When the malloc code is instrumented, the global _brksize variable
should not be transferred. However, when the malloc code is not
instrumented, failing to transfer _brksize would reset the heap
upon state transfer. In this patch, the magic pass stores the flag
indicating whether memory function instrumentation is disabled, in
the target process. This allows libmagic to check this flag during
state transfer, to see whether it should transfer _brksize or not.
This patch changes the VM makefile to specify that the magic pass is
to skip memory function instrumentation, and to transfer the data
variables of the malloc code (thus overriding the exception we made
for all other system services). We add two magic pass flags to
achieve this. Since the magic pass is a big bowl of spaghetti code,
ignoring whitespace changes while viewing this patch is recommended.
The NetBSD libc malloc implementation uses a memory-mapped area for
its page directory. Since the process heap is reconstructed upon
state transfer for live update, this memory-mapped area must not be
transferred to the new process. However, as the new instance of the
process being updated inherits all memory-mapped areas of the old
instance, it also automatically inherits the malloc implementation's
page directory. Thus, we must explicitly free this area in order to
avoid a memory leak.
The magic pass already detects (de)allocation functions called from
within other (de)allocation functions, which is why the mmap(2) and
munmap(2) calls of the malloc code are not instrumented as it is.
This patch changes that particular case to allow a different hook
function to be called for such "nested" allocation calls, for a
particular set of nested calls. In particular, the malloc(3) code's
mmap(2) and munmap(2) calls are replaced with magic_nested_mmap and
magic_nested_munmap calls, respectively. The magic library then
tracks memory mapping allocations of the malloc code by providing an
implementation for these two wrappers, and frees the allocations upon
state transfer.
This approach was chosen over various alternatives:
- While it appears that nesting could be established by setting a
flag while the malloc(3) wrapper is active, and testing the flag in
the mmap(2)/munmap(2) wrappers, this approach would fail to detect
memory-mapped allocations made from uninstrumented malloc(3) calls,
and therefore not a viable option.
- It would be possible to obtain the value of the variables that
store the information about the memory-mapped area in the malloc
code. However, this is rather difficult in practice due to the way
the libc malloc implementation stores the size of the are, and it
would make the solution more dependent on the specific libc malloc
implementation.
- It would be possible to use the special "nested" instrumentation
for allocations made from certain marked sections. Since we mark
the data section of the malloc code already, this would not be hard
to do. Switching to this alternative would change very little, and
if for any reason this approach yields more advantages in the
future, we can still choose to do so.
Since the heap is reconstructed upon state transfer, the old malloc
state is discarded. In order to avoid state transfer errors, we can
and in fact must discard the internal state of the malloc
implementation. This patch achieves this by using the sectionify
pass to mark the variables in the libminc malloc object as state that
must be skipped during state transfer.
RS/VM: proper preparation for multi-VM live update
Due to changed VM internals, more elaborate preparation is required
before a live update with multiple components including VM can take
place. This patch adds the essential preparation infrastructure to
VM and adapts RS to make use of it. As a side effect, it is no
longer necessary to supply RS as the last component (if at all)
during the set-up of a multicomponent live update operation.
During live update, the new instance of VM may make changes that,
after a rollback, have to be undone by the old instance of VM, in
particular because both instances share (read-write) all dynamically
allocated pages.
Make the passes we have so far, hello and WeakAliasModuleOverride,
use settings from a Makefile include file in the parent directory.
This change is in preparation of adding other passes.
Lionel Sambuc [Fri, 23 Jan 2015 17:30:39 +0000 (18:30 +0100)]
QEMU default command lines updates
- Fix a bug in clientctl which tried to test for kvm. This simply
remove this faulty test as the kvm command has been deprecated by the
QEMU project for a couple of years now.
- Specify by default 256M of RAM as this is the minimal amount required
for the whole-OS live update test to succeed.
- Update the default command printed out at the end of the x86_hdimage
script to be more generic, less focused on one use-case.
The filtering also exposed the risk that a process be killed or
swapped while on the list of VM memory requests. These cases are
now handled properly as well.
Lionel Sambuc [Mon, 26 Jan 2015 14:09:05 +0000 (15:09 +0100)]
VM: set recovery policy to restart
- Update proc to select restart policy for VM
- Update testrelpol to test the supported modes of recovery for VM
- Small code cleanups in testrelpol as well.
A missing check to see whether the range being transferred is sane
(with a starting address lower than an ending address) caused extra
memory to be marked erroneously as copy-on-write for some processes,
ultimately resulting in pagefaults on the stack during live update
rollback.
Dirk Vogt [Mon, 19 Jan 2015 14:20:30 +0000 (15:20 +0100)]
VM: live update - check for regions above stack
If the stack is not mapped at the VM_DATATOP (e.g. booted with
ac_layout = 1), there might be some more regions hiding above
the stack. We also have to transfer those.
