Conversation
When activating pages, the LRU flag needs to be set back. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Pass the current si_status (should be simply the signal that stopped the task). Fixes: d5bbf4b ("fork: Implement proper sys_clone semantics") Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Inactive works are also attached to a pool_workqueue. Therefore, ->data needs to be set. Fixes: 19ef128 ("workqueue: add initial version of workqueues") Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Disallow the same work running multiple times concurrently, by keeping track of every worker's executing work, plus a check when assigning work. If the work that's getting assigned is already running, then that same worker will run it after finishing. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
This all stemmed from PTY write paths, which recursed very awkwardly immediately. Since the workqueue implementation was introduced, replace DPC with it, and use it for every TTY driver, including serial, VT and PTY. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Instead of checking for egid == 0, it's supposed to check for euid == 0 (if the process is capable user-wise). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Instead of trying to manually piece together the options string (by parsing $'s), take the whole thing. Whoever is calling crypt() can just do that, and crypt() will grok its own options. This fixes hash calculation on some hash formats different than the static files used in the repo. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
On every fill_pbuf() call, the kernel dropped cmsgs by setting has_cmsg to false. This is wrong, and trivially explained by: sendmsg(Message 0, SCM_RIGHTS) \- creates new pbf, adds fds sendmsg(Message 1, SCM_RIGHTS) \- tries to merge with 0, calls fill_pbuf() \- gets 0, not an error (merge did not happen), clears has_cmsg \- allocates new pbf, no longer has cmsgs, silently dropped This caused problems in programs doing FD passing like OpenSSH sshd, which send various FDs in a row, without plain messages in between. Fix it by reworking this code with a unix_sendmsg_data. The consuming side (fill_pbuf()) consumes it by clearing msg after init. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Fix various weird oddities and bugs in FD passing partial failure, and socket destruction without consuming, by properly setting pbf->dtor to unix_pbf_free. This covers every pbf destruction path. Also cover any odd edgecases in unix_scm_rights() failing midway by clamping nfiles. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Logic pushing raw input bytes to the line discipline needs to properly handle backpressure (when the buffer is full). Do so by setting a flag, that input consumption will look at to further wake up tty_ldisc_input() work. This fixes dropped input in large PTY writes. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
In case something fires, lockdep will stop updating its internal data, and everyone looking at it will get stale content. Wrap the whole WARN_ON_ONCE with a debug_locks. If lockdep fires, then debug_locks will be set to 0. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Erroneously, the dgram recvmsg() logic was only discarding the packet once it reached the end. This is wrong and does not provide DGRAM semantics on the read side. A short read should discard the whole packet, instead of providing more of the same packet on the next recvmsg(). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Add SOCK_SEQPACKET support by using most of STREAM's connection-based logic, plus strict DGRAM boundary limits. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
For use in C code. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Not clearing the active flag causes confusion in the LRU, since it increments inactive_file/anon, and then on free it will decrement active_file/anon. This is obviously wrong. Clear the flag. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
swap_free() returned the swap in use, not free. Also, SwapTotal: was being printed twice, instead of the correct SwapFree: label. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
It's possible the batch was completely added to the LRU, which makes it so no one grabs a lock. This leads to a NULL deref when doing the final unlock. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
folio_batch_init() was broken (and using it would make the next folio_batch_add() store out of bounds). folio_batch_count() was also broken. And ultimately the folio_batch_add()'s return value was also broken (although it ended up being fine for the one case that the mm code cares about, which is "is this folio batch full"). Fix it. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Otherwise, if batchlen != 1 (which currently there isn't any caller that does so), the result ends up being incorrect (a batch filled with the exact same struct page). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Add them to the tail of the LRU, and not the head. Otherwise, the next reclaim run will find them (which results in not much of a rotation at all). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Dirty indirect blocks when doing truncation. This fixes data corruption on partial truncates (e.g truncates that don't get rid of the indirect block). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
It was accidentally jumping to out_release_write, so it released a write it was never able to grab. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Do tid clearing at exec time as well, before the mm is switched to the new one. This matches what happens in Linux. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Prepare reclaim for large folios. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Prepare for large folios. As a side effect, delete the page batching logic (in favour of the folio batch). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Most of it isn't safe for large folios yet (due to pagecache and swap reasons). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
