| Message ID | 1404147350-28904-2-git-send-email-somlo@cmu.edu |
|---|---|
| State | New |
| Headers | show |
On Mon, Jun 30, 2014 at 12:55:49PM -0400, Gabriel L. Somlo wrote: > set_phy_ctrl() runs the same autonegotiation availability checks > that were factored out into have_autoneg() in an earlier commit > (d7a4155265416a1c8f3067b59e68bf5fda1d6215), except it runs them > on the value about to be written into the phy_ctrl register, and > then does not actually write the register. > > This patch moves the have_autoneg() function in front of > set_phy_ctrl(), then updates the latter to actually write > the register with the given value, and use the factored-out > check to determine whether the link should be bounced. > > Also, fix indentation/line width when setting up the timer. > > Signed-off-by: Gabriel Somlo <somlo@cmu.edu> Weird. So does this mean have_autoneg is always false then? How does it work? Wrt to patch itself, let's do it properly pls: - bits 0:5 are RO on some cards - bit 9 is SC > --- > > I'm not 100% sure whether "val" in set_phy_ctrl() is not written > to phy_reg[PHY_CTRL] by mistake, or whether there's an ulterior > motive (in which case we should add a comment as to why that is). > > Assuming it's a mistake, and we *should* have been writing it (as > per my patch below), we can then use have_autoneg() to also simplify > the code significantly :) > > Thanks, > Gabriel > > hw/net/e1000.c | 27 ++++++++++++++------------- > 1 file changed, 14 insertions(+), 13 deletions(-) > > diff --git a/hw/net/e1000.c b/hw/net/e1000.c > index 0fc29a0..2376910 100644 > --- a/hw/net/e1000.c > +++ b/hw/net/e1000.c > @@ -186,6 +186,14 @@ e1000_link_up(E1000State *s) > s->phy_reg[PHY_STATUS] |= MII_SR_LINK_STATUS; > } > > +static bool > +have_autoneg(E1000State *s) > +{ > + return (s->compat_flags & E1000_FLAG_AUTONEG) && > + (s->phy_reg[PHY_CTRL] & MII_CR_AUTO_NEG_EN) && > + (s->phy_reg[PHY_CTRL] & MII_CR_RESTART_AUTO_NEG); > +} > + > static void > set_phy_ctrl(E1000State *s, int index, uint16_t val) > { > @@ -194,13 +202,14 @@ set_phy_ctrl(E1000State *s, int index, uint16_t val) > * migrate during auto negotiation, after migration the link will be > * down. > */ > - if (!(s->compat_flags & E1000_FLAG_AUTONEG)) { > - return; > - } > - if ((val & MII_CR_AUTO_NEG_EN) && (val & MII_CR_RESTART_AUTO_NEG)) { > + > + s->phy_reg[PHY_CTRL] = val; > + > + if (have_autoneg(s)) { > e1000_link_down(s); > DBGOUT(PHY, "Start link auto negotiation\n"); > - timer_mod(s->autoneg_timer, qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + 500); > + timer_mod(s->autoneg_timer, > + qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + 500); > } > } > > @@ -848,14 +857,6 @@ receive_filter(E1000State *s, const uint8_t *buf, int size) > return 0; > } > > -static bool > -have_autoneg(E1000State *s) > -{ > - return (s->compat_flags & E1000_FLAG_AUTONEG) && > - (s->phy_reg[PHY_CTRL] & MII_CR_AUTO_NEG_EN) && > - (s->phy_reg[PHY_CTRL] & MII_CR_RESTART_AUTO_NEG); > -} > - > static void > e1000_set_link_status(NetClientState *nc) > { > -- > 1.9.3
On Mon, Jun 30, 2014 at 09:04:12PM +0300, Michael S. Tsirkin wrote: > On Mon, Jun 30, 2014 at 12:55:49PM -0400, Gabriel L. Somlo wrote: > > set_phy_ctrl() runs the same autonegotiation availability checks > > that were factored out into have_autoneg() in an earlier commit > > (d7a4155265416a1c8f3067b59e68bf5fda1d6215), except it runs them > > on the value about to be written into the phy_ctrl register, and > > then does not actually write the register. > > > > This patch moves the have_autoneg() function in front of > > set_phy_ctrl(), then updates the latter to actually write > > the register with the given value, and use the factored-out > > check to determine whether the link should be bounced. > > > > Also, fix indentation/line width when setting up the timer. > > > > Signed-off-by: Gabriel Somlo <somlo@cmu.edu> > > Weird. > So does this mean have_autoneg is always false then? > How does it work? phy_reg[PHY_CTRL] is initialized (along with the rest of phy_reg) by a memmove from phy_reg_init, during e1000_reset(). So the question would be rather "does it make sense to look at anything