TLS offload drivers keep track of TCP seq numbers to make sure
the packets are fed into the HW in order.
When packets get dropped on the way through the stack, the driver
will get out of sync and have to use fallback encryption, but unless
TCP seq number is resynced it will never match the packets correctly
(or even worse - use incorrect record sequence number after TCP seq
wraps).
Existing drivers (mlx5) feed the entire record on every out-of-order
event, allowing FW/HW to always be in sync.
This patch adds an alternative, more akin to the RX resync. When
driver sees a frame which is past its expected sequence number the
stream must have gotten out of order (if the sequence number is
smaller than expected its likely a retransmission which doesn't
require resync). Driver will ask the stack to perform TX sync
before it submits the next full record, and fall back to software
crypto until stack has performed the sync.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently only RX direction is ever resynced, however, TX may
also get out of sequence if packets get dropped on the way to
the driver. Rename the resync callback and add a direction
parameter.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TLS offload device may lose sync with the TCP stream if packets
arrive out of order. Drivers can currently request a resync at
a specific TCP sequence number. When a record is found starting
at that sequence number kernel will inform the device of the
corresponding record number.
This requires the device to constantly scan the stream for a
known pattern (constant bytes of the header) after sync is lost.
This patch adds an alternative approach which is entirely under
the control of the kernel. Kernel tracks records it had to fully
decrypt, even though TLS socket is in TLS_HW mode. If multiple
records did not have any decrypted parts - it's a pretty strong
indication that the device is out of sync.
We choose the min number of fully encrypted records to be 2,
which should hopefully be more than will get retransmitted at
a time.
After kernel decides the device is out of sync it schedules a
resync request. If the TCP socket is empty the resync gets
performed immediately. If socket is not empty we leave the
record parser to resync when next record comes.
Before resync in message parser we peek at the TCP socket and
don't attempt the sync if the socket already has some of the
next record queued.
On resync failure (encrypted data continues to flow in) we
retry with exponential backoff, up to once every 128 records
(with a 16k record thats at most once every 2M of data).
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
handle_device_resync() doesn't describe the function very well.
The function checks if resync should be issued upon parsing of
a new record.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TLS offload code casts record number to a u64. The buffer
should be aligned to 8 bytes, but its actually a __be64, and
the rest of the TLS code treats it as big int. Make the
offload callbacks take a byte array, drivers can make the
choice to do the ugly cast if they want to.
Prepare for copying the record number onto the stack by
defining a constant for max size of the byte array.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some ISDN files that got removed in net-next had some changes
done in mainline, take the removals.
Signed-off-by: David S. Miller <davem@davemloft.net>
While offloading TLS connections, drivers need to handle the case where
out of order packets need to be transmitted.
Other drivers obtain the entire TLS record for the specific skb to
provide as context to hardware for encryption. However, other designs
may also want to keep the hardware state intact and perform the
out of order encryption entirely on the host.
To achieve this, export the already existing software encryption
fallback path so drivers could access this.
Signed-off-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently drivers have to ensure the alignment of their tls state
structure, which leads to unnecessary layers of getters and
encapsulated structures in each driver.
Simplify all this by marking the driver state as aligned (driver_state
members are currently aligned, so no hole is added, besides ALIGN in
TLS_OFFLOAD_CONTEXT_SIZE_RX/TX would reserve this extra space, anyway.)
With that we can add a common accessor to the core.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 bytes of driver state has been enough so far, but for drivers
which have to store 8 byte handle it's no longer practical to
store the state directly in the context.
Drivers generally don't need much extra state on RX side, while
TX side has to be tracking TCP sequence numbers. Split the
lengths of max driver state size on RX and TX.
The struct tls_offload_context_tx currently stands at 616 bytes and
struct tls_offload_context_rx stands at 368 bytes. Upcoming work
will consume extra 8 bytes in both for kernel-driven resync.
This means that we can bump TX side to 16 bytes and still fit
into the same number of cache lines but on RX side we would be 8
bytes over.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All callers pass prot->version as the last parameter
of tls_advance_record_sn(), yet tls_advance_record_sn()
itself needs a pointer to prot. Pass prot from callers.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct tls_context is slightly badly laid out. If we reorder things
right we can save 16 bytes (320 -> 304) but also make all fast path
data fit into two cache lines (one read only and one read/write,
down from four cache lines).
