Re: [PATCH v3] eventfd: convert global percpu eventfd_wake_count to ctx percpu eventfd_wake_count
From: Zhang, Qiang
Date: Thu Jun 10 2021 - 21:39:23 EST
Hello Jens
Excuse me, this change need to be review by you, I didn't think of a better way to avoid the problem I described, I hope I can get your opinion.
Thanks
Qiang
________________________________________
From: Zhang, Qiang <qiang.zhang@xxxxxxxxxxxxx>
Sent: Wednesday, 9 June 2021 16:15
To: axboe@xxxxxxxxx
Cc: viro@xxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-fsdevel@xxxxxxxxxxxxxxx
Subject: [PATCH v3] eventfd: convert global percpu eventfd_wake_count to ctx percpu eventfd_wake_count
From: Zqiang <qiang.zhang@xxxxxxxxxxxxx>
In RT system, the spinlock_irq be replaced by rt_mutex, when
call eventfd_signal(), if the current task is preempted after
increasing the current CPU eventfd_wake_count, when other task
run on this CPU and call eventfd_signal(), find this CPU
eventfd_wake_count is not zero, will trigger warning and direct
return, miss wakeup.
RIP: 0010:eventfd_signal+0x85/0xa0
vhost_add_used_and_signal_n+0x41/0x50 [vhost]
handle_rx+0xb9/0x9e0 [vhost_net]
handle_rx_net+0x15/0x20 [vhost_net]
vhost_worker+0x95/0xe0 [vhost]
kthread+0x19c/0x1c0
ret_from_fork+0x22/0x30
In no-RT system, even if the eventfd_signal() call is nested, if
if it's different eventfd_ctx object, it is not happen deadlock.
Fixes: b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth")
Reported-by: kernel test robot <lkp@xxxxxxxxx>
Signed-off-by: Zqiang <qiang.zhang@xxxxxxxxxxxxx>
---
v1->v2:
Modify submission information.
v2->v3:
Fix compilation error in riscv32.
fs/aio.c | 2 +-
fs/eventfd.c | 30 ++++++++++--------------------
include/linux/eventfd.h | 26 +++++++++++++++++++++-----
3 files changed, 32 insertions(+), 26 deletions(-)
diff --git a/fs/aio.c b/fs/aio.c
index 76ce0cc3ee4e..b45983d5d35a 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1695,7 +1695,7 @@ static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
list_del(&iocb->ki_list);
iocb->ki_res.res = mangle_poll(mask);
req->done = true;
- if (iocb->ki_eventfd && eventfd_signal_count()) {
+ if (iocb->ki_eventfd && eventfd_signal_count(iocb->ki_eventfd)) {
iocb = NULL;
INIT_WORK(&req->work, aio_poll_put_work);
schedule_work(&req->work);
diff --git a/fs/eventfd.c b/fs/eventfd.c
index e265b6dd4f34..b1df2c5720a7 100644
--- a/fs/eventfd.c
+++ b/fs/eventfd.c
@@ -25,26 +25,9 @@
#include <linux/idr.h>
#include <linux/uio.h>
-DEFINE_PER_CPU(int, eventfd_wake_count);
static DEFINE_IDA(eventfd_ida);
-struct eventfd_ctx {
- struct kref kref;
- wait_queue_head_t wqh;
- /*
- * Every time that a write(2) is performed on an eventfd, the
- * value of the __u64 being written is added to "count" and a
- * wakeup is performed on "wqh". A read(2) will return the "count"
- * value to userspace, and will reset "count" to zero. The kernel
- * side eventfd_signal() also, adds to the "count" counter and
- * issue a wakeup.
- */
- __u64 count;
- unsigned int flags;
- int id;
-};
-
/**
* eventfd_signal - Adds @n to the eventfd counter.
* @ctx: [in] Pointer to the eventfd context.
@@ -71,17 +54,17 @@ __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n)
* it returns true, the eventfd_signal() call should be deferred to a
* safe context.
