Testing padata with the tcrypt module on a 5.2 kernel...

    # modprobe tcrypt alg="pcrypt(rfc4106(gcm(aes)))" type=3
    # modprobe tcrypt mode=211 sec=1

...produces this splat:

    INFO: task modprobe:10075 blocked for more than 120 seconds.
          Not tainted 5.2.0-base+ #16
    modprobe        D    0 10075  10064 0x80004080
    Call Trace:
     ? __schedule+0x4dd/0x610
     ? ring_buffer_unlock_commit+0x23/0x100
     schedule+0x6c/0x90
     schedule_timeout+0x3b/0x320
     ? trace_buffer_unlock_commit_regs+0x4f/0x1f0
     wait_for_common+0x160/0x1a0
     ? wake_up_q+0x80/0x80
     { crypto_wait_req }             # entries in braces added by hand
     { do_one_aead_op }
     { test_aead_jiffies }
     test_aead_speed.constprop.17+0x681/0xf30 [tcrypt]
     do_test+0x4053/0x6a2b [tcrypt]
     ? 0xffffffffa00f4000
     tcrypt_mod_init+0x50/0x1000 [tcrypt]
     ...

The second modprobe command never finishes because in padata_reorder,
CPU0's load of reorder_objects is executed before the unlocking store in
spin_unlock_bh(pd->lock), causing CPU0 to miss CPU1's increment:

CPU0                                 CPU1

padata_reorder                       padata_do_serial
  LOAD reorder_objects  // 0
                                       INC reorder_objects  // 1
                                       padata_reorder
                                         TRYLOCK pd->lock   // failed
  UNLOCK pd->lock

CPU0 deletes the timer before returning from padata_reorder and since no
other job is submitted to padata, modprobe waits indefinitely.

Add a pair of full barriers to guarantee proper ordering:

CPU0                                 CPU1

padata_reorder                       padata_do_serial
  UNLOCK pd->lock
  smp_mb()
  LOAD reorder_objects
                                       INC reorder_objects
                                       smp_mb__after_atomic()
                                       padata_reorder
                                         TRYLOCK pd->lock

smp_mb__after_atomic is needed so the read part of the trylock operation
comes after the INC, as Andrea points out.   Thanks also to Andrea for
help with writing a litmus test.

Fixes: 16295bec ("padata: Generic parallelization/serialization interface")
Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Cc: <stable@vger.kernel.org>
Cc: Andrea Parri <andrea.parri@amarulasolutions.com>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Paul E. McKenney <paulmck@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: linux-arch@vger.kernel.org
Cc: linux-crypto@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

SEV_VERSION_GREATER_OR_EQUAL() will fail if upgrading from 2.2 to 3.1, for
example, because the minor version is not equal to or greater than the
major.

Fix this and move to a static inline function for appropriate type
checking.

Fixes: edd303ff ("crypto: ccp - Add DOWNLOAD_FIRMWARE SEV command")
Reported-by: Cfir Cohen <cfir@google.com>
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Tom Lendacky <thomas.lendacky@amd.com>
Acked-by: Gary R Hook <gary.hook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Avoid leaking GCM tag through timing side channel.

Fixes: 36cf515b ("crypto: ccp - Enable support for AES GCM on v5 CCPs")
Cc: <stable@vger.kernel.org> # v4.12+
Signed-off-by: Cfir Cohen <cfir@google.com>
Acked-by: Gary R Hook <ghook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The AES GCM function reuses an 'op' data structure, which members
contain values that must be cleared for each (re)use.

This fix resolves a crypto self-test failure:
alg: aead: gcm-aes-ccp encryption test failed (wrong result) on test vector 2, cfg="two even aligned splits"

Fixes: 36cf515b ("crypto: ccp - Enable support for AES GCM on v5 CCPs")
Cc: <stable@vger.kernel.org>
Signed-off-by: Gary R Hook <gary.hook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

There is a possible double free issue in ppc4xx_trng_probe():

85:	dev->trng_base = of_iomap(trng, 0);
86:	of_node_put(trng);          ---> released here
87:	if (!dev->trng_base)
88:		goto err_out;
...
110:	ierr_out:
111:		of_node_put(trng);  ---> double released here
...

This issue was detected by using the Coccinelle software.
We fix it by removing the unnecessary of_node_put().