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 38030d7cb7 ("net/tls: avoid NULL-deref on resync during device removal")
tried to fix a potential NULL-dereference by taking the
context rwsem. Unfortunately the RX resync may get called
from soft IRQ, so we can't use the rwsem to protect from
the device disappearing. Because we are guaranteed there
can be only one resync at a time (it's called from strparser)
use a bit to indicate resync is busy and make device
removal wait for the bit to get cleared.
Note that there is a leftover "flags" field in struct
tls_context already.
Fixes: 4799ac81e5 ("tls: Add rx inline crypto offload")
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To avoid a sparse warning byteswap the be32 sequence number
before it's stored in the atomic value. While at it drop
unnecessary brackets and use kernel's u64 type.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There seems to be no reason for tls_ops to be defined in netdevice.h
which is included in a lot of places. Don't wrap the struct/enum
declaration in ifdefs, it trickles down unnecessary ifdefs into
driver code.
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tls_device_sk_destruct being set on a socket used to indicate
that socket is a kTLS device one. That is no longer true -
now we use sk_validate_xmit_skb pointer for that purpose.
Remove the export. tls_device_attach() needs to be moved.
While at it, remove the dead declaration of tls_sk_destruct().
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Unlike '&&' operator, the '&' does not have short-circuit
evaluation semantics. IOW both sides of the operator always
get evaluated. Fix the wrong operator in
tls_is_sk_tx_device_offloaded(), which would lead to
out-of-bounds access for for non-full sockets.
Fixes: 4799ac81e5 ("tls: Add rx inline crypto offload")
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
David reports that tls triggers warnings related to
sk->sk_forward_alloc not being zero at destruction time:
WARNING: CPU: 5 PID: 6831 at net/core/stream.c:206 sk_stream_kill_queues+0x103/0x110
WARNING: CPU: 5 PID: 6831 at net/ipv4/af_inet.c:160 inet_sock_destruct+0x15b/0x170
When sender fills up the write buffer and dies from
SIGPIPE. This is due to the device implementation
not cleaning up the partially_sent_record.
This is because commit a42055e8d2 ("net/tls: Add support for async encryption of records for performance")
moved the partial record cleanup to the SW-only path.
Fixes: a42055e8d2 ("net/tls: Add support for async encryption of records for performance")
Reported-by: David Beckett <david.beckett@netronome.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Dirk van der Merwe <dirk.vandermerwe@netronome.com>
Reviewed-by: Simon Horman <simon.horman@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Added support for AES128-CCM based record encryption. AES128-CCM is
similar to AES128-GCM. Both of them have same salt/iv/mac size. The
notable difference between the two is that while invoking AES128-CCM
operation, the salt||nonce (which is passed as IV) has to be prefixed
with a hardcoded value '2'. Further, CCM implementation in kernel
requires IV passed in crypto_aead_request() to be full '16' bytes.
Therefore, the record structure 'struct tls_rec' has been modified to
reserve '16' bytes for IV. This works for both GCM and CCM based cipher.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TLS device cannot use the sw context. This patch returns the original
tls device write space handler and moves the sw/device specific portions
to the relevant files.
Also, we remove the write_space call for the tls_sw flow, because it
handles partial records in its delayed tx work handler.
Fixes: a42055e8d2 ("net/tls: Add support for async encryption of records for performance")
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch enables returning 'type' in msghdr for records that are
retrieved with MSG_PEEK in recvmsg. Further it prevents records peeked
from socket from getting clubbed with any other record of different
type when records are subsequently dequeued from strparser.
For each record, we now retain its type in sk_buff's control buffer
cb[]. Inside control buffer, record's full length and offset are already
stored by strparser in 'struct strp_msg'. We store record type after
'struct strp_msg' inside 'struct tls_msg'. For tls1.2, the type is
stored just after record dequeue. For tls1.3, the type is stored after
record has been decrypted.
Inside process_rx_list(), before processing a non-data record, we check
that we must be able to return back the record type to the user
application. If not, the decrypted records in tls context's rx_list is
left there without consuming any data.
