net: Add Life of a URLRequest documentation.

net/docs/life-of-a-url-request.md

Index: net/docs/life-of-a-url-request.md
diff --git a/net/docs/life-of-a-url-request.md b/net/docs/life-of-a-url-request.md
new file mode 100644
index 0000000000000000000000000000000000000000..6a3b5286ba92726929419c36b0e4dc3f2850cbbd
--- /dev/null
+++ b/net/docs/life-of-a-url-request.md
@@ -0,0 +1,480 @@
+# Life of a URLRequest

[Randy Smith (Not in Mondays), 2015/07/08 20:36:25]

Could you add this doc (and maybe the other docs in this directory, if you are
so moved) into net/net.gypi:net_docs_sources so that HTML gets automatically
created in people's output directories?  Arguably we don't need to do this
(because source view auto-translates) but it does make reviewing easier.  I also
believe in consistency--i.e. we should do it for all docs or none.

[mmenke, 2015/07/09 19:38:28]

Done.  I was unaware of that target...and suspect almost everyone else is, too,
which may be a problem.

[Randy Smith (Not in Mondays), 2015/07/13 15:39:30]

Agreed; not sure what to do about it, though.

+
+This document is intended as an overview of the core layers of the network
+stack, their basic responsibilities, how they fit together, and where some of
+the pain points are, without going into too much detail. Though it touches a
+bit on the renderer process and the content/loader stack, the focus is on net/
+itself.
+
+It's particularly targeted at people new to the Chrome network stack, but
+should also be useful for team members who may be experts at some parts of the
+stack, but are largely unfamiliar with other components. It starts by walking
+through how a basic request issued by Blink works its way through the network
+stack, and then moves on to discuss how various components plug in.
+
+
+# Anatomy of the Network Stack
+
+The main top-level network stack object is the URLRequextContext. The context
+has non-owning pointers to everything needed to create and issue a URLRequest.
+The context must outlive all requests that use it. Creating a context is a
+rather complicated process, and it's recommended that most embedders use
+URLRequestContextBuilder to do this. Chrome itself has several request
+contexts that the network stack team owns:
+
+* The proxy URLRequestContext, owned by the IOThread and used to get PAC
+scripts while avoiding re-entrancy.
+* The system URLRequestContext, also owned by the IOThread, used for requests
+that aren't associated with a profile.
+* Each profile, including incognito profiles, has a number of URLRequestContexts
+that are created as needed:
+ * The main URLRequestContext is mostly created in ProfileIOData, though it

[Randy Smith (Not in Mondays), 2015/07/08 20:36:25]

These sub-items aren't showing up as indented/second level list in the HTML.

[mmenke, 2015/07/09 19:38:29]

Fixed.  Weird, they're fine when I paste them into snippets (See earlier comment
about not knowing about net_docs target).  Can't we just standardize on one
version of markdown for all of Google?  :(

Looks like this version of markdown needs 4 space indent (I meant to use two,
but when I switched to 1 space between sentences, accidentally reduced these to
1 space, too).

+ has a couple components that are passed in from content's StoragePartition
+ code. Several other components are shared with the system URLRequestContext,
+ like the HostResolver.
+ * Each non-incognito profile also has a media request context, which uses a
+ different on-disk cache than the main request context. This prevents a single
+ huge media file from evicting everything else in the cache.
+ * On desktop platforms, each profile has a request context for extensions.
+ * Each profile has two contexts for each isolated app (One for media, one for
+ everything else).
+
+The "HttpNetworkSession" is another major network stack object. It has
+pointers to the network stack objects that more directly deal with sockets, and
+their dependendencies. Its main objects are the HttpStreamFactory, the socket
+pools, and the SPDY/QUIC session pools.
+
+This document does not mention either of these objects much, but at layers
+above the HttpStreamFactory, objects often grab their dependencies from the
+URLRequestContext, while the HttpStreamFactory and layers below it generally
+get their dependencies from the HttpNetworkSession.
+
+
+# How many "Delegates"?
+
+The network stack informs the embedder of important events for a request using
+two main interfaces: The URLRequest::Delegate interface and the NetworkDelegate
+interface.
+
+The URLRequest::Delegate interface consists of small set callbacks needed to let
+the embedder drive a request forward. URLRequest::Delegates generally own the
+URLRequest.
+
+The NetworkDelegate is geerally a single object shared by all requests, and
+consists includes callbacks corresponding to most of the URLRequest::Delegate's
+callbacks, as well as an assortment of other methods. The NetworkDelegate is
+optional, the URLRequest::Delegate is not.
+
+
+# Life of a "Simple" URLRequest
+
+Consider a simple request issued by the renderer process. Suppose it's an HTTP
+request, the response is uncompressed, has no request body (i.e. is not an
+upload), and no matching entry in the cache.
+
+
+## Overview

