Jann Horn reported following analysis that could potentially result
in a very hard to trigger (if not impossible) UAF race, to quote his
event timeline:

 - Set up a process with threads T1, T2 and T3
 - Let T1 set up a socket filter F1 that invokes another filter F2
   through a BPF map [tail call]
 - Let T1 trigger the socket filter via a unix domain socket write,
   don't wait for completion
 - Let T2 call PERF_EVENT_IOC_SET_BPF with F2, don't wait for completion
 - Now T2 should be behind bpf_prog_get(), but before bpf_prog_put()
 - Let T3 close the file descriptor for F2, dropping the reference
   count of F2 to 2
 - At this point, T1 should have looked up F2 from the map, but not
   finished executing it
 - Let T3 remove F2 from the BPF map, dropping the reference count of
   F2 to 1
 - Now T2 should call bpf_prog_put() (wrong BPF program type), dropping
   the reference count of F2 to 0 and scheduling bpf_prog_free_deferred()
   via schedule_work()
 - At this point, the BPF program could be freed
 - BPF execution is still running in a freed BPF program

While at PERF_EVENT_IOC_SET_BPF time it's only guaranteed that the perf
event fd we're doing the syscall on doesn't disappear from underneath us
for whole syscall time, it may not be the case for the bpf fd used as
an argument only after we did the put. It needs to be a valid fd pointing
to a BPF program at the time of the call to make the bpf_prog_get() and
while T2 gets preempted, F2 must have dropped reference to 1 on the other
CPU. The fput() from the close() in T3 should also add additionally delay
to the reference drop via exit_task_work() when bpf_prog_release() gets
called as well as scheduling bpf_prog_free_deferred().

That said, it makes nevertheless sense to move the BPF prog destruction
generally after RCU grace period to guarantee that such scenario above,
but also others as recently fixed in ceb56070 ("bpf, perf: delay release
of BPF prog after grace period") with regards to tail calls won't happen.
Integrating bpf_prog_free_deferred() directly into the RCU callback is
not allowed since the invocation might happen from either softirq or
process context, so we're not permitted to block. Reviewing all bpf_prog_put()
invocations from eBPF side (note, cBPF -> eBPF progs don't use this for
their destruction) with call_rcu() look good to me.

Since we don't know whether at the time of attaching the program, we're
already part of a tail call map, we need to use RCU variant. However, due
to this, there won't be severely more stress on the RCU callback queue:
situations with above bpf_prog_get() and bpf_prog_put() combo in practice
normally won't lead to releases, but even if they would, enough effort/
cycles have to be put into loading a BPF program into the kernel already.
Reported-by: NJann Horn <jannh@google.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The behavior of perf event arrays are quite different from all
others as they are tightly coupled to perf event fds, f.e. shown
recently by commit e03e7ee3 ("perf/bpf: Convert perf_event_array
to use struct file") to make refcounting on perf event more robust.
A remaining issue that the current code still has is that since
additions to the perf event array take a reference on the struct
file via perf_event_get() and are only released via fput() (that
cleans up the perf event eventually via perf_event_release_kernel())
when the element is either manually removed from the map from user
space or automatically when the last reference on the perf event
map is dropped. However, this leads us to dangling struct file's
when the map gets pinned after the application owning the perf
event descriptor exits, and since the struct file reference will
in such case only be manually dropped or via pinned file removal,
it leads to the perf event living longer than necessary, consuming
needlessly resources for that time.

Relations between perf event fds and bpf perf event map fds can be
rather complex. F.e. maps can act as demuxers among different perf
event fds that can possibly be owned by different threads and based
on the index selection from the program, events get dispatched to
one of the per-cpu fd endpoints. One perf event fd (or, rather a
per-cpu set of them) can also live in multiple perf event maps at
the same time, listening for events. Also, another requirement is
that perf event fds can get closed from application side after they
have been attached to the perf event map, so that on exit perf event
map will take care of dropping their references eventually. Likewise,
when such maps are pinned, the intended behavior is that a user
application does bpf_obj_get(), puts its fds in there and on exit
when fd is released, they are dropped from the map again, so the map
acts rather as connector endpoint. This also makes perf event maps
inherently different from program arrays as described in more detail
in commit c9da161c ("bpf: fix clearing on persistent program
array maps").

