Google Git
Sign in
akaros / upstream / 6ad62964edcdf2e026d465e5e34381119cf42f8f / . / Documentation / process-internals.txt
blob: 78aae558156151829fbe853d968f35098bf292d0 [file] [log] [blame]
process-internals.txt
Barret Rhoden
This discusses core issues with process design and implementation. Most of this
info is available in the source in the comments (but may not be in the future).
For now, it's a dumping ground for topics that people ought to understand before
they muck with how processes work.
Contents:
1. Reference Counting
2. When Do We Really Leave "Process Context"?
3. Leaving the Kernel Stack
4. Preemption and Notification Issues
5. current_ctx and owning_proc
6. Locking!
7. TLB Coherency
8. Process Management
9. On the Ordering of Messages
10. TBD
1. Reference Counting
===========================
1.1 Basics:
---------------------------
Reference counts exist to keep a process alive. We use krefs for this, similar
to Linux's kref:
- Can only incref if the current value is greater than 0, meaning there is
already a reference to it. It is a bug to try to incref on something that has
no references, so always make sure you incref something that you know has a
reference. If you don't know, you need to get it from pid2proc (which is a
careful way of doing this - pid2proc kref_get_not_zero()s on the reference that is
stored inside it). If you incref and there are 0 references, the kernel will
panic. Fix your bug / don't incref random pointers.
- Can always decref.
- When the decref returns 0, perform some operation. This does some final
cleanup on the object.
- Process code is trickier since we frequently make references from 'current'
(which isn't too bad), but also because we often do not return and need to be
careful about the references we passed in to a no-return function.
1.2 Brief History of the Refcnt:
---------------------------
Originally, the refcnt was created to keep page tables from being destroyed (in
proc_free()) while cores were still using them, which is what was happens during
an ARSC (async remote syscall). It was then defined to be a count of places in
the kernel that had an interest in the process staying alive, practically just
to protect current/cr3. This 'interest' actually extends to any code holding a
pointer to the proc, such as one acquired via pid2proc(), which is its current
meaning.
1.3 Quick Aside: The current Macro:
---------------------------
current is a pointer to the proc that is currently loaded/running on any given
core. It is stored in the per_cpu_info struct, and set/managed by low-level
process code. It is necessary for the kernel to quickly figure out who is
running on its core, especially when servicing interrupts and traps. current is
protected by a refcnt.
current does not say which process owns / will-run on a core. The per-cpu
variable 'owning_proc' covers that. 'owning_proc' should be treated like
'current' (aka, 'cur_proc') when it comes to reference counting. Like all
refcnts, you can use it, but you can't consume it without atomically either
upping the refcnt or passing the reference (clearing the variable storing the
reference). Don't pass it to a function that will consume it and not return
without upping it.
1.4 Reference Counting Rules:
---------------------------
+1 for existing.
- The fact that the process is supposed to exist is worth +1. When it is time
to die, we decref, and it will eventually be cleaned up. This existence is
explicitly kref_put()d in proc_destroy().
- The hash table is a bit tricky. We need to kref_get_not_zero() when it is
locked, so we know we aren't racing with proc_free freeing the proc and
removing it from the list. After removing it from the hash, we don't need to
kref_put it, since it was an internal ref. The kref (i.e. external) isn't for
being on the hash list, it's for existing. This separation allows us to
remove the proc from the hash list in the "release" function. See kref.txt
for more details.
+1 for someone using it or planning to use it.
- This includes simply having a pointer to the proc, since presumably you will
use it. pid2proc() will incref for you. When you are done, decref.
- Functions that create a process and return a pointer (like proc_create() or
kfs_proc_create()) will also up the refcnt for you. Decref when you're done.
- If the *proc is stored somewhere where it will be used again, such as in an IO
continuation, it needs to be refcnt'd. Note that if you already had a
reference from pid2proc(), simply don't decref after you store the pointer.
+1 for current.
- current counts as someone using it (expressing interest in the core), but is
also a source of the pointer, so its a bit different. Note that all kref's
are sources of a pointer. When we are running on a core that has current
loaded, the ref is both for its usage as well as for being the current
process.
- You have a reference from current and can use it without refcnting, but
anything that needs to eat a reference or store/use it needs an incref first.
To be clear, your reference is *NOT* edible. It protects the cr3, guarantees
the process won't die, and serves as a bootstrappable reference.
- Specifically, if you get a ref from current, but then save it somewhere (like
an IO continuation request), then clearly you must incref, since it's both
current and stored/used.
- If you don't know what might be downstream from your function, then incref
before passing the reference, and decref when it returns. We used to do this
for all syscalls, but now only do it for calls that might not return and
expect to receive reference (like proc_yield).
All functions that take a *proc have a refcnt'd reference, though it may not be
edible (it could be current). It is the callers responsibility to make sure
it'd edible if it necessary. It is the callees responsibility to incref if it
stores or makes a copy of the reference.
1.5 Functions That Don't or Might Not Return:
---------------------------
Refcnting and especially decreffing gets tricky when there are functions that
MAY not return. proc_restartcore() does not return (it pops into userspace).
proc_run() used to not return, if the core it was called on would pop into
userspace (if it was a _S, or if the core is part of the vcoremap for a _M).
This doesn't happen anymore, since we have cur_ctx in the per-cpu info.
Functions that MAY not return will "eat" your reference *IF* they do not return.
This means that you must have a reference when you call them (like always), and
that reference will be consumed / decref'd for you if the function doesn't
return. Or something similarly appropriate.
Arguably, for functions that MAY not return, but will always be called with
current's reference (proc_yield()), we could get away without giving it an
edible reference, and then never eating the ref. Yield needs to be reworked
anyway, so it's not a bit deal yet.
We do this because when the function does not return, you will not have the
chance to decref (your decref code will never run). We need the reference when
going in to keep the object alive (like with any other refcnt). We can't have
the function always eat the reference, since you cannot simply re-incref the
pointer (not allowed to incref unless you know you had a good reference). You'd
have to do something like p = pid2proc(p_pid); It's clunky to do that, easy to
screw up, and semantically, if the function returns, then we may still have an
interest in p and should decref later.
The downside is that functions need to determine if they will return or not,
which can be a pain (for an out-of-date example: a linear time search when
running an _M, for instance, which can suck if we are trying to use a
broadcast/logical IPI).
As the caller, you usually won't know if the function will return or not, so you
need to provide a consumable reference. Current doesn't count. For example,
proc_run() requires a reference. You can proc_run(p), and use p afterwards, and
later decref. You need to make sure you have a reference, so things like
proc_run(pid2proc(55)) works, since pid2proc() increfs for you. But you cannot
proc_run(current), unless you incref current in advance. Incidentally,
proc_running current doesn't make a lot of sense.
1.6 Runnable List:
---------------------------
Procs on the runnable list need to have a refcnt (other than the +1 for
existing). It's something that cares that the process exists. We could have
had it implicitly be refcnt'd (the fact that it's on the list is enough, sort of
as if it was part of the +1 for existing), but that complicates things. For
instance, it is a source of a reference (for the scheduler) and you could not
proc_run() a process from the runnable list without worrying about increfing it
before hand. This isn't true anymore, but the runnable lists are getting
overhauled anyway. We'll see what works nicely.
1.7 Internal Details for Specific Functions:
---------------------------
proc_run()/__proc_give_cores(): makes sure enough refcnts are in place for all
places that will install owning_proc. This also makes it easier on the system
(one big incref(n), instead of n increfs of (1) from multiple cores).
__set_proc_current() is a helper that makes sure p is the cur_proc. It will
incref if installing a new reference to p. If it removed an old proc, it will
decref.
__proc_startcore(): assumes all references to p are sorted. It will not
return, and you should not pass it a reference you need to decref(). Passing
it 'owning_proc' works, since you don't want to decref owning_proc.
proc_destroy(): it used to not return, and back then if your reference was
from 'current', you needed to incref. Now that proc_destroy() returns, it
isn't a big deal. Just keep in mind that if you have a function that doesn't
return, there's no way for the function to know if it's passed reference is
edible. Even if p == current, proc_destroy() can't tell if you sent it p (and
had a reference) or current and didn't.
proc_yield(): when this doesn't return, it eats your reference. It will also
decref twice. Once when it clears_owning_proc, and again when it calls
abandon_core() (which clears cur_proc).
abandon_core(): it was not given a reference, so it doesn't eat one. It will
decref when it unloads the cr3. Note that this is a potential performance
issue. When preempting or killing, there are n cores that are fighting for the
cacheline to decref. An alternative would be to have one core decref for all n
cores, after it knows all cores unloaded the cr3. This would be a good use of
the checklist (possibly with one cacheline per core). It would take a large
amount of memory for better scalability.
