| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
It turns out that udev under certain circumstances will concurrently try
to load the same modules over-and-over excessively. This isn't a kernel
bug, but it ends up affecting the kernel, to the point that under
certain circumstances we can fail to boot, because the kernel uses a lot
of memory to read all the module data all at once.
Note that it isn't a memory leak, it's just basically a thundering herd
problem happening at bootup with a lot of CPUs, with the worst cases
then being pretty bad.
Admittedly the worst situations are somewhat contrived: lots and lots of
CPUs, not a lot of memory, and KASAN enabled to make it all slower and
as such (unintentionally) exacerbate the problem.
Luis explains: [1]
"My best assessment of the situation is that each CPU in udev ends up
triggering a load of duplicate set of modules, not just one, but *a
lot*. Not sure what heuristics udev uses to load a set of modules per
CPU."
Petr Pavlu chimes in: [2]
"My understanding is that udev workers are forked. An initial kmod
context is created by the main udevd process but no sharing happens
after the fork. It means that the mentioned memory pool logic doesn't
really kick in.
Multiple parallel load requests come from multiple udev workers, for
instance, each handling an udev event for one CPU device and making
the exactly same requests as all others are doing at the same time.
The optimization idea would be to recognize these duplicate requests
at the udevd/kmod level and converge them"
Note that module loading has tried to mitigate this issue before, see
for example commit 064f4536d139 ("module: avoid allocation if module is
already present and ready"), which has a few ASCII graphs on memory use
due to this same issue.
However, while that noticed that the module was already loaded, and
exited with an error early before spending any more time on setting up
the module, it didn't handle the case of multiple concurrent module
loads all being active - but not complete - at the same time.
Yes, one of them will eventually win the race and finalize its copy, and
the others will then notice that the module already exists and error
out, but while this all happens, we have tons of unnecessary concurrent
work being done.
Again, the real fix is for udev to not do that (maybe it should use
threads instead of fork, and have actual shared data structures and not
cause duplicate work). That real fix is apparently not trivial.
But it turns out that the kernel already has a pretty good model for
dealing with concurrent access to the same file: the i_writecount of the
inode.
In fact, the module loading already indirectly uses 'i_writecount' ,
because 'kernel_file_read()' will in fact do
ret = deny_write_access(file);
if (ret)
return ret;
...
allow_write_access(file);
around the read of the file data. We do not allow concurrent writes to
the file, and return -ETXTBUSY if the file was open for writing at the
same time as the module data is loaded from it.
And the solution to the reader concurrency problem is to simply extend
this "no concurrent writers" logic to simply be "exclusive access".
Note that "exclusive" in this context isn't really some absolute thing:
it's only exclusion from writers and from other "special readers" that
do this writer denial. So we simply introduce a variation of that
"deny_write_access()" logic that not only denies write access, but also
requires that this is the _only_ such access that denies write access.
Which means that you can't start loading a module that is already being
loaded as a module by somebody else, or you will get the same -ETXTBSY
error that you would get if there were writers around.
[ It also means that you can't try to load a currently executing
executable as a module, for the same reason: executables do that same
"deny_write_access()" thing, and that's obviously where the whole
ETXTBSY logic traditionally came from.
This is not a problem for kernel modules, since the set of normal
executable files and kernel module files is entirely disjoint. ]
This new function is called "exclusive_deny_write_access()", and the
implementation is trivial, in that it's just an atomic decrement of
i_writecount if it was 0 before.
To use that new exclusivity check, all we then do is wrap the module
loading with that exclusive_deny_write_access()() / allow_write_access()
pair. The actual patch is a bit bigger than that, because we want to
surround not just the "load file data" part, but the whole module setup,
to get maximum exclusion.
So this ends up splitting up "finit_module()" into a few helper
functions to make it all very clear and legible.
In Luis' test-case (bringing up 255 vcpu's in a virtual machine [3]),
the "wasted vmalloc" space (ie module data read into a vmalloc'ed area
in order to be loaded as a module, but then discarded because somebody
else loaded the same module instead) dropped from 1.8GiB to 474kB. Yes,
that's gigabytes to kilobytes.
It doesn't drop completely to zero, because even with this change, you
can still end up having completely serial pointless module loads, where
one udev process has loaded a module fully (and thus the kernel has
released that exclusive lock on the module file), and then another udev
process tries to load the same module again.
So while we cannot fully get rid of the fundamental bug in user space,
we _can_ get rid of the excessive concurrent thundering herd effect.
A couple of final side notes on this all:
- This tweak only affects the "finit_module()" system call, which gives
the kernel a file descriptor with the module data.
You can also just feed the module data as raw data from user space
with "init_module()" (note the lack of 'f' at the beginning), and
obviously for that case we do _not_ have any "exclusive read" logic.
So if you absolutely want to do things wrong in user space, and try
to load the same module multiple times, and error out only later when
the kernel ends up saying "you can't load the same module name
twice", you can still do that.
And in fact, some distros will do exactly that, because they will
uncompress the kernel module data in user space before feeding it to
the kernel (mainly because they haven't started using the new kernel
side decompression yet).
So this is not some absolute "you can't do concurrent loads of the
same module". It's literally just a very simple heuristic that will
catch it early in case you try to load the exact same module file at
the same time, and in that case avoid a potentially nasty situation.
- There is another user of "deny_write_access()": the verity code that
enables fs-verity on a file (the FS_IOC_ENABLE_VERITY ioctl).
