Merge pull request #7915 from poettering/pids-max-tweak

TODO

@@ -43,6 +43,15 @@ Features:
   sd_id128_get_machine_app_specific(). After all on long-running systems both
   IDs have similar properties.
 
+* emulate properties of the root cgroup on controllers that don't support such
+  properties natively on cpu/io/memory, the way we already do it for
+  "pids". Also, add the same logic to cgtop.
+
+* set TasksAccounting=1 on the root slice if we are running on the root cgroup,
+  and similar for the others, as soon as we emulate them properly. After all,
+  Linux keeps these system-wide stats anyway, and it costs nothing to expose
+  them.
+
 * sd-bus: add vtable flag, that may be used to request client creds implicitly
   and asynchronously before dispatching the operation
 

src/basic/meson.build

@@ -150,6 +150,8 @@ basic_sources = files('''
         proc-cmdline.h
         process-util.c
         process-util.h
+        procfs-util.c
+        procfs-util.h
         random-util.c
         random-util.h
         ratelimit.c

src/basic/process-util.h

@@ -184,3 +184,22 @@ static inline int safe_fork(const char *name, ForkFlags flags, pid_t *ret_pid) {
 }
 
 int fork_agent(const char *name, const int except[], unsigned n_except, pid_t *pid, const char *path, ...);
+
+#if SIZEOF_PID_T == 4
+/* The highest possibly (theoretic) pid_t value on this architecture. */
+#define PID_T_MAX ((pid_t) INT32_MAX)
+/* The maximum number of concurrent processes Linux allows on this architecture, as well as the highest valid PID value
+ * the kernel will potentially assign. This reflects a value compiled into the kernel (PID_MAX_LIMIT), and sets the
+ * upper boundary on what may be written to the /proc/sys/kernel/pid_max sysctl (but do note that the sysctl is off by
+ * 1, since PID 0 can never exist and there can hence only be one process less than the limit would suggest). Since
+ * these values are documented in proc(5) we feel quite confident that they are stable enough for the near future at
+ * least to define them here too. */
+#define TASKS_MAX 4194303U
+#elif SIZEOF_PID_T == 2
+#define PID_T_MAX ((pid_t) INT16_MAX)
+#define TASKS_MAX 32767U
+#else
+#error "Unknown pid_t size"
+#endif
+
+assert_cc(TASKS_MAX <= (unsigned long) PID_T_MAX)

