Andrey Ignatov says:
====================
v2->v3:
- rebase due to conflicts
- fix ipv6=m build
v1->v2:
- support expected_attach_type at prog load time so that prog (incl.
context accesses and calls to helpers) can be validated with regard to
specific attach point it is supposed to be attached to.
Later, at attach time, attach type is checked so that it must be same as
at load time if it was provided
- reworked hooks to rely on expected_attach_type, and reduced number of new
prog types from 6 to just 1: BPF_PROG_TYPE_CGROUP_SOCK_ADDR
- reused BPF_PROG_TYPE_CGROUP_SOCK for sys_bind post-hooks
- add selftests for post-sys_bind hook
For our container management we've been using complicated and fragile setup
consisting of LD_PRELOAD wrapper intercepting bind and connect calls from
all containerized applications. Unfortunately it doesn't work for apps that
don't use glibc and changing all applications that run in the datacenter
is not possible due to 3rd party code and libraries (despite being
open source code) and sheer amount of legacy code that has to be rewritten
(we're rewriting what we can in parallel)
These applications are written without containers in mind and have
builtin assumptions about network services. Like an application X
expects to connect localhost:special_port and find service Y in there.
To move application X and service Y into two different containers
LD_PRELOAD approach is used to help one service connect to another
without rewriting them.
Moving these two applications into different L2 (netns) or L3 (vrf)
network isolation scopes doesn't help to solve the problem, since
applications need to see each other like they were running on
the host without containers.
So if app X and app Y would run in different netns something
would need to punch a connectivity hole in those namespaces.
That would be real layering violation (with corresponding
network debugging pains), since clean l2, l3 abstraction would
suddenly support something that breaks through the layers.
Instead we used LD_PRELOAD (and now bpf programs) at bind/connect
time to help applications discover and connect to each other.
All applications are running in init_nens and there are no vrfs.
After bind/connect the normal fib/neighbor core networking
logic works as it should always do and the whole system is
clean from network point of view and can be debugged with
standard tools.
We also considered resurrecting Hannes's afnetns work,
but all hierarchical namespace abstraction don't work due
to these builtin networking assumptions inside the apps.
To run an application inside cgroup container that was not written
with containers in mind we have to make an illusion of running
in non-containerized environment.
In some cases we remember the port and container id in the post-bind hook
in a bpf map and when some other task in a different container is trying
to connect to a service we need to know where this service is running.
It can be remote and can be local. Both client and service may or may not
be written with containers in mind and this sockaddr rewrite is providing
connectivity and load balancing feature.
BPF+cgroup looks to be the best solution for this problem.
Hence we introduce 3 hooks:
- at entry into sys_bind and sys_connect
to let bpf prog look and modify 'struct sockaddr' provided
by user space and fail bind/connect when appropriate
- post sys_bind after port is allocated
The approach works great and has zero overhead for anyone who doesn't
use it and very low overhead when deployed.
Different use case for this feature is to do low overhead firewall
that doesn't need to inspect all packets and works at bind/connect time.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Add selftest for attach types `BPF_CGROUP_INET4_POST_BIND` and
`BPF_CGROUP_INET6_POST_BIND`.
The main things tested are:
* prog load behaves as expected (valid/invalid accesses in prog);
* prog attach behaves as expected (load- vs attach-time attach types);
* `BPF_CGROUP_INET_SOCK_CREATE` can be attached in a backward compatible
way;
* post-hooks return expected result and errno.
Example:
# ./test_sock
Test case: bind4 load with invalid access: src_ip6 .. [PASS]
Test case: bind4 load with invalid access: mark .. [PASS]
Test case: bind6 load with invalid access: src_ip4 .. [PASS]
Test case: sock_create load with invalid access: src_port .. [PASS]
Test case: sock_create load w/o expected_attach_type (compat mode) ..
[PASS]
Test case: sock_create load w/ expected_attach_type .. [PASS]
Test case: attach type mismatch bind4 vs bind6 .. [PASS]
Test case: attach type mismatch bind6 vs bind4 .. [PASS]
Test case: attach type mismatch default vs bind4 .. [PASS]
Test case: attach type mismatch bind6 vs sock_create .. [PASS]
Test case: bind4 reject all .. [PASS]
Test case: bind6 reject all .. [PASS]
Test case: bind6 deny specific IP & port .. [PASS]
Test case: bind4 allow specific IP & port .. [PASS]
Test case: bind4 allow all .. [PASS]
Test case: bind6 allow all .. [PASS]
Summary: 16 PASSED, 0 FAILED
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
"Post-hooks" are hooks that are called right before returning from
sys_bind. At this time IP and port are already allocated and no further
changes to `struct sock` can happen before returning from sys_bind but
BPF program has a chance to inspect the socket and change sys_bind
result.
