| /* Copyright (c) 2009, 2010 The Regents of the University of California |
| * Barret Rhoden <brho@cs.berkeley.edu> |
| * See LICENSE for details. */ |
| |
| #include <event.h> |
| #include <arch/arch.h> |
| #include <bitmask.h> |
| #include <process.h> |
| #include <atomic.h> |
| #include <smp.h> |
| #include <pmap.h> |
| #include <trap.h> |
| #include <umem.h> |
| #include <schedule.h> |
| #include <manager.h> |
| #include <stdio.h> |
| #include <assert.h> |
| #include <time.h> |
| #include <hashtable.h> |
| #include <slab.h> |
| #include <sys/queue.h> |
| #include <monitor.h> |
| #include <elf.h> |
| #include <arsc_server.h> |
| #include <kmalloc.h> |
| #include <ros/procinfo.h> |
| #include <init.h> |
| #include <rcu.h> |
| |
| struct kmem_cache *proc_cache; |
| |
| /* Other helpers, implemented later. */ |
| 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 |
| static DECL_BITMASK(pid_bmask, PID_MAX + 1); |
| spinlock_t pid_bmask_lock = SPINLOCK_INITIALIZER; |
| struct hashtable *pid_hash; |
| spinlock_t pid_hash_lock; // initialized in proc_init |
| |
| /* Finds the next free entry (zero) entry in the pid_bitmask. Set means busy. |
| * PID 0 is reserved (in proc_init). A return value of 0 is a failure (and |
| * you'll also see a warning, for now). Consider doing this with atomics. */ |
| static pid_t get_free_pid(void) |
| { |
| static pid_t next_free_pid = 1; |
| pid_t my_pid = 0; |
| |
| spin_lock(&pid_bmask_lock); |
| // atomically (can lock for now, then change to atomic_and_return |
| FOR_CIRC_BUFFER(next_free_pid, PID_MAX + 1, i) { |
| // always points to the next to test |
| next_free_pid = (next_free_pid + 1) % (PID_MAX + 1); |
| if (!GET_BITMASK_BIT(pid_bmask, i)) { |
| SET_BITMASK_BIT(pid_bmask, i); |
| my_pid = i; |
| break; |
| } |
| } |
| spin_unlock(&pid_bmask_lock); |
| if (!my_pid) |
| warn("Unable to find a PID! You need to deal with this!\n"); |
| return my_pid; |
| } |
| |
| /* Return a pid to the pid bitmask */ |
| static void put_free_pid(pid_t pid) |
| { |
| spin_lock(&pid_bmask_lock); |
| CLR_BITMASK_BIT(pid_bmask, pid); |
| 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. */ |
| int __proc_set_state(struct proc *p, uint32_t state) |
| { |
| uint32_t curstate = p->state; |
| /* Valid transitions: |
| * C -> RBS |
| * C -> D |
| * RBS -> RGS |
| * RGS -> RBS |
| * RGS -> W |
| * RGM -> W |
| * W -> RBS |
| * W -> RGS |
| * W -> RBM |
| * W -> D |
| * RGS -> RBM |
| * RBM -> RGM |
| * RGM -> RBM |
| * RGM -> RBS |
| * RGS -> D |
| * RGM -> D |
| * D -> DA |
| * |
| * These ought to be implemented later (allowed, not thought through |
| * yet). |
| * RBS -> D |
| * RBM -> D |
| */ |
| #if 1 // some sort of correctness flag |
| switch (curstate) { |
| case PROC_CREATED: |
| if (!(state & (PROC_RUNNABLE_S | PROC_DYING))) |
| goto invalid_state_transition; |
| break; |
| case PROC_RUNNABLE_S: |
| if (!(state & (PROC_RUNNING_S | PROC_DYING))) |
| goto invalid_state_transition; |
| break; |
| case PROC_RUNNING_S: |
| if (!(state & (PROC_RUNNABLE_S | PROC_RUNNABLE_M | PROC_WAITING |
| | PROC_DYING))) |
| goto invalid_state_transition; |
| break; |
| case PROC_WAITING: |
| if (!(state & (PROC_RUNNABLE_S | PROC_RUNNING_S | |
| PROC_RUNNABLE_M | PROC_DYING))) |
| goto invalid_state_transition; |
| break; |
| case PROC_DYING: |
| if (state != PROC_DYING_ABORT) |
| goto invalid_state_transition; |
| break; |
| case PROC_DYING_ABORT: |
| goto invalid_state_transition; |
| case PROC_RUNNABLE_M: |
| if (!(state & (PROC_RUNNING_M | PROC_DYING))) |
| goto invalid_state_transition; |
| break; |
| case PROC_RUNNING_M: |
| if (!(state & (PROC_RUNNABLE_S | PROC_RUNNABLE_M | PROC_WAITING |
| | PROC_DYING))) |
| goto invalid_state_transition; |
| break; |
| invalid_state_transition: |
| panic("Invalid State Transition! %s to %02x", |
| procstate2str(state), state); |
| } |
| #endif |
| p->state = state; |
| return 0; |
| } |
| |
| /* Returns a pointer to the proc with the given pid, 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 *pid2proc(pid_t pid) |
| { |
| spin_lock(&pid_hash_lock); |
| struct proc *p = hashtable_search(pid_hash, (void*)(long)pid); |
| |
| if (p) |
| if (!kref_get_not_zero(&p->p_kref, 1)) |
| p = 0; |
| spin_unlock(&pid_hash_lock); |
| 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. */ |
| void proc_init(void) |
| { |
| /* 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, NULL, 0, |
| 0, NULL); |
| /* Init PID mask and hash. pid 0 is reserved. */ |
| SET_BITMASK_BIT(pid_bmask, 0); |
| spinlock_init(&pid_hash_lock); |
| spin_lock(&pid_hash_lock); |
| pid_hash = create_hashtable(100, __generic_hash, __generic_eq); |
| spin_unlock(&pid_hash_lock); |
| schedule_init(); |
| |
| 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. */ |
| 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 = __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->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)); |
| memset(&p->procinfo->pcoremap, 0, sizeof(p->procinfo->pcoremap)); |
| p->procinfo->num_vcores = 0; |
| p->procinfo->is_mcp = FALSE; |
| p->procinfo->coremap_seqctr = SEQCTR_INITIALIZER; |
| /* 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); |
| } |
| } |
| |
| 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 for processes that block before initing uthread code (like rtld). |
| */ |
| 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, 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(*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); |
| /* Initialize the address space */ |
| if ((r = env_setup_vm(p)) < 0) { |
| kmem_cache_free(proc_cache, p); |
| return r; |
| } |
| if (!(p->pid = get_free_pid())) { |
| 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); |
| /* so we can see processes killed by the kernel */ |
| p->exitcode = 1337; |
| 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); |
| /* shouldn't go through state machine for init */ |
| p->state = PROC_CREATED; |
| p->env_flags = 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. */ |
| TAILQ_INIT(&p->online_vcs); |
| TAILQ_INIT(&p->bulk_preempted_vcs); |
| TAILQ_INIT(&p->inactive_vcs); |
| /* Init procinfo/procdata. Procinfo's argp/argb are 0'd */ |
| proc_init_procinfo(p); |
| proc_init_procdata(p); |
| |
| /* Initialize the generic sysevent ring buffer */ |
| SHARED_RING_INIT(&p->procdata->syseventring); |