If you use fs-verity and you care about verifying the kernel modules
(which does make sense), you should do it *before* loading said
kernel module. That may sound obvious, but now the implementation
basically requires it. Because if you try to do it concurrently, the
kernel may refuse to load the module file that is being set up by the
fs-verity code.
- This all will obviously mean that if you insist on loading the same
module in parallel, only one module load will succeed, and the others
will return with an error.
That was true before too, but what is different is that the -ETXTBSY
error can be returned *before* the success case of another process
fully loading and instantiating the module.
Again, that might sound obvious, and it is indeed the whole point of
the whole change: we are much quicker to notice the whole "you're
already in the process of loading this module".
So it's very much intentional, but it does mean that if you just
spray the kernel with "finit_module()", and expect that the module is
immediately loaded afterwards without checking the return value, you
are doing something horribly horribly wrong.
I'd like to say that that would never happen, but the whole _reason_
for this commit is that udev is currently doing something horribly
horribly wrong, so ...
Link: https://lore.kernel.org/all/ZEGopJ8VAYnE7LQ2@bombadil.infradead.org/ [1]
Link: https://lore.kernel.org/all/23bd0ce6-ef78-1cd8-1f21-0e706a00424a@suse.com/ [2]
Link: https://lore.kernel.org/lkml/ZG%2Fa+nrt4%2FAAUi5z@bombadil.infradead.org/ [3]
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Petr Pavlu <petr.pavlu@suse.com>
Tested-by: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Smatch warns:
kernel/module/stats.c:394 read_file_mod_stats()
warn: passing freed memory 'buf'
We are passing 'buf' to simple_read_from_buffer() after freeing it.
Fix this by changing the order of 'simple_read_from_buffer' and 'kfree'.
Fixes: df3e764d8e5c ("module: add debug stats to help identify memory pressure")
Signed-off-by: Harshit Mogalapalli <harshit.m.mogalapalli@oracle.com>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
Two newly introduced functions are declared in a header that is not
included before the definition, causing a warning with sparse or
'make W=1':
kernel/module/dups.c:118:6: error: no previous prototype for 'kmod_dup_request_exists_wait' [-Werror=missing-prototypes]
118 | bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret)
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~
kernel/module/dups.c:220:6: error: no previous prototype for 'kmod_dup_request_announce' [-Werror=missing-prototypes]
220 | void kmod_dup_request_announce(char *module_name, int ret)
| ^~~~~~~~~~~~~~~~~~~~~~~~~
Add an explicit include to ensure the prototypes match.
Fixes: 8660484ed1cf ("module: add debugging auto-load duplicate module support")
Reported-by: kernel test robot <lkp@intel.com>
Link: https://lore.kernel.org/oe-kbuild-all/202304141440.DYO4NAzp-lkp@intel.com/
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/mcgrof/linux
Pull module updates from Luis Chamberlain:
"The summary of the changes for this pull requests is:
- Song Liu's new struct module_memory replacement
- Nick Alcock's MODULE_LICENSE() removal for non-modules
- My cleanups and enhancements to reduce the areas where we vmalloc
module memory for duplicates, and the respective debug code which
proves the remaining vmalloc pressure comes from userspace.
Most of the changes have been in linux-next for quite some time except
the minor fixes I made to check if a module was already loaded prior
to allocating the final module memory with vmalloc and the respective
debug code it introduces to help clarify the issue. Although the
functional change is small it is rather safe as it can only *help*
reduce vmalloc space for duplicates and is confirmed to fix a bootup
issue with over 400 CPUs with KASAN enabled. I don't expect stable
kernels to pick up that fix as the cleanups would have also had to
have been picked up. Folks on larger CPU systems with modules will
want to just upgrade if vmalloc space has been an issue on bootup.
Given the size of this request, here's some more elaborate details:
The functional change change in this pull request is the very first
patch from Song Liu which replaces the 'struct module_layout' with a
new 'struct module_memory'. The old data structure tried to put
together all types of supported module memory types in one data
structure, the new one abstracts the differences in memory types in a
module to allow each one to provide their own set of details. This
paves the way in the future so we can deal with them in a cleaner way.
If you look at changes they also provide a nice cleanup of how we
handle these different memory areas in a module. This change has been
in linux-next since before the merge window opened for v6.3 so to
provide more than a full kernel cycle of testing. It's a good thing as
quite a bit of fixes have been found for it.
Jason Baron then made dynamic debug a first class citizen module user
by using module notifier callbacks to allocate / remove module
specific dynamic debug information.
Nick Alcock has done quite a bit of work cross-tree to remove module
license tags from things which cannot possibly be module at my request
so to:
a) help him with his longer term tooling goals which require a
deterministic evaluation if a piece a symbol code could ever be
part of a module or not. But quite recently it is has been made
clear that tooling is not the only one that would benefit.
Disambiguating symbols also helps efforts such as live patching,
kprobes and BPF, but for other reasons and R&D on this area is
active with no clear solution in sight.
b) help us inch closer to the now generally accepted long term goal
of automating all the MODULE_LICENSE() tags from SPDX license tags
In so far as a) is concerned, although module license tags are a no-op
for non-modules, tools which would want create a mapping of possible
modules can only rely on the module license tag after the commit
8b41fc4454e ("kbuild: create modules.builtin without
Makefile.modbuiltin or tristate.conf").
