Skip to contentSkip to conversation

PATCH: virtio_console: Fix poll blocking even though there is data to read

PATCH Hans de Goede 3 participants 15 messages Sep 15, 2010 – Sep 16, 2010

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

Hi All,

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

virtio_console's port_fops_poll checks port->inbuf != NULL to determine if read won't block. However if an application reads enough bytes from inbuf through port_fops_read, to empty the current port->inbuf, port->inbuf will be NULL even though there may be buffers left in the virtqueue.

This causes poll() to block even though there is data to be read, this patch fixes this by using the alredy defined will_read_block utility function instead of the port->inbuf != NULL check.

Regards,

Hans

Subject: virtio_console: Fix poll blocking even though there is data to read
From: Hans de Goede <[email protected]>

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

virtio_console's port_fops_poll checks port->inbuf != NULL to determine if read won't block. However if an application reads enough bytes from inbuf through port_fops_read, to empty the current port->inbuf, port->inbuf will be NULL even though there may be buffers left in the virtqueue.

This causes poll() to block even though there is data to be read, this patch fixes this by using the alredy defined will_read_block utility function instead of the port->inbuf != NULL check.

drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
diff -up linux-2.6.35.x86_64/drivers/char/virtio_console.c~ linux-2.6.35.x86_64/drivers/char/virtio_console.c
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp, &port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))
 		ret |= POLLIN | POLLRDNORM;
 	if (!will_write_block(port))
 		ret |= POLLOUT;
#1 lore raw
Quoted 29 lines · Hans de Goede

Hi All,

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

virtio_console's port_fops_poll checks port->inbuf != NULL to determine if read won't block. However if an application reads enough bytes from inbuf through port_fops_read, to empty the current port->inbuf, port->inbuf will be NULL even though there may be buffers left in the virtqueue.

This causes poll() to block even though there is data to be read, this patch fixes this by using the alredy defined will_read_block utility function instead of the port->inbuf != NULL check.

Regards,

Hans

diff -up linux-2.6.35.x86_64/drivers/char/virtio_console.c~ linux-2.6.35.x86_64/drivers/char/virtio_console.c

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 5 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp, &port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

if (port_has_data(port))

instead.

will_read_block() also tests if the host is connected, which is not what we want for POLLIN to be set.

Amit

#2 lore raw
In reply to Hans de Goede
Quoted 29 lines · Hans de Goede

Hi All,

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

virtio_console's port_fops_poll checks port->inbuf != NULL to determine if read won't block. However if an application reads enough bytes from inbuf through port_fops_read, to empty the current port->inbuf, port->inbuf will be NULL even though there may be buffers left in the virtqueue.

This causes poll() to block even though there is data to be read, this patch fixes this by using the alredy defined will_read_block utility function instead of the port->inbuf != NULL check.

Regards,

Hans

diff -up linux-2.6.35.x86_64/drivers/char/virtio_console.c~ linux-2.6.35.x86_64/drivers/char/virtio_console.c

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 5 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp, &port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

if (port_has_data(port))

instead.

will_read_block() also tests if the host is connected, which is not what we want for POLLIN to be set.

Amit

#3 lore raw

Hi,

Quoted 38 lines · Amit Shah
Quoted 29 lines · Hans de Goede

Hi All,

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

virtio_console's port_fops_poll checks port->inbuf != NULL to determine if read won't block. However if an application reads enough bytes from inbuf through port_fops_read, to empty the current port->inbuf, port->inbuf will be NULL even though there may be buffers left in the virtqueue.

This causes poll() to block even though there is data to be read, this patch fixes this by using the alredy defined will_read_block utility function instead of the port->inbuf != NULL check.

Regards,

Hans

diff -up linux-2.6.35.x86_64/drivers/char/virtio_console.c~ linux-2.6.35.x86_64/drivers/char/virtio_console.c

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 5 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
  	poll_wait(filp,&port->waitqueue, wait);
 
  	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Regards,

Hans
_______________________________________________
Spice-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/spice-devel

#14 lore raw
In reply to Amit Shah

Hi,

Quoted 38 lines · Amit Shah
Quoted 29 lines · Hans de Goede

Hi All,

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

I found this while working on a Linux agent for spice, the symptom I was seeing was select blocking on the spice vdagent virtio serial port even though there were messages queued up there.

virtio_console's port_fops_poll checks port->inbuf != NULL to determine if read won't block. However if an application reads enough bytes from inbuf through port_fops_read, to empty the current port->inbuf, port->inbuf will be NULL even though there may be buffers left in the virtqueue.

This causes poll() to block even though there is data to be read, this patch fixes this by using the alredy defined will_read_block utility function instead of the port->inbuf != NULL check.

Regards,

Hans

diff -up linux-2.6.35.x86_64/drivers/char/virtio_console.c~ linux-2.6.35.x86_64/drivers/char/virtio_console.c

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 5 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
  	poll_wait(filp,&port->waitqueue, wait);
 
  	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Regards,

Hans
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/virtualization

#4 lore raw
In reply to Hans de Goede
Quoted 18 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Quoted 8 lines

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

poll(2) says:

POLLIN There is data to read.

