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>
S390 bpf_jit.S is removed in net-next and had changes in 'net',
since that code isn't used any more take the removal.
TLS data structures split the TX and RX components in 'net-next',
put the new struct members from the bug fix in 'net' into the RX
part.
The 'net-next' tree had some reworking of how the ERSPAN code works in
the GRE tunneling code, overlapping with a one-line headroom
calculation fix in 'net'.
Overlapping changes in __sock_map_ctx_update_elem(), keep the bits
that read the prog members via READ_ONCE() into local variables
before using them.
Signed-off-by: David S. Miller <davem@davemloft.net>
scatterlist code expects virt_to_page() to work, which fails with
CONFIG_VMAP_STACK=y.
Fixes: c46234ebb4 ("tls: RX path for ktls")
Signed-off-by: Matt Mullins <mmullins@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is reported that in some cases, write_space may be called in
do_tcp_sendpages, such that we recursively invoke do_tcp_sendpages again:
[ 660.468802] ? do_tcp_sendpages+0x8d/0x580
[ 660.468826] ? tls_push_sg+0x74/0x130 [tls]
[ 660.468852] ? tls_push_record+0x24a/0x390 [tls]
[ 660.468880] ? tls_write_space+0x6a/0x80 [tls]
...
tls_push_sg already does a loop over all sending sg's, so ignore
any tls_write_space notifications until we are done sending.
We then have to call the previous write_space to wake up
poll() waiters after we are done with the send loop.
Reported-by: Andre Tomt <andre@tomt.net>
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a generic infrastructure to offload TLS crypto to a
network device. It enables the kernel TLS socket to skip encryption
and authentication operations on the transmit side of the data path.
Leaving those computationally expensive operations to the NIC.
The NIC offload infrastructure builds TLS records and pushes them to
the TCP layer just like the SW KTLS implementation and using the same
API.
TCP segmentation is mostly unaffected. Currently the only exception is
that we prevent mixed SKBs where only part of the payload requires
offload. In the future we are likely to add a similar restriction
following a change cipher spec record.
The notable differences between SW KTLS and NIC offloaded TLS
implementations are as follows:
1. The offloaded implementation builds "plaintext TLS record", those
records contain plaintext instead of ciphertext and place holder bytes
instead of authentication tags.
2. The offloaded implementation maintains a mapping from TCP sequence
number to TLS records. Thus given a TCP SKB sent from a NIC offloaded
TLS socket, we can use the tls NIC offload infrastructure to obtain
enough context to encrypt the payload of the SKB.
A TLS record is released when the last byte of the record is ack'ed,
this is done through the new icsk_clean_acked callback.
The infrastructure should be extendable to support various NIC offload
implementations. However it is currently written with the
implementation below in mind:
The NIC assumes that packets from each offloaded stream are sent as
plaintext and in-order. It keeps track of the TLS records in the TCP
stream. When a packet marked for offload is transmitted, the NIC
encrypts the payload in-place and puts authentication tags in the
relevant place holders.
The responsibility for handling out-of-order packets (i.e. TCP
retransmission, qdisc drops) falls on the netdev driver.
The netdev driver keeps track of the expected TCP SN from the NIC's
perspective. If the next packet to transmit matches the expected TCP
SN, the driver advances the expected TCP SN, and transmits the packet
with TLS offload indication.
If the next packet to transmit does not match the expected TCP SN. The
driver calls the TLS layer to obtain the TLS record that includes the
TCP of the packet for transmission. Using this TLS record, the driver
posts a work entry on the transmit queue to reconstruct the NIC TLS
state required for the offload of the out-of-order packet. It updates
the expected TCP SN accordingly and transmits the now in-order packet.
The same queue is used for packet transmission and TLS context
reconstruction to avoid the need for flushing the transmit queue before
issuing the context reconstruction request.
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Aviad Yehezkel <aviadye@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In TLS inline crypto, we can have one direction in software
and another in hardware. Thus, we split the TLS configuration to separate
structures for receive and transmit.
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Facility to register Inline TLS drivers to net/tls. Setup
TLS_HW_RECORD prot to listen on offload device.
Cases handled
- Inline TLS device exists, setup prot for TLS_HW_RECORD
- Atleast one Inline TLS exists, sets TLS_HW_RECORD.