To tackle this, map entries are marked by the map struct file that
added the element to the map. And when the last reference to that map
struct file is released from user space, then the tracked entries
are purged from the map. This is okay, because new map struct files
instances resp. frontends to the anon inode are provided via
bpf_map_new_fd() that is called when we invoke bpf_obj_get_user()
for retrieving a pinned map, but also when an initial instance is
created via map_create(). The rest is resolved by the vfs layer
automatically for us by keeping reference count on the map's struct
file. Any concurrent updates on the map slot are fine as well, it
just means that perf_event_fd_array_release() needs to delete less
of its own entires.
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch extends map_fd_get_ptr() callback that is used by fd array
maps, so that struct file pointer from the related map can be passed
in. It's safe to remove map_update_elem() callback for the two maps since
this is only allowed from syscall side, but not from eBPF programs for these
two map types. Like in per-cpu map case, bpf_fd_array_map_update_elem()
needs to be called directly here due to the extra argument.
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Suggested-by: NDaniel Borkmann <daniel@iogearbox.net>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The functions bpf_map_lookup_elem(map, key, value) and
bpf_map_update_elem(map, key, value, flags) need to get/set
values from all-cpus for per-cpu hash and array maps,
so that user space can aggregate/update them as necessary.

Example of single counter aggregation in user space:
  unsigned int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
  long values[nr_cpus];
  long value = 0;

  bpf_lookup_elem(fd, key, values);
  for (i = 0; i < nr_cpus; i++)
    value += values[i];

The user space must provide round_up(value_size, 8) * nr_cpus
array to get/set values, since kernel will use 'long' copy
of per-cpu values to try to copy good counters atomically.
It's a best-effort, since bpf programs and user space are racing
to access the same memory.
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Primary use case is a histogram array of latency
where bpf program computes the latency of block requests or other
events and stores histogram of latency into array of 64 elements.
All cpus are constantly running, so normal increment is not accurate,
bpf_xadd causes cache ping-pong and this per-cpu approach allows
fastest collision-free counters.
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Robustify refcounting.
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: Wang Nan <wangnan0@huawei.com>
Cc: vince@deater.net
Link: http://lkml.kernel.org/r/20160126045947.GA40151@ast-mbp.thefacebook.comSigned-off-by: NIngo Molnar <mingo@kernel.org>

For large map->value_size the user space can trigger memory allocation warnings like:
WARNING: CPU: 2 PID: 11122 at mm/page_alloc.c:2989
__alloc_pages_nodemask+0x695/0x14e0()
Call Trace:
 [<     inline     >] __dump_stack lib/dump_stack.c:15
 [<ffffffff82743b56>] dump_stack+0x68/0x92 lib/dump_stack.c:50
 [<ffffffff81244ec9>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:460
 [<ffffffff812450f9>] warn_slowpath_null+0x29/0x30 kernel/panic.c:493
 [<     inline     >] __alloc_pages_slowpath mm/page_alloc.c:2989
 [<ffffffff81554e95>] __alloc_pages_nodemask+0x695/0x14e0 mm/page_alloc.c:3235
 [<ffffffff816188fe>] alloc_pages_current+0xee/0x340 mm/mempolicy.c:2055
 [<     inline     >] alloc_pages include/linux/gfp.h:451
 [<ffffffff81550706>] alloc_kmem_pages+0x16/0xf0 mm/page_alloc.c:3414
 [<ffffffff815a1c89>] kmalloc_order+0x19/0x60 mm/slab_common.c:1007
 [<ffffffff815a1cef>] kmalloc_order_trace+0x1f/0xa0 mm/slab_common.c:1018
 [<     inline     >] kmalloc_large include/linux/slab.h:390
 [<ffffffff81627784>] __kmalloc+0x234/0x250 mm/slub.c:3525
 [<     inline     >] kmalloc include/linux/slab.h:463
 [<     inline     >] map_update_elem kernel/bpf/syscall.c:288
 [<     inline     >] SYSC_bpf kernel/bpf/syscall.c:744

