The system_callback() function in trace_events.c is only used within that
file. The "static" annotation was missed.
Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode") Reported-by: kernel test robot <lkp@intel.com> Closes: https://lore.kernel.org/oe-kbuild-all/202310051743.y9EobbUr-lkp@intel.com/ Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
My mainline commit c48a7c44a1d0 ("docs: kernel_feat.py: fix potential
command injection") contains a bug which can manifests like this when
building the documentation:
Sphinx parallel build error:
UnboundLocalError: local variable 'fname' referenced before assignment
make[2]: *** [Documentation/Makefile:102: htmldocs] Error 2
However, this only appears when there exists a '.. kernel-feat::'
directive that points to a non-existent file, which isn't the case in
mainline.
When this commit was backported to stable 6.6, it didn't change
Documentation/arch/ia64/features.rst since ia64 was removed in 6.7 in
commit cf8e8658100d ("arch: Remove Itanium (IA-64) architecture"). This
lead to the build failure seen above -- but only in stable kernels.
This patch fixes the backport and should only be applied to kernels where
Documentation/arch/ia64/features.rst exists and commit c48a7c44a1d0 has
also been applied.
A second patch will follow to fix kernel_feat.py in mainline so that it
doesn't error out when the '.. kernel-feat::' directive points to a
nonexistent file.
AMD does not have the requirement for a synchronization barrier when
acccessing a certain group of MSRs. Do not incur that unnecessary
penalty there.
There will be a CPUID bit which explicitly states that a MFENCE is not
needed. Once that bit is added to the APM, this will be extended with
it.
While at it, move to processor.h to avoid include hell. Untangling that
file properly is a matter for another day.
Some notes on the performance aspect of why this is relevant, courtesy
of Kishon VijayAbraham <Kishon.VijayAbraham@amd.com>:
On a AMD Zen4 system with 96 cores, a modified ipi-bench[1] on a VM
shows x2AVIC IPI rate is 3% to 4% lower than AVIC IPI rate. The
ipi-bench is modified so that the IPIs are sent between two vCPUs in the
same CCX. This also requires to pin the vCPU to a physical core to
prevent any latencies. This simulates the use case of pinning vCPUs to
the thread of a single CCX to avoid interrupt IPI latency.
In order to avoid run-to-run variance (for both x2AVIC and AVIC), the
below configurations are done:
1) Disable Power States in BIOS (to prevent the system from going to
lower power state)
2) Run the system at fixed frequency 2500MHz (to prevent the system
from increasing the frequency when the load is more)
With the above configuration:
*) Performance measured using ipi-bench for AVIC:
Average Latency: 1124.98ns [Time to send IPI from one vCPU to another vCPU]
Cumulative throughput: 42.6759M/s [Total number of IPIs sent in a second from
48 vCPUs simultaneously]
*) Performance measured using ipi-bench for x2AVIC:
Average Latency: 1172.42ns [Time to send IPI from one vCPU to another vCPU]
Cumulative throughput: 40.9432M/s [Total number of IPIs sent in a second from
48 vCPUs simultaneously]
From above, x2AVIC latency is ~4% more than AVIC. However, the expectation is
x2AVIC performance to be better or equivalent to AVIC. Upon analyzing
the perf captures, it is observed significant time is spent in
weak_wrmsr_fence() invoked by x2apic_send_IPI().
With the fix to skip weak_wrmsr_fence()
*) Performance measured using ipi-bench for x2AVIC:
Average Latency: 1117.44ns [Time to send IPI from one vCPU to another vCPU]
Cumulative throughput: 42.9608M/s [Total number of IPIs sent in a second from
48 vCPUs simultaneously]
Comparing the performance of x2AVIC with and without the fix, it can be seen
the performance improves by ~4%.
Performance captured using an unmodified ipi-bench using the 'mesh-ipi' option
with and without weak_wrmsr_fence() on a Zen4 system also showed significant
performance improvement without weak_wrmsr_fence(). The 'mesh-ipi' option ignores
CCX or CCD and just picks random vCPU.
Average throughput (10 iterations) with weak_wrmsr_fence(),
Cumulative throughput: 4933374 IPI/s
Average throughput (10 iterations) without weak_wrmsr_fence(),
Cumulative throughput: 6355156 IPI/s
According to a syzbot report, end_buffer_async_write(), which handles the
completion of block device writes, may detect abnormal condition of the
buffer async_write flag and cause a BUG_ON failure when using nilfs2.
Nilfs2 itself does not use end_buffer_async_write(). But, the async_write
flag is now used as a marker by commit 7f42ec394156 ("nilfs2: fix issue
with race condition of competition between segments for dirty blocks") as
a means of resolving double list insertion of dirty blocks in
nilfs_lookup_dirty_data_buffers() and nilfs_lookup_node_buffers() and the
resulting crash.
This modification is safe as long as it is used for file data and b-tree
node blocks where the page caches are independent. However, it was
irrelevant and redundant to also introduce async_write for segment summary
and super root blocks that share buffers with the backing device. This
led to the possibility that the BUG_ON check in end_buffer_async_write
would fail as described above, if independent writebacks of the backing
device occurred in parallel.
The use of async_write for segment summary buffers has already been
removed in a previous change.
Fix this issue by removing the manipulation of the async_write flag for
the remaining super root block buffer.
Link: https://lkml.kernel.org/r/20240203161645.4992-1-konishi.ryusuke@gmail.com Fixes: 7f42ec394156 ("nilfs2: fix issue with race condition of competition between segments for dirty blocks") Signed-off-by: Ryusuke Konishi <konishi.ryusuke@gmail.com> Reported-by: syzbot+5c04210f7c7f897c1e7f@syzkaller.appspotmail.com Closes: https://lkml.kernel.org/r/00000000000019a97c05fd42f8c8@google.com Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Similar to the existing "ports" node name, coresight device tree bindings
have added "in-ports" and "out-ports" as standard node names for a
collection of ports.
Add support for these name to of_graph_get_port_parent() so that
remote-endpoint parsing can find the correct parent node for these
coresight ports too.
On some systems, sys_membarrier can be very expensive, causing overall
slowdowns for everything. So put a lock on the path in order to
serialize the accesses to prevent the ability for this to be called at
too high of a frequency and saturate the machine.
The .compat section is a dummy PE section that contains the address of
the 32-bit entrypoint of the 64-bit kernel image if it is bootable from
32-bit firmware (i.e., CONFIG_EFI_MIXED=y)
This section is only 8 bytes in size and is only referenced from the
loader, and so it is placed at the end of the memory view of the image,
to avoid the need for padding it to 4k, which is required for sections
appearing in the middle of the image.
Unfortunately, this violates the PE/COFF spec, and even if most EFI
loaders will work correctly (including the Tianocore reference
implementation), PE loaders do exist that reject such images, on the
basis that both the file and memory views of the file contents should be
described by the section headers in a monotonically increasing manner
without leaving any gaps.
So reorganize the sections to avoid this issue. This results in a slight
padding overhead (< 4k) which can be avoided if desired by disabling
CONFIG_EFI_MIXED (which is only needed in rare cases these days)
Align x86 with other EFI architectures, and increase the section
alignment to the EFI page size (4k), so that firmware is able to honour
the section permission attributes and map code read-only and data
non-executable.
There are a number of requirements that have to be taken into account:
- the sign tools get cranky when there are gaps between sections in the
file view of the image
- the virtual offset of each section must be aligned to the image's
section alignment
- the file offset *and size* of each section must be aligned to the
image's file alignment
- the image size must be aligned to the section alignment
- each section's virtual offset must be greater than or equal to the
size of the headers.