RSS and VIRT need to be given in pages, not in kilobytes. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Always reset compound page metadata on new allocations. This makes sure that new allocations don't accidentally pick up stale info (and accidentally make page_compound_head() point into a stale head). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Properly unlock the page (by doign nr_ios = 0 sooner) if ext2_map_page fails. Otherwise, the if (nr_ios == 0) check in out_err is false, and the page is never unlocked. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
If the reader is shut down, it's supposed to get 0 (EOF) instead of -EPIPE. -EPIPE only applies to writers. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signal peer_nowr to the peer on disconnects. This provides the proper EOF semantics on the reader's side when the peer disconnects. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Instead of relying on userspace to do it, do it in the kernel. This makes it so init= is actually useful (and can use e.g /bin/bash directly). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Sometimes, when the stars align, the following hang early in boot can be
observed:
atomic<unsigned long>::load (this=0xffffffff81c801b0, order=mem_order::seq_cst) at include/onyx/atomic.hpp:51
atomic<unsigned long>::operator unsigned long (this=0xffffffff81c801b0) at include/onyx/atomic.hpp:175
smp::internal::sync_call_cntrlblk::wait (this=0xffffffff81c801a0, local=<optimized out>, context=<optimized out>) at kernel/smp.cpp:121
0xffffffff8103f144 in smp::sync_call_with_local (f=f@entry=0xffffffff81103b10 <x86_invalidate_tlb(void*)>, context=context@entry=0xffffffff81c80270, mask_=...,
local=local@entry=0xffffffff81103b10 <x86_invalidate_tlb(void*)>, context2=context2@entry=0xffffffff81c80270, flags=flags@entry=0) at kernel/smp.cpp:245
0xffffffff81103ebf in mmu_invalidate_range (addr=<optimized out>, pages=<optimized out>, mm=<optimized out>) at arch/x86_64/mmu.cpp:740
0xffffffff81076b01 in vm_invalidate_range (addr=<optimized out>, pages=<optimized out>) at kernel/mm/vm.c:2339
0xffffffff8107fe0b in tlbi_end_batch (tlbi=tlbi@entry=0xffffffff81c80308) at kernel/mm/memory.c:441
0xffffffff81080f58 in vm_mmu_unmap (mm=<optimized out>, addr=<optimized out>, pages=<optimized out>, vma=<optimized out>) at kernel/mm/memory.c:927
0xffffffff81078070 in __vfree (ptr=<optimized out>, is_mmiounmap=<optimized out>) at kernel/mm/vmalloc.cpp:335
0xffffffff81107757 in hpet_timer::~hpet_timer (this=0xffffd001270eef60) at arch/x86_64/hpet.cpp:108
unique_ptr<hpet_timer>::delete_mem (this=<synthetic pointer>) at include/onyx/memory.hpp:302
unique_ptr<hpet_timer>::~unique_ptr (this=<synthetic pointer>) at include/onyx/memory.hpp:377
hpet_init () at arch/x86_64/hpet.cpp:185
0xffffffff81030abc in do_init_level (level=3) at kernel/init.cpp:192
kernel_main () at kernel/init.cpp:233
0xffffffff810fdcd9 in x86_start () at arch/x86_64/boot.S:434
0x0000000000000000 in ?? ()
In other words, somewhere in the (fully no-op) HPET driver, the timer is
destroyed, the mmio mapping is freed, and the TLB shootdown hangs.
Further inspection reveals that the smp::sync_call_with_local is hanging
waiting for a CPU that never receives the message. Anecdotally looking
at the debug cpumask (with DEBUG_SMP_SYNC_CALL) shows it's usually the
last CPU to be brought up (e.g 0b1000 for a 4-cpu system).
The cause is simple: the BSP sets CPUs as online as soon as they set
boot_done. boot_done is set very early in boot. So, the following can
happen:
CPU 0 | CPU N
wake up CPU N |
| runs early bootstrap
receive boot_done |
set_online(N) |
... |
rest of init | smpboot_main()
... |
smp_sync_call |
[sends messages] |
| lapic_init_per_cpu()
where CPU N misses the message sent from CPU 0.
Instead of being subject to this, spin in the BSP on the online mask,
and make the secondary core set itself online after doing bringup (just
before entering idle). This fully sidesteps the issue.
The fix only applies to x86 and should probably be done on the riscv
side as well.
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Instead of conditionally creating it, create it for every inode. This makes for less broken edgecases. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
writeback_inode() doesn't need ->writepages() if data isn't dirty. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
This fixes funny edgecases (which I no longer recall, because I've carried this fix for a good while). Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Otherwise the destructor in exec_state() will free it. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Add group support and surrounding support code for rtnetlink notifications. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Add infrastructure for broadcasting link status changes via workqueues and netlink. This is required by userspace daemons and tooling. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Add libmnl (per d24c88693be13ae77843e18e3388ca9142efdc31) and lightly modified for Onyx. All credit goes to Pablo Neira Ayuso & other authors. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Listen for, and track, link status changes and new devices via netlink RTM_NEWLINK messages. This introduces a new dependency on libnl, a lightweight library for manipulating netlink. The inner workings are the following: instead of blindly using eth0 (despite it being present or not), create an rtnetlink socket bound to the LINK group. netctld then initiates a GETLINK dump, while it listens for changes on the LINK group. If an interface was down, but now it's up, it starts configuration (via normal dhcp, ipv6 procedures (already pre-existing). Not much changed there. If an interface is new and up, it also starts configuration. Doing interface teardown on RTM_DELLINK (or link down) is an exercise left to the reader. While at it, relicense the whole code to GPLv2. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Add peak RSS tracking, as per expected in getrusage, wait4, etc. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Disable it. Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.