To avoid never succeeding kmalloc with order >= MAX_ORDER check that
elem->value_size and computed elem_size are within limits for both hash and
array type maps.
Also add __GFP_NOWARN to kmalloc(value_size | elem_size) to avoid OOM warnings.
Note kmalloc(key_size) is highly unlikely to trigger OOM, since key_size <= 512,
so keep those kmalloc-s as-is.

Large value_size can cause integer overflows in elem_size and map.pages
formulas, so check for that as well.

Fixes: aaac3ba9 ("bpf: charge user for creation of BPF maps and programs")
Reported-by: NDmitry Vyukov <dvyukov@google.com>
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

During own review but also reported by Dmitry's syzkaller [1] it has been
noticed that we trigger a heap out-of-bounds access on eBPF array maps
when updating elements. This happens with each map whose map->value_size
(specified during map creation time) is not multiple of 8 bytes.

In array_map_alloc(), elem_size is round_up(attr->value_size, 8) and
used to align array map slots for faster access. However, in function
array_map_update_elem(), we update the element as ...

memcpy(array->value + array->elem_size * index, value, array->elem_size);

... where we access 'value' out-of-bounds, since it was allocated from
map_update_elem() from syscall side as kmalloc(map->value_size, GFP_USER)
and later on copied through copy_from_user(value, uvalue, map->value_size).
Thus, up to 7 bytes, we can access out-of-bounds.

Same could happen from within an eBPF program, where in worst case we
access beyond an eBPF program's designated stack.

Since 1be7f75d ("bpf: enable non-root eBPF programs") didn't hit an
official release yet, it only affects priviledged users.

In case of array_map_lookup_elem(), the verifier prevents eBPF programs
from accessing beyond map->value_size through check_map_access(). Also
from syscall side map_lookup_elem() only copies map->value_size back to
user, so nothing could leak.

  [1] http://github.com/google/syzkaller

Fixes: 28fbcfa0 ("bpf: add array type of eBPF maps")
Reported-by: NDmitry Vyukov <dvyukov@google.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Fix safety checks for bpf_perf_event_read():
- only non-inherited events can be added to perf_event_array map
  (do this check statically at map insertion time)
- dynamically check that event is local and !pmu->count
Otherwise buggy bpf program can cause kernel splat.

Also fix error path after perf_event_attrs()
and remove redundant 'extern'.

Fixes: 35578d79 ("bpf: Implement function bpf_perf_event_read() that get the selected hardware PMU conuter")
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Tested-by: NWang Nan <wangnan0@huawei.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This helper is used to send raw data from eBPF program into
special PERF_TYPE_SOFTWARE/PERF_COUNT_SW_BPF_OUTPUT perf_event.
User space needs to perf_event_open() it (either for one or all cpus) and
store FD into perf_event_array (similar to bpf_perf_event_read() helper)
before eBPF program can send data into it.

Today the programs triggered by kprobe collect the data and either store
it into the maps or print it via bpf_trace_printk() where latter is the debug
facility and not suitable to stream the data. This new helper replaces
such bpf_trace_printk() usage and allows programs to have dedicated
channel into user space for post-processing of the raw data collected.
Signed-off-by: NAlexei Starovoitov <ast@kernel.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

since eBPF programs and maps use kernel memory consider it 'locked' memory
from user accounting point of view and charge it against RLIMIT_MEMLOCK limit.
This limit is typically set to 64Kbytes by distros, so almost all
bpf+tracing programs would need to increase it, since they use maps,
but kernel charges maximum map size upfront.
For example the hash map of 1024 elements will be charged as 64Kbyte.
It's inconvenient for current users and changes current behavior for root,
but probably worth doing to be consistent root vs non-root.

