akaros/Documentation/process-internals.txt
<<
>>
Prefs
   1process-internals.txt
   2Barret Rhoden
   3
   4This discusses core issues with process design and implementation.  Most of this
   5info is available in the source in the comments (but may not be in the future).
   6For now, it's a dumping ground for topics that people ought to understand before
   7they muck with how processes work.
   8
   9Contents:
  101. Reference Counting
  112. When Do We Really Leave "Process Context"?
  123. Leaving the Kernel Stack
  134. Preemption and Notification Issues
  145. current_ctx and owning_proc
  156. Locking!
  167. TLB Coherency
  178. Process Management
  189. On the Ordering of Messages
  1910. TBD
  20
  211. Reference Counting
  22===========================
  231.1 Basics:
  24---------------------------
  25Reference counts exist to keep a process alive.  We use krefs for this, similar
  26to Linux's kref:
  27- Can only incref if the current value is greater than 0, meaning there is
  28  already a reference to it.  It is a bug to try to incref on something that has
  29  no references, so always make sure you incref something that you know has a
  30  reference.  If you don't know, you need to get it from pid2proc (which is a
  31  careful way of doing this - pid2proc kref_get_not_zero()s on the reference that is
  32  stored inside it).  If you incref and there are 0 references, the kernel will
  33  panic.  Fix your bug / don't incref random pointers.
  34- Can always decref.
  35- When the decref returns 0, perform some operation.  This does some final
  36  cleanup on the object.
  37- Process code is trickier since we frequently make references from 'current'
  38  (which isn't too bad), but also because we often do not return and need to be
  39  careful about the references we passed in to a no-return function.
  40
  411.2 Brief History of the Refcnt:
  42---------------------------
  43Originally, the refcnt was created to keep page tables from being destroyed (in
  44proc_free()) while cores were still using them, which is what was happens during
  45an ARSC (async remote syscall).  It was then defined to be a count of places in
  46the kernel that had an interest in the process staying alive, practically just
  47to protect current/cr3.  This 'interest' actually extends to any code holding a
  48pointer to the proc, such as one acquired via pid2proc(), which is its current
  49meaning.
  50
  511.3 Quick Aside: The current Macro:
  52---------------------------
  53current is a pointer to the proc that is currently loaded/running on any given
  54core.  It is stored in the per_cpu_info struct, and set/managed by low-level
  55process code.  It is necessary for the kernel to quickly figure out who is
  56running on its core, especially when servicing interrupts and traps.  current is
  57protected by a refcnt.
  58
  59current does not say which process owns / will-run on a core.  The per-cpu
  60variable 'owning_proc' covers that.  'owning_proc' should be treated like
  61'current' (aka, 'cur_proc') when it comes to reference counting.  Like all
  62refcnts, you can use it, but you can't consume it without atomically either
  63upping the refcnt or passing the reference (clearing the variable storing the
  64reference).  Don't pass it to a function that will consume it and not return
  65without upping it.
  66
  671.4 Reference Counting Rules:
  68---------------------------
  69+1 for existing.
  70- The fact that the process is supposed to exist is worth +1.  When it is time
  71  to die, we decref, and it will eventually be cleaned up.  This existence is
  72  explicitly kref_put()d in proc_destroy().
  73- The hash table is a bit tricky.  We need to kref_get_not_zero() when it is
  74  locked, so we know we aren't racing with proc_free freeing the proc and
  75  removing it from the list.  After removing it from the hash, we don't need to
  76  kref_put it, since it was an internal ref.  The kref (i.e. external) isn't for
  77  being on the hash list, it's for existing.  This separation allows us to
  78  remove the proc from the hash list in the "release" function.  See kref.txt
  79  for more details.
  80
  81+1 for someone using it or planning to use it.
  82- This includes simply having a pointer to the proc, since presumably you will
  83  use it.  pid2proc() will incref for you.  When you are done, decref.
  84- Functions that create a process and return a pointer (like proc_create() or
  85  kfs_proc_create()) will also up the refcnt for you.  Decref when you're done.
  86- If the *proc is stored somewhere where it will be used again, such as in an IO
  87  continuation, it needs to be refcnt'd.  Note that if you already had a
  88  reference from pid2proc(), simply don't decref after you store the pointer.
  89
  90+1 for current.
  91- current counts as someone using it (expressing interest in the core), but is
  92  also a source of the pointer, so its a bit different.  Note that all kref's
  93  are sources of a pointer.  When we are running on a core that has current
  94  loaded, the ref is both for its usage as well as for being the current
  95  process.
  96- You have a reference from current and can use it without refcnting, but
  97  anything that needs to eat a reference or store/use it needs an incref first.
  98  To be clear, your reference is *NOT* edible.  It protects the cr3, guarantees
  99  the process won't die, and serves as a bootstrappable reference.
 100- Specifically, if you get a ref from current, but then save it somewhere (like
 101  an IO continuation request), then clearly you must incref, since it's both
 102  current and stored/used.
 103- If you don't know what might be downstream from your function, then incref
 104  before passing the reference, and decref when it returns.  We used to do this
 105  for all syscalls, but now only do it for calls that might not return and
 106  expect to receive reference (like proc_yield).
 107
 108All functions that take a *proc have a refcnt'd reference, though it may not be
 109edible (it could be current).  It is the callers responsibility to make sure
 110it'd edible if it necessary.  It is the callees responsibility to incref if it
 111stores or makes a copy of the reference.
 112
 1131.5 Functions That Don't or Might Not Return:
 114---------------------------
 115Refcnting and especially decreffing gets tricky when there are functions that
 116MAY not return.  proc_restartcore() does not return (it pops into userspace).
 117proc_run() used to not return, if the core it was called on would pop into
 118userspace (if it was a _S, or if the core is part of the vcoremap for a _M).
 119This doesn't happen anymore, since we have cur_ctx in the per-cpu info.
 120
 121Functions that MAY not return will "eat" your reference *IF* they do not return.
 122This means that you must have a reference when you call them (like always), and
 123that reference will be consumed / decref'd for you if the function doesn't
 124return.  Or something similarly appropriate.
 125
 126Arguably, for functions that MAY not return, but will always be called with
 127current's reference (proc_yield()), we could get away without giving it an
 128edible reference, and then never eating the ref.  Yield needs to be reworked
 129anyway, so it's not a bit deal yet.
 130
 131We do this because when the function does not return, you will not have the
 132chance to decref (your decref code will never run).  We need the reference when
 133going in to keep the object alive (like with any other refcnt).  We can't have
 134the function always eat the reference, since you cannot simply re-incref the
 135pointer (not allowed to incref unless you know you had a good reference).  You'd
 136have to do something like p = pid2proc(p_pid);  It's clunky to do that, easy to
 137screw up, and semantically, if the function returns, then we may still have an
 138interest in p and should decref later.
 139
 140The downside is that functions need to determine if they will return or not,
 141which can be a pain (for an out-of-date example: a linear time search when
 142running an _M, for instance, which can suck if we are trying to use a
 143broadcast/logical IPI).
 144
 145As the caller, you usually won't know if the function will return or not, so you
 146need to provide a consumable reference.  Current doesn't count.  For example,
 147proc_run() requires a reference.  You can proc_run(p), and use p afterwards, and
 148later decref.  You need to make sure you have a reference, so things like
 149proc_run(pid2proc(55)) works, since pid2proc() increfs for you.  But you cannot
 150proc_run(current), unless you incref current in advance.  Incidentally,
 151proc_running current doesn't make a lot of sense.
 152
 1531.6 Runnable List:
 154---------------------------
 155Procs on the runnable list need to have a refcnt (other than the +1 for
 156existing).  It's something that cares that the process exists.  We could have
 157had it implicitly be refcnt'd (the fact that it's on the list is enough, sort of
 158as if it was part of the +1 for existing), but that complicates things.  For
 159instance, it is a source of a reference (for the scheduler) and you could not
 160proc_run() a process from the runnable list without worrying about increfing it
 161before hand.  This isn't true anymore, but the runnable lists are getting
 162overhauled anyway.  We'll see what works nicely.
 163
 1641.7 Internal Details for Specific Functions:
 165---------------------------
 166proc_run()/__proc_give_cores(): makes sure enough refcnts are in place for all
 167places that will install owning_proc.  This also makes it easier on the system
 168(one big incref(n), instead of n increfs of (1) from multiple cores). 
 169
 170__set_proc_current() is a helper that makes sure p is the cur_proc.  It will
 171incref if installing a new reference to p.  If it removed an old proc, it will
 172decref.
 173
 174__proc_startcore(): assumes all references to p are sorted.  It will not
 175return, and you should not pass it a reference you need to decref().  Passing
 176it 'owning_proc' works, since you don't want to decref owning_proc.
 177
 178proc_destroy(): it used to not return, and back then if your reference was
 179from 'current', you needed to incref.  Now that proc_destroy() returns, it
 180isn't a big deal.  Just keep in mind that if you have a function that doesn't
 181return, there's no way for the function to know if it's passed reference is
 182edible.  Even if p == current, proc_destroy() can't tell if you sent it p (and
 183had a reference) or current and didn't.
 184
 185proc_yield(): when this doesn't return, it eats your reference.  It will also
 186decref twice.  Once when it clears_owning_proc, and again when it calls
 187abandon_core() (which clears cur_proc).
