This fixes the following problems:
1) Typematic-repeat of 'enter' gives warning message.
2) Use of 'keypad enter' gives warning message.
3) Lag on the order of seconds between "break" and "make" when
expecting the enter "break" code. Seen under virtualized
environments such as VMware ESX.
Explanations:
1) Holding down 'enter' will not set a repeating sequence
of 0x1c(make)-0x9c(make), but a repeating sequence
of make codes, followed by one break code when the key
is released. Thus, it's wrong to expect the break code
after seeing the 'enter' make code.
2) Keypad enter generates different make/break, namely
0xe0 0x1c and 0xe0 0x9c. The 'generic' logic handles
the 0xe0 escape already, but the special 'enter' logic
always expects '0x9c' and not '0xe0 0x9c', so you get
a warning message, again.
3) When expecting the 'enter' break code, the code polls
the status register in a tight loop, like so -
> while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0);
However, it really should do something like -
> while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
> cpu_relax(); /* pause */
Basically, it's a common optimization to have a fast
path for accessing often accessed and slow changing I/O
in a virtualized environment. The tight spinning in KDB
seems to run against the logic by ESX keyboard virtualization
code to detect when the fast path or the slow path should
be used to satisfy the keyboard status read, leading to
multi-second timeouts before the 'real' status comes through.
Without knowing ESX internals, it's hard to say if this is
an ESX bug or not, but letting the VM be explicitely descheduled
seems to resolve the problem. I've seen something similar with
shared MMIO buffers with VMs on Hyper-V.
Anyway, given (3), (2) and (1), we might as well blow away the
entire special casing for 'enter'. The break codes will already
be handled correctly, and we get rid of the bugs with repeat
enters and keypad enter key. And of course, there is no
need to AND with 0x7f when checking for 'enter', because we'll
never ever get to this code with a break code (checked for much
earlier).
I tried to figure out the history behind the 'enter' key special
casing code, and it seems to have come from whatever the original
KDB patch was. Perhaps someone can chime in.
Tested on ESX 5.0 and QEMU.
Signed-off-by: Andrei Warkentin <[email protected]>
---
kernel/debug/kdb/kdb_keyboard.c | 28 +---------------------------
1 files changed, 1 insertions(+), 27 deletions(-)
diff --git a/kernel/debug/kdb/kdb_keyboard.c b/kernel/debug/kdb/kdb_keyboard.c
index 4bca634..ed4a2f9 100644
--- a/kernel/debug/kdb/kdb_keyboard.c
+++ b/kernel/debug/kdb/kdb_keyboard.c
@@ -178,34 +178,8 @@ int kdb_get_kbd_char(void)
return -1; /* ignore unprintables */
}
- if ((scancode & 0x7f) == 0x1c) {
- /*
- * enter key. All done. Absorb the release scancode.
- */
- while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
- ;
-
- /*
- * Fetch the scancode
- */
- scancode = inb(KBD_DATA_REG);
- scanstatus = inb(KBD_STATUS_REG);
-
- while (scanstatus & KBD_STAT_MOUSE_OBF) {
- scancode = inb(KBD_DATA_REG);
- scanstatus = inb(KBD_STATUS_REG);
- }
-
- if (scancode != 0x9c) {
- /*
- * Wasn't an enter-release, why not?
- */
- kdb_printf("kdb: expected enter got 0x%x status 0x%x\n",
- scancode, scanstatus);
- }
-
+ if (scancode == 0x1c)
return 13;
- }
return keychar & 0xff;
}
--
1.7.4.1
Hi,
----- Original Message -----
> From: "Andrei Warkentin" <[email protected]>
> To: [email protected]
> Cc: [email protected], "jason wessel" <[email protected]>, [email protected]
> Sent: Friday, February 17, 2012 6:52:53 PM
> Subject: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
> This fixes the following problems:
> 1) Typematic-repeat of 'enter' gives warning message.
> 2) Use of 'keypad enter' gives warning message.
> 3) Lag on the order of seconds between "break" and "make" when
> expecting the enter "break" code. Seen under virtualized
> environments such as VMware ESX.
>
Does this seem reasonable :-)?
A
On 02/17/2012 05:52 PM, Andrei Warkentin wrote:
> This fixes the following problems:
> 1) Typematic-repeat of 'enter' gives warning message.
> 2) Use of 'keypad enter' gives warning message.
> 3) Lag on the order of seconds between "break" and "make" when
> expecting the enter "break" code. Seen under virtualized
> environments such as VMware ESX.
>
> Explanations:
> 1) Holding down 'enter' will not set a repeating sequence
> of 0x1c(make)-0x9c(make), but a repeating sequence
> of make codes, followed by one break code when the key
> is released. Thus, it's wrong to expect the break code
> after seeing the 'enter' make code.
> 2) Keypad enter generates different make/break, namely
> 0xe0 0x1c and 0xe0 0x9c. The 'generic' logic handles
> the 0xe0 escape already, but the special 'enter' logic
> always expects '0x9c' and not '0xe0 0x9c', so you get
> a warning message, again.