The 'memory' service has holes in its data section, which causes
problems during state transfer. Since VM cannot handle page faults
during a multicomponent-with-VM live update, the state transfer must
ensure that no page faults occur during copying. Therefore, we now
query VM about the regions to copy, thus skipping holes. While the
solution is not ideal, it is sufficiently generic that it can be used
for the data section state transfer of all processes, and possibly
for state transfer of other regions in the future as well.
Ben Gras [Thu, 15 Jan 2015 15:47:46 +0000 (16:47 +0100)]
vm: restartability improvements (#1)
Two bugs fixed wrt vm restartability.
. make sure pagetable data is only allocated
using dynamic data instead of static spare pages
(bootstrap pages). They are needed for bootstrap
but now repeat some of the initialization so only
dynamic data remains. This solves the problem of
physical addresses changing (as static pages are
re-allocated for the new instance) after update.
. pt_ptalloc has to be specified in bytes instead of
pde slot numbers. leaving pt_pt NULL causes mapping
transfers to fail because NULL happens to be mapped in
then and updates then happen there.
. added some sanity checks against the above happening.
The new state is that VM can update many times, but the system
isn't fully reliable afterwards yet.
Ben Gras [Sun, 28 Jun 2015 22:07:29 +0000 (00:07 +0200)]
Kernel: delivermsg improvements
. make arch-independent, and local to proc.c, reduce code duplication
. make vm_suspend public but unduplicated in proc.c
. ask VM for handling once, 2nd time SIGSEGV process
. remove debug printfs
. test case for bogus sendrec() address argument
Allow extra space for in-band metadata when allocating cache blocks.
Edited by David van Moolenbroek: since this effectively halves the
potential size of the typical file system cache, do this only when
compiling with instrumentation.
Edited by David van Moolenbroek to deallocate the guard page as well.
Note that while the new approach is better in theory (previously, the
hole could end up being filled by another allocated page), guard page
protection is now broken in practice, because VM does not support
setting specific page permissions (in this case, PROT_NONE).
VM used to call sendrec to send a boot-time RS_INIT reply to RS, but
RS could run into a pagefault at the same time, thus spawning a
message to VM, resulting in a deadlock. We resolve this situation by
making VM acknowledge RS_INIT asynchronously at boot time, while
retaining the synchronous sendrec for subsequent RS_INIT responses.
The following services have been updated to support stateful restarts:
- Drivers: tty
- Filesystems: isofs, mfs, pfs, libvtreefs-based file servers
- System servers: tty, ds, pm, vfs, vm
Previously, RS would clean up dead services only when it is idle.
During shutdown, all services are marked with the 'exiting' flag,
and these flags lead RS to conclude it is not idle. Therefore, at
shutdown time, no services were cleaned up anymore, leading to
deadlock situations. For example, VFS could end up waiting for a
service that was already dead, or one driver could end up waiting
for an interrupt on a line shared with another dead driver.
While it may be possible to ignore RS_EXITING when checking idle
status, other flags may have the same ultimate effect. Therefore,
this patch skips the idle check altogether when in shutdown mode.
While in a multicomponent live update that includes RS, the new RS
instance may receive heartbeat replies which, after a rollback, the
old RS instance will then never see. As a result, the rolled-back
RS instance may end up killing well-behaving services.
Some select queries require a response from device drivers. If a
select call is nonblocking (with a zero timeout), the response to
the caller may have to be deferred until all involved drivers have
responded to the initial query. This is handled just fine.
However, if the select call has a timeout that is so short that it
triggers before all the involved drivers have responded, the
resulting alarm would be discarded, possibly resulting in the call
blocking forever. This fix changes the alarm handler such that if
the alarm triggers too early, the select call is further handled
as though it was nonblocking.
This fix resolves a test77 deadlock on really slow systems.
- ping(1) triggers warnings about unimplemented exceptions for select;
even if there could be a useful implementation (which is doubtful),
the warnings are not helping anyone right now;
- the clock_t data type has changed.
Fix /dev/tty-related issues in tmux(1) by hardcoding the PTY major
in VFS in addition to the TTY major. Even though this is exactly
what we did NOT want to have to do, the actual fix for this issue
is going to take a little longer.
The previous approach of storing pointers to messages structures for
thread-blocking sendrec operations relied on several assumptions,
which if violated could lead to odd cases of memory corruption.
With this patch, VFS resets pointers right after use, avoiding that
any dangling pointers are accidentally dereferenced later. This
approach was already used in some cases, but not all of them.
For dynamically linked executables, the interpreter is passed a
file descriptor of the binary being executed. To this end, VFS
opens the target executable, but opening the file fails if it is
not readable, even when it is executable. With this patch, when
opening the executable, it verifies the X bit rather than the R
bit on the file, thus allowing the execution of dynamically
linked binaries that are executable but not readable.