[Randy Smith (Not in Mondays), 2015/07/08 20:36:25]

I'd go for a (short!) paragraph before you dive into the bulleted list giving an
even more abstract description for a request.  Maybe something like:

"A request for data is normally dispatched from the renderer to the browser
process.  There a URLRequest is created to represent it.  An job specific to the
protocol (e.g. HTTP) is attached to the request.  That job first checks the
cache then establishes a network connection object to actually fetch the data. 
That connection object interacts with network socket pools to potentially re-use
sockets; the socket pools create and connect a socket if there is no appropriate
existing socket.  Once that socket exists, the HTTP request is dispatched, the
response read and parsed, and the result returned back up the stack and sent
over to the renderer.

Of course, it's never quite that simple :-}."

And then dive into the details.  That'll give the reader a bit more
understanding of the motivation for the different levels, and some sense of not
being completely lost as they wander through those levels.

+
+### Request Starts
+
+* ResourceDispatcher to creates an IPCResourceLoaderBridge.
+* The IPCResourceLoaderBridge asks ResourceDispatcher to start the request.
+* ResourceDispatcher sends an IPC to the ResourceDispatcherHost in the browser

[Randy Smith (Not in Mondays), 2015/07/08 20:36:24]

I'd add a line with some visual distinction after this one to call out the
transition to the browser process.

[mmenke, 2015/07/09 19:38:28]

I split up the section.  Let me know if you think I should do more.

+process.
+* ResourceDispatcherHost uses the URLRequestContext to create the URLRequest.
+* ResourceDispatcherHost creates a ResourceLoader and ResourceHandler chain to

[Randy Smith (Not in Mondays), 2015/07/08 20:36:24]

In this line and the next one, I'd refer to the "URLRequest" rather than the
"Request" to make clear the top-level concept versus lower level object
distinction.

[mmenke, 2015/07/09 19:38:28]

Done.

+manage the request.
+* ResourceLoader starts the request.
+
+### Request is Issued

[Randy Smith (Not in Mondays), 2015/07/08 20:36:24]

This section is intimidating in the number of bullet items it has just hammering
at the reader.  (At least, it's intimidating to me, and I've repeatedly created
this same bulleted list for my own education.)  I'd break this into
sub-sections, maybe URLRequest->HttpNetworkTransaction,
HTTPNetworkTransaction->TransportClientSocketPool,
TransportClientSocketPool->HttpStreamFactoryImpl::Job/HttpBasicStream.  This
might be done with nesting, to make unwinding easier, but if there's a clear way
to have it in separate sections, that's somewhat better.

[Randy Smith (Not in Mondays), 2015/07/08 20:36:25]

I'd put some verbiage here indicating that what's below is the common case, but
there may be variations at each level, usually at the "asks the X to create a Y,
which is a Z" that sometimes it's not a Z but some other subclass of Y.

[mmenke, 2015/07/09 19:38:29]

I don't want to hammer everywhere that this is a "simple" URLRequest.  I've
rearranged things slightly, tell me what you think.

+
+* The URLRequest asks the URLRequestJobFactory to create a URLRequest[Http]Job.

[Randy Smith (Not in Mondays), 2015/07/08 20:36:24]

Especially if you're targeting this document at folks new to the network stack,
I think "URLRequest[Http]Job" is too abbreviated a way of saying "A
URLRequestJob.  Usually, this will be the URLRequestJob subclass,
URLRequestHttpJob", so I'd spell that out.