> 3) When expecting the 'enter' break code, the code polls
> the status register in a tight loop, like so -
> > while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0);
>
> However, it really should do something like -
> > while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
> > cpu_relax(); /* pause */
>
> Basically, it's a common optimization to have a fast
> path for accessing often accessed and slow changing I/O
> in a virtualized environment. The tight spinning in KDB
> seems to run against the logic by ESX keyboard virtualization
> code to detect when the fast path or the slow path should
> be used to satisfy the keyboard status read, leading to
> multi-second timeouts before the 'real' status comes through.
> Without knowing ESX internals, it's hard to say if this is
> an ESX bug or not, but letting the VM be explicitely descheduled
> seems to resolve the problem. I've seen something similar with
> shared MMIO buffers with VMs on Hyper-V.
>
> Anyway, given (3), (2) and (1), we might as well blow away the
> entire special casing for 'enter'. The break codes will already
> be handled correctly, and we get rid of the bugs with repeat
> enters and keypad enter key. And of course, there is no
> need to AND with 0x7f when checking for 'enter', because we'll
> never ever get to this code with a break code (checked for much
> earlier).
>
> I tried to figure out the history behind the 'enter' key special
> casing code, and it seems to have come from whatever the original
> KDB patch was. Perhaps someone can chime in.
>
I did not write the original code, but I can explain why there was some special logic.
When you restore the system back to the running state you do not want to send any other key scan codes back to the kernel. The idea being that you type "go" and press enter to resume kernel execution. At that point you do not want to send a random scan code back to the kernel, ideally you want to leave everything as it was. This also handled the case where there was a PS/2 style mouse attached.
I do have a question about part of the section you deleted.
> Tested on ESX 5.0 and QEMU.
>
> Signed-off-by: Andrei Warkentin <[email protected]>
> ---
> kernel/debug/kdb/kdb_keyboard.c | 28 +---------------------------
> 1 files changed, 1 insertions(+), 27 deletions(-)
>
> diff --git a/kernel/debug/kdb/kdb_keyboard.c b/kernel/debug/kdb/kdb_keyboard.c
> index 4bca634..ed4a2f9 100644
> --- a/kernel/debug/kdb/kdb_keyboard.c
> +++ b/kernel/debug/kdb/kdb_keyboard.c
> @@ -178,34 +178,8 @@ int kdb_get_kbd_char(void)
> return -1; /* ignore unprintables */
> }
>
> - if ((scancode & 0x7f) == 0x1c) {
> - /*
> - * enter key. All done. Absorb the release scancode.
> - */
> - while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
> - ;
Seems there is a bug here. There is a cpu_relax() missing.
> -
> - /*
> - * Fetch the scancode
> - */
> - scancode = inb(KBD_DATA_REG);
> - scanstatus = inb(KBD_STATUS_REG);
> -
> - while (scanstatus & KBD_STAT_MOUSE_OBF) {
There should also be a cpu_relax() right here.
> - scancode = inb(KBD_DATA_REG);
> - scanstatus = inb(KBD_STATUS_REG);
> - }
> -
If you put the two cpu_relax() pieces in do you still end up with a problem? If this does not work for you the possibility to exists to clear the keyboard/mouse state on the kdb exit.
Thanks,
Jason.
On 02/26/2012 07:10 AM, Jason Wessel wrote:
> On 02/17/2012 05:52 PM, Andrei Warkentin wrote:
>> This fixes the following problems:
>> 1) Typematic-repeat of 'enter' gives warning message.
>> 2) Use of 'keypad enter' gives warning message.
>> 3) Lag on the order of seconds between "break" and "make" when
>> expecting the enter "break" code. Seen under virtualized
>> environments such as VMware ESX.
>>
>> Explanations:
>> 1) Holding down 'enter' will not set a repeating sequence
>> of 0x1c(make)-0x9c(make), but a repeating sequence
>> of make codes, followed by one break code when the key
>> is released. Thus, it's wrong to expect the break code
>> after seeing the 'enter' make code.
>> 2) Keypad enter generates different make/break, namely
>> 0xe0 0x1c and 0xe0 0x9c. The 'generic' logic handles
>> the 0xe0 escape already, but the special 'enter' logic
>> always expects '0x9c' and not '0xe0 0x9c', so you get
>> a warning message, again.
>> 3) When expecting the 'enter' break code, the code polls
>> the status register in a tight loop, like so -
>> > while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0);
>>
>> However, it really should do something like -
>> > while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
>> > cpu_relax(); /* pause */
>>
>> Basically, it's a common optimization to have a fast
>> path for accessing often accessed and slow changing I/O
>> in a virtualized environment. The tight spinning in KDB
>> seems to run against the logic by ESX keyboard virtualization
>> code to detect when the fast path or the slow path should
>> be used to satisfy the keyboard status read, leading to
>> multi-second timeouts before the 'real' status comes through.
>> Without knowing ESX internals, it's hard to say if this is
>> an ESX bug or not, but letting the VM be explicitely descheduled
>> seems to resolve the problem. I've seen something similar with
>> shared MMIO buffers with VMs on Hyper-V.
>>
>> Anyway, given (3), (2) and (1), we might as well blow away the
>> entire special casing for 'enter'. The break codes will already
>> be handled correctly, and we get rid of the bugs with repeat
>> enters and keypad enter key. And of course, there is no
>> need to AND with 0x7f when checking for 'enter', because we'll
>> never ever get to this code with a break code (checked for much
>> earlier).