Specifically it can e.g. inspect what port was allocated and if it
doesn't satisfy some policy, BPF program can force sys_bind to fail and
return EPERM to user.
Another example of usage is recording the IP:port pair to some map to
use it in later calls to sys_connect. E.g. if some TCP server inside
cgroup was bound to some IP:port_n, it can be recorded to a map. And
later when some TCP client inside same cgroup is trying to connect to
127.0.0.1:port_n, BPF hook for sys_connect can override the destination
and connect application to IP:port_n instead of 127.0.0.1:port_n. That
helps forcing all applications inside a cgroup to use desired IP and not
break those applications if they e.g. use localhost to communicate
between each other.
== Implementation details ==
Post-hooks are implemented as two new attach types
`BPF_CGROUP_INET4_POST_BIND` and `BPF_CGROUP_INET6_POST_BIND` for
existing prog type `BPF_PROG_TYPE_CGROUP_SOCK`.
Separate attach types for IPv4 and IPv6 are introduced to avoid access
to IPv6 field in `struct sock` from `inet_bind()` and to IPv4 field from
`inet6_bind()` since those fields might not make sense in such cases.
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
== The problem ==
See description of the problem in the initial patch of this patch set.
== The solution ==
The patch provides much more reliable in-kernel solution for the 2nd
part of the problem: making outgoing connecttion from desired IP.
It adds new attach types `BPF_CGROUP_INET4_CONNECT` and
`BPF_CGROUP_INET6_CONNECT` for program type
`BPF_PROG_TYPE_CGROUP_SOCK_ADDR` that can be used to override both
source and destination of a connection at connect(2) time.
Local end of connection can be bound to desired IP using newly
introduced BPF-helper `bpf_bind()`. It allows to bind to only IP though,
and doesn't support binding to port, i.e. leverages
`IP_BIND_ADDRESS_NO_PORT` socket option. There are two reasons for this:
* looking for a free port is expensive and can affect performance
significantly;
* there is no use-case for port.
As for remote end (`struct sockaddr *` passed by user), both parts of it
can be overridden, remote IP and remote port. It's useful if an
application inside cgroup wants to connect to another application inside
same cgroup or to itself, but knows nothing about IP assigned to the
cgroup.
Support is added for IPv4 and IPv6, for TCP and UDP.
IPv4 and IPv6 have separate attach types for same reason as sys_bind
hooks, i.e. to prevent reading from / writing to e.g. user_ip6 fields
when user passes sockaddr_in since it'd be out-of-bound.
== Implementation notes ==
The patch introduces new field in `struct proto`: `pre_connect` that is
a pointer to a function with same signature as `connect` but is called
before it. The reason is in some cases BPF hooks should be called way
before control is passed to `sk->sk_prot->connect`. Specifically
`inet_dgram_connect` autobinds socket before calling
`sk->sk_prot->connect` and there is no way to call `bpf_bind()` from
hooks from e.g. `ip4_datagram_connect` or `ip6_datagram_connect` since
it'd cause double-bind. On the other hand `proto.pre_connect` provides a
flexible way to add BPF hooks for connect only for necessary `proto` and
call them at desired time before `connect`. Since `bpf_bind()` is
allowed to bind only to IP and autobind in `inet_dgram_connect` binds
only port there is no chance of double-bind.
bpf_bind() sets `force_bind_address_no_port` to bind to only IP despite
of value of `bind_address_no_port` socket field.
bpf_bind() sets `with_lock` to `false` when calling to __inet_bind()
and __inet6_bind() since all call-sites, where bpf_bind() is called,
already hold socket lock.
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Refactor `bind()` code to make it ready to be called from BPF helper
function `bpf_bind()` (will be added soon). Implementation of
`inet_bind()` and `inet6_bind()` is separated into `__inet_bind()` and
`__inet6_bind()` correspondingly. These function can be used from both
`sk_prot->bind` and `bpf_bind()` contexts.
New functions have two additional arguments.
`force_bind_address_no_port` forces binding to IP only w/o checking
`inet_sock.bind_address_no_port` field. It'll allow to bind local end of
a connection to desired IP in `bpf_bind()` w/o changing
`bind_address_no_port` field of a socket. It's useful since `bpf_bind()`
can return an error and we'd need to restore original value of
`bind_address_no_port` in that case if we changed this before calling to
the helper.