Fixes: 5343e674 ("crypto4xx: integrate ppc4xx-rng into crypto4xx")
Signed-off-by: Wen Yang <wen.yang99@zte.com.cn>
Cc: <stable@vger.kernel.org>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Allison Randal <allison@lohutok.net>
Cc: Armijn Hemel <armijn@tjaldur.nl>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: linux-crypto@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Acked-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

This patch fixes a warning when compiling stm32 because %d is being
used on a size_t argument instead of %zd.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

When DMA is used, waiting for completion must not be
interruptible as it can generate an error that is not handle
by the driver. There is no need to put the completion
interruptible in this driver.
Signed-off-by: Lionel Debieve <lionel.debieve@st.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Correct condition for the second hmac loop. Key must be only
set in the first loop. Initial condition was wrong,
HMAC_KEY flag was not properly checked.
Signed-off-by: Lionel Debieve <lionel.debieve@st.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Use the same naming convention for all stm32 crypto
drivers.
Signed-off-by: Lionel Debieve <lionel.debieve@st.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

In commit af7ddd8a
("Merge tag 'dma-mapping-4.21' of git://git.infradead.org/users/hch/dma-mapping"),
dma_alloc_coherent has already zeroed the memory.
So memset is not needed.
Signed-off-by: Fuqian Huang <huangfq.daxian@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Add an SPDX identifier and remove any specific statements.
Signed-off-by: Gary R Hook <gary.hook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The error code read from the queue status register is only 6 bits wide,
but we need to verify its value is within range before indexing the error
messages.

Fixes: 81422bad ("crypto: ccp - Make syslog errors human-readable")
Cc: <stable@vger.kernel.org>
Reported-by: Cfir Cohen <cfir@google.com>
Signed-off-by: Gary R Hook <gary.hook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Tidy up the formatting/grammar in crypto_engine.rst. Use bulleted lists
where appropriate.
Signed-off-by: Gary R Hook <gary.hook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Fill in missing parameter descriptions for the compression algorithm,
then pick them up to document for the compression_alg structure.
Signed-off-by: Gary R Hook <gary.hook@amd.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

This implements 5-way interleaving for ECB, CBC decryption and CTR,
resulting in a speedup of ~11% on Marvell ThunderX2, which has a
very deep pipeline and therefore a high issue latency for NEON
instructions operating on the same registers.

Note that XTS is left alone: implementing 5-way interleave there
would either involve spilling of the calculated tweaks to the
stack, or recalculating them after the encryption operation, and
doing either of those would most likely penalize low end cores.

For ECB, this is not a concern at all, given that we have plenty
of spare registers. For CTR and CBC decryption, we take advantage
of the fact that v16 is not used by the CE version of the code
(which is the only one targeted by the optimization), and so we
can reshuffle the code a bit and avoid having to spill to memory
(with the exception of one extra reload in the CBC routine)
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

In preparation of tweaking the accelerated AES chaining mode routines
to be able to use a 5-way stride, implement the core routines to
support processing 5 blocks of input at a time. While at it, drop
the 2 way versions, which have been unused for a while now.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

icv_ool is not used anymore, drop it.

Fixes: e345177d ("crypto: talitos - fix AEAD processing.")
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

On SEC1, hash provides wrong result when performing hashing in several
steps with input data SG list has more than one element. This was
detected with CONFIG_CRYPTO_MANAGER_EXTRA_TESTS:

[   44.185947] alg: hash: md5-talitos test failed (wrong result) on test vector 6, cfg="random: may_sleep use_finup src_divs=[<reimport>25.88%@+8063, <flush>24.19%@+9588, 28.63%@+16333, <reimport>4.60%@+6756, 16.70%@+16281] dst_divs=[71.61%@alignmask+16361, 14.36%@+7756, 14.3%@+"
[   44.325122] alg: hash: sha1-talitos test failed (wrong result) on test vector 3, cfg="random: inplace use_final src_divs=[<flush,nosimd>16.56%@+16378, <reimport>52.0%@+16329, 21.42%@alignmask+16380, 10.2%@alignmask+16380] iv_offset=39"
[   44.493500] alg: hash: sha224-talitos test failed (wrong result) on test vector 4, cfg="random: use_final nosimd src_divs=[<reimport>52.27%@+7401, <reimport>17.34%@+16285, <flush>17.71%@+26, 12.68%@+10644] iv_offset=43"
[   44.673262] alg: hash: sha256-talitos test failed (wrong result) on test vector 4, cfg="random: may_sleep use_finup src_divs=[<reimport>60.6%@+12790, 17.86%@+1329, <reimport>12.64%@alignmask+16300, 8.29%@+15, 0.40%@+13506, <reimport>0.51%@+16322, <reimport>0.24%@+16339] dst_divs"

This is due to two issues:
- We have an overlap between the buffer used for copying the input
data (SEC1 doesn't do scatter/gather) and the chained descriptor.
- Data copy is wrong when the previous hash left less than one
blocksize of data to hash, implying a complement of the previous
block with a few bytes from the new request.

Fix it by:
- Moving the second descriptor after the buffer, as moving the buffer
after the descriptor would make it more complex for other cipher
operations (AEAD, ABLKCIPHER)
- Skip the bytes taken from the new request to complete the previous
one by moving the SG list forward.

