CVS commit: [netbsd-10] src

Module Name:	src
Committed By:	martin
Date:		Sat Jun 27 15:25:57 UTC 2026

Modified Files:
	src/distrib/sets/lists/debug [netbsd-10]: mi
	src/distrib/sets/lists/tests [netbsd-10]: mi
	src/sys/opencrypto [netbsd-10]: cryptodev.c cryptodev_internal.h
	    cryptosoft.c ocryptodev.c
	src/sys/rump/dev/lib/libopencrypto [netbsd-10]: OPENCRYPTO.ioconf
	    opencrypto_component.c
	src/tests/crypto/opencrypto [netbsd-10]: Makefile t_opencrypto.sh
Added Files:
	src/tests/crypto/opencrypto [netbsd-10]: h_thread.c

Log Message:
Pull up following revision(s) (requested by riastradh in ticket #1278):

	sys/opencrypto/cryptodev.c: revision 1.130
	sys/opencrypto/cryptodev.c: revision 1.131
	sys/opencrypto/cryptodev.c: revision 1.132
	sys/opencrypto/cryptodev.c: revision 1.133
	tests/crypto/opencrypto/Makefile: revision 1.6
	sys/opencrypto/cryptosoft.c: revision 1.67
	sys/opencrypto/ocryptodev.c: revision 1.19
	tests/crypto/opencrypto/h_thread.c: revision 1.1
	sys/rump/dev/lib/libopencrypto/OPENCRYPTO.ioconf: revision 1.2
	tests/crypto/opencrypto/h_thread.c: revision 1.2
	sys/rump/dev/lib/libopencrypto/opencrypto_component.c: revision 1.7
	tests/crypto/opencrypto/t_opencrypto.sh: revision 1.12
	tests/crypto/opencrypto/t_opencrypto.sh: revision 1.13
	sys/opencrypto/cryptodev_internal.h: revision 1.5
	tests/crypto/opencrypto/t_opencrypto.sh: revision 1.14
	sys/opencrypto/cryptodev.c: revision 1.129
	distrib/sets/lists/debug/mi: revision 1.512
	distrib/sets/lists/tests/mi: revision 1.1417

Add a threaded test to test for races in cryptodev.
new thread cryptodev test

crypto(4): Make test more reliable, and test more.
1. New thread to concurrently create and destroy sessions.
   (There should really be multiple threads to concurrently compete
   with each other to create and destroy sessions, but this is
   already surfacing more crashes, as I expected.)
2. Handle EBUSY in CIOCFSESSION in case there is a concurrent
   CIOCCRYPT, as we are trying to test.
3. Handle CIOCCRYPT failure if a concurrent CIOCFSESSION beat us to
   it, as we are trying to test
4. Dump core if the threads get stuck for too long.
5. Provide stack traces from the test program or rump server if they
   dump core.
PR kern/60281: crypto(4): bugs in reference counting and test

crypto(4): Take reference _before_ releasing the lock.

Otherwise nothing ensures the object will still exist by the time we
try to take the reference.

Also guard against too many references, since this is only a 32-bit
reference count.
PR kern/60281: crypto(4): bugs in reference counting and test

crypto(4): Fix missing membars on reference count release.

If two threads A and B both hold references, we need to ensure that
memory ops in thread A happen before memory free in thread B in:
thread A                thread B                notes
--------                --------                -----
memory ops
atomic_dec(&refcnt)                             goes from 2 to 1
                        atomic_dec(&refcnt)     goes from 1 to 0
                        memory free

This requires a membar_release in thread A before the atomic_dec (or
atomic_dec with memory_order_release), and a membar_acquire in thread
B after the atomic_dec is found to have brought the reference count
down to zero (or atomic_dec wiht memory_order_acquire).
kern/60281: crypto(4): bugs in reference counting and test

crypto(4): Disentangle initialization and attachment goo.
Lotta unnecessary boilerplate deleted here!

Disable module unloading: can't be done safely.  Explain precisely
why it can't be done safely.

This also fixes annoying `crypto: unable to register devsw, error 17'
messages in rump dmesg by having exactly one path to devsw_attach.
PR kern/60281: crypto(4): bugs in reference counting and test

crypto(4): Omit needless locking in fcrypt_dtor.

We must have exclusive access to the object for this function to work
at all, so if removing the locks appeared to cause issues, it would
necessarily happen only because there is a bug somewhere else.
PR kern/60281: crypto(4): bugs in reference counting and test
crypto(4): Nix spurious mutex_exit; add missing bounds checks.

Consistently use `foo = kmem_alloc(n * sizeof(*foo), ...)' instead of
`sizeof(struct whatever_foo_is)'.  Makes it easier for a reader to
notice a discrepancy this way.

Move CRYPTODEV_OPS_MAX to cryptodev_internal.h so it can be used by
the compat ocryptodev.c shims too.  I think this is waaaaaaaaaaaaay
too high, by the way.  For example, it looks like qat(4) puts a limit
of 16384 on the number of sessions.  Other devices like hifn(4) look
like they're limited to numbers of sessions ranging from 2 to around
256.
PR kern/60281: crypto(4): bugs in reference counting and test


To generate a diff of this commit:
cvs rdiff -u -r1.394.2.15 -r1.394.2.16 src/distrib/sets/lists/debug/mi
cvs rdiff -u -r1.1238.2.19 -r1.1238.2.20 src/distrib/sets/lists/tests/mi
cvs rdiff -u -r1.125.4.1 -r1.125.4.2 src/sys/opencrypto/cryptodev.c
cvs rdiff -u -r1.3.40.1 -r1.3.40.2 src/sys/opencrypto/cryptodev_internal.h
cvs rdiff -u -r1.64.4.2 -r1.64.4.3 src/sys/opencrypto/cryptosoft.c
cvs rdiff -u -r1.17.4.1 -r1.17.4.2 src/sys/opencrypto/ocryptodev.c
cvs rdiff -u -r1.1 -r1.1.52.1 \
    src/sys/rump/dev/lib/libopencrypto/OPENCRYPTO.ioconf
cvs rdiff -u -r1.6 -r1.6.24.1 \
    src/sys/rump/dev/lib/libopencrypto/opencrypto_component.c
cvs rdiff -u -r1.5 -r1.5.8.1 src/tests/crypto/opencrypto/Makefile
cvs rdiff -u -r0 -r1.2.4.2 src/tests/crypto/opencrypto/h_thread.c
cvs rdiff -u -r1.9 -r1.9.8.1 src/tests/crypto/opencrypto/t_opencrypto.sh

Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.