Nick has been working on this *for years* and AFAICT I was the only
one to suggest two alternatives to this approach for tooling. The
complexity in one of my suggested approaches lies in that we'd need a
possible-obj-m and a could-be-module which would check if the object
being built is part of any kconfig build which could ever lead to it
being part of a module, and if so define a new define
-DPOSSIBLE_MODULE [0].
A more obvious yet theoretical approach I've suggested would be to
have a tristate in kconfig imply the same new -DPOSSIBLE_MODULE as
well but that means getting kconfig symbol names mapping to modules
always, and I don't think that's the case today. I am not aware of
Nick or anyone exploring either of these options. Quite recently Josh
Poimboeuf has pointed out that live patching, kprobes and BPF would
benefit from resolving some part of the disambiguation as well but for
other reasons. The function granularity KASLR (fgkaslr) patches were
mentioned but Joe Lawrence has clarified this effort has been dropped
with no clear solution in sight [1].
In the meantime removing module license tags from code which could
never be modules is welcomed for both objectives mentioned above. Some
developers have also welcomed these changes as it has helped clarify
when a module was never possible and they forgot to clean this up, and
so you'll see quite a bit of Nick's patches in other pull requests for
this merge window. I just picked up the stragglers after rc3. LWN has
good coverage on the motivation behind this work [2] and the typical
cross-tree issues he ran into along the way. The only concrete blocker
issue he ran into was that we should not remove the MODULE_LICENSE()
tags from files which have no SPDX tags yet, even if they can never be
modules. Nick ended up giving up on his efforts due to having to do
this vetting and backlash he ran into from folks who really did *not
understand* the core of the issue nor were providing any alternative /
guidance. I've gone through his changes and dropped the patches which
dropped the module license tags where an SPDX license tag was missing,
it only consisted of 11 drivers. To see if a pull request deals with a
file which lacks SPDX tags you can just use:
./scripts/spdxcheck.py -f \
$(git diff --name-only commid-id | xargs echo)
You'll see a core module file in this pull request for the above, but
that's not related to his changes. WE just need to add the SPDX
license tag for the kernel/module/kmod.c file in the future but it
demonstrates the effectiveness of the script.
Most of Nick's changes were spread out through different trees, and I
just picked up the slack after rc3 for the last kernel was out. Those
changes have been in linux-next for over two weeks.
The cleanups, debug code I added and final fix I added for modules
were motivated by David Hildenbrand's report of boot failing on a
systems with over 400 CPUs when KASAN was enabled due to running out
of virtual memory space. Although the functional change only consists
of 3 lines in the patch "module: avoid allocation if module is already
present and ready", proving that this was the best we can do on the
modules side took quite a bit of effort and new debug code.
The initial cleanups I did on the modules side of things has been in
linux-next since around rc3 of the last kernel, the actual final fix
for and debug code however have only been in linux-next for about a
week or so but I think it is worth getting that code in for this merge
window as it does help fix / prove / evaluate the issues reported with
larger number of CPUs. Userspace is not yet fixed as it is taking a
bit of time for folks to understand the crux of the issue and find a
proper resolution. Worst come to worst, I have a kludge-of-concept [3]
of how to make kernel_read*() calls for modules unique / converge
them, but I'm currently inclined to just see if userspace can fix this
instead"
Link: https://lore.kernel.org/all/Y/kXDqW+7d71C4wz@bombadil.infradead.org/ [0]
Link: https://lkml.kernel.org/r/025f2151-ce7c-5630-9b90-98742c97ac65@redhat.com [1]
Link: https://lwn.net/Articles/927569/ [2]
Link: https://lkml.kernel.org/r/20230414052840.1994456-3-mcgrof@kernel.org [3]
* tag 'modules-6.4-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/mcgrof/linux: (121 commits)
module: add debugging auto-load duplicate module support
module: stats: fix invalid_mod_bytes typo
module: remove use of uninitialized variable len
module: fix building stats for 32-bit targets
module: stats: include uapi/linux/module.h
module: avoid allocation if module is already present and ready
module: add debug stats to help identify memory pressure
module: extract patient module check into helper
modules/kmod: replace implementation with a semaphore
Change DEFINE_SEMAPHORE() to take a number argument
module: fix kmemleak annotations for non init ELF sections
module: Ignore L0 and rename is_arm_mapping_symbol()
module: Move is_arm_mapping_symbol() to module_symbol.h
module: Sync code of is_arm_mapping_symbol()
scripts/gdb: use mem instead of core_layout to get the module address
interconnect: remove module-related code
interconnect: remove MODULE_LICENSE in non-modules
zswap: remove MODULE_LICENSE in non-modules
zpool: remove MODULE_LICENSE in non-modules
x86/mm/dump_pagetables: remove MODULE_LICENSE in non-modules
...
