diff options
Diffstat (limited to 'sysdep/unix')
-rw-r--r-- | sysdep/unix/Makefile | 2 | ||||
-rw-r--r-- | sysdep/unix/alloc.c | 235 | ||||
-rw-r--r-- | sysdep/unix/domain.c | 116 | ||||
-rw-r--r-- | sysdep/unix/io-loop.c | 646 | ||||
-rw-r--r-- | sysdep/unix/io-loop.h | 56 | ||||
-rw-r--r-- | sysdep/unix/io.c | 200 | ||||
-rw-r--r-- | sysdep/unix/krt.Y | 5 | ||||
-rw-r--r-- | sysdep/unix/krt.c | 440 | ||||
-rw-r--r-- | sysdep/unix/krt.h | 10 | ||||
-rw-r--r-- | sysdep/unix/log.c | 45 | ||||
-rw-r--r-- | sysdep/unix/main.c | 13 |
11 files changed, 1262 insertions, 506 deletions
diff --git a/sysdep/unix/Makefile b/sysdep/unix/Makefile index d0d36b5f..6f6b0d26 100644 --- a/sysdep/unix/Makefile +++ b/sysdep/unix/Makefile @@ -1,4 +1,4 @@ -src := alloc.c io.c krt.c log.c main.c random.c +src := alloc.c io.c io-loop.c krt.c log.c main.c random.c domain.c obj := $(src-o-files) $(all-daemon) $(cf-local) diff --git a/sysdep/unix/alloc.c b/sysdep/unix/alloc.c index e7c4e6b0..3ea10c32 100644 --- a/sysdep/unix/alloc.c +++ b/sysdep/unix/alloc.c @@ -10,6 +10,7 @@ #include "lib/resource.h" #include "lib/lists.h" #include "lib/event.h" +#include "lib/rcu.h" #include <errno.h> #include <stdlib.h> @@ -22,55 +23,55 @@ long page_size = 0; #ifdef HAVE_MMAP -#define KEEP_PAGES_MAIN_MAX 256 -#define KEEP_PAGES_MAIN_MIN 8 -#define CLEANUP_PAGES_BULK 256 +#define KEEP_PAGES_MAX 512 +#define KEEP_PAGES_MIN 32 +#define KEEP_PAGES_MAX_LOCAL 16 +#define ALLOC_PAGES_AT_ONCE 8 -STATIC_ASSERT(KEEP_PAGES_MAIN_MIN * 4 < KEEP_PAGES_MAIN_MAX); +STATIC_ASSERT(KEEP_PAGES_MIN * 4 < KEEP_PAGES_MAX); +STATIC_ASSERT(ALLOC_PAGES_AT_ONCE < KEEP_PAGES_MAX_LOCAL); static _Bool use_fake = 0; +static _Bool initialized = 0; #if DEBUGGING struct free_page { node unused[42]; - node n; + struct free_page * _Atomic next; }; #else struct free_page { - node n; + struct free_page * _Atomic next; }; #endif #define EP_POS_MAX ((page_size - OFFSETOF(struct empty_pages, pages)) / sizeof (void *)) struct empty_pages { - node n; + struct empty_pages *next; uint pos; void *pages[0]; }; -struct free_pages { - list pages; /* List of (struct free_page) keeping free pages without releasing them (hot) */ - list empty; /* List of (struct empty_pages) keeping invalidated pages mapped for us (cold) */ - u16 min, max; /* Minimal and maximal number of free pages kept */ - uint cnt; /* Number of free pages in list */ - event cleanup; -}; +DEFINE_DOMAIN(resource); +static DOMAIN(resource) empty_pages_domain; +static struct empty_pages *empty_pages = NULL; -static void global_free_pages_cleanup_event(void *); +static struct free_page * _Atomic page_stack = NULL; +static _Thread_local struct free_page * local_page_stack = NULL; -static struct free_pages global_free_pages = { - .min = KEEP_PAGES_MAIN_MIN, - .max = KEEP_PAGES_MAIN_MAX, - .cleanup = { .hook = global_free_pages_cleanup_event }, -}; +static void page_cleanup(void *); +static event page_cleanup_event = { .hook = page_cleanup, }; +#define SCHEDULE_CLEANUP do if (initialized && !shutting_down) ev_send(&global_event_list, &page_cleanup_event); while (0) -uint *pages_kept = &global_free_pages.cnt; +_Atomic int pages_kept = 0; +_Atomic int pages_kept_locally = 0; +static int pages_kept_here = 0; static void * alloc_sys_page(void) { - void *ptr = mmap(NULL, page_size, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); + void *ptr = mmap(NULL, page_size * ALLOC_PAGES_AT_ONCE, PROT_WRITE | PROT_READ, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); if (ptr == MAP_FAILED) bug("mmap(%lu) failed: %m", page_size); @@ -100,37 +101,55 @@ alloc_page(void) } #ifdef HAVE_MMAP - struct free_pages *fps = &global_free_pages; - - /* If there is any free page kept hot, we use it. */ - if (fps->cnt) + /* If there is any free page kept hot in this thread, we use it. */ + struct free_page *fp = local_page_stack; + if (fp) { - struct free_page *fp = SKIP_BACK(struct free_page, n, HEAD(fps->pages)); - rem_node(&fp->n); + local_page_stack = atomic_load_explicit(&fp->next, memory_order_acquire); + atomic_fetch_sub_explicit(&pages_kept_locally, 1, memory_order_relaxed); + pages_kept_here--; + return fp; + } - /* If the hot-free-page cache is getting short, request the cleanup routine to replenish the cache */ - if ((--fps->cnt < fps->min) && !shutting_down) - ev_schedule(&fps->cleanup); + /* If there is any free page kept hot in global storage, we use it. */ + rcu_read_lock(); + fp = atomic_load_explicit(&page_stack, memory_order_acquire); + while (fp && !atomic_compare_exchange_strong_explicit( + &page_stack, &fp, atomic_load_explicit(&fp->next, memory_order_acquire), + memory_order_acq_rel, memory_order_acquire)) + ; + rcu_read_unlock(); + if (fp) + { + atomic_fetch_sub_explicit(&pages_kept, 1, memory_order_relaxed); return fp; } /* If there is any free page kept cold, we use that. */ - if (!EMPTY_LIST(fps->empty)) - { - struct empty_pages *ep = HEAD(fps->empty); + LOCK_DOMAIN(resource, empty_pages_domain); + if (empty_pages) { + if (empty_pages->pos) + /* Either the keeper page contains at least one cold page pointer, return that */ + fp = empty_pages->pages[--empty_pages->pos]; + else + { + /* Or the keeper page has no more cold page pointer, return the keeper page */ + fp = (struct free_page *) empty_pages; + empty_pages = empty_pages->next; + } + } + UNLOCK_DOMAIN(resource, empty_pages_domain); - /* Either the keeper page contains at least one cold page pointer, return that */ - if (ep->pos) - return ep->pages[--ep->pos]; + if (fp) + return fp; - /* Or the keeper page has no more cold page pointer, return the keeper page */ - rem_node(&ep->n); - return ep; - } + /* And in the worst case, allocate some new pages by mmap() */ + void *ptr = alloc_sys_page(); + for (int i=1; i<ALLOC_PAGES_AT_ONCE; i++) + free_page(ptr + page_size * i); - /* And in the worst case, allocate a new page by mmap() */ - return alloc_sys_page(); + return ptr; #endif } @@ -145,58 +164,108 @@ free_page(void *ptr) } #ifdef HAVE_MMAP - struct free_pages *fps = &global_free_pages; + /* We primarily try to keep the pages locally. */ struct free_page *fp = ptr; + if (shutting_down || (pages_kept_here < KEEP_PAGES_MAX_LOCAL)) + { + atomic_store_explicit(&fp->next, local_page_stack, memory_order_relaxed); + atomic_fetch_add_explicit(&pages_kept_locally, 1, memory_order_relaxed); + pages_kept_here++; + return; + } + + /* If there are too many local pages, we add the free page to the global hot-free-page list */ + rcu_read_lock(); + struct free_page *next = atomic_load_explicit(&page_stack, memory_order_acquire); - /* Otherwise, we add the free page to the hot-free-page list */ - fp->n = (node) {}; - add_tail(&fps->pages, &fp->n); + do atomic_store_explicit(&fp->next, next, memory_order_release); + while (!atomic_compare_exchange_strong_explicit( + &page_stack, &next, fp, + memory_order_acq_rel, memory_order_acquire)); + rcu_read_unlock(); - /* And if there are too many hot free pages, we ask for page cleanup */ - if ((++fps->cnt > fps->max) && !shutting_down) - ev_schedule(&fps->cleanup); + /* And if there are too many global hot free pages, we ask for page cleanup */ + if (atomic_fetch_add_explicit(&pages_kept, 1, memory_order_relaxed) >= KEEP_PAGES_MAX) + SCHEDULE_CLEANUP; #endif } +/* When the routine is going to sleep for a long time, we flush the local + * hot page cache to not keep dirty pages for nothing. */ +void +flush_local_pages(void) +{ + if (use_fake || !local_page_stack || shutting_down) + return; + + /* We first count the pages to enable consistency