From: David Laight Date: Thu, 26 Mar 2020 11:08:52 +0000 Subject: [Intel-wired-lan] [PATCH net 1/1] e1000e: Stop tx/rx setup spinning for upwards of 300us. In-Reply-To: References: <6ef1e257642743a786c8ddd39645bba3@AcuMS.aculab.com> Message-ID: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: intel-wired-lan@osuosl.org List-ID: From: Jeff Kirsher > Sent: 04 March 2020 18:04 > On Tue, 2020-03-03 at 17:06 +0000, David Laight wrote: > > Instead of spinning waiting for the ME to be idle defer the ring > > tail updates until one of the following: > > - The next update for that ring. > > - The receive frame processing. > > - The next timer tick. > > > > Reduce the delay between checks for the ME being idle from 50us > > to uus. > > > > Part fix for bdc125f7. > > > > Signed-off-by: David Laight > > Added intel-wired-lan at lists.osuosl.org mailing list, so the right > people can review your patch. I don't see any sign of anyone looking at this. Is the code so bad everyone has buried their head in the sand? Am I right in thinking that the actual hardware problem is that PCIe writes are 'posted' in the hardware and then lost if the ME does a write while the PCIe write is still pending? In which case a much simpler patch that does a readback after every write and retries if the value is different will solve the problem without ever needing a delay(). The only 'problem' register would be the interrupt mask which the hardware appears to change itself. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)