On Wed, May 08, 2024 at 03:29:12PM +0300, Dan Carpenter wrote:
> Hello David Woodhouse,
>
> Commit 82980b1622d9 ("rcu: Kill rnp->ofl_seq and use only
> rcu_state.ofl_lock for exclusion") from Feb 16, 2021 (linux-next),
> leads to the following Smatch static checker warning:
>
> kernel/rcu/tree.c:1844 rcu_gp_init()
> warn: mixing irq and irqsave
There are actually cases where this does make sense, one example being
where some called function (for example, rcu_report_qs_rnp() below)
needs a flags argument.
But...
> kernel/rcu/srcutree.c:910 srcu_gp_end()
> warn: mixing irq and irqsave
>
> kernel/rcu/tree.c
> 1782 static noinline_for_stack bool rcu_gp_init(void)
> 1783 {
> 1784 unsigned long flags;
> 1785 unsigned long oldmask;
> 1786 unsigned long mask;
> 1787 struct rcu_data *rdp;
> 1788 struct rcu_node *rnp = rcu_get_root();
> 1789 bool start_new_poll;
> 1790
> 1791 WRITE_ONCE(rcu_state.gp_activity, jiffies);
> 1792 raw_spin_lock_irq_rcu_node(rnp);
> 1793 if (!rcu_state.gp_flags) {
> 1794 /* Spurious wakeup, tell caller to go back to sleep. */
> 1795 raw_spin_unlock_irq_rcu_node(rnp);
> 1796 return false;
> 1797 }
> 1798 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
> 1799
> 1800 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
> 1801 /*
> 1802 * Grace period already in progress, don't start another.
> 1803 * Not supposed to be able to happen.
> 1804 */
> 1805 raw_spin_unlock_irq_rcu_node(rnp);
> 1806 return false;
> 1807 }
> 1808
> 1809 /* Advance to a new grace period and initialize state. */
> 1810 record_gp_stall_check_time();
> 1811 /* Record GP times before starting GP, hence rcu_seq_start(). */
> 1812 rcu_seq_start(&rcu_state.gp_seq);
> 1813 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
> 1814 start_new_poll = rcu_sr_normal_gp_init();
> 1815 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
> 1816 rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap);
> 1817 raw_spin_unlock_irq_rcu_node(rnp);
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> Enables IRQs
>
> 1818
> 1819 /*
> 1820 * The "start_new_poll" is set to true, only when this GP is not able
> 1821 * to handle anything and there are outstanding users. It happens when
> 1822 * the rcu_sr_normal_gp_init() function was not able to insert a dummy
> 1823 * separator to the llist, because there were no left any dummy-nodes.
> 1824 *
> 1825 * Number of dummy-nodes is fixed, it could be that we are run out of
> 1826 * them, if so we start a new pool request to repeat a try. It is rare
> 1827 * and it means that a system is doing a slow processing of callbacks.
> 1828 */
> 1829 if (start_new_poll)
> 1830 (void) start_poll_synchronize_rcu();
> 1831
> 1832 /*
> 1833 * Apply per-leaf buffered online and offline operations to
> 1834 * the rcu_node tree. Note that this new grace period need not
> 1835 * wait for subsequent online CPUs, and that RCU hooks in the CPU
> 1836 * offlining path, when combined with checks in this function,
> 1837 * will handle CPUs that are currently going offline or that will
> 1838 * go offline later. Please also refer to "Hotplug CPU" section
> 1839 * of RCU's Requirements documentation.
> 1840 */
> 1841 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
> 1842 /* Exclude CPU hotplug operations. */
> 1843 rcu_for_each_leaf_node(rnp) {
> --> 1844 local_irq_save(flags);
> ^^^^^^^^^^^^^^^^^^^^^
> It doesn't make sense to mix irq and irqsave. Either it can be called
> with IRQs disabled or it can't...
... that is not the case here. We could use local_irq_disable().
I can queue a patch.
However...