[mmenke, 2015/07/09 19:38:28]

Done.

+* The URLRequestHttpJob asks the HttpCache to create an HttpTransaction (always
+an HttpCache::Transaction).
+* The HttpCache::Transaction sees there's no cache entry for the request, and
+creates an HttpNetworkTransaction.
+* The HttpNetworkTransaction calls into the HttpStreamFactory to request an
+HttpStream.
+* HttpStreamFactory creates an HttpStreamFactoryImpl::Job.
+* HttpStreamFactoryImpl::Job calls into the TransportClientSocketPool to
+populate an ClientSocketHandle.
+* TransportClientSocketPool has no idle sockets, so it creates a
+TransportConnectJob and starts it.
+* TransportConnectJob creates a StreamSocket and establishes a connection.
+* TransportClientSocketPool puts the StreamSocket in the ClientSocketHandle,
+and calls into HttpStreamFactoryImpl::Job.
+* HttpStreamFactoryImpl::Job creates an HttpBasicStream, which takes ownership
+of the ClientSocketHandle.
+* It returns the HttpBasicStream to the HttpNetworkTransaction.
+* HttpNetworkTransaction gives the request headers to the HttpBasicStream, and
+tells it to start the request.
+* HttpBasicStream sends the request, and waits for the response.
+* The HttpBasicStream sends the response headers back to the
+HttpNetworkTransaction.
+* Headers are sent up to the URLRequest, to the ResourceLoader, through the
+ResourceHandler stack.
+* They're then send by the AsyncResourceHandler to the ResourceDispatcher.
+
+### Response body is read
+
+* AsyncResourceHandler allocates a 512k ring buffer of shared memory to read
+the body of the request.
+* AsyncResourceHandler tells the ResourceLoader to read the response body to
+the buffer, 32kB at a time.
+* AsyncResourceHandler informs the ResourceDispatcher of each read.
+* ResourceDispatcher tells the AsyncResourceHandler when it's done with the
+data with each read, so it knows when parts of the buffer can be reused.
+
+### URLRequest is Destroyed
+
+* When complete, the RDH deletes the ResourceLoader, which deletes the
+URLRequest.
+* During destruction, the HttpNetworkTransaction determines if the socket is
+reusable, and if so, tells the HttpBasicStream to return it to the socket pool.
+
+## Details

[Randy Smith (Not in Mondays), 2015/07/08 20:36:24]

I'd suggest breaking this section out into subsections; I think it'll be a lot
more readable.  Maybe "Process Architecture" for RD/RDH, "URLRequest Creation
and Plumbing" for URLRequestContext->ResourceLoader/Handler, "Http Transactions"
for the cache and network transactions, and maybe "Socket pools", "Http Parsing
and Data Processing" for the rest.  It would be really useful if this section
was broken out in the same way as the above section (you don't need the labels
on the above section, but if the reader has already bucketed the concepts in a
certain way, they'll read this section more easily).

[mmenke, 2015/07/09 19:38:28]

I've split this into the same sections, with the same names, as the first part,
which I think is a big improvment.  Open to ideas for how to better split
things.

