Hi,
The commit 12e84c71b7d4ee (tty: export tty_open_by_driver) exports
tty_open_by_device to allow tty to be opened from inside kernel which
works fine except that it doesn't handle contention with user space or
another kernel-space open of the same tty. For example, opening a tty
from user space while it is kernel opened results in failure and a
kernel log message about mismatch between tty->count and tty's file open
count.
I suggest we make kernel access to tty exclusive so that if a user
process or kernel opens a kernel opened tty, it gets -EBUSY. Below is
a patch which does this by adding TTY_KOPENED flag to tty->flags. When
this flag is set, tty_open_by_driver returns -EBUSY. Instead of
overlaoding tty_open_by_driver for both kernel and user space, this
patch creates a separate function tty_kopen which closely follows
tty_open_by_driver.
I am far from an expert on tty and this patch might contain the wrong
approach. But it should convey what I mean.
Thanks,
Okash
---
drivers/staging/speakup/spk_ttyio.c | 2 -
drivers/tty/tty_io.c | 56 ++++++++++++++++++++++++++++++++++--
include/linux/tty.h | 7 +---
3 files changed, 58 insertions(+), 7 deletions(-)
--- a/drivers/staging/speakup/spk_ttyio.c
+++ b/drivers/staging/speakup/spk_ttyio.c
@@ -164,7 +164,7 @@ static int spk_ttyio_initialise_ldisc(st
if (ret)
return ret;
- tty = tty_open_by_driver(dev, NULL, NULL);
+ tty = tty_kopen(dev);
if (IS_ERR(tty))
return PTR_ERR(tty);
--- a/drivers/tty/tty_io.c
+++ b/drivers/tty/tty_io.c
@@ -1786,6 +1786,53 @@ static struct tty_driver *tty_lookup_dri
}
/**
+ * tty_kopen - open a tty device for kernel
+ * @device: dev_t of device to open
+ *
+ * Opens tty exclusively for kernel. Performs the driver lookup,
+ * makes sure it's not already opened and performs the first-time
+ * tty initialization.
+ *
+ * Returns the locked initialized &tty_struct
+ *
+ * Claims the global tty_mutex to serialize:
+ * - concurrent first-time tty initialization
+ * - concurrent tty driver removal w/ lookup
+ * - concurrent tty removal from driver table
+ */
+struct tty_struct *tty_kopen(dev_t device)
+{
+ struct tty_struct *tty;
+ struct tty_driver *driver = NULL;
+ int index = -1;
+
+ mutex_lock(&tty_mutex);
+ driver = tty_lookup_driver(device, NULL, &index);
+ if (IS_ERR(driver)) {
+ mutex_unlock(&tty_mutex);
+ return ERR_CAST(driver);
+ }
+
+ /* check whether we're reopening an existing tty */
+ tty = tty_driver_lookup_tty(driver, NULL, index);
+ if (IS_ERR(tty))
+ goto out;
+
+ if (tty) {
+ tty_kref_put(tty); /* drop kref from tty_driver_lookup_tty() */
+ tty = ERR_PTR(-EBUSY);
+ } else { /* Returns with the tty_lock held for now */
+ tty = tty_init_dev(driver, index);
+ set_bit(TTY_KOPENED, &tty->flags);
+ }
+out:
+ mutex_unlock(&tty_mutex);
+ tty_driver_kref_put(driver);
+ return tty;
+}
+EXPORT_SYMBOL(tty_kopen);
+
+/**
* tty_open_by_driver - open a tty device
* @device: dev_t of device to open
* @inode: inode of device file
@@ -1801,7 +1848,7 @@ static struct tty_driver *tty_lookup_dri
* - concurrent tty driver removal w/ lookup
* - concurrent tty removal from driver table
*/
-struct tty_struct *tty_open_by_driver(dev_t device, struct inode *inode,
+static struct tty_struct *tty_open_by_driver(dev_t device, struct inode *inode,
struct file *filp)
{
struct tty_struct *tty;
@@ -1824,6 +1871,12 @@ struct tty_struct *tty_open_by_driver(de
}
if (tty) {
+ if (test_bit(TTY_KOPENED, &tty->flags)) {
+ tty_kref_put(tty);
+ mutex_unlock(&tty_mutex);
+ tty = ERR_PTR(-EBUSY);
+ goto out;
+ }
mutex_unlock(&tty_mutex);
retval = tty_lock_interruptible(tty);
tty_kref_put(tty); /* drop kref from tty_driver_lookup_tty() */
@@ -1846,7 +1899,6 @@ out:
tty_driver_kref_put(driver);
return tty;
}
-EXPORT_SYMBOL_GPL(tty_open_by_driver);
/**
* tty_open - open a tty device
--- a/include/linux/tty.h
+++ b/include/linux/tty.h
@@ -362,6 +362,7 @@ struct tty_file_private {
#define TTY_NO_WRITE_SPLIT 17 /* Preserve write boundaries to driver */
#define TTY_HUPPED 18 /* Post driver->hangup() */