In order to meet all these requirements, while avoiding the need for
lots of padding to accommodate the .compat section, the latter is placed
at an arbitrary offset towards the end of the image, but aligned to the
minimum file alignment (512 bytes). The space before the .text section
is therefore distributed between the PE header, the .setup section and
the .compat section, leaving no gaps in the file coverage, making the
signing tools happy.
Describe the code and data of the decompressor binary using separate
.text and .data PE/COFF sections, so that we will be able to map them
using restricted permissions once we increase the section and file
alignment sufficiently. This avoids the need for memory mappings that
are writable and executable at the same time, which is something that
is best avoided for security reasons.
Ancient buggy EFI loaders may have required a .reloc section to be
present at some point in time, but this has not been true for a long
time so the .reloc section can just be dropped.
Now that the size of the setup block is visible to the assembler, it is
possible to populate the PE/COFF header fields from the asm code
directly, instead of poking the values into the binary using the build
tool. This will make it easier to reorganize the section layout without
having to tweak the build tool in lockstep.
This change has no impact on the resulting bzImage binary.
Tweak the linker script so that the value of _edata represents the
decompressor binary's file size rounded up to the appropriate alignment.
This removes the need to calculate it in the build tool, and will make
it easier to refer to the file size from the header directly in
subsequent changes to the PE header layout.
While adding _edata to the sed regex that parses the compressed
vmlinux's symbol list, tweak the regex a bit for conciseness.
This change has no impact on the resulting bzImage binary when
configured with CONFIG_EFI_STUB=y.
The setup block contains the real mode startup code that is used when
booting from a legacy BIOS, along with the boot_params/setup_data that
is used by legacy x86 bootloaders to pass the command line and initial
ramdisk parameters, among other things.
The setup block also contains the PE/COFF header of the entire combined
image, which includes the compressed kernel image, the decompressor and
the EFI stub.
This PE header describes the layout of the executable image in memory,
and currently, the fact that the setup block precedes it makes it rather
fiddly to get the right values into the right place in the final image.
Let's make things a bit easier by defining the setup_size in the linker
script so it can be referenced from the asm code directly, rather than
having to rely on the build tool to calculate it. For the time being,
add 64 bytes of fixed padding for the .reloc and .compat sections - this
will be removed in a subsequent patch after the PE/COFF header has been
reorganized.
This change has no impact on the resulting bzImage binary when
configured with CONFIG_EFI_MIXED=y.
The offsets of the EFI handover entrypoints are available to the
assembler when constructing the header, so there is no need to set them
from the build tool afterwards.
This change has no impact on the resulting bzImage binary.
Instead of parsing zoffset.h and poking the kernel_info offset value
into the header from the build tool, just grab the value directly in the
asm file that describes this header.
This change has no impact on the resulting bzImage binary.
The x86 boot image generation tool assign a default value to startup_64
and subsequently parses the actual value from zoffset.h but it never
actually uses the value anywhere. So remove this code.
This change has no impact on the resulting bzImage binary.
Now that the EFI stub decompresses the kernel and hands over to the
decompressed image directly, there is no longer a need to provide a
decompression buffer as part of the .BSS allocation of the PE/COFF
image. It also means the PE/COFF image can be loaded anywhere in memory,
and setting the preferred image base is unnecessary. So drop the
handling of this from the header and from the build tool.
Ancient (pre-2003) x86 kernels could boot from a floppy disk straight from
the BIOS, using a small real mode boot stub at the start of the image
where the BIOS would expect the boot record (or boot block) to appear.
Due to its limitations (kernel size < 1 MiB, no support for IDE, USB or
El Torito floppy emulation), this support was dropped, and a Linux aware
bootloader is now always required to boot the kernel from a legacy BIOS.
To smoothen this transition, the boot stub was not removed entirely, but
replaced with one that just prints an error message telling the user to
install a bootloader.
As it is unlikely that anyone doing direct floppy boot with such an
ancient kernel is going to upgrade to v6.5+ and expect that this boot
method still works, printing this message is kind of pointless, and so
it should be possible to remove the logic that emits it.
Let's free up this space so it can be used to expand the PE header in a
subsequent patch.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: H. Peter Anvin (Intel) <hpa@zytor.com> Link: https://lore.kernel.org/r/20230912090051.4014114-21-ardb@google.com Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The section header flags for alignment are documented in the PE/COFF
spec as being applicable to PE object files only, not to PE executables
such as the Linux bzImage, so let's drop them from the PE header.
The native EFI entrypoint does not take a struct boot_params from the
loader, but instead, it constructs one from scratch, using the setup
header data placed at the start of the image.
This setup header is placed in a way that permits legacy loaders to
manipulate the contents (i.e., to pass the kernel command line or the
address and size of an initial ramdisk), but EFI boot does not use it in
that way - it only copies the contents that were placed there at build
time, but EFI loaders will not (and should not) manipulate the setup
header to configure the boot. (Commit 63bf28ceb3ebbe76 "efi: x86: Wipe
setup_data on pure EFI boot" deals with some of the fallout of using
setup_data in a way that breaks EFI boot.)
Given that none of the non-zero values that are copied from the setup
header into the EFI stub's struct boot_params are relevant to the boot
now that the EFI stub no longer enters via the legacy decompressor, the
copy can be omitted altogether.
Now that the EFI stub always zero inits its BSS section upon entry,
there is no longer a need to place the BSS symbols carried by the stub
into the .data section.
A recent change to check_for_locks() changed it to take ->flc_lock while
holding ->fi_lock. This creates a lock inversion (reported by lockdep)
because there is a case where ->fi_lock is taken while holding
->flc_lock.
->flc_lock is held across ->fl_lmops callbacks, and
nfsd_break_deleg_cb() is one of those and does take ->fi_lock. However
it doesn't need to.
Prior to v4.17-rc1~110^2~22 ("nfsd: create a separate lease for each
delegation") nfsd_break_deleg_cb() would walk the ->fi_delegations list
and so needed the lock. Since then it doesn't walk the list and doesn't
need the lock.
Two actions are performed under the lock. One is to call
nfsd_break_one_deleg which calls nfsd4_run_cb(). These doesn't act on
the nfs4_file at all, so don't need the lock.
The other is to set ->fi_had_conflict which is in the nfs4_file.
This field is only ever set here (except when initialised to false)
so there is no possible problem will multiple threads racing when
setting it.
The field is tested twice in nfs4_set_delegation(). The first test does
not hold a lock and is documented as an opportunistic optimisation, so
it doesn't impose any need to hold ->fi_lock while setting
->fi_had_conflict.
The second test in nfs4_set_delegation() *is* make under ->fi_lock, so
removing the locking when ->fi_had_conflict is set could make a change.
The change could only be interesting if ->fi_had_conflict tested as
false even though nfsd_break_one_deleg() ran before ->fi_lock was
unlocked. i.e. while hash_delegation_locked() was running.
As hash_delegation_lock() doesn't interact in any way with nfs4_run_cb()
there can be no importance to this interaction.
So this patch removes the locking from nfsd_break_one_deleg() and moves
the final test on ->fi_had_conflict out of the locked region to make it
clear that locking isn't important to the test. It is still tested
*after* vfs_setlease() has succeeded. This might be significant and as
vfs_setlease() takes ->flc_lock, and nfsd_break_one_deleg() is called
under ->flc_lock this "after" is a true ordering provided by a spinlock.
Fixes: edcf9725150e ("nfsd: fix RELEASE_LOCKOWNER") Signed-off-by: NeilBrown <neilb@suse.de> Reviewed-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The directory link count in eventfs was somewhat bogus. It was only being
updated when a directory child was being looked up and not on creation.