>>
>> I tried to figure out the history behind the 'enter' key special
>> casing code, and it seems to have come from whatever the original
>> KDB patch was. Perhaps someone can chime in.
>>
>
Andrei, if you agree with the attached patch, I'll put it in the merge queue. If you find problems we can go another iteration. :-)
Cheers,
Jason.
Hi Jason,
Thank you for the review. Comments inline.
----- Original Message -----
> From: "Jason Wessel" <[email protected]>
> To: "Andrei Warkentin" <[email protected]>
> Cc: [email protected], [email protected]
> Sent: Sunday, February 26, 2012 8:10:56 AM
> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
> I did not write the original code, but I can explain why there was
> some special logic.
>
> When you restore the system back to the running state you do not want
> to send any other key scan codes back to the kernel. The idea being
> that you type "go" and press enter to resume kernel execution. At
> that point you do not want to send a random scan code back to the
> kernel, ideally you want to leave everything as it was. This also
> handled the case where there was a PS/2 style mouse attached.
The input device is already reset on kdb exit to a safe state, however, by the
code inside kgdboc_post_exp_handler, which calls kgdboc_restore_input().
The existing logic also already correctly will handle the case of a PS/2 mouse, i.e.:
/*
* Ignore mouse events.
*/
if (scanstatus & KBD_STAT_MOUSE_OBF)
return -1;
...so there is nothing that the non-ENTER-specific code already doesn't deal with.
That special ENTER code also doesn't correctly handle typematic repeat, nor does it handle
key-pad enter (as I mentioned before in the patch description, this has a different
make and break codes, prefixed with 0xe0).
>
> I do have a question about part of the section you deleted.
>
>
> > - if ((scancode & 0x7f) == 0x1c) {
> > - /*
> > - * enter key. All done. Absorb the release scancode.
> > - */
> > - while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
> > - ;
>
> Seems there is a bug here. There is a cpu_relax() missing.
Right, that's exactly what I said in the patch :-).
>
> > -
> > - /*
> > - * Fetch the scancode
> > - */
> > - scancode = inb(KBD_DATA_REG);
> > - scanstatus = inb(KBD_STATUS_REG);
> > -
> > - while (scanstatus & KBD_STAT_MOUSE_OBF) {
>
> There should also be a cpu_relax() right here.
Correct.
>
> > - scancode = inb(KBD_DATA_REG);
> > - scanstatus = inb(KBD_STATUS_REG);
> > - }
> > -
>
>
> If you put the two cpu_relax() pieces in do you still end up with a
> problem? If this does not work for you the possibility to exists to
> clear the keyboard/mouse state on the kdb exit.
You will solve the multi-second hangs on ESX. However, this still doesn't fix the
bogus messages during typematic repeat (holding down ENTER) or keypad enter key use.
Basically, if you nuke the entire ENTER-key special case block, the existing key will
correctly handle both typematic, break codes, and KP ENTER key use (as well as any
mouse events). Cleanup of mouse/keyboard state is done already on kdb exit by kgdboc, which
is the only consumer of the kdb_keyboard code.
A
Hi,
----- Original Message -----
> From: "Jason Wessel" <[email protected]>
> To: "Andrei Warkentin" <[email protected]>
> Cc: [email protected], [email protected]
> Sent: Sunday, February 26, 2012 8:58:37 AM
> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
> Andrei, if you agree with the attached patch, I'll put it in the
> merge queue. If you find problems we can go another iteration. :-)
>
Nak. As I've mentioned in my original email, I went down this route as well, only
to ask the question of why the special ENTER code was there, since all it did
was introduce problems elsewhere (like repeat ENTER handling, KP ENTER key...).
This is exactly why my solution was to get rid of the special ENTER code altogether.
The code in kgdboc already will do input device reset on kdb exit, so there is no need
to explicitely empty out the keyboard FIFO on every ENTER.
Thanks,
A
On 02/26/2012 07:04 PM, Andrei Warkentin wrote:
> Hi,
>
> ----- Original Message -----
>> From: "Jason Wessel" <[email protected]>
>> To: "Andrei Warkentin" <[email protected]>
>> Cc: [email protected], [email protected]
>> Sent: Sunday, February 26, 2012 8:58:37 AM
>> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
>>
>> Andrei, if you agree with the attached patch, I'll put it in the
>> merge queue. If you find problems we can go another iteration. :-)
>>
>
> Nak. As I've mentioned in my original email, I went down this route as well, only
> to ask the question of why the special ENTER code was there, since all it did
> was introduce problems elsewhere (like repeat ENTER handling, KP ENTER key...).
> This is exactly why my solution was to get rid of the special ENTER code altogether.
>
> The code in kgdboc already will do input device reset on kdb exit, so there is no need
> to explicitely empty out the keyboard FIFO on every ENTER.
>
Did you try the patch I sent? It might not address the key repeat
problem, but this is some thing I had not yet duplicated. Outside of
the warning message (which was killed off in the patch), was there
garbage characters or some notable behavior? Was it something you
could see in qemu or only in ESX?