| /* Initialize the frontend of the sysevent ring buffer */ |
| FRONT_RING_INIT(&p->syseventfrontring, |
| &p->procdata->syseventring, |
| SYSEVENTRINGSIZE); |
| |
| /* Init FS structures TODO: cleanup (might pull this out) */ |
| 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); |
| 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); |
| *pp = p; |
| return 0; |
| } |
| |
| /* We have a bunch of different ways to make processes. Call this once the |
| * process is ready to be used by the rest of the system. For now, this just |
| * means when it is ready to be named via the pidhash. In the future, we might |
| * push setting the state to CREATED into here. */ |
| void __proc_ready(struct 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); |
| } |
| |
| /* 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 /* 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) |
| { |
| 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)); |
| |
| 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 */ |
| /* now we'll finally decref files for the file-backed vmrs */ |
| unmap_and_destroy_vmrs(p); |
| /* Remove us from the pid_hash and give our PID back (in that order). */ |
| spin_lock(&pid_hash_lock); |
| hash_ret = hashtable_remove(pid_hash, (void*)(long)p->pid); |
| spin_unlock(&pid_hash_lock); |
| /* 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); |
| kpages_free(p->procinfo, PROCINFO_NUM_PAGES * PGSIZE); |
| kpages_free(p->procdata, PROCDATA_NUM_PAGES * PGSIZE); |
| |
| env_pagetable_free(p); |
| arch_pgdir_clear(&p->env_pgdir); |
| p->env_cr3 = 0; |
| |
| atomic_dec(&num_envs); |
| |
| /* Dealloc the struct proc */ |
| kmem_cache_free(proc_cache, p); |
| } |
| |
| /* 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 |
| * ref counting. Note that pid2proc doesn't use this interface. */ |
| void proc_incref(struct proc *p, unsigned int val) |
| { |
| kref_get(&p->p_kref, val); |
| } |
| |
| /* Helper to decref for debugging. Don't directly kref_put() for now. */ |
| void proc_decref(struct proc *p) |
| { |
| kref_put(&p->p_kref); |
| } |
| |
| /* Helper, makes p the 'current' process, dropping the old current/cr3. This no |
| * longer assumes the passed in reference already counted 'current'. It will |
| * incref internally when needed. */ |
| static void __set_proc_current(struct proc *p) |
| { |
| /* 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); |
| lcr3(p->env_cr3); |
| /* This is "leaving the process context" of the previous proc. |
| * The previous lcr3 unloaded the previous proc's context. This |
| * should rarely happen, since we usually proactively leave |
| * process context, but this is the fallback. */ |
| old_proc = pcpui->cur_proc; |
| pcpui->cur_proc = p; |
| if (old_proc) |
| proc_decref(old_proc); |
| } |
| } |
| |
| /* Flag says if vcore context is not ready, which is set in init_procdata. The |
| * process must turn off this flag on vcore0 at some point. It's off by default |
| * on all other vcores. */ |
| static bool scp_is_vcctx_ready(struct preempt_data *vcpd) |
| { |
| return !(atomic_read(&vcpd->flags) & VC_SCP_NOVCCTX); |
| } |
| |
| /* Dispatches a _S process to run on the current core. This should never be |
| * 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 |
| * 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) |
| { |
| uint32_t coreid = core_id(); |
| struct per_cpu_info *pcpui = &per_cpu_info[coreid]; |
| struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0]; |
| |
| 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: async death\n", |
| p->pid); |
| return; |
| case (PROC_RUNNABLE_S): |
| __proc_set_state(p, PROC_RUNNING_S); |
| /* 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; |
| __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); |
| /* 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); |
| /* set us up as owning_proc. ksched bug if there is already |
| * one, for now. can simply clear_owning if we want to. */ |
| assert(!pcpui->owning_proc); |
| pcpui->owning_proc = p; |
| 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 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 kernel. They won't even turn off |
| * notif_pending, so we'll do that for them. */ |
| if (!scp_is_vcctx_ready(vcpd)) |
| vcpd->notif_pending = FALSE; |
| /* this is one of the few times cur_ctx != &actual_ctx*/ |
| pcpui->cur_ctx = &p->scp_ctx; |
| } |
| /* When the calling core idles, it'll call restartcore and run |
| * the _S process's context. */ |
| return; |
| default: |
| spin_unlock(&p->proc_lock); |
| panic("Invalid process state %p in %s()!!", p->state, |
| __FUNCTION__); |
| } |
| } |
| |
| /* Helper: sends preempt messages to all vcores on the bulk preempt list, and |
| * moves them to the inactive list. */ |
| static void __send_bulkp_events(struct proc *p) |
| { |
| struct vcore *vc_i, *vc_temp; |
| struct event_msg preempt_msg = {0}; |
| |
| /* Whenever we send msgs with the proc locked, we need at least 1 online |
| */ |
| assert(!TAILQ_EMPTY(&p->online_vcs)); |
| /* Send preempt messages for any left on the BP list. No need to set |
| * any flags, it all was done on the real preempt. Now we're just |
| * telling the process about any that didn't get restarted and are still |
| * preempted. */ |
| TAILQ_FOREACH_SAFE(vc_i, &p->bulk_preempted_vcs, list, vc_temp) { |
| /* Note that if there are no active vcores, send_k_e will post |
| * to our own vcore, the last of which will be put on the |
| * inactive list and be the first to be started. We could have |
| * issues with deadlocking, since send_k_e() could grab the |
| * proclock (if there are no active vcores) */ |
| preempt_msg.ev_type = EV_VCORE_PREEMPT; |
| preempt_msg.ev_arg2 = vcore2vcoreid(p, vc_i); /* arg2 32 bits */ |
| send_kernel_event(p, &preempt_msg, 0); |
| /* TODO: we may want a TAILQ_CONCAT_HEAD, or something that does |
| * that. We need a loop for the messages, but not necessarily |
| * for the list changes. */ |
| TAILQ_REMOVE(&p->bulk_preempted_vcs, vc_i, list); |
| TAILQ_INSERT_HEAD(&p->inactive_vcs, vc_i, list); |
| } |
| } |
| |
| /* Run an _M. Can be called safely on one that is already running. Hold the |
| * lock before calling. Other than state checks, this just starts up the _M's |
| * vcores, much like the second part of give_cores_running. More specifically, |
