The return value need to be either ignored or acted upon, otherwise 'deadstore'
clang check would yell at us. I think it's better to just ignore what this
particular call of set_registers() returns. The adapter defaults are sane and
it would be operational even if the register write fail.
Signed-off-by: Petko Manolov <redacted>
---
drivers/net/usb/pegasus.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
From: Jakub Kicinski <kuba@kernel.org> Date: 2021-08-13 23:24:42
On Thu, 12 Aug 2021 11:23:51 +0300 Petko Manolov wrote:
quoted hunk
The return value need to be either ignored or acted upon, otherwise 'deadstore'
clang check would yell at us. I think it's better to just ignore what this
particular call of set_registers() returns. The adapter defaults are sane and
it would be operational even if the register write fail.
Signed-off-by: Petko Manolov <redacted>
---
drivers/net/usb/pegasus.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
This one is not added by the recent changes as I initially thought,
the driver has always checked this return value. The recent changes
did this:
ret = set_registers(pegasus, EthCtrl0, 3, data);
if (usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS2 ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_DLINK) {
u16 auxmode;
- read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ ret = read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ if (ret < 0)
+ goto fail;
auxmode |= 4;
write_mii_word(pegasus, 0, 0x1b, &auxmode);
}
+ return 0;
+fail:
+ netif_dbg(pegasus, drv, pegasus->net, "%s failed\n", __func__);
return ret;
}
now the return value of set_registeres() is ignored.
Seems like a better fix would be to bring back the error checking,
why not?
Please remember to add a fixes tag.
From: Pavel Skripkin <hidden> Date: 2021-08-14 13:19:08
On 8/14/21 2:24 AM, Jakub Kicinski wrote:
On Thu, 12 Aug 2021 11:23:51 +0300 Petko Manolov wrote:
quoted
The return value need to be either ignored or acted upon, otherwise 'deadstore'
clang check would yell at us. I think it's better to just ignore what this
particular call of set_registers() returns. The adapter defaults are sane and
it would be operational even if the register write fail.
Signed-off-by: Petko Manolov <redacted>
---
drivers/net/usb/pegasus.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
This one is not added by the recent changes as I initially thought,
the driver has always checked this return value. The recent changes
did this:
ret = set_registers(pegasus, EthCtrl0, 3, data);
if (usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS2 ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_DLINK) {
u16 auxmode;
- read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ ret = read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ if (ret < 0)
+ goto fail;
auxmode |= 4;
write_mii_word(pegasus, 0, 0x1b, &auxmode);
}
+ return 0;
+fail:
+ netif_dbg(pegasus, drv, pegasus->net, "%s failed\n", __func__);
return ret;
}
now the return value of set_registeres() is ignored.
Seems like a better fix would be to bring back the error checking,
why not?
Please remember to add a fixes tag.
Hi, Jakub!
I've suggested to handle this error, but Petko said that device won't
stop working, it will just get in non-optimal state.
https://lore.kernel.org/lkml/YRF1t5kx6hTrv5LC@carbon/
With regards,
Pavel Skripkin
On Thu, 12 Aug 2021 11:23:51 +0300 Petko Manolov wrote:
quoted
The return value need to be either ignored or acted upon, otherwise 'deadstore'
clang check would yell at us. I think it's better to just ignore what this
particular call of set_registers() returns. The adapter defaults are sane and
it would be operational even if the register write fail.
Signed-off-by: Petko Manolov <redacted>
---
drivers/net/usb/pegasus.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
This one is not added by the recent changes as I initially thought,
the driver has always checked this return value. The recent changes
did this:
ret = set_registers(pegasus, EthCtrl0, 3, data);
if (usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS2 ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_DLINK) {
u16 auxmode;
- read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ ret = read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ if (ret < 0)
+ goto fail;
auxmode |= 4;
write_mii_word(pegasus, 0, 0x1b, &auxmode);
}
+ return 0;
+fail:
+ netif_dbg(pegasus, drv, pegasus->net, "%s failed\n", __func__);
return ret;
}
now the return value of set_registeres() is ignored.
Seems like a better fix would be to bring back the error checking,
why not?
Mostly because for this particular adapter checking the read failure makes much
more sense than write failure.
Checking the return value of set_register(s) is often usless because device's
default register values are sane enough to get a working ethernet adapter even
without much prodding. There are exceptions, though, one of them being
set_ethernet_addr().