src/basic/procfs-util.c

@@ -0,0 +1,138 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <errno.h>
+
+#include "alloc-util.h"
+#include "fileio.h"
+#include "parse-util.h"
+#include "process-util.h"
+#include "procfs-util.h"
+#include "stdio-util.h"
+#include "string-util.h"
+
+int procfs_tasks_get_limit(uint64_t *ret) {
+        _cleanup_free_ char *value = NULL;
+        uint64_t pid_max, threads_max;
+        int r;
+
+        assert(ret);
+
+        /* So there are two sysctl files that control the system limit of processes:
+         *
+         * 1. kernel.threads-max: this is probably the sysctl that makes more sense, as it directly puts a limit on
+         *    concurrent tasks.
+         *
+         * 2. kernel.pid_max: this limits the numeric range PIDs can take, and thus indirectly also limits the number
+         *    of concurrent threads. AFAICS it's primarily a compatibility concept: some crappy old code used a signed
+         *    16bit type for PIDs, hence the kernel provides a way to ensure the PIDs never go beyond INT16_MAX by
+         *    default.
+         *
+         * By default #2 is set to much lower values than #1, hence the limit people come into contact with first, as
+         * it's the lowest boundary they need to bump when they want higher number of processes.
+         *
+         * Also note the weird definition of #2: PIDs assigned will be kept below this value, which means the number of
+         * tasks that can be created is one lower, as PID 0 is not a valid process ID. */
+
+        r = read_one_line_file("/proc/sys/kernel/pid_max", &value);
+        if (r < 0)
+                return r;
+
+        r = safe_atou64(value, &pid_max);
+        if (r < 0)
+                return r;
+
+        value = mfree(value);
+        r = read_one_line_file("/proc/sys/kernel/threads-max", &value);
+        if (r < 0)
+                return r;
+
+        r = safe_atou64(value, &threads_max);
+        if (r < 0)
+                return r;
+
+        /* Subtract one from pid_max, since PID 0 is not a valid PID */
+        *ret = MIN(pid_max-1, threads_max);
+        return 0;
+}
+
+int procfs_tasks_set_limit(uint64_t limit) {
+        char buffer[DECIMAL_STR_MAX(uint64_t)+1];
+        _cleanup_free_ char *value = NULL;
+        uint64_t pid_max;
+        int r;
+
+        if (limit == 0) /* This makes no sense, we are userspace and hence count as tasks too, and we want to live,
+                         * hence the limit conceptually has to be above 0. Also, most likely if anyone asks for a zero
+                         * limit he/she probably means "no limit", hence let's better refuse this to avoid
+                         * confusion. */
+                return -EINVAL;
+
+        /* The Linux kernel doesn't allow this value to go below 20, hence don't allow this either, higher values than
+         * TASKS_MAX are not accepted by the pid_max sysctl. We'll treat anything this high as "unbounded" and hence
+         * set it to the maximum. */
+        limit = CLAMP(limit, 20U, TASKS_MAX);
+
+        r = read_one_line_file("/proc/sys/kernel/pid_max", &value);
+        if (r < 0)
+                return r;
+        r = safe_atou64(value, &pid_max);
+        if (r < 0)
+                return r;
+
+        /* As pid_max is about the numeric pid_t range we'll bump it if necessary, but only ever increase it, never
+         * decrease it, as threads-max is the much more relevant sysctl. */
+        if (limit > pid_max-1) {
+                sprintf(buffer, "%" PRIu64, limit+1); /* Add one, since PID 0 is not a valid PID */
+                r = write_string_file("/proc/sys/kernel/pid_max", buffer, WRITE_STRING_FILE_DISABLE_BUFFER);
+                if (r < 0)
+                        return r;
+        }
+
+        sprintf(buffer, "%" PRIu64, limit);
+        r = write_string_file("/proc/sys/kernel/threads-max", buffer, WRITE_STRING_FILE_DISABLE_BUFFER);
+        if (r < 0) {
+                uint64_t threads_max;
+
+                /* Hmm, we couldn't write this? If so, maybe it was already set properly? In that case let's not
+                 * generate an error */
+
+                value = mfree(value);
+                if (read_one_line_file("/proc/sys/kernel/threads-max", &value) < 0)
+                        return r; /* return original error */
+
+                if (safe_atou64(value, &threads_max) < 0)
+                        return r; /* return original error */
+
+                if (MIN(pid_max-1, threads_max) != limit)
+                        return r; /* return original error */
+
+                /* Yay! Value set already matches what we were trying to set, hence consider this a success. */
+        }
+
+        return 0;
+}
+
+int procfs_tasks_get_current(uint64_t *ret) {
+        _cleanup_free_ char *value = NULL;
+        const char *p, *nr;
+        size_t n;
+        int r;
+
+        assert(ret);
+
+        r = read_one_line_file("/proc/loadavg", &value);
+        if (r < 0)
+                return r;
+
+        /* Look for the second part of the fourth field, which is separated by a slash from the first part. None of the
+         * earlier fields use a slash, hence let's use this to find the right spot. */
+        p = strchr(value, '/');
+        if (!p)
+                return -EINVAL;
+
+        p++;
+        n = strspn(p, DIGITS);
+        nr = strndupa(p, n);
+
+        return safe_atou64(nr, ret);
+}

src/basic/procfs-util.h

@@ -0,0 +1,8 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+#pragma once
+
+#include <inttypes.h>
+
+int procfs_tasks_get_limit(uint64_t *ret);
+int procfs_tasks_set_limit(uint64_t limit);
+int procfs_tasks_get_current(uint64_t *ret);