checking. + * Also, we need to know the last page. */ + struct free_page *last = local_page_stack, *next; + int check_count = 1; + while (next = atomic_load_explicit(&last->next, memory_order_acquire)) + { + check_count++; + last = next; + } + + /* The actual number of pages must be equal to the counter value. */ + ASSERT_DIE(check_count == pages_kept_here); + + /* Repeatedly trying to insert the whole page list into global page stack at once. */ + rcu_read_lock(); + next = atomic_load_explicit(&page_stack, memory_order_acquire); + + /* First we set the outwards pointer (from our last), + * then we try to set the inwards pointer to our first page. */ + do atomic_store_explicit(&last->next, next, memory_order_release); + while (!atomic_compare_exchange_strong_explicit( + &page_stack, &next, local_page_stack, + memory_order_acq_rel, memory_order_acquire)); + rcu_read_unlock(); + + /* Finished. Now the local stack is empty. */ + local_page_stack = NULL; + pages_kept_here = 0; + + /* Check the state of global page cache and maybe schedule its cleanup. */ + atomic_fetch_sub_explicit(&pages_kept_locally, check_count, memory_order_relaxed); + if (atomic_fetch_add_explicit(&pages_kept, check_count, memory_order_relaxed) >= KEEP_PAGES_MAX) + SCHEDULE_CLEANUP; +} + #ifdef HAVE_MMAP static void -global_free_pages_cleanup_event(void *data UNUSED) +page_cleanup(void *_ UNUSED) { /* Cleanup on shutdown is ignored. All pages may be kept hot, OS will take care. */ if (shutting_down) return; - struct free_pages *fps = &global_free_pages; + struct free_page *stack = atomic_exchange_explicit(&page_stack, NULL, memory_order_acq_rel); + if (!stack) + return; - /* Cleanup may get called when hot free page cache is short of pages. Replenishing. */ - while (fps->cnt / 2 < fps->min) - { - struct free_page *fp = alloc_sys_page(); - fp->n = (node) {}; - add_tail(&fps->pages, &fp->n); - fps->cnt++; - } + /* Cleanup gets called when hot free page cache is too big. + * Moving some pages to the cold free page cache. */ - /* Or the hot free page cache is too big. Moving some pages to the cold free page cache. */ - for (int limit = CLEANUP_PAGES_BULK; limit && (fps->cnt > fps->max / 2); fps->cnt--, limit--) - { - struct free_page *fp = SKIP_BACK(struct free_page, n, TAIL(fps->pages)); - rem_node(&fp->n); + do { + synchronize_rcu(); + struct free_page *fp = stack; + stack = atomic_load_explicit(&fp->next, memory_order_acquire); + LOCK_DOMAIN(resource, empty_pages_domain); /* Empty pages are stored as pointers. To store them, we need a pointer block. */ - struct empty_pages *ep; - if (EMPTY_LIST(fps->empty) || ((ep = HEAD(fps->empty))->pos == EP_POS_MAX)) + if (!empty_pages || (empty_pages->pos == EP_POS_MAX)) { /* There is either no pointer block or the last block is full. We use this block as a pointer block. */ - ep = (struct empty_pages *) fp; - *ep = (struct empty_pages) {}; - add_head(&fps->empty, &ep->n); + empty_pages = (struct empty_pages *) fp; + *empty_pages = (struct empty_pages) {}; } else { /* We store this block as a pointer into the first free place * and tell the OS that the underlying memory is trash. */ - ep->pages[ep->pos++] = fp; + empty_pages->pages[empty_pages->pos++] = fp; if (madvise(fp, page_size, #ifdef CONFIG_MADV_DONTNEED_TO_FREE MADV_DONTNEED @@ -206,12 +275,18 @@ global_free_pages_cleanup_event(void *data UNUSED) ) < 0) bug("madvise(%p) failed: %m", fp); } + UNLOCK_DOMAIN(resource, empty_pages_domain); } + while ((atomic_fetch_sub_explicit(&pages_kept, 1, memory_order_relaxed) >= KEEP_PAGES_MAX / 2) && stack); - /* If the hot free page cleanup hit the limit, re-schedule this routine - * to allow for other routines to run. */ - if (fps->cnt > fps->max) - ev_schedule(&fps->cleanup); + while (stack) + { + struct free_page *f = stack; + stack = atomic_load_explicit(&f->next, memory_order_acquire); + free_page(f); + + atomic_fetch_sub_explicit(&pages_kept, 1, memory_order_relaxed); + } } #endif @@ -219,8 +294,6 @@ void resource_sys_init(void) { #ifdef HAVE_MMAP - ASSERT_DIE(global_free_pages.cnt == 0); - /* Check what page size the system supports */ if (!(page_size = sysconf(_SC_PAGESIZE))) die("System page size must be non-zero"); @@ -230,11 +303,8 @@ resource_sys_init(void) /* We assume that page size has only one bit and is between 1K and 256K (incl.). * Otherwise, the assumptions in lib/slab.c (sl_head's num_full range) aren't met. */ - struct free_pages *fps = &global_free_pages; - - init_list(&fps->pages); - init_list(&fps->empty); - global_free_pages_cleanup_event(NULL); + empty_pages_domain = DOMAIN_NEW(resource, "Empty Pages"); + initialized = 1; return; } @@ -244,4 +314,5 @@ resource_sys_init(void) #endif page_size = 4096; + initialized = 1; } diff --git a/sysdep/unix/domain.c b/sysdep/unix/domain.c new file mode 100644 index 00000000..0a5858a6 --- /dev/null +++ b/sysdep/unix/domain.c @@ -0,0 +1,116 @@ +/* + * BIRD Locking + * + * (c) 2020 Maria Matejka <mq@jmq.cz> + * + * Can be freely distributed and used under the terms of the GNU GPL. + */ + +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#endif + +#undef LOCAL_DEBUG + +#undef DEBUG_LOCKING + +#include "lib/birdlib.h" +#include "lib/locking.h" +#include "lib/resource.h" +#include "lib/timer.h" + +#include "conf/conf.h" + +#include <errno.h> +#include <fcntl.h> +#include <poll.h> +#include <pthread.h> +#include <semaphore.h> +#include <stdatomic.h> +#include <stdlib.h> +#include <string.h> +#include <unistd.h> + +/* + * Locking subsystem + */ + +_Thread_local struct lock_order locking_stack = {}; +_Thread_local struct domain_generic **last_locked = NULL; + +#define ASSERT_NO_LOCK ASSERT_DIE(last_locked == NULL) + +struct domain_generic { + pthread_mutex_t mutex; + uint order; + struct domain_generic **prev; + struct lock_order *locked_by; + const char *name; +}; + +#define DOMAIN_INIT(_name, _order) { .mutex = PTHREAD_MUTEX_INITIALIZER, .name = _name, .order = _order } + +static struct domain_generic the_bird_domain_gen = DOMAIN_INIT("The BIRD", OFFSETOF(struct lock_order, the_bird)); + +DOMAIN(the_bird) the_bird_domain = { .the_bird = &the_bird_domain_gen }; + +struct domain_generic * +domain_new(const char *name, uint order) +{ + ASSERT_DIE(order < sizeof(struct lock_order)); + struct domain_generic *dg = xmalloc(sizeof(struct domain_generic)); + *dg = (struct domain_generic) DOMAIN_INIT(name, order); + return dg; +} + +void +domain_free(struct domain_generic *dg) +{ + pthread_mutex_destroy(&dg->mutex); + xfree(dg); +} + +uint dg_order(struct domain_generic *dg) +{ + return dg->order; +} + +void do_lock(struct domain_generic *dg, struct domain_generic **lsp) +{ + if ((char *) lsp - (char *) &locking_stack != dg->order) + bug("Trying to lock on bad position: order=%u, lsp=%p, base=%p", dg->order, lsp, &locking_stack); + + if (lsp <= last_locked) + bug("Trying to lock in a bad order"); + if (*lsp) + bug("Inconsistent locking stack state on lock"); + + btime lock_begin = current_time(); + pthread_mutex_lock(&dg->mutex); + btime duration = current_time() - lock_begin; + if (config && (duration > config->watchdog_warning)) + log(L_WARN "Locking of %s took %d ms", dg->name, (int) (duration TO_MS)); + + if (dg->prev || dg->locked_by) + bug("Previous unlock not finished correctly"); + dg->prev = last_locked; + *lsp = dg; + last_locked = lsp; + dg->locked_by = &locking_stack; +} + +void do_unlock(struct domain_generic *dg, struct domain_generic **lsp) +{ + if ((char *) lsp - (char *) &locking_stack != dg->order) + bug("Trying to unlock on bad position: order=%u, lsp=%p, base=%p", dg->order, lsp, &locking_stack); + + if (dg->locked_by != &locking_stack) + bug("Inconsistent domain state on unlock"); + if ((last_locked != lsp) || (*lsp != dg)) + bug("Inconsistent locking stack state on unlock"); + dg->locked_by = NULL; + last_locked = dg->prev; + *lsp = NULL; + dg->prev = NULL; + pthread_mutex_unlock(&dg->mutex); +} diff --git a/sysdep/unix/io-loop.c b/sysdep/unix/io-loop.c new file mode 100644 index 00000000..5ce2d350 --- /dev/null +++ b/sysdep/unix/io-loop.c @@ -0,0 +1,646 @@ +/* + * BIRD -- I/O and event loop + * + * Can be freely distributed and used under the terms of the GNU GPL. + */ + +#include <stdio.h> +#include <stdlib.h> +#include <unistd.h> +#include <errno.h> +#include <fcntl.h> +#include <poll.h> +#include <pthread.h> +#include <time.h> +#include <sys/time.h> + +#include "nest/bird.h" + +#include "lib/buffer.h" +#include "lib/lists.h" +#include "lib/resource.h" +#include "lib/event.h" +#include "lib/timer.h" +#include "lib/socket.h" + +#include "lib/io-loop.h" +#include "sysdep/unix/io-loop.h" +#include "conf/conf.h" + +#define THREAD_STACK_SIZE 65536 /* To be lowered in near future */ + +/* + * Current thread context + */ + +_Thread_local struct birdloop *birdloop_current; +static _Thread_local struct birdloop *birdloop_wakeup_masked; +static _Thread_local uint birdloop_wakeup_masked_count; + +event_list * +birdloop_event_list(struct birdloop *loop) +{ + return &loop->event_list; +} + +struct timeloop * +birdloop_time_loop(struct birdloop *loop) +{ + return &loop->time; +} + +_Bool +birdloop_inside(struct birdloop *loop) +{ + for (struct birdloop *c = birdloop_current; c; c = c->prev_loop) + if (loop == c) + return 1; + + return 0; +} + +void +birdloop_flag(struct birdloop *loop, u32 flag) +{ + atomic_fetch_or_explicit(&loop->flags, flag, memory_order_acq_rel); + birdloop_ping(loop); +} + +void +birdloop_flag_set_handler(struct birdloop *loop, struct birdloop_flag_handler *fh) +{ + ASSERT_DIE(birdloop_inside(loop)); + loop->flag_handler = fh; +} + +static int +birdloop_process_flags(struct birdloop *loop) +{ + if (!loop->flag_handler) + return 0; + + u32 flags = atomic_exchange_explicit(&loop->flags, 0, memory_order_acq_rel); + loop->flag_handler->hook(loop->flag_handler, flags); + return !!flags; +} + +static int +birdloop_run_events(struct birdloop *loop) +{ + btime begin = current_time(); + while (current_time() - begin < 5 MS) + { + if (!ev_run_list(&loop->event_list)) + return 0; + + times_update(); + } + + return 1; +} + +/* + * Wakeup code for birdloop + */ + +void +pipe_new(struct pipe *p) +{ + int rv = pipe(p->fd); + if (rv < 0) + die("pipe: %m"); + + if (fcntl(p->fd[0], F_SETFL, O_NONBLOCK) < 0) + die("fcntl(O_NONBLOCK): %m"); + + if (fcntl(p->fd[1], F_SETFL, O_NONBLOCK) < 0) + die("fcntl(O_NONBLOCK): %m"); +} + +void +pipe_drain(struct pipe *p) +{ + while (1) { + char buf[64]; + int rv = read(p->fd[0], buf, sizeof(buf)); + if ((rv < 0) && (errno == EAGAIN)) + return; + + if (rv == 0) + bug("wakeup read eof"); + if ((rv < 0) && (errno != EINTR)) + bug("wakeup read: %m"); + } +} + +int +pipe_read_one(struct pipe *p) +{ + while (1) { + char v; + int rv = read(p->fd[0], &v, sizeof(v)); + if (rv == 1) + return 1; + if ((rv < 0) && (errno == EAGAIN)) + return 0; + if (rv > 1) + bug("wakeup read more bytes than expected: %d", rv); + if (rv == 0) + bug("wakeup read eof"); + if (errno != EINTR) + bug("wakeup read: %m"); + } +} + +void +pipe_kick(struct pipe *p) +{ + char v = 1; + int rv; + + while (1) { + rv = write(p->fd[1], &v, sizeof(v)); + if ((rv >= 0) || (errno == EAGAIN)) + return; + if (errno != EINTR) + bug("wakeup write: %m"); + } +} + +void +pipe_pollin(struct pipe *p, struct pollfd *pfd) +{ + pfd->fd = p->fd[0]; + pfd->events = POLLIN; + pfd->revents = 0; +} + +static inline void +wakeup_init(struct birdloop *loop) +{ + pipe_new(&loop->wakeup); +} + +static inline void +wakeup_drain(struct birdloop *loop) +{ + pipe_drain(&loop->wakeup); +} + +static inline void +wakeup_do_kick(struct birdloop *loop) +{ + pipe_kick(&loop->wakeup); +} + +static inline void +birdloop_do_ping(struct birdloop *loop) +{ + if (atomic_fetch_add_explicit(&loop->ping_sent, 1, memory_order_acq_rel)) + return; + + if (loop == birdloop_wakeup_masked) + birdloop_wakeup_masked_count++; + else + wakeup_do_kick(loop); +} + +void +birdloop_ping(struct birdloop *loop) +{ + if (birdloop_inside(loop) && !loop->ping_pending) + loop->ping_pending++; + else + birdloop_do_ping(loop); +} + + +/* + * Sockets + */ + +static void +sockets_init(struct birdloop *loop) +{ + init_list(&loop->sock_list); + loop->sock_num = 0; + + BUFFER_INIT(loop->poll_sk, loop->pool, 4); + BUFFER_INIT(loop->poll_fd, loop->pool, 4); + loop->poll_changed = 1; /* add wakeup fd */ +} + +static void +sockets_add(struct birdloop *loop, sock *s) +{ + add_tail(&loop->sock_list, &s->n); + loop->sock_num++; + + s->index = -1; + loop->poll_changed = 1; + + birdloop_ping(loop); +} + +void +sk_start(sock *s) +{ + ASSERT_DIE(birdloop_current != &main_birdloop); + sockets_add(birdloop_current, s); +} + +static void +sockets_remove(struct birdloop *loop, sock *s) +{ + if (!enlisted(&s->n)) + return; + + rem_node(&s->n); + loop->sock_num--; + + if (s->index >= 0) + { + loop->poll_sk.data[s->index] = NULL; + s->index = -1; + loop->poll_changed = 1; + loop->close_scheduled = 1; + birdloop_ping(loop); + } + else + close(s->fd); +} + +void +sk_stop(sock *s) +{ + sockets_remove(birdloop_current, s); +} + +static inline uint sk_want_events(sock *s) +{ return (s->rx_hook ? POLLIN : 0) | ((s->ttx != s->tpos) ? POLLOUT : 0); } + +/* +FIXME: this should be called from sock code + +static void +sockets_update(struct birdloop *loop, sock *s) +{ + if (s->index >= 0) + loop->poll_fd.data[s->index].events = sk_want_events(s); +} +*/ + +static void +sockets_prepare(struct birdloop *loop) +{ + BUFFER_SET(loop->poll_sk, loop->sock_num + 1); + BUFFER_SET(loop->poll_fd, loop->sock_num + 1); + + struct pollfd *pfd = loop->poll_fd.data; + sock **psk = loop->poll_sk.data; + uint i = 0; + node *n; + + WALK_LIST(n, loop->sock_list) + { + sock *s = SKIP_BACK(sock, n, n); + + ASSERT(i < loop->sock_num); + + s->index = i; + *psk = s; + pfd->fd = s->fd; + pfd->events = sk_want_events(s); + pfd->revents = 0; + + pfd++; + psk++; + i++; + } + + ASSERT(i == loop->sock_num); + + /* Add internal wakeup fd */ + *psk = NULL; + pipe_pollin(&loop->wakeup, pfd); + + loop->poll_changed = 0; +} + +static void +sockets_close_fds(struct birdloop *loop) +{ + struct pollfd *pfd = loop->poll_fd.data; + sock **psk = loop->poll_sk.data; + int poll_num = loop->poll_fd.used - 1; + + int i; + for (i = 0; i < poll_num; i++) + if (psk[i] == NULL) + close(pfd[i].fd); + + loop->close_scheduled = 0; +} + +int sk_read(sock *s, int revents); +int sk_write(sock *s); + +static void +sockets_fire(struct birdloop *loop) +{ + struct pollfd *pfd = loop->poll_fd.data; + sock **psk = loop->poll_sk.data; + int poll_num = loop->poll_fd.used - 1; + + times_update(); + + /* Last fd is internal wakeup fd */ + if (pfd[poll_num].revents & POLLIN) + wakeup_drain(loop); + + int i; + for (i = 0; i < poll_num; pfd++, psk++, i++) + { + int e = 1; + + if (! pfd->revents) + continue; + + if (pfd->revents & POLLNVAL) + bug("poll: invalid fd %d", pfd->fd); + + if (pfd->revents & POLLIN) + while (e && *psk && (*psk)->rx_hook) + e = sk_read(*psk, pfd->revents); + + e = 1; + if (pfd->revents & POLLOUT) + { + loop->poll_changed = 1; + while (e && *psk) + e = sk_write(*psk); + } + } +} + + +/* + * Birdloop + */ + +struct birdloop main_birdloop; + +static void birdloop_enter_locked(struct birdloop *loop); + +void +birdloop_init(void) +{ + wakeup_init(&main_birdloop); + + main_birdloop.time.domain = the_bird_domain.the_bird; + main_birdloop.time.loop = &main_birdloop; + + times_update(); + timers_init(&main_birdloop.time, &root_pool); + + birdloop_enter_locked(&main_birdloop); +} + +static void *birdloop_main(void *arg); + +struct birdloop * +birdloop_new(pool *pp, uint order, const char *name) +{ + struct domain_generic *dg = domain_new(name, order); + + pool *p = rp_new(pp, name); + struct birdloop *loop = mb_allocz(p, sizeof(struct birdloop)); + loop->pool = p; + + loop->time.domain = dg; + loop->time.loop = loop; + + birdloop_enter(loop); + + wakeup_init(loop); + ev_init_list(&loop->event_list, loop, name); + timers_init(&loop->time, p); + sockets_init(loop); + + int e = 0; + + if (e = pthread_attr_init(&loop->thread_attr)) + die("pthread_attr_init() failed: %M", e); + + if (e = pthread_attr_setstacksize(&loop->thread_attr, THREAD_STACK_SIZE)) + die("pthread_attr_setstacksize(%u) failed: %M", THREAD_STACK_SIZE, e); + + if (e = pthread_attr_setdetachstate(&loop->thread_attr, PTHREAD_CREATE_DETACHED)) + die("pthread_attr_setdetachstate(PTHREAD_CREATE_DETACHED) failed: %M", e); + + if (e = pthread_create(&loop->thread_id, &loop->thread_attr, birdloop_main, loop)) + die("pthread_create() failed: %M", e); + + birdloop_leave(loop); + + return loop; +} + +static void +birdloop_do_stop(struct birdloop *loop, void (*stopped)(void *data), void *data) +{ + loop->stopped = stopped; + loop->stop_data = data; + wakeup_do_kick(loop); +} + +void +birdloop_stop(struct birdloop *loop, void (*stopped)(void *data), void *data) +{ + DG_LOCK(loop->time.domain); + birdloop_do_stop(loop, stopped, data); + DG_UNLOCK(loop->time.domain); +} + +void +birdloop_stop_self(struct birdloop *loop, void (*stopped)(void *data), void *data) +{ + ASSERT_DIE(loop == birdloop_current); + ASSERT_DIE(DG_IS_LOCKED(loop->time.domain)); + + birdloop_do_stop(loop, stopped, data); +} + +void +birdloop_free(struct birdloop *loop) +{ + ASSERT_DIE(loop->links == 0); + ASSERT_DIE(pthread_equal(pthread_self(), loop->thread_id)); + + rcu_birdloop_stop(&loop->rcu); + pthread_attr_destroy(&loop->thread_attr); + + domain_free(loop->time.domain); + rfree(loop->pool); +} + +static void +birdloop_enter_locked(struct birdloop *loop) +{ + ASSERT_DIE(DG_IS_LOCKED(loop->time.domain)); + ASSERT_DIE(!birdloop_inside(loop)); + + /* Store the old context */ + loop->prev_loop = birdloop_current; + + /* Put the new context */ + birdloop_current = loop; +} + +void +birdloop_enter(struct birdloop *loop) +{ + DG_LOCK(loop->time.domain); + return birdloop_enter_locked(loop); +} + +static void +birdloop_leave_locked(struct birdloop *loop) +{ + /* Check the current context */ + ASSERT_DIE(birdloop_current == loop); + + /* Send pending pings */ + if (loop->ping_pending) + { + loop->ping_pending = 0; + birdloop_do_ping(loop); + } + + /* Restore the old context */ + birdloop_current = loop->prev_loop; +} + +void +birdloop_leave(struct birdloop *loop) +{ + birdloop_leave_locked(loop); + DG_UNLOCK(loop->time.domain); +} + +void +birdloop_mask_wakeups(struct birdloop *loop) +{ + ASSERT_DIE(birdloop_wakeup_masked == NULL); + birdloop_wakeup_masked = loop; +} + +void +birdloop_unmask_wakeups(struct birdloop *loop) +{ + ASSERT_DIE(birdloop_wakeup_masked == loop); + birdloop_wakeup_masked = NULL; + if (birdloop_wakeup_masked_count) + wakeup_do_kick(loop); + + birdloop_wakeup_masked_count = 0; +} + +void +birdloop_link(struct birdloop *loop) +{ + ASSERT_DIE(birdloop_inside(loop)); + loop->links++; +} + +void +birdloop_unlink(struct birdloop *loop) +{ + ASSERT_DIE(birdloop_inside(loop)); + loop->links--; +} + +static void * +birdloop_main(void *arg) +{ + struct birdloop *loop = arg; + timer *t; + int rv, timeout; + + rcu_birdloop_start(&loop->rcu); + + btime loop_begin = current_time(); + + tmp_init(loop->pool); + + birdloop_enter(loop); + while (1) + { + timers_fire(&loop->time, 0); + if (birdloop_process_flags(loop) + birdloop_run_events(loop)) + timeout = 0; + else if (t = timers_first(&loop->time)) + timeout = (tm_remains(t) TO_MS) + 1; + else + timeout = -1; + + if (loop->poll_changed) + sockets_prepare(loop); + else + if ((timeout < 0) || (timeout > 5000)) + flush_local_pages(); + + btime duration = current_time() - loop_begin; + if (duration > config->watchdog_warning) + log(L_WARN "I/O loop cycle took %d ms", (int) (duration TO_MS)); + + birdloop_leave(loop); + + try: + rv = poll(loop->poll_fd.data, loop->poll_fd.used, timeout); + if (rv < 0) + { + if (errno == EINTR || errno == EAGAIN) + goto try; + bug("poll: %m"); + } + + birdloop_enter(loop); + + if (loop->close_scheduled) + sockets_close_fds(loop); + + if (loop->stopped) + break; + + loop_begin = current_time(); + + if (rv && !loop->poll_changed) + sockets_fire(loop); + + atomic_exchange_explicit(&loop->ping_sent, 0, memory_order_acq_rel); + } + + /* Flush remaining events */ + ASSERT_DIE(!ev_run_list(&loop->event_list)); + + /* Drop timers */ + while (t = timers_first(&loop->time)) + tm_stop(t); + + /* No sockets allowed */ + ASSERT_DIE(EMPTY_LIST(loop->sock_list)); + ASSERT_DIE(loop->sock_num == 0); + + birdloop_leave(loop); + loop->stopped(loop->stop_data); + + flush_local_pages(); + return NULL; +} + +void +birdloop_yield(void) +{ + usleep(100); +} diff --git a/sysdep/unix/io-loop.h b/sysdep/unix/io-loop.h new file mode 100644 index 00000000..29ca96d6 --- /dev/null +++ b/sysdep/unix/io-loop.h @@ -0,0 +1,56 @@ +/* + * BIRD -- I/O and event loop + * + * Can be freely distributed and used under the terms of the GNU GPL. + */ + +#ifndef _BIRD_SYSDEP_UNIX_IO_LOOP_H_ +#define _BIRD_SYSDEP_UNIX_IO_LOOP_H_ + +#include "lib/rcu.h" + +struct pipe +{ + int fd[2]; +}; + +void pipe_new(struct pipe *); +void pipe_pollin(struct pipe *, struct pollfd *); +void pipe_drain(struct pipe *); +void pipe_kick(struct pipe *); + +struct birdloop +{ + pool *pool; + + struct timeloop time; + event_list event_list; + list sock_list; + uint sock_num; + + BUFFER(sock *) poll_sk; + BUFFER(struct pollfd) poll_fd; + u8 poll_changed; + u8 close_scheduled; + + uint ping_pending; + _Atomic u32 ping_sent; + struct pipe wakeup; + + pthread_t thread_id; + pthread_attr_t thread_attr; + + struct rcu_birdloop rcu; + + uint links; + + _Atomic u32 flags; + struct birdloop_flag_handler *flag_handler; + + void (*stopped)(void *data); + void *stop_data; + + struct birdloop *prev_loop; +}; + +#endif diff --git a/sysdep/unix/io.c b/sysdep/unix/io.c index 810e782d..c91bd597 100644 --- a/sysdep/unix/io.c +++ b/sysdep/unix/io.c @@ -36,12 +36,14 @@ #include "lib/resource.h" #include "lib/socket.h" #include "lib/event.h" +#include "lib/locking.h" #include "lib/timer.h" #include "lib/string.h" #include "nest/iface.h" #include "conf/conf.h" #include "sysdep/unix/unix.h" +#include "sysdep/unix/io-loop.h" #include CONFIG_INCLUDE_SYSIO_H /* Maximum number of calls of tx handler for one socket in one @@ -122,55 +124,50 @@ rf_fileno(struct rfile *f) btime boot_time; + void -times_init(struct timeloop *loop) +times_update(void) { struct timespec ts; int rv; + btime old_time = current_time(); + btime old_real_time = current_real_time(); + rv = clock_gettime(CLOCK_MONOTONIC, &ts); if (rv < 0) die("Monotonic clock is missing"); if ((ts.tv_sec < 0) || (((u64) ts.tv_sec) > ((u64) 1 << 40))) log(L_WARN "Monotonic clock is crazy"); - - loop->last_time = ts.tv_sec S + ts.tv_nsec NS; - loop->real_time = 0; -} - -void -times_update(struct timeloop *loop) -{ - struct timespec ts; - int rv; - - rv = clock_gettime(CLOCK_MONOTONIC, &ts); - if (rv < 0) - die("clock_gettime: %m"); - + btime new_time = ts.tv_sec S + ts.tv_nsec NS; - if (new_time < loop->last_time) + if (new_time < old_time) log(L_ERR "Monotonic clock is broken"); - loop->last_time = new_time; - loop->real_time = 0; -} - -void -times_update_real_time(struct timeloop *loop) -{ - struct timespec ts; - int rv; - rv = clock_gettime(CLOCK_REALTIME, &ts); if (rv < 0) die("clock_gettime: %m"); - loop->real_time = ts.tv_sec S + ts.tv_nsec NS; -} + btime new_real_time = ts.tv_sec S + ts.tv_nsec NS; + if (!atomic_compare_exchange_strong_explicit( + &last_time, + &old_time, + new_time, + memory_order_acq_rel, + memory_order_relaxed)) + DBG("Time update collision: last_time"); + + if (!atomic_compare_exchange_strong_explicit( + &real_time, + &old_real_time, + new_real_time, + memory_order_acq_rel, + memory_order_relaxed)) + DBG("Time update collision: real_time"); +} /** * DOC: Sockets @@ -777,8 +774,7 @@ sk_ssh_free(sock *s) if (ssh->channel) { - if (ssh_channel_is_open(ssh->channel)) - ssh_channel_close(ssh->channel); + ssh_channel_close(ssh->channel); ssh_channel_free(ssh->channel); ssh->channel = NULL; } @@ -804,18 +800,16 @@ sk_free(resource *r) sk_ssh_free(s); #endif - if (s->fd < 0) + if ((s->fd < 0) || (s->flags & SKF_THREAD)) return; - /* FIXME: we should call sk_stop() for SKF_THREAD sockets */ - if (!(s->flags & SKF_THREAD)) - { - if (s == current_sock) - current_sock = sk_next(s); - if (s == stored_sock) - stored_sock = sk_next(s); + if (s == current_sock) + current_sock = sk_next(s); + if (s == stored_sock) + stored_sock = sk_next(s); + + if (enlisted(&s->n)) rem_node(&s->n); - } if (s->type != SK_SSH && s->type != SK_SSH_ACTIVE) close(s->fd); @@ -1108,7 +1102,11 @@ sk_passive_connected(sock *s, int type) return 1; } - sk_insert(t); + if (s->flags & SKF_PASSIVE_THREAD) + t->flags |= SKF_THREAD; + else + sk_insert(t); + sk_alloc_bufs(t); s->rx_hook(t, 0); return 1; @@ -1153,34 +1151,45 @@ sk_ssh_connect(sock *s) { int server_identity_is_ok = 1; +#ifdef HAVE_SSH_OLD_SERVER_VALIDATION_API +#define ssh_session_is_known_server ssh_is_server_known +#define SSH_KNOWN_HOSTS_OK SSH_SERVER_KNOWN_OK +#define SSH_KNOWN_HOSTS_UNKNOWN SSH_SERVER_NOT_KNOWN +#define SSH_KNOWN_HOSTS_CHANGED SSH_SERVER_KNOWN_CHANGED +#define SSH_KNOWN_HOSTS_NOT_FOUND SSH_SERVER_FILE_NOT_FOUND +#define SSH_KNOWN_HOSTS_ERROR SSH_SERVER_ERROR +#define SSH_KNOWN_HOSTS_OTHER SSH_SERVER_FOUND_OTHER +#endif + /* Check server identity */ - switch (ssh_is_server_known(s->ssh->session)) + switch (ssh_session_is_known_server(s->ssh->session)) { #define LOG_WARN_ABOUT_SSH_SERVER_VALIDATION(s,msg,args...) log(L_WARN "SSH Identity %s@%s:%u: " msg, (s)->ssh->username, (s)->host, (s)->dport, ## args); - case SSH_SERVER_KNOWN_OK: + case SSH_KNOWN_HOSTS_OK: /* The server is known and has not changed. */ break; - case SSH_SERVER_NOT_KNOWN: + case SSH_KNOWN_HOSTS_UNKNOWN: LOG_WARN_ABOUT_SSH_SERVER_VALIDATION(s, "The server is unknown, its public key was not found in the known host file %s", s->ssh->server_hostkey_path); + server_identity_is_ok = 0; break; - case SSH_SERVER_KNOWN_CHANGED: + case SSH_KNOWN_HOSTS_CHANGED: LOG_WARN_ABOUT_SSH_SERVER_VALIDATION(s, "The server key has changed. Either you are under attack or the administrator changed the key."); server_identity_is_ok = 0; break; - case SSH_SERVER_FILE_NOT_FOUND: + case SSH_KNOWN_HOSTS_NOT_FOUND: LOG_WARN_ABOUT_SSH_SERVER_VALIDATION(s, "The known host file %s does not exist", s->ssh->server_hostkey_path); server_identity_is_ok = 0; break; - case SSH_SERVER_ERROR: + case SSH_KNOWN_HOSTS_ERROR: LOG_WARN_ABOUT_SSH_SERVER_VALIDATION(s, "Some error happened"); server_identity_is_ok = 0; break; - case SSH_SERVER_FOUND_OTHER: + case SSH_KNOWN_HOSTS_OTHER: LOG_WARN_ABOUT_SSH_SERVER_VALIDATION(s, "The server gave use a key of a type while we had an other type recorded. " \ "It is a possible attack."); server_identity_is_ok = 0; @@ -1516,6 +1525,36 @@ sk_open_unix(sock *s, char *name) return 0; } +static void +sk_reloop_hook(void *_vs) +{ + sock *s = _vs; + if (birdloop_inside(&main_birdloop)) + { + s->flags &= ~SKF_THREAD; + sk_insert(s); + } + else + { + s->flags |= SKF_THREAD; + sk_start(s); + } +} + +void +sk_reloop(sock *s, struct birdloop *loop) +{ + if (enlisted(&s->n)) + rem_node(&s->n); + + s->reloop = (event) { + .hook = sk_reloop_hook, + .data = s, + }; + + ev_send_loop(loop, &s->reloop); +} + #define CMSG_RX_SPACE MAX(CMSG4_SPACE_PKTINFO+CMSG4_SPACE_TTL, \ CMSG6_SPACE_PKTINFO+CMSG6_SPACE_TTL) @@ -2037,30 +2076,17 @@ struct event_log_entry static struct event_log_entry event_log[EVENT_LOG_LENGTH]; static struct event_log_entry *event_open; static int event_log_pos, event_log_num, watchdog_active; -static btime last_time; +static btime last_io_time; static btime loop_time; static void io_update_time(void) { - struct timespec ts; - int rv; - - /* - * This is third time-tracking procedure (after update_times() above and - * times_update() in BFD), dedicated to internal event log and latency - * tracking. Hopefully, we consolidate these sometimes. - */ - - rv = clock_gettime(CLOCK_MONOTONIC, &ts); - if (rv < 0) - die("clock_gettime: %m"); - - last_time = ts.tv_sec S + ts.tv_nsec NS; + last_io_time = current_time(); if (event_open) { - event_open->duration = last_time - event_open->timestamp; + event_open->duration = last_io_time - event_open->timestamp; if (event_open->duration > config->latency_limit) log(L_WARN "Event 0x%p 0x%p took %u.%03u ms", @@ -2089,7 +2115,7 @@ io_log_event(void *hook, void *data) en->hook = hook; en->data = data; - en->timestamp = last_time; + en->timestamp = last_io_time; en->duration = 0; event_log_num++; @@ -2117,14 +2143,14 @@ io_log_dump(void) struct event_log_entry *en = event_log + (event_log_pos + i) % EVENT_LOG_LENGTH; if (en->hook) log(L_DEBUG " Event 0x%p 0x%p at %8d for %d ms", en->hook, en->data, - (int) ((last_time - en->timestamp) TO_MS), (int) (en->duration TO_MS)); + (int) ((last_io_time - en->timestamp) TO_MS), (int) (en->duration TO_MS)); } } void watchdog_sigalrm(int sig UNUSED) { - /* Update last_time and duration, but skip latency check */ + /* Update last_io_time and duration, but skip latency check */ config->latency_limit = 0xffffffff; io_update_time(); @@ -2137,7 +2163,7 @@ watchdog_start1(void) { io_update_time(); - loop_time = last_time; + loop_time = last_io_time; } static inline void @@ -2145,7 +2171,7 @@ watchdog_start(void) { io_update_time(); - loop_time = last_time; + loop_time = last_io_time; event_log_num = 0; if (config->watchdog_timeout) @@ -2166,7 +2192,7 @@ watchdog_stop(void) watchdog_active = 0; } - btime duration = last_time - loop_time; + btime duration = last_io_time - loop_time; if (duration > config->watchdog_warning) log(L_WARN "I/O loop cycle took %u.%03u ms for %d events", (uint) (duration TO_MS), (uint) (duration % 1000), event_log_num); @@ -2181,8 +2207,9 @@ void io_init(void) { init_list(&sock_list); - init_list(&global_event_list); - init_list(&global_work_list); + ev_init_list(&global_event_list, &main_birdloop, "Global event list"); + ev_init_list(&global_work_list, &main_birdloop, "Global work list"); + ev_init_list(&main_birdloop.event_list, &main_birdloop, "Global fast event list"); krt_io_init(); // XXX init_times(); // XXX update_times(); @@ -2210,22 +2237,27 @@ io_loop(void) watchdog_start1(); for(;;) { - times_update(&main_timeloop); + times_update(); events = ev_run_list(&global_event_list); events = ev_run_list_limited(&global_work_list, WORK_EVENTS_MAX) || events; - timers_fire(&main_timeloop); + events = ev_run_list(&main_birdloop.event_list) || events; + timers_fire(&main_birdloop.time, 1); io_close_event(); // FIXME poll_tout = (events ? 