Use buffer traits over specific IDs

[platform/core/system/dlog.git] / src / libdlog / log.c

/*  MIT License
 *
 * Copyright (c) 2012-2020 Samsung Electronics Co., Ltd
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is furnished
 * to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 * THE SOFTWARE. */

// C
#include <assert.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>

// POSIX
#include <pthread.h>
#include <sys/wait.h>
#include <unistd.h>

// Tizen
#include <tizen.h>

// DLog
#include <buffer_traits.h>
#include "deduplicate.h"
#include <dynamic_config.h>
#include <libdlog.h>
#include "logconfig.h"
#include "loglimiter.h"

#define DEFAULT_CONFIG_LIMITER false
#define DEFAULT_CONFIG_PLOG true
#define DEFAULT_CONFIG_DEBUGMODE 0
#define DEFAULT_CONFIG_LIMITER_APPLY_TO_ALL_BUFFERS 0


/* A pseudo-backend that does nothing. Useful for removing the overhead of dlog
 * for debugging and measurement purposes, also it simplifies some checking as
 * the `write_to_log` pointer below never needs to be NULL. Note that features
 * independent of the backend (such as limiter or dynamic config) are controlled
 * separately if dlog is to be disabled completely and that minimal overhead is
 * always present (e.g. building the message via vprintf). */
static int write_to_log_null (log_id_t log_id, log_priority prio, const char *tag, const char *msg, struct timespec *tp_mono, int32_t pid, int32_t tid)
{
        return DLOG_ERROR_NONE;
}

/**
 * @brief Points to a function which writes a log message
 * @details The function pointed to depends on the backend used
 * @param[in] log_id ID of the buffer to log to. Belongs to (LOG_ID_INVALID, LOG_ID_MAX) non-inclusive
 * @param[in] prio Priority of the message.
 * @param[in] tag The message tag, identifies the sender.
 * @param[in] msg The contents of the message.
 * @return Returns the number of bytes written on success and a negative error value on error.
 * @see __dlog_init_backend
 */
int (*write_to_log)(log_id_t log_id, log_priority prio, const char *tag, const char *msg, struct timespec *tp_mono, int32_t pid, int32_t tid) = write_to_log_null;
void (*destroy_backend)(void);

int (*stash_failed_log)(log_id_t log_id, log_priority prio, const char *tag, const char *msg) = NULL;
#ifndef UNIT_TEST
static int stash_critical(log_id_t log_id, log_priority prio, const char *tag, const char *msg);
#endif

pthread_rwlock_t log_limiter_lock = PTHREAD_RWLOCK_INITIALIZER;
static pthread_mutex_t log_construction_lock = PTHREAD_MUTEX_INITIALIZER;
static bool is_initialized = false;

extern void __dlog_init_pipe(const struct log_config *conf);
extern void __dlog_init_android(const struct log_config *conf);
extern void __dlog_init_zero_copy();

extern void prepend_container_tag_if_in_container(size_t buf_size, char buf[static buf_size], int *len);

bool limiter;
struct limiter_data *limiter_data;
static bool dynamic_config;
static bool plog[LOG_ID_MAX];
static bool plog_default_values[LOG_ID_MAX];
static bool enable_secure_logs = true;
static bool enable_critical = true;

static bool should_disable_cancels;

static int debugmode;
static int fatal_assert;
static int limiter_apply_to_all_buffers;
static _Atomic log_priority priority_filter_level = DLOG_VERBOSE;

/* Cache pid and tid to avoid up to two syscalls per log. The update function
 * is registered to happen when new threads are made (including a full fork). */
static int32_t cached_pid = 0;
_Thread_local int32_t cached_tid = 0;

static inline int32_t get_cached_pid()
{
        return (cached_pid = cached_pid ?: getpid());
}

static inline int32_t get_cached_tid()
{
        return (cached_tid = cached_tid ?: gettid());
}

static void update_thread_local_ids(void)
{
        cached_pid = 0;
        cached_tid = 0;
}