#define TTY_LDISC_HALTED 22 /* Line discipline is halted */
+#define TTY_KOPENED 23 /* TTY exclusively opened by kernel */
/* Values for tty->flow_change */
#define TTY_THROTTLE_SAFE 1
@@ -399,8 +400,7 @@ extern struct tty_struct *get_current_tt
/* tty_io.c */
extern int __init tty_init(void);
extern const char *tty_name(const struct tty_struct *tty);
-extern struct tty_struct *tty_open_by_driver(dev_t device, struct inode *inode,
- struct file *filp);
+extern struct tty_struct *tty_kopen(dev_t device);
extern int tty_dev_name_to_number(const char *name, dev_t *number);
#else
static inline void tty_kref_put(struct tty_struct *tty)
@@ -422,8 +422,7 @@ static inline int __init tty_init(void)
{ return 0; }
static inline const char *tty_name(const struct tty_struct *tty)
{ return "(none)"; }
-static inline struct tty_struct *tty_open_by_driver(dev_t device,
- struct inode *inode, struct file *filp)
+static inline struct tty_struct *tty_kopen(dev_t device)
{ return NULL; }
static inline int tty_dev_name_to_number(const char *name, dev_t *number)
{ return -ENOTSUPP; }
On Fri, Jul 07, 2017 at 09:28:41PM +0100, Okash Khawaja wrote:
> Hi,
>
> The commit 12e84c71b7d4ee (tty: export tty_open_by_driver) exports
> tty_open_by_device to allow tty to be opened from inside kernel which
> works fine except that it doesn't handle contention with user space or
> another kernel-space open of the same tty. For example, opening a tty
> from user space while it is kernel opened results in failure and a
> kernel log message about mismatch between tty->count and tty's file open
> count.
>
> I suggest we make kernel access to tty exclusive so that if a user
> process or kernel opens a kernel opened tty, it gets -EBUSY. Below is
> a patch which does this by adding TTY_KOPENED flag to tty->flags. When
> this flag is set, tty_open_by_driver returns -EBUSY. Instead of
> overlaoding tty_open_by_driver for both kernel and user space, this
> patch creates a separate function tty_kopen which closely follows
> tty_open_by_driver.
>
> I am far from an expert on tty and this patch might contain the wrong
> approach. But it should convey what I mean.
>
> Thanks,
> Okash
>
> ---
> drivers/staging/speakup/spk_ttyio.c | 2 -
> drivers/tty/tty_io.c | 56 ++++++++++++++++++++++++++++++++++--
> include/linux/tty.h | 7 +---
> 3 files changed, 58 insertions(+), 7 deletions(-)
>
> --- a/drivers/staging/speakup/spk_ttyio.c
> +++ b/drivers/staging/speakup/spk_ttyio.c
> @@ -164,7 +164,7 @@ static int spk_ttyio_initialise_ldisc(st
> if (ret)
> return ret;
>
> - tty = tty_open_by_driver(dev, NULL, NULL);
> + tty = tty_kopen(dev);
> if (IS_ERR(tty))
> return PTR_ERR(tty);
>
> --- a/drivers/tty/tty_io.c
> +++ b/drivers/tty/tty_io.c
> @@ -1786,6 +1786,53 @@ static struct tty_driver *tty_lookup_dri
> }
>
> /**
> + * tty_kopen - open a tty device for kernel
> + * @device: dev_t of device to open
If nothing else, this api call is much simpler overall, as your use
above shows :)
> + *
> + * Opens tty exclusively for kernel. Performs the driver lookup,
> + * makes sure it's not already opened and performs the first-time
> + * tty initialization.
> + *
> + * Returns the locked initialized &tty_struct
> + *
> + * Claims the global tty_mutex to serialize:
> + * - concurrent first-time tty initialization
> + * - concurrent tty driver removal w/ lookup
> + * - concurrent tty removal from driver table
> + */
> +struct tty_struct *tty_kopen(dev_t device)
> +{
> + struct tty_struct *tty;
> + struct tty_driver *driver = NULL;
> + int index = -1;
> +
> + mutex_lock(&tty_mutex);
> + driver = tty_lookup_driver(device, NULL, &index);
> + if (IS_ERR(driver)) {
> + mutex_unlock(&tty_mutex);
> + return ERR_CAST(driver);
> + }
> +
> + /* check whether we're reopening an existing tty */
> + tty = tty_driver_lookup_tty(driver, NULL, index);
> + if (IS_ERR(tty))
> + goto out;
> +
> + if (tty) {
> + tty_kref_put(tty); /* drop kref from tty_driver_lookup_tty() */
Put the comment above the line?