1.8 Things I Could Have Done But Didn't And Why:
---------------------------
Q: Could we have the first reference (existence) mean it could be on the runnable
list or otherwise in the proc system (but not other subsystems)? In this case,
proc_run() won't need to eat a reference at all - it will just incref for every
current it will set up.
New A: Maybe, now that proc_run() returns.
Old A: No: if you pid2proc(), then proc_run() but never return, you have (and
lose) an extra reference. We need proc_run() to eat the reference when it
does not return. If you decref between pid2proc() and proc_run(), there's a
(rare) race where the refcnt hits 0 by someone else trying to kill it. While
proc_run() will check to see if someone else is trying to kill it, there's a
slight chance that the struct will be reused and recreated. It'll probably
never happen, but it could, and out of principle we shouldn't be referencing
memory after it's been deallocated. Avoiding races like this is one of the
reasons for our refcnt discipline.
Q: (Moot) Could proc_run() always eat your reference, which would make it
easier for its implementation?
A: Yeah, technically, but it'd be a pain, as mentioned above. You'd need to
reaquire a reference via pid2proc, and is rather easy to mess up.
Q: (Moot) Could we have made proc_destroy() take a flag, saying whether or not
it was called on current and needed a decref instead of wasting an incref?
A: We could, but won't. This is one case where the external caller is the one
that knows the function needs to decref or not. But it breaks the convention a
bit, doesn't mirror proc_create() as well, and we need to pull in the cacheline
with the refcnt anyways. So for now, no.
Q: (Moot) Could we make __proc_give_cores() simply not return if an IPI is
coming?
A: I did this originally, and manually unlocked and __wait_for_ipi()d. Though
we'd then need to deal with it like that for all of the related functions, which
doesn't work if you wanted to do something afterwards (like schedule(p)). Also
these functions are meant to be internal helpers, so returning the bool makes
more sense. It eventually led to having __proc_unlock_ipi_pending(), which made
proc_destroy() much cleaner and helped with a general model of dealing with
these issues. Win-win.
2. When Do We Really Leave "Process Context"?
===========================
2.1 Overview
---------------------------
First off, it's not really "process context" in the way Linux deals with it. We
aren't operating in kernel mode on behalf of the process (always). We are
specifically talking about when a process's cr3 is loaded on a core. Usually,
current is also set (the exception for now is when processing ARSCs).
There are a couple different ways to do this. One is to never unload a context
until something new is being run there (handled solely in __proc_startcore()).
Another way is to always explicitly leave the core, like by abandon_core()ing.
The issue with the former is that you could have contexts sitting around for a
while, and also would have a bit of extra latency when __proc_free()ing during
someone *else's* __proc_startcore() (though that could be avoided if it becomes
a real issue, via some form of reaping). You'll also probably have excessive
decrefs (based on the interactions between proc_run() and __startcore()).
The issue with the latter is excessive TLB shootdowns and corner cases. There
could be some weird cases (in core_request() for example) where the core you are
running on has the context loaded for proc A on a mgmt core, but decides to give
it to proc B.
If no process is running there, current == 0 and boot_cr3 is loaded, meaning no
process's context is loaded.
All changes to cur_proc, owning_proc, and cur_ctx need to be done with
interrupts disabled, since they change in interrupt handlers.
2.2 Here's how it is done now:
---------------------------
All code is capable of 'spamming' cur_proc (with interrupts disabled!). If it
is 0, feel free to set it to whatever process you want. All code that
requires current to be set will do so (like __proc_startcore()). The
smp_idle() path will make sure current is clear when it halts. So long as you
don't change other concurrent code's expectations, you're okay. What I mean
by that is you don't clear cur_proc while in an interrupt handler. But if it
is already 0, __startcore is allowed to set it to it's future proc (which is
an optimization). Other code didn't have any expectations of it (it was 0).
Likewise, kthread code when we sleep_on() doesn't have to keep cur_proc set.
A kthread is somewhat an isolated block (codewise), and leaving current set
when it is done is solely to avoid a TLB flush (at the cost of an incref).
In general, we try to proactively leave process context, but have the ability
to stay in context til __proc_startcore() to handle the corner cases (and to
maybe cut down the TLB flushes later). To stop proactively leaving, just
change abandon_core() to not do anything with current/cr3. You'll see weird
things like processes that won't die until their old cores are reused. The
reason we proactively leave context is to help with sanity for these issues,
and also to avoid decref's in __startcore().
A couple other details: __startcore() sorts the extra increfs, and
__proc_startcore() sorts leaving the old context. Anytime a __startcore kernel
message is sent, the sender increfs in advance for the owning_proc refcnt. As
an optimization, we can also incref to *attempt* to set current. If current
was 0, we set it. If it was already something else, we failed and need to
decref. __proc_startcore(), which the last moment before we *must* have the
cr3/current issues sorted, does the actual check if there was an old process
there or not, while it handles the lcr3 (if necessary). In general, lcr3's
ought to have refcnts near them, or else comments explaining why not.
So we leave process context when told to do so (__death/abandon_core()) or if
another process is run there. The _M code is such that a proc will stay on its
core until it receives a message, and that message would cleanup/restore a
generic context (boot_cr3). A _S could stay on its core until another _S came
in. This is much simpler for cases when a timer interrupt goes off to force a
schedule() decision. It also avoids a TLB flush in case the scheduler picked
that same proc to run again. This could also happen to an _M, if for some
reason it was given a management core (!!!) or some other event happened that
caused some management/scheduling function to run on one of it's cores (perhaps
it asked).
proc_yield() abandons the core / leaves context.
2.3 Other issues:
---------------------------
Note that dealing with interrupting processes that are in the kernel is tricky.
There is no true process context, so we can't leave a core until the kernel is
in a "safe place", i.e. it's state is bundled enough that it can be recontinued
later. Calls of this type are routine kernel messages, executed at a convenient
time (specifically, before we return to userspace in proc_restartcore().
This same thing applies to __death messages. Even though a process is dying, it
doesn't mean we can just drop whatever the kernel was doing on its behalf. For
instance, it might be holding a reference that will never get decreffed if its
stack gets dropped.
3. Leaving the Kernel Stack:
===========================
Just because a message comes in saying to kill a process, it does not mean we
should immediately abandon_core(). The problem is more obvious when there is
a preempt message, instead of a death message, but either way there is state
that needs cleaned up (refcnts that need downed, etc).
The solution to this is rather simple: don't abandon right away. That was
always somewhat the plan for preemption, but was never done for death. And
there are several other cases to worry about too. To enforce this, we expand
the old "active messages" into a generic work execution message (a kernel
message) that can be delayed or shipped to another core. These types of
messages will not be executed immediately on the receiving pcore - instead they
are on the queue for "when there's nothing else to do in the kernel", which is
checked in smp_idle() and before returning to userspace in proc_restartcore().
Additionally, these kernel messages can also be queued on an alarm queue,
delaying their activation as part of a generic kernel alarm facility.
One subtlety is that __proc_startcore() shouldn't check for messages, since it
is called by __startcore (a message). Checking there would run the messages out
of order, which is exactly what we are trying to avoid (total chaos). No one
should call __proc_startcore, other than proc_restartcore() or __startcore().
If we ever have functions that do so, if they are not called from a message,
they must check for outstanding messages.
This last subtlety is why we needed to change proc_run()'s _S case to use a
local message instead of calling proc_starcore (and why no one should ever call
proc_startcore()). We could unlock, thereby freeing another core to change the
proc state and send a message to us, then try to proc_startcore, and then
reading the message before we had installed current or had a userspace TF to
preempt, and probably a few other things. Treating _S as a local message is
cleaner, begs to be merged in the code with _M's code, and uses the messaging
infrastructure to avoid all the races that it was created to handle.