Fixes: 37b5e889 ("crypto: talitos - chain in buffered data for ahash on SEC1")
Cc: stable@vger.kernel.org
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Moves struct talitos_edesc into talitos.h so that it can be used
from any place in talitos.c

It will be required for next patch ("crypto: talitos - fix hash
on SEC1")
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Cc: stable@vger.kernel.org
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

All mapping iterator logic is based on the assumption that sg->offset
is always lower than PAGE_SIZE.

But there are situations where sg->offset is such that the SG item
is on the second page. In that case sg_copy_to_buffer() fails
properly copying the data into the buffer. One of the reason is
that the data will be outside the kmapped area used to access that
data.

This patch fixes the issue by adjusting the mapping iterator
offset and pgoffset fields such that offset is always lower than
PAGE_SIZE.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Fixes: 4225fc85 ("lib/scatterlist: use page iterator in the mapping iterator")
Cc: stable@vger.kernel.org
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

System gets checkstop if RxFIFO overruns with more requests than the
maximum possible number of CRBs in FIFO at the same time. The max number
of requests per window is controlled by window credits. So find max
CRBs from FIFO size and set it to receive window credits.

Fixes: b0d6c9ba ("crypto/nx: Add P9 NX support for 842 compression engine")
CC: stable@vger.kernel.org # v4.14+
Signed-off-by: Haren Myneni <haren@us.ibm.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Build testing with some core crypto options disabled revealed
a few modules that are missing CRYPTO_HASH:

crypto/asymmetric_keys/x509_public_key.o: In function `x509_get_sig_params':
x509_public_key.c:(.text+0x4c7): undefined reference to `crypto_alloc_shash'
x509_public_key.c:(.text+0x5e5): undefined reference to `crypto_shash_digest'
crypto/asymmetric_keys/pkcs7_verify.o: In function `pkcs7_digest.isra.0':
pkcs7_verify.c:(.text+0xab): undefined reference to `crypto_alloc_shash'
pkcs7_verify.c:(.text+0x1b2): undefined reference to `crypto_shash_digest'
pkcs7_verify.c:(.text+0x3c1): undefined reference to `crypto_shash_update'
pkcs7_verify.c:(.text+0x411): undefined reference to `crypto_shash_finup'

This normally doesn't show up in randconfig tests because there is
a large number of other options that select CRYPTO_HASH.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The same bug that gcc hit in the past is apparently now showing
up with clang, which decides to inline __serpent_setkey_sbox:

crypto/serpent_generic.c:268:5: error: stack frame size of 2112 bytes in function '__serpent_setkey' [-Werror,-Wframe-larger-than=]

Marking it 'noinline' reduces the stack usage from 2112 bytes to
192 and 96 bytes, respectively, and seems to generate more
useful object code.

Fixes: c871c10e ("crypto: serpent - improve __serpent_setkey with UBSAN")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Eric Biggers <ebiggers@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The largest stack object in this file is now the shash descriptor.
Since there are many other stack variables, this can push it
over the 1024 byte warning limit, in particular with clang and
KASAN:

crypto/testmgr.c:1693:12: error: stack frame size of 1312 bytes in function '__alg_test_hash' [-Werror,-Wframe-larger-than=]

Make test_hash_vs_generic_impl() do the same thing as the
corresponding eaed and skcipher functions by allocating the
descriptor dynamically. We can still do better than this,
but it brings us well below the 1024 byte limit.
Suggested-by: Eric Biggers <ebiggers@kernel.org>
Fixes: 9a8a6b3f ("crypto: testmgr - fuzz hashes against their generic implementation")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Eric Biggers <ebiggers@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

On arm32, we get warnings about high stack usage in some of the functions:

crypto/testmgr.c:2269:12: error: stack frame size of 1032 bytes in function 'alg_test_aead' [-Werror,-Wframe-larger-than=]
static int alg_test_aead(const struct alg_test_desc *desc, const char *driver,
           ^
crypto/testmgr.c:1693:12: error: stack frame size of 1312 bytes in function '__alg_test_hash' [-Werror,-Wframe-larger-than=]
static int __alg_test_hash(const struct hash_testvec *vecs,
           ^

On of the larger objects on the stack here is struct testvec_config, so
change that to dynamic allocation.

Fixes: 40153b10 ("crypto: testmgr - fuzz AEADs against their generic implementation")
Fixes: d435e10e ("crypto: testmgr - fuzz skciphers against their generic implementation")
Fixes: 9a8a6b3f ("crypto: testmgr - fuzz hashes against their generic implementation")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Eric Biggers <ebiggers@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

When building for SEC1 only, talitos2_done functions are unneeded
and should go away.