src/basic/util.c

@@ -52,6 +52,7 @@
 #include "parse-util.h"
 #include "path-util.h"
 #include "process-util.h"
+#include "procfs-util.h"
 #include "set.h"
 #include "signal-util.h"
 #include "stat-util.h"
@@ -473,31 +474,22 @@ uint64_t physical_memory_scale(uint64_t v, uint64_t max) {
 
 uint64_t system_tasks_max(void) {
 
-#if SIZEOF_PID_T == 4
-#define TASKS_MAX ((uint64_t) (INT32_MAX-1))
-#elif SIZEOF_PID_T == 2
-#define TASKS_MAX ((uint64_t) (INT16_MAX-1))
-#else
-#error "Unknown pid_t size"
-#endif
-
-        _cleanup_free_ char *value = NULL, *root = NULL;
         uint64_t a = TASKS_MAX, b = TASKS_MAX;
+        _cleanup_free_ char *root = NULL;
 
         /* Determine the maximum number of tasks that may run on this system. We check three sources to determine this
          * limit:
          *
-         * a) the maximum value for the pid_t type
+         * a) the maximum tasks value the kernel allows on this architecture
          * b) the cgroups pids_max attribute for the system
-         * c) the kernel's configure maximum PID value
+         * c) the kernel's configured maximum PID value
          *
          * And then pick the smallest of the three */
 
-        if (read_one_line_file("/proc/sys/kernel/pid_max", &value) >= 0)
-                (void) safe_atou64(value, &a);
+        (void) procfs_tasks_get_limit(&a);
 
         if (cg_get_root_path(&root) >= 0) {
-                value = mfree(value);
+                _cleanup_free_ char *value = NULL;
 
                 if (cg_get_attribute("pids", root, "pids.max", &value) >= 0)
                         (void) safe_atou64(value, &b);

src/cgtop/cgtop.c

@@ -40,7 +40,9 @@
 #include "parse-util.h"
 #include "path-util.h"
 #include "process-util.h"
+#include "procfs-util.h"
 #include "stdio-util.h"
+#include "strv.h"
 #include "terminal-util.h"
 #include "unit-name.h"
 #include "util.h"
@@ -193,26 +195,33 @@ static int process(
                         g->n_tasks_valid = true;
 
         } else if (streq(controller, "pids") && arg_count == COUNT_PIDS) {
-                _cleanup_free_ char *p = NULL, *v = NULL;
 
-                r = cg_get_path(controller, path, "pids.current", &p);
-                if (r < 0)
-                        return r;
+                if (isempty(path) || path_equal(path, "/")) {
+                        r = procfs_tasks_get_current(&g->n_tasks);
+                        if (r < 0)
+                                return r;
+                } else {
+                        _cleanup_free_ char *p = NULL, *v = NULL;
 
-                r = read_one_line_file(p, &v);
-                if (r == -ENOENT)
-                        return 0;
-                if (r < 0)
-                        return r;
+                        r = cg_get_path(controller, path, "pids.current", &p);
+                        if (r < 0)
+                                return r;
 
-                r = safe_atou64(v, &g->n_tasks);
-                if (r < 0)
-                        return r;
+                        r = read_one_line_file(p, &v);
+                        if (r == -ENOENT)
+                                return 0;
+                        if (r < 0)
+                                return r;
+
+                        r = safe_atou64(v, &g->n_tasks);
+                        if (r < 0)
+                                return r;
+                }
 
                 if (g->n_tasks > 0)
                         g->n_tasks_valid = true;
 
-        } else if (streq(controller, "cpu") || streq(controller, "cpuacct")) {
+        } else if (STR_IN_SET(controller, "cpu", "cpuacct")) {
                 _cleanup_free_ char *p = NULL, *v = NULL;
                 uint64_t new_usage;
                 nsec_t timestamp;

src/core/cgroup.c

@@ -32,6 +32,7 @@
 #include "parse-util.h"
 #include "path-util.h"
 #include "process-util.h"
+#include "procfs-util.h"
 #include "special.h"
 #include "stdio-util.h"
 #include "string-table.h"
@@ -39,6 +40,18 @@
 