Incidentally, we don't need to worry about missing messages while trying to pop
back to userspace from __proc_startcore, since an IPI will be on the way
(possibly a self-ipi caused by the __kernel_message() handler). This is also
why we needed to make process_routine_kmsg() keep interrupts disabled when it
stops (there's a race between checking the queue and disabling ints).
4. Preemption and Notification Issues:
===========================
4.1: Message Ordering and Local Calls:
---------------------------
Since we go with the model of cores being told what to do, there are issues
with messages being received in the wrong order. That is why we have the
kernel messages (guaranteed, in-order delivery), with the proc-lock protecting
the send order. However, this is not enough for some rare races.
Local calls can also perform the same tasks as messages (calling
proc_destroy() while a death IPI is on its way). We refer to these calls as
messing with "local fate" (compared to global state (we're clever).
Preempting a single vcore doesn't change the process's state). These calls
are a little different, because they also involve a check to see if it should
perform the function or other action (e.g., death just idling and waiting for
an IPI instead of trying to kill itself), instead of just blindly doing
something.
4.1.1: Possible Solutions
----------------
There are two ways to deal with this. One (and the better one, I think) is to
check state, and determine if it should proceed or abort. This requires that
all local-fate dependent calls always have enough state to do its job. In the
past, this meant that any function that results in sending a directive to a
vcore store enough info in the proc struct that a local call can determine if
it should take action or abort. In the past, we used the vcore/pcoremap as a
way to send info to the receiver about what vcore they are (or should be).
Now, we store that info in pcpui (for '__startcore', we send it as a
parameter. Either way, the general idea is still true: local calls can
proceed when they are called, and not self-ipi'd to a nebulous later time.
The other way is to send the work (including the checks) in a self-ipi kernel
message. This will guarantee that the message is executed after any existing
messages (making the k_msg queue the authority for what should happen to a
core). The check is also performed later (when the k_msg executes). There
are a couple issues with this: if we allow the local core to send itself an
k_msg that could be out of order (meaning it should not be sent, and is only
sent due to ignorance of its sealed fate), AND if we return the core to the
idle-core-list once its fate is sealed, we need to detect that the message is
for the wrong process and that the process is in the wrong state. To do this,
we probably need local versioning on the pcore so it can detect that the
message is late/wrong. We might get by with just the proc* (though that is
tricky with death and proc reuse), so long as we don't allow new startcores
for a proc until AFTER the preemption is completed.
4.2: Preempt-Served Flag
----------------
We want to be able to consider a pcore free once its owning proc has dealt
with removing it. This allows a scheduler-like function to easily take a core
and then give it to someone else, without waiting for each vcore to respond,
saying that the pcore is free/idle.
We used to not unmap until we were in '__preempt' or '__death', and we needed
a flag to tell yield-like calls that a message was already on the way and to
not rely on the vcoremap. This is pretty fucked up for a number of reasons,
so we changed that. But we still wanted to know when a preempt was in
progress so that the kernel could avoid giving out the vcore until the preempt
was complete.
Here's the scenario: we send a '__startcore' to core 3 for VC5->PC3. Then we
quickly send a '__preempt' to 3, and then a '__startcore' to core 4 (a
different pcore) for VC5->PC4. Imagine all of this happens before the first
'__startcore' gets processed (IRQ delay, fast ksched, whatever). We need to
not run the second '__startcore' on pcore 4 before the preemption has saved
all of the state of the VC5. So we spin on preempt_served (which may get
renamed to preempt_in_progress). We need to do this in the sender, and not
the receiver (not in the kmsg), because the kmsgs can't tell which one they
are. Specifically, the first '__startcore' on core 3 runs the same code as
the '__startcore' on core 4, working on the same vcore. Anything we tell VC5
will be seen by both PC3 and PC4. We'd end up deadlocking on PC3 while it
spins waiting for the preempt message that also needs to run on PC3.
The preempt_pending flag is actual a timestamp, with the expiration time of
the core at which the message will be sent. We could try to use that, but
since alarms aren't fired at exactly the time they are scheduled, the message
might not actually be sent yet (though it will, really soon). Still, we'll
just go with the preempt-served flag for now.
4.3: Impending Notifications
----------------
It's also possible that there is an impending notification. There's no change
in fate (though there could be a fate-changing preempt on its way), just the
user wants a notification handler to run. We need a flag anyways for this
(discussed below), so proc_yield() or whatever other local call we have can
check this flag as well.
Though for proc_yield(), it doesn't care if a notification is on its way (can
be dependent on a flag to yield from userspace, based on the nature of the
yield (which still needs to be sorted)). If the yield is in response to a
preempt_pending, it actually should yield and not receive the notification.
So it should destroy its vcoreid->pcoreid mapping and abandon_core(). When
that notification hits, it will be for a proc that isn't current, and will be
ignored (it will get run the next time that vcore fires up, handled below).
There is a slight chance that the same proc will run on that pcore, but with a
different vcoreid. In the off chance this happens, the new vcore will get a
spurious notification. Userspace needs to be able to handle spurious
notifications anyways, (there are a couple other cases, and in general it's
not hard to do), so this is not a problem. Instead of trying to have the
kernel ignore the notification, we just send a spurious one. A crappy
alternative would be to send the vcoreid with the notification, but that would
mean we can't send a generic message (broadcast) to a bunch of cores, which
will probably be a problem later.
Note that this specific case is because the "local work message" gets
processed out of order with respect to the notification. And we want this in
that case, since that proc_yield() is more important than the notification.
4.4: Preemption / Allocation Phases and Alarm Delays
---------------------------
A per-vcore preemption phase starts when the kernel marks the core's
preempt_pending flag/counter and can includes the time when an alarm is
waiting to go off to reclaim the core. The phase ends when the vcore's pcore
is reclaimed, either as a result of the kernel taking control, or because a
process voluntarily yielded.
Specifically, the preempt_pending variable is actually a timestamp for when
the core will be revoked (this assumes some form of global time, which we need
anyways). If its value is 0, then there is no preempt-pending, it is not in a
phase, and the vcore can be given out again.
When a preempt alarm goes off, the alarm only means to check a process for
expired vcores. If the vcore has been yielded while the alarm was pending,
the preempt_pending flag will be reset to 0. To speed up the search for
vcores to preempt, there's a circular buffer corelist in the proc struct, with
vcoreids of potential suspects. Or at least this will exist at some point.
Also note that the preemption list isn't bound to a specific alarm: you can
check the list at any time (not necessarily on a specific alarm), and you can
have spurious alarms (the list is empty, so it'll be a noop).
Likewise, a global preemption phase is when an entire MCP is getting
gang_prempted, and the global deadline is set. A function can quickly check
to see if the process responded, since the list of vcores with preemptions
pending will be empty.
It seems obvious, but we do not allow allocation of a vcore during its
preemption phase. The main reason is that it can potentially break
assumptions about the vcore->pcore mapping and can result in multiple
instances of the same vcore on different pcores. Imagine a preempt message
sent to a pcore (after the alarm goes off), meanwhile that vcore/pcore yields
and the vcore reactivates somewhere else. There is a potential race on the
vcore_ctx state: the new vcore is reading while the old is writing. This
issue is sorted naturally: the vcore entry in the vcoremap isn't cleared until
the vcore/pcore is actually yielded/taken away, so the code looking for a free
vcoreid slot will not try to use it.
Note that if we didn't design the alarm system to simply check for
preemptions (perhaps it has a stored list of vcores to preempt), then we
couldn't end the preempt-phase until the alarm was sorted. If that is the
case, we could easily give out a vcore that had been yielded but was still in
a preempt-phase. Stopping an alarm would be tricky too, since there could be
lots of vcores in different states that need to be sorted by the alarm (so
ripping it out isn't enough). Setting a flag might not be enough either.
Vcore version numbers/history (as well as global proc histories) is a pain I'd
like to avoid too. So don't change the alarm / delayed preemption system
without thinking about this.
Also, allowing a vcore to restart while preemptions are pending also mucks
with keeping the vcore mapping "old" (while the message is in flight). A
pcore will want to use that to determine which vcore is running on it. It
would be possible to keep a pcoremap for the reverse mapping out of sync, but
that seems like a bad idea. In general, having the pcoremap is a good idea
(whenever we talk about a vcoremap, we're usually talking about both
directions: "the vcore->pcore mapping").
