APACHE 2.x ROADMAP ================== Last modified at [$Date$] WORKS IN PROGRESS ----------------- * Source code should follow style guidelines. OK, we all agree pretty code is good. Probably best to clean this up by hand immediately upon branching a 2.1 tree. Status: Justin volunteers to hand-edit the entire source tree ;) Justin says: Recall when the release plan for 2.0 was written: Absolute Enforcement of an "Apache Style" for code. Watch this slip into 3.0. David says: The style guide needs to be reviewed before this can be done. http://httpd.apache.org/dev/styleguide.html The current file is dated April 20th 1998! OtherBill offers: It's survived since '98 because it's welldone :-) Suggest we simply follow whatever is documented in styleguide.html as we branch the next tree. Really sort of straightforward, if you dislike a bit within that doc, bring it up on the dev@httpd list prior to the next branch. So Bill sums up ... let's get the code cleaned up in CVS head. Remember, it just takes cvs diff -b (that is, --ignore-space-change) to see the code changes and ignore that cruft. Get editing Justin :) * revamp the input filter syntax to provide for ordering of filters created with the Set{Input|Output}Filter and the Add{Input|Output}Filter directives. A 'relative to filterx' syntax is definately preferable. * Platforms that do not support fork (primarily Win32 and AS/400) Architect start-up code that avoids initializing all the modules in the parent process on platforms that do not support fork. . Better yet - not only inform the startup of which phase it's in, but allow the parent 'process' to initialize shared memory, etc, and create a module-by-module stream to pass to the child, so the parent can actually arbitrate the important stuff. * Replace stat [deferred open] with open/fstat in directory_walk. Justin, Ian, OtherBill all interested in this. Implies setting up the apr_file_t member in request_rec, and having all modules use that file, and allow the cleanup to close it [if it isn't a shared, cached file handle.] * The Async Apache Server implemented in terms of APR. [Bill Stoddard's pet project.] Message-ID: <008301c17d42$9b446970$01000100@sashimi> (dev@apr) OtherBill notes that this can proceed in two parts... Async accept, setup, and tear-down of the request e.g. dealing with the incoming request headers, prior to dispatching the request to a thread for processing. This doesn't need to wait for a 2.x/3.0 bump. Async delegation of the entire request processing chain Too many handlers use stack storage and presume it is available for the life of the request, so a complete async implementation would need to happen 3.0 release. Brian notes that async writes will provide a bigger scalability win than async reads for most servers. We may want to try a hybrid sync-read/async-write MPM as a next step. This should be relatively easy to build: start with the current worker or leader/followers model, but hand off each response brigade to a "completion thread" that multiplexes writes on many connections, so that the worker thread doesn't have to wait around for the sendfile to complete. * Add a string "class" that combines a char* with a length and a reference count. This will help reduce the number of strlen and strdup operations during request processing. Including both the length and allocation will save us a ton of reallocation we do today, in terms of string manipulation. OtherBill asks if this is really an APR issue, not an HTTPD issue? Brian notes that the performance optimization work in 2.0 has all but eliminated the original motiviation for this idea. The httpd doesn't spend that much time in strlen calls any more. MAKING APACHE REPOSITORY-AGNOSTIC (or: remove knowledge of the filesystem) [ 2002/10/01: discussion in progress on items below; this isn't planned yet ] * dav_resource concept for an HTTP resource ("ap_resource") * r->filename, r->canonical_filename, r->finfo need to disappear. All users need to use new APIs on the ap_resource object. (backwards compat: today, when this occurs with mod_dav and a custom backend, the above items refer to the topmost directory mapped by a location; e.g. docroot) Need to preserve a 'filename'-like string for mime-by-name sorts of operations. But this only needs to be the name itself and not a full path. Justin: Can we leverage the path info, or do we not trust the user? gstein: well, it isn't the "path info", but the actual URI of the resource. And of course we trust the user... that is the resource they requested. dav_resource->uri is the field you want. path_info might still exist, but that portion might be related to the CGI concept of "path translated" or some other further resolution. To continue, I would suggest that "path translated" and having *any* path info is Badness. It means that you did