One solution would be to update in get_attr() the link count by iterating
the ei->children list and then adding 2. But that could slow down simple
stat() calls, especially if it's done on all directories in eventfs.
Another solution would be to add a parent pointer to the eventfs_inode
and keep track of the number of sub directories it has on creation. But
this adds overhead for something not really worthwhile.
The solution decided upon is to keep all directory links in eventfs as 1.
This tells user space not to rely on the hard links of directories. Which
in this case it shouldn't.
Link: https://lore.kernel.org/linux-trace-kernel/20240201002719.GS2087318@ZenIV/ Link: https://lore.kernel.org/linux-trace-kernel/20240201161617.339968298@goodmis.org Cc: stable@vger.kernel.org Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Christian Brauner <brauner@kernel.org> Cc: Al Viro <viro@ZenIV.linux.org.uk> Cc: Ajay Kaher <ajay.kaher@broadcom.com> Fixes: c1504e510238 ("eventfs: Implement eventfs dir creation functions") Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The dentries and inodes are created when referenced in the lookup code.
There's no reason to call fsnotify_*() functions when they are created by
a reference. It doesn't make any sense.
Link: https://lore.kernel.org/linux-trace-kernel/20240201002719.GS2087318@ZenIV/ Link: https://lore.kernel.org/linux-trace-kernel/20240201161617.166973329@goodmis.org Cc: stable@vger.kernel.org Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Christian Brauner <brauner@kernel.org> Cc: Al Viro <viro@ZenIV.linux.org.uk> Cc: Ajay Kaher <ajay.kaher@broadcom.com> Fixes: a376007917776 ("eventfs: Implement functions to create files and dirs when accessed"); Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There should never be a case where an evenfs_inode is being freed without
is_freed being set. Add a WARN_ON_ONCE() if it ever happens. That would
mean there was one too many put_ei()s.
Link: https://lore.kernel.org/linux-trace-kernel/20240201161616.843551963@goodmis.org Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Christian Brauner <brauner@kernel.org> Cc: Al Viro <viro@ZenIV.linux.org.uk> Cc: Ajay Kaher <ajay.kaher@broadcom.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs inode had pointers to dentries (and child dentries) without
actually holding a refcount on said pointer. That is fundamentally
broken, and while eventfs tried to then maintain coherence with dentries
going away by hooking into the '.d_iput' callback, that doesn't actually
work since it's not ordered wrt lookups.
There were two reasonms why eventfs tried to keep a pointer to a dentry:
- the creation of a 'events' directory would actually have a stable
dentry pointer that it created with tracefs_start_creating().
And it needed that dentry when tearing it all down again in
eventfs_remove_events_dir().
This use is actually ok, because the special top-level events
directory dentries are actually stable, not just a temporary cache of
the eventfs data structures.
- the 'eventfs_inode' (aka ei) needs to stay around as long as there
are dentries that refer to it.
It then used these dentry pointers as a replacement for doing
reference counting: it would try to make sure that there was only
ever one dentry associated with an event_inode, and keep a child
dentry array around to see which dentries might still refer to the
parent ei.
This gets rid of the invalid dentry pointer use, and renames the one
valid case to a different name to make it clear that it's not just any
random dentry.
The magic child dentry array that is kind of a "reverse reference list"
is simply replaced by having child dentries take a ref to the ei. As
does the directory dentries. That makes the broken use case go away.
In order for the dentries to stay up-to-date with the eventfs changes,
just add a 'd_revalidate' function that checks the 'is_freed' bit.
Also, clean up the dentry release to actually use d_release() rather
than the slightly odd d_iput() function. We don't care about the inode,
all we want to do is to get rid of the refcount to the eventfs data
added by dentry->d_fsdata.
It would probably be cleaner to make eventfs its own filesystem, or at
least set its own dentry ops when looking up eventfs files. But as it
is, only eventfs dentries use d_fsdata, so we don't really need to split
these things up by use.
Another thing that might be worth doing is to make all eventfs lookups
mark their dentries as not worth caching. We could do that with
d_delete(), but the DCACHE_DONTCACHE flag would likely be even better.
As it is, the dentries are all freeable, but they only tend to get freed
at memory pressure rather than more proactively. But that's a separate
issue.
The dentry lookup for eventfs files was very broken, and had lots of
signs of the old situation where the filesystem names were all created
statically in the dentry tree, rather than being looked up dynamically
based on the eventfs data structures.
You could see it in the naming - how it claimed to "create" dentries
rather than just look up the dentries that were given it.
You could see it in various nonsensical and very incorrect operations,
like using "simple_lookup()" on the dentries that were passed in, which
only results in those dentries becoming negative dentries. Which meant
that any other lookup would possibly return ENOENT if it saw that
negative dentry before the data was then later filled in.
You could see it in the immense amount of nonsensical code that didn't
actually just do lookups.
The eventfs_find_events() code tries to walk up the tree to find the
event directory that a dentry belongs to, in order to then find the
eventfs inode that is associated with that event directory.
However, it uses an odd combination of walking the dentry parent,
looking up the eventfs inode associated with that, and then looking up
the dentry from there. Repeat.
But the code shouldn't have back-pointers to dentries in the first
place, and it should just walk the dentry parenthood chain directly.
Similarly, 'set_top_events_ownership()' looks up the dentry from the
eventfs inode, but the only reason it wants a dentry is to look up the
superblock in order to look up the root dentry.
But it already has the real filesystem inode, which has that same
superblock pointer. So just pass in the superblock pointer using the
information that's already there, instead of looking up extraneous data
that is irrelevant.
eventfs uses the tracefs_inode and assumes that it's already initialized
to zero. That is, it doesn't set fields to zero (like ti->private) after
getting its tracefs_inode. This causes bugs due to stale values.
Just initialize the entire structure to zero on allocation so there isn't
any more surprises.
This is a partial fix to access to ti->private. The assignment still needs
to be made before the dentry is instantiated.
Link: https://lore.kernel.org/linux-trace-kernel/20240131185512.315825944@goodmis.org Cc: stable@vger.kernel.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Christian Brauner <brauner@kernel.org> Cc: Al Viro <viro@ZenIV.linux.org.uk> Cc: Ajay Kaher <ajay.kaher@broadcom.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode") Reported-by: kernel test robot <oliver.sang@intel.com> Closes: https://lore.kernel.org/oe-lkp/202401291043.e62e89dc-oliver.sang@intel.com Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs inodes and directories are allocated when referenced. But this
leaves the issue of keeping consistent inode numbers and the number is
only saved in the inode structure itself. When the inode is no longer
referenced, it can be freed. When the file that the inode was representing
is referenced again, the inode is once again created, but the inode number
needs to be the same as it was before.
Just making the inode numbers the same for all files is fine, but that
does not work with directories. The find command will check for loops via
the inode number and having the same inode number for directories triggers:
# find /sys/kernel/tracing
find: File system loop detected;
'/sys/kernel/debug/tracing/events/initcall/initcall_finish' is part of the same file system loop as
'/sys/kernel/debug/tracing/events/initcall'.
[..]
Linus pointed out that the eventfs_inode structure ends with a single
32bit int, and on 64 bit machines, there's likely a 4 byte hole due to
alignment. We can use this hole to store the inode number for the
eventfs_inode. All directories in eventfs are represented by an
eventfs_inode and that data structure can hold its inode number.
That last int was also purposely placed at the end of the structure to
prevent holes from within. Now that there's a 4 byte number to hold the
inode, both the inode number and the last integer can be moved up in the
structure for better cache locality, where the llist and rcu fields can be
moved to the end as they are only used when the eventfs_inode is being
deleted.