4.5: Global Preemption Flags
---------------------------
If we are trying to preempt an entire process at the same time, instead of
playing with the circular buffer of vcores pending preemption, we could have a
global timer as well. This avoids some O(n) operations, though it means that
userspace needs to check two "flags" (expiration dates) when grabbing its
preempt-critical locks.
4.6: Notifications Mixed with Preemption and Sleeping
---------------------------
It is possible that notifications will mix with preemptions or come while a
process is not running. Ultimately, the process wants to be notified on a
given vcore. Whenever we send an active notification, we set a flag in procdata
(notif_pending). If the vcore is offline, we don't bother sending the IPI/notif
message. The kernel will make sure it runs the notification handler (as well as
restoring the vcore_ctx) the next time that vcore is restarted. Note that
userspace can toggle this, so they can handle the notifications from a different
core if it likes, or they can independently send a notification.
Note we use notif_pending to detect if an IPI was missed while notifs were
disabled (this is done in pop_user_ctx() by userspace). The overall meaning
of notif_pending is that a vcore wants to be IPI'd. The IPI could be
in-flight, or it could be missed. Since notification IPIs can be spurious,
when we have potential races, we err on the side of sending. This happens
when pop_user_ctx() notifies itself, and when the kernel makes sure to start a
vcore in vcore context if a notif was pending. This was simplified a bit over
the years by having uthreads always get saved into the uthread_ctx (formerly
the notif_tf), instead of in the old preempt_tf (which is now the vcore_ctx).
If a vcore has a preempt_pending, we will still send the active notification
(IPI). The core ought to get a notification for the preemption anyway, so we
need to be able to send one. Additionally, once the vcore is handling that
preemption notification, it will have notifs disabled, which will prevent us
from sending any extra notifications anyways.
4.7: Notifs While a Preempt Message is Served
---------------------------
It is possible to have the kernel handling a notification k_msg and to have a
preempt k_msg in the queue (preempt-served flag is set). Ultimately, what we
want is for the core to be preempted and the notification handler to run on
the next execution. Both messages are in the k_msg queue for "a convenient
time to leave the kernel" (I'll have a better name for that later). What we
do is execute the notification handler and jump to userspace. Since there is
still an k_msg in the queue (and we self_ipi'd ourselves, it's part of how
k_msgs work), the IPI will fire and push us right back into the kernel to
execute the preemption, and the notif handler's context will be saved in the
vcore_ctx (ready to go when the vcore gets started again).
We could try to just leave the notif_pending flag set and ignore the message,
but that would involve inspecting the queue for the preempt k_msg.
Additionally, a preempt k_msg can arrive anyway. Finally, it's possible to have
another message in the queue between the notif and the preempt, and it gets ugly
quickly trying to determine what to do.
4.8: When a Pcore is "Free"
---------------------------
There are a couple ways to handle pcores. One approach would be to not
consider them free and able to be given to another process until the old
process is completely removed (abandon_core()). Another approach is to free
the core once its fate is sealed (which we do). This probably gives more
flexibility in schedule()-like functions (no need to wait to give the core
out), quicker dispatch latencies, less contention on shared structs (like the
idle-core-map), etc.
This 'freeing' of the pcore is from the perspective of the kernel scheduler
and the proc struct. Contrary to all previous announcements, vcores are
unmapped from pcores when sending k_msgs (technically right after), while
holding the lock. The pcore isn't actually not-running-the-proc until the
kmsg completes and we abandon_core(). Previously, we used the vcoremap to
communicate to other cores in a lock-free manner, but that was pretty shitty
and now we just store the vcoreid in pcpu info.
Another tricky part is the seq_ctr used to signal userspace of changes to the
coremap or num_vcores (coremap_seqctr). While we may not even need this in the
long run, it still seems like it could be useful. The trickiness comes from
when we update the seq_ctr when we are unmapping vcores on the receive side of a
message (like __death or __preempt). We'd rather not have each pcore contend on
the seq_ctr cache line (let alone any locking) while they perform a somewhat
data-parallel task. So we continue to have the sending core handle the seq_ctr
upping and downing. This works, since the "unlocking" happens after messages
are sent, which means the receiving core is no longer in userspace (if there is
a delay, it is because the remote core is in the kernel, possibly with
interrupts disabled). Because of this, userspace will be unable to read the new
value of the seq_ctr before the IPI hits and does the unmapping that the seq_ctr
protects/advertises. This is most likely true. It wouldn't be if the "last IPI
was sent" flag clears before the IPI actually hit the other core.
4.9: Future Broadcast/Messaging Needs
---------------------------
Currently, messaging is serialized. Broadcast IPIs exist, but the kernel
message system is based on adding an k_msg to a list in a pcore's
per_cpu_info. Further the sending of these messages is in a loop. In the
future, we would like to have broadcast messaging of some sort (literally a
broadcast, like the IPIs, and if not that, then a communication tree of
sorts).
In the past, (OLD INFO): given those desires, we wanted to make sure that no
message we send needs details specific to a pcore (such as the vcoreid running
on it, a history number, or anything like that). Thus no k_msg related to
process management would have anything that cannot apply to the entire
process. At this point, most just have a struct proc *. A pcore was be able
to figure out what is happening based on the pcoremap, information in the
struct proc, and in the preempt struct in procdata.
In more recent revisions, the coremap no longer is meant to be used across
kmsgs, so some messages ('__startcore') send the vcoreid. This means we can't
easily broadcast the message. However, many broadcast mechanisms wouldn't
handle '__startcore' naturally. For instance, logical IPIs need something
already set in the LAPIC, or maybe need to be sent to a somewhat predetermined
group (again, bits in the LAPIC). If we tried this for '__startcore', we
could add something in to the messaging to carry these vcoreids. More likely,
we'll have a broadcast tree. Keeping vcoreid (or any arg) next to whoever
needs to receive the message is a very small amount of bookkeeping on a struct
that already does bookkeeping.
4.10: Other Things We Thought of but Don't Like
---------------------------
All local fate-related work is sent as a self k_msg, to enforce ordering.
It doesn't capture the difference between a local call and a remote k_msg.
The k_msg has already considered state and made its decision. The local call
is an attempt. It is also unnecessary, if we put in enough information to
make a decision in the proc struct. Finally, it caused a few other problems
(like needing to detect arbitrary stale messages).
Overall message history: doesn't work well when you do per-core stuff, since
it will invalidate other messages for the process. We then though of a pcore
history counter to detect stale messages. Don't like that either. We'd have
to send the history in the message, since it's a per-message, per-core
expiration. There might be other ways around this, but this doesn't seem
necessary.
Alarms have pointers to a list of which cores should be preempted when that
specific alarm goes off (saved with the alarm). Ugh. It gets ugly with
multiple outstanding preemptions and cores getting yielded while the alarms
sleep (and possibly could get reallocated, though we'd make a rule to prevent
that). Like with notifications, being able to handle spurious alarms and
thinking of an alarm as just a prod to check somewhere is much more flexible
and simple. It is similar to generic messages that have the actual important
information stored somewhere else (as with allowing broadcasts, with different
receivers performing slightly different operations).
Synchrony for messages (wanting a response to a preempt k_msg, for example)
sucks. Just encode the state of impending fate in the proc struct, where it
belongs. Additionally, we don't want to hold the proc lock even longer than
we do now (which is probably too long as it is). Finally, it breaks a golden
rule: never wait while holding a lock: you will deadlock the system (e.g. if
the receiver is already in the kernel spinning on the lock). We'd have to
send messages, unlock (which might cause a message to hit the calling pcore,
as in the case of locally called proc_destroy()), and in the meantime some
useful invariant might be broken.
We also considered using the transition stack as a signal that a process is in
a notification handler. The kernel can inspect the stack pointer to determine
this. It's possible, but unnecessary.
Using the pcoremap as a way to pass info with kmsgs: it worked a little, but
had some serious problems, as well as making life difficult. It had two
purposes: help with local fate calls (yield) and allow broadcast messaging.