`with_lock` specifies whether to lock socket when working with `struct
sk` or not. The argument is set to `true` for `sk_prot->bind`, i.e. old
behavior is preserved. But it will be set to `false` for `bpf_bind()`
use-case. The reason is all call-sites, where `bpf_bind()` will be
called, already hold that socket lock.
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Add selftest to work with bpf_sock_addr context from
`BPF_PROG_TYPE_CGROUP_SOCK_ADDR` programs.
Try to bind(2) on IP:port and apply:
* loads to make sure context can be read correctly, including narrow
loads (byte, half) for IP and full-size loads (word) for all fields;
* stores to those fields allowed by verifier.
All combination from IPv4/IPv6 and TCP/UDP are tested.
Both scenarios are tested:
* valid programs can be loaded and attached;
* invalid programs can be neither loaded nor attached.
Test passes when expected data can be read from context in the
BPF-program, and after the call to bind(2) socket is bound to IP:port
pair that was written by BPF-program to the context.
Example:
# ./test_sock_addr
Attached bind4 program.
Test case #1 (IPv4/TCP):
Requested: bind(192.168.1.254, 4040) ..
Actual: bind(127.0.0.1, 4444)
Test case #2 (IPv4/UDP):
Requested: bind(192.168.1.254, 4040) ..
Actual: bind(127.0.0.1, 4444)
Attached bind6 program.
Test case #3 (IPv6/TCP):
Requested: bind(face:b00c:1234:5678::abcd, 6060) ..
Actual: bind(::1, 6666)
Test case #4 (IPv6/UDP):
Requested: bind(face:b00c:1234:5678::abcd, 6060) ..
Actual: bind(::1, 6666)
### SUCCESS
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
== The problem ==
There is a use-case when all processes inside a cgroup should use one
single IP address on a host that has multiple IP configured. Those
processes should use the IP for both ingress and egress, for TCP and UDP
traffic. So TCP/UDP servers should be bound to that IP to accept
incoming connections on it, and TCP/UDP clients should make outgoing
connections from that IP. It should not require changing application
code since it's often not possible.
Currently it's solved by intercepting glibc wrappers around syscalls
such as `bind(2)` and `connect(2)`. It's done by a shared library that
is preloaded for every process in a cgroup so that whenever TCP/UDP
server calls `bind(2)`, the library replaces IP in sockaddr before
passing arguments to syscall. When application calls `connect(2)` the
library transparently binds the local end of connection to that IP
(`bind(2)` with `IP_BIND_ADDRESS_NO_PORT` to avoid performance penalty).
Shared library approach is fragile though, e.g.:
* some applications clear env vars (incl. `LD_PRELOAD`);
* `/etc/ld.so.preload` doesn't help since some applications are linked
with option `-z nodefaultlib`;
* other applications don't use glibc and there is nothing to intercept.
== The solution ==
The patch provides much more reliable in-kernel solution for the 1st
part of the problem: binding TCP/UDP servers on desired IP. It does not
depend on application environment and implementation details (whether
glibc is used or not).
It adds new eBPF program type `BPF_PROG_TYPE_CGROUP_SOCK_ADDR` and
attach types `BPF_CGROUP_INET4_BIND` and `BPF_CGROUP_INET6_BIND`
(similar to already existing `BPF_CGROUP_INET_SOCK_CREATE`).
The new program type is intended to be used with sockets (`struct sock`)
in a cgroup and provided by user `struct sockaddr`. Pointers to both of
them are parts of the context passed to programs of newly added types.
The new attach types provides hooks in `bind(2)` system call for both
IPv4 and IPv6 so that one can write a program to override IP addresses
and ports user program tries to bind to and apply such a program for
whole cgroup.
== Implementation notes ==
[1]
Separate attach types for `AF_INET` and `AF_INET6` are added
intentionally to prevent reading/writing to offsets that don't make
sense for corresponding socket family. E.g. if user passes `sockaddr_in`
it doesn't make sense to read from / write to `user_ip6[]` context
fields.
[2]
The write access to `struct bpf_sock_addr_kern` is implemented using
special field as an additional "register".
There are just two registers in `sock_addr_convert_ctx_access`: `src`
with value to write and `dst` with pointer to context that can't be
changed not to break later instructions. But the fields, allowed to
write to, are not available directly and to access them address of
corresponding pointer has to be loaded first. To get additional register
the 1st not used by `src` and `dst` one is taken, its content is saved
to `bpf_sock_addr_kern.tmp_reg`, then the register is used to load
address of pointer field, and finally the register's content is restored
from the temporary field after writing `src` value.
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Support setting `expected_attach_type` at prog load time in both
`bpf/bpf.h` and `bpf/libbpf.h`.
