Re: [PATCH] n_tty: Fix unordered accesses to lockless read buffer

From: Peter Hurley
Date: Fri Jan 09 2015 - 14:09:59 EST


[ reviewing my own patch :) ]

On 12/30/2014 07:05 AM, Peter Hurley wrote:
> Add commit_head buffer index, which the producer-side publishes
> after input processing. This ensures the consumer-side observes
> correctly-ordered writes in raw mode (ie., the buffer data is
> written before the buffer index is advanced).
>
> Further, remove read_cnt() and expand inline, using ACCESS_ONCE()
> on the relevant buffer index; read_tail from the producer-side
> and canon_head/commit_head from the consumer-side, or both in shared
> paths such as receive_room().
>
> Based on work by Christian Riesch <christian.riesch@xxxxxxxxxx>
>
> NB: Exclusive access is still guaranteed with termios_rwsem write
> lock, eg. by n_tty_set_termios() and in n_tty_ioctl(); in this
> circumstance, commit_head is equivalent to read_head.
>
> Cc: Christian Riesch <christian.riesch@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx> # v3.14.x (will need backport to v3.12.x)
> Signed-off-by: Peter Hurley <peter@xxxxxxxxxxxxxxxxxx>
> ---
> drivers/tty/n_tty.c | 80 ++++++++++++++++++++++++++---------------------------
> 1 file changed, 40 insertions(+), 40 deletions(-)
>
> diff --git a/drivers/tty/n_tty.c b/drivers/tty/n_tty.c
> index d2b4967..a618b10 100644
> --- a/drivers/tty/n_tty.c
> +++ b/drivers/tty/n_tty.c
> @@ -90,6 +90,7 @@
> struct n_tty_data {
> /* producer-published */
> size_t read_head;
> + size_t commit_head; /* == read_head when not receiving */

commit_head is now the observable producer-side buffer index when
termios is !icanon || L_EXTPROC mode.

> size_t canon_head;
> size_t echo_head;
> size_t echo_commit;
> @@ -127,11 +128,6 @@ struct n_tty_data {
> struct mutex output_lock;
> };
>
> -static inline size_t read_cnt(struct n_tty_data *ldata)
> -{
> - return ldata->read_head - ldata->read_tail;
> -}
> -

Can keep read_cnt(). Will continue to be used in several places. See
other comments.

> static inline unsigned char read_buf(struct n_tty_data *ldata, size_t i)
> {
> return ldata->read_buf[i & (N_TTY_BUF_SIZE - 1)];
> @@ -164,15 +160,17 @@ static inline int tty_put_user(struct tty_struct *tty, unsigned char x,
> static int receive_room(struct tty_struct *tty)
> {
> struct n_tty_data *ldata = tty->disc_data;
> + size_t head = ACCESS_ONCE(ldata->commit_head);
> + size_t tail = ACCESS_ONCE(ldata->read_tail);

Producer-side receive_room() needs to be special-cased for the
call from the n_tty_receive_buf_common() i/o loop, because the
read_tail access needs to be:

size_t tail = smp_load_acquire(ldata->read_tail);

Paired with the consumer-side smp_store_release(read_tail), will
guarantee that the consumer has finished loading from the
read_buf before the producer may overwrite that data.

The producer-side receive_room() doesn't need the ACCESS_ONCE()s because
the commit_head and canon_head are only modified by itself (as the
producer).

The consumer-side receive_room() doesn't need the ACCESS_ONCE()s either.
At least one compiler barrier is passed through on each loop in
n_tty_read(), and at least once before starting the loop.