The main issue is that it was using a global struct to pass info with
messages, but it was based on the snapshot of state at the time the message
was sent. When you send a bunch of messages, certain state may have changed
between messages, and the old snapshot isn't there anymore by the time the
message gets there. To avoid this, we went through some hoops and had some
fragile code that would use other signals to avoid those scenarios where the
global state change would send the wrong message. It was tough to understand,
and not clear it was correct (hint: it wasn't). Here's an example (on one
pcore): if we send a preempt and we then try to map that pcore to another
vcore in the same process before the preempt call checks its pcoremap, we'll
clobber the pcore->vcore mapping (used by startcore) and the preempt will
think it is the new vcore, not the one it was when the message was sent.
While this is a bit convoluted, I can imagine a ksched doing this, and
perhaps with weird IRQ delays, the messages might get delayed enough for it to
happen. I'd rather not have to have the ksched worry about this just because
proc code was old and ghetto. Another reason we changed all of this was so
that you could trust the vcoremap while holding the lock. Otherwise, it's
actually non-trivial to know the state of a vcore (need to check a combination
of preempt_served and is_mapped), and even if you do that, there are some
complications with doing this in the ksched.
5. current_ctx and owning_proc
===========================
Originally, current_tf was a per-core macro that returns a struct trapframe *
that points back on the kernel stack to the user context that was running on
the given core when an interrupt or trap happened. Saving the reference to
the TF helps simplify code that needs to do something with the TF (like save
it and pop another TF). This way, we don't need to pass the context all over
the place, especially through code that might not care.
Then, current_tf was more broadly defined as the user context that should be
run when the kernel is ready to run a process. In the older case, it was when
the kernel tries to return to userspace from a trap/interrupt. current_tf
could be set by an IPI/KMSG (like '__startcore') so that when the kernel wants
to idle, it will find a current_tf that it needs to run, even though we never
trapped in on that context in the first place.
Finally, current_tf was changed to current_ctx, and instead of tracking a
struct trapframe (equivalent to a hw_trapframe), it now tracked a struct
user_context, which could be either a HW or a SW trapframe.
Further, we now have 'owning_proc', which tells the kernel which process
should be run. 'owning_proc' is a bigger deal than 'current_ctx', and it is
what tells us to run cur_ctx.
Process management KMSGs now simply modify 'owning_proc' and cur_ctx, as if we
had interrupted a process. Instead of '__startcore' forcing the kernel to
actually run the process and trapframe, it will just mean we will eventually
run it. In the meantime a '__notify' or a '__preempt' can come in, and they
will apply to the owning_proc/cur_ctx. This greatly simplifies process code
and code calling process code (like the scheduler), since we no longer need to
worry about whether or not we are getting a "stack killing" kernel message.
Before this, code needed to care where it was running when managing _Ms.
Note that neither 'current_ctx' nor 'owning_proc' rely on 'current'/'cur_proc'.
'current' is just what process context we're in, not what process (and which
trapframe) we will eventually run.
cur_ctx does not point to kernel trapframes, which is important when we
receive an interrupt in the kernel. At one point, we were (hypothetically)
clobbering the reference to the user trapframe, and were unable to recover.
We can get away with this because the kernel always returns to its previous
context from a nested handler (via iret on x86).
In the future, we may need to save kernel contexts and may not always return
via iret. At which point, if the code path is deep enough that we don't want
to carry the TF pointer, we may revisit this. Until then, current_ctx is just
for userspace contexts, and is simply stored in per_cpu_info.
Brief note from the future (months after this paragraph was written): cur_ctx
has two aspects/jobs:
1) tell the kernel what we should do (trap, fault, sysc, etc), how we came
into the kernel (the fact that it is a user tf), which is why we copy-out
early on
2) be a vehicle for us to restart the process/vcore
We've been focusing on the latter case a lot, since that is what gets
removed when preempted, changed during a notify, created during a startcore,
etc. Don't forget it was also an instruction of sorts. The former case is
always true throughout the life of the syscall. The latter only happens to be
true throughout the life of a *non-blocking* trap since preempts are routine
KMSGs. But if we block in a syscall, the cur_ctx is no longer the TF we came
in on (and possibly the one we are asked to operate on), and that old cur_ctx
has probably restarted.
(Note that cur_ctx is a pointer, and syscalls/traps actually operate on the TF
they came in on regardless of what happens to cur_ctx or pcpui->actual_tf.)
6. Locking!
===========================
6.1: proc_lock
---------------------------
Currently, all locking is done on the proc_lock. It's main goal is to protect
the vcore mapping (vcore->pcore and vice versa). As of Apr 2010, it's also used
to protect changes to the address space and the refcnt. Eventually the refcnt
will be handled with atomics, and the address space will have it's own MM lock.
We grab the proc_lock all over the place, but we try to avoid it whereever
possible - especially in kernel messages or other places that will be executed
in parallel. One place we do grab it but would like to not is in proc_yield().
We don't always need to grab the proc lock. Here are some examples:
6.1.1: Lockless Notifications:
-------------
We don't lock when sending a notification. We want the proc_lock to not be an
irqsave lock (discussed below). Since we might want to send a notification from
interrupt context, we can't grab the proc_lock if it's a regular lock.
This is okay, since the proc_lock is only protecting the vcoremapping. We could
accidentally send the notification to the wrong pcore. The __notif handler
checks to make sure it is the right process, and all _M processes should be able
to handle spurious notifications. This assumes they are still _M.
If we send it to the wrong pcore, there is a danger of losing the notif, since
it didn't go to the correct vcore. That would happen anyway, (the vcore is
unmapped, or in the process of mapping). The notif_pending flag will be caught
when the vcore is started up next time (and that flag was set before reading the
vcoremap).
6.1.2: Local get_vcoreid():
-------------
It's not necessary to lock while checking the vcoremap if you are checking for
the core you are running on (e.g. pcoreid == core_id()). This is because all
unmappings of a vcore are done on the receive side of a routine kmsg, and that
code cannot run concurrently with the code you are running.
6.2: irqsave
---------------------------
The proc_lock used to be an irqsave lock (meaning it disables interrupts and can
be grabbed from interrupt context). We made it a regular lock for a couple
reasons. The immediate one was it was causing deadlocks due to some other
ghetto things (blocking on the frontend server, for instance). More generally,
we don't want to disable interrupts for long periods of time, so it was
something worth doing anyway.
This means that we cannot grab the proc_lock from interrupt context. This
includes having schedule called from an interrupt handler (like the
timer_interrupt() handler), since it will call proc_run. Right now, we actually
do this, which we shouldn't, and that will eventually get fixed. The right
answer is that the actual work of running the scheduler should be a routine
kmsg, similar to how Linux sets a bit in the kernel that it checks on the way
out to see if it should run the scheduler or not.
7. TLB Coherency
===========================
When changing or removing memory mappings, we need to do some form of a TLB
shootdown. Normally, this will require sending an IPI (immediate kmsg) to
every vcore of a process to unmap the affected page. Before allocating that
page back out, we need to make sure that every TLB has been flushed.
One reason to use a kmsg over a simple handler is that we often want to pass a
virtual address to flush for those architectures (like x86) that can
invalidate a specific page. Ideally, we'd use a broadcast kmsg (doesn't exist
yet), though we already have simple broadcast IPIs.
7.1 Initial Stuff
---------------------------
One big issue is whether or not to wait for a response from the other vcores
that they have unmapped. There are two concerns: 1) Page reuse and 2) User
semantics. We cannot give out the physical page while it may still be in a
TLB (even to the same process. Ask us about the pthread_test bug).
The second case is a little more detailed. The application may not like it if
it thinks a page is unmapped or protected, and it does not generate a fault.
I am less concerned about this, especially since we know that even if we don't
wait to hear from every vcore, we know that the message was delivered and the
IPI sent. Any cores that are in userspace will have trapped and eventually
handle the shootdown before having a chance to execute other user code. The
delays in the shootdown response are due to being in the kernel with
interrupts disabled (it was an IMMEDIATE kmsg).
7.2 RCU
---------------------------
One approach is similar to RCU. Unmap the page, but don't put it on the free
list. Instead, don't reallocate it until we are sure every core (possibly
just affected cores) had a chance to run its kmsg handlers. This time is
similar to the RCU grace periods. Once the period is over, we can then truly
free the page.