/* Here, static_config is the original config from /etc/dlog.conf{,.d} which can be overriden,
 * but comes back if the override is removed. both_config additionally contains dynamic rules
 * (by default from /run/dlog/filters.d) which can be changed in the runtime.
 * The static_config has to be kept separately, so that we can go back to it when dynamic rules change.
 * Note that most functions only use static_config, since the parameters can't be changed in runtime. */
static void __configure_limiter(struct log_config *static_config, struct log_config *both_config)
{
        assert(static_config);
        if (dynamic_config)
                assert(both_config);

        if (!limiter)
                return;

        limiter_data = __log_limiter_create(static_config);
        if (limiter_data && dynamic_config)
                __log_limiter_update(limiter_data, both_config);
        limiter = (bool)limiter_data;

        should_disable_cancels |= limiter; // due to locks
}

static int __configure_backend(struct log_config *config)
{
        assert(config);

        const char *const backend = log_config_claim_backend(config);
        if (!backend)
                return 0;

        if (!strcmp(backend, "pipe")) {
                __dlog_init_pipe(config);
                should_disable_cancels = true;
        } else if (!strcmp(backend, "logger"))
                __dlog_init_android(config);
        else if (!strcmp(backend, "zero-copy"))
                __dlog_init_zero_copy(config);
        else if (!strcmp(backend, "null"))
                ; // already the default
        else
                return 0;

        return 1;
}

static void __set_plog_default_values(void)
{
        for (size_t i = 0; i < NELEMS(plog); ++i)
                plog_default_values[i] = plog[i];
}

static void __initialize_plog(const struct log_config *config)
{
        assert(config);

        /* plog stands for "platform logging", which is why
         * non-platform (i.e. application) buffers ignore it */
        const bool plog_default = log_config_get_boolean(config, "plog", DEFAULT_CONFIG_PLOG);
        for (size_t i = 0; i < NELEMS(plog); ++i)
                plog[i] = is_platform_buffer(i) ? plog_default : true;

        __set_plog_default_values();
}

static void __configure_parameters(struct log_config *static_config, struct log_config *both_config)
{
        assert(static_config);
        assert(both_config);

        __initialize_plog(static_config);
        __update_plog(static_config);
        /* Like in __configure_limiter, we also check the dynamic rules. However, we make sure to
     * set the default values to the ones generated by the static rules first. */
        __set_plog_default_values();
        __update_plog(both_config);

        enable_secure_logs = log_config_get_boolean(both_config, "enable_secure_logs", enable_secure_logs);
        enable_critical = log_config_get_boolean(both_config, "enable_critical", enable_critical);
        debugmode = log_config_get_int(both_config, "debugmode", DEFAULT_CONFIG_DEBUGMODE);
        fatal_assert = access(DEBUGMODE_FILE, F_OK) != -1;
        limiter = log_config_get_boolean(both_config, "limiter", DEFAULT_CONFIG_LIMITER);
        limiter_apply_to_all_buffers = log_config_get_int(both_config,
                                                                        "limiter_apply_to_all_buffers",
                                                                        DEFAULT_CONFIG_LIMITER_APPLY_TO_ALL_BUFFERS);

        stash_failed_log = NULL;
        const char *stash_failed_log_method = log_config_get(both_config, "stash_failed_log_method");
        if (stash_failed_log_method) {
#ifndef UNIT_TEST
                if (strcmp(stash_failed_log_method, "critical") == 0)
                        stash_failed_log = stash_critical;
#endif
        }
}

void __update_plog(const struct log_config *conf)
{
        assert(conf);

        for (size_t i = 0; i < NELEMS(plog); ++i) {
                char key[MAX_CONF_KEY_LEN];
                const int r = snprintf(key, sizeof key, "enable_%s", log_name_by_id((log_id_t)i));
                if (r < 0)
                        continue;
                plog[i] = log_config_get_boolean(conf, key, plog_default_values[i]);
        }
}

/**
 * @brief Configure the library
 * @details Reads relevant config values
 * @remarks This is more or less a constructor, but there are some obstacles
 *          to using it as such (i.e. with attribute constructor):
 *
 *  - some important pieces of the system link to dlog, they start very early
 *    such that dlog can't properly initialize (which lasts for program lifetime)
 *    but don't actually log anything until later on and would be fine under lazy
 *    initialisation. The way to do it "properly" would be to expose this function
 *    into the API so that people can manually call it when they're ready, but
 *    one of the design goals of the current API is that it requires absolutely no
 *    other calls than `dlog_print`. Changing it would require somebody with a
 *    bird's eye view of the system to produce a design so I wouldn't count on it.
 *
 *  - the constructor would need to have as high of a priority as possible (so as
 *    to minimize the risk of another library's constructor using uninitialized data)
 *    but at the same time others might want some room to wrap functions before
 *    dlog uses them (think mprobe/mcheck). This would also require a design pass.
 */
#ifndef UNIT_TEST
static
#endif
bool __configure(void)
{
        __attribute__((cleanup(log_config_free))) struct log_config static_config = {};
        __attribute__((cleanup(log_config_free))) struct log_config both_config = {};