For this, use has_ftr_sec1() which will always return true when only
SEC1 support is being built, allowing GCC to drop TALITOS2 functions.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Horia Geantă <horia.geanta@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

After the latest addition, the stack usage of sun4i_ss_cipher_poll
grew beyond the warning limit when KASAN is enabled:

drivers/crypto/sunxi-ss/sun4i-ss-cipher.c:118:12: error: stack frame size of 1152 bytes in function 'sun4i_ss_cipher_poll' [-Werror,-Wframe-larger-than=]
static int sun4i_ss_cipher_poll(struct skcipher_request *areq)

Reduce it in three ways:

- split out the new code into a separate function so its stack
  usage can overlap that of the sun4i_ss_opti_poll() code path
- mark both special cases as noinline_for_stack, which should
  ideally result in a tail call that frees the rest of the
  stack
- move the buf and obuf variables into the code blocks in
  which they are used.

The three separate functions now use 144, 640 and 304 bytes of kernel
stack, respectively.

Fixes: 0ae1f46c ("crypto: sun4i-ss - fallback when length is not multiple of blocksize")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Tested-by: Corentin LABBE <clabbe.montjoie@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Add check to verify the stated device tree HW configuration
matches the HW.
Signed-off-by: Gilad Ben-Yossef <gilad@benyossef.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

ccree irq may be shared with other devices, in order to prevent ccree isr
handling while device maybe suspended we added a check to verify that the
device is not suspended.
Signed-off-by: Ofir Drang <ofir.drang@arm.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

In case of driver probe and pm resume we need to check that the cryptocell
hardware reset cycle is completed. during the reset cycle that Cryptocell
provide read only access to the APB interface which allows to verify
through the CC registers that the reset is completed. Until reset
completion we assume that any write/crypto operation is blocked.
Signed-off-by: Ofir Drang <ofir.drang@arm.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Signed-off-by: Ofir Drang <ofir.drang@arm.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

When calling debugfs functions, there is no need to ever check the
return value.  The function can work or not, but the code logic should
never do something different based on this.

Also, there is no need to store the individual debugfs file names,
especially as the whole directiry is deleted at once, so remove the
unneeded structure entirely.

Cc: "Breno Leitão" <leitao@debian.org>
Cc: Nayna Jain <nayna@linux.ibm.com>
Cc: Paulo Flabiano Smorigo <pfsmorigo@gmail.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: linux-crypto@vger.kernel.org
Cc: linuxppc-dev@lists.ozlabs.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The CIFS code uses the sync skcipher API to invoke the ecb(arc4) skcipher,
of which only a single generic C code implementation exists. This means
that going through all the trouble of using scatterlists etc buys us
very little, and we're better off just invoking the arc4 library directly.

This also reverts commit 5f4b5569 ("CIFS: Fix BUG() in calc_seckey()"),
since it is no longer necessary to allocate sec_key on the heap.

Cc: linux-cifs@vger.kernel.org
Cc: Steve French <sfrench@samba.org>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Acked-by: Steve French <stfrench@microsoft.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The MPPE code uses the sync skcipher to invoke the ecb(arc4) skcipher,
of which only a single generic C code implementation exists. This means
that going through all the trouble of using scatterlists etc buys us
very little, and we're better off just invoking the arc4 library directly.

Note that the SHA1 shash used by this driver has several accelerated
implementations for various architectures, so retaining that part does
make sense.

Cc: linux-ppp@vger.kernel.org
Cc: Paul Mackerras <paulus@samba.org>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

There are no remaining users of the cipher implementation, and there
are no meaningful ways in which the arc4 cipher can be combined with
templates other than ECB (and the way we do provide that combination
is highly dubious to begin with).

So let's drop the arc4 cipher altogether, and only keep the ecb(arc4)
skcipher, which is used in various places in the kernel.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The crypto API abstraction is not very useful for invoking ciphers
directly, especially in the case of arc4, which only has a generic
implementation in C. So let's invoke the library code directly.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The crypto API abstraction is not very useful for invoking ciphers
directly, especially in the case of arc4, which only has a generic
implementation in C. So let's invoke the library code directly.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

The WEP code in the mac80211 subsystem currently uses the crypto
API to access the arc4 (RC4) cipher, which is overly complicated,
and doesn't really have an upside in this particular case, since
ciphers are always synchronous and therefore always implemented in
software. Given that we have no accelerated software implementations
either, it is much more straightforward to invoke a generic library
interface directly.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Refactor the core rc4 handling so we can move most users to a library
interface, permitting us to drop the cipher interface entirely in a
future patch. This is part of an effort to simplify the crypto API
and improve its robustness against incorrect use.
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

Below commit came with a typo in the CONFIG_ symbol, leading
to a permanently reduced max key size regarless of the driver
capabilities.
Reported-by: Horia Geantă <horia.geanta@nxp.com>
Fixes: b8fbdc2b ("crypto: talitos - reduce max key size for SEC1")
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Horia Geantă <horia.geanta@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>