Fixes: 692d7b5d1f ("tls: Fix recvmsg() to be able to peek across multiple records")
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Each tls context maintains two cipher contexts (one each for tx and rx
directions). For each tls session, the constants such as protocol
version, ciphersuite, iv size, associated data size etc are same for
both the directions and need to be stored only once per tls context.
Hence these are moved from 'struct cipher_context' to 'struct
tls_prot_info' and stored only once in 'struct tls_context'.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently we don't zerocopy if the crypto framework async bit is set.
However some crypto algorithms (such as x86 AESNI) support async,
but in the context of sendmsg, will never run asynchronously. Instead,
check for actual EINPROGRESS return code before assuming algorithm is
async.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TLS 1.3 has minor changes from TLS 1.2 at the record layer.
* Header now hardcodes the same version and application content type in
the header.
* The real content type is appended after the data, before encryption (or
after decryption).
* The IV is xored with the sequence number, instead of concatinating four
bytes of IV with the explicit IV.
* Zero-padding: No exlicit length is given, we search backwards from the
end of the decrypted data for the first non-zero byte, which is the
content type. Currently recv supports reading zero-padding, but there
is no way for send to add zero padding.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TLS 1.3 has a different AAD size, use a variable in the code to
make TLS 1.3 support easy.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Wire up support for 256 bit keys from the setsockopt to the crypto
framework
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
aead_request_set_crypt takes an iv pointer, and we change the iv
soon after setting it. Some async crypto algorithms don't save the iv,
so we need to save it in the tls_rec for async requests.
Found by hardcoding x64 aesni to use async crypto manager (to test the async
codepath), however I don't think this combination can happen in the wild.
Presumably other hardware offloads will need this fix, but there have been
no user reports.
Fixes: a42055e8d2 ("Add support for async encryption of records...")
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This fixes recvmsg() to be able to peek across multiple tls records.
Without this patch, the tls's selftests test case
'recv_peek_large_buf_mult_recs' fails. Each tls receive context now
maintains a 'rx_list' to retain incoming skb carrying tls records. If a
tls record needs to be retained e.g. for peek case or for the case when
the buffer passed to recvmsg() has a length smaller than decrypted
record length, then it is added to 'rx_list'. Additionally, records are
added in 'rx_list' if the crypto operation runs in async mode. The
records are dequeued from 'rx_list' after the decrypted data is consumed
by copying into the buffer passed to recvmsg(). In case, the MSG_PEEK
flag is used in recvmsg(), then records are not consumed or removed
from the 'rx_list'.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Borkmann says:
====================
pull-request: bpf-next 2018-12-21
The following pull-request contains BPF updates for your *net-next* tree.
There is a merge conflict in test_verifier.c. Result looks as follows:
[...]
},
{
"calls: cross frame pruning",
.insns = {
[...]
.prog_type = BPF_PROG_TYPE_SOCKET_FILTER,
.errstr_unpriv = "function calls to other bpf functions are allowed for root only",
.result_unpriv = REJECT,
.errstr = "!read_ok",
.result = REJECT,
},
{
"jset: functional",
.insns = {
[...]
{
"jset: unknown const compare not taken",
.insns = {
BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
BPF_FUNC_get_prandom_u32),
BPF_JMP_IMM(BPF_JSET, BPF_REG_0, 1, 1),
BPF_LDX_MEM(BPF_B, BPF_REG_8, BPF_REG_9, 0),
BPF_EXIT_INSN(),
},
.prog_type = BPF_PROG_TYPE_SOCKET_FILTER,
.errstr_unpriv = "!read_ok",
.result_unpriv = REJECT,
.errstr = "!read_ok",
.result = REJECT,
},
[...]
{
"jset: range",
.insns = {
[...]
},
.prog_type = BPF_PROG_TYPE_SOCKET_FILTER,
.result_unpriv = ACCEPT,
.result = ACCEPT,
},
The main changes are:
1) Various BTF related improvements in order to get line info
working. Meaning, verifier will now annotate the corresponding
BPF C code to the error log, from Martin and Yonghong.
2) Implement support for raw BPF tracepoints in modules, from Matt.