        if (log_config_read(&static_config) < 0)
                return false;
        log_config_copy(&both_config, &static_config);

        dynamic_config = __dynamic_config_create(&both_config);

        __configure_parameters(&static_config, &both_config);

        if (!__configure_backend(&both_config)) {
                __dynamic_config_destroy();
                dynamic_config = false;
                return false;
        }

        __configure_deduplicate(&both_config);
        __configure_limiter(&static_config, &both_config);

        /* The fork handler also runs when a new thread is made (not just
         * when the whole process forks), so we can cache both pid and tid. */
        pthread_atfork(NULL, NULL, update_thread_local_ids);

        return true;
}

static void __attribute__((constructor(101))) __install_pipe_handler(void)
{
        /* We mask SIGPIPE signal because most applications do not install their
         * own SIGPIPE handler. Default behaviour in SIGPIPE case is to abort the
         * process. SIGPIPE occurs when e.g. dlog daemon closes read pipe endpoint.
         *
         * We do this in the library constructor (at maximum priority) and not
         * during regular (lazy) initialisation so as to prevent overwriting the
         * program's actual signal handler, if it has one.
         *
         * In theory this is not required for the Android logger backend; however,
         * this early we don't yet know the backend and also it is good to behave
         * consistently in this regard anyway.
         *
         * We don't revert this in a destructor because Unix signals are bonkers
         * and we have no way to do this cleanly. Most libdlog users don't use
         * runtime linking so this would mostly done at program exit either way. */
        signal(SIGPIPE, SIG_IGN);
}

static void __attribute__((constructor(102))) __set_output_buffering(void)
{
        /* If stdout and/or stderr is redirected to dlog (service, driver)
         * it is best if we buffer lines, otherwise the following can happen:
         *  - no buffering: service/driver recives every single byte (precisely
         *    every single write*(2) call), causing it to either buffer the data
         *    anyway (service), or print logs at random places (driver)
         *  - full buffering: service/driver receives several hundred lines,
         *    which results in one giant entry being added (driver) or long delays
         *    before logs appear (service) */
        if (getenv(DLOG_ENV_STDOUT_LINE_BUFFERED))
                setlinebuf(stdout);

        if (getenv(DLOG_ENV_STDERR_LINE_BUFFERED))
                setlinebuf(stderr);
}

static bool first = true;
static bool initialize(void)
{
        if (is_initialized)
                return true;

        /* The mutex acts as a barrier, but otherwise the C language's
         * machine abstraction is single-threaded. This means that the
         * compiler is free to rearrange calls inside the mutex according
         * to the as-if rule because it doesn't care if another thread can
         * access it in parallel. In particular, `is_initialized = true`
         * directly after `__configure()` could be rearranged to go in
         * front of it because it is not touched inside that function
         * if the compiler thinks it helps somehow (not unlikely: since
         * it is checked before the mutex, it is very probable for it to
         * still be in the CPU register or something like that). On top
         * of that, some architectures (in particular, armv7l) don't have
         * strict memory guarantees and can reorder actual memory stores
         * on their own, even if the compiler didn't do anything fancy
         * when creating machine code. For more info about the issue,
         * see https://www.aristeia.com/Papers/DDJ_Jul_Aug_2004_revised.pdf
         *
         * Ultimately this means that there needs to be some sort of
         * barrier between `__configure` and `is_initialized = true`,
         * and the simplest way to achieve that is to just wait until
         * the second entry into the mutex. */

        bool ret;
        pthread_mutex_lock(&log_construction_lock);
                if (first)
                        first = !__configure();
                else
                        is_initialized = true;
                ret = !first;
        pthread_mutex_unlock(&log_construction_lock);
        return ret;
}

