Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
> On Jan 4, 2019, at 2:23 AM, Tetsuo Handa > wrote: > > But why not to clarify what are appropriate sanity checks? > ... > want a cleanup for scripts/checkpatch.pl . These are good goals. I avoid purely cosmetic patches. I do not object to cosmetic patches from others that do not change behavior. The checks that concern you deal with changing tty line disciplines. Dealing with line discipline changes has been an ongoing issue since n_hdlc was derived from other line disciplines 20 years ago, with major overhauls along the way. It is complex: driver layers shifting during operation while dealing properly with opens, closes, hangups, and sleeping operations. Patches have been added to the latest unreleased kernel to address line discipline changes, it is still evolving. Why are the existing line discipline checks in n_hdlc where they are? Becasue that’s how they evolved from where they started to accomodate these changes. There are not many and their function is known: has the line discipline changed at that point? I know that is not satisfying but coming up with a definitive comment saying a check is absolutely required in one place and not in another requires more insight into the long history of a moving target than I have. Without that insight I would not alter existing checks in code that is not causing problems.
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
On 2019/01/04 0:57, Paul Fulghum wrote: >> On Jan 3, 2019, at 3:32 AM, Tetsuo Handa >> wrote: >> >> On 2019/01/03 18:09, Jiri Slaby wrote: >>> On 02. 01. 19, 16:04, Tetsuo Handa wrote: + if (wait_event_interruptible(tty->read_wait, + (ret = -EIO, test_bit(TTY_OTHER_CLOSED, >flags)) || + (ret = 0, tty_hung_up_p(file)) || + (rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list)) != NULL || + (ret = -EAGAIN, tty_io_nonblock(tty, file + return -EINTR; >>> >>> Oh, that looks really ugly. Could you move all this to a function >>> returning a bool and taking and as parameters? >>> >> >> OK. Something like this? > > > I agree with Jiri that placing all the conditional logic in a single > expression is difficult to read. > > But exchanging that many locals with a separate function defeats the original > purpose of > simplifying code and this implementation changes the logic (write no > longer checks for line discipline changing under it during wait). Not only defeating the original purpose but also increasing object size. > > Converting to wait_event_interruptible where possible is a good goal but this > instance > may be better left alone. The current structure mirrors the existing n_tty > line discipline. But why not to clarify what are appropriate sanity checks? Currently, read side does not check "n_hdlc->magic != HDLC_MAGIC" case while write side does. If it is intended for catching memory corruption etc. then both sides should do the same thing. Currently, write side does not check "tty != n_hdlc->tty" case before schedule() while does after schedule(). If "tty != n_hdlc->tty" case can ever happen, what prevents n_hdlc_tty_write() from being called again after n_hdlc_tty_write() once returned with -EIO due to "tty != n_hdlc->tty" case? If it is intended for catching memory corruption etc. then both sides should check "tty != n_hdlc->tty" case. Current logic is unclear, in addition to want a cleanup for scripts/checkpatch.pl . total: 158 errors, 95 warnings, 994 lines checked
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
> On Jan 3, 2019, at 3:32 AM, Tetsuo Handa > wrote: > > On 2019/01/03 18:09, Jiri Slaby wrote: >> On 02. 01. 19, 16:04, Tetsuo Handa wrote: >>> + if (wait_event_interruptible(tty->read_wait, >>> +(ret = -EIO, test_bit(TTY_OTHER_CLOSED, >flags)) || >>> +(ret = 0, tty_hung_up_p(file)) || >>> +(rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list)) != NULL || >>> +(ret = -EAGAIN, tty_io_nonblock(tty, file >>> + return -EINTR; >> >> Oh, that looks really ugly. Could you move all this to a function >> returning a bool and taking and as parameters? >> > > OK. Something like this? I agree with Jiri that placing all the conditional logic in a single expression is difficult to read. But exchanging that many locals with a separate function defeats the original purpose of simplifying code and this implementation changes the logic (write no longer checks for line discipline changing under it during wait). Converting to wait_event_interruptible where possible is a good goal but this instance may be better left alone. The current structure mirrors the existing n_tty line discipline.