| * give_cores_runnable puts cores on the online list, which this then sends |
| * messages to. give_cores_running immediately puts them on the list and sends |
| * the message. the two-step style may go out of fashion soon. |
| * |
| * This expects that the "instructions" for which core(s) to run this on will be |
| * in the vcoremap, which needs to be set externally (give_cores()). */ |
| void __proc_run_m(struct proc *p) |
| { |
| struct vcore *vc_i; |
| 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; |
| case (PROC_RUNNABLE_M): |
| /* vcoremap[i] holds the coreid of the physical core allocated |
| * to this process. It is set outside proc_run. */ |
| if (p->procinfo->num_vcores) { |
| __send_bulkp_events(p); |
| __proc_set_state(p, PROC_RUNNING_M); |
| /* Up the refcnt, to avoid the n refcnt upping on the |
| * destination cores. Keep in sync with __startcore */ |
| proc_incref(p, p->procinfo->num_vcores * 2); |
| /* Send kernel messages to all online vcores (which were |
| * added to the list and mapped in __proc_give_cores()), |
| * making them turn online */ |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) { |
| send_kernel_message(vc_i->pcoreid, __startcore, |
| (long)p, |
| (long)vcore2vcoreid(p, vc_i), |
| (long)vc_i->nr_preempts_sent, |
| KMSG_ROUTINE); |
| } |
| } else { |
| warn("Tried to proc_run() an _M with no vcores!"); |
| } |
| /* There a subtle race avoidance here (when we unlock after |
| * sending the message). __proc_startcore can handle a death |
| * message, but we can't have the startcore come after the death |
| * message. Otherwise, it would look like a new process. So we |
| * hold the lock til after we send our message, which prevents a |
| * possible death message. |
| * - Note there is no guarantee this core's interrupts were on, |
| * so it may not get the message for a while... */ |
| return; |
| case (PROC_RUNNING_M): |
| return; |
| default: |
| /* unlock just so the monitor can call something that might |
| * lock*/ |
| spin_unlock(&p->proc_lock); |
| panic("Invalid process state %p in %s()!!", p->state, |
| __FUNCTION__); |
| } |
| } |
| |
| /* 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. |
| * |
| * A note on refcnting: this function will not return, and your proc reference |
| * 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(!is_ktask(pcpui->cur_kthread)); |
| __set_proc_current(p); |
| __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. |
| * |
| * 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. |
| * |
| * 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) |
| { |
| 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); |
| } |
| 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 |
| * 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) |
| * 1. set state to dying. that keeps the kernel from doing anything for the |
| * process (like proc_running it). |
| * 2. figure out where the process is running (cross-core/async or RUNNING_M) |
| * 3. IPI to clean up those cores (decref, etc). |
| * 4. Unlock |
| * 5. Clean up your core, if applicable |
| * (Last core/kernel thread to decref cleans up and deallocates resources.) |
| * |
| * Note that some cores can be processing async calls, but will eventually |
| * decref. Should think about this more, like some sort of callback/revocation. |
| * |
| * This function will now always return (it used to not return if the calling |
| * core was dying). However, when it returns, a kernel message will eventually |
| * 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. */ |
| void proc_destroy(struct proc *p) |
| { |
| uint32_t nr_cores_revoked = 0; |
| struct kthread *sleeper; |
| struct proc *child_i, *temp; |
| |
| 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: |
| case PROC_RUNNABLE_S: |
| 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 |
| // here's how to do it manually |
| if (current == p) { |
| lcr3(boot_cr3); |
| current = NULL; |
| proc_decref(p); /* this decref is for the cr3 */ |
| } |
| #endif |
| send_kernel_message(get_pcoreid(p, 0), __death, (long)p, 0, 0, |
| KMSG_ROUTINE); |
| __seq_start_write(&p->procinfo->coremap_seqctr); |
| __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; |
| default: |
| warn("Weird state(%s) in %s()", procstate2str(p->state), |
| __FUNCTION__); |
| 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); |
| 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); |
| /* 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. 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_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) */ |
| proc_signal_parent(p); |
| } |
| |
| /* Can use this to signal anything that might cause a parent to wait on the |
| * 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); |
| /* 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. |
| * |
| * 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 */ |
| 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; |
| 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. */ |
| int proc_change_to_m(struct proc *p) |
| { |
| int retval = 0; |
| spin_lock(&p->proc_lock); |
| /* in case userspace erroneously tries to change more than once */ |
| if (__proc_is_mcp(p)) |
| 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 */ |
| if ((current != p) || (get_pcoreid(p, 0) != core_id())) |
| panic("We don't handle async RUNNING_S core requests"); |
| struct preempt_data *vcpd = &p->procdata->vcore_preempt_data[0]; |
| |
| 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_vc_fp_state(vcpd); |
| /* Userspace needs to not fuck with notif_disabled before |
| * transitioning to _M. */ |
| if (vcpd->notif_disabled) { |
| printk("[kernel] user bug: notifs disabled for vcore 0\n"); |
| vcpd->notif_disabled = FALSE; |
| } |
| /* in the async case, we'll need to remotely stop and bundle |
| * vcore0's TF. this is already done for the sync case (local |
| * syscall). */ |
| /* 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: (ACR) will need to unmap remotely (receive-side) |
| __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; |
| 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"); |
| goto error_out; |
| case (PROC_DYING): |
| case (PROC_DYING_ABORT): |
| warn("Dying, core request coming from %d\n", core_id()); |
| goto error_out; |