Since both headers already have API to load programs, new functions are
added not to break backward compatibility for existing ones:
* `bpf_load_program_xattr()` is added to `bpf/bpf.h`;
* `bpf_prog_load_xattr()` is added to `bpf/libbpf.h`.
Both new functions accept structures, `struct bpf_load_program_attr` and
`struct bpf_prog_load_attr` correspondingly, where new fields can be
added in the future w/o changing the API.
Standard `_xattr` suffix is used to name the new API functions.
Since `bpf_load_program_name()` is not used as heavily as
`bpf_load_program()`, it was removed in favor of more generic
`bpf_load_program_xattr()`.
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
== The problem ==
There are use-cases when a program of some type can be attached to
multiple attach points and those attach points must have different
permissions to access context or to call helpers.
E.g. context structure may have fields for both IPv4 and IPv6 but it
doesn't make sense to read from / write to IPv6 field when attach point
is somewhere in IPv4 stack.
Same applies to BPF-helpers: it may make sense to call some helper from
some attach point, but not from other for same prog type.
== The solution ==
Introduce `expected_attach_type` field in in `struct bpf_attr` for
`BPF_PROG_LOAD` command. If scenario described in "The problem" section
is the case for some prog type, the field will be checked twice:
1) At load time prog type is checked to see if attach type for it must
be known to validate program permissions correctly. Prog will be
rejected with EINVAL if it's the case and `expected_attach_type` is
not specified or has invalid value.
2) At attach time `attach_type` is compared with `expected_attach_type`,
if prog type requires to have one, and, if they differ, attach will
be rejected with EINVAL.
The `expected_attach_type` is now available as part of `struct bpf_prog`
in both `bpf_verifier_ops->is_valid_access()` and
`bpf_verifier_ops->get_func_proto()` () and can be used to check context
accesses and calls to helpers correspondingly.
Initially the idea was discussed by Alexei Starovoitov <ast@fb.com> and
Daniel Borkmann <daniel@iogearbox.net> here:
https://marc.info/?l=linux-netdev&m=152107378717201&w=2
Signed-off-by: Andrey Ignatov <rdna@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Add explicit checks in ext4_xattr_block_get() just in case the
e_value_offs and e_value_size fields in the the xattr block are
corrupted in memory after the buffer_verified bit is set on the xattr
block.
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@kernel.org
The missing error handling in add_extent_changeset was hidden, so make
it at least visible in the callers.
Signed-off-by: David Sterba <dsterba@suse.com>
When mount fails to read trees like fs tree, checksum tree, extent
tree, etc, there is not enough information about where went wrong.
With this, messages like
"BTRFS warning (device sdf): failed to read root (objectid=7): -5"
would help us a bit.
Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
All users pass a local unsigned int and not the __uXX types that are
supposed to be used for userspace interfaces.
Signed-off-by: David Sterba <dsterba@suse.com>
The current calls are unclear in what way btrfs_dev_replace_lock takes
the locks, so drop the argument, split the helpers and use similar
naming as for read and write locks.
Signed-off-by: David Sterba <dsterba@suse.com>
The fs_mutex has been killed in 2008, a213501153 ("Btrfs: Replace
the big fs_mutex with a collection of other locks"), still remembered in
some comments.
We don't have any extra needs for locking in the ACL handlers.
Signed-off-by: David Sterba <dsterba@suse.com>
The show_devname callback is used to print device name in
/proc/self/mounts, we need to traverse the device list consistently and
read the name that's copied to a seq buffer so we don't need further
locking.
If the first device is being deleted at the same time, the RCU will
allow us to read the device name, though it will become stale right
after the RCU protection ends. This is unavoidable and the user can
expect that the device will disappear from the filesystem's list at some
point.
The device_list_mutex was pretty heavy as it is used eg. for writing
superblock and a few other IO related contexts. This can stall any
application that reads the proc file for no reason.
Reviewed-by: Anand Jain <anand.jain@oracle.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Once there was a simple int force_cow that was used with the plain
barriers, and then converted to a bit, so we should use the appropriate
barrier helper.
Other variables in the complex if condition do not depend on a barrier,
so we should be fine in case the atomic barrier becomes a no-op.
Signed-off-by: David Sterba <dsterba@suse.com>
We have several reports about node pointer points to incorrect child
tree blocks, which could have even wrong owner and level but still with
valid generation and checksum.
Although btrfs check could handle it and print error message like:
leaf parent key incorrect 60670574592
Kernel doesn't have enough check on this type of corruption correctly.
At least add such check to read_tree_block() and btrfs_read_buffer(),
where we need two new parameters @level and @first_key to verify the
child tree block.