3) Add several improvements to verifier state logic, namely speeding
up stacksafe check, optimizations for stack state equivalence
test and safety checks for liveness analysis, from Alexei.
4) Teach verifier to make use of BPF_JSET instruction, add several
test cases to kselftests and remove nfp specific JSET optimization
now that verifier has awareness, from Jakub.
5) Improve BPF verifier's slot_type marking logic in order to
allow more stack slot sharing, from Jiong.
6) Add sk_msg->size member for context access and add set of fixes
and improvements to make sock_map with kTLS usable with openssl
based applications, from John.
7) Several cleanups and documentation updates in bpftool as well as
auto-mount of tracefs for "bpftool prog tracelog" command,
from Quentin.
8) Include sub-program tags from now on in bpf_prog_info in order to
have a reliable way for user space to get all tags of the program
e.g. needed for kallsyms correlation, from Song.
9) Add BTF annotations for cgroup_local_storage BPF maps and
implement bpf fs pretty print support, from Roman.
10) Fix bpftool in order to allow for cross-compilation, from Ivan.
11) Update of bpftool license to GPLv2-only + BSD-2-Clause in order
to be compatible with libbfd and allow for Debian packaging,
from Jakub.
12) Remove an obsolete prog->aux sanitation in dump and get rid of
version check for prog load, from Daniel.
13) Fix a memory leak in libbpf's line info handling, from Prashant.
14) Fix cpumap's frame alignment for build_skb() so that skb_shared_info
does not get unaligned, from Jesper.
15) Fix test_progs kselftest to work with older compilers which are less
smart in optimizing (and thus throwing build error), from Stanislav.
16) Cleanup and simplify AF_XDP socket teardown, from Björn.
17) Fix sk lookup in BPF kselftest's test_sock_addr with regards
to netns_id argument, from Andrey.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
A sockmap program that redirects through a kTLS ULP enabled socket
will not work correctly because the ULP layer is skipped. This
fixes the behavior to call through the ULP layer on redirect to
ensure any operations required on the data stream at the ULP layer
continue to be applied.
To do this we add an internal flag MSG_SENDPAGE_NOPOLICY to avoid
calling the BPF layer on a redirected message. This is
required to avoid calling the BPF layer multiple times (possibly
recursively) which is not the current/expected behavior without
ULPs. In the future we may add a redirect flag if users _do_
want the policy applied again but this would need to work for both
ULP and non-ULP sockets and be opt-in to avoid breaking existing
programs.
Also to avoid polluting the flag space with an internal flag we
reuse the flag space overlapping MSG_SENDPAGE_NOPOLICY with
MSG_WAITFORONE. Here WAITFORONE is specific to recv path and
SENDPAGE_NOPOLICY is only used for sendpage hooks. The last thing
to verify is user space API is masked correctly to ensure the flag
can not be set by user. (Note this needs to be true regardless
because we have internal flags already in-use that user space
should not be able to set). But for completeness we have two UAPI
paths into sendpage, sendfile and splice.
In the sendfile case the function do_sendfile() zero's flags,
./fs/read_write.c:
static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
size_t count, loff_t max)
{
...
fl = 0;
#if 0
/*
* We need to debate whether we can enable this or not. The
* man page documents EAGAIN return for the output at least,
* and the application is arguably buggy if it doesn't expect
* EAGAIN on a non-blocking file descriptor.
*/
if (in.file->f_flags & O_NONBLOCK)
fl = SPLICE_F_NONBLOCK;
#endif
file_start_write(out.file);
retval = do_splice_direct(in.file, &pos, out.file, &out_pos, count, fl);
}
In the splice case the pipe_to_sendpage "actor" is used which
masks flags with SPLICE_F_MORE.
./fs/splice.c:
static int pipe_to_sendpage(struct pipe_inode_info *pipe,
struct pipe_buffer *buf, struct splice_desc *sd)
{
...
more = (sd->flags & SPLICE_F_MORE) ? MSG_MORE : 0;
...
}
Confirming what we expect that internal flags are in fact internal
to socket side.
Fixes: d3b18ad31f ("tls: add bpf support to sk_msg handling")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
HW unhash within mutex for registered tls devices cause sleep
when called from tcp_set_state for TCP_CLOSE. Release lock and
re-acquire after function call with ref count incr/dec.
defined kref and fp release for tls_device to ensure device
is not released outside lock.