Having authored the keyboard/input handler with some suggestions from
Dmitry, I was fairly certain it had no way to prevent the leak of the
up keystroke unless you have the KDB specific capture where it waits
to act on the "key up" event for an enter keystroke. The cleanup
handler only cleans up state events where keys were down at the time
the debug core became active. It will not prevent the leak of key
strokes or state changes while the kernel was stopped. The goofy
enter handler was supposed to take care of that.
I conclusively proved there is event leakage from using the original
patch you provided. Here is the test case, which you should be able to
execute with qemu or kvm (since you also mentioned that was what you
were previously using).
1) boot the qemu
2) Use kgdboc=kbd
3) break into the debugger to the kdb prompt
4) type "g" but to not press return
5) Connect to the qemu back end debug connection with gdb and set a
breakpoint at line 402 of atkbd.c which should be the call to:
input_event(dev, EV_MSC, MSC_RAW, code);
6) continue the gdb connection
7) In the qemu monitor enter the command:
sendkey ret 4000
After 4 seconds when the key is released you will catch the leaked
event in the atkbd.c, and if you had X running it propagates all the
way up the input chain.
Jason.
Hi,
----- Original Message -----
> From: "Jason Wessel" <[email protected]>
> To: "Andrei Warkentin" <[email protected]>
> Cc: [email protected], [email protected], "Andrei Warkentin" <[email protected]>
> Sent: Monday, February 27, 2012 5:50:59 PM
> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
> Did you try the patch I sent? It might not address the key repeat
> problem, but this is some thing I had not yet duplicated. Outside of
> the warning message (which was killed off in the patch), was there
> garbage characters or some notable behavior? Was it something you
> could see in qemu or only in ESX?
Let me elaborate. The cpu_relax() you added solves the hanging.
>
> Having authored the keyboard/input handler with some suggestions from
> Dmitry, I was fairly certain it had no way to prevent the leak of the
> up keystroke unless you have the KDB specific capture where it waits
> to act on the "key up" event for an enter keystroke. The cleanup
> handler only cleans up state events where keys were down at the time
> the debug core became active. It will not prevent the leak of key
> strokes or state changes while the kernel was stopped. The goofy
> enter handler was supposed to take care of that.
>
> I conclusively proved there is event leakage from using the original
> patch you provided. Here is the test case, which you should be able
> to
> execute with qemu or kvm (since you also mentioned that was what you
> were previously using).
>
> 1) boot the qemu
> 2) Use kgdboc=kbd
> 3) break into the debugger to the kdb prompt
> 4) type "g" but to not press return
> 5) Connect to the qemu back end debug connection with gdb and set a
> breakpoint at line 402 of atkbd.c which should be the call to:
> input_event(dev, EV_MSC, MSC_RAW, code);
> 6) continue the gdb connection
> 7) In the qemu monitor enter the command:
> sendkey ret 4000
>
> After 4 seconds when the key is released you will catch the leaked
> event in the atkbd.c, and if you had X running it propagates all the
> way up the input chain.
I know I should have read deeper into the input code. I assumed it performed a reset
on the input device. Assumptions are bad. Sorry.
So here is the thing. With the patch you sent, you would still leak the break
keypress for the keypad enter. This is because the break code for KP ENTER is 0xe0 0x9c,
and the inb(KBD_DATA_REG) done in the goofy handler will read just the 0xe0 part, leaving
0x9c behind, which will be leaked out.
So here is my question - do you want me to fix the existing goofy handler to properly
handle KP ENTER and ENTER repeats? Or do you think it's appropriate to empty out the
FIFO during KDB exit as part of post_exp kgdb_io op? Or reset the input device?
Thanks for investigating this.
A
On 02/27/2012 05:18 PM, Andrei Warkentin wrote:
> Hi,
>
> ----- Original Message -----
>> From: "Jason Wessel" <[email protected]>
>> To: "Andrei Warkentin" <[email protected]>
>> Cc: [email protected], [email protected], "Andrei Warkentin" <[email protected]>
>> Sent: Monday, February 27, 2012 5:50:59 PM
>> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>>
>> Did you try the patch I sent? It might not address the key repeat
>> problem, but this is some thing I had not yet duplicated. Outside of
>> the warning message (which was killed off in the patch), was there
>> garbage characters or some notable behavior? Was it something you
>> could see in qemu or only in ESX?
> Let me elaborate. The cpu_relax() you added solves the hanging.
>
>> Having authored the keyboard/input handler with some suggestions from
>> Dmitry, I was fairly certain it had no way to prevent the leak of the
>> up keystroke unless you have the KDB specific capture where it waits
>> to act on the "key up" event for an enter keystroke. The cleanup
>> handler only cleans up state events where keys were down at the time
>> the debug core became active. It will not prevent the leak of key
>> strokes or state changes while the kernel was stopped. The goofy
>> enter handler was supposed to take care of that.
>>
>> I conclusively proved there is event leakage from using the original
>> patch you provided. Here is the test case, which you should be able
>> to
>> execute with qemu or kvm (since you also mentioned that was what you
>> were previously using).
>>
>> 1) boot the qemu
>> 2) Use kgdboc=kbd
>> 3) break into the debugger to the kdb prompt
>> 4) type "g" but to not press return
>> 5) Connect to the qemu back end debug connection with gdb and set a
>> breakpoint at line 402 of atkbd.c which should be the call to:
>> input_event(dev, EV_MSC, MSC_RAW, code);
>> 6) continue the gdb connection
>> 7) In the qemu monitor enter the command:
>> sendkey ret 4000
>>
>> After 4 seconds when the key is released you will catch the leaked
>> event in the atkbd.c, and if you had X running it propagates all the
>> way up the input chain.