 188
 189abandon_core(): it was not given a reference, so it doesn't eat one.  It will
 190decref when it unloads the cr3.  Note that this is a potential performance
 191issue.  When preempting or killing, there are n cores that are fighting for the
 192cacheline to decref.  An alternative would be to have one core decref for all n
 193cores, after it knows all cores unloaded the cr3.  This would be a good use of
 194the checklist (possibly with one cacheline per core).  It would take a large
 195amount of memory for better scalability.
 196
 1971.8 Things I Could Have Done But Didn't And Why:
 198---------------------------
 199Q: Could we have the first reference (existence) mean it could be on the runnable
 200list or otherwise in the proc system (but not other subsystems)?  In this case,
 201proc_run() won't need to eat a reference at all - it will just incref for every
 202current it will set up.
 203
 204New A: Maybe, now that proc_run() returns.
 205
 206Old A: No: if you pid2proc(), then proc_run() but never return, you have (and
 207lose) an extra reference.  We need proc_run() to eat the reference when it
 208does not return.  If you decref between pid2proc() and proc_run(), there's a
 209(rare) race where the refcnt hits 0 by someone else trying to kill it.  While
 210proc_run() will check to see if someone else is trying to kill it, there's a
 211slight chance that the struct will be reused and recreated.  It'll probably
 212never happen, but it could, and out of principle we shouldn't be referencing
 213memory after it's been deallocated.  Avoiding races like this is one of the
 214reasons for our refcnt discipline.
 215
 216Q: (Moot) Could proc_run() always eat your reference, which would make it
 217easier for its implementation?
 218
 219A: Yeah, technically, but it'd be a pain, as mentioned above.  You'd need to
 220reaquire a reference via pid2proc, and is rather easy to mess up.
 221
 222Q: (Moot) Could we have made proc_destroy() take a flag, saying whether or not
 223it was called on current and needed a decref instead of wasting an incref?
 224
 225A: We could, but won't.  This is one case where the external caller is the one
 226that knows the function needs to decref or not.  But it breaks the convention a
 227bit, doesn't mirror proc_create() as well, and we need to pull in the cacheline
 228with the refcnt anyways.  So for now, no.
 229
 230Q: (Moot) Could we make __proc_give_cores() simply not return if an IPI is
 231coming?
 232
 233A: I did this originally, and manually unlocked and __wait_for_ipi()d.  Though
 234we'd then need to deal with it like that for all of the related functions, which
 235doesn't work if you wanted to do something afterwards (like schedule(p)).  Also
 236these functions are meant to be internal helpers, so returning the bool makes
 237more sense.  It eventually led to having __proc_unlock_ipi_pending(), which made
 238proc_destroy() much cleaner and helped with a general model of dealing with
 239these issues.  Win-win.
 240
 2412. When Do We Really Leave "Process Context"?
 242===========================
 2432.1 Overview
 244---------------------------
 245First off, it's not really "process context" in the way Linux deals with it.  We
 246aren't operating in kernel mode on behalf of the process (always).  We are
 247specifically talking about when a process's cr3 is loaded on a core.  Usually,
 248current is also set (the exception for now is when processing ARSCs).
 249
 250There are a couple different ways to do this.  One is to never unload a context
 251until something new is being run there (handled solely in __proc_startcore()).
 252Another way is to always explicitly leave the core, like by abandon_core()ing.
 253
 254The issue with the former is that you could have contexts sitting around for a
 255while, and also would have a bit of extra latency when __proc_free()ing during
 256someone *else's* __proc_startcore() (though that could be avoided if it becomes
 257a real issue, via some form of reaping).  You'll also probably have excessive
 258decrefs (based on the interactions between proc_run() and __startcore()).
 259
 260The issue with the latter is excessive TLB shootdowns and corner cases.  There
 261could be some weird cases (in core_request() for example) where the core you are
 262running on has the context loaded for proc A on a mgmt core, but decides to give
 263it to proc B.
 264
 265If no process is running there, current == 0 and boot_cr3 is loaded, meaning no
 266process's context is loaded.
 267
 268All changes to cur_proc, owning_proc, and cur_ctx need to be done with
 269interrupts disabled, since they change in interrupt handlers.
 270
 2712.2 Here's how it is done now:
 272---------------------------
 273All code is capable of 'spamming' cur_proc (with interrupts disabled!).  If it
 274is 0, feel free to set it to whatever process you want.  All code that
 275requires current to be set will do so (like __proc_startcore()).  The
 276smp_idle() path will make sure current is clear when it halts.  So long as you
 277don't change other concurrent code's expectations, you're okay.  What I mean
 278by that is you don't clear cur_proc while in an interrupt handler.  But if it
 279is already 0, __startcore is allowed to set it to it's future proc (which is
 280an optimization).  Other code didn't have any expectations of it (it was 0).
 281Likewise, kthread code when we sleep_on() doesn't have to keep cur_proc set.
 282A kthread is somewhat an isolated block (codewise), and leaving current set
 283when it is done is solely to avoid a TLB flush (at the cost of an incref).
 284
 285In general, we try to proactively leave process context, but have the ability
 286to stay in context til __proc_startcore() to handle the corner cases (and to
 287maybe cut down the TLB flushes later).  To stop proactively leaving, just
 288change abandon_core() to not do anything with current/cr3.  You'll see weird
 289things like processes that won't die until their old cores are reused.  The
 290reason we proactively leave context is to help with sanity for these issues,
 291and also to avoid decref's in __startcore().
 292
 293A couple other details: __startcore() sorts the extra increfs, and
 294__proc_startcore() sorts leaving the old context.  Anytime a __startcore kernel
 295message is sent, the sender increfs in advance for the owning_proc refcnt.  As
 296an optimization, we can also incref to *attempt* to set current.  If current
 297was 0, we set it.  If it was already something else, we failed and need to
 298decref.  __proc_startcore(), which the last moment before we *must* have the
 299cr3/current issues sorted, does the actual check if there was an old process
 300there or not, while it handles the lcr3 (if necessary).  In general, lcr3's
 301ought to have refcnts near them, or else comments explaining why not.
 302
 303So we leave process context when told to do so (__death/abandon_core()) or if
 304another process is run there.  The _M code is such that a proc will stay on its
 305core until it receives a message, and that message would cleanup/restore a
 306generic context (boot_cr3).  A _S could stay on its core until another _S came
 307in.  This is much simpler for cases when a timer interrupt goes off to force a
 308schedule() decision.  It also avoids a TLB flush in case the scheduler picked
 309that same proc to run again.  This could also happen to an _M, if for some
 310reason it was given a management core (!!!) or some other event happened that
 311caused some management/scheduling function to run on one of it's cores (perhaps
 312it asked).
 313
 314proc_yield() abandons the core / leaves context.
 315
 3162.3 Other issues:
 317---------------------------
 318Note that dealing with interrupting processes that are in the kernel is tricky.
 319There is no true process context, so we can't leave a core until the kernel is
 320in a "safe place", i.e. it's state is bundled enough that it can be recontinued
 321later.  Calls of this type are routine kernel messages, executed at a convenient
 322time (specifically, before we return to userspace in proc_restartcore().
 323
 324This same thing applies to __death messages.  Even though a process is dying, it
 325doesn't mean we can just drop whatever the kernel was doing on its behalf.  For
 326instance, it might be holding a reference that will never get decreffed if its
 327stack gets dropped.
 328
 3293. Leaving the Kernel Stack:
 330===========================
 331Just because a message comes in saying to kill a process, it does not mean we
 332should immediately abandon_core().  The problem is more obvious when there is
 333a preempt message, instead of a death message, but either way there is state
 334that needs cleaned up (refcnts that need downed, etc).
 335
 336The solution to this is rather simple: don't abandon right away.  That was
 337always somewhat the plan for preemption, but was never done for death.  And
 338there are several other cases to worry about too.  To enforce this, we expand
 339the old "active messages" into a generic work execution message (a kernel
 340message) that can be delayed or shipped to another core.  These types of
 341messages will not be executed immediately on the receiving pcore - instead they
 342are on the queue for "when there's nothing else to do in the kernel", which is
 343checked in smp_idle() and before returning to userspace in proc_restartcore().
 344Additionally, these kernel messages can also be queued on an alarm queue,
 345delaying their activation as part of a generic kernel alarm facility.
 346
 347One subtlety is that __proc_startcore() shouldn't check for messages, since it
 348is called by __startcore (a message).  Checking there would run the messages out
 349of order, which is exactly what we are trying to avoid (total chaos).  No one
 350should call __proc_startcore, other than proc_restartcore() or __startcore().
 351If we ever have functions that do so, if they are not called from a message,
 352they must check for outstanding messages.
 353
 354This last subtlety is why we needed to change proc_run()'s _S case to use a
 355local message instead of calling proc_starcore (and why no one should ever call
 356proc_startcore()).  We could unlock, thereby freeing another core to change the
 357proc state and send a message to us, then try to proc_startcore, and then
 358reading the message before we had installed current or had a userspace TF to
 359preempt, and probably a few other things.  Treating _S as a local message is
 360cleaner, begs to be merged in the code with _M's code, and uses the messaging
 361infrastructure to avoid all the races that it was created to handle.