> + tty = ERR_PTR(-EBUSY);
> + } else { /* Returns with the tty_lock held for now */
I don't understand this comment.
> + tty = tty_init_dev(driver, index);
> + set_bit(TTY_KOPENED, &tty->flags);
> + }
> +out:
> + mutex_unlock(&tty_mutex);
> + tty_driver_kref_put(driver);
> + return tty;
> +}
> +EXPORT_SYMBOL(tty_kopen);
EXPORT_SYMBOL_GPL()? :)
> +
> +/**
> * tty_open_by_driver - open a tty device
> * @device: dev_t of device to open
> * @inode: inode of device file
> @@ -1801,7 +1848,7 @@ static struct tty_driver *tty_lookup_dri
> * - concurrent tty driver removal w/ lookup
> * - concurrent tty removal from driver table
> */
> -struct tty_struct *tty_open_by_driver(dev_t device, struct inode *inode,
> +static struct tty_struct *tty_open_by_driver(dev_t device, struct inode *inode,
> struct file *filp)
> {
> struct tty_struct *tty;
> @@ -1824,6 +1871,12 @@ struct tty_struct *tty_open_by_driver(de
> }
>
> if (tty) {
> + if (test_bit(TTY_KOPENED, &tty->flags)) {
> + tty_kref_put(tty);
> + mutex_unlock(&tty_mutex);
> + tty = ERR_PTR(-EBUSY);
> + goto out;
> + }
> mutex_unlock(&tty_mutex);
> retval = tty_lock_interruptible(tty);
> tty_kref_put(tty); /* drop kref from tty_driver_lookup_tty() */
> @@ -1846,7 +1899,6 @@ out:
> tty_driver_kref_put(driver);
> return tty;
> }
> -EXPORT_SYMBOL_GPL(tty_open_by_driver);
>
> /**
> * tty_open - open a tty device
> --- a/include/linux/tty.h
> +++ b/include/linux/tty.h
> @@ -362,6 +362,7 @@ struct tty_file_private {
> #define TTY_NO_WRITE_SPLIT 17 /* Preserve write boundaries to driver */
> #define TTY_HUPPED 18 /* Post driver->hangup() */
> #define TTY_LDISC_HALTED 22 /* Line discipline is halted */
> +#define TTY_KOPENED 23 /* TTY exclusively opened by kernel */
Minor nit, please use tabs here.
Overall, the idea looks sane to me. Keeping userspace from opening a
tty that the kernel has opened internally makes sense, hopefully
userspace doesn't get too confused when that happens. I don't think we
normally return -EBUSY from an open call, have you seen what happens
with apps when you do this (like minicom?)
thanks,
greg k-h
On Sat, Jul 08, 2017 at 10:38:03AM +0200, Greg Kroah-Hartman wrote:
> > +
> > + if (tty) {
> > + tty_kref_put(tty); /* drop kref from tty_driver_lookup_tty() */
>
> Put the comment above the line?
>
Sure
> > + tty = ERR_PTR(-EBUSY);
> > + } else { /* Returns with the tty_lock held for now */
>
> I don't understand this comment.
>
This is same comment in tty_open_by_driver. Basically, tty_init_dev
returns tty locked so that it doesn't dissapear from underneath the
caller. I am not sure about the "for now" part of the comment. I'll
remove it.
> > + tty = tty_init_dev(driver, index);
> > + set_bit(TTY_KOPENED, &tty->flags);
> > + }
> > +out:
> > + mutex_unlock(&tty_mutex);
> > + tty_driver_kref_put(driver);
> > + return tty;
> > +}
> > +EXPORT_SYMBOL(tty_kopen);
>
> EXPORT_SYMBOL_GPL()? :)
>
Of course, will update
> > /**
> > * tty_open - open a tty device
> > --- a/include/linux/tty.h
> > +++ b/include/linux/tty.h
> > @@ -362,6 +362,7 @@ struct tty_file_private {
> > #define TTY_NO_WRITE_SPLIT 17 /* Preserve write boundaries to driver */
> > #define TTY_HUPPED 18 /* Post driver->hangup() */
> > #define TTY_LDISC_HALTED 22 /* Line discipline is halted */
> > +#define TTY_KOPENED 23 /* TTY exclusively opened by kernel */
>
> Minor nit, please use tabs here.