As noted in the "Deprecated Interfaces, Language Features, Attributes,
and Conventions" documentation [1], size calculations (especially
multiplication) should not be performed in memory allocator (or similar)
function arguments due to the risk of them overflowing. This could lead
to values wrapping around and a smaller allocation being made than the
caller was expecting. Using those allocations could lead to linear
overflows of heap memory and other misbehaviors.
So, use the purpose specific kcalloc() function instead of the argument
size * count in the kzalloc() function.
The original eventfs code added a wrapper around the dcache_readdir open
callback and created all the dentries and inodes at open, and increment
their ref count. A wrapper was added around the dcache_readdir release
function to decrement all the ref counts of those created inodes and
dentries. But this proved to be buggy[1] for when a kprobe was created
during a dir read, it would create a dentry between the open and the
release, and because the release would decrement all ref counts of all
files and directories, that would include the kprobe directory that was
not there to have its ref count incremented in open. This would cause the
ref count to go to negative and later crash the kernel.
To solve this, the dentries and inodes that were created and had their ref
count upped in open needed to be saved. That list needed to be passed from
the open to the release, so that the release would only decrement the ref
counts of the entries that were incremented in the open.
Unfortunately, the dcache_readdir logic was already using the
file->private_data, which is the only field that can be used to pass
information from the open to the release. What was done was the eventfs
created another descriptor that had a void pointer to save the
dcache_readdir pointer, and it wrapped all the callbacks, so that it could
save the list of entries that had their ref counts incremented in the
open, and pass it to the release. The wrapped callbacks would just put
back the dcache_readdir pointer and call the functions it used so it could
still use its data[2].
But Linus had an issue with the "hijacking" of the file->private_data
(unfortunately this discussion was on a security list, so no public link).
Which we finally agreed on doing everything within the iterate_shared
callback and leave the dcache_readdir out of it[3]. All the information
needed for the getents() could be created then.
But this ended up being buggy too[4]. The iterate_shared callback was not
the right place to create the dentries and inodes. Even Christian Brauner
had issues with that[5].
An attempt was to go back to creating the inodes and dentries at
the open, create an array to store the information in the
file->private_data, and pass that information to the other callbacks.[6]
The difference between that and the original method, is that it does not
use dcache_readdir. It also does not up the ref counts of the dentries and
pass them. Instead, it creates an array of a structure that saves the
dentry's name and inode number. That information is used in the
iterate_shared callback, and the array is freed in the dir release. The
dentries and inodes created in the open are not used for the iterate_share
or release callbacks. Just their names and inode numbers.
Linus did not like that either[7] and just wanted to remove the dentries
being created in iterate_shared and use the hard coded inode numbers.
[ All this while Linus enjoyed an unexpected vacation during the merge
window due to lack of power. ]
The dentries and inodes are created in the readdir for the sole purpose of
getting a consistent inode number. Linus stated that is unnecessary, and
that all inodes can have the same inode number. For a virtual file system
they are pretty meaningless.
Instead use a single unique inode number for all files and one for all
directories.
As the ei->entries array is fixed for the duration of the eventfs_inode,
it can be used to skip over already read entries in eventfs_iterate().
That is, if ctx->pos is greater than zero, there's no reason in doing the
loop across the ei->entries array for the entries less than ctx->pos.
Instead, start the lookup of the entries at the current ctx->pos.
In order to apply a shortcut to skip over the current ctx->pos
immediately, by using the ei->entries array, the reading of that array
should be first. Moving the array reading before the linked list reading
will make the shortcut change diff nicer to read.
The ctx->pos was only updated when it added an entry, but the "skip to
current pos" check (c--) happened for every loop regardless of if the
entry was added or not. This inconsistency caused readdir to be incorrect.
It was due to:
for (i = 0; i < ei->nr_entries; i++) {
if (c > 0) {
c--;
continue;
}
mutex_lock(&eventfs_mutex);
/* If ei->is_freed then just bail here, nothing more to do */
if (ei->is_freed) {
mutex_unlock(&eventfs_mutex);
goto out;
}
r = entry->callback(name, &mode, &cdata, &fops);
mutex_unlock(&eventfs_mutex);
[..]
ctx->pos++;
}
But this can cause the iterator to return a file that was already read.
That's because of the way the callback() works. Some events may not have
all files, and the callback can return 0 to tell eventfs to skip the file
for this directory.
for instance, we have:
# ls /sys/kernel/tracing/events/ftrace/function
format hist hist_debug id inject
and
# ls /sys/kernel/tracing/events/sched/sched_switch/
enable filter format hist hist_debug id inject trigger
Where the function directory is missing "enable", "filter" and
"trigger". That's because the callback() for events has:
/*
* Only event directories that can be enabled should have
* triggers or filters, with the exception of the "print"
* event that can have a "trigger" file.
*/
if (!(call->flags & TRACE_EVENT_FL_IGNORE_ENABLE)) {
if (call->class->reg && strcmp(name, "enable") == 0) {
*mode = TRACE_MODE_WRITE;
*fops = &ftrace_enable_fops;
return 1;
}
Where the function event has the TRACE_EVENT_FL_IGNORE_ENABLE set.
This means that the entries array elements for "enable", "filter" and
"trigger" when called on the function event will have the callback return
0 and not 1, to tell eventfs to skip these files for it.
Because the "skip to current ctx->pos" check happened for all entries, but
the ctx->pos++ only happened to entries that exist, it would confuse the
reading of a directory. Which would cause:
# ls /sys/kernel/tracing/events/ftrace/function/
format hist hist hist_debug hist_debug id inject inject
The missing "enable", "filter" and "trigger" caused ls to show "hist",
"hist_debug" and "inject" twice.
Update the ctx->pos for every iteration to keep its update and the "skip"
update consistent. This also means that on error, the ctx->pos needs to be
decremented if it was incremented without adding something.
Link: https://lore.kernel.org/all/20240104150500.38b15a62@gandalf.local.home/ Link: https://lore.kernel.org/linux-trace-kernel/20240104220048.172295263@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Fixes: 493ec81a8fb8e ("eventfs: Stop using dcache_readdir() for getdents()") Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If ei->is_freed is set in eventfs_iterate(), it means that the directory
that is being iterated on is in the process of being freed. Just exit the
loop immediately when that is ever detected, and separate out the return
of the entry->callback() from ei->is_freed.
Link: https://lore.kernel.org/linux-trace-kernel/20240104220048.016261289@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Instead of walking the dentries on mount/remount to update the gid values of
all the dentries if a gid option is specified on mount, just update the root
inode. Add .getattr, .setattr, and .permissions on the tracefs inode
operations to update the permissions of the files and directories.
For all files and directories in the top level instance:
/sys/kernel/tracing/*
It will use the root inode as the default permissions. The inode that
represents: /sys/kernel/tracing (or wherever it is mounted).
When an instance is created:
mkdir /sys/kernel/tracing/instance/foo
The directory "foo" and all its files and directories underneath will use
the default of what foo is when it was created. A remount of tracefs will
not affect it.
If a user were to modify the permissions of any file or directory in
tracefs, it will also no longer be modified by a change in ownership of a
remount.
The events directory, if it is in the top level instance, will use the
tracefs root inode as the default ownership for itself and all the files and
directories below it.
For the events directory in an instance ("foo"), it will keep the ownership
of what it was when it was created, and that will be used as the default
ownership for the files and directories beneath it.
The eventfs creates dynamically allocated dentries and inodes. Using the
dcache_readdir() logic for its own directory lookups requires hiding the
cursor of the dcache logic and playing games to allow the dcache_readdir()
to still have access to the cursor while the eventfs saved what it created
and what it needs to release.