> int left;
>
> if (I_PARMRK(tty)) {
> - /* Multiply read_cnt by 3, since each byte might take up to
> + /* Multiply count by 3, since each byte might take up to
> * three times as many spaces when PARMRK is set (depending on
> * its flags, e.g. parity error). */
> - left = N_TTY_BUF_SIZE - read_cnt(ldata) * 3 - 1;
> + left = N_TTY_BUF_SIZE - (head - tail) * 3 - 1;
> } else
> - left = N_TTY_BUF_SIZE - read_cnt(ldata) - 1;
> + left = N_TTY_BUF_SIZE - (head - tail) - 1;
>
> /*
> * If we are doing input canonicalization, and there are no
> @@ -181,7 +179,7 @@ static int receive_room(struct tty_struct *tty)
> * characters will be beeped.
> */
> if (left <= 0)
> - left = ldata->icanon && ldata->canon_head == ldata->read_tail;
> + left = ldata->icanon && ACCESS_ONCE(ldata->canon_head) == tail;
>
> return left;
> }
> @@ -224,9 +222,9 @@ static ssize_t chars_in_buffer(struct tty_struct *tty)
> ssize_t n = 0;
>
> if (!ldata->icanon)
> - n = read_cnt(ldata);
> + n = ACCESS_ONCE(ldata->commit_head) - ldata->read_tail;
> else
> - n = ldata->canon_head - ldata->read_tail;
> + n = ACCESS_ONCE(ldata->canon_head) - ldata->read_tail;

Existing compiler barriers in the consumer-side i/o loop make
these ACCESS_ONCE()s unnecessary (as with receive_room() and input_available_p()).

> return n;
> }
>
> @@ -313,10 +311,6 @@ static void n_tty_check_unthrottle(struct tty_struct *tty)
> *
> * n_tty_receive_buf()/producer path:
> * caller holds non-exclusive termios_rwsem
> - * modifies read_head
> - *
> - * read_head is only considered 'published' if canonical mode is
> - * not active.
> */
>
> static inline void put_tty_queue(unsigned char c, struct n_tty_data *ldata)
> @@ -340,6 +334,7 @@ static void reset_buffer_flags(struct n_tty_data *ldata)
> {
> ldata->read_head = ldata->canon_head = ldata->read_tail = 0;
> ldata->echo_head = ldata->echo_tail = ldata->echo_commit = 0;
> + ldata->commit_head = 0;

Ok.

> ldata->echo_mark = 0;
> ldata->line_start = 0;
>
> @@ -987,10 +982,6 @@ static inline void finish_erasing(struct n_tty_data *ldata)
> *
> * n_tty_receive_buf()/producer path:
> * caller holds non-exclusive termios_rwsem
> - * modifies read_head
> - *
> - * Modifying the read_head is not considered a publish in this context
> - * because canonical mode is active -- only canon_head publishes
> */
>
> static void eraser(unsigned char c, struct tty_struct *tty)
> @@ -1139,7 +1130,7 @@ static void isig(int sig, struct tty_struct *tty)
> *
> * n_tty_receive_buf()/producer path:
> * caller holds non-exclusive termios_rwsem
> - * publishes read_head via put_tty_queue()
> + * publishes read_head as commit_head
> *
> * Note: may get exclusive termios_rwsem if flushing input buffer
> */
> @@ -1166,6 +1157,8 @@ static void n_tty_receive_break(struct tty_struct *tty)
> put_tty_queue('\0', ldata);
> }
> put_tty_queue('\0', ldata);
> + /* publish read_head to consumer */
> + smp_store_release(&ldata->commit_head, ldata->read_head);
> if (waitqueue_active(&tty->read_wait))
> wake_up_interruptible_poll(&tty->read_wait, POLLIN);

I intend to remove the wake_up (along with the comment changes above)
rather than add the publish.

It turns out that the wake_up cannot cause a userspace observable change,
that the wake_up() in __receive_buf() would not also guarantee.

Analysis:

Case 1) ICANON == true && L_EXTPROC == 0

The canon_head has not advanced because there is no newline so
input_available_p() is false for both n_tty_read() and n_tty_poll(),
so nothing happens.