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
On 2019/01/03 18:09, Jiri Slaby wrote: > On 02. 01. 19, 16:04, Tetsuo Handa wrote: >> +if (wait_event_interruptible(tty->read_wait, >> + (ret = -EIO, test_bit(TTY_OTHER_CLOSED, >flags)) || >> + (ret = 0, tty_hung_up_p(file)) || >> + (rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list)) != NULL || >> + (ret = -EAGAIN, tty_io_nonblock(tty, file >> +return -EINTR; > > Oh, that looks really ugly. Could you move all this to a function > returning a bool and taking and as parameters? > OK. Something like this? >From 725c55be437b6ce3b578a045cc7ddeeb2bbeb4b3 Mon Sep 17 00:00:00 2001 From: Tetsuo Handa Date: Thu, 3 Jan 2019 20:29:49 +0900 Subject: [PATCH] tty/n_hdlc: Use wait_event_interruptible() and same sanity checks. We can use wait_event_interruptible() in order to make it easier to understand. Also, since the reason of using different level/frequency of sanity checks for read and write is unclear while nowadays we have rich fuzzing/sanitizing tools, let's use same sanity checks for read and write. Signed-off-by: Tetsuo Handa --- drivers/tty/n_hdlc.c | 154 --- 1 file changed, 61 insertions(+), 93 deletions(-) diff --git a/drivers/tty/n_hdlc.c b/drivers/tty/n_hdlc.c index 8223d02..c497ef1 100644 --- a/drivers/tty/n_hdlc.c +++ b/drivers/tty/n_hdlc.c @@ -548,6 +548,27 @@ static void n_hdlc_tty_receive(struct tty_struct *tty, const __u8 *data, } /* end of n_hdlc_tty_receive() */ +static bool __n_hdlc_tty_read(struct tty_struct *tty, struct file *file, + struct n_hdlc *n_hdlc, int *ret, + struct n_hdlc_buf **rbuf) +{ + if (test_bit(TTY_OTHER_CLOSED, >flags)) { + *ret = -EIO; + return true; + } + if (tty_hung_up_p(file)) + return true; + *rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list); + if (*rbuf) + return true; + /* no data */ + if (tty_io_nonblock(tty, file)) { + *ret = -EAGAIN; + return true; + } + return false; +} + /** * n_hdlc_tty_read - Called to retrieve one frame of data (if available) * @tty - pointer to tty instance data @@ -562,14 +583,13 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, { struct n_hdlc *n_hdlc = tty2n_hdlc(tty); int ret = 0; - struct n_hdlc_buf *rbuf; - DECLARE_WAITQUEUE(wait, current); + struct n_hdlc_buf *rbuf = NULL; if (debuglevel >= DEBUG_LEVEL_INFO) printk("%s(%d)n_hdlc_tty_read() called\n",__FILE__,__LINE__); /* Validate the pointers */ - if (!n_hdlc) + if (!n_hdlc || n_hdlc->magic != HDLC_MAGIC) return -EIO; /* verify user access to buffer */ @@ -579,60 +599,41 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, return -EFAULT; } - add_wait_queue(>read_wait, ); - - for (;;) { - if (test_bit(TTY_OTHER_CLOSED, >flags)) { - ret = -EIO; - break; - } - if (tty_hung_up_p(file)) - break; - - set_current_state(TASK_INTERRUPTIBLE); - - rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list); - if (rbuf) { - if (rbuf->count > nr) { - /* too large for caller's buffer */ - ret = -EOVERFLOW; - } else { - __set_current_state(TASK_RUNNING); - if (copy_to_user(buf, rbuf->buf, rbuf->count)) - ret = -EFAULT; - else - ret = rbuf->count; - } - - if (n_hdlc->rx_free_buf_list.count > - DEFAULT_RX_BUF_COUNT) - kfree(rbuf); - else - n_hdlc_buf_put(_hdlc->rx_free_buf_list, rbuf); - break; - } - - /* no data */ - if (tty_io_nonblock(tty, file)) { - ret = -EAGAIN; - break; - } - - schedule(); - - if (signal_pending(current)) { - ret = -EINTR; - break; - } + if (wait_event_interruptible(tty->read_wait, +__n_hdlc_tty_read(tty, file, n_hdlc, , + ))) + return -EINTR; + if (rbuf) { + if (rbuf->count > nr) + /* too large for caller's buffer */ +