 362
 363Incidentally, we don't need to worry about missing messages while trying to pop
 364back to userspace from __proc_startcore, since an IPI will be on the way
 365(possibly a self-ipi caused by the __kernel_message() handler).  This is also
 366why we needed to make process_routine_kmsg() keep interrupts disabled when it
 367stops (there's a race between checking the queue and disabling ints).
 368
 3694. Preemption and Notification Issues:
 370===========================
 3714.1: Message Ordering and Local Calls:
 372---------------------------
 373Since we go with the model of cores being told what to do, there are issues
 374with messages being received in the wrong order.  That is why we have the
 375kernel messages (guaranteed, in-order delivery), with the proc-lock protecting
 376the send order.  However, this is not enough for some rare races.
 377
 378Local calls can also perform the same tasks as messages (calling
 379proc_destroy() while a death IPI is on its way). We refer to these calls as
 380messing with "local fate" (compared to global state (we're clever).
 381Preempting a single vcore doesn't change the process's state).  These calls
 382are a little different, because they also involve a check to see if it should
 383perform the function or other action (e.g., death just idling and waiting for
 384an IPI instead of trying to kill itself), instead of just blindly doing
 385something.
 386
 3874.1.1: Possible Solutions
 388----------------
 389There are two ways to deal with this.  One (and the better one, I think) is to
 390check state, and determine if it should proceed or abort.  This requires that
 391all local-fate dependent calls always have enough state to do its job.  In the
 392past, this meant that any function that results in sending a directive to a
 393vcore store enough info in the proc struct that a local call can determine if
 394it should take action or abort.  In the past, we used the vcore/pcoremap as a
 395way to send info to the receiver about what vcore they are (or should be).
 396Now, we store that info in pcpui (for '__startcore', we send it as a
 397parameter.  Either way, the general idea is still true: local calls can
 398proceed when they are called, and not self-ipi'd to a nebulous later time.
 399
 400The other way is to send the work (including the checks) in a self-ipi kernel
 401message.  This will guarantee that the message is executed after any existing
 402messages (making the k_msg queue the authority for what should happen to a
 403core).  The check is also performed later (when the k_msg executes).  There
 404are a couple issues with this: if we allow the local core to send itself an
 405k_msg that could be out of order (meaning it should not be sent, and is only
 406sent due to ignorance of its sealed fate), AND if we return the core to the
 407idle-core-list once its fate is sealed, we need to detect that the message is
 408for the wrong process and that the process is in the wrong state.  To do this,
 409we probably need local versioning on the pcore so it can detect that the
 410message is late/wrong.  We might get by with just the proc* (though that is
 411tricky with death and proc reuse), so long as we don't allow new startcores
 412for a proc until AFTER the preemption is completed.
 413
 4144.2: Preempt-Served Flag
 415----------------
 416We want to be able to consider a pcore free once its owning proc has dealt
 417with removing it.  This allows a scheduler-like function to easily take a core
 418and then give it to someone else, without waiting for each vcore to respond,
 419saying that the pcore is free/idle.
 420
 421We used to not unmap until we were in '__preempt' or '__death', and we needed
 422a flag to tell yield-like calls that a message was already on the way and to
 423not rely on the vcoremap.  This is pretty fucked up for a number of reasons,
 424so we changed that.  But we still wanted to know when a preempt was in
 425progress so that the kernel could avoid giving out the vcore until the preempt
 426was complete.
 427
 428Here's the scenario: we send a '__startcore' to core 3 for VC5->PC3.  Then we
 429quickly send a '__preempt' to 3, and then a '__startcore' to core 4 (a
 430different pcore) for VC5->PC4.  Imagine all of this happens before the first
 431'__startcore' gets processed (IRQ delay, fast ksched, whatever).  We need to
 432not run the second '__startcore' on pcore 4 before the preemption has saved
 433all of the state of the VC5.  So we spin on preempt_served (which may get
 434renamed to preempt_in_progress).  We need to do this in the sender, and not
 435the receiver (not in the kmsg), because the kmsgs can't tell which one they
 436are.  Specifically, the first '__startcore' on core 3 runs the same code as
 437the '__startcore' on core 4, working on the same vcore.  Anything we tell VC5
 438will be seen by both PC3 and PC4.  We'd end up deadlocking on PC3 while it
 439spins waiting for the preempt message that also needs to run on PC3.
 440
 441The preempt_pending flag is actual a timestamp, with the expiration time of
 442the core at which the message will be sent.  We could try to use that, but
 443since alarms aren't fired at exactly the time they are scheduled, the message
 444might not actually be sent yet (though it will, really soon).  Still, we'll
 445just go with the preempt-served flag for now.
 446
 4474.3: Impending Notifications
 448----------------
 449It's also possible that there is an impending notification.  There's no change
 450in fate (though there could be a fate-changing preempt on its way), just the
 451user wants a notification handler to run.  We need a flag anyways for this
 452(discussed below), so proc_yield() or whatever other local call we have can
 453check this flag as well.  
 454
 455Though for proc_yield(), it doesn't care if a notification is on its way (can
 456be dependent on a flag to yield from userspace, based on the nature of the
 457yield (which still needs to be sorted)).  If the yield is in response to a
 458preempt_pending, it actually should yield and not receive the notification.
 459So it should destroy its vcoreid->pcoreid mapping and abandon_core().  When
 460that notification hits, it will be for a proc that isn't current, and will be
 461ignored (it will get run the next time that vcore fires up, handled below).
 462
 463There is a slight chance that the same proc will run on that pcore, but with a
 464different vcoreid.  In the off chance this happens, the new vcore will get a
 465spurious notification.  Userspace needs to be able to handle spurious
 466notifications anyways, (there are a couple other cases, and in general it's
 467not hard to do), so this is not a problem.  Instead of trying to have the
 468kernel ignore the notification, we just send a spurious one.  A crappy
 469alternative would be to send the vcoreid with the notification, but that would
 470mean we can't send a generic message (broadcast) to a bunch of cores, which
 471will probably be a problem later.
 472
 473Note that this specific case is because the "local work message" gets
 474processed out of order with respect to the notification.  And we want this in
 475that case, since that proc_yield() is more important than the notification.
 476
 4774.4: Preemption / Allocation Phases and Alarm Delays
 478---------------------------
 479A per-vcore preemption phase starts when the kernel marks the core's
 480preempt_pending flag/counter and can includes the time when an alarm is
 481waiting to go off to reclaim the core.  The phase ends when the vcore's pcore
 482is reclaimed, either as a result of the kernel taking control, or because a
 483process voluntarily yielded.
 484
 485Specifically, the preempt_pending variable is actually a timestamp for when
 486the core will be revoked (this assumes some form of global time, which we need
 487anyways).  If its value is 0, then there is no preempt-pending, it is not in a
 488phase, and the vcore can be given out again. 
 489
 490When a preempt alarm goes off, the alarm only means to check a process for
 491expired vcores.  If the vcore has been yielded while the alarm was pending,
 492the preempt_pending flag will be reset to 0.  To speed up the search for
 493vcores to preempt, there's a circular buffer corelist in the proc struct, with
 494vcoreids of potential suspects.  Or at least this will exist at some point.
 495Also note that the preemption list isn't bound to a specific alarm: you can
 496check the list at any time (not necessarily on a specific alarm), and you can
 497have spurious alarms (the list is empty, so it'll be a noop).
 498
 499Likewise, a global preemption phase is when an entire MCP is getting
 500gang_prempted, and the global deadline is set.  A function can quickly check
 501to see if the process responded, since the list of vcores with preemptions
 502pending will be empty.
 503
 504It seems obvious, but we do not allow allocation of a vcore during its
 505preemption phase.  The main reason is that it can potentially break
 506assumptions about the vcore->pcore mapping and can result in multiple
 507instances of the same vcore on different pcores.  Imagine a preempt message
 508sent to a pcore (after the alarm goes off), meanwhile that vcore/pcore yields
 509and the vcore reactivates somewhere else.  There is a potential race on the
 510vcore_ctx state: the new vcore is reading while the old is writing.  This
 511issue is sorted naturally: the vcore entry in the vcoremap isn't cleared until
 512the vcore/pcore is actually yielded/taken away, so the code looking for a free
 513vcoreid slot will not try to use it.
 514
 515Note that if we didn't design the alarm system to simply check for
 516preemptions (perhaps it has a stored list of vcores to preempt), then we
 517couldn't end the preempt-phase until the alarm was sorted.  If that is the
 518case, we could easily give out a vcore that had been yielded but was still in
 519a preempt-phase.  Stopping an alarm would be tricky too, since there could be
 520lots of vcores in different states that need to be sorted by the alarm (so
 521ripping it out isn't enough).  Setting a flag might not be enough either.
 522Vcore version numbers/history (as well as global proc histories) is a pain I'd
 523like to avoid too.  So don't change the alarm / delayed preemption system
 524without thinking about this.
 525
 526Also, allowing a vcore to restart while preemptions are pending also mucks
 527with keeping the vcore mapping "old" (while the message is in flight).  A
 528pcore will want to use that to determine which vcore is running on it.  It
 529would be possible to keep a pcoremap for the reverse mapping out of sync, but
 530that seems like a bad idea.  In general, having the pcoremap is a good idea
 531(whenever we talk about a vcoremap, we're usually talking about both
 532directions: "the vcore->pcore mapping").