Case 2) ICANON == true && L_EXTPROC != 0

The wake_up in __receive_buf() still happens when the input block
finishes processing, which is an insignificant delay.

Case 3) ICANON == false and the following matrix applies:

MIN_CHAR == 0 | MIN_CHAR == 0
TIME_CHAR != 0 | TIME_CHAR == 0
|
minimum_to_wake => 1 | minimum_to_wake => 1
|
## same as Case 2 ## | ## same as Case 2 ##
|
--------------------------------------------------------------------
|
MIN_CHAR != 0 | MIN_CHAR != 0
TIME_CHAR != 0 | TIME_CHAR == 0
|
minimum_to_wake => 1 | minimum_to_wake => MIN_CHAR()
|
## same as Case 2 ## | ## see below ##
|

When MIN_CHAR != 0 && TIME_CHAR == 0, input_available_p() is false when
called from n_tty_poll() until minimum_to_wake # of chars has accumulated,
which is the same condition under which the wake_up will occur in __receive_buf().

Similarly, although input_available_p() is true when called from n_tty_read(),
the reader i/o loop will not return the buffer back to userspace until
minimum_to_wake # of chars has accumulated, which is the same condition under which
the wake_up will occur in __receive_buf().

> }
> @@ -1209,7 +1202,7 @@ static void n_tty_receive_overrun(struct tty_struct *tty)
> *
> * n_tty_receive_buf()/producer path:
> * caller holds non-exclusive termios_rwsem
> - * publishes read_head via put_tty_queue()
> + * publishes read_head as commit_head
> */
> static void n_tty_receive_parity_error(struct tty_struct *tty, unsigned char c)
> {
> @@ -1226,6 +1219,8 @@ static void n_tty_receive_parity_error(struct tty_struct *tty, unsigned char c)
> put_tty_queue('\0', ldata);
> } else
> put_tty_queue(c, ldata);
> + /* publish read_head to consumer */
> + smp_store_release(&ldata->commit_head, ldata->read_head);
> if (waitqueue_active(&tty->read_wait))
> wake_up_interruptible_poll(&tty->read_wait, POLLIN);

Same as above.

> }
> @@ -1263,7 +1258,6 @@ n_tty_receive_signal_char(struct tty_struct *tty, int signal, unsigned char c)
> * n_tty_receive_buf()/producer path:
> * caller holds non-exclusive termios_rwsem
> * publishes canon_head if canonical mode is active
> - * otherwise, publishes read_head via put_tty_queue()
> *
> * Returns 1 if LNEXT was received, else returns 0
> */
> @@ -1376,7 +1370,7 @@ n_tty_receive_char_special(struct tty_struct *tty, unsigned char c)
> handle_newline:
> set_bit(ldata->read_head & (N_TTY_BUF_SIZE - 1), ldata->read_flags);
> put_tty_queue(c, ldata);
> - ldata->canon_head = ldata->read_head;
> + smp_store_release(&ldata->canon_head, ldata->read_head);

Required.

> kill_fasync(&tty->fasync, SIGIO, POLL_IN);
> if (waitqueue_active(&tty->read_wait))
> wake_up_interruptible_poll(&tty->read_wait, POLLIN);
> @@ -1526,7 +1520,7 @@ n_tty_receive_char_lnext(struct tty_struct *tty, unsigned char c, char flag)
> *
> * n_tty_receive_buf()/producer path:
> * claims non-exclusive termios_rwsem
> - * publishes read_head and canon_head
> + * publishes canon_head or commit_head

Ok.

> */
>
> static void
> @@ -1534,10 +1528,11 @@ n_tty_receive_buf_real_raw(struct tty_struct *tty, const unsigned char *cp,
> char *fp, int count)
> {
> struct n_tty_data *ldata = tty->disc_data;
> + size_t tail = ACCESS_ONCE(ldata->read_tail);

Forcing the compiler to load from memory is not necessary here
since the read_tail will be loaded on each loop through
producer_side receive_room().