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next
Pull networking updates from Paolo Abeni:
"Core:
- Introduce a config option to tweak MAX_SKB_FRAGS. Increasing the
default value allows for better BIG TCP performances
- Reduce compound page head access for zero-copy data transfers
- RPS/RFS improvements, avoiding unneeded NET_RX_SOFTIRQ when
possible
- Threaded NAPI improvements, adding defer skb free support and
unneeded softirq avoidance
- Address dst_entry reference count scalability issues, via false
sharing avoidance and optimize refcount tracking
- Add lockless accesses annotation to sk_err[_soft]
- Optimize again the skb struct layout
- Extends the skb drop reasons to make it usable by multiple
subsystems
- Better const qualifier awareness for socket casts
BPF:
- Add skb and XDP typed dynptrs which allow BPF programs for more
ergonomic and less brittle iteration through data and
variable-sized accesses
- Add a new BPF netfilter program type and minimal support to hook
BPF programs to netfilter hooks such as prerouting or forward
- Add more precise memory usage reporting for all BPF map types
- Adds support for using {FOU,GUE} encap with an ipip device
operating in collect_md mode and add a set of BPF kfuncs for
controlling encap params
- Allow BPF programs to detect at load time whether a particular
kfunc exists or not, and also add support for this in light
skeleton
- Bigger batch of BPF verifier improvements to prepare for upcoming
BPF open-coded iterators allowing for less restrictive looping
capabilities
- Rework RCU enforcement in the verifier, add kptr_rcu and enforce
BPF programs to NULL-check before passing such pointers into kfunc
- Add support for kptrs in percpu hashmaps, percpu LRU hashmaps and
in local storage maps
- Enable RCU semantics for task BPF kptrs and allow referenced kptr
tasks to be stored in BPF maps
- Add support for refcounted local kptrs to the verifier for allowing
shared ownership, useful for adding a node to both the BPF list and
rbtree
- Add BPF verifier support for ST instructions in
convert_ctx_access() which will help new -mcpu=v4 clang flag to
start emitting them
- Add ARM32 USDT support to libbpf
- Improve bpftool's visual program dump which produces the control
flow graph in a DOT format by adding C source inline annotations
Protocols:
- IPv4: Allow adding to IPv4 address a 'protocol' tag. Such value
indicates the provenance of the IP address
- IPv6: optimize route lookup, dropping unneeded R/W lock acquisition
- Add the handshake upcall mechanism, allowing the user-space to
implement generic TLS handshake on kernel's behalf
- Bridge: support per-{Port, VLAN} neighbor suppression, increasing
resilience to nodes failures
- SCTP: add support for Fair Capacity and Weighted Fair Queueing
schedulers
- MPTCP: delay first subflow allocation up to its first usage. This
will allow for later better LSM interaction
- xfrm: Remove inner/outer modes from input/output path. These are
not needed anymore
- WiFi:
- reduced neighbor report (RNR) handling for AP mode
- HW timestamping support
- support for randomized auth/deauth TA for PASN privacy
- per-link debugfs for multi-link
- TC offload support for mac80211 drivers
- mac80211 mesh fast-xmit and fast-rx support
- enable Wi-Fi 7 (EHT) mesh support
Netfilter:
- Add nf_tables 'brouting' support, to force a packet to be routed
instead of being bridged
- Update bridge netfilter and ovs conntrack helpers to handle IPv6
Jumbo packets properly, i.e. fetch the packet length from
hop-by-hop extension header. This is needed for BIT TCP support
- The iptables 32bit compat interface isn't compiled in by default
anymore
- Move ip(6)tables builtin icmp matches to the udptcp one. This has
the advantage that icmp/icmpv6 match doesn't load the
iptables/ip6tables modules anymore when iptables-nft is used
- Extended netlink error report for netdevice in flowtables and
netdev/chains. Allow for incrementally add/delete devices to netdev
basechain. Allow to create netdev chain without device
Driver API:
- Remove redundant Device Control Error Reporting Enable, as PCI core
has already error reporting enabled at enumeration time
- Move Multicast DB netlink handlers to core, allowing devices other
then bridge to use them
- Allow the page_pool to directly recycle the pages from safely
localized NAPI
- Implement lockless TX queue stop/wake combo macros, allowing for
further code de-duplication and sanitization
- Add YNL support for user headers and struct attrs
- Add partial YNL specification for devlink
- Add partial YNL specification for ethtool
- Add tc-mqprio and tc-taprio support for preemptible traffic classes
- Add tx push buf len param to ethtool, specifies the maximum number
of bytes of a transmitted packet a driver can push directly to the
underlying device
- Add basic LED support for switch/phy
- Add NAPI documentation, stop relaying on external links
- Convert dsa_master_ioctl() to netdev notifier. This is a
preparatory work to make the hardware timestamping layer selectable
by user space
- Add transceiver support and improve the error messages for CAN-FD
controllers
New hardware / drivers:
- Ethernet:
- AMD/Pensando core device support
- MediaTek MT7981 SoC
- MediaTek MT7988 SoC
- Broadcom BCM53134 embedded switch
- Texas Instruments CPSW9G ethernet switch
- Qualcomm EMAC3 DWMAC ethernet
- StarFive JH7110 SoC
- NXP CBTX ethernet PHY
- WiFi:
- Apple M1 Pro/Max devices
- RealTek rtl8710bu/rtl8188gu
- RealTek rtl8822bs, rtl8822cs and rtl8821cs SDIO chipset
- Bluetooth:
- Realtek RTL8821CS, RTL8851B, RTL8852BS
- Mediatek MT7663, MT7922
- NXP w8997
- Actions Semi ATS2851
- QTI WCN6855