| default: |
| goto error_out; |
| } |
| error_out: |
| spin_unlock(&p->proc_lock); |
| return -EINVAL; |
| } |
| |
| /* Old code to turn a RUNNING_M to a RUNNING_S, with the calling context |
| * becoming the new 'thread0'. Don't use this. Caller needs to send in a |
| * pc_arr big enough for all vcores. Will return the number of cores given up |
| * by the proc. */ |
| uint32_t __proc_change_to_s(struct proc *p, uint32_t *pc_arr) |
| { |
| 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); |
| 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 |
| * this case. */ |
| num_revoked = __proc_take_allcores(p, pc_arr, FALSE); |
| __proc_set_state(p, PROC_RUNNABLE_S); |
| return num_revoked; |
| } |
| |
| /* Helper function. Is the given pcore a mapped vcore? No locking involved, be |
| * careful. */ |
| static bool is_mapped_vcore(struct proc *p, uint32_t pcoreid) |
| { |
| return p->procinfo->pcoremap[pcoreid].valid; |
| } |
| |
| /* Helper function. Find the vcoreid for a given physical core id for proc p. |
| * No locking involved, be careful. Panics on failure. */ |
| static uint32_t get_vcoreid(struct proc *p, uint32_t pcoreid) |
| { |
| assert(is_mapped_vcore(p, pcoreid)); |
| return p->procinfo->pcoremap[pcoreid].vcoreid; |
| } |
| |
| /* Helper function. Try to find the pcoreid for a given virtual core id for |
| * proc p. No locking involved, be careful. Use this when you can tolerate a |
| * stale or otherwise 'wrong' answer. */ |
| static uint32_t try_get_pcoreid(struct proc *p, uint32_t vcoreid) |
| { |
| return p->procinfo->vcoremap[vcoreid].pcoreid; |
| } |
| |
| /* Helper function. Find the pcoreid for a given virtual core id for proc p. |
| * No locking involved, be careful. Panics on failure. */ |
| static uint32_t get_pcoreid(struct proc *p, uint32_t vcoreid) |
| { |
| assert(vcore_is_mapped(p, vcoreid)); |
| return try_get_pcoreid(p, vcoreid); |
| } |
| |
| /* 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) |
| { |
| 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. |
| * - If RUNNING_S, you just give up your time slice and will eventually return, |
| * possibly after WAITING on an event. |
| * - If RUNNING_M, you give up the current vcore (which never returns), and |
| * adjust the amount of cores wanted/granted. |
| * - If you have only one vcore, you switch to WAITING. There's no 'classic |
| * yield' for MCPs (at least not now). When you run again, you'll have one |
| * guaranteed core, starting from the entry point. |
| * |
| * If the call is being nice, it means different things for SCPs and MCPs. For |
| * MCPs, it means that it is in response to a preemption (which needs to be |
| * checked). If there is no preemption pending, just return. For SCPs, it |
| * means the proc wants to give up the core, but still has work to do. If not, |
| * the proc is trying to wait on an event. It's not being nice to others, it |
| * just has no work to do. |
| * |
| * This usually does not return (smp_idle()), so it will eat your reference. |
| * 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_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; |
| |
| /* 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]; |
| /* 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 WAITING. one (or both) of us will see |
| * and make sure the proc wakes up. */ |
| __proc_set_state(p, PROC_WAITING); |
| /* don't let the state write pass the notif read */ |
| wrmb(); |
| 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); |
| 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); |
| spin_unlock(&p->proc_lock); |
| /* immediately wake up the proc (makes it runnable) */ |
| proc_wakeup(p); |
| } |
| goto out_yield_core; |
| 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: |
| panic("Weird state(%s) in %s()", procstate2str(p->state), |
| __FUNCTION__); |
| } |
| /* 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; |
| /* 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. */ |
| /* If there's a preempt pending, we don't need to preempt later since we |
| * are yielding (nice or otherwise). If not, this is just a regular |
| * yield. */ |
| if (vc->preempt_pending) { |
| vc->preempt_pending = 0; |
| } else { |
| /* Optional: on a normal yield, check to see if we are putting |
| * them below amt_wanted (help with user races) and bail. */ |
| if (p->procdata->res_req[RES_CORES].amt_wanted >= |
| p->procinfo->num_vcores) |
| 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_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 |
| * posting). */ |
| if (vcpd->notif_pending) |
| 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, 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); |
| /* Now that we're off the online list, check to see if an alert made |
| * it through (event.c sets this) */ |
| wrmb(); /* prev write must hit before reading notif_pending */ |
| /* 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. 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 |
| * include the TAILQs. */ |
| __seq_start_write(&p->procinfo->coremap_seqctr); |
| /* Next time the vcore starts, it starts fresh */ |
| vcpd->notif_disabled = FALSE; |
| __unmap_vcore(p, vcoreid); |
| 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; |
| 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); |
| return; |
| out_yield_core: /* successfully yielded the core */ |
| proc_decref(p); /* need to eat the ref passed in */ |
| /* Clean up the core and idle. */ |
| clear_owning_proc(pcoreid); /* so we don't restart */ |
| abandon_core(); |
| smp_idle(); |
| } |
| |
| /* Sends a notification (aka active notification, aka IPI) to p's vcore. We |
| * only send a notification if one they are enabled. There's a bunch of weird |
| * cases with this, and how pending / enabled are signals between the user and |
| * kernel - check the documentation. Note that pending is more about messages. |
| * The process needs to be in vcore_context, and the reason is usually a |
| * message. We set pending here in case we were called to prod them into vcore |
| * context (like via a sys_self_notify). Also note that this works for _S |
| * procs, if you send to vcore 0 (and the proc is running). */ |
| 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) { |
| /* GIANT WARNING: we aren't using the proc-lock to protect the |