 533
 5344.5: Global Preemption Flags
 535---------------------------
 536If we are trying to preempt an entire process at the same time, instead of
 537playing with the circular buffer of vcores pending preemption, we could have a
 538global timer as well.  This avoids some O(n) operations, though it means that
 539userspace needs to check two "flags" (expiration dates) when grabbing its
 540preempt-critical locks.
 541
 5424.6: Notifications Mixed with Preemption and Sleeping
 543---------------------------
 544It is possible that notifications will mix with preemptions or come while a
 545process is not running.  Ultimately, the process wants to be notified on a
 546given vcore.  Whenever we send an active notification, we set a flag in procdata
 547(notif_pending).  If the vcore is offline, we don't bother sending the IPI/notif
 548message.  The kernel will make sure it runs the notification handler (as well as
 549restoring the vcore_ctx) the next time that vcore is restarted.  Note that
 550userspace can toggle this, so they can handle the notifications from a different
 551core if it likes, or they can independently send a notification.
 552
 553Note we use notif_pending to detect if an IPI was missed while notifs were
 554disabled (this is done in pop_user_ctx() by userspace).  The overall meaning
 555of notif_pending is that a vcore wants to be IPI'd.  The IPI could be
 556in-flight, or it could be missed.  Since notification IPIs can be spurious,
 557when we have potential races, we err on the side of sending.  This happens
 558when pop_user_ctx() notifies itself, and when the kernel makes sure to start a
 559vcore in vcore context if a notif was pending.  This was simplified a bit over
 560the years by having uthreads always get saved into the uthread_ctx (formerly
 561the notif_tf), instead of in the old preempt_tf (which is now the vcore_ctx).
 562
 563If a vcore has a preempt_pending, we will still send the active notification
 564(IPI).  The core ought to get a notification for the preemption anyway, so we
 565need to be able to send one.  Additionally, once the vcore is handling that
 566preemption notification, it will have notifs disabled, which will prevent us
 567from sending any extra notifications anyways.
 568 
 5694.7: Notifs While a Preempt Message is Served
 570---------------------------
 571It is possible to have the kernel handling a notification k_msg and to have a
 572preempt k_msg in the queue (preempt-served flag is set).  Ultimately, what we
 573want is for the core to be preempted and the notification handler to run on
 574the next execution.  Both messages are in the k_msg queue for "a convenient
 575time to leave the kernel" (I'll have a better name for that later).  What we
 576do is execute the notification handler and jump to userspace.  Since there is
 577still an k_msg in the queue (and we self_ipi'd ourselves, it's part of how
 578k_msgs work), the IPI will fire and push us right back into the kernel to
 579execute the preemption, and the notif handler's context will be saved in the
 580vcore_ctx (ready to go when the vcore gets started again).
 581
 582We could try to just leave the notif_pending flag set and ignore the message,
 583but that would involve inspecting the queue for the preempt k_msg.
 584Additionally, a preempt k_msg can arrive anyway.  Finally, it's possible to have
 585another message in the queue between the notif and the preempt, and it gets ugly
 586quickly trying to determine what to do.
 587
 5884.8: When a Pcore is "Free"
 589---------------------------
 590There are a couple ways to handle pcores.  One approach would be to not
 591consider them free and able to be given to another process until the old
 592process is completely removed (abandon_core()).  Another approach is to free
 593the core once its fate is sealed (which we do).  This probably gives more
 594flexibility in schedule()-like functions (no need to wait to give the core
 595out), quicker dispatch latencies, less contention on shared structs (like the
 596idle-core-map), etc.
 597
 598This 'freeing' of the pcore is from the perspective of the kernel scheduler
 599and the proc struct.  Contrary to all previous announcements, vcores are
 600unmapped from pcores when sending k_msgs (technically right after), while
 601holding the lock.  The pcore isn't actually not-running-the-proc until the
 602kmsg completes and we abandon_core().  Previously, we used the vcoremap to
 603communicate to other cores in a lock-free manner, but that was pretty shitty
 604and now we just store the vcoreid in pcpu info.
 605
 606Another tricky part is the seq_ctr used to signal userspace of changes to the
 607coremap or num_vcores (coremap_seqctr).  While we may not even need this in the
 608long run, it still seems like it could be useful.  The trickiness comes from
 609when we update the seq_ctr when we are unmapping vcores on the receive side of a
 610message (like __death or __preempt).  We'd rather not have each pcore contend on
 611the seq_ctr cache line (let alone any locking) while they perform a somewhat
 612data-parallel task.  So we continue to have the sending core handle the seq_ctr
 613upping and downing.  This works, since the "unlocking" happens after messages
 614are sent, which means the receiving core is no longer in userspace (if there is
 615a delay, it is because the remote core is in the kernel, possibly with
 616interrupts disabled).  Because of this, userspace will be unable to read the new
 617value of the seq_ctr before the IPI hits and does the unmapping that the seq_ctr
 618protects/advertises.  This is most likely true.  It wouldn't be if the "last IPI
 619was sent" flag clears before the IPI actually hit the other core.
 620
 6214.9: Future Broadcast/Messaging Needs
 622---------------------------
 623Currently, messaging is serialized.  Broadcast IPIs exist, but the kernel
 624message system is based on adding an k_msg to a list in a pcore's
 625per_cpu_info.  Further the sending of these messages is in a loop.  In the
 626future, we would like to have broadcast messaging of some sort (literally a
 627broadcast, like the IPIs, and if not that, then a communication tree of
 628sorts).  
 629
 630In the past, (OLD INFO): given those desires, we wanted to make sure that no
 631message we send needs details specific to a pcore (such as the vcoreid running
 632on it, a history number, or anything like that).  Thus no k_msg related to
 633process management would have anything that cannot apply to the entire
 634process.  At this point, most just have a struct proc *.  A pcore was be able
 635to figure out what is happening based on the pcoremap, information in the
 636struct proc, and in the preempt struct in procdata.
 637
 638In more recent revisions, the coremap no longer is meant to be used across
 639kmsgs, so some messages ('__startcore') send the vcoreid.  This means we can't
 640easily broadcast the message.  However, many broadcast mechanisms wouldn't
 641handle '__startcore' naturally.  For instance, logical IPIs need something
 642already set in the LAPIC, or maybe need to be sent to a somewhat predetermined
 643group (again, bits in the LAPIC).  If we tried this for '__startcore', we
 644could add something in to the messaging to carry these vcoreids.  More likely,
 645we'll have a broadcast tree.  Keeping vcoreid (or any arg) next to whoever
 646needs to receive the message is a very small amount of bookkeeping on a struct
 647that already does bookkeeping.
 648
 6494.10: Other Things We Thought of but Don't Like
 650---------------------------
 651All local fate-related work is sent as a self k_msg, to enforce ordering.
 652It doesn't capture the difference between a local call and a remote k_msg.
 653The k_msg has already considered state and made its decision.  The local call
 654is an attempt.  It is also unnecessary, if we put in enough information to
 655make a decision in the proc struct.  Finally, it caused a few other problems
 656(like needing to detect arbitrary stale messages).
 657
 658Overall message history: doesn't work well when you do per-core stuff, since
 659it will invalidate other messages for the process.  We then though of a pcore
 660history counter to detect stale messages.  Don't like that either.  We'd have
 661to send the history in the message, since it's a per-message, per-core
 662expiration.  There might be other ways around this, but this doesn't seem
 663necessary.
 664
 665Alarms have pointers to a list of which cores should be preempted when that
 666specific alarm goes off (saved with the alarm).  Ugh.  It gets ugly with
 667multiple outstanding preemptions and cores getting yielded while the alarms
 668sleep (and possibly could get reallocated, though we'd make a rule to prevent
 669that).  Like with notifications, being able to handle spurious alarms and
 670thinking of an alarm as just a prod to check somewhere is much more flexible
 671and simple.  It is similar to generic messages that have the actual important
 672information stored somewhere else (as with allowing broadcasts, with different
 673receivers performing slightly different operations).
 674
 675Synchrony for messages (wanting a response to a preempt k_msg, for example)
 676sucks.  Just encode the state of impending fate in the proc struct, where it
 677belongs.  Additionally, we don't want to hold the proc lock even longer than
 678we do now (which is probably too long as it is).  Finally, it breaks a golden
 679rule: never wait while holding a lock: you will deadlock the system (e.g. if
 680the receiver is already in the kernel spinning on the lock).  We'd have to
 681send messages, unlock (which might cause a message to hit the calling pcore,
 682as in the case of locally called proc_destroy()), and in the meantime some
 683useful invariant might be broken.
 684
 685We also considered using the transition stack as a signal that a process is in
 686a notification handler.  The kernel can inspect the stack pointer to determine
 687this.  It's possible, but unnecessary.
 688
 689Using the pcoremap as a way to pass info with kmsgs: it worked a little, but
 690had some serious problems, as well as making life difficult.  It had two
 691purposes: help with local fate calls (yield) and allow broadcast messaging.