> I know I should have read deeper into the input code. I assumed it performed a reset
> on the input device. Assumptions are bad. Sorry.
>
> So here is the thing. With the patch you sent, you would still leak the break
> keypress for the keypad enter. This is because the break code for KP ENTER is 0xe0 0x9c,
> and the inb(KBD_DATA_REG) done in the goofy handler will read just the 0xe0 part, leaving
> 0x9c behind, which will be leaked out.
>
> So here is my question - do you want me to fix the existing goofy handler to properly
> handle KP ENTER and ENTER repeats? Or do you think it's appropriate to empty out the
> FIFO during KDB exit as part of post_exp kgdb_io op? Or reset the input device?
Thanks for the explanation. I agree that we still have a leak with the patch I had sent you. The old kdb code simply was incomplete and likely assumed people didn't ever use anything except the main keyboard's enter key.
I think the right choice is to correctly handle any that generates enter/return and do not leave the kdb keyboard handler until the release scan code is processed. We have to do this because the physical release delay will always generate another scan code later on regardless of clearing the fifo or resetting the device. The computer is simply faster than the human behind the console in this case. :-)
If you have some idea how to correctly fix up the KP ENTER please send a patch, else I will have a look at it sometime soon. It is definitely worth fixing.
Cheers,
Jason.
Hi,
----- Original Message -----
> From: "Jason Wessel" <[email protected]>
> To: "Andrei Warkentin" <[email protected]>
> Cc: [email protected], [email protected], "Andrei Warkentin" <[email protected]>
> Sent: Monday, February 27, 2012 6:28:20 PM
> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
> If you have some idea how to correctly fix up the KP ENTER please
> send a patch, else I will have a look at it sometime soon. It is
> definitely worth fixing.
>
Will send an updated patch. Thanks again.
A
This fixes the following problems:
1) Typematic-repeat of 'enter' gives warning message
and leaks make/break if KDB exits. Repeats
look something like 0x1c 0x1c .... 0x9c
2) Use of 'keypad enter' gives warning message and
leaks the ENTER break/make code out if KDB exits.
KP ENTER repeats look someting like 0xe0 0x1c
0xe0 0x1c ... 0xe0 0x9c.
3) Lag on the order of seconds between "break" and "make" when
expecting the enter "break" code. Seen under virtualized
environments such as VMware ESX.
The existing special enter handler tries to glob the
enter break code, but this fails if the other (KP) enter
was used, or if there was a key repeat. It also fails
if you mashed some keys along with enter, and you ended
up with a non-enter make or non-enter break code coming
after the enter make code. So first, we modify the handler
to handle these cases. But performing these actions on
every enter is annoying since now you can't hold ENTER down
to scroll <more>d messages in KDB. Since this special
behaviour is only necessary to handle the exiting KDB
('g' + ENTER) without leaking scancodes to the OS, we
perform the cleanup in kgdboc_post_exp_handler path.
Tested on QEMU. Set a bp on atkbd.c to verify no scan
code was leaked.
Cc: Andrei Warkentin <[email protected]>
Cc: Jason Wessel <[email protected]>
Signed-off-by: Andrei Warkentin <[email protected]>
---
drivers/tty/serial/kgdboc.c | 15 +++++++
include/linux/kdb.h | 1 +
kernel/debug/kdb/kdb_keyboard.c | 87 +++++++++++++++++++++++++++++---------
3 files changed, 82 insertions(+), 21 deletions(-)
diff --git a/drivers/tty/serial/kgdboc.c b/drivers/tty/serial/kgdboc.c
index 2b42a01..c74b47b 100644
--- a/drivers/tty/serial/kgdboc.c
+++ b/drivers/tty/serial/kgdboc.c
@@ -284,6 +284,8 @@ static void kgdboc_pre_exp_handler(void)
static void kgdboc_post_exp_handler(void)
{
+ int i;
+
/* decrement the module count when the debugger detaches */
if (!kgdb_connected)
module_put(THIS_MODULE);
@@ -291,6 +293,19 @@ static void kgdboc_post_exp_handler(void)
dbg_restore_graphics = 0;
con_debug_leave();
}
+
+ /*
+ * For i8042 keyboard, we still have the ENTER
+ * break code in the buffer after KDB processed
+ * the 'g' command and exited. Clean it up.
+ */
+ for (i = 0; i < kdb_poll_idx; i++) {
+ if (kdb_poll_funcs[i] == kdb_get_kbd_char) {
+ kdb_kbd_cleanup_state();
+ break;
+ }
+ }
+
kgdboc_restore_input();
}
diff --git a/include/linux/kdb.h b/include/linux/kdb.h
index 0647258..9bf65cc 100644
--- a/include/linux/kdb.h
+++ b/include/linux/kdb.h
@@ -124,6 +124,7 @@ extern void kdb_init(int level);
typedef int (*get_char_func)(void);
extern get_char_func kdb_poll_funcs[];
extern int kdb_get_kbd_char(void);
+extern void kdb_kbd_cleanup_state(void);
static inline
int kdb_process_cpu(const struct task_struct *p)
diff --git a/kernel/debug/kdb/kdb_keyboard.c b/kernel/debug/kdb/kdb_keyboard.c
index 4bca634..2671f5d 100644
--- a/kernel/debug/kdb/kdb_keyboard.c
+++ b/kernel/debug/kdb/kdb_keyboard.c
@@ -178,35 +178,80 @@ int kdb_get_kbd_char(void)
return -1; /* ignore unprintables */
}
- if ((scancode & 0x7f) == 0x1c) {
- /*
- * enter key. All done. Absorb the release scancode.