Similar accounting logic is done by mmap of perf_event.
Signed-off-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Commit ea317b26 ("bpf: Add new bpf map type to store the pointer
to struct perf_event") added perf_event.h to the main eBPF header, so
it gets included for all users. perf_event.h is actually only needed
from array map side, so lets sanitize this a bit.
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Cc: Kaixu Xia <xiakaixu@huawei.com>
Acked-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Introduce a new bpf map type 'BPF_MAP_TYPE_PERF_EVENT_ARRAY'.
This map only stores the pointer to struct perf_event. The
user space event FDs from perf_event_open() syscall are converted
to the pointer to struct perf_event and stored in map.
Signed-off-by: NKaixu Xia <xiakaixu@huawei.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

All the map backends are of generic nature. In order to avoid
adding much special code into the eBPF core, rewrite part of
the bpf_prog_array map code and make it more generic. So the
new perf_event_array map type can reuse most of code with
bpf_prog_array map and add fewer lines of special code.
Signed-off-by: NWang Nan <wangnan0@huawei.com>
Signed-off-by: NKaixu Xia <xiakaixu@huawei.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Normally the program attachment place (like sockets, qdiscs) takes
care of rcu protection and calls bpf_prog_put() after a grace period.
The programs stored inside prog_array may not be attached anywhere,
so prog_array needs to take care of preserving rcu protection.
Otherwise bpf_tail_call() will race with bpf_prog_put().
To solve that introduce bpf_prog_put_rcu() helper function and use
it in 3 places where unattached program can decrement refcnt:
closing program fd, deleting/replacing program in prog_array.

Fixes: 04fd61ab ("bpf: allow bpf programs to tail-call other bpf programs")
Reported-by: NMartin Schwidefsky <schwidefsky@de.ibm.com>
Signed-off-by: NAlexei Starovoitov <ast@plumgrid.com>
Acked-by: NDaniel Borkmann <daniel@iogearbox.net>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

introduce bpf_tail_call(ctx, &jmp_table, index) helper function
which can be used from BPF programs like:
int bpf_prog(struct pt_regs *ctx)
{
  ...
  bpf_tail_call(ctx, &jmp_table, index);
  ...
}
that is roughly equivalent to:
int bpf_prog(struct pt_regs *ctx)
{
  ...
  if (jmp_table[index])
    return (*jmp_table[index])(ctx);
  ...
}
The important detail that it's not a normal call, but a tail call.
The kernel stack is precious, so this helper reuses the current
stack frame and jumps into another BPF program without adding
extra call frame.
It's trivially done in interpreter and a bit trickier in JITs.
In case of x64 JIT the bigger part of generated assembler prologue
is common for all programs, so it is simply skipped while jumping.
Other JITs can do similar prologue-skipping optimization or
do stack unwind before jumping into the next program.

bpf_tail_call() arguments:
ctx - context pointer
jmp_table - one of BPF_MAP_TYPE_PROG_ARRAY maps used as the jump table
index - index in the jump table

Since all BPF programs are idenitified by file descriptor, user space
need to populate the jmp_table with FDs of other BPF programs.
If jmp_table[index] is empty the bpf_tail_call() doesn't jump anywhere
and program execution continues as normal.

New BPF_MAP_TYPE_PROG_ARRAY map type is introduced so that user space can
populate this jmp_table array with FDs of other bpf programs.
Programs can share the same jmp_table array or use multiple jmp_tables.

The chain of tail calls can form unpredictable dynamic loops therefore
tail_call_cnt is used to limit the number of calls and currently is set to 32.