 692The main issue is that it was using a global struct to pass info with
 693messages, but it was based on the snapshot of state at the time the message
 694was sent.  When you send a bunch of messages, certain state may have changed
 695between messages, and the old snapshot isn't there anymore by the time the
 696message gets there.  To avoid this, we went through some hoops and had some
 697fragile code that would use other signals to avoid those scenarios where the
 698global state change would send the wrong message.  It was tough to understand,
 699and not clear it was correct (hint: it wasn't).  Here's an example (on one
 700pcore): if we send a preempt and we then try to map that pcore to another
 701vcore in the same process before the preempt call checks its pcoremap, we'll
 702clobber the pcore->vcore mapping (used by startcore) and the preempt will
 703think it is the new vcore, not the one it was when the message was sent.
 704While this is a bit convoluted, I can imagine a ksched doing this, and
 705perhaps with weird IRQ delays, the messages might get delayed enough for it to
 706happen.  I'd rather not have to have the ksched worry about this just because
 707proc code was old and ghetto.  Another reason we changed all of this was so
 708that you could trust the vcoremap while holding the lock.  Otherwise, it's
 709actually non-trivial to know the state of a vcore (need to check a combination
 710of preempt_served and is_mapped), and even if you do that, there are some
 711complications with doing this in the ksched.
 712
 7135. current_ctx and owning_proc
 714===========================
 715Originally, current_tf was a per-core macro that returns a struct trapframe *
 716that points back on the kernel stack to the user context that was running on
 717the given core when an interrupt or trap happened.  Saving the reference to
 718the TF helps simplify code that needs to do something with the TF (like save
 719it and pop another TF).  This way, we don't need to pass the context all over
 720the place, especially through code that might not care.
 721
 722Then, current_tf was more broadly defined as the user context that should be
 723run when the kernel is ready to run a process.  In the older case, it was when
 724the kernel tries to return to userspace from a trap/interrupt.  current_tf
 725could be set by an IPI/KMSG (like '__startcore') so that when the kernel wants
 726to idle, it will find a current_tf that it needs to run, even though we never
 727trapped in on that context in the first place.
 728
 729Finally, current_tf was changed to current_ctx, and instead of tracking a
 730struct trapframe (equivalent to a hw_trapframe), it now tracked a struct
 731user_context, which could be either a HW or a SW trapframe.
 732
 733Further, we now have 'owning_proc', which tells the kernel which process
 734should be run.  'owning_proc' is a bigger deal than 'current_ctx', and it is
 735what tells us to run cur_ctx.
 736
 737Process management KMSGs now simply modify 'owning_proc' and cur_ctx, as if we
 738had interrupted a process.  Instead of '__startcore' forcing the kernel to
 739actually run the process and trapframe, it will just mean we will eventually
 740run it.  In the meantime a '__notify' or a '__preempt' can come in, and they
 741will apply to the owning_proc/cur_ctx.  This greatly simplifies process code
 742and code calling process code (like the scheduler), since we no longer need to
 743worry about whether or not we are getting a "stack killing" kernel message.
 744Before this, code needed to care where it was running when managing _Ms.
 745
 746Note that neither 'current_ctx' nor 'owning_proc' rely on 'current'/'cur_proc'.
 747'current' is just what process context we're in, not what process (and which
 748trapframe) we will eventually run.
 749
 750cur_ctx does not point to kernel trapframes, which is important when we
 751receive an interrupt in the kernel.  At one point, we were (hypothetically)
 752clobbering the reference to the user trapframe, and were unable to recover.
 753We can get away with this because the kernel always returns to its previous
 754context from a nested handler (via iret on x86).  
 755
 756In the future, we may need to save kernel contexts and may not always return
 757via iret.  At which point, if the code path is deep enough that we don't want
 758to carry the TF pointer, we may revisit this.  Until then, current_ctx is just
 759for userspace contexts, and is simply stored in per_cpu_info.
 760
 761Brief note from the future (months after this paragraph was written): cur_ctx
 762has two aspects/jobs:
 7631) tell the kernel what we should do (trap, fault, sysc, etc), how we came
 764into the kernel (the fact that it is a user tf), which is why we copy-out
 765early on
 7662) be a vehicle for us to restart the process/vcore
 767
 768We've been focusing on the latter case a lot, since that is what gets
 769removed when preempted, changed during a notify, created during a startcore,
 770etc.  Don't forget it was also an instruction of sorts.  The former case is
 771always true throughout the life of the syscall.  The latter only happens to be
 772true throughout the life of a *non-blocking* trap since preempts are routine
 773KMSGs.  But if we block in a syscall, the cur_ctx is no longer the TF we came
 774in on (and possibly the one we are asked to operate on), and that old cur_ctx
 775has probably restarted.
 776
 777(Note that cur_ctx is a pointer, and syscalls/traps actually operate on the TF
 778they came in on regardless of what happens to cur_ctx or pcpui->actual_tf.)
 779
 7806. Locking!
 781===========================
 7826.1: proc_lock
 783---------------------------
 784Currently, all locking is done on the proc_lock.  It's main goal is to protect
 785the vcore mapping (vcore->pcore and vice versa).  As of Apr 2010, it's also used
 786to protect changes to the address space and the refcnt.  Eventually the refcnt
 787will be handled with atomics, and the address space will have it's own MM lock.  
 788
 789We grab the proc_lock all over the place, but we try to avoid it whereever
 790possible - especially in kernel messages or other places that will be executed
 791in parallel.  One place we do grab it but would like to not is in proc_yield().  
 792We don't always need to grab the proc lock.  Here are some examples:
 793
 7946.1.1: Lockless Notifications:
 795-------------
 796We don't lock when sending a notification.  We want the proc_lock to not be an
 797irqsave lock (discussed below).  Since we might want to send a notification from
 798interrupt context, we can't grab the proc_lock if it's a regular lock.  
 799
 800This is okay, since the proc_lock is only protecting the vcoremapping.  We could
 801accidentally send the notification to the wrong pcore.  The __notif handler
 802checks to make sure it is the right process, and all _M processes should be able
 803to handle spurious notifications.  This assumes they are still _M.
 804
 805If we send it to the wrong pcore, there is a danger of losing the notif, since
 806it didn't go to the correct vcore.  That would happen anyway, (the vcore is
 807unmapped, or in the process of mapping).  The notif_pending flag will be caught
 808when the vcore is started up next time (and that flag was set before reading the
 809vcoremap).
 810
 8116.1.2: Local get_vcoreid():
 812-------------
 813It's not necessary to lock while checking the vcoremap if you are checking for
 814the core you are running on (e.g. pcoreid == core_id()).  This is because all
 815unmappings of a vcore are done on the receive side of a routine kmsg, and that
 816code cannot run concurrently with the code you are running.  
 817
 8186.2: irqsave
 819---------------------------
 820The proc_lock used to be an irqsave lock (meaning it disables interrupts and can
 821be grabbed from interrupt context).  We made it a regular lock for a couple
 822reasons.  The immediate one was it was causing deadlocks due to some other
 823ghetto things (blocking on the frontend server, for instance).  More generally,
 824we don't want to disable interrupts for long periods of time, so it was
 825something worth doing anyway.  
 826
 827This means that we cannot grab the proc_lock from interrupt context.  This
 828includes having schedule called from an interrupt handler (like the
 829timer_interrupt() handler), since it will call proc_run.  Right now, we actually
 830do this, which we shouldn't, and that will eventually get fixed.  The right
 831answer is that the actual work of running the scheduler should be a routine
 832kmsg, similar to how Linux sets a bit in the kernel that it checks on the way
 833out to see if it should run the scheduler or not.
 834
 8357. TLB Coherency
 836===========================
 837When changing or removing memory mappings, we need to do some form of a TLB
 838shootdown.  Normally, this will require sending an IPI (immediate kmsg) to
 839every vcore of a process to unmap the affected page.  Before allocating that
 840page back out, we need to make sure that every TLB has been flushed.  
 841
 842One reason to use a kmsg over a simple handler is that we often want to pass a
 843virtual address to flush for those architectures (like x86) that can
 844invalidate a specific page.  Ideally, we'd use a broadcast kmsg (doesn't exist
 845yet), though we already have simple broadcast IPIs.
 846
 8477.1 Initial Stuff
 848---------------------------
 849One big issue is whether or not to wait for a response from the other vcores
 850that they have unmapped.  There are two concerns: 1) Page reuse and 2) User
 851semantics.  We cannot give out the physical page while it may still be in a
 852TLB (even to the same process.  Ask us about the pthread_test bug).
 853
 854The second case is a little more detailed.  The application may not like it if
 855it thinks a page is unmapped or protected, and it does not generate a fault.
 856I am less concerned about this, especially since we know that even if we don't
 857wait to hear from every vcore, we know that the message was delivered and the
 858IPI sent.  Any cores that are in userspace will have trapped and eventually
 859handle the shootdown before having a chance to execute other user code.  The
 860delays in the shootdown response are due to being in the kernel with
 861interrupts disabled (it was an IMMEDIATE kmsg).
 862
 8637.2 RCU
 864---------------------------
 865One approach is similar to RCU.  Unmap the page, but don't put it on the free
 866list.  Instead, don't reallocate it until we are sure every core (possibly
 867just affected cores) had a chance to run its kmsg handlers.  This time is
 868similar to the RCU grace periods.  Once the period is over, we can then truly
 869free the page.
 870
 871This would require some sort of RCU-like mechanism and probably a per-core
 872variable that has the timestamp of the last quiescent period.  Code caring
 873about when this page (or pages) can be freed would have to check on all of the
 874cores (probably in a bitmask for what needs to be freed).  It would make sense
 875to amortize this over several RCU-like operations.