+
+Each child process has at most one ResourceDispatcher, which is responsible for
+all URL request-related communication with the browser process. When something
+in the renderer needs to issue a resource request, it calls into the
+ResourceDispatcher, which returns an IPCResourceLoaderBridge to the caller.
+The caller uses the bridge to start a request. When started, the
+ResourceDispatcher assigns the request a per-renderer ID, and then sends the
+ID, along with all information needed to issue the request, to the
+ResourceDispatcherHost in the browser process.
+
+The ResourceDispatcherHost (RDH), along with most of the network stack, lives
+on the browser process's IO thread. The browser process only has one RDH,
+which is responsible for handling all network requests initiated by
+ResourceDispatchers in all child processes, not just renderer process.
+Browser-initiated don't go through the RDH, with some exceptions.
+
+When the RDH sees the request, it calls into a URLRequestContext to create the
+URLRequest. The URLRequestContext has pointers to all the network stack
+objects needed to issue the request over the network, such as the cache, cookie
+store, and host resolver. The RDH then creates a chain of ResourceHandlers
+each of which can monitor/modify/delay/cancel the URLRequest and the
+information it returns. The only one of these I'll talk about here is the
+AsyncResourceHandler, which is the last ResourceHandler in the chain. The RDH
+then creates a ResourceLoader (Which is the URLRequest::Delegate), passes
+ownership of the URLRequest and the ResourceHandler chain to it, and then starts
+the ResourceLoader.
+
+The ResourceLoader checks that none of the ResourceHandlers want to cancel,
+modify, or delay the request, and then finally starts the URLRequest. The
+URLRequest then calls into the URLRequestJobFactory to create a URLRequestJob
+and then starts it. In the case of an HTTP or HTTPS request, this will be a
+URLRequestHttpJob.
+
+The URLRequestHttpJob calls into the HttpCache to create an
+HttpCache::Transaction. If there's no matching entry in the cache, the
+HttpCache::Transaction will just call into the HttpNetworkLayer to create an
+HttpNetworkTransaction, and transparently wrap it.
+
+The HttpNetworkTransaction calls into the HttpStreamFactory to request an
+HttpStream to the server. The HttpStreamFactoryImpl::Job creates a
+ClientSocketHandle to hold a socket, once connected, and passes it in to the
+ClientSocketPoolManager. The ClientSocketPoolManager assembles the
+TransportSocketParams needed to establish the connection and creates a group
+name ("host:port") used to identify sockets that can be used interchangeably.
+
+The ClientSocketPoolManager directs the request to the
+TransportClientSocketPool, since there's no proxy and it will be using
+HTTP/1.x. The pool sends it on to the
+ClientSocketPoolBase<TransportSocketParams> it wraps, which sends it on to its
+ClientSocketPoolBaseHelper, which actually manages the socket pool. If there
+isn't already an idle connection, and there are available socket slots, the
+ClientSocketPoolBaseHelper will create a new TransportConnectJob using the
+aforementioned params object. The Job will do the actual DNS lookup by calling
+into the HostResolverImpl, if needed, and then finally establish a connection.
+
+When the socket is connected, ownership of the socket is passed to the
+ClientSocketHandle. The HttpStreamFactoryImpl::Job is informed the
+connection attempt succeeded, and it then creates an HttpBasicStream, which
+takes ownership of the ClientSocketHandle. It then passes ownership of the
+HttpBasicStream back to the HttpNetworkTransaction. The Transaction passes
+the request headers to the HttpBasicStream, which uses an HttpStreamParser to
+(finally) format the request headers and send them to the server. The
+HttpStreamParser waits to receive the response and then parses the HTTP/1.x
+response headers, and then passes them up through both Transaction classes
+to the URLRequestHttpJob, which passes them up to the URLRequest and on to
+the ResourceLoader.
+
+The ResourceLoader passes them through the chain of ResourceHandlers, and then
+they make their way to the AsyncResourceHandler. The AsyncResourceHandler uses
+the renderer process ID ("child ID") to figure out which process the request
+was associated with, and then sends the headers along with the request ID to
+that process's ResourceDispatcher. The ResourceDispatcher uses the ID to
+figure out which IPCResourceLoaderBridge the headers should be sent to, which
+sends them on to whatever created the IPCResourceLoaderBridge in the first
+place.
+
+Without waiting to hear back from the ResourceDispatcher, the ResourceLoader
+tells its ResourceHandler chain to allocate memory to receive the response
+body. The AsyncResourceHandler creates a 512KB ring buffer of shared memory,
+and then passes the first 32KB of it to the ResourceLoader for the first read.
+The ResourceLoader then passes a 32KB body read request down through the
+URLRequest all the way down to the HttpResponseParser. Once some data is read,
+possibly less than 32KB, the number of bytes read makes its way back to the