/**
 * @brief Fatal assertion
 * @details Conditionally crash the sucka who sent the log
 * @param[in] prio Priority of the log
 */
static void __dlog_fatal_assert(int prio)
{
        assert(!fatal_assert || (prio != DLOG_FATAL));
}

/**
 * @brief Check log validity
 * @details Checks whether the log is valid and eligible for printing
 * @param[in] log_id The target buffer ID
 * @param[in] prio The log's priority
 * @param[in] tag The log's tag
 * @return DLOG_ERROR_NONE on success, else an error code.
 * @retval DLOG_ERROR_INVALID_PARAMETER Invalid parameter
 */
static int dlog_check_validity(log_id_t log_id, int prio, const char *tag)
{
        (void) prio;
        if (!tag)
                return DLOG_ERROR_INVALID_PARAMETER;

        if (!is_buffer_valid(log_id))
                return DLOG_ERROR_INVALID_PARAMETER;

        return DLOG_ERROR_NONE;
}

/**
 * @brief Check log against limiter rules
 * @details Checks whether the log passes current limiter rules
 * @param[in] log_id The target buffer ID
 * @param[in] prio The log's priority
 * @param[in] tag The log's tag
 * @return DLOG_ERROR_NONE on success, else an error code.
 * @retval DLOG_ERROR_NOT_PERMITTED Not permitted
 */
static int dlog_check_limiter(log_id_t log_id, int prio, const char *tag)
{
        if (!debugmode && prio <= DLOG_DEBUG)
                return DLOG_ERROR_NOT_PERMITTED;

        if (dynamic_config)
                __dynamic_config_update(limiter_data);

        if (limiter) {
                struct pass_log_result should_log = { .decision = DECISION_DENIED };

                /* Since the only `wrlock` is done by the dynamic config, we can avoid
                 * the `rdlock` entirely if the config is static. This sounds unsafe
                 * but lets us save an entire syscall, which is a lot (both comparatively
                 * and because it compounds). */
                if (!dynamic_config || !pthread_rwlock_rdlock(&log_limiter_lock)) {
                        should_log = __log_limiter_pass_log_pid(limiter_data, tag, prio, get_cached_pid());
                        if (dynamic_config)
                                pthread_rwlock_unlock(&log_limiter_lock);
                }

                switch (should_log.decision) {
                        case DECISION_DENIED:
                                return DLOG_ERROR_NOT_PERMITTED;

                        case DECISION_TAG_LIMIT_EXCEEDED_MESSAGE:
                        case DECISION_PID_LIMIT_EXCEEDED_MESSAGE: {
                                struct timespec tp;
                                int result = clock_gettime(CLOCK_MONOTONIC, &tp);
                                if (result < 0)
                                        return DLOG_ERROR_NOT_PERMITTED;
                                char buf[100] = {};
                                snprintf(buf, sizeof(buf),
                                                "Your log has been blocked due to per-%s limit of %d logs per %d seconds.",
                                                should_log.decision == DECISION_TAG_LIMIT_EXCEEDED_MESSAGE ? "tag" : "PID",
                                                should_log.logs_per_period, should_log.period_s);
                                write_to_log(log_id, prio, tag, buf, &tp, get_cached_pid(), get_cached_tid());
                                return DLOG_ERROR_NOT_PERMITTED;
                        }

                        case DECISION_ALLOWED:
                                break;
                }
        }

        /* This can change due to __dynamic_config_update(), but is atomic and its
         * value implies nothing else so does not need to be under a lock. */
        if (!plog[log_id])
                return DLOG_ERROR_NOT_PERMITTED;

        return DLOG_ERROR_NONE;
}

static int __write_to_log_critical_section(log_id_t log_id, int prio, const char *tag, const char *fmt, va_list ap, bool check_should_log)
{
        if ((check_should_log || limiter_apply_to_all_buffers) && (dlog_check_limiter(log_id, prio, tag) < 0))
                return DLOG_ERROR_NONE;

        char buf[LOG_MAX_PAYLOAD_SIZE];
        int len = vsnprintf(buf, sizeof buf, fmt, ap);
        if (len < 0)
                return DLOG_ERROR_NONE;
        else if (len >= sizeof buf)
                len = sizeof buf - 1;