 876
 8777.3 Checklist
 878---------------------------
 879It might not suck that much to wait for a response if you already sent an IPI,
 880though it incurs some more cache misses.  If you wanted to ensure all vcores
 881ran the shootdown handler, you'd have them all toggle their bit in a checklist
 882(unused for a while, check smp.c).  The only one who waits would be the
 883caller, but there still are a bunch of cache misses in the handlers.  Maybe
 884this isn't that big of a deal, and the RCU thing is an unnecessary
 885optimization.
 886
 8877.4 Just Wait til a Context Switch
 888---------------------------
 889Another option is to not bother freeing the page until the entire process is
 890descheduled.  This could be a very long time, and also will mess with
 891userspace's semantics.  They would be running user code that could still
 892access the old page, so in essence this is a lazy munmap/mprotect.  The
 893process basically has the page in pergatory: it can't be reallocated, and it
 894might be accessible, but can't be guaranteed to work.
 895
 896The main benefit of this is that you don't need to send the TLB shootdown IPI
 897at all - so you don't interfere with the app.  Though in return, they have
 898possibly weird semantics.  One aspect of these weird semantics is that the
 899same virtual address could map to two different pages - that seems like a
 900disaster waiting to happen.  We could also block that range of the virtual
 901address space from being reallocated, but that gets even more tricky.
 902
 903One issue with just waiting and RCU is memory pressure.  If we actually need
 904the page, we will need to enforce an unmapping, which sucks a little.
 905
 9067.5 Bulk vs Single
 907---------------------------
 908If there are a lot of pages being shot down, it'd be best to amortize the cost
 909of the kernel messages, as well as the invlpg calls (single page shootdowns).
 910One option would be for the kmsg to take a range, and not just a single
 911address.  This would help with bulk munmap/mprotects.  Based on the number of
 912these, perhaps a raw tlbflush (the entire TLB) would be worth while, instead
 913of n single shots.  Odds are, that number is arch and possibly workload
 914specific.
 915
 916For now, the plan will be to send a range and have them individually shot
 917down.
 918
 9197.6 Don't do it
 920---------------------------
 921Either way, munmap/mprotect sucks in an MCP.  I recommend not doing it, and
 922doing the appropriate mmap/munmap/mprotects in _S mode.  Unfortunately, even
 923our crap pthread library munmaps on demand as threads are created and
 924destroyed.  The vcore code probably does in the bowels of glibc's TLS code
 925too, though at least that isn't on every user context switch.
 926
 9277.7 Local memory
 928---------------------------
 929Private local memory would help with this too.  If each vcore has its own
 930range, we won't need to send TLB shootdowns for those areas, and we won't have
 931to worry about weird application semantics.  The downside is we would need to
 932do these mmaps in certain ranges in advance, and might not easily be able to
 933do them remotely.  More on this when we actually design and build it.
 934
 9357.8 Future Hardware Support
 936---------------------------
 937It would be cool and interesting if we had the ability to remotely shootdown
 938TLBs.  For instance, all cores with cr3 == X, shootdown range Y..Z.  It's
 939basically what we'll do with the kernel message and the vcoremap, but with
 940magic hardware.
 941
 9427.9 Current Status
 943---------------------------
 944For now, we just send a kernel message to all vcores to do a full TLB flush,
 945and not to worry about checklists, waiting, or anything.  This is due to being
 946short on time and not wanting to sort out the issue with ranges.  The way
 947it'll get changed to is to send the kmsg with the range to the appropriate
 948cores, and then maybe put the page on the end of the freelist (instead of the
 949head).  More to come.
 950
 9518. Process Management
 952===========================
 9538.1 Vcore lists
 954---------------------------
 955We have three lists to track a process's vcores.  The vcores themselves sit in
 956the vcoremap in procinfo; they aren't dynamically allocated (memory) or
 957anything like that.  The lists greatly eases vcore discovery and management.
 958
 959A vcore is on exactly one of three lists: online (mapped and running vcores,
 960sometimes called 'active'), bulk_preempt (was online when the process was bulk
 961preempted (like a timeslice)), and inactive (yielded, hasn't come on yet,
 962etc).  When writes are complete (unlocked), either the online list or the
 963bulk_preempt list should be empty.
 964
 965List modifications are protected by the proc_lock.  You can concurrently read,
 966but note you may get some weird behavior, such as a vcore on multiple lists, a
 967vcore on no lists, online and bulk_preempt both having items, etc.  Currently,
 968event code will read these lists when hunting for a suitable core, and will
 969have to be careful about races.  I want to avoid event FALLBACK code from
 970grabbing the proc_lock.
 971
 972Another slight thing to be careful of is that the vcore lists don't always
 973agree with the vcore mapping.  However, it will always agree with what the
 974state of the process will be when all kmsgs are processed (fate).
 975Specifically, when we take vcores, the unmapping happens with the lock not
 976held on the vcore itself (as discussed elsewhere).  The vcore lists represent
 977the result of those pending unmaps.
 978
 979Before we used the lists, we scanned the vcoremap in a painful, clunky manner.
 980In the old style, when you asked for a vcore, the first one you got was the
 981first hole in the vcoremap.  Ex: Vcore0 would always be granted if it was
 982offline.  That's no longer true; the most recent vcore yielded will be given
 983out next.  This will help with cache locality, and also cuts down on the
 984scenarios on which the kernel gives out a vcore that userspace wasn't
 985expecting.  This can still happen if they ask for more vcores than they set up
 986for, or if a vcore doesn't *want* to come online (there's a couple scenarios
 987with preemption recovery where that may come up).
 988
 989So the plan with the bulk preempt list is that vcores on it were preempted,
 990and the kernel will attempt to restart all of them (and move them to the online
 991list).  Any leftovers will be moved to the inactive list, and have preemption
 992recovery messages sent out.  Any shortages (they want more vcores than were
 993bulk_preempted) will be taken from the yield list.  This all means that
 994whether or not a vcore needs to be preempt-recovered or if there is a message
 995out about its preemption doesn't really affect which list it is on.  You could
 996have a vcore on the inactive list that was bulk preempted (and not turned back
 997on), and then that vcore gets granted in the next round of vcore_requests().
 998The preemption recovery handlers will need to deal with concurrent handlers
 999and the vcore itself starting back up.
1000
10019. On the Ordering of Messages and Bugs with Old State
1002===========================
1003This is a sordid tale involving message ordering, message delivery times, and
1004finding out (sometimes too late) that the state you expected is gone and
1005having to deal with that error.
1006
1007A few design issues:
1008- being able to send messages and have them execute in the order they are
1009  sent
1010- having message handlers resolve issues with global state.  Some need to know
1011  the correct 'world view', and others need to know what was the state at the
1012  time they were sent.
1013- realizing syscalls, traps, faults, and any non-IRQ entry into the kernel is
1014  really a message.
1015
1016Process management messages have alternated from ROUTINE to IMMEDIATE and now
1017back to ROUTINE.  These messages include such family favorites as
1018'__startcore', '__preempt', etc.  Meanwhile, syscalls were coming in that
1019needed to know about the core and the process's state (specifically, yield,
1020change_to, and get_vcoreid).  Finally, we wanted to avoid locking, esp in
1021KMSGs handlers (imagine all cores grabbing the lock to check the vcoremap or
1022something).
1023
1024Incidentally, events were being delivered concurretly to vcores, though that
1025actually didn't matter much (check out async_events.txt for more on that).
1026
10279.1: Design Guidelines
1028---------------------------
1029Initially, we wanted to keep broadcast messaging available as an option.  As
1030noted elsewhere, we can't really do this well for startcore, since most
1031hardware broadcast options need some initial per-core setup, and any sort of
1032broadcast tree we make should be able to handle a small message.  Anyway, this
1033desire in the early code to keep all messages identical lead to a few
1034problems.
1035
1036Another objective of the kernel messaging was to avoid having the message
1037handlers grab any locks, especially the same lock (the proc lock is used to
1038protect the vcore map, for instance).
1039
1040Later on, a few needs popped up that motivated the changes discussed below:
1041- Being able to find out which proc/vcore was on a pcore
1042- Not having syscalls/traps require crazy logic if the carpet was pulled out
1043  from under them.
1044- Having proc management calls return.  This one was sorted out by making all
1045  kmsg handlers return.  It would be a nightmare making a ksched without this.
1046
10479.2: Looking at Old State: a New Bug for an Old Problem
1048---------------------------
1049We've always had issues with syscalls coming in and already had the fate of a
1050core determined.  This is referred to in a few places as "predetermined fate"
1051vs "local state".  A remote lock holder (ksched) already determined a core
1052should be unmapped and sent a message.  Only later does some call like
1053proc_yield() realize its core is already *unmapped*. (I use that term poorly
1054here).  This sort of code had to realize it was working on an old version of
1055state and just abort.  This was usually safe, though looking at the vcoremap
1056was a bad idea.  Initially, we used preempt_served as the signal, which was
1057okay.  Around 12b06586 yield started to use the vcoremap, which turned out to
1058be wrong.