>
Will do
> Overall, the idea looks sane to me. Keeping userspace from opening a
> tty that the kernel has opened internally makes sense, hopefully
> userspace doesn't get too confused when that happens. I don't think we
> normally return -EBUSY from an open call, have you seen what happens
> with apps when you do this (like minicom?)
>
Yes, returning -EBUSY is a change in the interface. I will test against
applications like minicom before resubmitting this.
Thanks,
Okash
On Sat, Jul 08, 2017 at 10:38:03AM +0200, Greg Kroah-Hartman wrote:
> Overall, the idea looks sane to me. Keeping userspace from opening a
> tty that the kernel has opened internally makes sense, hopefully
> userspace doesn't get too confused when that happens. I don't think we
> normally return -EBUSY from an open call, have you seen what happens
> with apps when you do this (like minicom?)
>
I tested this wil minincom, picocom and commands like "echo foo >
/dev/ttyS0". They all correctly report "Device or resource busy".
I have addressed all the comments you made. I have also split the patch
into three. Following is summary of each.
Patch 1: introduces the tty_kopen function and checks for TTY_KOPENED
Patch 2: updates speakup code to use tty_kopen instead of
tty_open_by_driver
Patch 3: reverses the export of tty_open_by_driver
Thanks,
Okash
On Sun, Jul 09, 2017 at 12:41:53PM +0100, Okash Khawaja wrote:
> On Sat, Jul 08, 2017 at 10:38:03AM +0200, Greg Kroah-Hartman wrote:
> > Overall, the idea looks sane to me. Keeping userspace from opening a
> > tty that the kernel has opened internally makes sense, hopefully
> > userspace doesn't get too confused when that happens. I don't think we
> > normally return -EBUSY from an open call, have you seen what happens
> > with apps when you do this (like minicom?)
> >
> I tested this wil minincom, picocom and commands like "echo foo >
> /dev/ttyS0". They all correctly report "Device or resource busy".
>
> I have addressed all the comments you made. I have also split the patch
> into three. Following is summary of each.
>
> Patch 1: introduces the tty_kopen function and checks for TTY_KOPENED
> Patch 2: updates speakup code to use tty_kopen instead of
> tty_open_by_driver
> Patch 3: reverses the export of tty_open_by_driver
Looks great, only one tiny comment about the return value from me, and
really, it's not a big deal at all, you can send a patch 4/3 to change
that up if you want. No one really runs without the tty layer enabled :)
If there are no other objections, I'll queue this up after 4.13-rc1 is
out, nice job.
greg k-h
When TTY is not built, tty_kopen should return an error. This way
calling code remains consistent, regardless of whether tty is built or
not. This patch returns -ENODEV when there is no tty.
Signed-off-by: Okash Khawaja <[email protected]>
---
include/linux/tty.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
--- a/include/linux/tty.h
+++ b/include/linux/tty.h
@@ -423,7 +423,7 @@ static inline int __init tty_init(void)
static inline const char *tty_name(const struct tty_struct *tty)
{ return "(none)"; }
static inline struct tty_struct *tty_kopen(dev_t device)
-{ return NULL; }
+{ return ERR_PTR(-ENODEV); }
static inline int tty_dev_name_to_number(const char *name, dev_t *number)
{ return -ENOTSUPP; }
#endif
On Sun, 09 Jul 2017 12:41:53 +0100
Okash Khawaja <[email protected]> wrote:
> On Sat, Jul 08, 2017 at 10:38:03AM +0200, Greg Kroah-Hartman wrote:
> > Overall, the idea looks sane to me. Keeping userspace from opening a
> > tty that the kernel has opened internally makes sense, hopefully
> > userspace doesn't get too confused when that happens. I don't think we
> > normally return -EBUSY from an open call, have you seen what happens
> > with apps when you do this (like minicom?)
> >
> I tested this wil minincom, picocom and commands like "echo foo >
> /dev/ttyS0". They all correctly report "Device or resource busy".
>
> I have addressed all the comments you made. I have also split the patch
> into three. Following is summary of each.
If the tty counts are being misreported then it would be better to fix
the code to actually manage the counts properly. The core tty code is
telling you that the tty is not in a valid state. While this is of
itself a good API to have, the underlying reference miscounting ought
IMHO to be fixed as well.
Also you don't need a new TTY_KOPENED flag as far as I can see. All tty's
have a usage count and active bit flags already. Use those.
