On Mon, Sep 13 2021, Timothe Litt via curl-library wrote:
> [[PGP Signed Part:Undecided]] > On 13-Sep-21 07:01, Daniel Stenberg via curl-library wrote: > > # Feedback > > I'm all ears. Especially if you have alternative solutions to suggest or if > you have an opinion on which way to go. > > This is not a problem we must solve *right now*, but I would feel better if > we have an idea about how to address it when we get there. Because I'm > convinced we will reach this point eventually. > > Here's an approach that has some short-term pain, but solves the problem > permanently - including for protocol 65, 129, ... > > Switch to an expandable array of bits, similar to select()'s FDSETS: > > Deprecate the existing CURLOPT_PROTOCOLS and CURLOPT_REDIR_PROTOCOLS, replace > with CURLOPT_PROTOCOLS_EXT and CURLOPT_REDIR_PROTOCOLS_EXT. > > Use indices rather than bitmasks for CURLPROTO_* (e.g. add CURLPROTO_DICTn, > CURLPROTO_FTPn, etc) > > Switch from a long to a pointer to typedef struct { unsigned int size; uint8 > bits[(CURL_PROTO_MAXn + 7)/8]} CURLPROTO. > > Provide some macros along the lines of FD_CLR/FD_ISSET/FD_SET/FD_ZERO, but > instead of the FD_SETSIZE hack, use the 'size' value of the structure, which > will increase every time you add 8 more protocols. But clients compiled > earlier will have a smaller "size", so will not inadvertently enable new > protocols. > > e.g. the user-visible functions might be something like: > > #define CURLPROTO_SET( str, bit ) do { ASSERT((bit) <= CURL_PROTO_MAXn && > (bit) <= (str)->size); (str)->bits[(bit)>>3] |= 1u<<((bit)&7); } while(0) /* > Could also provide a vararg function to set multiple */ > > #define CURLPROTO_ISSET( str, bit ) ( ((bit) > CURL_PROTO_MAXn || (bit) > > (str)->size))? 0 : (str)->bits[(bit)>>3] & 1u<<((bit)&7) ) > > So specifying protocols looks something like: > > CURLPROTO allow = { sizeof( CURLPROTO ) }; /* The initialization could be a > macro - e.g CURL_PROTO_DECL(allow); */ > > CURLPROTO_ZERO(&allow); /* If a stack or malloc()'d variable */ > > CURLPROTO_SET( &allow, CURL_PROTO_FTPn ); > > CURLPROTO_SET(&allow, CURL_PROTO_HTTPSn); > > curl_easy_setopt( handle, CURLOPT_PROTOCOLS_EXT, &allow); > > Internally, use the current (usually larger) size so you don't have to > bounds-check every reference; just memcpy min(libraryMAX, 'size' provided) to > an > internal structure. Convert the deprecated functions to set(or clear) the > first few bits in the internal structure as specified; they should zero all > bits 32+. (Be careful about endianisms.) > > There ought to be a function to return, in the same format, a structure > listing all the protocols implemented by the current library. > > This scheme provides backward compatibility with infinite expandability. > There's some overhead for the client, but these aren't critical path - they're > probably setup once and tested never. In the library, the assertions will > optimize out, and a compiler will optimize the bit references to be no more > expensive than the current bit tests. The compatibility layer is pretty thin > - it probably ends up being a cast & possible byteswap. > > With bit more thought (pun intended), you might be able to avoid introducing > the new CURLPROTO_*n symbols - but at first blush, it seems expensive to do > that while also exposing the existing API. > > Polishing is left as an exercise for the reader... This sounds already off-topic for what Daniel asked about, which is how to do this in a backwards-compatible way. I can't think of a good way to do that off-hand. But just on this: Isn't this a rather elaborate way to do what C gives you these days (including I think, in C89) with a struct where you specify the bits a given unsigned int variable should occupy? I.e.: struct curl_protocols { unsigned int http:1, https:1, ftp:1, ftps:1; }; etc., you can keep adding to that at will, trusting the compiler to expand it for you, and accessing it is going to be: proto->https = !!enabled; I believe that's quite portable, e.g. it's widely used in the git.git codebase, the first occurance made it in-tree in 2006. I think curl's more widely ported than that, so YMMV. -- Unsubscribe: https://lists.haxx.se/listinfo/curl-library Etiquette: https://curl.haxx.se/mail/etiquette.html