and flags with an sxlock. This leads to a significant and measurable
performance improvement as a result of access to shared locking for
frequent lookup operations, reduced general overhead, and reduced overhead
in the event of contention. All of these are imported for threaded
applications where simultaneous access to a shared file descriptor array
occurs frequently. Kris has reported 2x-4x transaction rate improvements
on 8-core MySQL benchmarks; smaller improvements can be expected for many
workloads as a result of reduced overhead.
- Generally eliminate the distinction between "fast" and regular
acquisisition of the filedesc lock; the plan is that they will now all
be fast. Change all locking instances to either shared or exclusive
locks.
- Correct a bug (pointed out by kib) in fdfree() where previously msleep()
was called without the mutex held; sx_sleep() is now always called with
the sxlock held exclusively.
- Universally hold the struct file lock over changes to struct file,
rather than the filedesc lock or no lock. Always update the f_ops
field last. A further memory barrier is required here in the future
(discussed with jhb).
- Improve locking and reference management in linux_at(), which fails to
properly acquire vnode references before using vnode pointers. Annotate
improper use of vn_fullpath(), which will be replaced at a future date.
In fcntl(), we conservatively acquire an exclusive lock, even though in
some cases a shared lock may be sufficient, which should be revisited.
The dropping of the filedesc lock in fdgrowtable() is no longer required
as the sxlock can be held over the sleep operation; we should consider
removing that (pointed out by attilio).
Tested by: kris
Discussed with: jhb, kris, attilio, jeff
o make all crypto drivers have a device_t; pseudo drivers like the s/w
crypto driver synthesize one
o change the api between the crypto subsystem and drivers to use kobj;
cryptodev_if.m defines this api
o use the fact that all crypto drivers now have a device_t to add support
for specifying which of several potential devices to use when doing
crypto operations
o add new ioctls that allow user apps to select a specific crypto device
to use (previous ioctls maintained for compatibility)
o overhaul crypto subsystem code to eliminate lots of cruft and hide
implementation details from drivers
o bring in numerous fixes from Michale Richardson/hifn; mostly for
795x parts
o add an optional mechanism for mmap'ing the hifn 795x public key h/w
to user space for use by openssl (not enabled by default)
o update crypto test tools to use new ioctl's and add cmd line options
to specify a device to use for tests
These changes will also enable much future work on improving the core
crypto subsystem; including proper load balancing and interposing code
between the core and drivers to dispatch small operations to the s/w
driver as appropriate.
These changes were instigated by the work of Michael Richardson.
Reviewed by: pjd
Approved by: re
Before the change if a hardware crypto driver was loaded after
the software crypto driver, calling crypto_newsession() with
hard=0, will always choose software crypto.
- Add defines with block length for each HMAC algorithm.
- Add AES_BLOCK_LEN define which is an alias for RIJNDAEL128_BLOCK_LEN.
- Add NULL_BLOCK_LEN define.
Checking if the queues are empty is not enough for the crypto_proc thread
(it is enough for the crypto_ret_thread), because drivers can be marked
as blocked. In a situation where we have operations related to different
crypto drivers in the queue, it is possible that one driver is marked as
blocked. In this case, the queue will not be empty and we won't wakeup
the crypto_proc thread to execute operations for the others drivers.
Simply setting a global variable to 1 when we goes to sleep and setting
it back to 0 when we wake up is sufficient. The variable is protected
with the queue lock.
Before the change if the thread was working on symmetric operation, we
would send unnecessary wakeup after adding asymmetric operation (when
asym queue was empty) and vice versa.
twice if we call crypto_kinvoke() from crypto_proc thread.
This change also removes unprotected access to cc_kqblocked field
(CRYPTO_Q_LOCK() should be used for protection).
where crypto_invoke() returns ERESTART and before we set cc_qblocked to 1,
crypto_unblock() is called and sets it to 0. This way we mark device as
blocked forever.
Fix it by not setting cc_qblocked in the fast path and by protecting
crypto_invoke() in the crypto_proc thread with CRYPTO_Q_LOCK().
This won't slow things down, because there is no contention - we have
only one crypto thread. Actually it can be slightly faster, because we
save two atomic ops per crypto request.
The fast code path remains lock-less.
or SHA512, the blocksize is 128 bytes, not 64 bytes as anywhere else.
The bug also exists in NetBSD, OpenBSD and various other independed
implementations I look at.