> size_t n, head;
>
> head = ldata->read_head & (N_TTY_BUF_SIZE - 1);
> - n = N_TTY_BUF_SIZE - max(read_cnt(ldata), head);
> + n = N_TTY_BUF_SIZE - max(ldata->read_head - tail, head);

Not required since above change is unnecessary.

> n = min_t(size_t, count, n);
> memcpy(read_buf_addr(ldata, head), cp, n);
> ldata->read_head += n;
> @@ -1545,7 +1540,7 @@ n_tty_receive_buf_real_raw(struct tty_struct *tty, const unsigned char *cp,
> count -= n;
>
> head = ldata->read_head & (N_TTY_BUF_SIZE - 1);
> - n = N_TTY_BUF_SIZE - max(read_cnt(ldata), head);
> + n = N_TTY_BUF_SIZE - max(ldata->read_head - tail, head);

Same.

> n = min_t(size_t, count, n);
> memcpy(read_buf_addr(ldata, head), cp, n);
> ldata->read_head += n;
> @@ -1649,6 +1644,7 @@ static void __receive_buf(struct tty_struct *tty, const unsigned char *cp,
> {
> struct n_tty_data *ldata = tty->disc_data;
> bool preops = I_ISTRIP(tty) || (I_IUCLC(tty) && L_IEXTEN(tty));
> + size_t c;

Unnecessary. See next 2 comments.
>
> if (ldata->real_raw)
> n_tty_receive_buf_real_raw(tty, cp, fp, count);
> @@ -1676,8 +1672,11 @@ static void __receive_buf(struct tty_struct *tty, const unsigned char *cp,
> tty->ops->flush_chars(tty);
> }
>
> - if ((!ldata->icanon && (read_cnt(ldata) >= ldata->minimum_to_wake)) ||
> - L_EXTPROC(tty)) {
> + /* publish read_head to consumer */
> + smp_store_release(&ldata->commit_head, ldata->read_head);
> + c = ldata->read_head - ACCESS_ONCE(ldata->read_tail);

ACCESS_ONCE(read_tail) is not required since the smp_store_release() is a
compiler barrier, so there can be no local cache of ldata->read_tail for
the compiler to skip loading read_tail.

> +
> + if ((!ldata->icanon && c >= ldata->minimum_to_wake) || L_EXTPROC(tty)) {

At this point, the read_head == commit_head on this cpu so read_cnt() is
equivalent.

> kill_fasync(&tty->fasync, SIGIO, POLL_IN);
> if (waitqueue_active(&tty->read_wait))
> wake_up_interruptible_poll(&tty->read_wait, POLLIN);
> @@ -1755,13 +1754,13 @@ static void n_tty_set_termios(struct tty_struct *tty, struct ktermios *old)
> if (!old || (old->c_lflag ^ tty->termios.c_lflag) & ICANON) {
> bitmap_zero(ldata->read_flags, N_TTY_BUF_SIZE);
> ldata->line_start = ldata->read_tail;
> - if (!L_ICANON(tty) || !read_cnt(ldata)) {
> + if (!L_ICANON(tty) || ldata->commit_head == ldata->read_tail) {
> ldata->canon_head = ldata->read_tail;
> ldata->push = 0;
> } else {
> - set_bit((ldata->read_head - 1) & (N_TTY_BUF_SIZE - 1),
> + set_bit((ldata->commit_head - 1) & (N_TTY_BUF_SIZE - 1),
> ldata->read_flags);
> - ldata->canon_head = ldata->read_head;
> + ldata->canon_head = ldata->commit_head;

None of this is necessary because read_head == commit_head when the termios_rwsem
is write locked (as is the case here).