- If non-inline device establish connection, move to TLS_SW_TX
Signed-off-by: Atul Gupta <atul.gupta@chelsio.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add rx path for tls software implementation.
recvmsg, splice_read, and poll implemented.
An additional sockopt TLS_RX is added, with the same interface as
TLS_TX. Either TLX_RX or TLX_TX may be provided separately, or
together (with two different setsockopt calls with appropriate keys).
Control messages are passed via CMSG in a similar way to transmit.
If no cmsg buffer is passed, then only application data records
will be passed to userspace, and EIO is returned for other types of
alerts.
EBADMSG is passed for decryption errors, and EMSGSIZE is passed for
framing too big, and EBADMSG for framing too small (matching openssl
semantics). EINVAL is returned for TLS versions that do not match the
original setsockopt call. All are unrecoverable.
strparser is used to parse TLS framing. Decryption is done directly
in to userspace buffers if they are large enough to support it, otherwise
sk_cow_data is called (similar to ipsec), and buffers are decrypted in
place and copied. splice_read always decrypts in place, since no
buffers are provided to decrypt in to.
sk_poll is overridden, and only returns POLLIN if a full TLS message is
received. Otherwise we wait for strparser to finish reading a full frame.
Actual decryption is only done during recvmsg or splice_read calls.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several config variables are prefixed with tx, drop the prefix
since these will be used for both tx and rx.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pass EBADMSG explicitly to tls_err_abort. Receive path will
pass additional codes - EMSGSIZE if framing is larger than max
TLS record size, EINVAL if TLS version mismatch.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Separate tx crypto parameters to a separate cipher_context struct.
The same parameters will be used for rx using the same struct.
tls_advance_record_sn is modified to only take the cipher info.
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Async crypto accelerators (e.g. drivers/crypto/caam) support offloading
GCM operation. If they are enabled, crypto_aead_encrypt() return error
code -EINPROGRESS. In this case tls_do_encryption() needs to wait on a
completion till the time the response for crypto offload request is
received.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
sendfile() calls can hang endless with using Kernel TLS if a socket error occurs.
Socket error codes must be inverted by Kernel TLS before returning because
they are stored with positive sign. If returned non-inverted they are
interpreted as number of bytes sent, causing endless looping of the
splice mechanic behind sendfile().
Signed-off-by: Robert Hering <r.hering@avm.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move inclusion of a private kernel header <net/tcp.h>
from uapi/linux/tls.h to its only user - net/tls.h,
to fix the following linux/tls.h userspace compilation error:
/usr/include/linux/tls.h:41:21: fatal error: net/tcp.h: No such file or directory
As to this point uapi/linux/tls.h was totaly unusuable for userspace,
cleanup this header file further by moving other redundant includes
to net/tls.h.
Fixes: 3c4d755915 ("tls: kernel TLS support")
Cc: <stable@vger.kernel.org> # v4.13+
Signed-off-by: Dmitry V. Levin <ldv@altlinux.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
move tls_make_aad as it is going to be reused
by the device offload code and rx path.
Remove unused recv parameter.
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Previously the TLS ulp context would leak if we attached a TLS ulp
to a socket but did not use the TLS_TX setsockopt,
or did use it but it failed.
This patch solves the issue by overriding prot[TLS_BASE_TX].close
and fixing tls_sk_proto_close to work properly
when its called with ctx->tx_conf == TLS_BASE_TX.
This patch also removes ctx->free_resources as we can use ctx->tx_conf
to obtain the relevant information.
Fixes: 3c4d755915 ('tls: kernel TLS support')
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The tx configuration is now stored in ctx->tx_conf.
And sk->sk_prot is updated trough a function
This will simplify things when we add rx
and support for different possible
tx and rx cross configurations.
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Software implementation of transport layer security, implemented using ULP
infrastructure. tcp proto_ops are replaced with tls equivalents of sendmsg and
sendpage.
Only symmetric crypto is done in the kernel, keys are passed by setsockopt
after the handshake is complete. All control messages are supported via CMSG
data - the actual symmetric encryption is the same, just the message type needs
to be passed separately.
For user API, please see Documentation patch.
Pieces that can be shared between hw and sw implementation
are in tls_main.c
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Aviad Yehezkel <aviadye@mellanox.com>
Signed-off-by: Dave Watson <davejwatson@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