- Marvell 88W8997
- Can:
- STMicroelectronics bxcan stm32f429
Drivers:
- Ethernet NICs:
- Intel (1G, icg):
- add tracking and reporting of QBV config errors
- add support for configuring max SDU for each Tx queue
- Intel (100G, ice):
- refactor mailbox overflow detection to support Scalable IOV
- GNSS interface optimization
- Intel (i40e):
- support XDP multi-buffer
- nVidia/Mellanox:
- add the support for linux bridge multicast offload
- enable TC offload for egress and engress MACVLAN over bond
- add support for VxLAN GBP encap/decap flows offload
- extend packet offload to fully support libreswan
- support tunnel mode in mlx5 IPsec packet offload
- extend XDP multi-buffer support
- support MACsec VLAN offload
- add support for dynamic msix vectors allocation
- drop RX page_cache and fully use page_pool
- implement thermal zone to report NIC temperature
- Netronome/Corigine:
- add support for multi-zone conntrack offload
- Solarflare/Xilinx:
- support offloading TC VLAN push/pop actions to the MAE
- support TC decap rules
- support unicast PTP
- Other NICs:
- Broadcom (bnxt): enforce software based freq adjustments only on
shared PHC NIC
- RealTek (r8169): refactor to addess ASPM issues during NAPI poll
- Micrel (lan8841): add support for PTP_PF_PEROUT
- Cadence (macb): enable PTP unicast
- Engleder (tsnep): add XDP socket zero-copy support
- virtio-net: implement exact header length guest feature
- veth: add page_pool support for page recycling
- vxlan: add MDB data path support
- gve: add XDP support for GQI-QPL format
- geneve: accept every ethertype
- macvlan: allow some packets to bypass broadcast queue
- mana: add support for jumbo frame
- Ethernet high-speed switches:
- Microchip (sparx5): Add support for TC flower templates
- Ethernet embedded switches:
- Broadcom (b54):
- configure 6318 and 63268 RGMII ports
- Marvell (mv88e6xxx):
- faster C45 bus scan
- Microchip:
- lan966x:
- add support for IS1 VCAP
- better TX/RX from/to CPU performances
- ksz9477: add ETS Qdisc support
- ksz8: enhance static MAC table operations and error handling
- sama7g5: add PTP capability
- NXP (ocelot):
- add support for external ports
- add support for preemptible traffic classes
- Texas Instruments:
- add CPSWxG SGMII support for J7200 and J721E
- Intel WiFi (iwlwifi):
- preparation for Wi-Fi 7 EHT and multi-link support
- EHT (Wi-Fi 7) sniffer support
- hardware timestamping support for some devices/firwmares
- TX beacon protection on newer hardware
- Qualcomm 802.11ax WiFi (ath11k):
- MU-MIMO parameters support
- ack signal support for management packets
- RealTek WiFi (rtw88):
- SDIO bus support
- better support for some SDIO devices (e.g. MAC address from
efuse)
- RealTek WiFi (rtw89):
- HW scan support for 8852b
- better support for 6 GHz scanning
- support for various newer firmware APIs
- framework firmware backwards compatibility
- MediaTek WiFi (mt76):
- P2P support
- mesh A-MSDU support
- EHT (Wi-Fi 7) support
- coredump support"
* tag 'net-next-6.4' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next: (2078 commits)
net: phy: hide the PHYLIB_LEDS knob
net: phy: marvell-88x2222: remove unnecessary (void*) conversions
tcp/udp: Fix memleaks of sk and zerocopy skbs with TX timestamp.
net: amd: Fix link leak when verifying config failed
net: phy: marvell: Fix inconsistent indenting in led_blink_set
lan966x: Don't use xdp_frame when action is XDP_TX
tsnep: Add XDP socket zero-copy TX support
tsnep: Add XDP socket zero-copy RX support
tsnep: Move skb receive action to separate function
tsnep: Add functions for queue enable/disable
tsnep: Rework TX/RX queue initialization
tsnep: Replace modulo operation with mask
net: phy: dp83867: Add led_brightness_set support
net: phy: Fix reading LED reg property
drivers: nfc: nfcsim: remove return value check of `dev_dir`
net: phy: dp83867: Remove unnecessary (void*) conversions
net: ethtool: coalesce: try to make user settings stick twice
net: mana: Check if netdev/napi_alloc_frag returns single page
net: mana: Rename mana_refill_rxoob and remove some empty lines
net: veth: add page_pool stats
...
|
|
The finit_module() system call can in the worst case use up to more than
twice of a module's size in virtual memory. Duplicate finit_module()
system calls are non fatal, however they unnecessarily strain virtual
memory during bootup and in the worst case can cause a system to fail
to boot. This is only known to currently be an issue on systems with
larger number of CPUs.
To help debug this situation we need to consider the different sources for
finit_module(). Requests from the kernel that rely on module auto-loading,
ie, the kernel's *request_module() API, are one source of calls. Although
modprobe checks to see if a module is already loaded prior to calling
finit_module() there is a small race possible allowing userspace to
trigger multiple modprobe calls racing against modprobe and this not
seeing the module yet loaded.
This adds debugging support to the kernel module auto-loader (*request_module()
calls) to easily detect duplicate module requests. To aid with possible bootup
failure issues incurred by this, it will converge duplicates requests to a
single request. This avoids any possible strain on virtual memory during
bootup which could be incurred by duplicate module autoloading requests.
Folks debugging virtual memory abuse on bootup can and should enable
this to see what pr_warn()s come on, to see if module auto-loading is to
blame for their wores. If they see duplicates they can further debug this
by enabling the module.enable_dups_trace kernel parameter or by enabling
CONFIG_MODULE_DEBUG_AUTOLOAD_DUPS_TRACE.
