On 10/09/18 12:51, Willy Tarreau wrote: > On Tue, Oct 09, 2018 at 12:19:04PM -0700, H. Peter Anvin wrote: >> [Resending to a wider audience] >> >> In trying to get the termios2 interface actually implemented in glibc, >> the question came up if we will ever care about baud rates in excess of >> 4 Gbps, even in the relatively remote future. >> >> If this is something we care about *at all*, I would like to suggest >> that rather than defining yet another kernel interface, we steal some >> bits from the MSB of the speed fields, alternatively one of the c_cc >> bytes (all likearchitectures seem to have c_cc[18] free) or some field, >> if we can find them, in c_cflags, to indicate an exponent. >> >> With 5 bits from the top of the speed fields, the current values would >> be identical up to 248 Gbps, and values up to ~288 Pbps would be >> encodable ±2 ppb. >> >> In the short term, all we would have to do in the kernel would be >> erroring out on baud rates higher than 0x0fffffff (2^28-1 due to >> implicit one aliasing rhe first bit of a 5-bit exponent - less than 2^27 >> are functionally denorms.) However, I'd like to put the glibc >> infrastructure for this now if this is something we may ever be >> interested in. >> >> Thoughts? > > Just my two cents, maybe we can conclude that for now we don't care > thus don't implement anything, but that everything you identified as > a possible place to steal bits should be marked "reserved for future > use, must be sent as zero". This will leave you ample room later to > decide how to proceed (and maybe it will not be the bps that you'll > want to change but the number of lanes, or word size, or bit encoding, > especially at 4 Gbps). > Well, it would be nice to be able to pre-enable it in glibc as much as possible. What I'm thinking of doing is to use a 64-bit "baud_t" type in glibc, and reserve the upper 4 bits of the speed field as must be zero (which is de facto the case anyway.) In other to avoid a *huge* user space ABI versioning mess we need to be able to encode the baud rate inside a 32-bit speed_t in glibc, and given that I believe it would be a Very Nice Thing if we could squeeze the information into 32 bits on the kernel side as well. So reserving the upper 4 bits I think is The Right Thing. I think that is actually a null change. I'm not sure if it would be a good idea to make the kernel -EINVAL on currently-unused c_cc bytes or c_*flags; I can see pros and cons (the latter being in no small part that that is not legacy behavior.) -hpa