BUG: sleeping function called from invalid context at
kernel/locking/mutex.c:748
in_atomic(): 1, irqs_disabled(): 0, pid: 0, name: swapper/7
INFO: lockdep is turned off.
CPU: 7 PID: 0 Comm: swapper/7 Tainted: G W O
Call Trace:
<IRQ>
dump_stack+0x5e/0x8b
___might_sleep+0x222/0x260
__mutex_lock+0x5c/0xa50
? vprintk_emit+0x1f3/0x440
? kmem_cache_free+0x22d/0x2a0
? tls_hw_unhash+0x2f/0x80
? printk+0x52/0x6e
? tls_hw_unhash+0x2f/0x80
tls_hw_unhash+0x2f/0x80
tcp_set_state+0x5f/0x180
tcp_done+0x2e/0xe0
tcp_rcv_state_process+0x92c/0xdd3
? lock_acquire+0xf5/0x1f0
? tcp_v4_rcv+0xa7c/0xbe0
? tcp_v4_do_rcv+0x70/0x1e0
Signed-off-by: Atul Gupta <atul.gupta@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of re-implementing poll routine use the poll callback to
trigger read from kTLS, we reuse the stream_memory_read callback
which is simpler and achieves the same. This helps to align sockmap
and kTLS so we can more easily embed BPF in kTLS.
Joint work with Daniel.
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Convert kTLS over to make use of sk_msg interface for plaintext and
encrypted scattergather data, so it reuses all the sk_msg helpers
and data structure which later on in a second step enables to glue
this to BPF.
This also allows to remove quite a bit of open coded helpers which
are covered by the sk_msg API. Recent changes in kTLs 80ece6a03a
("tls: Remove redundant vars from tls record structure") and
4e6d47206c ("tls: Add support for inplace records encryption")
changed the data path handling a bit; while we've kept the latter
optimization intact, we had to undo the former change to better
fit the sk_msg model, hence the sg_aead_in and sg_aead_out have
been brought back and are linked into the sk_msg sgs. Now the kTLS
record contains a msg_plaintext and msg_encrypted sk_msg each.
In the original code, the zerocopy_from_iter() has been used out
of TX but also RX path. For the strparser skb-based RX path,
we've left the zerocopy_from_iter() in decrypt_internal() mostly
untouched, meaning it has been moved into tls_setup_from_iter()
with charging logic removed (as not used from RX). Given RX path
is not based on sk_msg objects, we haven't pursued setting up a
dummy sk_msg to call into sk_msg_zerocopy_from_iter(), but it
could be an option to prusue in a later step.
Joint work with John.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Presently, for non-zero copy case, separate pages are allocated for
storing plaintext and encrypted text of records. These pages are stored
in sg_plaintext_data and sg_encrypted_data scatterlists inside record
structure. Further, sg_plaintext_data & sg_encrypted_data are passed
to cryptoapis for record encryption. Allocating separate pages for
plaintext and encrypted text is inefficient from both required memory
and performance point of view.
This patch adds support of inplace encryption of records. For non-zero
copy case, we reuse the pages from sg_encrypted_data scatterlist to
copy the application's plaintext data. For the movement of pages from
sg_encrypted_data to sg_plaintext_data scatterlists, we introduce a new
function move_to_plaintext_sg(). This function add pages into
sg_plaintext_data from sg_encrypted_data scatterlists.
tls_do_encryption() is modified to pass the same scatterlist as both
source and destination into aead_request_set_crypt() if inplace crypto
has been enabled. A new ariable 'inplace_crypto' has been introduced in
record structure to signify whether the same scatterlist can be used.
By default, the inplace_crypto is enabled in get_rec(). If zero-copy is
used (i.e. plaintext data is not copied), inplace_crypto is set to '0'.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Reviewed-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Structure 'tls_rec' contains sg_aead_in and sg_aead_out which point
to a aad_space and then chain scatterlists sg_plaintext_data,
sg_encrypted_data respectively. Rather than using chained scatterlists
for plaintext and encrypted data in aead_req, it is efficient to store
aad_space in sg_encrypted_data and sg_plaintext_data itself in the
first index and get rid of sg_aead_in, sg_aead_in and further chaining.