This would require some sort of RCU-like mechanism and probably a per-core
variable that has the timestamp of the last quiescent period. Code caring
about when this page (or pages) can be freed would have to check on all of the
cores (probably in a bitmask for what needs to be freed). It would make sense
to amortize this over several RCU-like operations.
7.3 Checklist
---------------------------
It might not suck that much to wait for a response if you already sent an IPI,
though it incurs some more cache misses. If you wanted to ensure all vcores
ran the shootdown handler, you'd have them all toggle their bit in a checklist
(unused for a while, check smp.c). The only one who waits would be the
caller, but there still are a bunch of cache misses in the handlers. Maybe
this isn't that big of a deal, and the RCU thing is an unnecessary
optimization.
7.4 Just Wait til a Context Switch
---------------------------
Another option is to not bother freeing the page until the entire process is
descheduled. This could be a very long time, and also will mess with
userspace's semantics. They would be running user code that could still
access the old page, so in essence this is a lazy munmap/mprotect. The
process basically has the page in pergatory: it can't be reallocated, and it
might be accessible, but can't be guaranteed to work.
The main benefit of this is that you don't need to send the TLB shootdown IPI
at all - so you don't interfere with the app. Though in return, they have
possibly weird semantics. One aspect of these weird semantics is that the
same virtual address could map to two different pages - that seems like a
disaster waiting to happen. We could also block that range of the virtual
address space from being reallocated, but that gets even more tricky.
One issue with just waiting and RCU is memory pressure. If we actually need
the page, we will need to enforce an unmapping, which sucks a little.
7.5 Bulk vs Single
---------------------------
If there are a lot of pages being shot down, it'd be best to amortize the cost
of the kernel messages, as well as the invlpg calls (single page shootdowns).
One option would be for the kmsg to take a range, and not just a single
address. This would help with bulk munmap/mprotects. Based on the number of
these, perhaps a raw tlbflush (the entire TLB) would be worth while, instead
of n single shots. Odds are, that number is arch and possibly workload
specific.
For now, the plan will be to send a range and have them individually shot
down.
7.6 Don't do it
---------------------------
Either way, munmap/mprotect sucks in an MCP. I recommend not doing it, and
doing the appropriate mmap/munmap/mprotects in _S mode. Unfortunately, even
our crap pthread library munmaps on demand as threads are created and
destroyed. The vcore code probably does in the bowels of glibc's TLS code
too, though at least that isn't on every user context switch.
7.7 Local memory
---------------------------
Private local memory would help with this too. If each vcore has its own
range, we won't need to send TLB shootdowns for those areas, and we won't have
to worry about weird application semantics. The downside is we would need to
do these mmaps in certain ranges in advance, and might not easily be able to
do them remotely. More on this when we actually design and build it.
7.8 Future Hardware Support
---------------------------
It would be cool and interesting if we had the ability to remotely shootdown
TLBs. For instance, all cores with cr3 == X, shootdown range Y..Z. It's
basically what we'll do with the kernel message and the vcoremap, but with
magic hardware.
7.9 Current Status
---------------------------
For now, we just send a kernel message to all vcores to do a full TLB flush,
and not to worry about checklists, waiting, or anything. This is due to being
short on time and not wanting to sort out the issue with ranges. The way
it'll get changed to is to send the kmsg with the range to the appropriate
cores, and then maybe put the page on the end of the freelist (instead of the
head). More to come.
8. Process Management
===========================
8.1 Vcore lists
---------------------------
We have three lists to track a process's vcores. The vcores themselves sit in
the vcoremap in procinfo; they aren't dynamically allocated (memory) or
anything like that. The lists greatly eases vcore discovery and management.
A vcore is on exactly one of three lists: online (mapped and running vcores,
sometimes called 'active'), bulk_preempt (was online when the process was bulk
preempted (like a timeslice)), and inactive (yielded, hasn't come on yet,
etc). When writes are complete (unlocked), either the online list or the
bulk_preempt list should be empty.
List modifications are protected by the proc_lock. You can concurrently read,
but note you may get some weird behavior, such as a vcore on multiple lists, a
vcore on no lists, online and bulk_preempt both having items, etc. Currently,
event code will read these lists when hunting for a suitable core, and will
have to be careful about races. I want to avoid event FALLBACK code from
grabbing the proc_lock.
Another slight thing to be careful of is that the vcore lists don't always
agree with the vcore mapping. However, it will always agree with what the
state of the process will be when all kmsgs are processed (fate).
Specifically, when we take vcores, the unmapping happens with the lock not
held on the vcore itself (as discussed elsewhere). The vcore lists represent
the result of those pending unmaps.
Before we used the lists, we scanned the vcoremap in a painful, clunky manner.
In the old style, when you asked for a vcore, the first one you got was the
first hole in the vcoremap. Ex: Vcore0 would always be granted if it was
offline. That's no longer true; the most recent vcore yielded will be given
out next. This will help with cache locality, and also cuts down on the
scenarios on which the kernel gives out a vcore that userspace wasn't
expecting. This can still happen if they ask for more vcores than they set up
for, or if a vcore doesn't *want* to come online (there's a couple scenarios
with preemption recovery where that may come up).
So the plan with the bulk preempt list is that vcores on it were preempted,
and the kernel will attempt to restart all of them (and move them to the online
list). Any leftovers will be moved to the inactive list, and have preemption
recovery messages sent out. Any shortages (they want more vcores than were
bulk_preempted) will be taken from the yield list. This all means that
whether or not a vcore needs to be preempt-recovered or if there is a message
out about its preemption doesn't really affect which list it is on. You could
have a vcore on the inactive list that was bulk preempted (and not turned back
on), and then that vcore gets granted in the next round of vcore_requests().
The preemption recovery handlers will need to deal with concurrent handlers
and the vcore itself starting back up.
9. On the Ordering of Messages and Bugs with Old State
===========================
This is a sordid tale involving message ordering, message delivery times, and
finding out (sometimes too late) that the state you expected is gone and
having to deal with that error.
A few design issues:
- being able to send messages and have them execute in the order they are
sent
- having message handlers resolve issues with global state. Some need to know
the correct 'world view', and others need to know what was the state at the
time they were sent.
- realizing syscalls, traps, faults, and any non-IRQ entry into the kernel is
really a message.
Process management messages have alternated from ROUTINE to IMMEDIATE and now
back to ROUTINE. These messages include such family favorites as
'__startcore', '__preempt', etc. Meanwhile, syscalls were coming in that
needed to know about the core and the process's state (specifically, yield,
change_to, and get_vcoreid). Finally, we wanted to avoid locking, esp in
KMSGs handlers (imagine all cores grabbing the lock to check the vcoremap or
something).
Incidentally, events were being delivered concurretly to vcores, though that
actually didn't matter much (check out async_events.txt for more on that).
9.1: Design Guidelines
---------------------------
Initially, we wanted to keep broadcast messaging available as an option. As
noted elsewhere, we can't really do this well for startcore, since most
hardware broadcast options need some initial per-core setup, and any sort of
broadcast tree we make should be able to handle a small message. Anyway, this
desire in the early code to keep all messages identical lead to a few
problems.
Another objective of the kernel messaging was to avoid having the message
handlers grab any locks, especially the same lock (the proc lock is used to
protect the vcore map, for instance).
Later on, a few needs popped up that motivated the changes discussed below:
- Being able to find out which proc/vcore was on a pcore
- Not having syscalls/traps require crazy logic if the carpet was pulled out
from under them.
- Having proc management calls return. This one was sorted out by making all
kmsg handlers return. It would be a nightmare making a ksched without this.
9.2: Looking at Old State: a New Bug for an Old Problem
---------------------------
We've always had issues with syscalls coming in and already had the fate of a
core determined. This is referred to in a few places as "predetermined fate"
vs "local state". A remote lock holder (ksched) already determined a core
should be unmapped and sent a message. Only later does some call like
proc_yield() realize its core is already *unmapped*. (I use that term poorly
here). This sort of code had to realize it was working on an old version of
state and just abort. This was usually safe, though looking at the vcoremap
was a bad idea. Initially, we used preempt_served as the signal, which was
okay. Around 12b06586 yield started to use the vcoremap, which turned out to
be wrong.
