On Thu, Sep 9, 2010 at 9:38 AM, Johannes Berg <johannes@xxxxxxxxxxxxxxxx> wrote: > > On Wed, 8 Sep 2010 16:25:25 -0700, Steve deRosier <steve@xxxxxxxxxxx> > wrote: >> mac80211 requires that the MAC address be known and set before calling >> ieee80211_register_hw(). If this isn't done, we see bad MAC addresses >> in our packet headers. In order to make this happen, I had to > restructure >> to have if_sdio_probe load the firmware and get the hardware specs. >> >> I had to add a if_sdio_update_hw_spec function as if_sdio can't use the >> standard >> command as several required variables aren't setup yet. >> if_sdio_update_hw_spec essentially uses polled io to get the hw spec >> command response from the card. > > You should probably keep this for development purposes, but it will break > if your code is built into the kernel. The working model we've adopted to > solve this is to use request_firmware_nowait() in probe, and then only > continue doing work when the firmware loading returns (and then register > with mac80211 etc) > Johannes, Before doing any of this stuff (back in patch 1), my driver goes through and loads the firmware using the request_firmware() calls anyway. In other words, I haven't changed that part. Is the problem specifically with my "get the MAC from the hardware before setting up mac80211" change, or with waiting on firmware loading? For fun, I just tried compiling the driver into the kernel. Mixed results: with two cards on the system, the first card found failed on the call to request_firmware() with a -ENOENT, while the second card came up just fine. When done as modules, both work fine. So there's clearly something to what you're saying. Is it that the filesystem is not up and available yet when we try to load the firmware for the first card? >From looking through the drivers, iwl and ar9170_usb are the only ones to use the request_firmware_nowait(). I don't know if those are the only mac80211 cards with firmware or not everyone's following the model. Do you have specific pointers of which code I should look at as a model, or a better way to solve the problem? In the short term, can we limit libertas_tf to build as module only via Kconfig, and I can add the request_firmware_nowait() as a new feature later with a new patch? Looking at the examples, it looks like the changes for this would be fairly involved and I would prefer to break such a change out separately. I'd like to get libertas_tf_sdio accepted as a base so I can then do smaller change sets going forward. Is that a reasonable plan or just plain silly? BTW, thanks for pointing out this problem. I hadn't tried this built into the kernel and it never occurred to me that this would be an issue. Thanks, - Steve -- To unsubscribe from this list: send the line "unsubscribe linux-wireless" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