This requires increasing size of sg_encrypted_data & sg_plaintext_data
arrarys by 1 to accommodate entry for aad_space. The code which uses
sg_encrypted_data and sg_plaintext_data has been modified to skip first
index as it points to aad_space.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On processors with multi-engine crypto accelerators, it is possible that
multiple records get encrypted in parallel and their encryption
completion is notified to different cpus in multicore processor. This
leads to the situation where tls_encrypt_done() starts executing in
parallel on different cores. In current implementation, encrypted
records are queued to tx_ready_list in tls_encrypt_done(). This requires
addition to linked list 'tx_ready_list' to be protected. As
tls_decrypt_done() could be executing in irq content, it is not possible
to protect linked list addition operation using a lock.
To fix the problem, we remove linked list addition operation from the
irq context. We do tx_ready_list addition/removal operation from
application context only and get rid of possible multiple access to
the linked list. Before starting encryption on the record, we add it to
the tail of tx_ready_list. To prevent tls_tx_records() from transmitting
it, we mark the record with a new flag 'tx_ready' in 'struct tls_rec'.
When record encryption gets completed, tls_encrypt_done() has to only
update the 'tx_ready' flag to true & linked list add operation is not
required.
The changed logic brings some other side benefits. Since the records
are always submitted in tls sequence number order for encryption, the
tx_ready_list always remains sorted and addition of new records to it
does not have to traverse the linked list.
Lastly, we renamed tx_ready_list in 'struct tls_sw_context_tx' to
'tx_list'. This is because now, the some of the records at the tail are
not ready to transmit.
Fixes: a42055e8d2 ("net/tls: Add support for async encryption")
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In current implementation, tls records are encrypted & transmitted
serially. Till the time the previously submitted user data is encrypted,
the implementation waits and on finish starts transmitting the record.
This approach of encrypt-one record at a time is inefficient when
asynchronous crypto accelerators are used. For each record, there are
overheads of interrupts, driver softIRQ scheduling etc. Also the crypto
accelerator sits idle most of time while an encrypted record's pages are
handed over to tcp stack for transmission.
This patch enables encryption of multiple records in parallel when an
async capable crypto accelerator is present in system. This is achieved
by allowing the user space application to send more data using sendmsg()
even while previously issued data is being processed by crypto
accelerator. This requires returning the control back to user space
application after submitting encryption request to accelerator. This
also means that zero-copy mode of encryption cannot be used with async
accelerator as we must be done with user space application buffer before
returning from sendmsg().
There can be multiple records in flight to/from the accelerator. Each of
the record is represented by 'struct tls_rec'. This is used to store the
memory pages for the record.
After the records are encrypted, they are added in a linked list called
tx_ready_list which contains encrypted tls records sorted as per tls
sequence number. The records from tx_ready_list are transmitted using a
newly introduced function called tls_tx_records(). The tx_ready_list is
polled for any record ready to be transmitted in sendmsg(), sendpage()
after initiating encryption of new tls records. This achieves parallel
encryption and transmission of records when async accelerator is
present.
There could be situation when crypto accelerator completes encryption
later than polling of tx_ready_list by sendmsg()/sendpage(). Therefore
we need a deferred work context to be able to transmit records from
tx_ready_list. The deferred work context gets scheduled if applications
are not sending much data through the socket. If the applications issue
sendmsg()/sendpage() in quick succession, then the scheduling of
tx_work_handler gets cancelled as the tx_ready_list would be polled from
application's context itself. This saves scheduling overhead of deferred
work.
The patch also brings some side benefit. We are able to get rid of the
concept of CLOSED record. This is because the records once closed are
either encrypted and then placed into tx_ready_list or if encryption
fails, the socket error is set. This simplifies the kernel tls
sendpath. However since tls_device.c is still using macros, accessory
functions for CLOSED records have been retained.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Two new tls tests added in parallel in both net and net-next.
Used Stephen Rothwell's linux-next resolution.