+AsyncResourceHandler, which passes the shared memory buffer and the offset and
+amount of data read to the renderer process.
+
+The AsyncResourceHandler relies on ACKs from the renderer to prevent it from
+overwriting data that the rendererer has yet to consume. This process repeats
+until the response body is completely read. When the URLRequest informs the
+ResourceLoader it's complete, the ResourceLoader tells the ResourceHandlers,
+and the AsyncResourceHandler tells the ResourceDispatcher the request is
+complete. The RDH then deletes ResourceLoader, which deletes the URLRequest.
+
+When the HttpNetworkTransaction is being torn down, it figures out if the
+socket is reusable. If not, it tells the HttpBasicStream to close the socket.
+Either way, the ClientSocketHandle returns the socket is then returned to the
+socket pool, either for reuse or so the socket pool knows it has another free
+socket slot.
+
+
+# Additional Topics
+
+## HTTP Cache
+
+The HttpCache::Transaction sits between the URLRequestHttpJob and the
+HttpNetworkTransaction, and implements the HttpTransaction interface, just like
+the HttpNetworkTransaction. The HttpCache::Transaction checks if a request can
+be served out of the cache. If a request needs to be revalidated, it handles
+sending a 204 revalidation request over the network. It may also break a range
+request into multiple cached and non-cached contiguous chunks, and may issue
+multiple network requests for a single range URLRequest.
+
+One important detail is that it has a read/write lock for each URL. The lock
+technically allows multiple reads at once, but since an HttpCache::Transaction
+always grabs the lock for writing and reading before downgrading it to a read
+only lock, all requests for the same URL are effectively done serially. Blink
+merges requests for the same URL in many cases, which mitigates this problem to
+some extent.
+
+The HttpCache::Transaction uses one of three disk_cache::Backends to actually
+store the cache's index and files: The in memory backend, the blockfile cache
+backend, and the simple cache backend. The first is used in incognito. The
+latter two are both stored on disk, and are used on different platforms.
+
+## Cancellation
+
+A request can be cancelled by the renderer process Blink or by any of the
+ResourceHandlers through the ResourceLoader. When the cancellation message
+reaches the URLRequest, it passes on the fact it's been cancelled back to the
+ResourceLoader, which then sends the message down the ResourceHandler chain.
+
+When an HttpNetworkTransaction for a canelled request is being torn down, it
+figures out if the socket the HttpStream owns can potentially be reused, based
+on the protocol (HTTP / SPDY / QUIC) and any received headers. If the socket
+potentially can be reused, an HttpResponseBodyDrainer is created to try and
+read any remaining body bytes of the HttpStream, if any, before returning the
+socket to the SocketPool. If this takes too long, or there's an error, the
+socket is closed instead. Since this all happens at the layer below the cache,
+any drained bytes are not written to the cache, and as far as the cache layer is
+concerned, it only has a partial response.
+
+## Redirects
+
+The URLRequestHttpJob checks if headers indicate a redirect when it receives
+them from the next layer down (Typically the HttpCache::Transaction). If they
+indicate a redirect, it tells the cache the response is complete, ignoring the
+body, so the cache only has the headers. The cache then treats it as a complete
+entry, even if the headers indicated there will be a body.
+
+The URLRequestHttpJob then checks if the URLRequest if the request should be
+followed. First it checks the scheme. Then it informs the ResourceLoader
+about the redirect, to give it a chance to cancel the request. The information
+makes its way down through the AsyncResourceHandler to the ResourceDispatcher
+and on into Blink, which checks if the redirect should be followed.
+
+The ResourceDispatcher then asynchronously sends a message back to either
+follow the redirect or cancel the request. In either case, the old
+HttpTransaction is destroyed, and the HttpNetworkTransaction attempts to drain
+the socket for reuse, just as in the cancellation case. If the redirect is
+followed, the URLRequest calls into the URLRequestJobFactory to create a new
+URLRequestJob, and then starts it.
+
+## Filters (gzip, SDCH, etc)
+
+When the URLRequestHttpJob receives headers, it sends a list of all Content-
+Encoding values to Filter::Factory, which creates a (possibly empty) chain of
+filters. As body bytes are received, they're passed through the filters at the
+URLRequestJob layer and the decoded bytes are passed back to the embedder.
+
+Since this is done above the cache layer, the cache stores the responses prior
+to decompression. As a result, if files aren't compressed over the wire, they
+aren't compressed in the cache, either. This behavior can also create problems
+when responses are SDCH compressed, as a dictionary may be evicted from the
+cache independently of the response that was compressed with it.