Alan
On Mon, Jul 10, 2017 at 12:52:33PM +0100, Alan Cox wrote:
> On Sun, 09 Jul 2017 12:41:53 +0100
> Okash Khawaja <[email protected]> wrote:
>
> > On Sat, Jul 08, 2017 at 10:38:03AM +0200, Greg Kroah-Hartman wrote:
> > > Overall, the idea looks sane to me. Keeping userspace from opening a
> > > tty that the kernel has opened internally makes sense, hopefully
> > > userspace doesn't get too confused when that happens. I don't think we
> > > normally return -EBUSY from an open call, have you seen what happens
> > > with apps when you do this (like minicom?)
> > >
> > I tested this wil minincom, picocom and commands like "echo foo >
> > /dev/ttyS0". They all correctly report "Device or resource busy".
> >
> > I have addressed all the comments you made. I have also split the patch
> > into three. Following is summary of each.
>
> If the tty counts are being misreported then it would be better to fix
> the code to actually manage the counts properly. The core tty code is
> telling you that the tty is not in a valid state. While this is of
> itself a good API to have, the underlying reference miscounting ought
> IMHO to be fixed as well.
When opening from kernel, we don't use file pointer. The count mismatch
is between tty->count and #fd's. So opening from kernel leads to #fd's
being less than tty->count. I thought this difference is relevant to
user-space opening of tty, and not to kernel opening of tty. Can you
suggest how to address this mismatch?
>
> Also you don't need a new TTY_KOPENED flag as far as I can see. All tty's
> have a usage count and active bit flags already. Use those.
Ah may be I didn't notice the active bit. Is it one of the #defines in
tty.h? Can usage count and active bit be used to differentiate between
whether the tty was opened by kernel or user?
Thanks,
Okash
On Mon, Jul 10, 2017 at 01:33:07PM +0100, Okash Khawaja wrote:
> > If the tty counts are being misreported then it would be better to fix
> > the code to actually manage the counts properly. The core tty code is
> > telling you that the tty is not in a valid state. While this is of
> > itself a good API to have, the underlying reference miscounting ought
> > IMHO to be fixed as well.
> When opening from kernel, we don't use file pointer. The count mismatch
> is between tty->count and #fd's. So opening from kernel leads to #fd's
> being less than tty->count. I thought this difference is relevant to
> user-space opening of tty, and not to kernel opening of tty. Can you
> suggest how to address this mismatch?
Idea is tty_kopen only ever returns a newly initialised tty - returned
by tty_init_dev. Since the access is exclusive, tty->count shouldn't
matter. When the caller is done with it, it calls tty_release_struct
which frees up the tty itself.
But yes, all the while a tty is kopened, its tty->count and #fds will be
unequal.
Thanks,
Okash
> When opening from kernel, we don't use file pointer. The count mismatch
> is between tty->count and #fd's. So opening from kernel leads to #fd's
> being less than tty->count. I thought this difference is relevant to
> user-space opening of tty, and not to kernel opening of tty. Can you
> suggest how to address this mismatch?
Your kernel reference is the same as having a file open reference so I
think this actually needs addressing in the maths. In other words count
the number of kernel references and also add that into the test for
check_tty_count (kernel references + #fds == count).
I'd really like to keep this right because that check has a long history
of catching really nasty race conditions in the tty code. The
open/close/hangup code is really fragile so worth the debugability.
> Ah may be I didn't notice the active bit. Is it one of the #defines in
> tty.h? Can usage count and active bit be used to differentiate between
> whether the tty was opened by kernel or user?
It only tells you whether the port is currently active for some purpose,
not which. If you still want to implement exclusivity between kernel and
user then it needs another flag, but I think that flag should be in
port->flags as it is a property of the physical interface.
(Take a look at tty_port_open for example)
Alan
On Wed, Jul 12, 2017 at 07:20:28PM +0100, Alan Cox wrote:
>
> > When opening from kernel, we don't use file pointer. The count mismatch
> > is between tty->count and #fd's. So opening from kernel leads to #fd's
> > being less than tty->count. I thought this difference is relevant to
> > user-space opening of tty, and not to kernel opening of tty. Can you
> > suggest how to address this mismatch?
>
> Your kernel reference is the same as having a file open reference so I
> think this actually needs addressing in the maths. In other words count
> the number of kernel references and also add that into the test for
> check_tty_count (kernel references + #fds == count).
>
> I'd really like to keep this right because that check has a long history
> of catching really nasty race conditions in the tty code. The
> open/close/hangup code is really fragile so worth the debugability.
I see. Okay based this, check_tty_count can be easily updated to take
into account kernel references.
>
> > Ah may be I didn't notice the active bit. Is it one of the #defines in
> > tty.h? Can usage count and active bit be used to differentiate between
> > whether the tty was opened by kernel or user?
>
> It only tells you whether the port is currently active for some purpose,
> not which. If you still want to implement exclusivity between kernel and
> user then it needs another flag, but I think that flag should be in
> port->flags as it is a property of the physical interface.