Current evidence seems to point to only a few duplicates for module
auto-loading. And so the source for other duplicates creating heavy
virtual memory pressure due to larger number of CPUs should becoming
from another place (likely udev).
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
This was caught by randconfig builds but does not show up in
build testing without CONFIG_MODULE_DECOMPRESS:
kernel/module/stats.c: In function 'mod_stat_bump_invalid':
kernel/module/stats.c:229:42: error: 'invalid_mod_byte' undeclared (first use in this function); did you mean 'invalid_mod_bytes'?
229 | atomic_long_add(info->compressed_len, &invalid_mod_byte);
| ^~~~~~~~~~~~~~~~
| invalid_mod_bytes
Fixes: df3e764d8e5c ("module: add debug stats to help identify memory pressure")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Randy Dunlap <rdunlap@infradead.org>
Tested-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
clang build reports
kernel/module/stats.c:307:34: error: variable
'len' is uninitialized when used here [-Werror,-Wuninitialized]
len = scnprintf(buf + 0, size - len,
^~~
At the start of this sequence, neither the '+ 0', nor the '- len' are needed.
So remove them and fix using 'len' uninitalized.
Fixes: df3e764d8e5c ("module: add debug stats to help identify memory pressure")
Signed-off-by: Tom Rix <trix@redhat.com>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
The new module statistics code mixes 64-bit types and wordsized 'long'
variables, which leads to build failures on 32-bit architectures:
kernel/module/stats.c: In function 'read_file_mod_stats':
kernel/module/stats.c:291:29: error: passing argument 1 of 'atomic64_read' from incompatible pointer type [-Werror=incompatible-pointer-types]
291 | total_size = atomic64_read(&total_mod_size);
x86_64-linux-ld: kernel/module/stats.o: in function `read_file_mod_stats':
stats.c:(.text+0x2b2): undefined reference to `__udivdi3'
To fix this, the code has to use one of the two types consistently.
Change them all to word-size types here.
Fixes: df3e764d8e5c ("module: add debug stats to help identify memory pressure")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
MODULE_INIT_COMPRESSED_FILE is defined in the uapi header, which
is not included indirectly from the normal linux/module.h, but
has to be pulled in explicitly:
kernel/module/stats.c: In function 'mod_stat_bump_invalid':
kernel/module/stats.c:227:14: error: 'MODULE_INIT_COMPRESSED_FILE' undeclared (first use in this function)
227 | if (flags & MODULE_INIT_COMPRESSED_FILE)
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
The finit_module() system call can create unnecessary virtual memory
pressure for duplicate modules. This is because load_module() can in
the worse case allocate more than twice the size of a module in virtual
memory. This saves at least a full size of the module in wasted vmalloc
space memory by trying to avoid duplicates as soon as we can validate
the module name in the read module structure.
This can only be an issue if a system is getting hammered with userspace
loading modules. There are two ways to load modules typically on systems,
one is the kernel moduile auto-loading (*request_module*() calls in-kernel)
and the other is things like udev. The auto-loading is in-kernel, but that
pings back to userspace to just call modprobe. We already have a way to
restrict the amount of concurrent kernel auto-loads in a given time, however
that still allows multiple requests for the same module to go through
and force two threads in userspace racing to call modprobe for the same
exact module. Even though libkmod which both modprobe and udev does check
if a module is already loaded prior calling finit_module() races are
still possible and this is clearly evident today when you have multiple
CPUs.
To avoid memory pressure for such stupid cases put a stop gap for them.
The *earliest* we can detect duplicates from the modules side of things
is once we have blessed the module name, sadly after the first vmalloc
allocation. We can check for the module being present *before* a secondary
vmalloc() allocation.
There is a linear relationship between wasted virtual memory bytes and
the number of CPU counts. The reason is that udev ends up racing to call
tons of the same modules for each of the CPUs.
We can see the different linear relationships between wasted virtual
memory and CPU count during after boot in the following graph:
+----------------------------------------------------------------------------+
14GB |-+ + + + + *+ +-|
| **** |
| *** |
| ** |
12GB |-+ ** +-|
| ** |
| ** |
| ** |
| ** |
10GB |-+ ** +-|
| ** |
| ** |
| ** |
8GB |-+ ** +-|
waste | ** ### |
| ** #### |
| ** ####### |
6GB |-+ **** #### +-|
| * #### |
| * #### |
| ***** #### |
4GB |-+ ** #### +-|
| ** #### |
| ** #### |
| ** #### |
2GB |-+ ** ##### +-|
| * #### |
| * #### Before ******* |
| **## + + + + After ####### |
+----------------------------------------------------------------------------+
0 50 100 150 200 250 300
CPUs count
On the y-axis we can see gigabytes of wasted virtual memory during boot
due to duplicate module requests which just end up failing. Trying to
infer the slope this ends up being about ~463 MiB per CPU lost prior
to this patch. After this patch we only loose about ~230 MiB per CPU, for
a total savings of about ~233 MiB per CPU. This is all *just on bootup*!