| * vcoremap. We want to be able to use this from interrupt |
| * context, and don't want the proc_lock to be an irqsave. |
| * Spurious __notify() kmsgs are okay (it checks to see if the |
| * right receiver is current). */ |
| if (vcore_is_mapped(p, vcoreid)) { |
| 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_ROUTINE); |
| } |
| } |
| } |
| |
| /* 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) { |
| spin_unlock(&p->proc_lock); |
| return; |
| } |
| __proc_set_state(p, PROC_RUNNABLE_M); |
| spin_unlock(&p->proc_lock); |
| __sched_mcp_wakeup(p); |
| 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 are either benign (just go out), or potential |
| * bugs (_Ms) */ |
| switch (p->state) { |
| case (PROC_CREATED): |
| case (PROC_WAITING): |
| __proc_set_state(p, PROC_RUNNABLE_S); |
| break; |
| 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): |
| case (PROC_RUNNING_M): |
| warn("Weird state(%s) in %s()", procstate2str(p->state), |
| __FUNCTION__); |
| spin_unlock(&p->proc_lock); |
| return; |
| } |
| /* thanks, past brho! */ |
| printd("[kernel] FYI, waking up an _S proc\n"); |
| spin_unlock(&p->proc_lock); |
| __sched_scp_wakeup(p); |
| } |
| } |
| |
| /* Is the process in multi_mode / is an MCP or not? */ |
| bool __proc_is_mcp(struct proc *p) |
| { |
| /* in lieu of using the amount of cores requested, or having a bunch of |
| * states (like PROC_WAITING_M and _S), I'll just track it with a bool. |
| */ |
| 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. |
| * |
| * Careful about what takes a vcoreid and what takes a pcoreid. Also, there may |
| * be weird glitches with setting the state to RUNNABLE_M. It is somewhat in |
| * flux. The num_vcores is changed after take_cores, but some of the messages |
| * (or local traps) may not yet be ready to handle seeing their future state. |
| * But they should be, so fix those when they pop up. |
| * |
| * Another thing to do would be to make the _core functions take a pcorelist, |
| * and not just one pcoreid. */ |
| |
| /* Sets a preempt_pending warning for p's vcore, to go off 'when'. If you care |
| * about locking, do it before calling. Takes a vcoreid! */ |
| void __proc_preempt_warn(struct proc *p, uint32_t vcoreid, uint64_t when) |
| { |
| struct event_msg local_msg = {0}; |
| |
| /* danger with doing this unlocked: preempt_pending is set, but never |
| * 0'd, since it is unmapped and not dealt with (TODO)*/ |
| p->procinfo->vcoremap[vcoreid].preempt_pending = when; |
| |
| /* Send the event (which internally checks to see how they want it) */ |
| local_msg.ev_type = EV_PREEMPT_PENDING; |
| local_msg.ev_arg1 = vcoreid; |
| /* Whenever we send msgs with the proc locked, we need at least 1 |
| * online. Caller needs to make sure the core was online/mapped. */ |
| assert(!TAILQ_EMPTY(&p->online_vcs)); |
| send_kernel_event(p, &local_msg, vcoreid); |
| |
| /* TODO: consider putting in some lookup place for the alarm to find it. |
| * til then, it'll have to scan the vcoremap (O(n) instead of O(m)) */ |
| } |
| |
| /* Warns all active vcores of an impending preemption. Hold the lock if you |
| * care about the mapping (and you should). */ |
| void __proc_preempt_warnall(struct proc *p, uint64_t when) |
| { |
| struct vcore *vc_i; |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) |
| __proc_preempt_warn(p, vcore2vcoreid(p, vc_i), when); |
| /* TODO: consider putting in some lookup place for the alarm to find it. |
| * til then, it'll have to scan the vcoremap (O(n) instead of O(m)) */ |
| } |
| |
| // TODO: function to set an alarm, if none is outstanding |
| |
| /* Raw function to preempt a single core. If you care about locking, do it |
| * before calling. */ |
| void __proc_preempt_core(struct proc *p, uint32_t pcoreid) |
| { |
| uint32_t vcoreid = get_vcoreid(p, pcoreid); |
| struct event_msg preempt_msg = {0}; |
| /* 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 |
| * us up (deadlock), and hey don't need a message. the core we just |
| * took will be the first one to be restarted. It will look like a |
| * notif. in the future, we could send the event if we want, but the |
| * caller needs to do that (after unlocking). */ |
| if (!TAILQ_EMPTY(&p->online_vcs)) { |
| preempt_msg.ev_type = EV_VCORE_PREEMPT; |
| preempt_msg.ev_arg2 = vcoreid; |
| send_kernel_event(p, &preempt_msg, 0); |
| } |
| } |
| |
| /* Raw function to preempt every vcore. If you care about locking, do it before |
| * calling. */ |
| uint32_t __proc_preempt_all(struct proc *p, uint32_t *pc_arr) |
| { |
| 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->nr_preempts_sent++; |
| return __proc_take_allcores(p, pc_arr, TRUE); |
| } |
| |
| /* Warns and preempts a vcore from p. No delaying / alarming, or anything. The |
| * warning will be for u usec from now. Returns TRUE if the core belonged to |
| * the proc (and thus preempted), False if the proc no longer has the core. */ |
| bool proc_preempt_core(struct proc *p, uint32_t pcoreid, uint64_t usec) |
| { |
| uint64_t warn_time = read_tsc() + usec2tsc(usec); |
| bool retval = FALSE; |
| if (p->state != PROC_RUNNING_M) { |
| /* more of an FYI for brho. should be harmless to return. */ |
| warn("Tried to preempt from a non RUNNING_M proc!"); |
| return FALSE; |
| } |
| spin_lock(&p->proc_lock); |
| if (is_mapped_vcore(p, pcoreid)) { |
| __proc_preempt_warn(p, get_vcoreid(p, pcoreid), warn_time); |
| __proc_preempt_core(p, pcoreid); |
| /* we might have taken the last core */ |
| if (!p->procinfo->num_vcores) |
| __proc_set_state(p, PROC_RUNNABLE_M); |
| retval = TRUE; |
| } |
| spin_unlock(&p->proc_lock); |
| return retval; |
| } |
| |
| /* Warns and preempts all from p. No delaying / alarming, or anything. The |
| * warning will be for u usec from now. */ |
| void proc_preempt_all(struct proc *p, uint64_t usec) |
| { |
| uint64_t warn_time = read_tsc() + usec2tsc(usec); |
| uint32_t num_revoked = 0; |
| |
| 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]; |
| |
| /* DYING could be okay */ |
| if (p->state != PROC_RUNNING_M) { |
| warn("Tried to preempt from a non RUNNING_M proc!"); |
| spin_unlock(&p->proc_lock); |
| return; |
| } |
| __proc_preempt_warnall(p, warn_time); |
| num_revoked = __proc_preempt_all(p, pc_arr); |
| assert(!p->procinfo->num_vcores); |
| __proc_set_state(p, PROC_RUNNABLE_M); |