        // Temporary workaround, see temporary.c
        prepend_container_tag_if_in_container(sizeof buf, buf, &len);

        struct timespec tp;
        int r;
        if (deduplicate_func && !clock_gettime(CLOCK_MONOTONIC, &tp)) {
                dlog_deduplicate_e ret = deduplicate_func(buf, len, &tp);
                if (ret == DLOG_DEDUPLICATE)
                        return DLOG_ERROR_NONE;
                else if (ret == DLOG_DO_NOT_DEDUPLICATE_BUT_WARN)
                        deduplicate_warn(buf, sizeof buf, len);
                r = write_to_log(log_id, prio, tag, buf, &tp, get_cached_pid(), get_cached_tid());
        } else
                r = write_to_log(log_id, prio, tag, buf, NULL, get_cached_pid(), get_cached_tid());

        if (r < 0 && stash_failed_log)
                r = stash_failed_log(log_id, prio, tag, buf);

        return r;
}

static int __write_to_log(log_id_t log_id, int prio, const char *tag, const char *fmt, va_list ap, bool check_should_log, bool secure_log)
{
        int ret = dlog_check_validity(log_id, prio, tag);
        if (ret < 0)
                return ret;

        if (check_should_log && prio < priority_filter_level)
                return DLOG_ERROR_NONE;

        /* Threads can be cancelled before they give up a lock.
         * Therefore cancellation is temporarily disabled as
         * long as the current set of features uses a lock.
         *
         * This solution is comparatively simple and cheap.
         * The other solutions (cleanup handlers, robust mutexes)
         * would be much more complicated and also inflict larger
         * runtime costs. The downside of disabling cancellation
         * is not a problem in our case because it is temporary
         * and very brief so we don't keep an obsolete thread
         * for much longer than we otherwise would. */
        int old_cancel_state;
        if (should_disable_cancels)
                pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &old_cancel_state);

        if (!initialize())
                // TODO: We could consider stashing the failed log here
                ret = DLOG_ERROR_NOT_PERMITTED;
        else if (secure_log && !enable_secure_logs)
                ret = 0;
        else
                ret = __write_to_log_critical_section(log_id, prio, tag, fmt, ap, check_should_log);

        if (should_disable_cancels)
                pthread_setcancelstate(old_cancel_state, NULL);

        return ret;
}

int __critical_log_append_timestamp(char *buf, size_t buflen)
{
        /* NB: the timestamp may slightly differ from the one that gets
         * added onto the copy that goes into the regular buffer, and
         * timestamp acquisition is duplicated. This would ideally be
         * solved, but timestamps are currently added fairly deep in
         * backend-specific functions so for now this will have to do.
         * Also, since we're the sender, there is just this one set of
         * timestamps, i.e. the send timestamp! The usual alternative
         * set of receive timestamps will never have the opportunity
         * to get added to the entry since this log is supposed to end
         * up straight in the file (there's potentially the trusted
         * writer binary but we're trying to keep the set of actions
         * it needs to do to the minimum and those timestamps would
         * in practice be the same anyway). */

        struct timespec ts;
        clock_gettime(CLOCK_REALTIME, &ts);
        const time_t tt = ts.tv_sec;
        const long int real_millisec = ts.tv_nsec / 1000000;
        clock_gettime(CLOCK_MONOTONIC, &ts);
        struct tm tmBuf;
        struct tm *const ptm = localtime_r(&tt, &tmBuf);
        assert(ptm); // we're in a short lived fork so asserts are fine and make things simple

        int len = strftime(buf, buflen, "%m-%d %H:%M:%S", ptm);
        assert(len != 0);

        int tmp_len = snprintf(buf + len, buflen - len, ".%03ld", real_millisec);
        assert(tmp_len > 0);
        assert(tmp_len < buflen - len);
        len += tmp_len;

        tmp_len = strftime(buf + len, buflen - len, "%z ", ptm);
        assert(tmp_len != 0);
        len += tmp_len;

        tmp_len = snprintf(buf + len, buflen - len, "%5lu.%03ld", ts.tv_sec, ts.tv_nsec / 1000000);
        assert(tmp_len > 0);
        assert(tmp_len < buflen - len);
        len += tmp_len;