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
On 02. 01. 19, 16:04, Tetsuo Handa wrote: > + if (wait_event_interruptible(tty->read_wait, > + (ret = -EIO, test_bit(TTY_OTHER_CLOSED, >flags)) || > + (ret = 0, tty_hung_up_p(file)) || > + (rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list)) != NULL || > + (ret = -EAGAIN, tty_io_nonblock(tty, file > + return -EINTR; Oh, that looks really ugly. Could you move all this to a function returning a bool and taking and as parameters? > + if (rbuf) { > + if (rbuf->count > nr) > + /* too large for caller's buffer */ > + ret = -EOVERFLOW; > + else if (copy_to_user(buf, rbuf->buf, rbuf->count)) > + ret = -EFAULT; > + else > + ret = rbuf->count; > + if (n_hdlc->rx_free_buf_list.count > DEFAULT_RX_BUF_COUNT) > + kfree(rbuf); > + else > + n_hdlc_buf_put(_hdlc->rx_free_buf_list, rbuf); > } > - > - remove_wait_queue(>read_wait, ); > - __set_current_state(TASK_RUNNING); > - > return ret; > - > }/* end of n_hdlc_tty_read() */ > > /** > @@ -645,21 +612,13 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, > struct file *file, > static ssize_t n_hdlc_tty_write(struct tty_struct *tty, struct file *file, > const unsigned char *data, size_t count) > { > - struct n_hdlc *n_hdlc = tty2n_hdlc (tty); > + struct n_hdlc *n_hdlc; > int error = 0; > - DECLARE_WAITQUEUE(wait, current); > - struct n_hdlc_buf *tbuf; > + struct n_hdlc_buf *tbuf = NULL; > > if (debuglevel >= DEBUG_LEVEL_INFO) > printk("%s(%d)n_hdlc_tty_write() called count=%zd\n", > __FILE__,__LINE__,count); > - > - /* Verify pointers */ > - if (!n_hdlc) > - return -EIO; > - > - if (n_hdlc->magic != HDLC_MAGIC) > - return -EIO; > > /* verify frame size */ > if (count > maxframe ) { > @@ -670,40 +629,14 @@ static ssize_t n_hdlc_tty_write(struct tty_struct *tty, > struct file *file, > maxframe ); > count = maxframe; > } > - > - add_wait_queue(>write_wait, ); > > - for (;;) { > - set_current_state(TASK_INTERRUPTIBLE); > - > - tbuf = n_hdlc_buf_get(_hdlc->tx_free_buf_list); > - if (tbuf) > - break; > - > - if (tty_io_nonblock(tty, file)) { > - error = -EAGAIN; > - break; > - } > - schedule(); > - > - n_hdlc = tty2n_hdlc (tty); > - if (!n_hdlc || n_hdlc->magic != HDLC_MAGIC || > - tty != n_hdlc->tty) { > - printk("n_hdlc_tty_write: %p invalid after wait!