> 1845 arch_spin_lock(&rcu_state.ofl_lock);
> 1846 raw_spin_lock_rcu_node(rnp);
> 1847 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
> 1848 !rnp->wait_blkd_tasks) {
> 1849 /* Nothing to do on this leaf rcu_node structure. */
> 1850 raw_spin_unlock_rcu_node(rnp);
> 1851 arch_spin_unlock(&rcu_state.ofl_lock);
> 1852 local_irq_restore(flags);
> 1853 continue;
> 1854 }
> 1855
> 1856 /* Record old state, apply changes to ->qsmaskinit field. */
> 1857 oldmask = rnp->qsmaskinit;
> 1858 rnp->qsmaskinit = rnp->qsmaskinitnext;
> 1859
> 1860 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
> 1861 if (!oldmask != !rnp->qsmaskinit) {
> 1862 if (!oldmask) { /* First online CPU for rcu_node. */
> 1863 if (!rnp->wait_blkd_tasks) /* Ever offline? */
> 1864 rcu_init_new_rnp(rnp);
> 1865 } else if (rcu_preempt_has_tasks(rnp)) {
> 1866 rnp->wait_blkd_tasks = true; /* blocked tasks */
> 1867 } else { /* Last offline CPU and can propagate. */
> 1868 rcu_cleanup_dead_rnp(rnp);
> 1869 }
> 1870 }
> 1871
> 1872 /*
> 1873 * If all waited-on tasks from prior grace period are
> 1874 * done, and if all this rcu_node structure's CPUs are
> 1875 * still offline, propagate up the rcu_node tree and
> 1876 * clear ->wait_blkd_tasks. Otherwise, if one of this
> 1877 * rcu_node structure's CPUs has since come back online,
> 1878 * simply clear ->wait_blkd_tasks.
> 1879 */
> 1880 if (rnp->wait_blkd_tasks &&
> 1881 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
> 1882 rnp->wait_blkd_tasks = false;
> 1883 if (!rnp->qsmaskinit)
> 1884 rcu_cleanup_dead_rnp(rnp);
> 1885 }
> 1886
> 1887 raw_spin_unlock_rcu_node(rnp);
> 1888 arch_spin_unlock(&rcu_state.ofl_lock);
> 1889 local_irq_restore(flags);
> 1890 }
> 1891 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
> 1892
> 1893 /*
> 1894 * Set the quiescent-state-needed bits in all the rcu_node
> 1895 * structures for all currently online CPUs in breadth-first
> 1896 * order, starting from the root rcu_node structure, relying on the
> 1897 * layout of the tree within the rcu_state.node[] array. Note that
> 1898 * other CPUs will access only the leaves of the hierarchy, thus
> 1899 * seeing that no grace period is in progress, at least until the
> 1900 * corresponding leaf node has been initialized.
> 1901 *
> 1902 * The grace period cannot complete until the initialization
> 1903 * process finishes, because this kthread handles both.
> 1904 */
> 1905 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
> 1906 rcu_for_each_node_breadth_first(rnp) {
> 1907 rcu_gp_slow(gp_init_delay);
> 1908 raw_spin_lock_irqsave_rcu_node(rnp, flags);
... here we do need flags because ...
> 1909 rdp = this_cpu_ptr(&rcu_data);
> 1910 rcu_preempt_check_blocked_tasks(rnp);
> 1911 rnp->qsmask = rnp->qsmaskinit;
> 1912 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
> 1913 if (rnp == rdp->mynode)
> 1914 (void)__note_gp_changes(rnp, rdp);
> 1915 rcu_preempt_boost_start_gp(rnp);
> 1916 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
> 1917 rnp->level, rnp->grplo,
> 1918 rnp->grphi, rnp->qsmask);
> 1919 /* Quiescent states for tasks on any now-offline CPUs. */
> 1920 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
> 1921 rnp->rcu_gp_init_mask = mask;
> 1922 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
> 1923 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
... rcu_report_qs_rnp() needs them.
> 1924 else
> 1925 raw_spin_unlock_irq_rcu_node(rnp);
And we do have the odd ignoring of those flags here. Which is fine.
Thanx, Paul
> 1926 cond_resched_tasks_rcu_qs();
> 1927 WRITE_ONCE(rcu_state.gp_activity, jiffies);
> 1928 }
> 1929
> 1930 // If strict, make all CPUs aware of new grace period.
> 1931 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
> 1932 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
> 1933
> 1934 return true;
> 1935 }
>
> regards,
> dan carpenter