besides (s->compat_flags & E1000_FLAG_AUTONEG), since we never allow phy_reg[PHY_CTRL] to change after initialization, so MII_CR_AUTO_NEG_EN and MII_CR_RESTART_AUTO_NEG bits will always be set?" > Wrt to patch itself, let's do it properly pls: > - bits 0:5 are RO on some cards > - bit 9 is SC Will do, once we've sorted out the phy_reg[] mystery :) Thanks, --Gabriel > > > --- > > > > I'm not 100% sure whether "val" in set_phy_ctrl() is not written > > to phy_reg[PHY_CTRL] by mistake, or whether there's an ulterior > > motive (in which case we should add a comment as to why that is). > > > > Assuming it's a mistake, and we *should* have been writing it (as > > per my patch below), we can then use have_autoneg() to also simplify > > the code significantly :) > > > > Thanks, > > Gabriel > > > > hw/net/e1000.c | 27 ++++++++++++++------------- > > 1 file changed, 14 insertions(+), 13 deletions(-) > > > > diff --git a/hw/net/e1000.c b/hw/net/e1000.c > > index 0fc29a0..2376910 100644 > > --- a/hw/net/e1000.c > > +++ b/hw/net/e1000.c > > @@ -186,6 +186,14 @@ e1000_link_up(E1000State *s) > > s->phy_reg[PHY_STATUS] |= MII_SR_LINK_STATUS; > > } > > > > +static bool > > +have_autoneg(E1000State *s) > > +{ > > + return (s->compat_flags & E1000_FLAG_AUTONEG) && > > + (s->phy_reg[PHY_CTRL] & MII_CR_AUTO_NEG_EN) && > > + (s->phy_reg[PHY_CTRL] & MII_CR_RESTART_AUTO_NEG); > > +} > > + > > static void > > set_phy_ctrl(E1000State *s, int index, uint16_t val) > > { > > @@ -194,13 +202,14 @@ set_phy_ctrl(E1000State *s, int index, uint16_t val) > > * migrate during auto negotiation, after migration the link will be > > * down. > > */ > > - if (!(s->compat_flags & E1000_FLAG_AUTONEG)) { > > - return; > > - } > > - if ((val & MII_CR_AUTO_NEG_EN) && (val & MII_CR_RESTART_AUTO_NEG)) { > > + > > + s->phy_reg[PHY_CTRL] = val; > > + > > + if (have_autoneg(s)) { > > e1000_link_down(s); > > DBGOUT(PHY, "Start link auto negotiation\n"); > > - timer_mod(s->autoneg_timer, qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + 500); > > + timer_mod(s->autoneg_timer, > > + qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + 500); > > } > > } > > > > @@ -848,14 +857,6 @@ receive_filter(E1000State *s, const uint8_t *buf, int size) > > return 0; > > } > > > > -static bool > > -have_autoneg(E1000State *s) > > -{ > > - return (s->compat_flags & E1000_FLAG_AUTONEG) && > > - (s->phy_reg[PHY_CTRL] & MII_CR_AUTO_NEG_EN) && > > - (s->phy_reg[PHY_CTRL] & MII_CR_RESTART_AUTO_NEG); > > -} > > - > > static void > > e1000_set_link_status(NetClientState *nc) > > { > > -- > > 1.9.3
On Mon, Jun 30, 2014 at 02:12:51PM -0400, Gabriel L. Somlo wrote: > On Mon, Jun 30, 2014 at 09:04:12PM +0300, Michael S. Tsirkin wrote: > > On Mon, Jun 30, 2014 at 12:55:49PM -0400, Gabriel L. Somlo wrote: > > > set_phy_ctrl() runs the same autonegotiation availability checks > > > that were factored out into have_autoneg() in an earlier commit > > > (d7a4155265416a1c8f3067b59e68bf5fda1d6215), except it runs them > > > on the value about to be written into the phy_ctrl register, and > > > then does not actually write the register. > > > > > > This patch moves the have_autoneg() function in front of > > > set_phy_ctrl(), then updates the latter to actually write > > > the register with the given value, and use the factored-out > > > check to determine whether the link should be bounced. > > > > > > Also, fix indentation/line width when setting up the timer. > > > > > > Signed-off-by: Gabriel Somlo <somlo@cmu.edu> > > > > Weird. > > So does this mean have_autoneg is always false then? > > How does it work? Oh, I think I know what's going on now: In set_mdic() we have: ... } else { if (addr < NPHYWRITEOPS && phyreg_writeops[addr]) { phyreg_writeops[addr](s, index, data); } s->phy_reg[addr] = data; } ... Basically, phyreg_writeops[PHY_CTRL] == set_phy_ctrl() is the only one that applies, but it's like a "pre-write" writeop, the write happens *after* the write_op completes :) :) Fun stuff... Leaving it as-is is probably OK, although it does indeed not handle SC bits properly. I could rewrite have_autoneg() to accept the *value* of the phy control register, then we can use it from set_phy_ctrl() before the register is actually written. Not sure about whether that would be more or less ugly than the current form :) Thanks, --Gabriel