not fully resolve a resource for the given URI. The "abs_path" in a URI identifies a resource, and that should get fully resolved. None of this "resolve to and then we have a magical second resolution (inside the CGI script)" or somesuch. Justin: Well, let's consider mod_mbox for a second. It is sort of a virtual filesystem in its own right - as it introduces it's own notion of a URI space, but it is intrinsically tied to the filesystem to do the lookups. But, for the portion that isn't resolved on the file system, it has its own addressing scheme. Do we need the ability to layer resolution? * The translate_name hook goes away Wrowe altogether disagrees. translate_name today even operates on URIs ... this mechansim needs to be preserved. * The doc for map_to_storage is totally opaque to me. It has something to do with filesystems, but it also talks about security and per_dir_config and other stuff. I presume something needs to happen there -- at least better doc. Wrowe agrees and will write it up. * The directory_walk concept disappears. All configuration is tagged to Locations. The "mod_filesystem" module might have some internal concept of the same config appearing in multiple places, but that is handled internally rather than by Apache core. Wrowe suggests this is wrong, instead it's private to filesystem requests, and is already invoked from map_to_storage, not the core handler. and blocks are preserved as-is, but sections become specific to the filesystem handler alone. Because alternate filesystem schemes could be loaded, this should be exposed, from the core, for other file-based stores to share. Consider an archive store where the layers become -> -> Justin: How do we map Directory entries to Locations? * The "Location tree" is an in-memory representation of the URL namespace. Nodes of the tree have configuration specific to that location in the namespace. Something like: typedef struct { const char *name; /* name of this node relative to parent */ struct ap_conf_vector_t *locn_config; apr_hash_t *children; /* NULL if no child configs */ } ap_locn_node; The following config: SetHandler server-status Order deny,allow Deny from all Allow from 127.0.0.1 Creates a node with name=="server_status", and the node is a child of the "/" node. (hmm. node->name is redundant with the hash key; maybe drop node->name) In the config vector, mod_access has stored its Order, Deny, and Allow configs. mod_core has stored the SetHandler. During the Location walk, we merge the config vectors normally. Note that an Alias simply associates a filesystem path (in mod_filesystem) with that Location in the tree. Merging continues with child locations, but a merge is never done through filesystem locations. Config on a specific subdir needs to be mapped back into the corresponding point in the Location tree for proper merging. * Config is parsed into a tree, as we did for the 2.0 timeframe, but that tree is just a representation of the config (for multiple runs and for in-memory manipulation and usage). It is unrelated to the "Location tree". * Calls to apr_file_io functions generally need to be replaced with operations against the ap_resource. For example, rather than calling apr_dir_open/read/close(), a caller uses resource->repos->get_children() or somesuch. Note that things like mod_dir, mod_autoindex, and mod_negotation need to be converted to use these mechanisms so that their functions will work on logical repositories rather than just filesystems. * How do we handle CGI scripts? Especially when the resource may not be backed by a file? Ideally, we should be able to come up with some mechanism to allow CGIs to work in a repository-independent manner. - Writing the virtual data as a file and then executing it? - Can a shell be executed in a streamy manner? (Portably?) - Have an 'execute_resource' hook/func that allows the repository to choose its manner - be it exec() or whatever. - Won't this approach lead to duplication of code? Helper fns? gstein: PHP, Perl, and Python scripts are nominally executed by a filter inserted by mod_php/perl/python. I'd suggest that shell/batch scripts are similar. But to ask further: what if it is an executable *program* rather than just a script? Do we yank that out of the repository, drop it onto the filesystem, and run it? eeewwwww... I'll vote -0.9 for CGIs as a filter. Keep 'em handlers. Justin: So, do we give up executing CGIs from virtual repositories? That seems like a sad tradeoff to make. I'd like to have my CGI scripts under DAV (SVN) control. * How do we handle overlaying of Location and Directory entries? Right now, we have a problem when /cgi-bin/ is ScriptAlias'd and mod_dav has control over /. Some people believe that /cgi-bin/ shouldn't be under DAV control, while others do believe it should be. What's the right strategy?