This patch changes the prefetch API so that file systems must now
provide a set of block numbers, rather than a set of buffers. The
result is a leaner and more well-defined API; linear computation of
the range of blocks to prefetch; duplicates no longer interfering
with the prefetch process; guaranteed inclusion of the block needed
next into the prefetch range; and, limits and policy decisions better
established by libminixfs now actually being moved into libminixfs.
libminixfs: better support for read errors and EOF
- The lmfs_get_block*(3) API calls may now return an error. The idea
is to encourage a next generation of file system services to do a
better job at dealing with block read errors than the MFS-derived
implementations do. These existing file systems have been changed
to panic immediately upon getting a block read error, in order to
let unchecked errors cause corruption. Note that libbdev already
retries failing I/O operations a few times first.
- The libminixfs block device I/O module (bio.c) now deals properly
with end-of-file conditions on block devices. Since a device or
partition size may not be a multiple of the root file system's block
size, support for partial block retrival has been added, with a new
internal lmfs_get_partial_block(3) call. A new test program,
test85, tests the new handling of EOF conditions when reading,
writing, and memory-mapping a block device.
This patch changes the libminixfs API and implementation such that the
library is at all times aware of how many total and used blocks there
are in the file system. This removes the last upcall of libminixfs
into file systems (fs_blockstats). In the process, make this part of
the libminixfs API a little prettier and more robust. Change file
systems accordingly. Since this change only adds to MFS being unable
to deal with zones and blocks having different sizes, fail to mount
such file systems immediately rather than triggering an assert later.
This removes an implicit requirement for the way the libminixfs API is
to be used, namely that a block is to be marked as dirty only once its
contents have been fully updated, within a single get_block/put_block
window. The requirement may not be appropriate for all file systems.
With this change, the lmfs_get_block*(3) functions allow the caller to
specify that it only wants the block if it is in the cache or the
secondary VM cache. If the block is not found there, the functions
return NULL. Previously, the PREFETCH method would be used to this
end instead, which was both abuse in name and less efficient.
When VM asks a file system to provide a block to satisfy a page fault
on a file memory mapping, the file system previously had no way to
inform VM that the block is a hole, since there is no corresponding
block on the underlying device. To work around this, MFS and ext2
would actually allocate a block for the hole when asked by VM, which
not only defeats the point of holes in the first place, but also does
not work on read-only file systems. With this patch, a new libminixfs
call allows the file system to inform VM about holes. This issue does
raise the question as to whether the VM cache is using the right data
structures, since there are now two places where we have to fake a
device offset. This will have to be revisited in the future.
The patch changes file systems accordingly, and adds a test to test74.
This patch employs one solution to resolve two independent but related
issues. Both issues are the result of one fundamental aspect of the
way VM's memory mapping works: VM uses its cache to map in blocks for
memory-mapped file regions, and for blocks already in the VM cache, VM
does not go to the file system before mapping them in. To preserve
consistency between the FS and VM caches, VM relies on being informed
about all updates to file contents through the block cache. The two
issues are both the result of VM not being properly informed about
such updates:
1. Once a file system provides libminixfs with an inode association
(inode number + inode offset) for a disk block, this association
is not broken until a new inode association is provided for it.
If a block is freed and reallocated as a metadata (non-inode)
block, its old association is maintained, and may be supplied to
VM's secondary cache. Due to reuse of inodes, it is possible
that the same inode association becomes valid for an actual file
block again. In that case, when that new file is memory-mapped,
under certain circumstances, VM may end up using the metadata
block to satisfy a page fault on the file, due to the stale inode
association. The result is a corrupted memory mapping, with the
application seeing data other than the current file contents
mapped in at the file block.
2. When a hole is created in a file, the underlying block is freed
from the device, but VM is not informed of this update, and thus,
if VM's cache contains the block with its previous inode
association, this block will remain there. As a result, if an
application subsequently memory-maps the file, VM will map in the
old block at the position of the hole, rather than an all-zeroes
block. Thus, again, the result is a corrupted memory mapping.
This patch resolves both issues by making the file system inform the
minixfs library about blocks being freed, so that libminixfs can
break the inode association for that block, both in its own cache and
in the VM cache. Since libminixfs does not know whether VM has the
block in its cache or not, it makes a call to VM for each block being
freed. Thus, this change introduces more calls to VM, but it solves
the correctness issues at hand; optimizations may be introduced
later. On the upside, all freed blocks are now marked as clean,
which should result in fewer blocks being written back to the device,
and the blocks are removed from the caches entirely, which should
result in slightly better cache usage.
This patch is necessary but not sufficient to resolve the situation
with respect to memory mapping of file holes in general. Therefore,
this patch extends test 74 with a (rather particular but effective)
test for the first issue, but not yet with a test for the second one.
There is no reason to keep these tightly coupled data structures
separate. Moreover, there is no reason to have a union of file
descriptor and file pointer, since the second can be derived from
the first. The result are somewhat cleaner VFS internals.