Message ID | 4966DABD.1000203@pobox.com |
---|---|
State | Not Applicable, archived |
Delegated to: | David Miller |
Headers | show |
On Fri, Jan 09, 2009 at 12:03:57AM -0500, Jeff Garzik wrote: > > I think you misunderstand. You don't have touch any drivers at all... > see attached demonstration patch. > > The more general point is that it is silly to add two ethtool ioctls > each time you want to twiddle a single boolean flag (whatever that flag > may be, generic or driver-specific or whatnot). > > If you still desire separation from ->{get,set}_flags() ops, then at > least create an ETHTOOL_[GS]STACK_FLAGS. OK, however I'm still not convinced that this is a good idea. First of all we don't have a shortage in the ethtool name space, we've only used up two hex digits worth of a 32-bit integer field. More importantly, making multiple bit changes at the same time may create semantic nightmares in future. For example, imagine if we started out with this generic flag function and TX checksum offload, SG, TSO were done using it. Now the user issues a request changing all of these bits, we'd then have to either validate it for contradictory settings, or devise some ad-hoc ordering in which the settings are applied. This just seems to be unnecessary penny-pinching that doesn't save much but may end up costing us down the road. Cheers,
Herbert Xu wrote: > On Fri, Jan 09, 2009 at 12:03:57AM -0500, Jeff Garzik wrote: >> I think you misunderstand. You don't have touch any drivers at all... >> see attached demonstration patch. >> >> The more general point is that it is silly to add two ethtool ioctls >> each time you want to twiddle a single boolean flag (whatever that flag >> may be, generic or driver-specific or whatnot). >> >> If you still desire separation from ->{get,set}_flags() ops, then at >> least create an ETHTOOL_[GS]STACK_FLAGS. > > OK, however I'm still not convinced that this is a good idea. > First of all we don't have a shortage in the ethtool name space, > we've only used up two hex digits worth of a 32-bit integer field. > > More importantly, making multiple bit changes at the same time > may create semantic nightmares in future. > > For example, imagine if we started out with this generic flag > function and TX checksum offload, SG, TSO were done using it. > Now the user issues a request changing all of these bits, we'd > then have to either validate it for contradictory settings, or > devise some ad-hoc ordering in which the settings are applied. Actually it's just the opposite -- _the_ most common complaint from users and driver developers of the ethtool interface, over the years, has been that there is no way to collect all the modifications and then commit it to hardware all in one go. Each new ethtool command added often winds up pausing and resetting the hardware completely, and ETHTOOL_SGRO is no exception. Now consider what happens when you have a lot of settings to tweak... reset, reset, reset, reset, reset. That's a lot of needless hardware banging. If I could redesign everything from scratch, I would make the interface flexible enough to input a list of changed settings, and then commit those all at once. Lists of flags make that possible to a limited extent, at least. > This just seems to be unnecessary penny-pinching that doesn't > save much but may end up costing us down the road. This sort of interface was added because the current "multi-reset" interface has already cost us. A multi-flag interface is also more flexible and future-proof, as you can deploy new flags without having to update any userland utility... ethtool can permit input of the raw value as well as a more refined command interface. That's another facet of the driver-private flag interfaces, too. The ETHTOOL_[GS]PFLAGS interface is flexible enough to permit a wide variety of feature flags to be exported by various drivers; userland ethtool automatically supports all flags a driver supports, and never needs updating even if the list of supported driver-private flags change. Jeff -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Jan 09, 2009 at 12:30:21AM -0500, Jeff Garzik wrote: > > Actually it's just the opposite -- _the_ most common complaint from > users and driver developers of the ethtool interface, over the years, > has been that there is no way to collect all the modifications and then > commit it to hardware all in one go. Yes that's a problem for flags that require the drivers to reset itself. > Each new ethtool command added often winds up pausing and resetting the > hardware completely, and ETHTOOL_SGRO is no exception. But as I explained before, GRO (like GSO) is purely a software setting, it has nothing to do with the driver at all. In other words we don't need the driver to reset or do anything. If anything by going into the driver's set_flags function as you suggested may cause a spurious reset that wouldn't have happened otherwise. So for software flags like GSO/GRO at least, I don't see any benefit in going to a multi-bit interface. On the flip side, I see potential complications with a multi-bit interfaces that simply don't exist with a single-bit interface. Cheers,