The new @level check is mandatory and all call sites are already
modified to extract expected level from its call chain.
While @first_key is optional, the following call sites are skipping such
check:
1) Root node/leaf
As ROOT_ITEM doesn't contain the first key, skip @first_key check.
2) Direct backref
Only parent bytenr and level is known and we need to resolve the key
all by ourselves, skip @first_key check.
Another note of this verification is, it needs extra info from nodeptr
or ROOT_ITEM, so it can't fit into current tree-checker framework, which
is limited to node/leaf boundary.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
The extent tree of the test fs is like the following:
BTRFS info (device (null)): leaf 16327509003777336587 total ptrs 1 free space 3919
item 0 key (4096 168 4096) itemoff 3944 itemsize 51
extent refs 1 gen 1 flags 2
tree block key (68719476736 0 0) level 1
^^^^^^^
ref#0: tree block backref root 5
And it's using an empty tree for fs tree, so there is no way that its
level can be 1.
For REAL (created by mkfs) fs tree backref with no skinny metadata, the
result should look like:
item 3 key (30408704 EXTENT_ITEM 4096) itemoff 3845 itemsize 51
refs 1 gen 4 flags TREE_BLOCK
tree block key (256 INODE_ITEM 0) level 0
^^^^^^^
tree block backref root 5
Fix the level to 0, so it won't break later tree level checker.
Fixes: faa2dbf004 ("Btrfs: add sanity tests for new qgroup accounting code")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When logging an inode, at tree-log.c:copy_items(), if we call
btrfs_next_leaf() at the loop which checks for the need to log holes, we
need to make sure copy_items() returns the value 1 to its caller and
not 0 (on success). This is because the path the caller passed was
released and is now different from what is was before, and the caller
expects a return value of 0 to mean both success and that the path
has not changed, while a return value of 1 means both success and
signals the caller that it can not reuse the path, it has to perform
another tree search.
Even though this is a case that should not be triggered on normal
circumstances or very rare at least, its consequences can be very
unpredictable (especially when replaying a log tree).
Fixes: 16e7549f04 ("Btrfs: incompatible format change to remove hole extents")
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
When we have the no-holes mode enabled and fsync a file after punching a
hole in it, we can end up not logging the whole hole range in the log tree.
This happens if the file has extent items that span more than one leaf and
we punch a hole that covers a range that starts in a leaf but does not go
beyond the offset of the first extent in the next leaf.
Example:
$ mkfs.btrfs -f -O no-holes -n 65536 /dev/sdb
$ mount /dev/sdb /mnt
$ for ((i = 0; i <= 831; i++)); do
offset=$((i * 2 * 256 * 1024))
xfs_io -f -c "pwrite -S 0xab -b 256K $offset 256K" \
/mnt/foobar >/dev/null
done
$ sync
# We now have 2 leafs in our filesystem fs tree, the first leaf has an
# item corresponding the extent at file offset 216530944 and the second
# leaf has a first item corresponding to the extent at offset 217055232.
# Now we punch a hole that partially covers the range of the extent at
# offset 216530944 but does go beyond the offset 217055232.
$ xfs_io -c "fpunch $((216530944 + 128 * 1024 - 4000)) 256K" /mnt/foobar
$ xfs_io -c "fsync" /mnt/foobar
<power fail>
# mount to replay the log
$ mount /dev/sdb /mnt
# Before this patch, only the subrange [216658016, 216662016[ (length of
# 4000 bytes) was logged, leaving an incorrect file layout after log
# replay.
Fix this by checking if there is a hole between the last extent item that
we processed and the first extent item in the next leaf, and if there is
one, log an explicit hole extent item.
Fixes: 16e7549f04 ("Btrfs: incompatible format change to remove hole extents")
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
We have a nice helper to do proper casting of a qgroup to a ulist aux
value. And several places that could make use of it.
Signed-off-by: David Sterba <dsterba@suse.com>
This reverts commit 48a89bc4f2.
The idea to commit transaction and free some space after hitting qgroup
limit is good, although the problem is it can easily cause deadlocks.