1059
1060A similar issue happens for the vcore messages (startcore, preempt, etc).  The
1061way startcore used to work was that it would only know what pcore it was on,
1062and then look into the vcoremap to figure out what vcoreid it should be
1063running.  This was to keep broadcast messaging available as an option.  The
1064problem with it is that the vcoremap may have changed between when the
1065messages were sent and when they were executed.  Imagine a startcore followed
1066by a preempt, afterwhich the vcore was unmapped.  Well, to get around that, we
1067had the unmapping happen in the preempt or death handlers.  Yikes!  This was
1068the case back in the early days of ROS.  This meant the vcoremap wasn't
1069actually representative of the decisions the ksched made - we also needed to
1070look at the state we'd have after all outstanding messages executed.  And this
1071would differ from the vcore lists (which were correct for a lock holder).
1072
1073This was managable for a little while, until I tried to conclusively know who
1074owned a particular pcore.  This came up while making a provisioning scheduler.
1075Given a pcore, tell me which process/vcore (if any) were on it.  It was rather
1076tough.  Getting the proc wasn't too hard, but knowing which vcore was a little
1077tougher.  (Note the ksched doesn't care about which vcore is running, and the
1078process can change vcores on a pcore at will).  But once you start looking at
1079the process, you can't tell which vcore a certain pcore has.  The vcoremap may
1080be wrong, since a preempt is already on the way.  You would have had to scan
1081the vcore lists to see if the proc code thought that vcore was online or not
1082(which would mean there had been no preempts).  This is the pain I was talking
1083about back around commit 5343a74e0.
1084
1085So I changed things so that the vcoremap was always correct for lock holders,
1086and used pcpui to track owning_vcoreid (for preempt/notify), and used an extra
1087KMSG variable to tell startcore which vcoreid it should use.  In doing so, we
1088(re)created the issue that the delayed unmapping dealt with: the vcoremap
1089would represent *now*, and not the vcoremap of when the messages were first
1090sent.  However, this had little to do with the KMSGs, which I was originally
1091worried about.  No one was looking at the vcoremap without the lock, so the
1092KMSGs were okay, but remember: syscalls are like messages too.  They needed to
1093figure out what vcore they were on, i.e. what vcore userspace was making
1094requests on (viewing a trap/fault as a type of request).
1095
1096Now the problem was that we were using the vcoremap to figure out which vcore
1097we were supposed to be.  When a syscall finally ran, the vcoremap could be
1098completely wrong, and with immediate KMSGs (discussed below), the pcpui was
1099already changed!  We dealt with the problem for KMSGs, but not syscalls, and
1100basically reintroduced the bug of looking at current state and thinking it
1101represented the state from when the 'message' was sent (when we trapped into
1102the kernel, for a syscall/exception).
1103
11049.3: Message Delivery, Circular Waiting, and Having the Carpet Pulled Out
1105---------------------------
1106In-order message delivery was what drove me to build the kernel messaging
1107system in the first place.  It provides in-order messages to a particular
1108pcore.  This was enough for a few scenarios, such as preempts racing ahead of
1109startcores, or deaths racing a head of preempts, etc.  However, I also wanted
1110an ordering of messages related to a particular vcore, and this wasn't
1111apparent early on.
1112
1113The issue first popped up with a startcore coming quickly on the heals of a
1114preempt for the same VC, but on different PCs.  The startcore cannot proceed
1115until the preempt saved the TF into the VCPD.  The old way of dealing with
1116this was to spin in '__map_vcore()'.  This was problematic, since it meant we
1117were spinning while holding a lock, and resulted in some minor bugs and issues
1118with lock ordering and IRQ disabling (couldn't disable IRQs and then try to
1119grab the lock, since the lock holder could have sent you a message and is
1120waiting for you to handle the IRQ/IMMED KMSG).  However, it was doable.  But
1121what wasn't doable was to have the KMSGs be ROUTINE.  Any syscalls that tried
1122to grab the proc lock (lots of them) would deadlock, since the lock holder was
1123waiting on us to handle the preempt (same circular waiting issue as above).
1124
1125This was fine, albeit subpar, until a new issue showed up.  Sending IMMED
1126KMSGs worked fine if we were coming from userspace already, but if we were in
1127the kernel, those messages would run immediately (hence the name), just like
1128an IRQ handler, and could confuse syscalls that touched cur_ctx/pcpui.  If a
1129preempt came in during a syscall, the process/vcore could be changed before
1130the syscall took place.  Some syscalls could handle this, albeit poorly.
1131sys_proc_yield() and sys_change_vcore() delicately tried to detect if they
1132were still mapped or not and use that to determine if a preemption happened.
1133
1134As mentioned above, looking at the vcoremap only tells you what is currently
1135happening, and not what happened in the past.  Specifically, it doesn't tell
1136you the state of the mapping when a particular core trapped into the kernel
1137for a syscall (referred to as when the 'message' was sent up above).  Imagine
1138sys_get_vcoreid(): you trap in, then immediately get preempted, then startcore
1139for the same process but a different vcoreid.  The syscall would return with
1140the vcoreid of the new vcore, since it cannot tell there was a change.  The
1141async syscall would complete and we'd have a wrong answer.  While this never
1142happened to me, I had a similar issue while debugging some other bugs (I'd get
1143a vcoreid of 0xdeadbeef, for instance, which was the old poison value for an
1144unmapped vcoreid).  There are a bunch of other scenarios that trigger similar
1145disasters, and they are very hard to avoid.
1146
1147One way out of this was a per-core history counter, that changed whenever we
1148changed cur_ctx.  Then when we trapped in for a syscall, we could save the
1149value, enable_irqs(), and go about our business.  Later on, we'd have to
1150disable_irqs() and compare the counters.  If they were different, we'd have to
1151bail out some how.  This could have worked for change_to and yield, and some
1152others.  But any syscall that wanted to operate on cur_ctx in some way would
1153fail (imagine a hypothetical sys_change_stack_pointer()).  The context that
1154trapped has already returned on another core.  I guess we could just fail that
1155syscall, though it seems a little silly to not be able to do that.
1156
1157The previous example was a bit contrived, but lets also remember that it isn't
1158just syscalls: all exceptions have the same issue.  Faults might be fixable,
1159since if you restart a faulting context, it will start on the faulting
1160instruction.  However all traps (like syscall) restart on the next
1161instruction.  Hope we don't want to do anything fancy with breakpoint!  Note
1162that I had breakpointing contexts restart on other pcores and continue while I
1163was in the breakpoint handler (noticed while I was debugging some bugs with
1164lots of preempts).  Yikes.  And don't forget we eventually want to do some
1165complicated things with the page fault handler, and may want to turn on
1166interrupts / kthread during a page fault (imaging hitting disk).  Yikes.
1167
1168So I looked into going back to ROUTINE kernel messages.  With ROUTINE
1169messages, I didn't have to worry about having the carpet pulled out from under
1170syscalls and exceptions (traps, faults, etc).  The 'carpet' is stuff like
1171cur_ctx, owning_proc, owning_vcoreid, etc.  We still cannot trust the vcoremap,
1172unless we *know* there were no preempts or other KMSGs waiting for us.
1173(Incidentally, in the recent fix a93aa7559, we merely use the vcoremap as a
1174sanity check).
1175
1176However, we can't just switch back to ROUTINEs.  Remember: with ROUTINEs,
1177we will deadlock in '__map_vcore()', when it waits for the completion of
1178preempt.  Ideally, we would have had startcore spin on the signal.  Since we
1179already gave up on using x86-style broadcast IPIs for startcore (in
11805343a74e0), we might as well pass along a history counter, so it knows to wait
1181on preempt.
1182
11839.4: The Solution
1184---------------------------
1185To fix up all of this, we now detect preemptions in syscalls/traps and order
1186our kernel messages with two simple per-vcore counters.  Whenever we send a
1187preempt, we up one counter.  Whenever that preempt finishes, it ups another
1188counter.  When we send out startcores, we send a copy of the first counter.
1189This is a way of telling startcore where it belongs in the list of messages.
1190More specifically, it tells it which preempt happens-before it.
1191
1192Basically, I wanted a partial ordering on my messages, so that messages sent
1193to a particular vcore are handled in the order they were sent, even if those
1194messages run on different physical cores.
1195
1196It is not sufficient to use a seq counter (one integer, odd values for
1197'preempt in progress' and even values for 'preempt done').  It is possible to
1198have multiple preempts in flight for the same vcore, albeit with startcores in
1199between.  Still, there's no way to encode that scenario in just one counter.
1200
1201Here's a normal example of traffic to some vcore.  I note both the sending and
1202the execution of the kmsgs:
1203   nr_pre_sent    nr_pre_done    pcore     message sent/status
1204   -------------------------------------------------------------
1205   0              0              X         startcore (nr_pre_sent == 0)
1206   0              0              X         startcore (executes)
1207   1              0              X         preempt   (kmsg sent)
1208   1              1              Y         preempt   (executes)
1209   1              1              Y         startcore (nr_pre_sent == 1)
1210   1              1              Y         startcore (executes)
1211
1212Note the messages are always sent by the lockholder in the order of the
1213example above.
1214
1215Here's when the startcore gets ahead of the prior preempt:
1216   nr_pre_sent    nr_pre_done    pcore     message sent/status
1217   -------------------------------------------------------------
1218   0              0              X         startcore (nr_pre_sent == 0) 
1219   0              0              X         startcore (executes)
1220   1              0              X         preempt   (kmsg sent)
1221   1              0              Y         startcore (nr_pre_sent == 1)
1222   1              1              X         preempt   (executes)
1223   1              1              Y         startcore (executes)
1224
1225Note that this can only happen across cores, since KMSGs to a particular core
1226are handled in order (for a given class of message).  The startcore blocks on
1227the prior preempt.