| spin_unlock(&p->proc_lock); |
| /* TODO: when we revise this func, look at __put_idle */ |
| /* Return the cores to the ksched */ |
| if (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 |
| * this. The proc needs to be _M and prepared for it. the pcore needs to be |
| * free, etc. */ |
| void proc_give(struct proc *p, uint32_t pcoreid) |
| { |
| warn("Your idlecoremap is now screwed up"); /* TODO (IDLE) */ |
| spin_lock(&p->proc_lock); |
| // expects a pcorelist, we give it a list of one |
| __proc_give_cores(p, &pcoreid, 1); |
| spin_unlock(&p->proc_lock); |
| } |
| |
| /* Global version of the helper, for sys_get_vcoreid (might phase that syscall |
| * out). */ |
| uint32_t proc_get_vcoreid(struct proc *p) |
| { |
| 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 */ |
| bool vcore_is_mapped(struct proc *p, uint32_t vcoreid) |
| { |
| return p->procinfo->vcoremap[vcoreid].valid; |
| } |
| |
| /* Can do this, or just create a new field and save it in the vcoremap */ |
| uint32_t vcore2vcoreid(struct proc *p, struct vcore *vc) |
| { |
| return (vc - p->procinfo->vcoremap); |
| } |
| |
| struct vcore *vcoreid2vcore(struct proc *p, uint32_t vcoreid) |
| { |
| return &p->procinfo->vcoremap[vcoreid]; |
| } |
| |
| /********** Core granting (bulk and single) ***********/ |
| |
| /* Helper: gives pcore to the process, mapping it to the next available vcore |
| * from list vc_list. Returns TRUE if we succeeded (non-empty). If you pass in |
| * **vc, we'll tell you which vcore it was. */ |
| static bool __proc_give_a_pcore(struct proc *p, uint32_t pcore, |
| struct vcore_tailq *vc_list, struct vcore **vc) |
| { |
| struct vcore *new_vc; |
| |
| new_vc = TAILQ_FIRST(vc_list); |
| if (!new_vc) |
| return FALSE; |
| printd("setting vcore %d to pcore %d\n", vcore2vcoreid(p, new_vc), |
| 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); |
| if (vc) |
| *vc = new_vc; |
| return TRUE; |
| } |
| |
| static void __proc_give_cores_runnable(struct proc *p, uint32_t *pc_arr, |
| uint32_t num) |
| { |
| assert(p->state == PROC_RUNNABLE_M); |
| assert(num); /* catch bugs */ |
| /* add new items to the vcoremap */ |
| /* unncessary if offline */ |
| __seq_start_write(&p->procinfo->coremap_seqctr); |
| p->procinfo->num_vcores += num; |
| for (int i = 0; i < num; i++) { |
| /* Try from the bulk list first */ |
| if (__proc_give_a_pcore(p, pc_arr[i], &p->bulk_preempted_vcs, |
| 0)) |
| continue; |
| /* o/w, try from the inactive list. at one point, i thought |
| * there might be a legit way in which the inactive list could |
| * be empty, but that i wanted to catch it via an assert. */ |
| assert(__proc_give_a_pcore(p, pc_arr[i], &p->inactive_vcs, 0)); |
| } |
| __seq_end_write(&p->procinfo->coremap_seqctr); |
| } |
| |
| static void __proc_give_cores_running(struct proc *p, uint32_t *pc_arr, |
| uint32_t num) |
| { |
| struct vcore *vc_i; |
| /* Up the refcnt, since num cores are going to start using this |
| * process and have it loaded in their owning_proc and 'current'. */ |
| proc_incref(p, num * 2); /* keep in sync with __startcore */ |
| __seq_start_write(&p->procinfo->coremap_seqctr); |
| p->procinfo->num_vcores += num; |
| assert(TAILQ_EMPTY(&p->bulk_preempted_vcs)); |
| 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)vc_i->nr_preempts_sent, KMSG_ROUTINE); |
| } |
| __seq_end_write(&p->procinfo->coremap_seqctr); |
| } |
| |
| /* Gives process p the additional num cores listed in pcorelist. If the proc is |
| * not RUNNABLE_M or RUNNING_M, this will fail and allocate none of the core |
| * (and return -1). If you're RUNNING_M, this will startup your new cores at |
| * the entry point with their virtual IDs (or restore a preemption). If you're |
| * RUNNABLE_M, you should call __proc_run_m after this so that the process can |
| * start to use its cores. In either case, this returns 0. |
| * |
| * If you're *_S, make sure your core0's TF is set (which is done when coming in |
| * via arch/trap.c and we are RUNNING_S), change your state, then call this. |
| * Then call __proc_run_m(). |
| * |
| * The reason I didn't bring the _S cases from core_request over here is so we |
| * can keep this family of calls dealing with only *_Ms, to avoiding caring if |
| * this is called from another core, and to avoid the _S -> _M transition. |
| * |
| * WARNING: You must hold the proc_lock before calling this! */ |
| 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_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; |
| case (PROC_RUNNABLE_M): |
| __proc_give_cores_runnable(p, pc_arr, num); |
| break; |
| case (PROC_RUNNING_M): |
| __proc_give_cores_running(p, pc_arr, num); |
| break; |
| default: |
| panic("Weird state(%s) in %s()", procstate2str(p->state), |
| __FUNCTION__); |
| } |
| /* TODO: considering moving to the ksched (hard, due to yield) */ |
| p->procinfo->res_grant[RES_CORES] += num; |
| return 0; |
| } |
| |
| /********** Core revocation (bulk and single) ***********/ |
| |
| /* Revokes a single vcore from a process (unmaps or sends a KMSG to unmap). */ |
| static void __proc_revoke_core(struct proc *p, uint32_t vcoreid, bool preempt) |
| { |
| uint32_t pcoreid = get_pcoreid(p, vcoreid); |
| struct preempt_data *vcpd; |
| if (preempt) { |
| /* 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_ROUTINE); |
| } else { |
| send_kernel_message(pcoreid, __death, (long)p, 0, 0, |
| KMSG_ROUTINE); |
| } |
| } |
| |
| /* Revokes all cores from the process (unmaps or sends a KMSGS). */ |
| static void __proc_revoke_allcores(struct proc *p, bool preempt) |
| { |
| struct vcore *vc_i; |
| |
| /* TODO: if we ever get broadcast messaging, use it here (still need to |
| * lock the vcores' states for preemption) */ |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) |
| __proc_revoke_core(p, vcore2vcoreid(p, vc_i), preempt); |
| } |
| |
| /* Might be faster to scan the vcoremap than to walk the list... */ |
| static void __proc_unmap_allcores(struct proc *p) |
| { |
| struct vcore *vc_i; |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) |
| __unmap_vcore(p, vcore2vcoreid(p, vc_i)); |
| } |
| |
| /* Takes (revoke via kmsg or unmap) from process p the num cores listed in |
| * pc_arr. Will preempt if 'preempt' is set. o/w, no state will be saved, etc. |
| * Don't use this for taking all of a process's cores. |
| * |