\n", > n_hdlc); > - error = -EIO; > - break; > - } > - > - if (signal_pending(current)) { > - error = -EINTR; > - break; > - } > - } > - > - __set_current_state(TASK_RUNNING); > - remove_wait_queue(>write_wait, ); > - > - if (!error) { > + if (wait_event_interruptible(tty->write_wait, > + (error = -EIO, n_hdlc = tty2n_hdlc(tty), /* Verify pointers */ > + !n_hdlc || n_hdlc->magic != HDLC_MAGIC || tty != n_hdlc->tty) || > + (tbuf = n_hdlc_buf_get(_hdlc->tx_free_buf_list)) != NULL || > + (error = -EAGAIN, tty_io_nonblock(tty, file > + return -EINTR; This is even worse. So detto as above? > + if (tbuf) { > /* Retrieve the user's buffer */ > memcpy(tbuf->buf, data, count); > > @@ -711,8 +644,9 @@ static ssize_t n_hdlc_tty_write(struct tty_struct *tty, > struct file *file, > tbuf->count = error = count; > n_hdlc_buf_put(_hdlc->tx_buf_list,tbuf); > n_hdlc_send_frames(n_hdlc,tty); > + } else if (error == -EIO) { > + printk("n_hdlc_tty_write: %p invalid!\n", n_hdlc); > } > - > return error; > > }/* end of n_hdlc_tty_write() */ > thanks, -- js suse labs
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
> On Jan 2, 2019, at 7:04 AM, Tetsuo Handa > wrote: > > On 2019/01/01 12:11, Paul Fulghum wrote: >> NAK to this patch. It causes lost wakeups in both read and write paths. >> >> The write path does not need changing. >> >> The read path can be fixed by setting current to TASK_RUNNING at the top of >> the if (rbuf) block >> so the warning is not triggered by copy_to_user(). If this block runs the >> condition is satisfied >> and it breaks out of the polling loop where it is already being set to >> TASK_RUNNING and removed >> from the wait queue. This particular path just needs to account for the >> copy_to_user which occurs >> before breaking out. >> >> I’ll make a patch to do this when I have the ability to test it in a day or >> two. >> > > OK. Then, any chance it is rewritten using wait_event_interruptible() in > order to reduce lines? > ( wait_event_interruptible() automatically calls might_sleep(), but is it > acceptable for you? ) > This looks good to me. I applied it and tested blocking (sleep/no sleep) and non-blocking (success/EAGAIN) paths for both read and write.
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
On 2019/01/01 12:11, Paul Fulghum wrote: > NAK to this patch. It causes lost wakeups in both read and write paths. > > The write path does not need changing. > > The read path can be fixed by setting current to TASK_RUNNING at the top of > the if (rbuf) block > so the warning is not triggered by copy_to_user(). If this block runs the > condition is satisfied > and it breaks out of the polling loop where it is already being set to > TASK_RUNNING and removed > from the wait queue. This particular path just needs to account for the > copy_to_user which occurs > before breaking out. > > I’ll make a patch to do this when I have the ability to test it in a day or > two. > OK. Then, any chance it is rewritten using wait_event_interruptible() in order to reduce lines? ( wait_event_interruptible() automatically calls might_sleep(), but is it acceptable for you? ) --- drivers/tty/n_hdlc.c | 126 --- 1 file changed, 30 insertions(+), 96 deletions(-) diff --git a/drivers/tty/n_hdlc.c b/drivers/tty/n_hdlc.c index 8223d02..2e4ccf9 100644 --- a/drivers/tty/n_hdlc.c +++ b/drivers/tty/n_hdlc.c @@ -562,8 +562,7 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, { struct n_hdlc *n_hdlc = tty2n_hdlc(tty); int ret = 0; - struct n_hdlc_buf *rbuf; - DECLARE_WAITQUEUE(wait, current); + struct n_hdlc_buf *rbuf = NULL; if (debuglevel >= DEBUG_LEVEL_INFO) printk("%s(%d)n_hdlc_tty_read() called\n",__FILE__,__LINE__); @@ -579,58 +578,26 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, return -EFAULT; } - add_wait_queue(>read_wait, ); - - for (;;) { - if (test_bit(TTY_OTHER_CLOSED, >flags)) { - ret = -EIO; - break; - } - if (tty_hung_up_p(file)) - break; - - set_current_state(TASK_INTERRUPTIBLE); - - rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list); - if (rbuf) { - if (rbuf->count > nr) { - /* too large for caller's buffer */ - ret = -EOVERFLOW; - } else { - __set_current_state(TASK_RUNNING); - if (copy_to_user(buf, rbuf->buf, rbuf->count)) - ret = -EFAULT; - else - ret = rbuf->count; - } - - if (n_hdlc->rx_free_buf_list.count > - DEFAULT_RX_BUF_COUNT) - kfree(rbuf); - else - n_hdlc_buf_put(_hdlc->rx_free_buf_list, rbuf); - break; - } - - /* no data */ - if (tty_io_nonblock(tty, file)) { - ret = -EAGAIN; - break; - } - - schedule(); - - if (signal_pending(current)) { - ret = -EINTR; - break; - } + if (wait_event_interruptible(tty->read_wait, +(ret = -EIO, test_bit(TTY_OTHER_CLOSED, >flags)) || +(ret = 0, tty_hung_up_p(file)) || +(rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list)) != NULL || +(ret = -EAGAIN, tty_io_nonblock(tty, file + return -EINTR; + if (rbuf) { + if (rbuf->count > nr) + /* too large for caller's buffer */ + ret = -EOVERFLOW; + else if (copy_to_user(buf, rbuf->buf, rbuf->count)) + ret = -EFAULT; + else + ret = rbuf->count; + if (n_hdlc->rx_free_buf_list.count > DEFAULT_RX_BUF_COUNT) + kfree(rbuf); + else + n_hdlc_buf_put(_hdlc->rx_free_buf_list, rbuf); } - - remove_wait_queue(>read_wait, ); - __set_current_state(TASK_RUNNING); - return ret; - } /* end of n_hdlc_tty_read() */ /** @@ -645,21 +612,13 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, static ssize_t n_hdlc_tty_write(struct tty_struct *tty, struct file *file, const unsigned char *data, size_t count) { - struct n_hdlc *n_hdlc = tty2n_hdlc (tty); + struct n_hdlc *n_hdlc; int error = 0; - DECLARE_WAITQUEUE(wait, current); - struct n_hdlc_buf *tbuf; + struct n_hdlc_buf *tbuf = NULL; if (debuglevel >= DEBUG_LEVEL_INFO) printk("%s(%d)n_hdlc_tty_write() called count=%zd\n", __FILE__,__LINE__,count); -
Re: [PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
On Dec 31, 2018, at 7:11 PM, Paul Fulghum wrote: NAK to this patch. It causes lost wakeups in both read and write paths. The write path does not need changing. The read path can be fixed by setting current to TASK_RUNNING at the top of the if (rbuf) block so the warning is not triggered by copy_to_user(). If this block runs the condition is satisfied and it breaks out of the polling loop where it is already being set to TASK_RUNNING and removed from the wait queue. This particular path just needs to account for the copy_to_user which occurs before breaking out. I’ll make a patch to do this when I have the ability to test it in a day or two. > On Dec 29, 2018, at 3:48 AM, Tetsuo Handa > wrote: > > syzbot is hitting __might_sleep() warning [1], for commit 1035b63d3c6fc34a > ("n_hdlc: fix read and write locking") changed to set TASK_INTERRUPTIBLE > state before calling copy_to_user(). Let's set TASK_INTERRUPTIBLE state > immediately before calling