[Randy Smith (Not in Mondays), 2015/07/08 20:36:25]

nit: I'd leave out "from the cache" since SDCH implementation is, at least
currently, partially memory based around SDCH-specific storage.

[mmenke, 2015/07/09 19:38:29]

Reworded it a bit.

+
+TODO(mmenke): Discuss filter creation.
+
+## Socket Pools
+
+The ClientSocketPoolManager is responsible for assembling the parameters needed
+to connect a socket, and then sending the request to the right socket pool.
+Each socket request sent to a socket pool comes with a socket params object, a
+ClientSocketHandle, and a "group name". The params object contains all the
+information a ConnectJob needs to create a connection of a given type, and
+different types of socket pools take different params types. The
+ClientSocketHandle will take temporary ownership of the socket, once connected
+socket, and return it to the socket pool when done. All connections with the
+same group name in the same pool can be used to service the same connection
+request, so it consists of host, port, protocol, and whether "privacy mode" is
+used for requests using the socket or not.
+
+All socket pool classes derive from the ClientSocketPoolBase<SocketParamType>.
+The ClientSocketPoolBase handles managing sockets - which requests to create
+sockets for, which requests get connected sockets first, which sockets belong
+to which groups, connection limits per group, keeping track of and closing idle
+sockets, etc. Each ClientSocketPoolBase subclass has its own ConnectJob type,
+which establishes a connection using the socket params, before the pool hands
+out the connected socket.
+
+## Socket Pool Layering
+
+Some socket pools are layered on top other socket pools. This is done when a
+"socket" in a higher layer needs to establish a connection in a lower level
+pool and then take ownership of it as part of its connection process. See later
+sections for examples. There are a couple additional complexities here.
+
+From the perspective of the lower layer pool, all of its sockets that a higher
+layer pools owns are actively in use, even when the higher layer pool considers
+them idle. As a result, when a lower layer pool is at its connection limit and
+needs to make a new connection, it will ask any higher layer pools pools to
+close an idle connection if they have one, so it can make a new connection.
+
+Sockets in the higher layer pool must have their own distinct group name in the
+lower layer pool as well. This is needed so the lower layer pool won't, for
+example, group SSL and HTTP connections to the same port together.
+
+## SSL
+
+When an SSL connection is needed, the ClientSocketPoolManager assembles the
+parameters needed both to connect the TCP socket and establish an SSL
+connection. It then passes them to the SSLClientSocketPool, which creates
+an SSLConnectJob using them. The SSLConnectJob's first step is to call into the
+TransportSocketPool to establish a TCP connection.
+
+Once a connection is established by the lower layered pool, the SSLConnectJob
+then starts SSL negotiation. Once that's done, the SSL socket is passed back to
+the HttpStreamFactoryImpl::Job that initiated the request, and things proceed
+just as with HTTP. Whe complete, the socket is returned to the
+SSLClientSocketPool.
+
+## Proxy discovery
+
+The first step the HttpStreamFactoryImpl::Job performs, just before calling
+into the ClientSocketPoolManager to create a socket, is to check with the
+ProxyService to see if a proxy is needed for the URL it's been given. The
+ClientSocketPoolManager then uses this information to find the correct proxy
+socket pool to send the request to.
+
+TODO(mmenke): Discuss proxy configurations, WPAD, tracing proxy resolver.
+
+## Proxy Socket Pools
+
+Each SOCKS or HTTP proxy has its own completely independent set of socket
+pools. They have their own exclusive TransportSocketPool, their own protocol-
+specific pool above it, and their own SSLSocketPool above that. HTTPS proxies
+also have a second SSLSocketPool between the the HttpProxyClientSocketPool and
+the TransportSocketPool, since they can talk SSL to both the proxy and the
+destination server, layered on top of each other.
+
+## SPDY
+
+Once an SSL connection is established, the HttpStreamFactoryImpl::Job checks if
+SPDY was negotiated over the socket. If so, it creates a SpdySession using the
+socket, and a SpdyHttpStream. The SpdyHttpStream will be passed to the
+HttpNetworkTransaction, which drives the stream as usual.
+