 #define CGROUP_CPU_QUOTA_PERIOD_USEC ((usec_t) 100 * USEC_PER_MSEC)
 
+bool unit_has_root_cgroup(Unit *u) {
+        assert(u);
+
+        /* Returns whether this unit manages the root cgroup. Note that this is different from being named "-.slice",
+         * as inside of containers the root slice won't be identical to the root cgroup. */
+
+        if (!u->cgroup_path)
+                return false;
+
+        return isempty(u->cgroup_path) || path_equal(u->cgroup_path, "/");
+}
+
 static void cgroup_compat_warn(void) {
         static bool cgroup_compat_warned = false;
 
@@ -708,21 +721,17 @@ static void cgroup_context_apply(
 
         assert(u);
 
-        c = unit_get_cgroup_context(u);
-        path = u->cgroup_path;
-
-        assert(c);
-        assert(path);
-
         /* Nothing to do? Exit early! */
         if (apply_mask == 0 && !apply_bpf)
                 return;
 
-        /* Some cgroup attributes are not supported on the root cgroup,
-         * hence silently ignore */
-        is_root = isempty(path) || path_equal(path, "/");
-        if (is_root)
-                /* Make sure we don't try to display messages with an empty path. */
+        /* Some cgroup attributes are not supported on the root cgroup, hence silently ignore */
+        is_root = unit_has_root_cgroup(u);
+
+        assert_se(c = unit_get_cgroup_context(u));
+        assert_se(path = u->cgroup_path);
+
+        if (is_root) /* Make sure we don't try to display messages with an empty path. */
                 path = "/";
 
         /* We generally ignore errors caused by read-only mounted
@@ -1019,19 +1028,46 @@ static void cgroup_context_apply(
                 }
         }
 
-        if ((apply_mask & CGROUP_MASK_PIDS) && !is_root) {
+        if (apply_mask & CGROUP_MASK_PIDS) {
 
-                if (c->tasks_max != CGROUP_LIMIT_MAX) {
-                        char buf[DECIMAL_STR_MAX(uint64_t) + 2];
+                if (is_root) {
+                        /* So, the "pids" controller does not expose anything on the root cgroup, in order not to
+                         * replicate knobs exposed elsewhere needlessly. We abstract this away here however, and when
+                         * the knobs of the root cgroup are modified propagate this to the relevant sysctls. There's a
+                         * non-obvious asymmetry however: unlike the cgroup properties we don't really want to take
+                         * exclusive ownership of the sysctls, but we still want to honour things if the user sets
+                         * limits. Hence we employ sort of a one-way strategy: when the user sets a bounded limit
+                         * through us it counts. When the user afterwards unsets it again (i.e. sets it to unbounded)
+                         * it also counts. But if the user never set a limit through us (i.e. we are the default of
+                         * "unbounded") we leave things unmodified. For this we manage a global boolean that we turn on
+                         * the first time we set a limit. Note that this boolean is flushed out on manager reload,
+                         * which is desirable so that there's an offical way to release control of the sysctl from
+                         * systemd: set the limit to unbounded and reload. */
 
-                        sprintf(buf, "%" PRIu64 "\n", c->tasks_max);
-                        r = cg_set_attribute("pids", path, "pids.max", buf);
-                } else
-                        r = cg_set_attribute("pids", path, "pids.max", "max");
+                        if (c->tasks_max != CGROUP_LIMIT_MAX) {
+                                u->manager->sysctl_pid_max_changed = true;
+                                r = procfs_tasks_set_limit(c->tasks_max);
+                        } else if (u->manager->sysctl_pid_max_changed)
+                                r = procfs_tasks_set_limit(TASKS_MAX);
+                        else
+                                r = 0;
 