        return len;
}

int __critical_log_build_msg(char *buf, size_t buflen, pid_t main_pid, pid_t main_tid, log_id_t log_id, int prio, const char *tag, const char *fmt, va_list ap)
{
        int len = __critical_log_append_timestamp(buf, buflen);
        const int metadata_len = snprintf(buf + len, buflen - len, " P%5d T%5d B%-6s %c/%-8s: ",
                main_pid,
                main_tid,
                log_name_by_id(log_id),
                filter_pri_to_char(prio),
                tag ?: "CRITICAL_NO_TAG");
        assert(metadata_len > 0);
        if (metadata_len >= buflen - len)
                return buflen - 1; // can genuinely happen with an exceedingly large tag
        len += metadata_len;

        const int content_len = vsnprintf(buf + len, buflen - len, fmt, ap);
        assert(content_len >= 0); // 0 is legit with format == ""
        if (content_len >= buflen - len)
                return buflen - 1;
        len += content_len;

        return len;
}

#ifndef UNIT_TEST
__attribute__ ((noreturn))
#endif
void __critical_log_child(pid_t main_pid, pid_t main_tid, log_id_t log_id, int prio, const char *tag, const char *fmt, va_list ap)
{
        char buf[LOG_MAX_PAYLOAD_SIZE + 128]; // extra space for some metadata
        const int len = __critical_log_build_msg(buf, sizeof buf - 1, main_pid, main_tid, log_id, prio, tag, fmt, ap);
        buf[len] = '\n';
        buf[len + 1] = '\0';

        static const char *const path = "/usr/libexec/dlog-log-critical";
        execl(path, path /* argv[0] convention */, buf, (char *) NULL);

#ifndef UNIT_TEST
        /* Compilers are sometimes smart enough to recognize _exit's
         * noreturn attribute, even if we wrap it with something that
         * returns. This causes it to behave in unexpected ways, for
         * example it can blow up the program regardless or it can
         * optimize some conditionals out (and incorrectly enter them
         * after the exit call fails to actually exit). This makes it
         * unsuitable for tests. */

        _exit(1); // not the regular `exit` so as not to trigger any `atexit` handlers prematurely
#endif
}

#ifndef UNIT_TEST // contains forks and exits, these don't work well with wrapping (see above)
void __critical_log(log_id_t log_id, int prio, const char *tag, const char *fmt, va_list ap)
{
        /* Critical log functionality is mostly done in a separate binary
         * to handle security correctly (else every process would have to
         * possess the necessary privilege to write onto that file, which
         * would be opening a fairly nasty can of worms from the security
         * point of view). Our use of exec() is why a simple thread would
         * not suffice and we're resorting to a fork.
         *
         * The double fork, much like a double barreled 12 gauge shotgun,
         * is an elegant solution designed to stop a zombie army. We'd be
         * creating zombie processes if we didn't wait() for the children
         * we spawn, but we don't really want to do that since it results
         * in a needless delay. Instead, the writer process is actually a
         * grandchild, with our direct child exiting immediately just for
         * us to have something to wait on that is guaranteed not to take
         * too long. The orphaned grandchild is adopted by init, who will
         * take care to reap it when it dies. In addition to avoiding the
         * delay, the client will not have any unexpected children (which
         * could ruin logic in its own waits).
         *
         * Right after forks:
         * ┌───────┐   ┌─────────┐   ┌─────────────┐   ┌────────┐
         * │ pid 1 ├──>│ libdlog ├──>│ immediately ├──>│ execs  │
         * │ init  │   │ client  │   │    exits    │   │ writer │
         * └───────┘   └─────────┘   └─────────────┘   └────────┘
         *
         * Afterwards, libdlog has no children:
         * ┌───────┐   ┌─────────┐                     ┌────────┐
         * │ pid 1 ├──>│ libdlog │          ┌─────────>│ writer │
         * │ init  ├─┐ │ client  │          │          │ binary │
         * └───────┘ │ └─────────┘          │          └────────┘
         *           └──────────────────────┘
         */

        initialize();

        if (!enable_critical)
                return;

        const pid_t main_pid = getpid();
        const pid_t main_tid = gettid();

        const int temporary_exiter_pid = fork();
        if (temporary_exiter_pid < 0)
                return;
        if (temporary_exiter_pid != 0) {
                waitpid(temporary_exiter_pid, NULL, 0);
                return;
        }

        const int child_pid = fork();
        if (child_pid < 0)
                _exit(1);
        if (child_pid != 0)
                _exit(0);