>
> (Take a look at tty_port_open for example)
Okay I can add TTY_PORT_KOPENED to port->flags and that should work too.
However, can you please help me understand this:
Our use case requires kernel access to tty_struct and accordingly
tty_kopen returns tty_struct. The exclusivity between user and kernel
space is also meant to prevent one side from opening tty_struct while
another has it opened. In all this, it is tty_struct and not tty_port
which is the key resource we are concerned with. So shouldn't the
exclusivity flag belong to tty_struct?
Adding a the flag to port->flags but controlling it from code for
opening and closing tty will also mean we have tty_port_kopened,
tty_port_set_kopen etc inside tty open/close code.
Am I viewing this problem incorrectly?
Thanks,
Okash
On Thu, Jul 13, 2017 at 12:29:54PM +0100, Okash Khawaja wrote:
> On Wed, Jul 12, 2017 at 07:20:28PM +0100, Alan Cox wrote:
> >
> > > When opening from kernel, we don't use file pointer. The count mismatch
> > > is between tty->count and #fd's. So opening from kernel leads to #fd's
> > > being less than tty->count. I thought this difference is relevant to
> > > user-space opening of tty, and not to kernel opening of tty. Can you
> > > suggest how to address this mismatch?
> >
> > Your kernel reference is the same as having a file open reference so I
> > think this actually needs addressing in the maths. In other words count
> > the number of kernel references and also add that into the test for
> > check_tty_count (kernel references + #fds == count).
> >
> > I'd really like to keep this right because that check has a long history
> > of catching really nasty race conditions in the tty code. The
> > open/close/hangup code is really fragile so worth the debugability.
>
> I see. Okay based this, check_tty_count can be easily updated to take
> into account kernel references.
Ok, I'll drop this series from my "to-apply" queue and wait for you to
redo it.
thanks,
greg k-h
> On 17 Jul 2017, at 13:31, Greg Kroah-Hartman <[email protected]> wrote:
>
>> On Thu, Jul 13, 2017 at 12:29:54PM +0100, Okash Khawaja wrote:
>>> On Wed, Jul 12, 2017 at 07:20:28PM +0100, Alan Cox wrote:
>>>
>>>> When opening from kernel, we don't use file pointer. The count mismatch
>>>> is between tty->count and #fd's. So opening from kernel leads to #fd's
>>>> being less than tty->count. I thought this difference is relevant to
>>>> user-space opening of tty, and not to kernel opening of tty. Can you
>>>> suggest how to address this mismatch?
>>>
>>> Your kernel reference is the same as having a file open reference so I
>>> think this actually needs addressing in the maths. In other words count
>>> the number of kernel references and also add that into the test for
>>> check_tty_count (kernel references + #fds == count).
>>>
>>> I'd really like to keep this right because that check has a long history
>>> of catching really nasty race conditions in the tty code. The
>>> open/close/hangup code is really fragile so worth the debugability.
>>
>> I see. Okay based this, check_tty_count can be easily updated to take
>> into account kernel references.
>
> Ok, I'll drop this series from my "to-apply" queue and wait for you to
> redo it.
Sure. I can fix the tty->count mismatch based on Alan's suggestion. However I don't understand why the exclusivity flag should belong to tty_port and not tty_struct. It will be good to know why.
Thanks,
Okash
Hi,
I have reworked the previous patch set. These are the changes:
1. Patch 1 fixes tty->count mismatch reported by check_tty_count when a
tty is kopened.
2. Patch 1 incorporates patch 4 in the previous patch set - it returns
-ENODEV when tty is not configured.
Thanks,
Okash
> Sure. I can fix the tty->count mismatch based on Alan's suggestion. However I don't understand why the exclusivity flag should belong to tty_port and not tty_struct. It will be good to know why.
We are trying to move all the flags that we can and structs into the
tty_port, except any that are used internally within the struct tty
level code. The main reason for this is to make the object lifetimes and
locking simpler - because the tty_port lasts for the time the hardware is
present.
Alan
On Mon, Jul 17, 2017 at 11:04:38PM +0100, Alan Cox wrote:
>
> > Sure. I can fix the tty->count mismatch based on Alan's suggestion. However I don't understand why the exclusivity flag should belong to tty_port and not tty_struct. It will be good to know why.
>
> We are trying to move all the flags that we can and structs into the
> tty_port, except any that are used internally within the struct tty
> level code. The main reason for this is to make the object lifetimes and
> locking simpler - because the tty_port lasts for the time the hardware is
> present.
I see. That does make sense. I have appended a version of the patch which
adds the flag to tty_port and uses that inside tty_kopen.
