All of lore.kernel.org
 help / color / mirror / Atom feed
From: Andy Lutomirski <luto@kernel.org>
To: x86@kernel.org, Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: LKML <linux-kernel@vger.kernel.org>,
	Nicholas Piggin <npiggin@gmail.com>,
	Arnd Bergmann <arnd@arndb.de>, Anton Blanchard <anton@ozlabs.org>,
	Andy Lutomirski <luto@kernel.org>,
	stable@vger.kernel.org
Subject: [PATCH v3 4/4] membarrier: Execute SYNC_CORE on the calling thread
Date: Thu,  3 Dec 2020 21:07:06 -0800	[thread overview]
Message-ID: <250ded637696d490c69bef1877148db86066881c.1607058304.git.luto@kernel.org> (raw)
In-Reply-To: <cover.1607058304.git.luto@kernel.org>

membarrier()'s MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE is documented
as syncing the core on all sibling threads but not necessarily the
calling thread.  This behavior is fundamentally buggy and cannot be used
safely.  Suppose a user program has two threads.  Thread A is on CPU 0
and thread B is on CPU 1.  Thread A modifies some text and calls
membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE).  Then thread B
executes the modified code.  If, at any point after membarrier() decides
which CPUs to target, thread A could be preempted and replaced by thread
B on CPU 0.  This could even happen on exit from the membarrier()
syscall.  If this happens, thread B will end up running on CPU 0 without
having synced.

In principle, this could be fixed by arranging for the scheduler to
sync_core_before_usermode() whenever switching between two threads in
the same mm if there is any possibility of a concurrent membarrier()
call, but this would have considerable overhead.  Instead, make
membarrier() sync the calling CPU as well.

As an optimization, this avoids an extra smp_mb() in the default
barrier-only mode.

Cc: stable@vger.kernel.org
Signed-off-by: Andy Lutomirski <luto@kernel.org>
---
 kernel/sched/membarrier.c | 51 +++++++++++++++++++++++++--------------
 1 file changed, 33 insertions(+), 18 deletions(-)

diff --git a/kernel/sched/membarrier.c b/kernel/sched/membarrier.c
index 01538b31f27e..57266ab32ef9 100644
--- a/kernel/sched/membarrier.c
+++ b/kernel/sched/membarrier.c
@@ -333,7 +333,8 @@ static int membarrier_private_expedited(int flags, int cpu_id)
 			return -EPERM;
 	}
 
-	if (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1)
+	if (flags != MEMBARRIER_FLAG_SYNC_CORE &&
+	    (atomic_read(&mm->mm_users) == 1 || num_online_cpus() == 1))
 		return 0;
 
 	/*
@@ -352,8 +353,6 @@ static int membarrier_private_expedited(int flags, int cpu_id)
 
 		if (cpu_id >= nr_cpu_ids || !cpu_online(cpu_id))
 			goto out;
-		if (cpu_id == raw_smp_processor_id())
-			goto out;
 		rcu_read_lock();
 		p = rcu_dereference(cpu_rq(cpu_id)->curr);
 		if (!p || p->mm != mm) {
@@ -368,16 +367,6 @@ static int membarrier_private_expedited(int flags, int cpu_id)
 		for_each_online_cpu(cpu) {
 			struct task_struct *p;
 
-			/*
-			 * Skipping the current CPU is OK even through we can be
-			 * migrated at any point. The current CPU, at the point
-			 * where we read raw_smp_processor_id(), is ensured to
-			 * be in program order with respect to the caller
-			 * thread. Therefore, we can skip this CPU from the
-			 * iteration.
-			 */
-			if (cpu == raw_smp_processor_id())
-				continue;
 			p = rcu_dereference(cpu_rq(cpu)->curr);
 			if (p && p->mm == mm)
 				__cpumask_set_cpu(cpu, tmpmask);
@@ -385,12 +374,38 @@ static int membarrier_private_expedited(int flags, int cpu_id)
 		rcu_read_unlock();
 	}
 
-	preempt_disable();
-	if (cpu_id >= 0)
+	if (cpu_id >= 0) {
+		/*
+		 * smp_call_function_single() will call ipi_func() if cpu_id
+		 * is the calling CPU.
+		 */
 		smp_call_function_single(cpu_id, ipi_func, NULL, 1);
-	else
-		smp_call_function_many(tmpmask, ipi_func, NULL, 1);
-	preempt_enable();
+	} else {
+		/*
+		 * For regular membarrier, we can save a few cycles by
+		 * skipping the current cpu -- we're about to do smp_mb()
+		 * below, and if we migrate to a different cpu, this cpu
+		 * and the new cpu will execute a full barrier in the
+		 * scheduler.
+		 *
+		 * For CORE_SYNC, we do need a barrier on the current cpu --
+		 * otherwise, if we are migrated and replaced by a different
+		 * task in the same mm just before, during, or after
+		 * membarrier, we will end up with some thread in the mm
+		 * running without a core sync.
+		 *
+		 * For RSEQ, don't rseq_preempt() the caller.  User code
+		 * is not supposed to issue syscalls at all from inside an
+		 * rseq critical section.
+		 */
+		if (flags != MEMBARRIER_FLAG_SYNC_CORE) {
+			preempt_disable();
+			smp_call_function_many(tmpmask, ipi_func, NULL, true);
+			preempt_enable();
+		} else {
+			on_each_cpu_mask(tmpmask, ipi_func, NULL, true);
+		}
+	}
 
 out:
 	if (cpu_id < 0)
-- 
2.28.0


  parent reply	other threads:[~2020-12-04  5:08 UTC|newest]

Thread overview: 15+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-12-04  5:07 [PATCH v3 0/4] membarrier fixes Andy Lutomirski
2020-12-04  5:07 ` [PATCH v3 1/4] x86/membarrier: Get rid of a dubious optimization Andy Lutomirski
2020-12-09  4:10   ` Andy Lutomirski
2020-12-09  8:42   ` [tip: x86/urgent] " tip-bot2 for Andy Lutomirski
2020-12-04  5:07 ` [PATCH v3 2/4] membarrier: Add an actual barrier before rseq_preempt() Andy Lutomirski
2020-12-09  4:12   ` Andy Lutomirski
2020-12-09  8:42   ` [tip: x86/urgent] " tip-bot2 for Andy Lutomirski
2020-12-14 18:05     ` Andy Lutomirski
2020-12-04  5:07 ` [PATCH v3 3/4] membarrier: Explicitly sync remote cores when SYNC_CORE is requested Andy Lutomirski
2020-12-09  4:13   ` Andy Lutomirski
2020-12-09  8:42   ` [tip: x86/urgent] " tip-bot2 for Andy Lutomirski
2020-12-04  5:07 ` Andy Lutomirski [this message]
2020-12-04 19:35   ` [PATCH v3 4/4] membarrier: Execute SYNC_CORE on the calling thread Mathieu Desnoyers
2020-12-09  4:15   ` Andy Lutomirski
2020-12-09  8:42   ` [tip: x86/urgent] " tip-bot2 for Andy Lutomirski

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=250ded637696d490c69bef1877148db86066881c.1607058304.git.luto@kernel.org \
    --to=luto@kernel.org \
    --cc=anton@ozlabs.org \
    --cc=arnd@arndb.de \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mathieu.desnoyers@efficios.com \
    --cc=npiggin@gmail.com \
    --cc=stable@vger.kernel.org \
    --cc=x86@kernel.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.