Signed-off-by: David S. Miller <davem@davemloft.net>
When async support was added it needed to access the sk from the async
callback to report errors up the stack. The patch tried to use space
after the aead request struct by directly setting the reqsize field in
aead_request. This is an internal field that should not be used
outside the crypto APIs. It is used by the crypto code to define extra
space for private structures used in the crypto context. Users of the
API then use crypto_aead_reqsize() and add the returned amount of
bytes to the end of the request memory allocation before posting the
request to encrypt/decrypt APIs.
So this breaks (with general protection fault and KASAN error, if
enabled) because the request sent to decrypt is shorter than required
causing the crypto API out-of-bounds errors. Also it seems unlikely the
sk is even valid by the time it gets to the callback because of memset
in crypto layer.
Anyways, fix this by holding the sk in the skb->sk field when the
callback is set up and because the skb is already passed through to
the callback handler via void* we can access it in the handler. Then
in the handler we need to be careful to NULL the pointer again before
kfree_skb. I added comments on both the setup (in tls_do_decryption)
and when we clear it from the crypto callback handler
tls_decrypt_done(). After this selftests pass again and fixes KASAN
errors/warnings.
Fixes: 94524d8fc9 ("net/tls: Add support for async decryption of tls records")
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Reviewed-by: Vakul Garg <Vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This contains key material in crypto_send_aes_gcm_128 and
crypto_recv_aes_gcm_128.
Introduce union tls_crypto_context, and replace the two identical
unions directly embedded in struct tls_context with it. We can then
use this union to clean up the memory in the new tls_ctx_free()
function.
Fixes: 3c4d755915 ("tls: kernel TLS support")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
When tls records are decrypted using asynchronous acclerators such as
NXP CAAM engine, the crypto apis return -EINPROGRESS. Presently, on
getting -EINPROGRESS, the tls record processing stops till the time the
crypto accelerator finishes off and returns the result. This incurs a
context switch and is not an efficient way of accessing the crypto
accelerators. Crypto accelerators work efficient when they are queued
with multiple crypto jobs without having to wait for the previous ones
to complete.
The patch submits multiple crypto requests without having to wait for
for previous ones to complete. This has been implemented for records
which are decrypted in zero-copy mode. At the end of recvmsg(), we wait
for all the asynchronous decryption requests to complete.
The references to records which have been sent for async decryption are
dropped. For cases where record decryption is not possible in zero-copy
mode, asynchronous decryption is not used and we wait for decryption
crypto api to complete.
For crypto requests executing in async fashion, the memory for
aead_request, sglists and skb etc is freed from the decryption
completion handler. The decryption completion handler wakesup the
sleeping user context when recvmsg() flags that it has done sending
all the decryption requests and there are no more decryption requests
pending to be completed.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Reviewed-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For preparing decryption request, several memory chunks are required
(aead_req, sgin, sgout, iv, aad). For submitting the decrypt request to
an accelerator, it is required that the buffers which are read by the
accelerator must be dma-able and not come from stack. The buffers for
aad and iv can be separately kmalloced each, but it is inefficient.
This patch does a combined allocation for preparing decryption request
and then segments into aead_req || sgin || sgout || iv || aad.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch completes the generic infrastructure to offload TLS crypto to a
network device. It enables the kernel to skip decryption and
authentication of some skbs marked as decrypted by the NIC. In the fast
path, all packets received are decrypted by the NIC and the performance
is comparable to plain TCP.
This infrastructure doesn't require a TCP offload engine. Instead, the
NIC only decrypts packets that contain the expected TCP sequence number.
Out-Of-Order TCP packets are provided unmodified. As a result, at the
worst case a received TLS record consists of both plaintext and ciphertext
packets. These partially decrypted records must be reencrypted,
only to be decrypted.
The notable differences between SW KTLS Rx and this offload are as
follows:
1. Partial decryption - Software must handle the case of a TLS record
that was only partially decrypted by HW. This can happen due to packet
reordering.
2. Resynchronization - tls_read_size calls the device driver to
resynchronize HW after HW lost track of TLS record framing in
the TCP stream.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch splits tls_sw_release_resources_rx into two functions one
which releases all inner software tls structures and another that also
frees the containing structure.