Instead, just have eventfs have its own iterate_shared callback function
that will fill in the dent entries. This simplifies the code quite a bit.
Link: https://lore.kernel.org/linux-trace-kernel/20240104015435.682218477@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Ajay Kaher <akaher@vmware.com> Cc: Al Viro <viro@ZenIV.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The "lookup" parameter is a way to differentiate the call to
create_file/dir_dentry() from when it's just a lookup (no need to up the
dentry refcount) and accessed via a readdir (need to up the refcount).
But reality, it just makes the code more complex. Just up the refcount and
let the caller decide to dput() the result or not.
A flag was needed to denote which eventfs_inode was the "events"
directory, so a bit was taken from the "nr_entries" field, as there's not
that many entries, and 2^30 is plenty. But the bit number for nr_entries
was not updated to reflect the bit taken from it, which would add an
unnecessary integer to the structure.
If a getdents() is called on the tracefs directory but does not get all
the files, it can leave a "cursor" dentry in the d_subdirs list of tracefs
dentry. This cursor dentry does not have a d_inode for it. Before
referencing tracefs_inode from the dentry, the d_inode must first be
checked if it has content. If not, then it's not a tracefs_inode and can
be ignored.
That's because the 256 BUF_SIZE was not big enough to read all the
dentries of the tracefs file system and it left a "cursor" dentry in the
subdirs of the tracefs root inode. Then on remounting with "gid=1000",
it would cause an iteration of all dentries which hit:
ti = get_tracefs(dentry->d_inode);
if (ti && (ti->flags & TRACEFS_EVENT_INODE))
eventfs_update_gid(dentry, gid);
Which crashed because of the dereference of the cursor dentry which had a NULL
d_inode.
In the subdir loop of the dentry lookup of set_gid(), if a child has a
NULL d_inode, simply skip it.
It was reported that when mounting the tracefs file system with a gid
other than root, the ownership did not carry down to the eventfs directory
due to the dynamic nature of it.
A fix was done to solve this, but it had two issues.
(a) if the attr passed into update_inode_attr() was NULL, it didn't do
anything. This is true for files that have not had a chown or chgrp
done to itself or any of its sibling files, as the attr is allocated
for all children when any one needs it.
# ls -ld /mnt/events/sched
drwxr-xr-x 28 root rostedt 0 Dec 21 13:12 /mnt/events/sched/
# ls -ld /mnt/events/sched/sched_switch
drwxr-xr-x 2 root rostedt 0 Dec 21 13:12 /mnt/events/sched/sched_switch/
But when checking the files:
# ls -l /mnt/events/sched/sched_switch
total 0
-rw-r----- 1 root root 0 Dec 21 13:12 enable
-rw-r----- 1 root root 0 Dec 21 13:12 filter
-r--r----- 1 root root 0 Dec 21 13:12 format
-r--r----- 1 root root 0 Dec 21 13:12 hist
-r--r----- 1 root root 0 Dec 21 13:12 id
-rw-r----- 1 root root 0 Dec 21 13:12 trigger
(b) When the attr does not denote the UID or GID, it defaulted to using
the parent uid or gid. This is incorrect as changing the parent
uid or gid will automatically change all its children.
# chgrp tracing /mnt/events/timer
# ls -ld /mnt/events/timer
drwxr-xr-x 2 root tracing 0 Dec 21 14:34 /mnt/events/timer
# ls -l /mnt/events/timer
total 0
-rw-r----- 1 root root 0 Dec 21 14:35 enable
-rw-r----- 1 root root 0 Dec 21 14:35 filter
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 hrtimer_cancel
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 hrtimer_expire_entry
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 hrtimer_expire_exit
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 hrtimer_init
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 hrtimer_start
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 itimer_expire
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 itimer_state
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 tick_stop
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 timer_cancel
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 timer_expire_entry
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 timer_expire_exit
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 timer_init
drwxr-xr-x 2 root tracing 0 Dec 21 14:35 timer_start
At first it was thought that this could be easily fixed by just making the
default ownership of the superblock when it was mounted. But this does not
handle the case of:
# chgrp tracing instances
# mkdir instances/foo
If the superblock was used, then the group ownership would be that of what
it was when it was mounted, when it should instead be "tracing".
Instead, set a flag for the top level eventfs directory ("events") to flag
which eventfs_inode belongs to it.
Since the "events" directory's dentry and inode are never freed, it does
not need to use its attr field to restore its mode and ownership. Use the
this eventfs_inode's attr as the default ownership for all the files and
directories underneath it.
When the events eventfs_inode is created, it sets its ownership to its
parent uid and gid. As the events directory is created at boot up before
it gets mounted, this will always be uid=0 and gid=0. If it's created via
an instance, then it will take the ownership of the instance directory.
When the file system is mounted, it will update all the gids if one is
specified. This will have a callback to update the events evenfs_inode's
default entries.
When a file or directory is created under the events directory, it will
walk the ei->dentry parents until it finds the evenfs_inode that belongs
to the events directory to retrieve the default uid and gid values.
Link: https://lore.kernel.org/all/CAHk-=wiwQtUHvzwyZucDq8=Gtw+AnwScyLhpFswrQ84PjhoGsg@mail.gmail.com/ Link: https://lore.kernel.org/linux-trace-kernel/20231221190757.7eddbca9@gandalf.local.home Cc: stable@vger.kernel.org Cc: Mark Rutland <mark.rutland@arm.com> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Cc: Dongliang Cui <cuidongliang390@gmail.com> Cc: Hongyu Jin <hongyu.jin@unisoc.com> Fixes: 0dfc852b6fe3 ("eventfs: Have event files and directories default to parent uid and gid") Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Tested-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Reported-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
I found that in the latest version, the nodes of tracefs have been
changed to dynamically created.
This has caused me to encounter a problem where the gid I specified in
the mounting parameters cannot apply to all files, as in the following
situation:
/data/tmp/events # mount | grep tracefs
tracefs on /data/tmp type tracefs (rw,seclabel,relatime,gid=3012)
Eventfs uses simple_lookup(), however, it will fail if the name of the
entry is beyond NAME_MAX length. When this error is encountered, eventfs
still tries to create dentries instead of skipping the dentry creation.
When the dentry is attempted to be created in this state d_wait_lookup()
will loop forever, waiting for the lookup to be removed.
Fix eventfs to return the error in simple_lookup() back to the caller
instead of continuing to try to create the dentry.
Since the locking of the parent->d_inode has been moved outside the
creation of the files and directories (as it use to be locked via a
conditional), add a WARN_ON_ONCE() to the case that it's not locked.
Link: https://lkml.kernel.org/r/20231121231112.853962542@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs directory is dynamically created via the meta data supplied by
the existing trace events. All files and directories in eventfs has a
parent. Do not allow NULL to be passed into eventfs_start_creating() as
the parent because that should never happen. Warn if it does.
Link: https://lkml.kernel.org/r/20231121231112.693841807@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The both create_file_dentry() and create_dir_dentry() takes a boolean
parameter "lookup", as on lookup the inode_lock should already be taken,
but for dcache_dir_open_wrapper() it is not taken.
There's no reason that the dcache_dir_open_wrapper() can't take the
inode_lock before calling these functions. In fact, it's better if it
does, as the lock can be held throughout both directory and file
creations.
This also simplifies the code, and possibly prevents unexpected race
conditions when the lock is released.
Link: https://lkml.kernel.org/r/20231121231112.528544825@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode") Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If memory reclaim happens, it can reclaim file system pages. The file
system pages from eventfs may take the eventfs_mutex on reclaim. This
means that allocation while holding the eventfs_mutex must not call into
filesystem reclaim. A lockdep splat uncovered this.