>From readability point of view however, I think adding the flag to
tty_struct looks more intuitive. At least for now - until we move
other things from tty_struct to tty_port.
The patch is untested but I can work on this if that fits in with the
plans for tty.
Thanks,
Okash
---
drivers/tty/tty_io.c | 67 +++++++++++++++++++++++++++++++++++++++++++++++----
include/linux/tty.h | 17 ++++++++++++
2 files changed, 79 insertions(+), 5 deletions(-)
--- a/drivers/tty/tty_io.c
+++ b/drivers/tty/tty_io.c
@@ -280,7 +280,7 @@ static int check_tty_count(struct tty_st
{
#ifdef CHECK_TTY_COUNT
struct list_head *p;
- int count = 0;
+ int count = 0, kopen_count = 0;
spin_lock(&tty->files_lock);
list_for_each(p, &tty->tty_files) {
@@ -291,10 +291,12 @@ static int check_tty_count(struct tty_st
tty->driver->subtype == PTY_TYPE_SLAVE &&
tty->link && tty->link->count)
count++;
- if (tty->count != count) {
- tty_warn(tty, "%s: tty->count(%d) != #fd's(%d)\n",
- routine, tty->count, count);
- return count;
+ if (tty_port_kopened(tty->port))
+ kopen_count++;
+ if (tty->count != (count + kopen_count)) {
+ tty_warn(tty, "%s: tty->count(%d) != (#fd's(%d) + #kopen's(%d))\n",
+ routine, tty->count, count, kopen_count);
+ return (count + kopen_count);
}
#endif
return 0;
@@ -1522,6 +1524,7 @@ static int tty_release_checks(struct tty
*/
void tty_release_struct(struct tty_struct *tty, int idx)
{
+ // TODO: reset TTY_PORT_KOPENED on tty->port
/*
* Ask the line discipline code to release its structures
*/
@@ -1786,6 +1789,54 @@ static struct tty_driver *tty_lookup_dri
}
/**
+ * tty_kopen - open a tty device for kernel
+ * @device: dev_t of device to open
+ *
+ * Opens tty exclusively for kernel. Performs the driver lookup,
+ * makes sure it's not already opened and performs the first-time
+ * tty initialization.
+ *
+ * Returns the locked initialized &tty_struct
+ *
+ * Claims the global tty_mutex to serialize:
+ * - concurrent first-time tty initialization
+ * - concurrent tty driver removal w/ lookup
+ * - concurrent tty removal from driver table
+ */
+struct tty_struct *tty_kopen(dev_t device)
+{
+ struct tty_struct *tty;
+ struct tty_driver *driver = NULL;
+ int index = -1;
+
+ mutex_lock(&tty_mutex);
+ driver = tty_lookup_driver(device, NULL, &index);
+ if (IS_ERR(driver)) {
+ mutex_unlock(&tty_mutex);
+ return ERR_CAST(driver);
+ }
+
+ /* check whether we're reopening an existing tty */
+ tty = tty_driver_lookup_tty(driver, NULL, index);
+ if (IS_ERR(tty))
+ goto out;
+
+ if (tty) {
+ /* drop kref from tty_driver_lookup_tty() */
+ tty_kref_put(tty);
+ tty = ERR_PTR(-EBUSY);
+ } else { /* tty_init_dev returns tty with the tty_lock held */
+ tty = tty_init_dev(driver, index);
+ tty_port_set_kopened(tty->port, 1);
+ }
+out:
+ mutex_unlock(&tty_mutex);
+ tty_driver_kref_put(driver);
+ return tty;
+}
+EXPORT_SYMBOL_GPL(tty_kopen);
+
+/**
* tty_open_by_driver - open a tty device
* @device: dev_t of device to open
* @inode: inode of device file
@@ -1824,6 +1875,12 @@ struct tty_struct *tty_open_by_driver(de
}
if (tty) {
+ if (tty_port_kopened(tty->port)) {
+ tty_kref_put(tty);
+ mutex_unlock(&tty_mutex);
+ tty = ERR_PTR(-EBUSY);
+ goto out;
+ }
mutex_unlock(&tty_mutex);
retval = tty_lock_interruptible(tty);
tty_kref_put(tty); /* drop kref from tty_driver_lookup_tty() */
--- a/include/linux/tty.h
+++ b/include/linux/tty.h
@@ -261,6 +261,7 @@ struct tty_port {
*/
#define TTY_PORT_CTS_FLOW 3 /* h/w flow control enabled */
#define TTY_PORT_CHECK_CD 4 /* carrier detect enabled */
+#define TTY_PORT_KOPENED 5 /* device exclusively opened by kernel */
/*
* Where all of the state associated with a tty is kept while the tty
@@ -401,6 +402,7 @@ extern int __init tty_init(void);
extern const char *tty_name(const struct tty_struct *tty);