That makes it simple.

Also, we also set POLLHUP in case the host is disconnected, so if host is not connected while there's data to read, we'll get POLLIN|POLLHUP in revents, and further reads will be blocked.

I'm willing to go with the poll manpage and use port_has_data().

Amit
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/virtualization

#5 lore raw
In reply to Hans de Goede
Quoted 18 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Quoted 8 lines

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

poll(2) says:

POLLIN There is data to read.

That makes it simple.

Also, we also set POLLHUP in case the host is disconnected, so if host is not connected while there's data to read, we'll get POLLIN|POLLHUP in revents, and further reads will be blocked.

I'm willing to go with the poll manpage and use port_has_data().

Amit

#6 lore raw

Hi,

Quoted 30 lines · Amit Shah
Quoted 18 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
  	poll_wait(filp,&port->waitqueue, wait);
 
  	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Quoted 8 lines

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

poll(2) says:

Quoted 8 lines

POLLIN There is data to read.

That makes it simple.

Also, we also set POLLHUP in case the host is disconnected, so if host is not connected while there's data to read, we'll get POLLIN|POLLHUP in revents, and further reads will be blocked.

Ah, I completely missed the POLLHUP getting set thingie. I guess that will get used by the kernel to also not block if any fds have it said in select's readfds argument.

One new patch coming up!

Regards,

Hans
_______________________________________________
Spice-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/spice-devel

#7 lore raw
In reply to Amit Shah

Hi,

Quoted 30 lines · Amit Shah
Quoted 18 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
  	poll_wait(filp,&port->waitqueue, wait);
 
  	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Quoted 8 lines

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

poll(2) says:

Quoted 8 lines

POLLIN There is data to read.

That makes it simple.

Also, we also set POLLHUP in case the host is disconnected, so if host is not connected while there's data to read, we'll get POLLIN|POLLHUP in revents, and further reads will be blocked.

Ah, I completely missed the POLLHUP getting set thingie. I guess that will get used by the kernel to also not block if any fds have it said in select's readfds argument.

One new patch coming up!

Regards,

Hans
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/virtualization

#8 lore raw
In reply to Amit Shah
Quoted 28 lines · Amit Shah
Quoted 18 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Quoted 8 lines

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Quoted 5 lines

poll(2) says:

POLLIN There is data to read.

That makes it simple.

That's a documentation bug. On a pipe, POLLIN is set when the other end closes. read() then returns 0 immediately.

poll() sets POLLIN when read() won't block, and people count on it.

Let's do that, please.
Rusty.
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/virtualization

#9 lore raw
In reply to Amit Shah
Quoted 28 lines · Amit Shah
Quoted 18 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Quoted 8 lines

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Quoted 5 lines

poll(2) says:

POLLIN There is data to read.

That makes it simple.

That's a documentation bug. On a pipe, POLLIN is set when the other end closes. read() then returns 0 immediately.

poll() sets POLLIN when read() won't block, and people count on it.

Let's do that, please.
Rusty.

#10 lore raw
Quoted 29 lines · Rusty Russell
Quoted 28 lines · Amit Shah
Quoted 26 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Quoted 8 lines
Quoted 5 lines

poll(2) says:

POLLIN There is data to read.

That makes it simple.

That's a documentation bug. On a pipe, POLLIN is set when the other end closes. read() then returns 0 immediately.

Currently we don't set POLLIN when host goes down. I'll do a second patch for that.

Quoted 1 line

poll() sets POLLIN when read() won't block, and people count on it.

Yes, that's the behaviour with Hans's new patch as well -- that's not changing.

The will_read_block() function (and the comment on top of it) are causing this confusion:

/* The condition that must be true for polling to end */ static bool will_read_block(struct port *port)

{
	if (!port->guest_connected) {
		/* Port got hot-unplugged. Let's exit. */
		return false;
	}
	return !port_has_data(port) && port->host_connected;
}

This function is only called to unblock a blocking read() call. So the comment there has to be changed to read 'waiting' to end instead of 'polling' to end.

read() does return 0 immediately when the other end is not connected (and there's no data to read).

In effect, we need:
- Hans's patch
- a patch to set POLLIN when host goes down (in addition to POLLHUP and SIGIO)
- a patch to change the comment for will_read_block.

Thanks,

Amit
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/virtualization

#11 lore raw
In reply to Rusty Russell
Quoted 29 lines · Rusty Russell
Quoted 28 lines · Amit Shah
Quoted 26 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Quoted 8 lines
Quoted 5 lines

poll(2) says:

POLLIN There is data to read.

That makes it simple.

That's a documentation bug. On a pipe, POLLIN is set when the other end closes. read() then returns 0 immediately.

Currently we don't set POLLIN when host goes down. I'll do a second patch for that.

Quoted 1 line

poll() sets POLLIN when read() won't block, and people count on it.

Yes, that's the behaviour with Hans's new patch as well -- that's not changing.