+The SpdySession will be shared with other Jobs conecting to the same server,
+and future Jobs will find the SpdySession before they try to create a
+connection. HttpServerProperties also tracks which servers supported SPDY when
+we last talked to them. We only try to establish a single connection to servers
+we thing speak SPDY when multiple HttpStreamFactoryImpl::Jobs are trying to
+connect to them, to avoid wasting resources.
+
+## QUIC
+
+HttpServerProperties also tracks which servers have advertised QUIC support in
+the past. If a server hass advertisied QUIC support, a second
+HttpStreamFactoryImpl::Job will be created for SPDY, and will be raced against
+the one for HTTP/HTTPS connection. Whichever connects first will be used.
+Existing QUIC sessions will be reused if available.
+
+TODO(mmenke): Discuss SPDY/QUIC proxies?
+
+## Uploads
+
+Upload data is passed to a URLRequest using the UploadDataStream class. Since
+the over-the-wire format of uploads is determined by the HttpStream type, the
+upload body is read from the stream and prepared to be sent over the write by
+the HttpStream classes (HttpBasicStream, SpdyHttpStream, QuicHttpStream).
+UploadDataStreams have to be replayable, since redirects and retries may need
+to re-upload data.
+
+UploadDataStreams either have a length known in advance, or are "chunked".
+The main implementation for the non-chunked case is ElementsUploadDataStream,
+which consists of one or more UploadElementReader, each of which contains a
+fixed-size chunk of data, either in memory or in a file.
+
+ChunkedUploadDataStream is the main implementation for the chunked case.
+Chunked uploads are only used by Chrome internally, since many servers don't
+support them, and the length is always known in advance for web-initiated
+uploads. Chrome adds data bit by bit, and the HttpStream implementation
+sends data as long as more data is needed and it has more data to send.
+Because of the replayable requirement mentioned above, the entire content of
+these chunked requests must be buffered into memory.
+
+One weirdness is that reads from UploadDataStreams currently aren't allowed to
+fail. If a read from a file fails, then the contents of the file are replaced
+by 0's. Apparently this matched FireFox's behavior at the time of
+implementation.
+
+## Cookies
+
+Cookies are added to a request by the URLRequestHttpJob, and saved at that layer
+as well, once the response headers have been received. The CookieStore (The
+implementation of which is called "CookieMonster") handles storage of cookies,
+and can be used either as an in-memory store or with an on-disk store, backed by
+a SQLitePersistentCookieStore.
+
+The CookieStore is currently reference counted, and outlives the rest of the
+network stack, which has led to some lifetime issues.
+
+## Prioritization
+
+URLRequests are assigned a priority on creation. It only comes into play in
+a couple places:
+
+* DNS lookups are initiated based on the highest priority request for a lookup.
+* Socket pools hand out and create sockets on prioritization. However, idle
+sockets will be assigned to low priority requests in preference to creating new
+sockets for higher priority requests.
+* SPDY and QUIC both support sending priorities over-the-wire.
+
+At the socket pool layer, sockets are only assigned to socket requests once the
+socket is connected and SSL is negotiated, if needed. This is done so that if
+a higher priority request for a group reaches the socket pool before a
+connection is established, the first usable connection goes to the highest
+priority socket request.
+
+## ResourceScheduler
+
+In addition to net's use of priorities, requests issued by other processes go
+through the ResourceScheduler. The ResourceScheduler restricts the number of
+low priority URLRequests for a given page can be started at once, based on the
+presense of higher priority requests. The idea is to reduce bandwidth
+contention, and to reduce the chance of low priority resources, like images,
+of delaying high priority HTML, CSS, and blocking scripts, so the page is
+displayable and interactive sooner, even if it's missing images and the like.
+
+## Non-HTTP schemes
+
+The URLRequestJobFactory has a ProtocolHander for each supported scheme.
+Non-HTTP URLRequests have their own ProtocolHandlers. Some are implemented in
+net/, (like FTP, file, and data, though blink handles some data URLs
+internally), and others are implemented in content/ or chrome (like blob,
+chrome, and chrome-extension).