extern struct tty_struct *tty_open_by_driver(dev_t device, struct inode *inode,
struct file *filp);
+extern struct tty_struct *tty_kopen(dev_t device);
extern int tty_dev_name_to_number(const char *name, dev_t *number);
#else
static inline void tty_kref_put(struct tty_struct *tty)
@@ -425,6 +427,8 @@ static inline const char *tty_name(const
static inline struct tty_struct *tty_open_by_driver(dev_t device,
struct inode *inode, struct file *filp)
{ return NULL; }
+static inline struct tty_struct *tty_kopen(dev_t device)
+{ return NULL; }
static inline int tty_dev_name_to_number(const char *name, dev_t *number)
{ return -ENOTSUPP; }
#endif
@@ -652,6 +656,19 @@ static inline void tty_port_set_initiali
clear_bit(TTY_PORT_INITIALIZED, &port->iflags);
}
+static inline bool tty_port_kopened(struct tty_port *port)
+{
+ return test_bit(TTY_PORT_KOPENED, &port->iflags);
+}
+
+static inline void tty_port_set_kopened(struct tty_port *port, bool val)
+{
+ if (val)
+ set_bit(TTY_PORT_KOPENED, &port->iflags);
+ else
+ clear_bit(TTY_PORT_KOPENED, &port->iflags);
+}
+
extern struct tty_struct *tty_port_tty_get(struct tty_port *port);
extern void tty_port_tty_set(struct tty_port *port, struct tty_struct *tty);
extern int tty_port_carrier_raised(struct tty_port *port);
On Tue, Jul 18, 2017 at 12:29:52PM +0100, Okash Khawaja wrote:
> +struct tty_struct *tty_kopen(dev_t device)
> +{
> + struct tty_struct *tty;
> + struct tty_driver *driver = NULL;
> + int index = -1;
> +
> + mutex_lock(&tty_mutex);
> + driver = tty_lookup_driver(device, NULL, &index);
> + if (IS_ERR(driver)) {
> + mutex_unlock(&tty_mutex);
> + return ERR_CAST(driver);
> + }
> +
> + /* check whether we're reopening an existing tty */
> + tty = tty_driver_lookup_tty(driver, NULL, index);
> + if (IS_ERR(tty))
> + goto out;
> +
> + if (tty) {
> + /* drop kref from tty_driver_lookup_tty() */
> + tty_kref_put(tty);
> + tty = ERR_PTR(-EBUSY);
> + } else { /* tty_init_dev returns tty with the tty_lock held */
> + tty = tty_init_dev(driver, index);
> + tty_port_set_kopened(tty->port, 1);
^^^^^^^^^
tty_init_dev() can fail leading to an error pointer dereference here.
> + }
> +out:
> + mutex_unlock(&tty_mutex);
> + tty_driver_kref_put(driver);
> + return tty;
> +}
> +EXPORT_SYMBOL_GPL(tty_kopen);
regards,
dan carpenter
> The patch is untested but I can work on this if that fits in with the
> plans for tty.
I think this is the right direction.
Alan
On Tue, Jul 18, 2017 at 03:26:37PM +0300, Dan Carpenter wrote:
> > + if (tty) {
> > + /* drop kref from tty_driver_lookup_tty() */
> > + tty_kref_put(tty);
> > + tty = ERR_PTR(-EBUSY);
> > + } else { /* tty_init_dev returns tty with the tty_lock held */
> > + tty = tty_init_dev(driver, index);
> > + tty_port_set_kopened(tty->port, 1);
> ^^^^^^^^^
>
> tty_init_dev() can fail leading to an error pointer dereference here.
Thanks very much. I will check for error pointer here.
Okash
Hi,
I have updated the patches so that the exclusivity flag is in tty_port.
When closing the struct - by calling tty_release_struct - we also need
to reset the flag. One way to do that is to reset the flag inside
tty_release_struct function, regardless of whether the tty was opened
through tty_kopen or not. In order to keep the code clean, I have
instead created a separate function called tty_kclose which is the same
as tty_release_struct except that it also resets the exclusivity flag.
As a result, any changes to tty_release_struct won't be in tty_kclose
untless manually added. Please advise on this.
Here is a summary of changes when compared to v2:
- Patch 1 uses TTY_PORT_KOPENED flag instead of TTY_KOPENED and as a
result adds helper functions to read and change the flag
- Patch 1 adds tty_kclose function to close tty opened by tty_kopen
- Patch 2 calls tty_kclose instead of tty_release_struct
Thanks,
Okash