- We cannot decide which hash function to use for HMAC based on the key
length, because any HMAC function can use any key length.
To fix it split CRYPTO_SHA2_HMAC into three algorithm:
CRYPTO_SHA2_256_HMAC, CRYPTO_SHA2_384_HMAC and CRYPTO_SHA2_512_HMAC.
Those names are consistent with OpenBSD's naming.
- Remove authsize field from auth_hash structure.
- Allow consumer to define size of hash he wants to receive.
This allows to use HMAC not only for IPsec, where 96 bits MAC is requested.
The size of requested MAC is defined at newsession time in the cri_mlen
field - when 0, entire MAC will be returned.
- Add swcr_authprepare() function which prepares authentication key.
- Allow to provide key for every authentication operation, not only at
newsession time by honoring CRD_F_KEY_EXPLICIT flag.
- Make giving key at newsession time optional - don't try to operate on it
if its NULL.
- Extend COPYBACK()/COPYDATA() macros to handle CRYPTO_BUF_CONTIG buffer
type as well.
- Accept CRYPTO_BUF_IOV buffer type in swcr_authcompute() as we have
cuio_apply() now.
- 16 bits for key length (SW_klen) is more than enough.
Reviewed by: sam
crypto_invoke(). This allows to serve multiple crypto requests in
parallel and not bached requests are served lock-less.
Drivers should not depend on the queue lock beeing held around
crypto_invoke() and if they do, that's an error in the driver - it
should do its own synchronization.
- Don't forget to wakeup the crypto thread when new requests is
queued and only if both symmetric and asymmetric queues are empty.
- Symmetric requests use sessions and there is no way driver can
disappear when there is an active session, so we don't need to check
this, but assert this. This is also safe to not use the driver lock
in this case.
- Assymetric requests don't use sessions, so don't check the driver
in crypto_kinvoke().
- Protect assymetric operation with the driver lock, because if there
is no symmetric session, driver can disappear.
- Don't send assymetric request to the driver if it is marked as
blocked.
- Add an XXX comment, because I don't think migration to another driver
is safe when there are pending requests using freed session.
- Remove 'hint' argument from crypto_kinvoke(), as it serves no purpose.
- Don't hold the driver lock around kprocess method call, instead use
cc_koperations to track number of in-progress requests.
- Cleanup register/unregister code a bit.
- Other small simplifications and cleanups.
Reviewed by: sam
- Implement CUIO_SKIP() macro which is only responsible for skipping the given
number of bytes from iovec list. This allows to avoid duplicating the same
code in three functions.
Reviewed by: sam
and want to have crypto support loaded as KLD. By moving zlib to separate
module and adding MODULE_DEPEND directives, it is possible to use such
configuration without complication. Otherwise, since IPSEC is linked with
zlib (just like crypto.ko) you'll get following error:
interface zlib.1 already present in the KLD 'kernel'!
Approved by: cognet (mentor)
This is actually a local DoS, as every user can use /dev/crypto if there
is crypto hardware in the system and cryptodev.ko is loaded (or compiled
into the kernel).
Reported by: Mike Tancsa <mike@sentex.net>
MFC after: 1 day
It checked other algorithms against this bug and it seems they aren't
affected.
Reported by: Mike Tancsa <mike@sentex.net>
PR: i386/84860
Reviewed by: phk, cperciva(x2)
Don't grab Giant in the upper syscall/wrapper code
NET_LOCK_GIANT in the socket code (sockets/fifos).
mtx_lock(&Giant) in the vnode code.
mtx_lock(&Giant) in the opencrypto code. (This may actually not be
needed, but better safe than sorry).
Devfs grabs Giant if the driver is marked as needing Giant.
individual file object implementations can optionally acquire Giant if
they require it:
- soo_close(): depends on debug.mpsafenet
- pipe_close(): Giant not acquired
- kqueue_close(): Giant required
- vn_close(): Giant required
- cryptof_close(): Giant required (conservative)
Notes:
Giant is still acquired in close() even when closing MPSAFE objects
due to kqueue requiring Giant in the calling closef() code.
Microbenchmarks indicate that this removal of Giant cuts 3%-3% off
of pipe create/destroy pairs from user space with SMP compiled into
the kernel.
The cryptodev and opencrypto code appears MPSAFE, but I'm unable to
test it extensively and so have left Giant over fo_close(). It can
probably be removed given some testing and review.