Link: https://lkml.kernel.org/r/20231121231112.373501894@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Fixes: 28e12c09f5aa0 ("eventfs: Save ownership and mode") Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode") Reported-by: Mark Rutland <mark.rutland@arm.com> Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
With the call to simple_recursive_removal() on the entire eventfs sub
system when the directory is removed, it performs the d_invalidate on all
the dentries when it is removed. There's no need to do clean ups when a
dentry is being created while the directory is being deleted.
As dentries are cleaned up by the simpler_recursive_removal(), trying to
do d_invalidate() in these functions will cause the dentry to be
invalidated twice, and crash the kernel.
Link: https://lore.kernel.org/all/20231116123016.140576-1-naresh.kamboju@linaro.org/ Link: https://lkml.kernel.org/r/20231120235154.422970988@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Fixes: 407c6726ca71 ("eventfs: Use simple_recursive_removal() to clean up dentries") Reported-by: Mark Rutland <mark.rutland@arm.com> Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The logic to free the eventfs_inode (ei) use to set is_freed and clear the
"dentry" field under the eventfs_mutex. But that changed when a race was
found where the ei->dentry needed to be cleared when the last dput() was
called on it. But there was still logic that checked if ei->dentry was not
NULL and is_freed is set, and would warn if it was.
But since that situation was changed and the ei->dentry isn't cleared
until the last dput() is called on it while the ei->is_freed is set, do
not test for that condition anymore, and change the comments to reflect
that.
Link: https://lkml.kernel.org/r/20231120235154.265826243@goodmis.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Fixes: 020010fbfa20 ("eventfs: Delete eventfs_inode when the last dentry is freed") Reported-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Looking at how dentry is removed via the tracefs system, I found that
eventfs does not do everything that it did under tracefs. The tracefs
removal of a dentry calls simple_recursive_removal() that does a lot more
than a simple d_invalidate().
As it should be a requirement that any eventfs_inode that has a dentry, so
does its parent. When removing a eventfs_inode, if it has a dentry, a call
to simple_recursive_removal() on that dentry should clean up all the
dentries underneath it.
Add WARN_ON_ONCE() to check for the parent having a dentry if any children
do.
Link: https://lore.kernel.org/all/20231101022553.GE1957730@ZenIV/ Link: https://lkml.kernel.org/r/20231101172650.552471568@goodmis.org Cc: stable@vger.kernel.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Fixes: 5bdcd5f5331a2 ("eventfs: Implement removal of meta data from eventfs") Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The top level events directory is no longer special with regards to how it
should be delete. Remove the extra processing for it in
eventfs_set_ei_status_free().
Link: https://lkml.kernel.org/r/20231101172650.340876747@goodmis.org Cc: Ajay Kaher <akaher@vmware.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
There exists a race between holding a reference of an eventfs_inode dentry
and the freeing of the eventfs_inode. If user space has a dentry held long
enough, it may still be able to access the dentry's eventfs_inode after it
has been freed.
To prevent this, have he eventfs_inode freed via the last dput() (or via
RCU if the eventfs_inode does not have a dentry).
This means reintroducing the eventfs_inode del_list field at a temporary
place to put the eventfs_inode. It needs to mark it as freed (via the
list) but also must invalidate the dentry immediately as the return from
eventfs_remove_dir() expects that they are. But the dentry invalidation
must not be called under the eventfs_mutex, so it must be done after the
eventfs_inode is marked as free (put on a deletion list).
Link: https://lkml.kernel.org/r/20231101172650.123479767@goodmis.org Cc: stable@vger.kernel.org Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Ajay Kaher <akaher@vmware.com> Fixes: 5bdcd5f5331a2 ("eventfs: Implement removal of meta data from eventfs") Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The callback function that is used to create inodes and dentries is not
protected by anything and the data that is passed to it could become
stale. After eventfs_remove_dir() is called by the tracing system, it is
free to remove the events that are associated to that directory.
Unfortunately, that means the callbacks must not be called after that.
for (...) {
entry = &ei->entries[i];
r = entry->callback() {
call = data; // call == event_call above
if (call->flags ...)
[ USE AFTER FREE BUG ]
The safest way to protect this is to wrap the callback with:
mutex_lock(&eventfs_mutex);
if (!ei->is_freed)
r = entry->callback();
else
r = -1;
mutex_unlock(&eventfs_mutex);
This will make sure that the callback will not be called after it is
freed. But now it needs to be known that the callback is called while
holding internal eventfs locks, and that it must not call back into the
eventfs / tracefs system. There's no reason it should anyway, but document
that as well.
Link: https://lore.kernel.org/all/CA+G9fYu9GOEbD=rR5eMR-=HJ8H6rMsbzDC2ZY5=Y50WpWAE7_Q@mail.gmail.com/ Link: https://lkml.kernel.org/r/20231101172649.906696613@goodmis.org Cc: Ajay Kaher <akaher@vmware.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode") Reported-by: Linux Kernel Functional Testing <lkft@linaro.org> Reported-by: Naresh Kamboju <naresh.kamboju@linaro.org> Tested-by: Linux Kernel Functional Testing <lkft@linaro.org> Tested-by: Naresh Kamboju <naresh.kamboju@linaro.org> Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Now that inodes and dentries are created on the fly, they are also
reclaimed on memory pressure. Since the ownership and file mode are saved
in the inode, if they are freed, any changes to the ownership and mode
will be lost.
To counter this, if the user changes the permissions or ownership, save
them, and when creating the inodes again, restore those changes.
Link: https://lkml.kernel.org/r/20231101172649.691841445@goodmis.org Cc: stable@vger.kernel.org Cc: Ajay Kaher <akaher@vmware.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Fixes: 63940449555e7 ("eventfs: Implement eventfs lookup, read, open functions") Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs_inode (ei) is protected by SRCU, but the ei->dentry is not. It
is protected by the eventfs_mutex. Anytime the eventfs_mutex is released,
and access to the ei->dentry needs to be done, it should first check if
ei->is_freed is set under the eventfs_mutex. If it is, then the ei->dentry
is invalid and must not be used. The ei->dentry must only be accessed
under the eventfs_mutex and after checking if ei->is_freed is set.
When the ei is being freed, it will (under the eventfs_mutex) set is_freed
and at the same time move the dentry to a free list to be cleared after
the eventfs_mutex is released. This means that any access to the
ei->dentry must check first if ei->is_freed is set, because if it is, then
the dentry is on its way to be freed.
As the eventfs_inode is freed in two different locations, make a helper
function free_ei() to make sure all the allocated fields of the
eventfs_inode is freed.
This requires renaming the existing free_ei() which is called by the srcu
handler to free_rcu_ei() and have free_ei() just do the freeing, where
free_rcu_ei() will call it.
Link: https://lkml.kernel.org/r/20231101172649.265214087@goodmis.org Cc: Ajay Kaher <akaher@vmware.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs_inode->is_freed was a union with the rcu_head with the
assumption that when it was on the srcu list the head would contain a
pointer which would make "is_freed" true. But that was a wrong assumption
as the rcu head is a single link list where the last element is NULL.
Instead, split the nr_entries integer so that "is_freed" is one bit and
the nr_entries is the next 31 bits. As there shouldn't be more than 10
(currently there's at most 5 to 7 depending on the config), this should
not be a problem.
Link: https://lkml.kernel.org/r/20231101172649.049758712@goodmis.org Cc: stable@vger.kernel.org Cc: Mark Rutland <mark.rutland@arm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Ajay Kaher <akaher@vmware.com> Fixes: 63940449555e7 ("eventfs: Implement eventfs lookup, read, open functions") Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs_remove_rec() had some missing parameters in the kerneldoc
comment above it. Also, rephrase the description a bit more to have a bit
more correct grammar.