- */
+ if (scancode == 0x1c)
+ return 13;
+
+ return keychar & 0xff;
+}
+EXPORT_SYMBOL_GPL(kdb_get_kbd_char);
+
+/*
+ * Best effort cleanup of ENTER break codes
+ * on leaving KDB. Called on exiting KDB,
+ * when we know we processed an ENTER or
+ * KP ENTER scan code.
+ */
+void kdb_kbd_cleanup_state(void)
+{
+ int scancode, scanstatus;
+ /*
+ * Enter key. Need to absorb the break
+ * code here, lest it gets leaked out
+ * if we exit KDB as the result of
+ * processing 'g'.
+ *
+ * This has several interesting implications:
+ * + Need to handle KP ENTER, which has break
+ * code 0xe0 0x9c.
+ * + Need to handle repeat ENTER and repeat
+ * KP ENTER. Repeats only get a break code
+ * at the end of the repeated sequence. This
+ * means we can't propagate the repeated key
+ * press, and must swallow it away.
+ * + Need to handle possible PS/2 mouse input.
+ * + Need to handle mashed keys.
+ */
+
+ while (1) {
while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
- ;
+ cpu_relax();
/*
- * Fetch the scancode
+ * Fetch the scancode.
*/
scancode = inb(KBD_DATA_REG);
scanstatus = inb(KBD_STATUS_REG);
- while (scanstatus & KBD_STAT_MOUSE_OBF) {
- scancode = inb(KBD_DATA_REG);
- scanstatus = inb(KBD_STATUS_REG);
- }
+ /*
+ * Skip mouse input.
+ */
+ if (scanstatus & KBD_STAT_MOUSE_OBF)
+ continue;
- if (scancode != 0x9c) {
- /*
- * Wasn't an enter-release, why not?
- */
- kdb_printf("kdb: expected enter got 0x%x status 0x%x\n",
- scancode, scanstatus);
- }
+ /*
+ * If we see 0xe0, this is either
+ * a break code for KP ENTER, or a repeat
+ * make for KP ENTER. Either way, since
+ * the second byte is equivalent to an ENTER,
+ * skip the 0xe0 and try again.
+ *
+ * If we see 0x1c, this must be a repeat
+ * ENTER or KP ENTER (and we swallowed 0xe0
+ * before). Try again.
+ *
+ * We can also see make and break codes
+ * for other keys mashed before or after
+ * pressing ENTER. Thus, if we see anything
+ * other than 0x9c, we have to try again.
+ *
+ * Note, if you held some key as ENTER
+ * was depressed, that break code would
+ * get leaked out.
+ */
+ if (scancode != 0x9c)
+ continue;
- return 13;
+ return;
}
-
- return keychar & 0xff;
}
-EXPORT_SYMBOL_GPL(kdb_get_kbd_char);
+EXPORT_SYMBOL_GPL(kdb_kbd_cleanup_state);
--
1.7.8.3
This fixes the following problems:
1) Typematic-repeat of 'enter' gives warning message
and leaks make/break if KDB exits. Repeats
look something like 0x1c 0x1c .... 0x9c
2) Use of 'keypad enter' gives warning message and
leaks the ENTER break/make code out if KDB exits.
KP ENTER repeats look someting like 0xe0 0x1c
0xe0 0x1c ... 0xe0 0x9c.
3) Lag on the order of seconds between "break" and "make" when
expecting the enter "break" code. Seen under virtualized
environments such as VMware ESX.
The existing special enter handler tries to glob the
enter break code, but this fails if the other (KP) enter
was used, or if there was a key repeat. It also fails
if you mashed some keys along with enter, and you ended
up with a non-enter make or non-enter break code coming
after the enter make code. So first, we modify the handler
to handle these cases. But performing these actions on
every enter is annoying since now you can't hold ENTER down
to scroll <more>d messages in KDB. Since this special
behaviour is only necessary to handle the exiting KDB
('g' + ENTER) without leaking scancodes to the OS, we
perform the cleanup in kgdboc_post_exp_handler path.
Fixed previous regression where if kbd was not used
to 'g' + ENTER, the cleanup code would hang.
Tested on QEMU. Set a bp on atkbd.c to verify no scan
code was leaked.