1228
1229Finally, here's an example of what a seq ctr can't handle:
1230   nr_pre_sent    nr_pre_done    pcore     message sent/status
1231   -------------------------------------------------------------
1232   0              0              X         startcore (nr_pre_sent == 0) 
1233   1              0              X         preempt   (kmsg sent)
1234   1              0              Y         startcore (nr_pre_sent == 1)
1235   2              0              Y         preempt   (kmsg sent)
1236   2              0              Z         startcore (nr_pre_sent == 2)
1237   2              1              X         preempt   (executes (upped to 1))
1238   2              1              Y         startcore (executes (needed 1))
1239   2              2              Y         preempt   (executes (upped to 2))
1240   2              Z              Z         startcore (executes (needed 2))
1241
1242As a nice bonus, it is easy for syscalls that care about the vcoreid (yield,
1243change_to, get_vcoreid) to check if they have a preempt_served.  Just grab the
1244lock (to prevent further messages being sent), then check the counters.  If
1245they are equal, there is no preempt on its way.  This actually was the
1246original way we checked for preempts in proc_yield back in the day.  It was
1247just called preempt_served.  Now, it is split into two counters, instead of
1248just being a bool.  
1249
1250Regardless of whether or not we were preempted, we still can look at
1251pcpui->owning_proc and owning_vcoreid to figure out what the vcoreid of the
1252trap/syscall is, and we know that the cur_ctx is still the correct cur_ctx (no
1253carpet pulled out), since while there could be a preempt ROUTINE message
1254waiting for us, we simply haven't run it yet.  So calls like yield should
1255still fail (since your core has been unmapped and you need to bail out and run
1256the preempt handler), but calls like sys_change_stack_pointer can proceed.
1257More importantly than that old joke syscall, the page fault handler can try to
1258do some cool things without worrying about really crazy stuff.
1259
12609.5: Why We (probably) Don't Deadlock
1261---------------------------
1262It's worth thinking about why this setup of preempts and startcores can't
1263deadlock.  Anytime we spin in the kernel, we ought to do this.  Perhaps there
1264is some issue with other KMSGs for other processes, or other vcores, or
1265something like that that can cause a deadlock.
1266
1267Hypothetical case: pcore 1 has a startcore for vc1 which is stuck behind vc2's
1268startcore on PC2, with time going upwards.  In these examples, startcores are
1269waiting on particular preempts, subject to the nr_preempts_sent parameter sent
1270along with the startcores.
1271
1272^                       
1273|            _________                 _________
1274|           |         |               |         |
1275|           | pr vc 2 |               | pr vc 1 |
1276|           |_________|               |_________|
1277|
1278|            _________                 _________
1279|           |         |               |         |
1280|           | sc vc 1 |               | sc vc 2 |
1281|           |_________|               |_________|
1282t           
1283---------------------------------------------------------------------------
1284              ______                    ______
1285             |      |                  |      |
1286             | PC 1 |                  | PC 2 |
1287             |______|                  |______|
1288
1289Here's the same picture, but with certain happens-before arrows.  We'll use X --> Y to
1290mean X happened before Y, was sent before Y.  e.g., a startcore is sent after
1291a preempt.
1292
1293^                       
1294|            _________                 _________
1295|           |         |               |         |
1296|       .-> | pr vc 2 | --.    .----- | pr vc 1 | <-.
1297|       |   |_________|    \  /   &   |_________|   |  
1298|     * |                   \/                      | * 
1299|       |    _________      /\         _________    |  
1300|       |   |         |    /  \   &   |         |   |  
1301|       '-- | sc vc 1 | <-'    '----> | sc vc 2 | --'
1302|           |_________|               |_________|
1303t           
1304---------------------------------------------------------------------------
1305              ______                    ______
1306             |      |                  |      |
1307             | PC 1 |                  | PC 2 |
1308             |______|                  |______|
1309
1310The arrows marked with * are ordered like that due to the property of KMSGs,
1311in that we have in order delivery.  Messages are executed in the order in
1312which they were sent (serialized with a spinlock btw), so on any pcore,
1313messages that are further ahead in the queue were sent before (and thus will
1314be run before) other messages.
1315
1316The arrows marked with a & are ordered like that due to how the proc
1317management code works.  The kernel won't send out a startcore for a particular
1318vcore before it sent out a preempt.  (Note that techincally, preempts follow
1319startcores.  The startcores in this example are when we start up a vcore after
1320it had been preempted in the past.).
1321
1322Anyway, note that we have a cycle, where all events happened before each
1323other, which isn't possible.  The trick to connecting "unrelated" events like
1324this (unrelated meaning 'not about the same vcore') in a happens-before manner
1325is the in-order properties of the KMSGs.
1326
1327Based on this example, we can derive general rules.  Note that 'sc vc 2' could
1328be any kmsg that waits on another message placed behind 'sc vc 1'.  This would
1329require us having sent a KMSG that waits on a KMSGs that we send later.  Bad
1330idea!  (you could have sent that KMSGs to yourself, aside from just being
1331dangerous).  If you want to spin, make sure you send the work that should
1332happen-before actually-before the waiter.
1333
1334In fact, we don't even need 'sc vc 2' to be a KMSG.  It could be miscellaneous
1335kernel code, like a proc mgmt syscall.  Imagine if we did something like the
1336old '__map_vcore' call from within the ksched.  That would be code that holds
1337the lock, and then waits on the execution of a message handler.  That would
1338deadlock (which is why we don't do it anymore).
1339
1340Finally, in case this isn't clear, all of the startcores and preempts for
1341a given vcore exist in a happens-before relation, both in sending and in
1342execution.  The sending aspect is handled by proc mgmt code.  For execution,
1343preempts always follow startcores due to the KMSG ordering property.  For
1344execution of startcores, startcores always spin until the preempt they follow
1345is complete, ensuring the execution of the main part of their handler happens
1346after the prior preempt.
1347
1348Here's some good ideas for the ordering of locks/irqs/messages:
1349- You can't hold a spinlock of any sort and then wait on a routine kernel
1350  message.  The core where that runs may be waiting on you, or some scenario
1351  like above.
1352        - Similarly, think about how this works with kthreads.  A kthread
1353          restart is a routine KMSG.  You shouldn't be waiting on code that
1354          could end up kthreading, mostly because those calls block!
1355- You can hold a spinlock and wait on an IMMED kmsg, if the waiters of the
1356  spinlock have irqs enabled while spinning (this is what we used to do with
1357  the proc lock and IMMED kmsgs, and 54c6008 is an example of doing it wrong)
1358        - As a corollary, locks like this cannot be irqsave, since the other
1359          attempted locker will have irq disabled
1360- For broadcast trees, you'd have to send IMMEDs for the intermediates, and
1361  then it'd be okay to wait on those intermediate, immediate messages (if we
1362  wanted confirmation of the posting of RKM)
1363        - The main thing any broadcast mechanism needs to do is make sure all
1364          messages get delivered in order to particular pcores (the central
1365          premise of KMSGs) (and not deadlock due to waiting on a KMSG
1366          improperly)
1367- Alternatively, we could use routines for the intermediates if we didn't want
1368  to wait for RKMs to hit their destination, we'd need to always use the same
1369  proxy for the same destination pcore, e.g., core 16 always covers 16-31.
1370        - Otherwise, we couldn't guarantee the ordering of SC before PR before
1371          another SC (which the proc_lock and proc mgmt code does); we need the
1372          ordering of intermediate msgs on the message queues of a particular
1373          core.
1374        - All kmsgs would need to use this broadcasting style (couldn't mix
1375          regular direct messages with broadcast), so odds are this style would
1376          be of limited use.
1377        - since we're not waiting on execution of a message, we could use RKMs
1378          (while holding a spinlock)
1379- There might be some bad effects with kthreads delaying the reception of RKMS
1380  for a while, but probably not catastrophically.
1381
13829.6: Things That We Don't Handle Nicely
1383---------------------------
1384If for some reason a syscall or fault handler blocks *unexpectedly*, we could
1385have issues.  Imagine if change_to happens to block in some early syscall code
1386(like instrumentation, or who knows what, that blocks in memory allocation).
1387When the syscall kthread restarts, its old cur_ctx is gone.  It may or may not
1388be running on a core owned by the original process.  If it was, we probably
1389would accidentally yield that vcore (clearly a bug).  
1390
1391For now, any of these calls that care about cur_ctx/pcpui need to not block
1392without some sort of protection.  None of them do, but in the future we might
1393do something that causes them to block.  We could deal with it by having a
1394pcpu or per-kthread/syscall flag that says if it ever blocked, and possibly
1395abort.  We get into similar nasty areas as with preempts, but this time, we
1396can't solve it by making preempt a routine KMSG - we block as part of that
1397syscall/handler code.  Odds are, we'll just have to outlaw this, now and
1398forever.  Just note that if a syscall/handler blocks, the TF it came in on is
1399probably not cur_ctx any longer, and that old cur_ctx has probably restarted.
1400
140110. TBD
1402===========================
1403