You could read the discussing in the netdev ML, but the essence of it is that
set_ethernet_addr() should not give up if set_register(s) fail. Instead, the
driver should assign a valid, even if random, MAC address.
It is much the same situation with enable_net_traffic() - it should continue
regardless. There are two options to resolve this: a) remove the error check
altogether; b) do the check and print a debug message. I prefer a), but i am
also not strongly opposed to b). Comments?
This one is not added by the recent changes as I initially thought,
the driver has always checked this return value. The recent changes
did this:
ret = set_registers(pegasus, EthCtrl0, 3, data);
if (usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS2 ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_DLINK) {
u16 auxmode;
- read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ ret = read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ if (ret < 0)
+ goto fail;
auxmode |= 4;
write_mii_word(pegasus, 0, 0x1b, &auxmode);
}
+ return 0;
+fail:
+ netif_dbg(pegasus, drv, pegasus->net, "%s failed\n", __func__);
return ret;
}
now the return value of set_registeres() is ignored.
Seems like a better fix would be to bring back the error checking,
why not?
Mostly because for this particular adapter checking the read failure makes much
more sense than write failure.
This is not an either-or choice.
Checking the return value of set_register(s) is often usless because device's
default register values are sane enough to get a working ethernet adapter even
without much prodding. There are exceptions, though, one of them being
set_ethernet_addr().
You could read the discussing in the netdev ML, but the essence of it is that
set_ethernet_addr() should not give up if set_register(s) fail. Instead, the
driver should assign a valid, even if random, MAC address.
It is much the same situation with enable_net_traffic() - it should continue
regardless. There are two options to resolve this: a) remove the error check
altogether; b) do the check and print a debug message. I prefer a), but i am
also not strongly opposed to b). Comments?
c) keep propagating the error like the driver used to.
I don't understand why that's not the most obvious option.
The driver used to propagate the errors from the set_registers() call
in enable_net_traffic() since the beginning of the git era. This is
_not_ one of the error checking that you recently added.
This one is not added by the recent changes as I initially thought,
the driver has always checked this return value. The recent changes
did this:
ret = set_registers(pegasus, EthCtrl0, 3, data);
if (usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_LINKSYS2 ||
usb_dev_id[pegasus->dev_index].vendor == VENDOR_DLINK) {
u16 auxmode;
- read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ ret = read_mii_word(pegasus, 0, 0x1b, &auxmode);
+ if (ret < 0)
+ goto fail;
auxmode |= 4;
write_mii_word(pegasus, 0, 0x1b, &auxmode);
}
+ return 0;
+fail:
+ netif_dbg(pegasus, drv, pegasus->net, "%s failed\n", __func__);
return ret;
}
now the return value of set_registeres() is ignored.
Seems like a better fix would be to bring back the error checking,
why not?
Mostly because for this particular adapter checking the read failure makes much
more sense than write failure.
This is not an either-or choice.
quoted
Checking the return value of set_register(s) is often usless because device's
default register values are sane enough to get a working ethernet adapter even
without much prodding. There are exceptions, though, one of them being
set_ethernet_addr().
You could read the discussing in the netdev ML, but the essence of it is that
set_ethernet_addr() should not give up if set_register(s) fail. Instead, the
driver should assign a valid, even if random, MAC address.
It is much the same situation with enable_net_traffic() - it should continue
regardless. There are two options to resolve this: a) remove the error check
altogether; b) do the check and print a debug message. I prefer a), but i am
also not strongly opposed to b). Comments?
c) keep propagating the error like the driver used to.
If you carefully read the code, which dates back to at least 2005, you'll see
that on line 436 (v5.14-rc6) 'ret' is assigned with the return value of
set_registers(), but 'ret' is never evaluated and thus not acted upon.
I don't understand why that's not the most obvious option.
Which part of "this is not a fatal error" you did not understand?
The driver used to propagate the errors from the set_registers() call in
enable_net_traffic() since the beginning of the git era. This is _not_ one of
the error checking that you recently added.
The driver hasn't propagated an error at this particular location in the last 16
years. So how exactly removing this assignment will make the driver worse than
it is now?
Anyway, i'll add a warn() and be done with it.