One deadlock example is caused by trying to flush data while still
holding it:
Call Trace:
__schedule+0x49d/0x10f0
schedule+0xc6/0x290
schedule_timeout+0x187/0x1c0
wait_for_completion+0x204/0x3a0
btrfs_wait_ordered_extents+0xa40/0xaf0 [btrfs]
qgroup_reserve+0x913/0xa10 [btrfs]
btrfs_qgroup_reserve_data+0x3ef/0x580 [btrfs]
btrfs_check_data_free_space+0x96/0xd0 [btrfs]
__btrfs_buffered_write+0x3ac/0xd40 [btrfs]
btrfs_file_write_iter+0x62a/0xba0 [btrfs]
__vfs_write+0x320/0x430
vfs_write+0x107/0x270
SyS_write+0xbf/0x150
do_syscall_64+0x1b0/0x3d0
entry_SYSCALL64_slow_path+0x25/0x25
Another can be caused by trying to commit one transaction while nesting
with trans handle held by ourselves:
btrfs_start_transaction()
|- btrfs_qgroup_reserve_meta_pertrans()
|- qgroup_reserve()
|- btrfs_join_transaction()
|- btrfs_commit_transaction()
The retry is causing more problems than exppected when limit is enabled.
At least a graceful EDQUOT is way better than deadlock.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Now trace_qgroup_meta_reserve() will have extra type parameter.
And introduce two new trace events:
1) trace_qgroup_meta_free_all_pertrans()
For btrfs_qgroup_free_meta_all_pertrans()
2) trace_qgroup_meta_convert()
For btrfs_qgroup_convert_reserved_meta()
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For quota disabled->enable case, it's possible that at reservation time
quota was not enabled so no bytes were really reserved, while at release
time, quota was enabled so we will try to release some bytes we didn't
really own.
Such situation can cause metadata reserveation underflow, for both types,
also less possible for per-trans type since quota enable will commit
transaction.
To address this, record qgroup meta reserved bytes into
root::qgroup_meta_rsv_pertrans and ::prealloc.
So at releasing time we won't free any bytes we didn't reserve.
For DATA, it's already handled by io_tree, so nothing needs to be done
there.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Quite similar for delalloc, some modification to delayed-inode and
delayed-item reservation. Also needs extra parameter for release case
to distinguish normal release and error release.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Add some paranoia checks to make sure we don't stray beyond the end of
the valid memory region containing ext4 xattr entries while we are
scanning for a match.
Also rename the function to xattr_find_entry() since it is static and
thus only used in fs/ext4/xattr.c
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@kernel.org
Upon a new UMR post, check if the WQE buffer contains
a previous UMR WQE. If so, modify the dynamic fields
instead of a whole WQE overwrite. This saves a memcpy.
In current setting, after 2 WQ cycles (12 UMR posts),
this will always be the case.
No degradation sensed.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
All UMR WQEs of an RQ share many common fields. We use
pre-initialized structures to save calculations in datapath.
One field (xlt_offset) was the only reason we saved a pre-initialized
copy per WQE index.
Here we remove its initialization (move its calculation to datapath),
and reduce the number of copies to one-per-RQ.
A very small datapath calculation is added, it occurs once per a MPWQE
(i.e. once every 256KB), but reduces memory consumption and gives
better cache utilization.
Performance testing:
Tested packet rate, no degradation sensed.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
When many packets reside on the same page, the bulking of
page_ref modifications reduces the total number of atomic
operations executed.
Besides the necessary 2 operations on page alloc/free, we
have the following extra ops per page:
- one on WQE allocation (bump refcnt to maximum possible),
- zero ops for SKBs,
- one on WQE free,
a constant of two operations in total, no matter how many
packets/SKBs actually populate the page.
Without this bulking, we have:
- no ops on WQE allocation or free,
- one op per SKB,
Comparing the two methods when PAGE_SIZE is 4K:
- As mentioned above, bulking method always executes 2 operations,
not more, but not less.
- In the default MTU configuration (1500, stride size is 2K),
the non-bulking method execute 2 ops as well.
- For larger MTUs with stride size of 4K, non-bulking method
executes only a single op.
- For XDP (stride size of 4K, no SKBs), non-bulking method
executes no ops at all!
Hence, to optimize the flows with linear SKB and XDP over Striding RQ,
we here remove the page_ref bulking method.
Performance testing:
ConnectX-5, Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz.
Single core packet rate (64 bytes).
Early drop in TC: no degradation.
XDP_DROP:
before: 14,270,188 pps
after: 20,503,603 pps, 43% improvement.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Add XDP support over Striding RQ.
Now that linear SKB is supported over Striding RQ,
we can support XDP by setting stride size to PAGE_SIZE
and headroom to XDP_PACKET_HEADROOM.
Upon a MPWQE free, do not release pages that are being
XDP xmit, they will be released upon completions.
Striding RQ is capable of a higher packet-rate than
conventional RQ.
A performance gain is expected for all cases that had
a HW packet-rate bottleneck. This is the case whenever
using many flows that distribute to many cores.
Performance testing:
ConnectX-5, 24 rings, default MTU.
CQE compression ON (to reduce completions BW in PCI).