Herbert Xu wrote: > On Fri, Jan 09, 2009 at 12:30:21AM -0500, Jeff Garzik wrote: >> Actually it's just the opposite -- _the_ most common complaint from >> users and driver developers of the ethtool interface, over the years, >> has been that there is no way to collect all the modifications and then >> commit it to hardware all in one go. > > Yes that's a problem for flags that require the drivers to reset > itself. > >> Each new ethtool command added often winds up pausing and resetting the >> hardware completely, and ETHTOOL_SGRO is no exception. > > But as I explained before, GRO (like GSO) is purely a software > setting, it has nothing to do with the driver at all. In other Not quite true... it touches the driver's rx-csum hook. > If anything by going into the driver's set_flags function as you > suggested may cause a spurious reset that wouldn't have happened > otherwise. > > So for software flags like GSO/GRO at least, I don't see any > benefit in going to a multi-bit interface. On the flip side, > I see potential complications with a multi-bit interfaces that > simply don't exist with a single-bit interface. Well, whichever. Overall, if [GS]GRO remains I am happy to take patches supporting it in the userspace utility... Jeff -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Jan 09, 2009 at 01:28:22AM -0500, Jeff Garzik wrote: > > Not quite true... it touches the driver's rx-csum hook. Only because we never put the RX offload flag in a place that's independent of the driver. Also GRO never sets the RX offload flag (or clears it). > Well, whichever. Overall, if [GS]GRO remains I am happy to take patches > supporting it in the userspace utility... Thanks,
diff --git a/include/linux/ethtool.h b/include/linux/ethtool.h index 27c67a5..75fab70 100644 --- a/include/linux/ethtool.h +++ b/include/linux/ethtool.h @@ -285,6 +285,7 @@ struct ethtool_perm_addr { */ enum ethtool_flags { ETH_FLAG_LRO = (1 << 15), /* LRO is enabled */ + ETH_FLAG_GRO = (1 << 14), /* GRO is enabled */ }; struct ethtool_rxnfc { diff --git a/net/core/ethtool.c b/net/core/ethtool.c index 947710a..a114afa 100644 --- a/net/core/ethtool.c +++ b/net/core/ethtool.c @@ -864,6 +864,60 @@ static int ethtool_set_value(struct net_device *dev, char __user *useraddr, return actor(dev, edata.data); } +static int ethtool_get_generic_flags(struct net_device *dev, u32 *val_out) +{ + if (dev->features & NETIF_F_GRO) + *val_out |= ETH_FLAG_GRO; + + return 0; +} + +static int ethtool_set_generic_flags(struct net_device *dev, u32 val_in) +{ + if (val_in & ETH_FLAG_GRO) { + if (!dev->ethtool_ops->get_rx_csum || + !dev->ethtool_ops->get_rx_csum(dev)) + return -EINVAL; + dev->features |= NETIF_F_GRO; + } else + dev->features &= ~NETIF_F_GRO; + + return 0; +} + +static int ethtool_get_flags(struct net_device *dev, char __user *useraddr) +{ + struct ethtool_value edata = { ETHTOOL_GFLAGS }; + int rc; + + if (dev->ethtool_ops->get_flags) + edata.data = dev->ethtool_ops->get_flags(dev); + + rc = ethtool_get_generic_flags(dev, &edata.data); + if (rc) + return rc; + + if (copy_to_user(useraddr, &edata, sizeof(edata))) + return -EFAULT; + return 0; +} + +static int ethtool_set_flags(struct net_device *dev, char __user *useraddr) +{ + struct ethtool_value edata; + int rc; + + if (copy_from_user(&edata, useraddr, sizeof(edata))) + return -EFAULT; + + rc = ethtool_set_generic_flags(dev, edata.data); + + if (rc == 0 && dev->ethtool_ops->set_flags) + rc = dev->ethtool_ops->set_flags(dev, edata.data); + + return rc; +} + /* The main entry point in this file. Called from net/core/dev.c */ int dev_ethtool(struct net *net, struct ifreq *ifr) @@ -1036,12 +1090,10 @@ int dev_ethtool(struct net *net, struct ifreq *ifr) rc = ethtool_set_gso(dev, useraddr); break; case ETHTOOL_GFLAGS: - rc = ethtool_get_value(dev, useraddr, ethcmd, - dev->ethtool_ops->get_flags); + rc = ethtool_get_flags(dev, useraddr); break; case ETHTOOL_SFLAGS: - rc = ethtool_set_value(dev, useraddr, - dev->ethtool_ops->set_flags); + rc = ethtool_set_flags(dev, useraddr); break; case ETHTOOL_GPFLAGS: rc = ethtool_get_value(dev, useraddr, ethcmd,