*/
- if (WARN_ON_ONCE(this_cpu_read(eventfd_wake_count)))
+ if (WARN_ON_ONCE(this_cpu_read(*ctx->eventfd_wake_count)))
return 0;
spin_lock_irqsave(&ctx->wqh.lock, flags);
- this_cpu_inc(eventfd_wake_count);
+ this_cpu_inc(*ctx->eventfd_wake_count);
if (ULLONG_MAX - ctx->count < n)
n = ULLONG_MAX - ctx->count;
ctx->count += n;
if (waitqueue_active(&ctx->wqh))
wake_up_locked_poll(&ctx->wqh, EPOLLIN);
- this_cpu_dec(eventfd_wake_count);
+ this_cpu_dec(*ctx->eventfd_wake_count);
spin_unlock_irqrestore(&ctx->wqh.lock, flags);
return n;
@@ -92,6 +75,9 @@ static void eventfd_free_ctx(struct eventfd_ctx *ctx)
{
if (ctx->id >= 0)
ida_simple_remove(&eventfd_ida, ctx->id);
+
+ if (ctx->eventfd_wake_count)
+ free_percpu(ctx->eventfd_wake_count);
kfree(ctx);
}
@@ -421,6 +407,10 @@ static int do_eventfd(unsigned int count, int flags)
if (!ctx)
return -ENOMEM;
+ ctx->eventfd_wake_count = alloc_percpu(int);
+ if (!ctx->eventfd_wake_count)
+ goto err;
+
kref_init(&ctx->kref);
init_waitqueue_head(&ctx->wqh);
ctx->count = count;
diff --git a/include/linux/eventfd.h b/include/linux/eventfd.h
index fa0a524baed0..6311b931ac6f 100644
--- a/include/linux/eventfd.h
+++ b/include/linux/eventfd.h
@@ -14,6 +14,7 @@
#include <linux/err.h>
#include <linux/percpu-defs.h>
#include <linux/percpu.h>
+#include <linux/kref.h>
/*
* CAREFUL: Check include/uapi/asm-generic/fcntl.h when defining
@@ -29,11 +30,27 @@
#define EFD_SHARED_FCNTL_FLAGS (O_CLOEXEC | O_NONBLOCK)
#define EFD_FLAGS_SET (EFD_SHARED_FCNTL_FLAGS | EFD_SEMAPHORE)
-struct eventfd_ctx;
struct file;
#ifdef CONFIG_EVENTFD
+struct eventfd_ctx {
+ struct kref kref;
+ wait_queue_head_t wqh;
+ /*
+ * Every time that a write(2) is performed on an eventfd, the
+ * value of the __u64 being written is added to "count" and a
+ * wakeup is performed on "wqh". A read(2) will return the "count"
+ * value to userspace, and will reset "count" to zero. The kernel
+ * side eventfd_signal() also, adds to the "count" counter and
+ * issue a wakeup.
+ */
+ __u64 count;
+ unsigned int flags;
+ int id;
+ int __percpu *eventfd_wake_count;
+};
+
void eventfd_ctx_put(struct eventfd_ctx *ctx);
struct file *eventfd_fget(int fd);
struct eventfd_ctx *eventfd_ctx_fdget(int fd);
@@ -43,11 +60,10 @@ int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *w
__u64 *cnt);
void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt);
-DECLARE_PER_CPU(int, eventfd_wake_count);
-static inline bool eventfd_signal_count(void)
+static inline bool eventfd_signal_count(struct eventfd_ctx *ctx)
{
- return this_cpu_read(eventfd_wake_count);
+ return this_cpu_read(*ctx->eventfd_wake_count);
}
#else /* CONFIG_EVENTFD */
@@ -78,7 +94,7 @@ static inline int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx,
return -ENOSYS;
}
-static inline bool eventfd_signal_count(void)
+static inline bool eventfd_signal_count(struct eventfd_ctx *ctx)
{
return false;
}
--
2.17.1