        __critical_log_child(main_pid, main_tid, log_id, prio, tag, fmt, ap);
}

static void stash_critical_inner(log_id_t log_id, log_priority prio, const char *tag, const char *fmt, ...)
{
        va_list ap;

        va_start(ap, fmt);
        __critical_log(log_id, prio, tag, fmt, ap);
        va_end(ap);
}

static int stash_critical(log_id_t log_id, log_priority prio, const char *tag, const char *msg)
{
        stash_critical_inner(log_id, prio, tag, "FAILED TO LOG: %s", msg);
        return 0;
}

EXPORT_API int __dlog_critical_print(log_id_t log_id, int prio, const char *tag, const char *fmt, ...)
{
        va_list ap;

        va_start(ap, fmt);
        __critical_log(log_id, prio, tag, fmt, ap);
        va_end(ap);

        va_start(ap, fmt);
        int ret = __dlog_vprint(log_id, prio, tag, fmt, ap);
        va_end(ap);

        return ret;
}
#endif

EXPORT_API int dlog_set_minimum_priority(int priority)
{
        if (priority < DLOG_DEFAULT || priority > DLOG_PRIO_MAX)
                return DLOG_ERROR_INVALID_PARAMETER;

        priority_filter_level = priority;
        return DLOG_ERROR_NONE;
}

/**
 * @brief Print log
 * @details Print a log line
 * @param[in] log_id The target buffer ID
 * @param[in] prio Priority
 * @param[in] tag tag
 * @param[in] fmt Format (same as printf)
 * @param[in] ap Argument list
 * @return Bytes written, or negative error
 */
EXPORT_API int __dlog_vprint(log_id_t log_id, int prio, const char *tag, const char *fmt, va_list ap)
{
        int ret = __write_to_log(log_id, prio, tag, fmt, ap, true, false);
        __dlog_fatal_assert(prio);

        return ret;
}

/**
 * @brief Print log
 * @details Print a log line
 * @param[in] log_id The target buffer ID
 * @param[in] prio Priority
 * @param[in] tag tag
 * @param[in] fmt Format (same as printf)
 * @return Bytes written, or negative error
 */
EXPORT_API int __dlog_print(log_id_t log_id, int prio, const char *tag, const char *fmt, ...)
{
        va_list ap;

        va_start(ap, fmt);
        int ret = __dlog_vprint(log_id, prio, tag, fmt, ap);
        va_end(ap);

        return ret;
}

/**
 * @brief Print log
 * @details Print a log line
 * @param[in] log_id The target buffer ID
 * @param[in] prio Priority
 * @param[in] tag tag
 * @param[in] fmt Format (same as printf)
 * @return Bytes written, or negative error
 */
EXPORT_API int __dlog_sec_print(log_id_t log_id, int prio, const char *tag, const char *fmt, ...)
{
        if (!enable_secure_logs)
                return 0;

        va_list ap;

        va_start(ap, fmt);
        int ret = __write_to_log(log_id, prio, tag, fmt, ap, true, true);
        __dlog_fatal_assert(prio);
        va_end(ap);

        return ret;
}

EXPORT_API int dlog_vprint(log_priority prio, const char *tag, const char *fmt, va_list ap)
{
        return __write_to_log(LOG_ID_APPS, prio, tag, fmt, ap, false, false);
}

EXPORT_API int dlog_print(log_priority prio, const char *tag, const char *fmt, ...)
{
        va_list ap;

        va_start(ap, fmt);
        int ret = dlog_vprint(prio, tag, fmt, ap);
        va_end(ap);

        return ret;
}

/**
 * @brief Finalize DLog
 * @details Finalizes and deallocates the library
 * @notes Used directly in tests; brings back the pre-init state
 */
void __dlog_fini(void)
{
        if (destroy_backend) {
                destroy_backend();
                destroy_backend = NULL;
        }
        write_to_log = write_to_log_null;
        stash_failed_log = NULL;
        is_initialized = false;
        first = true;

        enable_secure_logs = true;
        enable_critical = false;
        __deduplicate_destroy();
        __log_limiter_destroy(limiter_data);
        limiter = false;
        __dynamic_config_destroy();
        should_disable_cancels = false;
}