Link: https://lore.kernel.org/linux-trace-kernel/20231030121523.0b2225a7@gandalf.local.home Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode"); Reported-by: kernel test robot <lkp@intel.com> Closes: https://lore.kernel.org/oe-kbuild-all/202310052216.4SgqasWo-lkp@intel.com/ Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The creation of the top events directory does a dget() at the end of the
creation in eventfs_create_events_dir() with a comment saying the final
dput() will happen when it is removed. The problem is that a dget() is
already done on the dentry when it was created with tracefs_start_creating()!
The dget() now just causes a memory leak of that dentry.
Remove the extra dget() as the final dput() in the deletion of the events
directory actually matches the one in tracefs_start_creating().
Link: https://lore.kernel.org/linux-trace-kernel/20231031124229.4f2e3fa1@gandalf.local.home Cc: Masami Hiramatsu <mhiramat@kernel.org> Cc: Mark Rutland <mark.rutland@arm.com> Fixes: 5790b1fb3d672 ("eventfs: Remove eventfs_file and just use eventfs_inode") Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The failure path of allocating ei goes to a path that dereferences ei.
Add another label that skips over the ei dereferences to do the rest of
the clean up.
The update to removing the eventfs_file changed the way the events top
level directory was handled. Instead of returning a dentry, it now returns
the eventfs_inode. In this changed, the removing of the events top level
directory is not much different than removing any of the other
directories. Because of this, the removal just called eventfs_remove_dir()
instead of eventfs_remove_events_dir().
Although eventfs_remove_dir() does the clean up, it misses out on the
dget() of the ei->dentry done in eventfs_create_events_dir(). It makes
more sense to match eventfs_create_events_dir() with a specific function
eventfs_remove_events_dir() and this specific function can then perform
the dput() to the dentry that had the dget() when it was created.
Fixes: 5790b1fb3d67 ("eventfs: Remove eventfs_file and just use eventfs_inode") Reported-by: kernel test robot <lkp@intel.com> Closes: https://lore.kernel.org/oe-kbuild-all/202310051743.y9EobbUr-lkp@intel.com/ Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Instead of having a descriptor for every file represented in the eventfs
directory, only have the directory itself represented. Change the API to
send in a list of entries that represent all the files in the directory
(but not other directories). The entry list contains a name and a callback
function that will be used to create the files when they are accessed.
Where the name is the name of the file and the callback gets called when
the file is being created. The callback passes in the name (in case the
same callback is used for multiple files), a pointer to the mode, data and
fops. The data will be pointing to the data that was passed in
eventfs_create_dir() or eventfs_create_events_dir() but may be overridden
to point to something else, as it will be used to point to the
inode->i_private that is created. The information passed back from the
callback is used to create the dentry/inode.
If the callback fills the data and the file should be created, it must
return a positive number. On zero or negative, the file is ignored.
This logic may also be used as a prototype to convert entire pseudo file
systems into just-in-time allocation.
The "show_events_dentry" file has been updated to show the directories,
and any files they have.
The eventfs was not designed properly and may have some hidden bugs in it.
Linus rewrote it properly and I trust his version more than this one. Revert
the backported patches for 6.6 and re-apply all the changes to make it
equivalent to Linus's version.
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs was not designed properly and may have some hidden bugs in it.
Linus rewrote it properly and I trust his version more than this one. Revert
the backported patches for 6.6 and re-apply all the changes to make it
equivalent to Linus's version.
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs was not designed properly and may have some hidden bugs in it.
Linus rewrote it properly and I trust his version more than this one. Revert
the backported patches for 6.6 and re-apply all the changes to make it
equivalent to Linus's version.
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs was not designed properly and may have some hidden bugs in it.
Linus rewrote it properly and I trust his version more than this one. Revert
the backported patches for 6.6 and re-apply all the changes to make it
equivalent to Linus's version.
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs was not designed properly and may have some hidden bugs in it.
Linus rewrote it properly and I trust his version more than this one. Revert
the backported patches for 6.6 and re-apply all the changes to make it
equivalent to Linus's version.
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The eventfs was not designed properly and may have some hidden bugs in it.
Linus rewrote it properly and I trust his version more than this one. Revert
the backported patches for 6.6 and re-apply all the changes to make it
equivalent to Linus's version.
Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The current exception handler implementation, which assists when accessing
user space memory, may exhibit random data corruption if the compiler decides
to use a different register than the specified register %r29 (defined in
ASM_EXCEPTIONTABLE_REG) for the error code. If the compiler choose another
register, the fault handler will nevertheless store -EFAULT into %r29 and thus
trash whatever this register is used for.
Looking at the assembly I found that this happens sometimes in emulate_ldd().
To solve the issue, the easiest solution would be if it somehow is
possible to tell the fault handler which register is used to hold the error
code. Using %0 or %1 in the inline assembly is not posssible as it will show
up as e.g. %r29 (with the "%r" prefix), which the GNU assembler can not
convert to an integer.
This patch takes another, better and more flexible approach:
We extend the __ex_table (which is out of the execution path) by one 32-word.
In this word we tell the compiler to insert the assembler instruction
"or %r0,%r0,%reg", where %reg references the register which the compiler
choosed for the error return code.
In case of an access failure, the fault handler finds the __ex_table entry and
can examine the opcode. The used register is encoded in the lowest 5 bits, and
the fault handler can then store -EFAULT into this register.
Since we extend the __ex_table to 3 words we can't use the BUILDTIME_TABLE_SORT
config option any longer.
The patch fdb8e12cc2cc ("netfilter: ipset: fix performance regression
in swap operation") missed to add the calls to gc cancellations
at the error path of create operations and at module unload. Also,
because the half of the destroy operations now executed by a
function registered by call_rcu(), neither NFNL_SUBSYS_IPSET mutex
or rcu read lock is held and therefore the checking of them results
false warnings.
The patch "netfilter: ipset: fix race condition between swap/destroy
and kernel side add/del/test", commit 28628fa9 fixes a race condition.
But the synchronize_rcu() added to the swap function unnecessarily slows
it down: it can safely be moved to destroy and use call_rcu() instead.
Eric Dumazet pointed out that simply calling the destroy functions as
rcu callback does not work: sets with timeout use garbage collectors
which need cancelling at destroy which can wait. Therefore the destroy
functions are split into two: cancelling garbage collectors safely at
executing the command received by netlink and moving the remaining
part only into the rcu callback.
Link: https://lore.kernel.org/lkml/C0829B10-EAA6-4809-874E-E1E9C05A8D84@automattic.com/ Fixes: 28628fa952fe ("netfilter: ipset: fix race condition between swap/destroy and kernel side add/del/test") Reported-by: Ale Crismani <ale.crismani@automattic.com> Reported-by: David Wang <00107082@163.com> Tested-by: David Wang <00107082@163.com> Signed-off-by: Jozsef Kadlecsik <kadlec@netfilter.org> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Partial completions of zone append request is not allowed but if a zone
append completion indicates a number of completed bytes different from
the original BIO size, only the BIO status is set to error. This leads
to bio_advance() not setting the BIO size to 0 and thus to not call
bio_endio() at the end of req_bio_endio().
Make sure a partially completed zone append is failed and completed
immediately by forcing the completed number of bytes (nbytes) to be
equal to the BIO size, thus ensuring that bio_endio() is called.