In TLS_DEVICE we will need to release the software structures without
freeeing the containing structure, which contains other information.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously, decrypt_skb also updated the TLS context.
Now, decrypt_skb only decrypts the payload using the current context,
while decrypt_skb_update also updates the state.
Later, in the tls_device Rx flow, we will use decrypt_skb directly.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For symmetry, we rename tls_offload_context to
tls_offload_context_tx before we add tls_offload_context_rx.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The poll() changes were not well thought out, and completely
unexplained. They also caused a huge performance regression, because
"->poll()" was no longer a trivial file operation that just called down
to the underlying file operations, but instead did at least two indirect
calls.
Indirect calls are sadly slow now with the Spectre mitigation, but the
performance problem could at least be largely mitigated by changing the
"->get_poll_head()" operation to just have a per-file-descriptor pointer
to the poll head instead. That gets rid of one of the new indirections.
But that doesn't fix the new complexity that is completely unwarranted
for the regular case. The (undocumented) reason for the poll() changes
was some alleged AIO poll race fixing, but we don't make the common case
slower and more complex for some uncommon special case, so this all
really needs way more explanations and most likely a fundamental
redesign.
[ This revert is a revert of about 30 different commits, not reverted
individually because that would just be unnecessarily messy - Linus ]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
While hacking on kTLS, I ran into the following panic from an
unprivileged netserver / netperf TCP session:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000000
PGD 800000037f378067 P4D 800000037f378067 PUD 3c0e61067 PMD 0
Oops: 0010 [#1] SMP KASAN PTI
CPU: 1 PID: 2289 Comm: netserver Not tainted 4.17.0+ #139
Hardware name: LENOVO 20FBCTO1WW/20FBCTO1WW, BIOS N1FET47W (1.21 ) 11/28/2016
RIP: 0010: (null)
Code: Bad RIP value.
RSP: 0018:ffff88036abcf740 EFLAGS: 00010246
RAX: dffffc0000000000 RBX: ffff88036f5f6800 RCX: 1ffff1006debed26
RDX: ffff88036abcf920 RSI: ffff8803cb1a4f00 RDI: ffff8803c258c280
RBP: ffff8803c258c280 R08: ffff8803c258c280 R09: ffffed006f559d48
R10: ffff88037aacea43 R11: ffffed006f559d49 R12: ffff8803c258c280
R13: ffff8803cb1a4f20 R14: 00000000000000db R15: ffffffffc168a350
FS: 00007f7e631f4700(0000) GS:ffff8803d1c80000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: ffffffffffffffd6 CR3: 00000003ccf64005 CR4: 00000000003606e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
? tls_sw_poll+0xa4/0x160 [tls]
? sock_poll+0x20a/0x680
? do_select+0x77b/0x11a0
? poll_schedule_timeout.constprop.12+0x130/0x130
? pick_link+0xb00/0xb00
? read_word_at_a_time+0x13/0x20
? vfs_poll+0x270/0x270
? deref_stack_reg+0xad/0xe0
? __read_once_size_nocheck.constprop.6+0x10/0x10
[...]
Debugging further, it turns out that calling into ctx->sk_poll() is
invalid since sk_poll itself is NULL which was saved from the original
TCP socket in order for tls_sw_poll() to invoke it.
Looks like the recent conversion from poll to poll_mask callback started
in 1525242310 ("net: add support for ->poll_mask in proto_ops") missed
to eventually convert kTLS, too: TCP's ->poll was converted over to the
->poll_mask in commit 2c7d3daceb ("net/tcp: convert to ->poll_mask")
and therefore kTLS wrongly saved the ->poll old one which is now NULL.
Convert kTLS over to use ->poll_mask instead. Also instead of POLLIN |
POLLRDNORM use the proper EPOLLIN | EPOLLRDNORM bits as the case in
tcp_poll_mask() as well that is mangled here.
Fixes: 2c7d3daceb ("net/tcp: convert to ->poll_mask")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Dave Watson <davejwatson@fb.com>
Tested-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Jakub Kicinski
David S. Miller
Dirk van der Merwe
Vakul Garg
Boris Pismenny
Dave Watson
John Fastabend
Atul Gupta
Daniel Borkmann
Sabrina Dubroca
Linus Torvalds