0 : 3000); /* Time in milliseconds */ - if (t = timers_first(&main_timeloop)) + if (t = timers_first(&main_birdloop.time)) { - times_update(&main_timeloop); + times_update(); timeout = (tm_remains(t) TO_MS) + 1; poll_tout = MIN(poll_tout, timeout); } - nfds = 0; + /* A hack to reload main io_loop() when something has changed asynchronously. */ + pipe_pollin(&main_birdloop.wakeup, &pfd[0]); + + nfds = 1; + WALK_LIST(n, sock_list) { pfd[nfds] = (struct pollfd) { .fd = -1 }; /* everything other set to 0 by this */ @@ -2284,18 +2316,28 @@ io_loop(void) /* And finally enter poll() to find active sockets */ watchdog_stop(); + birdloop_leave(&main_birdloop); pout = poll(pfd, nfds, poll_tout); + birdloop_enter(&main_birdloop); watchdog_start(); if (pout < 0) { if (errno == EINTR || errno == EAGAIN) continue; - die("poll: %m"); + bug("poll: %m"); } if (pout) { - times_update(&main_timeloop); + if (pfd[0].revents & POLLIN) + { + /* IO loop reload requested */ + pipe_drain(&main_birdloop.wakeup); + atomic_exchange_explicit(&main_birdloop.ping_sent, 0, memory_order_acq_rel); + continue; + } + + times_update(); /* guaranteed to be non-empty */ current_sock = SKIP_BACK(sock, n, HEAD(sock_list)); diff --git a/sysdep/unix/krt.Y b/sysdep/unix/krt.Y index 95b54d65..4ce9a328 100644 --- a/sysdep/unix/krt.Y +++ b/sysdep/unix/krt.Y @@ -29,7 +29,7 @@ kif_set_preferred(ip_addr ip) CF_DECLS -CF_KEYWORDS(KERNEL, PERSIST, SCAN, TIME, LEARN, DEVICE, ROUTES, GRACEFUL, RESTART, KRT_SOURCE, KRT_METRIC, MERGE, PATHS) +CF_KEYWORDS(KERNEL, PERSIST, SCAN, TIME, LEARN, DEVICE, ROUTES, GRACEFUL, RESTART, MERGE, PATHS) CF_KEYWORDS(INTERFACE, PREFERRED) %type <i> kern_mp_limit @@ -122,9 +122,6 @@ kif_iface: kif_iface_start iface_patt_list_nopx kif_iface_opt_list; -dynamic_attr: KRT_SOURCE { $$ = f_new_dynamic_attr(EAF_TYPE_INT, T_INT, EA_KRT_SOURCE); } ; -dynamic_attr: KRT_METRIC { $$ = f_new_dynamic_attr(EAF_TYPE_INT, T_INT, EA_KRT_METRIC); } ; - CF_CODE CF_END diff --git a/sysdep/unix/krt.c b/sysdep/unix/krt.c index cb0f1d4f..3b6c91cc 100644 --- a/sysdep/unix/krt.c +++ b/sysdep/unix/krt.c @@ -53,7 +53,7 @@ #include "nest/bird.h" #include "nest/iface.h" -#include "nest/route.h" +#include "nest/rt.h" #include "nest/protocol.h" #include "filter/filter.h" #include "conf/conf.h" @@ -232,7 +232,6 @@ kif_copy_config(struct proto_config *dest, struct proto_config *src) struct protocol proto_unix_iface = { .name = "Device", .template = "device%d", - .class = PROTOCOL_DEVICE, .proto_size = sizeof(struct kif_proto), .config_size = sizeof(struct kif_config), .preconfig = kif_preconfig, @@ -258,14 +257,14 @@ static inline void krt_trace_in(struct krt_proto *p, rte *e, char *msg) { if (p->p.debug & D_PACKETS) - log(L_TRACE "%s: %N: %s", p->p.name, e->net->n.addr, msg); + log(L_TRACE "%s: %N: %s", p->p.name, e->net, msg); } static inline void krt_trace_in_rl(struct tbf *f, struct krt_proto *p, rte *e, char *msg) { if (p->p.debug & D_PACKETS) - log_rl(f, L_TRACE "%s: %N: %s", p->p.name, e->net->n.addr, msg); + log_rl(f, L_TRACE "%s: %N: %s", p->p.name, e->net, msg); } /* @@ -287,7 +286,7 @@ static struct tbf rl_alien = TBF_DEFAULT_LOG_LIMITS; static inline u32 krt_metric(rte *a) { - eattr *ea = ea_find(a->attrs->eattrs, EA_KRT_METRIC); + eattr *ea = ea_find(a->attrs, &ea_krt_metric); return ea ? ea->u.data : 0; } @@ -304,242 +303,35 @@ krt_uptodate(rte *a, rte *b) } static void -krt_learn_announce_update(struct krt_proto *p, rte *e) -{ - net *n = e->net; - rta *aa = rta_clone(e->attrs); - rte *ee = rte_get_temp(aa, p->p.main_source); - rte_update(&p->p, n->n.addr, ee); -} - -static void -krt_learn_announce_delete(struct krt_proto *p, net *n) -{ - rte_update(&p->p, n->n.addr, NULL); -} - -static void krt_learn_alien_attr(struct channel *c, rte *e) { - ASSERT(!e->attrs->cached); - e->attrs->pref = c->preference; - - e->attrs = rta_lookup(e->attrs); + ea_set_attr_u32(&e->attrs, &ea_gen_preference, 0, c->preference); } /* Called when alien route is discovered during scan */ static void krt_learn_scan(struct krt_proto *p, rte *e) { - net *n0 = e->net; - net *n = net_get(p->krt_table, n0->n.addr); - rte *m, **mm; + rte e0 = { + .attrs = e->attrs, + .src = rt_get_source(&p->p, krt_metric(e)), + }; - krt_learn_alien_attr(p->p.main_channel, e); + krt_learn_alien_attr(p->p.main_channel, &e0); - for(mm=&n->routes; m = *mm; mm=&m->next) - if (krt_same_key(m, e)) - break; - if (m) - { - if (krt_uptodate(m, e)) - { - krt_trace_in_rl(&rl_alien, p, e, "[alien] seen"); - rte_free(e); - m->pflags |= KRT_REF_SEEN; - } - else - { - krt_trace_in(p, e, "[alien] updated"); - *mm = m->next; - rte_free(m); - m = NULL; - } - } - else - krt_trace_in(p, e, "[alien] created"); - if (!m) - { - e->next = n->routes; - n->routes = e; - e->pflags |= KRT_REF_SEEN; - } -} - -static void -krt_learn_prune(struct krt_proto *p) -{ - struct fib *fib = &p->krt_table->fib; - struct fib_iterator fit; - - KRT_TRACE(p, D_EVENTS, "Pruning inherited routes"); - - FIB_ITERATE_INIT(&fit, fib); -again: - FIB_ITERATE_START(fib, &fit, net, n) - { - rte *e, **ee, *best, **pbest, *old_best; - - /* - * Note that old_best may be NULL even if there was an old best route in - * the previous step, because it might be replaced in krt_learn_scan(). - * But in that case there is a new valid best route. - */ - - old_best = NULL; - best = NULL; - pbest = NULL; - ee = &n->routes; - while (e = *ee) - { - if (e->pflags & KRT_REF_BEST) - old_best = e; - - if (!(e->pflags & KRT_REF_SEEN)) - { - *ee = e->next; - rte_free(e); - continue; - } - - if (!best || krt_metric(best) > krt_metric(e)) - { - best = e; - pbest = ee; - } - - e->pflags &= ~(KRT_REF_SEEN | KRT_REF_BEST); - ee = &e->next; - } - if (!n->routes) - { - DBG("%I/%d: deleting\n", n->n.prefix, n->n.pxlen); - if (old_best) - krt_learn_announce_delete(p, n); - - FIB_ITERATE_PUT(&fit); - fib_delete(fib, n); - goto again; - } - - best->pflags |= KRT_REF_BEST; - *pbest = best->next; - best->next = n->routes; - n->routes = best; - - if ((best != old_best) || p->reload) - { - DBG("%I/%d: announcing (metric=%d)\n", n->n.prefix, n->n.pxlen, krt_metric(best)); - krt_learn_announce_update(p, best); - } - else - DBG("%I/%d: uptodate (metric=%d)\n", n->n.prefix, n->n.pxlen, krt_metric(best)); - } - FIB_ITERATE_END; - - p->reload = 0; + rte_update(p->p.main_channel, e->net, &e0, e0.src); + rt_unlock_source(e0.src); } static void krt_learn_async(struct krt_proto *p, rte *e, int new) { - net *n0 = e->net; - net *n = net_get(p->krt_table, n0->n.addr); - rte *g, **gg, *best, **bestp, *old_best; - - krt_learn_alien_attr(p->p.main_channel, e); - - old_best = n->routes; - for(gg=&n->routes; g = *gg; gg = &g->next) - if (krt_same_key(g, e)) - break; if (new) - { - if (g) - { - if (krt_uptodate(g, e)) - { - krt_trace_in(p, e, "[alien async] same"); - rte_free(e); - return; - } - krt_trace_in(p, e, "[alien async] updated"); - *gg = g->next; - rte_free(g); - } - else - krt_trace_in(p, e, "[alien async] created"); - - e->next = n->routes; - n->routes = e; - } - else if (!g) - { - krt_trace_in(p, e, "[alien async] delete failed"); - rte_free(e); - return; - } - else - { - krt_trace_in(p, e, "[alien async] removed"); - *gg = g->next; - rte_free(e); - rte_free(g); - } - best = n->routes; - bestp = &n->routes; - for(gg=&n->routes; g=*gg; gg=&g->next) - { - if (krt_metric(best) > krt_metric(g)) - { - best = g; - bestp = gg; - } - - g->pflags &= ~KRT_REF_BEST; - } - - if (best) - { - best->pflags |= KRT_REF_BEST; - *bestp = best->next; - best->next = n->routes; - n->routes = best; - } + return krt_learn_scan(p, e); - if (best != old_best) - { - DBG("krt_learn_async: distributing change\n"); - if (best) - krt_learn_announce_update(p, best); - else - krt_learn_announce_delete(p, n); - } -} - -static void -krt_learn_init(struct krt_proto *p) -{ - if (KRT_CF->learn) - { - struct rtable_config *cf = mb_allocz(p->p.pool, sizeof(struct rtable_config)); - cf->name = "Inherited"; - cf->addr_type = p->p.net_type; - cf->internal = 1; - - p->krt_table = rt_setup(p->p.pool, cf); - } -} - -static void -krt_dump(struct proto *P) -{ - struct krt_proto *p = (struct krt_proto *) P; - - if (!KRT_CF->learn) - return; - debug("KRT: Table of inheritable routes\n"); - rt_dump(p->krt_table); + struct rte_src *src = rt_get_source(&p->p, krt_metric(e)); + rte_update(p->p.main_channel, e->net, NULL, src); + rt_unlock_source(src); } #endif @@ -551,40 +343,60 @@ krt_dump(struct proto *P) static inline int krt_is_installed(struct krt_proto *p, net *n) { - return n->routes && bmap_test(&p->p.main_channel->export_map, n->routes->id); + return n->routes && bmap_test(&p->p.main_channel->export_map, n->routes->rte.id); } -static void -krt_flush_routes(struct krt_proto *p) +static uint +rte_feed_count(net *n) { - struct rtable *t = p->p.main_channel->table; + uint count = 0; + for (struct rte_storage *e = n->routes; e; e = e->next) + if (rte_is_valid(RTE_OR_NULL(e))) + count++; + return count; +} - KRT_TRACE(p, D_EVENTS, "Flushing kernel routes"); - FIB_WALK(&t->fib, net, n) +static void +rte_feed_obtain(net *n, rte **feed, uint count) +{ + uint i = 0; + for (struct rte_storage *e = n->routes; e; e = e->next) + if (rte_is_valid(RTE_OR_NULL(e))) { - if (krt_is_installed(p, n)) - { - /* FIXME: this does not work if gw is changed in export filter */ - krt_replace_rte(p, n, NULL, n->routes); - } + ASSERT_DIE(i < count); + feed[i++] = &e->rte; } - FIB_WALK_END; + ASSERT_DIE(i == count); } static struct rte * -krt_export_net(struct krt_proto *p, net *net, rte **rt_free) +krt_export_net(struct krt_proto *p, net *net) { + /* FIXME: Here we are calling filters in table-locked context when exporting + * to kernel. Here BIRD can crash if the user requested ROA check in kernel + * export filter. It doesn't make much sense to write the filters like this, + * therefore we may keep this unfinished piece of work here for later as it + * won't really affect anybody. */ + ASSERT_DIE(RT_IS_LOCKED(p->p.main_channel->table)); + struct channel *c = p->p.main_channel; const struct filter *filter = c->out_filter; - rte *rt; if (c->ra_mode == RA_MERGED) - return rt_export_merged(c, net, rt_free, krt_filter_lp, 1); + { + uint count = rte_feed_count(net); + if (!count) + return NULL; - rt = net->routes; - *rt_free = NULL; + rte **feed = alloca(count * sizeof(rte *)); + rte_feed_obtain(net, feed, count); + return rt_export_merged(c, feed, count, krt_filter_lp, 1); + } - if (!rte_is_valid(rt)) + static _Thread_local rte rt; + rt = net->routes->rte; + + if (!rte_is_valid(&rt)) return NULL; if (filter == FILTER_REJECT) @@ -595,33 +407,26 @@ krt_export_net(struct krt_proto *p, net *net, rte **rt_free) if (filter == FILTER_ACCEPT) goto accept; - if (f_run(filter, &rt, krt_filter_lp, FF_SILENT) > F_ACCEPT) + if (f_run(filter, &rt, FF_SILENT) > F_ACCEPT) goto reject; accept: - if (rt != net->routes) - *rt_free = rt; - return rt; + return &rt; reject: - if (rt != net->routes) - rte_free(rt); return NULL; } static int krt_same_dest(rte *k, rte *e) { - rta *ka = k->attrs, *ea = e->attrs; - - if (ka->dest != ea->dest) - return 0; + ea_list *ka = k->attrs, *ea = e->attrs; - if (ka->dest == RTD_UNICAST) - return nexthop_same(&(ka->nh), &(ea->nh)); + eattr *nhea_k = ea_find(ka, &ea_gen_nexthop); + eattr *nhea_e = ea_find(ea, &ea_gen_nexthop); - return 1; + return (!nhea_k == !nhea_e) && adata_same(nhea_k->u.ptr, nhea_e->u.ptr); } /* @@ -632,8 +437,7 @@ krt_same_dest(rte *k, rte *e) void krt_got_route(struct krt_proto *p, rte *e, s8 src) { - rte *new = NULL, *rt_free = NULL; - net *n = e->net; + rte *new = NULL; e->pflags = 0; #ifdef KRT_ALLOW_LEARN @@ -649,24 +453,28 @@ krt_got_route(struct krt_proto *p, rte *e, s8 src) if (KRT_CF->learn) krt_learn_scan(p, e); else - { - krt_trace_in_rl(&rl_alien, p, e, "[alien] ignored"); - rte_free(e); - } + krt_trace_in_rl(&rl_alien, p, e, "[alien] ignored"); return; } #endif /* The rest is for KRT_SRC_BIRD (or KRT_SRC_UNKNOWN) */ + RT_LOCKED(p->p.main_channel->table, tab) + { + + /* Deleting all routes if flush is requested */ + if (p->flush_routes) + goto delete; /* We wait for the initial feed to have correct installed state */ if (!p->ready) goto ignore; - if (!krt_is_installed(p, n)) + net *net = net_find(tab, e->net); + if (!net || !krt_is_installed(p, net)) goto delete; - new = krt_export_net(p, n, &rt_free); + new = krt_export_net(p, net); /* Rejected by filters */ if (!new) @@ -699,19 +507,16 @@ ignore: update: krt_trace_in(p, new, "updating"); - krt_replace_rte(p, n, new, e); + krt_replace_rte(p, e->net, new, e); goto done; delete: krt_trace_in(p, e, "deleting"); - krt_replace_rte(p, n, NULL, e); + krt_replace_rte(p, e->net, NULL, e); goto done; -done: - rte_free(e); - - if (rt_free) - rte_free(rt_free); +done:; + } lp_flush(krt_filter_lp); } @@ -725,43 +530,47 @@ krt_init_scan(struct krt_proto *p) static void krt_prune(struct krt_proto *p) { - struct rtable *t = p->p.main_channel->table; + RT_LOCKED(p->p.main_channel->table, t) + { KRT_TRACE(p, D_EVENTS, "Pruning table %s", t->name); FIB_WALK(&t->fib, net, n) { - if (p->ready && krt_is_installed(p, n) && !bmap_test(&p->seen_map, n->routes->id)) + if (p->ready && krt_is_installed(p, n) && !bmap_test(&p->seen_map, n->routes->rte.id)) { - rte *rt_free = NULL; - rte *new = krt_export_net(p, n, &rt_free); + rte *new = krt_export_net(p, n); if (new) { krt_trace_in(p, new, "installing"); - krt_replace_rte(p, n, new, NULL); + krt_replace_rte(p, n->n.addr, new, NULL); } - if (rt_free) - rte_free(rt_free); - lp_flush(krt_filter_lp); } } FIB_WALK_END; -#ifdef KRT_ALLOW_LEARN - if (KRT_CF->learn) - krt_learn_prune(p); -#endif - if (p->ready) p->initialized = 1; + + } +} + +static void +krt_flush_routes(struct krt_proto *p) +{ + KRT_TRACE(p, D_EVENTS, "Flushing kernel routes"); + p->flush_routes = 1; + krt_init_scan(p); + krt_do_scan(p); + /* No prune! */ + p->flush_routes = 0; } void krt_got_route_async(struct krt_proto *p, rte *e, int new, s8 src) { - net *net = e->net; e->pflags = 0; switch (src) @@ -774,7 +583,7 @@ krt_got_route_async(struct krt_proto *p, rte *e, int new, s8 src) if (new) { krt_trace_in(p, e, "[redirect] deleting"); - krt_replace_rte(p, net, NULL, e); + krt_replace_rte(p, e->net, NULL, e); } /* If !new, it is probably echo of our deletion */ break; @@ -788,7 +597,6 @@ krt_got_route_async(struct krt_proto *p, rte *e, int new, s8 src) } #endif } - rte_free(e); } @@ -910,8 +718,7 @@ krt_scan_timer_kick(struct krt_proto *p) static int krt_preexport(struct channel *C, rte *e) { - // struct krt_proto *p = (struct krt_proto *) P; - if (e->src->proto == C->proto) + if (e->src->owner == &C->proto->sources) return -1; if (!krt_capable(e)) @@ -921,8 +728,8 @@ krt_preexport(struct channel *C, rte *e) } static void -krt_rt_notify(struct proto *P, struct channel *ch UNUSED, net *net, - rte *new, rte *old) +krt_rt_notify(struct proto *P, struct channel *ch UNUSED, const net_addr *net, + rte *new, const rte *old) { struct krt_proto *p = (struct krt_proto *) P; @@ -985,6 +792,14 @@ krt_feed_end(struct channel *C) krt_scan_timer_kick(p); } +static int +krt_rte_better(rte *new, rte *old) +{ + u32 n = ea_get_int(new->attrs, &ea_krt_metric, IGP_METRIC_UNKNOWN); + u32 o = ea_get_int(old->attrs, &ea_krt_metric, IGP_METRIC_UNKNOWN); + + return (n < o); +} /* * Protocol glue @@ -1026,6 +841,10 @@ krt_postconfig(struct proto_config *CF) krt_sys_postconfig(cf); } +struct rte_owner_class krt_rte_owner_class = { + .rte_better = krt_rte_better, +}; + static struct proto * krt_init(struct proto_config *CF) { @@ -1040,6 +859,8 @@ krt_init(struct proto_config *CF) p->p.reload_routes = krt_reload_routes; p->p.feed_end = krt_feed_end; + p->p.sources.class = &krt_rte_owner_class; + krt_sys_init(p); return &p->p; } @@ -1064,10 +885,6 @@ krt_start(struct proto *P) bmap_init(&p->seen_map, p->p.pool, 1024); add_tail(&krt_proto_list, &p->krt_node); -#ifdef KRT_ALLOW_LEARN - krt_learn_init(p); -#endif - if (!krt_sys_start(p)) { rem_node(&p->krt_node); @@ -1147,24 +964,15 @@ krt_copy_config(struct proto_config *dest, struct proto_config *src) krt_sys_copy_config(d, s); } -static int -krt_get_attr(const eattr *a, byte *buf, int buflen) -{ - switch (a->id) - { - case EA_KRT_SOURCE: - bsprintf(buf, "source"); - return GA_NAME; - - case EA_KRT_METRIC: - bsprintf(buf, "metric"); - return GA_NAME; - - default: - return krt_sys_get_attr(a, buf, buflen); - } -} +struct ea_class ea_krt_source = { + .name = "krt_source", + .type = T_INT, +}; +struct ea_class ea_krt_metric = { + .name = "krt_metric", + .type = T_INT, +}; #ifdef CONFIG_IP6_SADR_KERNEL #define MAYBE_IP6_SADR NB_IP6_SADR @@ -1181,7 +989,6 @@ krt_get_attr(const eattr *a, byte *buf, int buflen) struct protocol proto_unix_kernel = { .name = "Kernel", .template = "kernel%d", - .class = PROTOCOL_KERNEL, .preference = DEF_PREF_INHERITED, .channel_mask = NB_IP | MAYBE_IP6_SADR | MAYBE_MPLS, .proto_size = sizeof(struct krt_proto), @@ -1193,14 +1000,15 @@ struct protocol proto_unix_kernel = { .shutdown = krt_shutdown, .reconfigure = krt_reconfigure, .copy_config = krt_copy_config, - .get_attr = krt_get_attr, -#ifdef KRT_ALLOW_LEARN - .dump = krt_dump, -#endif }; void krt_build(void) { proto_build(&proto_unix_kernel); + + EA_REGISTER_ALL( + &ea_krt_source, + &ea_krt_metric, + ); } diff --git a/sysdep/unix/krt.h b/sysdep/unix/krt.h index 18a206e6..9f7ebb4f 100644 --- a/sysdep/unix/krt.h +++ b/sysdep/unix/krt.h @@ -21,8 +21,7 @@ struct kif_proto; #define KRT_DEFAULT_ECMP_LIMIT 16 -#define EA_KRT_SOURCE EA_CODE(PROTOCOL_KERNEL, 0) -#define EA_KRT_METRIC EA_CODE(PROTOCOL_KERNEL, 1) +extern struct ea_class ea_krt_source, ea_krt_metric; #define KRT_REF_SEEN 0x1 /* Seen in table */ #define KRT_REF_BEST 0x2 /* Best in table */ @@ -51,10 +50,6 @@ struct krt_proto { struct proto p; struct krt_state sys; /* Sysdep state */ -#ifdef KRT_ALLOW_LEARN - struct rtable *krt_table; /* Internal table of inherited routes */ -#endif - timer *scan_timer; struct bmap sync_map; /* Keeps track which exported routes were successfully written to kernel */ struct bmap seen_map; /* Routes seen during last periodic scan */ @@ -63,6 +58,7 @@ struct krt_proto { byte ready; /* Initial feed has been finished */ byte initialized; /* First scan has been finished */ byte reload; /* Next scan is doing reload */ + byte flush_routes; /* Scanning to flush */ }; extern pool *krt_pool; @@ -141,7 +137,7 @@ void krt_sys_copy_config(struct krt_config *, struct krt_config *); int krt_capable(rte *e); void krt_do_scan(struct krt_proto *); -void krt_replace_rte(struct krt_proto *p, net *n, rte *new, rte *old); +void krt_replace_rte(struct krt_proto *p, const net_addr *n, rte *new, const rte *old); int krt_sys_get_attr(const eattr *a, byte *buf, int buflen); diff --git a/sysdep/unix/log.c b/sysdep/unix/log.c index 4e9df069..185231e8 100644 --- a/sysdep/unix/log.c +++ b/sysdep/unix/log.c @@ -15,6 +15,7 @@ * user's manual. */ +#include <stdatomic.h> #include <stdio.h> #include <stdlib.h> #include <stdarg.h> @@ -35,8 +36,10 @@ static FILE *dbgf; static list *current_log_list; static char *current_syslog_name; /* NULL -> syslog closed */ +static _Atomic uint max_thread_id = ATOMIC_VAR_INIT(1); +static _Thread_local uint this_thread_id; -#ifdef USE_PTHREADS +#define THIS_THREAD_ID (this_thread_id ?: (this_thread_id = atomic_fetch_add_explicit(&max_thread_id, 1, memory_order_acq_rel))) #include <pthread.h> @@ -48,15 +51,6 @@ static pthread_t main_thread; void main_thread_init(void) { main_thread = pthread_self(); } static int main_thread_self(void) { return pthread_equal(pthread_self(), main_thread); } -#else - -static inline void log_lock(void) { } -static inline void log_unlock(void) { } -void main_thread_init(void) { } -static int main_thread_self(void) { return 1; } - -#endif - #ifdef HAVE_SYSLOG_H #include <sys/syslog.h> @@ -189,7 +183,7 @@ log_commit(int class, buffer *buf) l->pos += msg_len; } - fprintf(l->fh, "%s <%s> ", tbuf, class_names[class]); + fprintf(l->fh, "%s [%04x] <%s> ", tbuf, THIS_THREAD_ID, class_names[class]); } fputs(buf->start, l->fh); fputc('\n', l->fh); @@ -299,6 +293,8 @@ die(const char *msg, ...) exit(1); } +static struct timespec dbg_time_start; + /** * debug - write to debug output * @msg: a printf-like message @@ -311,12 +307,33 @@ debug(const char *msg, ...) { #define MAX_DEBUG_BUFSIZE 16384 va_list args; - char buf[MAX_DEBUG_BUFSIZE]; + char buf[MAX_DEBUG_BUFSIZE], *pos = buf; + int max = MAX_DEBUG_BUFSIZE; va_start(args, msg); if (dbgf) { - if (bvsnprintf(buf, MAX_DEBUG_BUFSIZE, msg, args) < 0) + struct timespec dbg_time; + clock_gettime(CLOCK_MONOTONIC, &dbg_time); + uint nsec; + uint sec; + + if (dbg_time.tv_nsec > dbg_time_start.tv_nsec) + { + nsec = dbg_time.tv_nsec - dbg_time_start.tv_nsec; + sec = dbg_time.tv_sec - dbg_time_start.tv_sec; + } + else + { + nsec = 1000000000 + dbg_time.tv_nsec - dbg_time_start.tv_nsec; + sec = dbg_time.tv_sec - dbg_time_start.tv_sec - 1; + } + + int n = bsnprintf(pos, max, "%u.%09u: [%04x] ", sec, nsec, THIS_THREAD_ID); + pos += n; + max -= n; + + if (bvsnprintf(pos, max, msg, args) < 0) bug("Extremely long debug output, split it."); fputs(buf, dbgf); @@ -422,6 +439,8 @@ done: void log_init_debug(char *f) { + clock_gettime(CLOCK_MONOTONIC, &dbg_time_start); + if (dbgf && dbgf != stderr) fclose(dbgf); if (!f) diff --git a/sysdep/unix/main.c b/sysdep/unix/main.c index 84e7d889..bf9f2be0 100644 --- a/sysdep/unix/main.c +++ b/sysdep/unix/main.c @@ -28,9 +28,10 @@ #include "lib/resource.h" #include "lib/socket.h" #include "lib/event.h" +#include "lib/locking.h" #include "lib/timer.h" #include "lib/string.h" -#include "nest/route.h" +#include "nest/rt.h" #include "nest/protocol.h" #include "nest/iface.h" #include "nest/cli.h" @@ -56,7 +57,7 @@ async_dump(void) // XXXX tm_dump_all(); if_dump_all(); neigh_dump_all(); - rta_dump_all(); + ea_dump_all(); rt_dump_all(); protos_dump_all(); @@ -873,16 +874,19 @@ main(int argc, char **argv) dmalloc_debug(0x2f03d00); #endif + times_update(); parse_args(argc, argv); log_switch(1, NULL, NULL); + the_bird_lock(); + random_init(); net_init(); resource_init(); - timer_init(); + birdloop_init(); olock_init(); - io_init(); rt_init(); + io_init(); if_init(); // roa_init(); config_init(); @@ -927,6 +931,7 @@ main(int argc, char **argv) dup2(0, 2); } + main_thread_init(); write_pid_file(); |