XDP_DROP packet rate:
--------------------------------------------------
| pkt size | XDP rate | 100GbE linerate | pct% |
--------------------------------------------------
| 64byte | 126.2 Mpps | 148.0 Mpps | 85% |
| 128byte | 80.0 Mpps | 84.8 Mpps | 94% |
| 256byte | 42.7 Mpps | 42.7 Mpps | 100% |
| 512byte | 23.4 Mpps | 23.4 Mpps | 100% |
--------------------------------------------------
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Make the xdp_xmit indication available for Striding RQ
by taking it out of the type-specific union.
This refactor is a preparation for a downstream patch that
adds XDP support over Striding RQ.
In addition, use a bitmap instead of a boolean for possible
future flags.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Current Striding RQ HW feature utilizes the RX buffers so that
there is no wasted room between the strides. This maximises
the memory utilization.
This prevents the use of build_skb() (which requires headroom
and tailroom), and demands to memcpy the packets headers into
the skb linear part.
In this patch, whenever a set of conditions holds, we apply
an RQ configuration that allows combining the use of linear SKB
on top of a Striding RQ.
To use build_skb() with Striding RQ, the following must hold:
1. packet does not cross a page boundary.
2. there is enough headroom and tailroom surrounding the packet.
We can satisfy 1 and 2 by configuring:
stride size = MTU + headroom + tailoom.
This is possible only when:
a. (MTU - headroom - tailoom) does not exceed PAGE_SIZE.
b. HW LRO is turned off.
Using linear SKB has many advantages:
- Saves a memcpy of the headers.
- No page-boundary checks in datapath.
- No filler CQEs.
- Significantly smaller CQ.
- SKB data continuously resides in linear part, and not split to
small amount (linear part) and large amount (fragment).
This saves datapath cycles in driver and improves utilization
of SKB fragments in GRO.
- The fragments of a resulting GRO SKB follow the IP forwarding
assumption of equal-size fragments.
Some implementation details:
HW writes the packets to the beginning of a stride,
i.e. does not keep headroom. To overcome this we make sure we can
extend backwards and use the last bytes of stride i-1.
Extra care is needed for stride 0 as it has no preceding stride.
We make sure headroom bytes are available by shifting the buffer
pointer passed to HW by headroom bytes.
This configuration now becomes default, whenever capable.
Of course, this implies turning LRO off.
Performance testing:
ConnectX-5, single core, single RX ring, default MTU.
UDP packet rate, early drop in TC layer:
--------------------------------------------
| pkt size | before | after | ratio |
--------------------------------------------
| 1500byte | 4.65 Mpps | 5.96 Mpps | 1.28x |
| 500byte | 5.23 Mpps | 5.97 Mpps | 1.14x |
| 64byte | 5.94 Mpps | 5.96 Mpps | 1.00x |
--------------------------------------------
TCP streams: ~20% gain
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Before this patch, btrfs qgroup is mixing per-transcation meta rsv with
preallocated meta rsv, making it quite easy to underflow qgroup meta
reservation.
Since we have the new qgroup meta rsv types, apply it to delalloc
reservation.
Now for delalloc, most of its reserved space will use META_PREALLOC qgroup
rsv type.
And for callers reducing outstanding extent like btrfs_finish_ordered_io(),
they will convert corresponding META_PREALLOC reservation to
META_PERTRANS.
This is mainly due to the fact that current qgroup numbers will only be
updated in btrfs_commit_transaction(), that's to say if we don't keep
such placeholder reservation, we can exceed qgroup limitation.
And for callers freeing outstanding extent in error handler, we will
just free META_PREALLOC bytes.
This behavior makes callers of btrfs_qgroup_release_meta() or
btrfs_qgroup_convert_meta() to be aware of which type they are.
So in this patch, btrfs_delalloc_release_metadata() and its callers get
an extra parameter to info qgroup to do correct meta convert/release.
The good news is, even we use the wrong type (convert or free), it won't
cause obvious bug, as prealloc type is always in good shape, and the
type only affects how per-trans meta is increased or not.
So the worst case will be at most metadata limitation can be sometimes
exceeded (no convert at all) or metadata limitation is reached too soon
(no free at all).
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
For meta_prealloc reservation users, after btrfs_join_transaction()
caller will modify tree so part (or even all) meta_prealloc reservation
should be converted to meta_pertrans until transaction commit time.
This patch introduces a new function,
btrfs_qgroup_convert_reserved_meta() to do this for META_PREALLOC
reservation user.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since qgroup has seperate metadata reservation types now, we can
completely get rid of the old root->qgroup_meta_rsv, which mostly acts
as current META_PERTRANS reservation type.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Btrfs uses 2 different methods to reseve metadata qgroup space.