A similar issue happens for the vcore messages (startcore, preempt, etc). The
way startcore used to work was that it would only know what pcore it was on,
and then look into the vcoremap to figure out what vcoreid it should be
running. This was to keep broadcast messaging available as an option. The
problem with it is that the vcoremap may have changed between when the
messages were sent and when they were executed. Imagine a startcore followed
by a preempt, afterwhich the vcore was unmapped. Well, to get around that, we
had the unmapping happen in the preempt or death handlers. Yikes! This was
the case back in the early days of ROS. This meant the vcoremap wasn't
actually representative of the decisions the ksched made - we also needed to
look at the state we'd have after all outstanding messages executed. And this
would differ from the vcore lists (which were correct for a lock holder).
This was managable for a little while, until I tried to conclusively know who
owned a particular pcore. This came up while making a provisioning scheduler.
Given a pcore, tell me which process/vcore (if any) were on it. It was rather
tough. Getting the proc wasn't too hard, but knowing which vcore was a little
tougher. (Note the ksched doesn't care about which vcore is running, and the
process can change vcores on a pcore at will). But once you start looking at
the process, you can't tell which vcore a certain pcore has. The vcoremap may
be wrong, since a preempt is already on the way. You would have had to scan
the vcore lists to see if the proc code thought that vcore was online or not
(which would mean there had been no preempts). This is the pain I was talking
about back around commit 5343a74e0.
So I changed things so that the vcoremap was always correct for lock holders,
and used pcpui to track owning_vcoreid (for preempt/notify), and used an extra
KMSG variable to tell startcore which vcoreid it should use. In doing so, we
(re)created the issue that the delayed unmapping dealt with: the vcoremap
would represent *now*, and not the vcoremap of when the messages were first
sent. However, this had little to do with the KMSGs, which I was originally
worried about. No one was looking at the vcoremap without the lock, so the
KMSGs were okay, but remember: syscalls are like messages too. They needed to
figure out what vcore they were on, i.e. what vcore userspace was making
requests on (viewing a trap/fault as a type of request).
Now the problem was that we were using the vcoremap to figure out which vcore
we were supposed to be. When a syscall finally ran, the vcoremap could be
completely wrong, and with immediate KMSGs (discussed below), the pcpui was
already changed! We dealt with the problem for KMSGs, but not syscalls, and
basically reintroduced the bug of looking at current state and thinking it
represented the state from when the 'message' was sent (when we trapped into
the kernel, for a syscall/exception).
9.3: Message Delivery, Circular Waiting, and Having the Carpet Pulled Out
---------------------------
In-order message delivery was what drove me to build the kernel messaging
system in the first place. It provides in-order messages to a particular
pcore. This was enough for a few scenarios, such as preempts racing ahead of
startcores, or deaths racing a head of preempts, etc. However, I also wanted
an ordering of messages related to a particular vcore, and this wasn't
apparent early on.
The issue first popped up with a startcore coming quickly on the heals of a
preempt for the same VC, but on different PCs. The startcore cannot proceed
until the preempt saved the TF into the VCPD. The old way of dealing with
this was to spin in '__map_vcore()'. This was problematic, since it meant we
were spinning while holding a lock, and resulted in some minor bugs and issues
with lock ordering and IRQ disabling (couldn't disable IRQs and then try to
grab the lock, since the lock holder could have sent you a message and is
waiting for you to handle the IRQ/IMMED KMSG). However, it was doable. But
what wasn't doable was to have the KMSGs be ROUTINE. Any syscalls that tried
to grab the proc lock (lots of them) would deadlock, since the lock holder was
waiting on us to handle the preempt (same circular waiting issue as above).
This was fine, albeit subpar, until a new issue showed up. Sending IMMED
KMSGs worked fine if we were coming from userspace already, but if we were in
the kernel, those messages would run immediately (hence the name), just like
an IRQ handler, and could confuse syscalls that touched cur_ctx/pcpui. If a
preempt came in during a syscall, the process/vcore could be changed before
the syscall took place. Some syscalls could handle this, albeit poorly.
sys_proc_yield() and sys_change_vcore() delicately tried to detect if they
were still mapped or not and use that to determine if a preemption happened.
As mentioned above, looking at the vcoremap only tells you what is currently
happening, and not what happened in the past. Specifically, it doesn't tell
you the state of the mapping when a particular core trapped into the kernel
for a syscall (referred to as when the 'message' was sent up above). Imagine
sys_get_vcoreid(): you trap in, then immediately get preempted, then startcore
for the same process but a different vcoreid. The syscall would return with
the vcoreid of the new vcore, since it cannot tell there was a change. The
async syscall would complete and we'd have a wrong answer. While this never
happened to me, I had a similar issue while debugging some other bugs (I'd get
a vcoreid of 0xdeadbeef, for instance, which was the old poison value for an
unmapped vcoreid). There are a bunch of other scenarios that trigger similar
disasters, and they are very hard to avoid.
One way out of this was a per-core history counter, that changed whenever we
changed cur_ctx. Then when we trapped in for a syscall, we could save the
value, enable_irqs(), and go about our business. Later on, we'd have to
disable_irqs() and compare the counters. If they were different, we'd have to
bail out some how. This could have worked for change_to and yield, and some
others. But any syscall that wanted to operate on cur_ctx in some way would
fail (imagine a hypothetical sys_change_stack_pointer()). The context that
trapped has already returned on another core. I guess we could just fail that
syscall, though it seems a little silly to not be able to do that.
The previous example was a bit contrived, but lets also remember that it isn't
just syscalls: all exceptions have the same issue. Faults might be fixable,
since if you restart a faulting context, it will start on the faulting
instruction. However all traps (like syscall) restart on the next
instruction. Hope we don't want to do anything fancy with breakpoint! Note
that I had breakpointing contexts restart on other pcores and continue while I
was in the breakpoint handler (noticed while I was debugging some bugs with
lots of preempts). Yikes. And don't forget we eventually want to do some
complicated things with the page fault handler, and may want to turn on
interrupts / kthread during a page fault (imaging hitting disk). Yikes.
So I looked into going back to ROUTINE kernel messages. With ROUTINE
messages, I didn't have to worry about having the carpet pulled out from under
syscalls and exceptions (traps, faults, etc). The 'carpet' is stuff like
cur_ctx, owning_proc, owning_vcoreid, etc. We still cannot trust the vcoremap,
unless we *know* there were no preempts or other KMSGs waiting for us.
(Incidentally, in the recent fix a93aa7559, we merely use the vcoremap as a
sanity check).
However, we can't just switch back to ROUTINEs. Remember: with ROUTINEs,
we will deadlock in '__map_vcore()', when it waits for the completion of
preempt. Ideally, we would have had startcore spin on the signal. Since we
already gave up on using x86-style broadcast IPIs for startcore (in
5343a74e0), we might as well pass along a history counter, so it knows to wait
on preempt.
9.4: The Solution
---------------------------
To fix up all of this, we now detect preemptions in syscalls/traps and order
our kernel messages with two simple per-vcore counters. Whenever we send a
preempt, we up one counter. Whenever that preempt finishes, it ups another
counter. When we send out startcores, we send a copy of the first counter.
This is a way of telling startcore where it belongs in the list of messages.
More specifically, it tells it which preempt happens-before it.
Basically, I wanted a partial ordering on my messages, so that messages sent
to a particular vcore are handled in the order they were sent, even if those
messages run on different physical cores.
It is not sufficient to use a seq counter (one integer, odd values for
'preempt in progress' and even values for 'preempt done'). It is possible to
have multiple preempts in flight for the same vcore, albeit with startcores in
between. Still, there's no way to encode that scenario in just one counter.
Here's a normal example of traffic to some vcore. I note both the sending and
the execution of the kmsgs:
nr_pre_sent nr_pre_done pcore message sent/status
-------------------------------------------------------------
0 0 X startcore (nr_pre_sent == 0)
0 0 X startcore (executes)
1 0 X preempt (kmsg sent)
1 1 Y preempt (executes)
1 1 Y startcore (nr_pre_sent == 1)
1 1 Y startcore (executes)
Note the messages are always sent by the lockholder in the order of the
example above.