schedule(). > > [1] > https://syzkaller.appspot.com/bug?id=17d5de7f1fcab794cb8c40032f893f52de899324 > > Signed-off-by: Tetsuo Handa > Reported-by: syzbot > Cc: Paul Fulghum > Cc: Arnd Bergmann > Cc: Alan Cox > --- > drivers/tty/n_hdlc.c | 7 +++ > 1 file changed, 3 insertions(+), 4 deletions(-) > > diff --git a/drivers/tty/n_hdlc.c b/drivers/tty/n_hdlc.c > index dabb391..7835489 100644 > --- a/drivers/tty/n_hdlc.c > +++ b/drivers/tty/n_hdlc.c > @@ -589,8 +589,6 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, > struct file *file, > if (tty_hung_up_p(file)) > break; > > - set_current_state(TASK_INTERRUPTIBLE); > - > rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list); > if (rbuf) { > if (rbuf->count > nr) { > @@ -617,6 +615,7 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, > struct file *file, > break; > } > > + set_current_state(TASK_INTERRUPTIBLE); > schedule(); > > if (signal_pending(current)) { > @@ -673,8 +672,6 @@ static ssize_t n_hdlc_tty_write(struct tty_struct *tty, > struct file *file, > add_wait_queue(>write_wait, ); > > for (;;) { > - set_current_state(TASK_INTERRUPTIBLE); > - > tbuf = n_hdlc_buf_get(_hdlc->tx_free_buf_list); > if (tbuf) > break; > @@ -683,6 +680,8 @@ static ssize_t n_hdlc_tty_write(struct tty_struct *tty, > struct file *file, > error = -EAGAIN; > break; > } > + > + set_current_state(TASK_INTERRUPTIBLE); > schedule(); > > n_hdlc = tty2n_hdlc (tty); > -- > 1.8.3.1 > > -- Paul Fulghum MicroGate Systems, Ltd. =Customer Driven, by Design= (512) 345-7791 x102 (Voice) (512) 343-9046 (Fax) Central Time Zone (GMT -5h) www.microgate.com
[PATCH] tty/n_hdlc: fix sleep in !TASK_RUNNING state warning
syzbot is hitting __might_sleep() warning [1], for commit 1035b63d3c6fc34a ("n_hdlc: fix read and write locking") changed to set TASK_INTERRUPTIBLE state before calling copy_to_user(). Let's set TASK_INTERRUPTIBLE state immediately before calling schedule(). [1] https://syzkaller.appspot.com/bug?id=17d5de7f1fcab794cb8c40032f893f52de899324 Signed-off-by: Tetsuo Handa Reported-by: syzbot Cc: Paul Fulghum Cc: Arnd Bergmann Cc: Alan Cox --- drivers/tty/n_hdlc.c | 7 +++ 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/drivers/tty/n_hdlc.c b/drivers/tty/n_hdlc.c index dabb391..7835489 100644 --- a/drivers/tty/n_hdlc.c +++ b/drivers/tty/n_hdlc.c @@ -589,8 +589,6 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, if (tty_hung_up_p(file)) break; - set_current_state(TASK_INTERRUPTIBLE); - rbuf = n_hdlc_buf_get(_hdlc->rx_buf_list); if (rbuf) { if (rbuf->count > nr) { @@ -617,6 +615,7 @@ static ssize_t n_hdlc_tty_read(struct tty_struct *tty, struct file *file, break; } + set_current_state(TASK_INTERRUPTIBLE); schedule(); if (signal_pending(current)) { @@ -673,8 +672,6 @@ static ssize_t n_hdlc_tty_write(struct tty_struct *tty, struct file *file, add_wait_queue(>write_wait, ); for (;;) { - set_current_state(TASK_INTERRUPTIBLE); - tbuf = n_hdlc_buf_get(_hdlc->tx_free_buf_list); if (tbuf) break; @@ -683,6 +680,8 @@ static ssize_t n_hdlc_tty_write(struct tty_struct *tty, struct file *file, error = -EAGAIN; break; } + + set_current_state(TASK_INTERRUPTIBLE); schedule(); n_hdlc = tty2n_hdlc (tty); -- 1.8.3.1