1) Reserve at btrfs_start_transaction() time
This is quite straightforward, caller will use the trans handler
allocated to modify b-trees.
In this case, reserved metadata should be kept until qgroup numbers
are updated.
2) Reserve by using block_rsv first, and later btrfs_join_transaction()
This is more complicated, caller will reserve space using block_rsv
first, and then later call btrfs_join_transaction() to get a trans
handle.
In this case, before we modify trees, the reserved space can be
modified on demand, and after btrfs_join_transaction(), such reserved
space should also be kept until qgroup numbers are updated.
Since these two types behave differently, split the original "META"
reservation type into 2 sub-types:
META_PERTRANS:
For above case 1)
META_PREALLOC:
For reservations that happened before btrfs_join_transaction() of
case 2)
NOTE: This patch will only convert existing qgroup meta reservation
callers according to its situation, not ensuring all callers are at
correct timing.
Such fix will be added in later patches.
Signed-off-by: Qu Wenruo <wqu@suse.com>
[ update comments ]
Signed-off-by: David Sterba <dsterba@suse.com>
When modifying qgroup relationship, for qgroup which only owns exclusive
extents, we will go through quick update path.
In this path, we will add/subtract exclusive and reference number for
parent qgroup, since the source (child) qgroup only has exclusive
extents, destination (parent) qgroup will also own or lose those extents
exclusively.
The same should be the same for reservation, since later reservation
adding/releasing will also affect parent qgroup, without the reservation
carried from child, parent will underflow reservation or have dead
reservation which will never be freed.
However original code doesn't do the same thing for reservation.
It handles qgroup reservation quite differently:
It removes qgroup reservation, as it's allocating space from the
reserved qgroup for relationship adding.
But does nothing for qgroup reservation if we're removing a qgroup
relationship.
According to the original code, it looks just like because we're adding
qgroup->rfer, the code assumes we're writing new data, so it's follows
the normal write routine, by reducing qgroup->reserved and adding
qgroup->rfer/excl.
This old behavior is wrong, and should be fixed to follow the same
excl/rfer behavior.
Just fix it by using the correct behavior described above.
Fixes: 31193213f1 ("Btrfs: qgroup: Introduce a may_use to account space_info->bytes_may_use.")
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since most callers of qgroup_reserve() are already defined by type,
converting qgroup_reserve() is quite an easy work.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Introduce helpers to:
1) Get total reserved space
For limit calculation
2) Add/release reserved space for given type
With underflow detection and warning
3) Add/release reserved space according to child qgroup
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Instead of single qgroup->reserved, use a new structure btrfs_qgroup_rsv
to store different types of reservation.
This patch only updates the header needed to compile.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
btrfs_orphan_add() has had this case commented out since it was first
introduced in commit d68fc57b7e ("Btrfs: Metadata reservation for
orphan inodes"). Most of the orphan cleanup code has been rewritten
since then, so it's safe to say that this code isn't needed.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
[ switch to bool ]
Signed-off-by: David Sterba <dsterba@suse.com>
Any time the first block group of a new type is created, we add a new
kobject to sysfs to hold the attributes for that type. Kobject-internal
allocations always use GFP_KERNEL, making them prone to fs-reclaim races.
While it appears as if this can occur any time a block group is created,
the only times the first block group of a new type can be created in
memory is at mount and when we create the first new block group during
raid conversion.
This patch adds a new list to track pending kobject additions and then
handles them after we do chunk relocation. Between relocating the
target chunk (or forcing allocation of a new chunk in the case of data)
and removing the old chunk, we're in a safe place for fs-reclaim to
occur. We're holding the volume mutex, which is already held across
page faults, and the delete_unused_bgs_mutex, which will only stall
the cleaner thread.
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Since commit 2be12ef79 (btrfs: Separate space_info create/update), we've
separated out the creation and updating of the space info structures.
That commit was a straightforward refactoring of the two parts of
update_space_info, but we can go a step further. Since commits
c59021f84 (Btrfs: fix OOPS of empty filesystem after balance) and
b742bb82f (Btrfs: Link block groups of different raid types), we know
that the space_info structures will be created at mount and there will
only ever be, at most, three of them.
This patch cleans out the create_space_info calls after __find_space_info
returns NULL since __find_space_info *can't* return NULL.
The initial cause for reviewing this was the kobject_add calls from
create_space_info occuring in sites where fs-reclaim wasn't allowed. Now
we are certain they occur only early in the mount process and are safe.
Signed-off-by: Jeff Mahoney <jeffm@suse.com>
Reviewed-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>