The will_read_block() function (and the comment on top of it) are causing this confusion:

/* The condition that must be true for polling to end */ static bool will_read_block(struct port *port)

{
	if (!port->guest_connected) {
		/* Port got hot-unplugged. Let's exit. */
		return false;
	}
	return !port_has_data(port) && port->host_connected;
}

This function is only called to unblock a blocking read() call. So the comment there has to be changed to read 'waiting' to end instead of 'polling' to end.

read() does return 0 immediately when the other end is not connected (and there's no data to read).

In effect, we need:
- Hans's patch
- a patch to set POLLIN when host goes down (in addition to POLLHUP and SIGIO)
- a patch to change the comment for will_read_block.

Thanks,

Amit

#12 lore raw
Quoted 65 lines · Amit Shah
Quoted 29 lines · Rusty Russell
Quoted 28 lines · Amit Shah
Quoted 26 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Quoted 8 lines
Quoted 5 lines

poll(2) says:

POLLIN There is data to read.

That makes it simple.

That's a documentation bug. On a pipe, POLLIN is set when the other end closes. read() then returns 0 immediately.

Currently we don't set POLLIN when host goes down. I'll do a second patch for that.

Quoted 1 line

poll() sets POLLIN when read() won't block, and people count on it.

Yes, that's the behaviour with Hans's new patch as well -- that's not changing.

The will_read_block() function (and the comment on top of it) are causing this confusion:

/* The condition that must be true for polling to end */ static bool will_read_block(struct port *port)

{
	if (!port->guest_connected) {
		/* Port got hot-unplugged. Let's exit. */
		return false;
	}
	return !port_has_data(port) && port->host_connected;
}

This function is only called to unblock a blocking read() call. So the comment there has to be changed to read 'waiting' to end instead of 'polling' to end.

Quoted 8 lines

read() does return 0 immediately when the other end is not connected (and there's no data to read).

In effect, we need:
- Hans's patch
- a patch to set POLLIN when host goes down (in addition to POLLHUP and SIGIO)
- a patch to change the comment for will_read_block.

Uh, sorry for blabbering.

will_read_block() was written to be used in poll(), and the comment says as much. I should've used it right from the start. Hans's first patch was correct, I'll pick that up and that addresses all the "issues" that are present.

Sorry again!

Amit
_______________________________________________
Spice-devel mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/spice-devel

#13 lore raw
In reply to Amit Shah
Quoted 65 lines · Amit Shah
Quoted 29 lines · Rusty Russell
Quoted 28 lines · Amit Shah
Quoted 26 lines · Hans de Goede
Quoted 11 lines

drivers/char/virtio_console.c static unsigned int port_fops_poll(struc · line 645

Quoted 8 lines
drivers/char/virtio_console.c renamed from drivers/char/virtio_console.c~ +1 -1
@@ -642,7 +642,7 @@ static unsigned int port_fops_poll(struc
 	poll_wait(filp,&port->waitqueue, wait);
 
 	ret = 0;
-	if (port->inbuf)
+	if (!will_read_block(port))

Looks correct, but this should be

Quoted 3 lines

if (port_has_data(port))

instead.

That certainly works for me (as in will still fix the bug I'm hitting), but quoting from "man 2 select":

Three independent sets of file descriptors are watched. Those listed in readfds will be watched to see if characters become available for reading (more precisely, to see if a read will not block; in particu‐ lar, a file descriptor is also ready on end-of-file)

Notice the "a file descriptor is also ready on end-of-file", and port_fops_read treats the host not being connected as eof (it returns 0 in that case). So from an API pov I'm not sure what is correct.

Quoted 8 lines
Quoted 5 lines

poll(2) says:

POLLIN There is data to read.

That makes it simple.

That's a documentation bug. On a pipe, POLLIN is set when the other end closes. read() then returns 0 immediately.

Currently we don't set POLLIN when host goes down. I'll do a second patch for that.

Quoted 1 line

poll() sets POLLIN when read() won't block, and people count on it.

Yes, that's the behaviour with Hans's new patch as well -- that's not changing.

The will_read_block() function (and the comment on top of it) are causing this confusion:

/* The condition that must be true for polling to end */ static bool will_read_block(struct port *port)

{
	if (!port->guest_connected) {
		/* Port got hot-unplugged. Let's exit. */
		return false;
	}
	return !port_has_data(port) && port->host_connected;
}

This function is only called to unblock a blocking read() call. So the comment there has to be changed to read 'waiting' to end instead of 'polling' to end.

Quoted 8 lines

read() does return 0 immediately when the other end is not connected (and there's no data to read).

In effect, we need:
- Hans's patch
- a patch to set POLLIN when host goes down (in addition to POLLHUP and SIGIO)
- a patch to change the comment for will_read_block.

Uh, sorry for blabbering.

will_read_block() was written to be used in poll(), and the comment says as much. I should've used it right from the start. Hans's first patch was correct, I'll pick that up and that addresses all the "issues" that are present.

Sorry again!

Amit
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/virtualization