *p = pid2proc(pid);
struct vcore *vc_i;
if (!p) {
printk("Bad PID.\n");
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 ");
printk("Refcnt: %d\n", atomic_read(&p->p_kref.refcount) - 1);
printk("Flags: 0x%08x\n", p->env_flags);
- printk("CR3(phys): 0x%08x\n", p->env_cr3);
+ printk("CR3(phys): %p\n", p->env_cr3);
printk("Num Vcores: %d\n", p->procinfo->num_vcores);
printk("Vcore Lists (may be in flux w/o locking):\n----------------------\n");
printk("Online:\n");
printk("Inactive / Yielded:\n");
TAILQ_FOREACH(vc_i, &p->inactive_vcs, list)
printk("\tVcore %d\n", vcore2vcoreid(p, vc_i));
+ printk("Nsec Online, up to the last offlining:\n------------------------");
+ 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");
+ proc_decref(p);
+ return;
+ }
spin_lock(&files->lock);
- for (int i = 0; i < files->max_files; i++)
- if (files->fd_array[i].fd_file) {
- printk("\tFD: %02d, File: %08p, File name: %s\n", i,
- files->fd_array[i].fd_file,
- file_name(files->fd_array[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);
+ if (files->fd[i].fd_file) {
+ printk("File: %p, File name: %s\n", files->fd[i].fd_file,
+ file_name(files->fd[i].fd_file));
+ } else {
+ assert(files->fd[i].fd_chan);
+ print_chaninfo(files->fd[i].fd_chan);
+ }
}
+ }
spin_unlock(&files->lock);
- /* No one cares, and it clutters the terminal */
- //printk("Vcore 0's Last Trapframe:\n");
- //print_trapframe(&p->env_tf);
+ printk("Children: (PID (struct proc *))\n");
+ TAILQ_FOREACH(child, &p->children, sibling_link)
+ printk("\t%d (%p)\n", child->pid, child);
/* 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;
* interrupts, which should cause us to skip cpu_halt() */
if (!STAILQ_EMPTY(&pcpui->immed_amsgs))
continue;
- printk("Owned pcore (%d) has no owner, by %08p, vc %d!\n",
+ printk("Owned pcore (%d) has no owner, by %p, vc %d!\n",
core_id(), p, vcore2vcoreid(p, vc_i));
spin_unlock(&p->proc_lock);
spin_unlock(&pid_hash_lock);
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);
}
}
projects / akaros.git / blobdiff