Here's when the startcore gets ahead of the prior preempt:
nr_pre_sent nr_pre_done pcore message sent/status
-------------------------------------------------------------
0 0 X startcore (nr_pre_sent == 0)
0 0 X startcore (executes)
1 0 X preempt (kmsg sent)
1 0 Y startcore (nr_pre_sent == 1)
1 1 X preempt (executes)
1 1 Y startcore (executes)
Note that this can only happen across cores, since KMSGs to a particular core
are handled in order (for a given class of message). The startcore blocks on
the prior preempt.
Finally, here's an example of what a seq ctr can't handle:
nr_pre_sent nr_pre_done pcore message sent/status
-------------------------------------------------------------
0 0 X startcore (nr_pre_sent == 0)
1 0 X preempt (kmsg sent)
1 0 Y startcore (nr_pre_sent == 1)
2 0 Y preempt (kmsg sent)
2 0 Z startcore (nr_pre_sent == 2)
2 1 X preempt (executes (upped to 1))
2 1 Y startcore (executes (needed 1))
2 2 Y preempt (executes (upped to 2))
2 Z Z startcore (executes (needed 2))
As a nice bonus, it is easy for syscalls that care about the vcoreid (yield,
change_to, get_vcoreid) to check if they have a preempt_served. Just grab the
lock (to prevent further messages being sent), then check the counters. If
they are equal, there is no preempt on its way. This actually was the
original way we checked for preempts in proc_yield back in the day. It was
just called preempt_served. Now, it is split into two counters, instead of
just being a bool.
Regardless of whether or not we were preempted, we still can look at
pcpui->owning_proc and owning_vcoreid to figure out what the vcoreid of the
trap/syscall is, and we know that the cur_ctx is still the correct cur_ctx (no
carpet pulled out), since while there could be a preempt ROUTINE message
waiting for us, we simply haven't run it yet. So calls like yield should
still fail (since your core has been unmapped and you need to bail out and run
the preempt handler), but calls like sys_change_stack_pointer can proceed.
More importantly than that old joke syscall, the page fault handler can try to
do some cool things without worrying about really crazy stuff.
9.5: Why We (probably) Don't Deadlock
---------------------------
It's worth thinking about why this setup of preempts and startcores can't
deadlock. Anytime we spin in the kernel, we ought to do this. Perhaps there
is some issue with other KMSGs for other processes, or other vcores, or
something like that that can cause a deadlock.
Hypothetical case: pcore 1 has a startcore for vc1 which is stuck behind vc2's
startcore on PC2, with time going upwards. In these examples, startcores are
waiting on particular preempts, subject to the nr_preempts_sent parameter sent
along with the startcores.
^
| _________ _________
| | | | |
| | pr vc 2 | | pr vc 1 |
| |_________| |_________|
|
| _________ _________
| | | | |
| | sc vc 1 | | sc vc 2 |
| |_________| |_________|
t
---------------------------------------------------------------------------
______ ______
| | | |
| PC 1 | | PC 2 |
|______| |______|
Here's the same picture, but with certain happens-before arrows. We'll use X --> Y to
mean X happened before Y, was sent before Y. e.g., a startcore is sent after
a preempt.
^
| _________ _________
| | | | |
| .-> | pr vc 2 | --. .----- | pr vc 1 | <-.
| | |_________| \ / & |_________| |
| * | \/ | *
| | _________ /\ _________ |
| | | | / \ & | | |
| '-- | sc vc 1 | <-' '----> | sc vc 2 | --'
| |_________| |_________|
t
---------------------------------------------------------------------------
______ ______
| | | |
| PC 1 | | PC 2 |
|______| |______|
The arrows marked with * are ordered like that due to the property of KMSGs,
in that we have in order delivery. Messages are executed in the order in
which they were sent (serialized with a spinlock btw), so on any pcore,
messages that are further ahead in the queue were sent before (and thus will
be run before) other messages.
The arrows marked with a & are ordered like that due to how the proc
management code works. The kernel won't send out a startcore for a particular
vcore before it sent out a preempt. (Note that techincally, preempts follow
startcores. The startcores in this example are when we start up a vcore after
it had been preempted in the past.).
Anyway, note that we have a cycle, where all events happened before each
other, which isn't possible. The trick to connecting "unrelated" events like
this (unrelated meaning 'not about the same vcore') in a happens-before manner
is the in-order properties of the KMSGs.
Based on this example, we can derive general rules. Note that 'sc vc 2' could
be any kmsg that waits on another message placed behind 'sc vc 1'. This would
require us having sent a KMSG that waits on a KMSGs that we send later. Bad
idea! (you could have sent that KMSGs to yourself, aside from just being
dangerous). If you want to spin, make sure you send the work that should
happen-before actually-before the waiter.
In fact, we don't even need 'sc vc 2' to be a KMSG. It could be miscellaneous
kernel code, like a proc mgmt syscall. Imagine if we did something like the
old '__map_vcore' call from within the ksched. That would be code that holds
the lock, and then waits on the execution of a message handler. That would
deadlock (which is why we don't do it anymore).
Finally, in case this isn't clear, all of the startcores and preempts for
a given vcore exist in a happens-before relation, both in sending and in
execution. The sending aspect is handled by proc mgmt code. For execution,
preempts always follow startcores due to the KMSG ordering property. For
execution of startcores, startcores always spin until the preempt they follow
is complete, ensuring the execution of the main part of their handler happens
after the prior preempt.
Here's some good ideas for the ordering of locks/irqs/messages:
- You can't hold a spinlock of any sort and then wait on a routine kernel
message. The core where that runs may be waiting on you, or some scenario
like above.
- Similarly, think about how this works with kthreads. A kthread
restart is a routine KMSG. You shouldn't be waiting on code that
could end up kthreading, mostly because those calls block!
- You can hold a spinlock and wait on an IMMED kmsg, if the waiters of the
spinlock have irqs enabled while spinning (this is what we used to do with
the proc lock and IMMED kmsgs, and 54c6008 is an example of doing it wrong)
- As a corollary, locks like this cannot be irqsave, since the other
attempted locker will have irq disabled
- For broadcast trees, you'd have to send IMMEDs for the intermediates, and
then it'd be okay to wait on those intermediate, immediate messages (if we
wanted confirmation of the posting of RKM)
- The main thing any broadcast mechanism needs to do is make sure all
messages get delivered in order to particular pcores (the central
premise of KMSGs) (and not deadlock due to waiting on a KMSG
improperly)
- Alternatively, we could use routines for the intermediates if we didn't want
to wait for RKMs to hit their destination, we'd need to always use the same
proxy for the same destination pcore, e.g., core 16 always covers 16-31.
- Otherwise, we couldn't guarantee the ordering of SC before PR before
another SC (which the proc_lock and proc mgmt code does); we need the
ordering of intermediate msgs on the message queues of a particular
core.
- All kmsgs would need to use this broadcasting style (couldn't mix
regular direct messages with broadcast), so odds are this style would
be of limited use.
- since we're not waiting on execution of a message, we could use RKMs
(while holding a spinlock)
- There might be some bad effects with kthreads delaying the reception of RKMS
for a while, but probably not catastrophically.
9.6: Things That We Don't Handle Nicely
---------------------------
If for some reason a syscall or fault handler blocks *unexpectedly*, we could
have issues. Imagine if change_to happens to block in some early syscall code
(like instrumentation, or who knows what, that blocks in memory allocation).
When the syscall kthread restarts, its old cur_ctx is gone. It may or may not
be running on a core owned by the original process. If it was, we probably
would accidentally yield that vcore (clearly a bug).
For now, any of these calls that care about cur_ctx/pcpui need to not block
without some sort of protection. None of them do, but in the future we might
do something that causes them to block. We could deal with it by having a
pcpu or per-kthread/syscall flag that says if it ever blocked, and possibly
abort. We get into similar nasty areas as with preempts, but this time, we
can't solve it by making preempt a routine KMSG - we block as part of that
syscall/handler code. Odds are, we'll just have to outlaw this, now and
forever. Just note that if a syscall/handler blocks, the TF it came in on is
probably not cur_ctx any longer, and that old cur_ctx has probably restarted.
10. TBD
===========================