On a 8vcpu 8 GiB RAM system using kdevops and testing against selftests
kmod.sh -t 0008 I see a saving in the *highest* side of memory
consumption of up to ~ 84 MiB with the Linux kernel selftests kmod
test 0008. With the new stress-ng module test I see a 145 MiB difference
in max memory consumption with 100 ops. The stress-ng module ops tests can be
pretty pathalogical -- it is not realistic, however it was used to
finally successfully reproduce issues which are only reported to happen on
system with over 400 CPUs [0] by just usign 100 ops on a 8vcpu 8 GiB RAM
system. Running out of virtual memory space is no surprise given the
above graph, since at least on x86_64 we're capped at 128 MiB, eventually
we'd hit a series of errors and once can use the above graph to
guestimate when. This of course will vary depending on the features
you have enabled. So for instance, enabling KASAN seems to make this
much worse.
The results with kmod and stress-ng can be observed and visualized below.
The time it takes to run the test is also not affected.
The kmod tests 0008:
The gnuplot is set to a range from 400000 KiB (390 Mib) - 580000 (566 Mib)
given the tests peak around that range.
cat kmod.plot
set term dumb
set output fileout
set yrange [400000:580000]
plot filein with linespoints title "Memory usage (KiB)"
Before:
root@kmod ~ # /data/linux-next/tools/testing/selftests/kmod/kmod.sh -t 0008
root@kmod ~ # free -k -s 1 -c 40 | grep Mem | awk '{print $3}' > log-0008-before.txt ^C
root@kmod ~ # sort -n -r log-0008-before.txt | head -1
528732
So ~516.33 MiB
After:
root@kmod ~ # /data/linux-next/tools/testing/selftests/kmod/kmod.sh -t 0008
root@kmod ~ # free -k -s 1 -c 40 | grep Mem | awk '{print $3}' > log-0008-after.txt ^C
root@kmod ~ # sort -n -r log-0008-after.txt | head -1
442516
So ~432.14 MiB
That's about 84 ~MiB in savings in the worst case. The graphs:
root@kmod ~ # gnuplot -e "filein='log-0008-before.txt'; fileout='graph-0008-before.txt'" kmod.plot
root@kmod ~ # gnuplot -e "filein='log-0008-after.txt'; fileout='graph-0008-after.txt'" kmod.plot
root@kmod ~ # cat graph-0008-before.txt
580000 +-----------------------------------------------------------------+
| + + + + + + + |
560000 |-+ Memory usage (KiB) ***A***-|
| |
540000 |-+ +-|
| |
| *A *AA*AA*A*AA *A*AA A*A*A *AA*A*AA*A A |
520000 |-+A*A*AA *AA*A *A*AA*A*AA *A*A A *A+-|
|*A |
500000 |-+ +-|
| |
480000 |-+ +-|
| |
460000 |-+ +-|
| |
| |
440000 |-+ +-|
| |
420000 |-+ +-|
| + + + + + + + |
400000 +-----------------------------------------------------------------+
0 5 10 15 20 25 30 35 40
root@kmod ~ # cat graph-0008-after.txt
580000 +-----------------------------------------------------------------+
| + + + + + + + |
560000 |-+ Memory usage (KiB) ***A***-|
| |
540000 |-+ +-|
| |
| |
520000 |-+ +-|
| |
500000 |-+ +-|
| |
480000 |-+ +-|
| |
460000 |-+ +-|
| |
| *A *A*A |
440000 |-+A*A*AA*A A A*A*AA A*A*AA*A*AA*A*AA*A*AA*AA*A*AA*A*AA-|
|*A *A*AA*A |
420000 |-+ +-|
| + + + + + + + |
400000 +-----------------------------------------------------------------+
0 5 10 15 20 25 30 35 40
The stress-ng module tests:
This is used to run the test to try to reproduce the vmap issues
reported by David:
echo 0 > /proc/sys/vm/oom_dump_tasks
./stress-ng --module 100 --module-name xfs
Prior to this commit:
root@kmod ~ # free -k -s 1 -c 40 | grep Mem | awk '{print $3}' > baseline-stress-ng.txt
root@kmod ~ # sort -n -r baseline-stress-ng.txt | head -1
5046456
After this commit:
root@kmod ~ # free -k -s 1 -c 40 | grep Mem | awk '{print $3}' > after-stress-ng.txt
root@kmod ~ # sort -n -r after-stress-ng.txt | head -1
4896972
5046456 - 4896972
149484
149484/1024
145.98046875000000000000
So this commit using stress-ng reveals saving about 145 MiB in memory
using 100 ops from stress-ng which reproduced the vmap issue reported.