-                if (r < 0)
-                        log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
-                                      "Failed to set pids.max: %m");
+                        if (r < 0)
+                                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                              "Failed to write to tasks limit sysctls: %m");
+
+                } else {
+                        if (c->tasks_max != CGROUP_LIMIT_MAX) {
+                                char buf[DECIMAL_STR_MAX(uint64_t) + 2];
+
+                                sprintf(buf, "%" PRIu64 "\n", c->tasks_max);
+                                r = cg_set_attribute("pids", path, "pids.max", buf);
+                        } else
+                                r = cg_set_attribute("pids", path, "pids.max", "max");
+                        if (r < 0)
+                                log_unit_full(u, IN_SET(r, -ENOENT, -EROFS, -EACCES) ? LOG_DEBUG : LOG_WARNING, r,
+                                              "Failed to set pids.max: %m");
+                }
         }
 
         if (apply_bpf)
@@ -1062,7 +1098,7 @@ CGroupMask cgroup_context_get_mask(CGroupContext *c) {
                 mask |= CGROUP_MASK_DEVICES;
 
         if (c->tasks_accounting ||
-            c->tasks_max != (uint64_t) -1)
+            c->tasks_max != CGROUP_LIMIT_MAX)
                 mask |= CGROUP_MASK_PIDS;
 
         return mask;
@@ -2294,6 +2330,10 @@ int unit_get_tasks_current(Unit *u, uint64_t *ret) {
         if ((u->cgroup_realized_mask & CGROUP_MASK_PIDS) == 0)
                 return -ENODATA;
 
+        /* The root cgroup doesn't expose this information, let's get it from /proc instead */
+        if (unit_has_root_cgroup(u))
+                return procfs_tasks_get_current(ret);
+
         r = cg_get_attribute("pids", u->cgroup_path, "pids.current", &v);
         if (r == -ENOENT)
                 return -ENODATA;

src/core/cgroup.h

@@ -208,6 +208,8 @@ int unit_reset_ip_accounting(Unit *u);
         cc ? cc->name : false;                          \
         })
 
+bool unit_has_root_cgroup(Unit *u);
+
 int manager_notify_cgroup_empty(Manager *m, const char *group);
 
 void unit_invalidate_cgroup(Unit *u, CGroupMask m);

src/core/manager.h

@@ -269,6 +269,9 @@ struct Manager {
         /* Have we already printed the taint line if necessary? */
         bool taint_logged:1;
 
+        /* Have we ever changed the "kernel.pid_max" sysctl? */
+        bool sysctl_pid_max_changed:1;
+
         unsigned test_run_flags:8;
 
         /* If non-zero, exit with the following value when the systemd

src/test/meson.build

@@ -411,6 +411,10 @@ tests += [
          [],
          []],
 
+        [['src/test/test-procfs-util.c'],
+         [],
+         []],
+
         [['src/test/test-unaligned.c'],
          [],
          []],

src/test/test-procfs-util.c

@@ -0,0 +1,38 @@
+/* SPDX-License-Identifier: LGPL-2.1+ */
+
+#include <errno.h>
+
+#include "log.h"
+#include "procfs-util.h"
+
+int main(int argc, char *argv[]) {
+        uint64_t v;
+        int r;
+
+        log_parse_environment();
+        log_open();
+
+        assert_se(procfs_tasks_get_current(&v) >= 0);
+        log_info("Current number of tasks: %" PRIu64, v);
+
+        assert_se(procfs_tasks_get_limit(&v) >= 0);
+        log_info("Limit of tasks: %" PRIu64, v);
+        assert_se(v > 0);
+        assert_se(procfs_tasks_set_limit(v) >= 0);
+
+        if (v > 100) {
+                uint64_t w;
+                r = procfs_tasks_set_limit(v-1);
+                assert_se(IN_SET(r, 0, -EPERM, -EACCES, -EROFS));
+
+                assert_se(procfs_tasks_get_limit(&w) >= 0);
+                assert_se((r == 0 && w == v - 1) || (r < 0 && w == v));
+
+                assert_se(procfs_tasks_set_limit(v) >= 0);
+
+                assert_se(procfs_tasks_get_limit(&w) >= 0);
+                assert_se(v == w);
+        }
+
+        return 0;
+}