> ldata->push = 1;
> }
> ldata->erasing = 0;
> @@ -1903,9 +1902,9 @@ static inline int input_available_p(struct tty_struct *tty, int poll)
> int amt = poll && !TIME_CHAR(tty) && MIN_CHAR(tty) ? MIN_CHAR(tty) : 1;
>
> if (ldata->icanon && !L_EXTPROC(tty))
> - return ldata->canon_head != ldata->read_tail;
> + return ACCESS_ONCE(ldata->canon_head) != ldata->read_tail;

Unnecessary. At least one compiler barrier has been passed through to get here,
either by looping or first time.

> else
> - return read_cnt(ldata) >= amt;
> + return ACCESS_ONCE(ldata->commit_head) - ldata->read_tail >= amt;

The ACCESS_ONCE() is not required, for the same reason as above.
But using commit_head instead of read_head is required, so this should be:

return ldata->commit_head - ldata->read_tail >= amt;

> }
>
> /**
> @@ -1938,9 +1937,10 @@ static int copy_from_read_buf(struct tty_struct *tty,
> size_t n;
> bool is_eof;
> size_t tail = ldata->read_tail & (N_TTY_BUF_SIZE - 1);
> + size_t head = smp_load_acquire(&ldata->commit_head);
>
> retval = 0;
> - n = min(read_cnt(ldata), N_TTY_BUF_SIZE - tail);
> + n = min(head - ldata->read_tail, N_TTY_BUF_SIZE - tail);



> n = min(*nr, n);
> if (n) {
> retval = copy_to_user(*b, read_buf_addr(ldata, tail), n);
> @@ -1948,9 +1948,10 @@ static int copy_from_read_buf(struct tty_struct *tty,
> is_eof = n == 1 && read_buf(ldata, tail) == EOF_CHAR(tty);
> tty_audit_add_data(tty, read_buf_addr(ldata, tail), n,
> ldata->icanon);
> - ldata->read_tail += n;
> + smp_store_release(&ldata->read_tail, ldata->read_tail + n);

Absolutely required. Ensures that the read_buf data has actually been loaded
by this cpu before the other cpu running the producer side can observe the
read_tail advancing and overwriting the read_buf data.

> /* Turn single EOF into zero-length read */
> - if (L_EXTPROC(tty) && ldata->icanon && is_eof && !read_cnt(ldata))
> + if (L_EXTPROC(tty) && ldata->icanon && is_eof &&
> + (head == ldata->read_tail))

Required since commit_head is now the observed producer buffer index
(instead of read_head).

> n = 0;
> *b += n;
> *nr -= n;
> @@ -1993,7 +1994,7 @@ static int canon_copy_from_read_buf(struct tty_struct *tty,
> bool eof_push = 0;
>
> /* N.B. avoid overrun if nr == 0 */
> - n = min(*nr, read_cnt(ldata));
> + n = min(*nr, smp_load_acquire(&ldata->canon_head) - ldata->read_tail);

Absolutely required to guarantee the buffer data has been written by the
producer. Keep.

> if (!n)
> return 0;
>
> @@ -2043,8 +2044,7 @@ static int canon_copy_from_read_buf(struct tty_struct *tty,
>
> if (found)
> clear_bit(eol, ldata->read_flags);
> - smp_mb__after_atomic();
> - ldata->read_tail += c;
> + smp_store_release(&ldata->read_tail, ldata->read_tail + c);

Strictly speaking, this is the same thing so not really necessary, but
I think it does a better job of self-documenting. Keep.

>
> if (found) {
> if (!ldata->push)
> @@ -2458,7 +2458,7 @@ static int n_tty_ioctl(struct tty_struct *tty, struct file *file,
> if (L_ICANON(tty))
> retval = inq_canon(ldata);
> else
> - retval = read_cnt(ldata);
> + retval = ldata->commit_head - ldata->read_tail;

Unnecessary. read_head == commit_head when termios_rwsem is write-locked,
as is the case here.

> up_write(&tty->termios_rwsem);
> return put_user(retval, (unsigned int __user *) arg);
> default:
>

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/