Sorry for the repeated self-replies, I appear to be thinking in very small increments today :) On Mon, Jun 30, 2014 at 03:29:57PM -0400, Gabriel L. Somlo wrote: > > > Weird. > > > So does this mean have_autoneg is always false then? > > > How does it work? > > Oh, I think I know what's going on now: > > In set_mdic() we have: > > ... > } else { > if (addr < NPHYWRITEOPS && phyreg_writeops[addr]) { > phyreg_writeops[addr](s, index, data); > } > s->phy_reg[addr] = data; > } > ... Of course, we could change this to ... } else { if (addr < NPHYWRITEOPS && phyreg_writeops[addr]) { phyreg_writeops[addr](s, index, data); } else { s->phy_reg[addr] = data; } } and have phy_set_ctrl() be a proper write_op, expected to actually write the data as well as do other things. That would at least be a step toward the principle of least WTF, IMHO :) I'll send out an updated patch shortly... Thanks, --Gabriel > ... > > Basically, phyreg_writeops[PHY_CTRL] == set_phy_ctrl() is the only > one that applies, but it's like a "pre-write" writeop, the write > happens *after* the write_op completes :) :) Fun stuff... > > Leaving it as-is is probably OK, although it does indeed not handle > SC bits properly. > > I could rewrite have_autoneg() to accept the *value* of the phy control > register, then we can use it from set_phy_ctrl() before the register > is actually written. Not sure about whether that would be more or less > ugly than the current form :) > > Thanks, > --Gabriel
diff --git a/hw/net/e1000.c b/hw/net/e1000.c index 0fc29a0..2376910 100644 --- a/hw/net/e1000.c +++ b/hw/net/e1000.c @@ -186,6 +186,14 @@ e1000_link_up(E1000State *s) s->phy_reg[PHY_STATUS] |= MII_SR_LINK_STATUS; } +static bool +have_autoneg(E1000State *s) +{ + return (s->compat_flags & E1000_FLAG_AUTONEG) && + (s->phy_reg[PHY_CTRL] & MII_CR_AUTO_NEG_EN) && + (s->phy_reg[PHY_CTRL] & MII_CR_RESTART_AUTO_NEG); +} + static void set_phy_ctrl(E1000State *s, int index, uint16_t val) { @@ -194,13 +202,14 @@ set_phy_ctrl(E1000State *s, int index, uint16_t val) * migrate during auto negotiation, after migration the link will be * down. */ - if (!(s->compat_flags & E1000_FLAG_AUTONEG)) { - return; - } - if ((val & MII_CR_AUTO_NEG_EN) && (val & MII_CR_RESTART_AUTO_NEG)) { + + s->phy_reg[PHY_CTRL] = val; + + if (have_autoneg(s)) { e1000_link_down(s); DBGOUT(PHY, "Start link auto negotiation\n"); - timer_mod(s->autoneg_timer, qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + 500); + timer_mod(s->autoneg_timer, + qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + 500); } } @@ -848,14 +857,6 @@ receive_filter(E1000State *s, const uint8_t *buf, int size) return 0; } -static bool -have_autoneg(E1000State *s) -{ - return (s->compat_flags & E1000_FLAG_AUTONEG) && - (s->phy_reg[PHY_CTRL] & MII_CR_AUTO_NEG_EN) && - (s->phy_reg[PHY_CTRL] & MII_CR_RESTART_AUTO_NEG); -} - static void e1000_set_link_status(NetClientState *nc) {
set_phy_ctrl() runs the same autonegotiation availability checks that were factored out into have_autoneg() in an earlier commit (d7a4155265416a1c8f3067b59e68bf5fda1d6215), except it runs them on the value about to be written into the phy_ctrl register, and then does not actually write the register. This patch moves the have_autoneg() function in front of set_phy_ctrl(), then updates the latter to actually write the register with the given value, and use the factored-out check to determine whether the link should be bounced. Also, fix indentation/line width when setting up the timer. Signed-off-by: Gabriel Somlo <somlo@cmu.edu> --- I'm not 100% sure whether "val" in set_phy_ctrl() is not written to phy_reg[PHY_CTRL] by mistake, or whether there's an ulterior motive (in which case we should add a comment as to why that is). Assuming it's a mistake, and we *should* have been writing it (as per my patch below), we can then use have_autoneg() to also simplify the code significantly :) Thanks, Gabriel hw/net/e1000.c | 27 ++++++++++++++------------- 1 file changed, 14 insertions(+), 13 deletions(-)