Cc: Andrei Warkentin <[email protected]>
Cc: Jason Wessel <[email protected]>
Signed-off-by: Andrei Warkentin <[email protected]>
---
drivers/tty/serial/kgdboc.c | 15 ++++++
include/linux/kdb.h | 1 +
kernel/debug/kdb/kdb_keyboard.c | 104 ++++++++++++++++++++++++++++++---------
3 files changed, 98 insertions(+), 22 deletions(-)
diff --git a/drivers/tty/serial/kgdboc.c b/drivers/tty/serial/kgdboc.c
index 2b42a01..c74b47b 100644
--- a/drivers/tty/serial/kgdboc.c
+++ b/drivers/tty/serial/kgdboc.c
@@ -284,6 +284,8 @@ static void kgdboc_pre_exp_handler(void)
static void kgdboc_post_exp_handler(void)
{
+ int i;
+
/* decrement the module count when the debugger detaches */
if (!kgdb_connected)
module_put(THIS_MODULE);
@@ -291,6 +293,19 @@ static void kgdboc_post_exp_handler(void)
dbg_restore_graphics = 0;
con_debug_leave();
}
+
+ /*
+ * For i8042 keyboard, we still have the ENTER
+ * break code in the buffer after KDB processed
+ * the 'g' command and exited. Clean it up.
+ */
+ for (i = 0; i < kdb_poll_idx; i++) {
+ if (kdb_poll_funcs[i] == kdb_get_kbd_char) {
+ kdb_kbd_cleanup_state();
+ break;
+ }
+ }
+
kgdboc_restore_input();
}
diff --git a/include/linux/kdb.h b/include/linux/kdb.h
index 0647258..9bf65cc 100644
--- a/include/linux/kdb.h
+++ b/include/linux/kdb.h
@@ -124,6 +124,7 @@ extern void kdb_init(int level);
typedef int (*get_char_func)(void);
extern get_char_func kdb_poll_funcs[];
extern int kdb_get_kbd_char(void);
+extern void kdb_kbd_cleanup_state(void);
static inline
int kdb_process_cpu(const struct task_struct *p)
diff --git a/kernel/debug/kdb/kdb_keyboard.c b/kernel/debug/kdb/kdb_keyboard.c
index 4bca634..34a8722 100644
--- a/kernel/debug/kdb/kdb_keyboard.c
+++ b/kernel/debug/kdb/kdb_keyboard.c
@@ -25,6 +25,7 @@
#define KBD_STAT_MOUSE_OBF 0x20 /* Mouse output buffer full */
static int kbd_exists;
+static int kbd_last_ret;
/*
* Check if the keyboard controller has a keypress for us.
@@ -90,8 +91,11 @@ int kdb_get_kbd_char(void)
return -1;
}
- if ((scancode & 0x80) != 0)
+ if ((scancode & 0x80) != 0) {
+ if (scancode == 0x9c)
+ kbd_last_ret = 0;
return -1;
+ }
scancode &= 0x7f;
@@ -178,35 +182,91 @@ int kdb_get_kbd_char(void)
return -1; /* ignore unprintables */
}
- if ((scancode & 0x7f) == 0x1c) {
- /*
- * enter key. All done. Absorb the release scancode.
- */
+ if (scancode == 0x1c) {
+ kbd_last_ret = 1;
+ return 13;
+ }
+
+ return keychar & 0xff;
+}
+EXPORT_SYMBOL_GPL(kdb_get_kbd_char);
+
+/*
+ * Best effort cleanup of ENTER break codes
+ * on leaving KDB. Called on exiting KDB,
+ * when we know we processed an ENTER or
+ * KP ENTER scan code.
+ */
+void kdb_kbd_cleanup_state(void)
+{
+ int scancode, scanstatus;
+
+ /*
+ * Nothing to clean up, since either
+ * ENTER was never pressed, or has already
+ * gotten cleaned up.
+ */
+ if (!kbd_last_ret)
+ return;
+
+ /*
+ * Enter key. Need to absorb the break
+ * code here, lest it gets leaked out
+ * if we exit KDB as the result of
+ * processing 'g'.
+ *
+ * This has several interesting implications:
+ * + Need to handle KP ENTER, which has break
+ * code 0xe0 0x9c.
+ * + Need to handle repeat ENTER and repeat
+ * KP ENTER. Repeats only get a break code
+ * at the end of the repeated sequence. This
+ * means we can't propagate the repeated key
+ * press, and must swallow it away.
+ * + Need to handle possible PS/2 mouse input.
+ * + Need to handle mashed keys.
+ */
+
+ while (1) {
while ((inb(KBD_STATUS_REG) & KBD_STAT_OBF) == 0)
- ;
+ cpu_relax();
/*
- * Fetch the scancode
+ * Fetch the scancode.
*/
scancode = inb(KBD_DATA_REG);
scanstatus = inb(KBD_STATUS_REG);
- while (scanstatus & KBD_STAT_MOUSE_OBF) {
- scancode = inb(KBD_DATA_REG);
- scanstatus = inb(KBD_STATUS_REG);
- }
+ /*
+ * Skip mouse input.
+ */
+ if (scanstatus & KBD_STAT_MOUSE_OBF)
+ continue;
- if (scancode != 0x9c) {
- /*
- * Wasn't an enter-release, why not?
- */
- kdb_printf("kdb: expected enter got 0x%x status 0x%x\n",
- scancode, scanstatus);
- }
+ /*
+ * If we see 0xe0, this is either
+ * a break code for KP ENTER, or a repeat
+ * make for KP ENTER. Either way, since
+ * the second byte is equivalent to an ENTER,
+ * skip the 0xe0 and try again.