| * Make sure you hold the lock when you call this, and make sure that the pcore |
| * actually belongs to the proc, non-trivial due to other __preempt messages. */ |
| void __proc_take_corelist(struct proc *p, uint32_t *pc_arr, uint32_t num, |
| bool preempt) |
| { |
| struct vcore *vc; |
| uint32_t vcoreid; |
| assert(p->state & (PROC_RUNNING_M | PROC_RUNNABLE_M)); |
| __seq_start_write(&p->procinfo->coremap_seqctr); |
| for (int i = 0; i < num; i++) { |
| vcoreid = get_vcoreid(p, pc_arr[i]); |
| /* Sanity check */ |
| assert(pc_arr[i] == get_pcoreid(p, vcoreid)); |
| /* Revoke / unmap core */ |
| if (p->state == PROC_RUNNING_M) |
| __proc_revoke_core(p, vcoreid, preempt); |
| __unmap_vcore(p, vcoreid); |
| /* Change lists for the vcore. Note, the vcore is already |
| * unmapped and/or the messages are already in flight. The only |
| * code that looks at the lists without holding the lock is |
| * event code. */ |
| vc = vcoreid2vcore(p, vcoreid); |
| TAILQ_REMOVE(&p->online_vcs, vc, list); |
| /* even for single preempts, we use the inactive list. bulk |
| * preempt is only used for when we take everything. */ |
| TAILQ_INSERT_HEAD(&p->inactive_vcs, vc, list); |
| } |
| p->procinfo->num_vcores -= num; |
| __seq_end_write(&p->procinfo->coremap_seqctr); |
| p->procinfo->res_grant[RES_CORES] -= num; |
| } |
| |
| /* Takes all cores from a process (revoke via kmsg or unmap), putting them on |
| * the appropriate vcore list, and fills pc_arr with the pcores revoked, and |
| * returns the number of entries in pc_arr. |
| * |
| * Make sure pc_arr is big enough to handle num_vcores(). |
| * Make sure you hold the lock when you call this. */ |
| uint32_t __proc_take_allcores(struct proc *p, uint32_t *pc_arr, bool preempt) |
| { |
| struct vcore *vc_i, *vc_temp; |
| uint32_t num = 0; |
| assert(p->state & (PROC_RUNNING_M | PROC_RUNNABLE_M)); |
| __seq_start_write(&p->procinfo->coremap_seqctr); |
| /* Write out which pcores we're going to take */ |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) |
| pc_arr[num++] = vc_i->pcoreid; |
| /* Revoke if they are running, and unmap. Both of these need the online |
| * list to not be changed yet. */ |
| if (p->state == PROC_RUNNING_M) |
| __proc_revoke_allcores(p, preempt); |
| __proc_unmap_allcores(p); |
| /* Move the vcores from online to the head of the appropriate list */ |
| TAILQ_FOREACH_SAFE(vc_i, &p->online_vcs, list, vc_temp) { |
| /* TODO: we may want a TAILQ_CONCAT_HEAD, or something that does |
| * that */ |
| TAILQ_REMOVE(&p->online_vcs, vc_i, list); |
| /* Put the cores on the appropriate list */ |
| if (preempt) |
| TAILQ_INSERT_HEAD(&p->bulk_preempted_vcs, vc_i, list); |
| else |
| TAILQ_INSERT_HEAD(&p->inactive_vcs, vc_i, list); |
| } |
| assert(TAILQ_EMPTY(&p->online_vcs)); |
| assert(num == p->procinfo->num_vcores); |
| p->procinfo->num_vcores = 0; |
| __seq_end_write(&p->procinfo->coremap_seqctr); |
| p->procinfo->res_grant[RES_CORES] = 0; |
| return num; |
| } |
| |
| /* Helper to do the vcore->pcore and inverse mapping. Hold the lock when |
| * calling. */ |
| void __map_vcore(struct proc *p, uint32_t vcoreid, uint32_t pcoreid) |
| { |
| p->procinfo->vcoremap[vcoreid].pcoreid = pcoreid; |
| p->procinfo->vcoremap[vcoreid].valid = TRUE; |
| p->procinfo->pcoremap[pcoreid].vcoreid = vcoreid; |
| p->procinfo->pcoremap[pcoreid].valid = TRUE; |
| } |
| |
| /* Helper to unmap the vcore->pcore and inverse mapping. Hold the lock when |
| * calling. */ |
| void __unmap_vcore(struct proc *p, uint32_t vcoreid) |
| { |
| p->procinfo->pcoremap[p->procinfo->vcoremap[vcoreid].pcoreid].valid = |
| FALSE; |
| p->procinfo->vcoremap[vcoreid].valid = FALSE; |
| } |
| |
| /* 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. |
| * |
| * 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) { |
| __abandon_core(); |
| return true; |
| } |
| return false; |
| } |
| |
| /* Helper to clear the core's owning processor and manage refcnting. Pass in |
| * core_id() to save a couple core_id() calls. */ |
| void clear_owning_proc(uint32_t coreid) |
| { |
| 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) */ |
| if (p) |
| proc_decref(p); |
| } |
| |
| /* 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. |
| * 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 */ |
| 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 ret; |
| } |
| |
| /* 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) |
| lcr3(old_proc->env_cr3); |
| else |
| lcr3(boot_cr3); |
| } |
| } |
| |
| /* Will send a TLB shootdown message to every vcore in the main address space |
| * (aka, all vcores for now). The message will take the start and end virtual |
| * addresses as well, in case we want to be more clever about how much we |
| * shootdown and batching our messages. Should do the sanity about rounding up |
| * and down in this function too. |
| * |
| * Would be nice to have a broadcast kmsg at this point. Note this may send a |
| * message to the calling core (interrupting it, possibly while holding the |
| * proc_lock). We don't need to process routine messages since it's an |
| * 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. */ |
| spin_lock(&p->proc_lock); |
| switch (p->state) { |
| case (PROC_RUNNING_S): |
| tlbflush(); |
| break; |
| case (PROC_RUNNING_M): |
| /* 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; |
| default: |
| /* 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 __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_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]; |
| 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(); |
| /* read nr_done before any other rd or wr. CPU mb in the atomic. */ |
| cmb(); |
| /* 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 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) { |
| /* copy-in the tf we'll pop, then set all security-related |
| * fields */ |
| pcpui->actual_ctx = vcpd->vcore_ctx; |
| proc_secure_ctx(&pcpui->actual_ctx); |
| } else { /* not restarting from a preemption, use a fresh vcore */ |
| 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; |
| } |
| /* 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. |
| * |
| * 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}; |
| 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)) { |
| 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_M): |