Fixes: 297db731847e ("block: fix req_bio_endio append error handling") Cc: stable@kernel.vger.org Signed-off-by: Damien Le Moal <dlemoal@kernel.org> Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> Reviewed-by: Hannes Reinecke <hare@suse.de> Link: https://lore.kernel.org/r/20240110092942.442334-1-dlemoal@kernel.org Signed-off-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 5e2cf333b7bd ("md/raid5: Wait for MD_SB_CHANGE_PENDING in raid5d")
introduced a hung bug and will be reverted in next patch, since the issue
that commit is fixing is due to md superblock write is throttled by wbt,
to fix it, we can have superblock write bypass block layer throttle.
Fixes: 5e2cf333b7bd ("md/raid5: Wait for MD_SB_CHANGE_PENDING in raid5d") Cc: stable@vger.kernel.org # v5.19+ Suggested-by: Yu Kuai <yukuai3@huawei.com> Signed-off-by: Junxiao Bi <junxiao.bi@oracle.com> Reviewed-by: Logan Gunthorpe <logang@deltatee.com> Reviewed-by: Yu Kuai <yukuai3@huawei.com> Signed-off-by: Song Liu <song@kernel.org> Link: https://lore.kernel.org/r/20231108182216.73611-1-junxiao.bi@oracle.com Signed-off-by: Sasha Levin <sashal@kernel.org>
Where that s->saved_cmdlines allocation looks to be a dangling allocation
to kmemleak. That's because kmemleak only keeps track of kmalloc()
allocations. For allocations that use page_alloc() directly, the kmemleak
needs to be explicitly informed about it.
Add kmemleak_alloc() and kmemleak_free() around the page allocation so
that it doesn't give the following false positive:
Commit a8b9cf62ade1 ("ftrace: Fix DIRECT_CALLS to use SAVE_REGS by
default") attempted to fix an issue with direct trampolines on x86, see
its description for details. However, it wrongly referenced the
HAVE_DYNAMIC_FTRACE_WITH_REGS config option and the problem is still
present.
Add the missing "CONFIG_" prefix for the logic to work as intended.
Link: https://lore.kernel.org/linux-trace-kernel/20240213132434.22537-1-petr.pavlu@suse.com Fixes: a8b9cf62ade1 ("ftrace: Fix DIRECT_CALLS to use SAVE_REGS by default") Signed-off-by: Petr Pavlu <petr.pavlu@suse.com> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Patch series "fs/proc: do_task_stat: use sig->stats_".
do_task_stat() has the same problem as getrusage() had before "getrusage:
use sig->stats_lock rather than lock_task_sighand()": a hard lockup. If
NR_CPUS threads call lock_task_sighand() at the same time and the process
has NR_THREADS, spin_lock_irq will spin with irqs disabled O(NR_CPUS *
NR_THREADS) time.
This patch (of 3):
thread_group_cputime() does its own locking, we can safely shift
thread_group_cputime_adjusted() which does another for_each_thread loop
outside of ->siglock protected section.
Not only this removes for_each_thread() from the critical section with
irqs disabled, this removes another case when stats_lock is taken with
siglock held. We want to remove this dependency, then we can change the
users of stats_lock to not disable irqs.
The unused clock cleanup uses the _sync initcall to give all users at
earlier initcalls time to probe. Do the same to avoid leaving some PDs
dangling at "on" (which actually happened on qcom!).
Lock jsk->sk to prevent UAF when setsockopt(..., SO_J1939_FILTER, ...)
modifies jsk->filters while receiving packets.
Following trace was seen on affected system:
==================================================================
BUG: KASAN: slab-use-after-free in j1939_sk_recv_match_one+0x1af/0x2d0 [can_j1939]
Read of size 4 at addr ffff888012144014 by task j1939/350
A reasonable fix is to change j1939_socks_lock to an rwlock, since in
the rare situations where a write lock is required for the linked list
that j1939_socks_lock is protecting, the code does not attempt to
acquire any more locks. This would break the circular lock dependency,
where, for example, the current thread already locks j1939_socks_lock
and attempts to acquire sk_session_queue_lock, and at the same time,
another thread attempts to acquire j1939_socks_lock while holding
sk_session_queue_lock.
NOTE: This patch along does not fix the unregister_netdevice bug
reported by Syzbot; instead, it solves a deadlock situation to prepare
for one or more further patches to actually fix the Syzbot bug, which
appears to be a reference counting problem within the j1939 codebase.
The TDCO calculation was done using the currently applied data bittiming,
instead of the newly computed data bittiming, which means that the TDCO
had an invalid value unless setting the same data bittiming twice.
Fixes: d99755f71a80 ("can: netlink: add interface for CAN-FD Transmitter Delay Compensation (TDC)") Signed-off-by: Maxime Jayat <maxime.jayat@mobile-devices.fr> Reviewed-by: Vincent Mailhol <mailhol.vincent@wanadoo.fr> Link: https://lore.kernel.org/all/40579c18-63c0-43a4-8d4c-f3a6c1c0b417@munic.io Cc: stable@vger.kernel.org Signed-off-by: Marc Kleine-Budde <mkl@pengutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
If the directory passed to the '.. kernel-feat::' directive does not
exist or the get_feat.pl script does not find any files to extract
features from, Sphinx will report the following error:
Sphinx parallel build error:
UnboundLocalError: local variable 'fname' referenced before assignment
make[2]: *** [Documentation/Makefile:102: htmldocs] Error 2
This is due to how I changed the script in c48a7c44a1d0 ("docs:
kernel_feat.py: fix potential command injection"). Before that, the
filename passed along to self.nestedParse() in this case was weirdly
just the whole get_feat.pl invocation.
We can fix it by doing what kernel_abi.py does -- just pass
self.arguments[0] as 'fname'.
The detection of dirty-throttled tasks in blk-wbt has been subtly broken
since its beginning in 2016. Namely if we are doing cgroup writeback and
the throttled task is not in the root cgroup, balance_dirty_pages() will
set dirty_sleep for the non-root bdi_writeback structure. However
blk-wbt checks dirty_sleep only in the root cgroup bdi_writeback
structure. Thus detection of recently throttled tasks is not working in
this case (we noticed this when we switched to cgroup v2 and suddently
writeback was slow).
Since blk-wbt has no easy way to get to proper bdi_writeback and
furthermore its intention has always been to work on the whole device
rather than on individual cgroups, just move the dirty_sleep timestamp
from bdi_writeback to backing_dev_info. That fixes the checking for
recently throttled task and saves memory for everybody as a bonus.
CC: stable@vger.kernel.org Fixes: b57d74aff9ab ("writeback: track if we're sleeping on progress in balance_dirty_pages()") Signed-off-by: Jan Kara <jack@suse.cz> Link: https://lore.kernel.org/r/20240123175826.21452-1-jack@suse.cz
[axboe: fixup indentation errors] Signed-off-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The earlycon parameter is based on fixmap, and fixmap addresses are not
supposed to be shadowed by KASAN. So return the kasan_early_shadow_page
in kasan_mem_to_shadow() if the input address is above FIXADDR_START.
Otherwise earlycon cannot work after kasan_init().
For shared memory of type SHM_HUGETLB, hugetlb pages are reserved in
shmget() call. If SHM_NORESERVE flags is specified then the hugetlb pages
are not reserved. However when the shared memory is attached with the
shmat() call the hugetlb pages are getting reserved incorrectly for
SHM_HUGETLB shared memory created with SHM_NORESERVE which is a bug.
-------------------------------
Following test shows the issue.
$cat shmhtb.c
int main()
{
int shmflags = 0660 | IPC_CREAT | SHM_HUGETLB | SHM_NORESERVE;
int shmid;
projects / linux-stable / log