cat kmod.plot
set term dumb
set output fileout
set yrange [4700000:5070000]
plot filein with linespoints title "Memory usage (KiB)"
root@kmod ~ # gnuplot -e "filein='baseline-stress-ng.txt'; fileout='graph-stress-ng-before.txt'" kmod-simple-stress-ng.plot
root@kmod ~ # gnuplot -e "filein='after-stress-ng.txt'; fileout='graph-stress-ng-after.txt'" kmod-simple-stress-ng.plot
root@kmod ~ # cat graph-stress-ng-before.txt
+---------------------------------------------------------------+
5.05e+06 |-+ + A + + + + + + +-|
| * Memory usage (KiB) ***A*** |
| * A |
5e+06 |-+ ** ** +-|
| ** * * A |
4.95e+06 |-+ * * A * A* +-|
| * * A A * * * * A |
| * * * * * * *A * * * A * |
4.9e+06 |-+ * * * A*A * A*AA*A A *A **A **A*A *+-|
| A A*A A * A * * A A * A * ** |
| * ** ** * * * * * * * |
4.85e+06 |-+ A A A ** * * ** *-|
| * * * * ** * |
| * A * * * * |
4.8e+06 |-+ * * * A A-|
| * * * |
4.75e+06 |-+ * * * +-|
| * ** |
| * + + + + + + ** + |
4.7e+06 +---------------------------------------------------------------+
0 5 10 15 20 25 30 35 40
root@kmod ~ # cat graph-stress-ng-after.txt
+---------------------------------------------------------------+
5.05e+06 |-+ + + + + + + + +-|
| Memory usage (KiB) ***A*** |
| |
5e+06 |-+ +-|
| |
4.95e+06 |-+ +-|
| |
| |
4.9e+06 |-+ *AA +-|
| A*AA*A*A A A*AA*AA*A*AA*A A A A*A *AA*A*A A A*AA*AA |
| * * ** * * * ** * *** * |
4.85e+06 |-+* *** * * * * *** A * * +-|
| * A * * ** * * A * * |
| * * * * ** * * |
4.8e+06 |-+* * * A * * * +-|
| * * * A * * |
4.75e+06 |-* * * * * +-|
| * * * * * |
| * + * *+ + + + + * *+ |
4.7e+06 +---------------------------------------------------------------+
0 5 10 15 20 25 30 35 40
[0] https://lkml.kernel.org/r/20221013180518.217405-1-david@redhat.com
Reported-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
|
|
Loading modules with finit_module() can end up using vmalloc(), vmap()
and vmalloc() again, for a total of up to 3 separate allocations in the
worst case for a single module. We always kernel_read*() the module,
that's a vmalloc(). Then vmap() is used for the module decompression,
and if so the last read buffer is freed as we use the now decompressed
module buffer to stuff data into our copy module. The last allocation is
specific to each architectures but pretty much that's generally a series
of vmalloc() calls or a variation of vmalloc to handle ELF sections with
special permissions.
Evaluation with new stress-ng module support [1] with just 100 ops
is proving that you can end up using GiBs of data easily even with all
care we have in the kernel and userspace today in trying to not load modules
which are already loaded. 100 ops seems to resemble the sort of pressure a
system with about 400 CPUs can create on module loading. Although issues
relating to duplicate module requests due to each CPU inucurring a new
module reuest is silly and some of these are being fixed, we currently lack
proper tooling to help diagnose easily what happened, when it happened
and who likely is to blame -- userspace or kernel module autoloading.
Provide an initial set of stats which use debugfs to let us easily scrape
post-boot information about failed loads. This sort of information can
be used on production worklaods to try to optimize *avoiding* redundant
memory pressure using finit_module().
There's a few examples that can be provided:
A 255 vCPU system without the next patch in this series applied:
Startup finished in 19.143s (kernel) + 7.078s (userspace) = 26.221s
graphical.target reached after 6.988s in userspace
And 13.58 GiB of virtual memory space lost due to failed module loading:
root@big ~ # cat /sys/kernel/debug/modules/stats
Mods ever loaded 67
Mods failed on kread 0
Mods failed on decompress 0
Mods failed on becoming 0
Mods failed on load 1411
Total module size 11464704
Total mod text size 4194304
Failed kread bytes 0
Failed decompress bytes 0
Failed becoming bytes 0
Failed kmod bytes 14588526272
Virtual mem wasted bytes 14588526272
Average mod size 171115
Average mod text size 62602
Average fail load bytes 10339140
Duplicate failed modules:
module-name How-many-times Reason
kvm_intel 249 Load
kvm 249 Load
irqbypass 8 Load
crct10dif_pclmul 128 Load
ghash_clmulni_intel 27 Load
sha512_ssse3 50 Load
sha512_generic 200 Load
aesni_intel 249 Load
crypto_simd 41 Load
cryptd 131 Load
evdev 2 Load
serio_raw 1 Load
virtio_pci 3 Load
nvme 3 Load
nvme_core 3 Load
virtio_pci_legacy_dev 3 Load
virtio_pci_modern_dev 3 Load
t10_pi 3 Load
virtio 3 Load
crc32_pclmul 6 Load
crc64_rocksoft 3 Load
crc32c_intel 40 Load
virtio_ring 3 Load
crc64 3 Load
The following screen shot, of a simple 8vcpu 8 GiB KVM guest with the
next patch in this series applied, shows 226.53 MiB are wasted in virtual
memory allocations which due to duplicate module requests during boot.
It also shows an average module memory size of 167.10 KiB and an an
average module .text + .init.text size of 61.13 KiB. The end shows all
modules which were detected as duplicate requests and whether or not
they failed early after just the first kernel_read*() call or late after
we've already allocated the private space for the module in
layout_and_allocate(). A system with module decompression would reveal
more wasted virtual memory space.
We should put effort now into identifying the source of these duplicate
module requests and trimming these down as much possible. Larger systems
will obviously show much more wasted virtual memory allocations.
root@kmod ~ # cat /sys/kernel/debug/modules/stats
Mods ever loaded 67
Mods failed on kread 0
Mods failed on decompress 0
Mods failed on becoming 83
Mods failed on load 16
Total module size 11464704
Total mod text size 4194304
Failed kread bytes 0
Failed decompress bytes 0
Failed becoming bytes 228959096
Failed kmod bytes 8578080
Virtual mem wasted bytes 237537176
Average mod size 171115
Average mod text size 62602
Avg fail becoming bytes 2758544
Average fail load bytes 536130
Duplicate failed modules:
module-na |