| 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: |
| panic("Weird state(%s) in %s()", procstate2str(p->state), |
| __FUNCTION__); |
| } |
| /* 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 uth ctx\n"); |
| goto out_locked; |
| } |
| /* 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, |
| new_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 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 ctx so that it'll get |
| * restarted by __startcore, to make the caller look like it was |
| * 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); |
| /* We don't bother with the notif_pending race. note that notif_pending |
| * could still be set. this was a preempted vcore, and userspace will |
| * need to deal with missed messages (preempt_recover() will handle |
| * that) */ |
| TAILQ_INSERT_HEAD(&p->inactive_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 */ |
| __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); |
| 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. */ |
| preempt_msg.ev_type = (enable_my_notif ? EV_CHECK_MSGS : |
| EV_VCORE_PREEMPT); |
| preempt_msg.ev_arg2 = caller_vcoreid; /* arg2 is 32 bits */ |
| /* Whenever we send msgs with the proc locked, we need at least 1 |
| * online. 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); |
| /* 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); |
| 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(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 *)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 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 if no one else is there. this is an optimization, since we |
| * expect to send these __startcores to idles cores, and this saves a |
| * scramble to incref when all of the cores restartcore/startcore later. |
| * Keep in sync with __proc_give_cores() and __proc_run_m(). */ |
| if (!pcpui->cur_proc) { |
| pcpui->cur_proc = p_to_run; /* install the ref to cur_proc */ |
| lcr3(p_to_run->env_cr3); |
| } else { |
| proc_decref(p_to_run); |
| } |
| /* 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_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. 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]; |
| struct preempt_data *vcpd; |
| struct proc *p = (struct proc*)a0; |
| |
| /* Not the right proc */ |
| if (p != pcpui->owning_proc) |
| return; |
| /* 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. this is harmless for MCPS to check this */ |
| if (!scp_is_vcctx_ready(vcpd)) |
| return; |
| printd("received active notification for proc %d's vcore %d on pcore %d\n", |
| p->procinfo->pid, vcoreid, coreid); |
| /* sort signals. notifs are now masked, like an interrupt gate */ |
| if (vcpd->notif_disabled) |
| return; |
| 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. */ |
| 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(uint32_t srcid, long a0, long a1, long a2) |
| { |
| uint32_t vcoreid, coreid = core_id(); |
| struct per_cpu_info *pcpui = &per_cpu_info[coreid]; |
| struct preempt_data *vcpd; |
| struct proc *p = (struct proc*)a0; |
| |
| assert(p); |
| if (p != pcpui->owning_proc) { |
| panic("__preempt arrived for proc (%p) that was not owning (%p)!", |
| p, pcpui->owning_proc); |
| } |
| /* 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 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) |
| copy_current_ctx_to(&vcpd->vcore_ctx); |
| else |
| 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); |
| /* either __preempt or proc_yield() ends the preempt phase. */ |
| p->procinfo->vcoremap[vcoreid].preempt_pending = 0; |
| vcore_account_offline(p, vcoreid); |
| /* make sure everything else hits before we finish the preempt */ |
| wmb(); |
| /* 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) */ |
| clear_owning_proc(coreid); |
| } |
| |
| /* Kernel message handler to clean up the core when a process is dying. |
| * 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(uint32_t srcid, long a0, long a1, long a2) |
| { |
| uint32_t vcoreid, coreid = core_id(); |
| struct per_cpu_info *pcpui = &per_cpu_info[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 proc (%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 |
| * addresses from a0 to a1. */ |
| 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 *opaque) |
| { |
| struct proc *p = (struct proc*)item; |
| assert(p); |
| /* 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); |
| } |
| 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, 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 *), 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("Num Vcores: %d\n", p->procinfo->num_vcores); |
| printk("Vcore Lists (may be in flux w/o locking):\n----------------\n"); |
| printk("Online:\n"); |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) |
| printk("\tVcore %d -> Pcore %d\n", vcore2vcoreid(p, vc_i), |
| vc_i->pcoreid); |
| printk("Bulk Preempted:\n"); |
| TAILQ_FOREACH(vc_i, &p->bulk_preempted_vcs, list) |
| printk("\tVcore %d\n", vcore2vcoreid(p, vc_i)); |
| 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 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 (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); |
| 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); |
| } |
| |
| /* Debugging function, checks what (process, vcore) is supposed to run on this |
| * pcore. Meant to be called from smp_idle() before halting. */ |
| void check_my_owner(void) |
| { |
| struct per_cpu_info *pcpui = &per_cpu_info[core_id()]; |
| void shazbot(void *item, void *opaque) |
| { |
| struct proc *p = (struct proc*)item; |
| struct vcore *vc_i; |
| assert(p); |
| spin_lock(&p->proc_lock); |
| TAILQ_FOREACH(vc_i, &p->online_vcs, list) { |
| /* this isn't true, a __startcore could be on the way |
| * and we're already "online" */ |
| if (vc_i->pcoreid == core_id()) { |
| /* Immediate message was sent, we should get it |
| * when we enable interrupts, which should cause |
| * us to skip cpu_halt() */ |
| if (!STAILQ_EMPTY(&pcpui->immed_amsgs)) |
| continue; |
| 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); |
| monitor(0); |
| } |
| } |
| spin_unlock(&p->proc_lock); |
| } |
| assert(!irq_is_enabled()); |
| if (!booting && !pcpui->owning_proc) { |
| spin_lock(&pid_hash_lock); |
| hash_for_each(pid_hash, shazbot, NULL); |
| spin_unlock(&pid_hash_lock); |
| } |
| } |