+ *
+ * If we see 0x1c, this must be a repeat
+ * ENTER or KP ENTER (and we swallowed 0xe0
+ * before). Try again.
+ *
+ * We can also see make and break codes
+ * for other keys mashed before or after
+ * pressing ENTER. Thus, if we see anything
+ * other than 0x9c, we have to try again.
+ *
+ * Note, if you held some key as ENTER
+ * was depressed, that break code would
+ * get leaked out.
+ */
+ if (scancode != 0x9c)
+ continue;
- return 13;
+ return;
}
-
- return keychar & 0xff;
}
-EXPORT_SYMBOL_GPL(kdb_get_kbd_char);
+EXPORT_SYMBOL_GPL(kdb_kbd_cleanup_state);
--
1.7.9.2
I think we are a lot closer this time. I attached a new patch based
on your prior version.
On 02/27/2012 10:26 PM, Andrei Warkentin wrote:
> This fixes the following problems:
> 1) Typematic-repeat of 'enter' gives warning message
> and leaks make/break if KDB exits. Repeats
> look something like 0x1c 0x1c .... 0x9c
> 2) Use of 'keypad enter' gives warning message and
> leaks the ENTER break/make code out if KDB exits.
> KP ENTER repeats look someting like 0xe0 0x1c
> 0xe0 0x1c ... 0xe0 0x9c.
> 3) Lag on the order of seconds between "break" and "make" when
> expecting the enter "break" code. Seen under virtualized
> environments such as VMware ESX.
>
> The existing special enter handler tries to glob the
> enter break code, but this fails if the other (KP) enter
> was used, or if there was a key repeat. It also fails
> if you mashed some keys along with enter, and you ended
> up with a non-enter make or non-enter break code coming
> after the enter make code. So first, we modify the handler
> to handle these cases. But performing these actions on
> every enter is annoying since now you can't hold ENTER down
> to scroll <more>d messages in KDB. Since this special
> behaviour is only necessary to handle the exiting KDB
> ('g' + ENTER) without leaking scancodes to the OS, we
> perform the cleanup in kgdboc_post_exp_handler path.
This is the wrong place to do this. The cleanup needs to get executed
any time you are going to leave kdb_main(), because there are a few
conditions like cpu switch, and transition to kgdb where you would
leak the enter code on to the buffer handler without ever returning to
the OS. Also if you did not set CONFIG_KDB_KEYBOARD your patch did
not compile properly. We could have fixed this by moving the code to
the input cleanup in kgdboc, but this is not needed at all if you take
a look at the revised version.
>
>
> Fixed previous regression where if kbd was not used
> to 'g' + ENTER, the cleanup code would hang.
Was this a regression in the out of tree code or something in the
mainline kdb?
> index 4bca634..34a8722 100644
> --- a/kernel/debug/kdb/kdb_keyboard.c
> +++ b/kernel/debug/kdb/kdb_keyboard.c
> +void kdb_kbd_cleanup_state(void)
> +{
> + int scancode, scanstatus;
> +
> + /*
> + * Nothing to clean up, since either
> + * ENTER was never pressed, or has already
> + * gotten cleaned up.
> + */
> + if (!kbd_last_ret)
> + return;
I added kbd_last_ret = 0; right here so this cannot get triggered a
second time if the kdb main loop is entered and exited without going
to the shell (which can happen on a soft single step operation on some archs).
Jason.
PS We can iterate other patches you sent after we settle on this one.
Thanks!
Hi Jason,
----- Original Message -----
> From: "Jason Wessel" <[email protected]>
> To: "Andrei Warkentin" <[email protected]>
> Cc: [email protected], [email protected], "Andrei Warkentin" <[email protected]>
> Sent: Tuesday, February 28, 2012 8:53:20 AM
> Subject: Re: [PATCH] KDB: Fix usability issues relating to the 'enter' key.
>
>
> I think we are a lot closer this time. I attached a new patch based
> on your prior version.
>
> This is the wrong place to do this. The cleanup needs to get
> executed
> any time you are going to leave kdb_main(), because there are a few
> conditions like cpu switch, and transition to kgdb where you would
> leak the enter code on to the buffer handler without ever returning
> to
> the OS. Also if you did not set CONFIG_KDB_KEYBOARD your patch did
> not compile properly. We could have fixed this by moving the code to
> the input cleanup in kgdboc, but this is not needed at all if you
> take a look at the revised version.
>
Oh I see now. Okay. I'll keep this in mind.
> >
> >
> > Fixed previous regression where if kbd was not used
> > to 'g' + ENTER, the cleanup code would hang.
>
>
> Was this a regression in the out of tree code or something in the
> mainline kdb?
The first patch I sent out didn't have the kbd_last_ret bit. I found this
while using netkgdb and kgdboc together, but I'm certain this would
also apply if you ran kgdboc=ttyS0,kbd and broke in/resumed from serial
without ever touching the i8042 keyboard.
>
> I added kbd_last_ret = 0; right here so this cannot get triggered a
> second time if the kdb main loop is entered and exited without going
> to the shell (which can happen on a soft single step operation on
> some archs).
>
Looks good to me! I gave it a spin as well.
Acked-by: Andrei Warkentin <[email protected]>
A