Petko
From: Jakub Kicinski <kuba@kernel.org> Date: 2021-08-16 20:18:31
On Mon, 16 Aug 2021 22:14:47 +0300 Petko Manolov wrote:
On 21-08-16 07:06:40, Jakub Kicinski wrote:
quoted
On Sun, 15 Aug 2021 11:54:55 +0300 Petko Manolov wrote:
quoted
Mostly because for this particular adapter checking the read failure makes much
more sense than write failure.
This is not an either-or choice.
quoted
Checking the return value of set_register(s) is often usless because device's
default register values are sane enough to get a working ethernet adapter even
without much prodding. There are exceptions, though, one of them being
set_ethernet_addr().
You could read the discussing in the netdev ML, but the essence of it is that
set_ethernet_addr() should not give up if set_register(s) fail. Instead, the
driver should assign a valid, even if random, MAC address.
It is much the same situation with enable_net_traffic() - it should continue
regardless. There are two options to resolve this: a) remove the error check
altogether; b) do the check and print a debug message. I prefer a), but i am
also not strongly opposed to b). Comments?
c) keep propagating the error like the driver used to.
If you carefully read the code, which dates back to at least 2005, you'll see
that on line 436 (v5.14-rc6) 'ret' is assigned with the return value of
set_registers(), but 'ret' is never evaluated and thus not acted upon.
It's no longer evaluated because of your commit 8a160e2e9aeb ("net:
usb: pegasus: Check the return value of get_geristers() and friends;")
IOW v5.14-rc6 has your recent patch. Which I quoted earlier in this
thread. That commit was on Aug 3 2021. The error checking (now
accidentally removed) was introduced somewhere in 2.6.x days.
If you disagree with that please show me the code you're referring to,
because I just don't see it.
quoted
I don't understand why that's not the most obvious option.
Which part of "this is not a fatal error" you did not understand?
That's not the point. The error checking was removed accidentally,
it should be brought back in net to avoid introducing regressions.
If the error checking is not necessary you can remove it in net-next,
no problem.
Perhaps you did not intend commit 8a160e2e9aeb ("net: usb: pegasus:
Check the return value of get_geristers() and friends;") to be applied
as a fix but it was, and it was backported to stable trees if I'm not
mistaken.
quoted
The driver used to propagate the errors from the set_registers() call in
enable_net_traffic() since the beginning of the git era. This is _not_ one of
the error checking that you recently added.
The driver hasn't propagated an error at this particular location in the last 16
years. So how exactly removing this assignment will make the driver worse than
it is now?
On Mon, 16 Aug 2021 22:14:47 +0300 Petko Manolov wrote:
quoted
On 21-08-16 07:06:40, Jakub Kicinski wrote:
quoted
On Sun, 15 Aug 2021 11:54:55 +0300 Petko Manolov wrote:
quoted
Mostly because for this particular adapter checking the read failure makes much
more sense than write failure.
This is not an either-or choice.
quoted
Checking the return value of set_register(s) is often usless because device's
default register values are sane enough to get a working ethernet adapter even
without much prodding. There are exceptions, though, one of them being
set_ethernet_addr().
You could read the discussing in the netdev ML, but the essence of it is that
set_ethernet_addr() should not give up if set_register(s) fail. Instead, the
driver should assign a valid, even if random, MAC address.
It is much the same situation with enable_net_traffic() - it should continue
regardless. There are two options to resolve this: a) remove the error check
altogether; b) do the check and print a debug message. I prefer a), but i am
also not strongly opposed to b). Comments?
c) keep propagating the error like the driver used to.
If you carefully read the code, which dates back to at least 2005, you'll see
that on line 436 (v5.14-rc6) 'ret' is assigned with the return value of
set_registers(), but 'ret' is never evaluated and thus not acted upon.
It's no longer evaluated because of your commit 8a160e2e9aeb ("net:
usb: pegasus: Check the return value of get_geristers() and friends;")
IOW v5.14-rc6 has your recent patch. Which I quoted earlier in this
thread. That commit was on Aug 3 2021. The error checking (now
accidentally removed) was introduced somewhere in 2.6.x days.
If you disagree with that please show me the code you're referring to,
because I just don't see it.
quoted
quoted
I don't understand why that's not the most obvious option.
Which part of "this is not a fatal error" you did not understand?
That's not the point. The error checking was removed accidentally, it should
be brought back in net to avoid introducing regressions.
Not introducing regressions is the argument that sold me on. If i don't get too
lazy i may further change this part of the code, but that's in the future. I've
sent a patch that restores the original behavior in addition of a few small
tweaks.
Petko