Use cases:
Acked-by: NDaniel Borkmann <daniel@iogearbox.net>

==========
- simplify complex programs by splitting them into a sequence of small programs

- dispatch routine
  For tracing and future seccomp the program may be triggered on all system
  calls, but processing of syscall arguments will be different. It's more
  efficient to implement them as:
  int syscall_entry(struct seccomp_data *ctx)
  {
     bpf_tail_call(ctx, &syscall_jmp_table, ctx->nr /* syscall number */);
     ... default: process unknown syscall ...
  }
  int sys_write_event(struct seccomp_data *ctx) {...}
  int sys_read_event(struct seccomp_data *ctx) {...}
  syscall_jmp_table[__NR_write] = sys_write_event;
  syscall_jmp_table[__NR_read] = sys_read_event;

  For networking the program may call into different parsers depending on
  packet format, like:
  int packet_parser(struct __sk_buff *skb)
  {
     ... parse L2, L3 here ...
     __u8 ipproto = load_byte(skb, ... offsetof(struct iphdr, protocol));
     bpf_tail_call(skb, &ipproto_jmp_table, ipproto);
     ... default: process unknown protocol ...
  }
  int parse_tcp(struct __sk_buff *skb) {...}
  int parse_udp(struct __sk_buff *skb) {...}
  ipproto_jmp_table[IPPROTO_TCP] = parse_tcp;
  ipproto_jmp_table[IPPROTO_UDP] = parse_udp;

- for TC use case, bpf_tail_call() allows to implement reclassify-like logic

- bpf_map_update_elem/delete calls into BPF_MAP_TYPE_PROG_ARRAY jump table
  are atomic, so user space can build chains of BPF programs on the fly

Implementation details:
=======================
- high performance of bpf_tail_call() is the goal.
  It could have been implemented without JIT changes as a wrapper on top of
  BPF_PROG_RUN() macro, but with two downsides:
  . all programs would have to pay performance penalty for this feature and
    tail call itself would be slower, since mandatory stack unwind, return,
    stack allocate would be done for every tailcall.
  . tailcall would be limited to programs running preempt_disabled, since
    generic 'void *ctx' doesn't have room for 'tail_call_cnt' and it would
    need to be either global per_cpu variable accessed by helper and by wrapper
    or global variable protected by locks.

  In this implementation x64 JIT bypasses stack unwind and jumps into the
  callee program after prologue.

- bpf_prog_array_compatible() ensures that prog_type of callee and caller
  are the same and JITed/non-JITed flag is the same, since calling JITed
  program from non-JITed is invalid, since stack frames are different.
  Similarly calling kprobe type program from socket type program is invalid.

- jump table is implemented as BPF_MAP_TYPE_PROG_ARRAY to reuse 'map'
  abstraction, its user space API and all of verifier logic.
  It's in the existing arraymap.c file, since several functions are
  shared with regular array map.
Signed-off-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We can move bpf_map_ops and bpf_verifier_ops and other structs into ro
section, bpf_map_type_list and bpf_prog_type_list into read mostly.
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

- fix NULL pointer dereference:
kernel/bpf/arraymap.c:41 array_map_alloc() error: potential null dereference 'array'.  (kzalloc returns null)
kernel/bpf/arraymap.c:41 array_map_alloc() error: we previously assumed 'array' could be null (see line 40)

- integer overflow check was missing in arraymap
(hashmap checks for overflow via kmalloc_array())

- arraymap can round_up(value_size, 8) to zero. check was missing.

- hashmap was missing zero size check as well, since roundup_pow_of_two() can
truncate into zero

- found a typo in the arraymap comment and unnecessary empty line

Fix all of these issues and make both overflow checks explicit U32 in size.
Reported-by: Nkbuild test robot <fengguang.wu@intel.com>
Signed-off-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

add new map type BPF_MAP_TYPE_ARRAY and its implementation

- optimized for fastest possible lookup()
  . in the future verifier/JIT may recognize lookup() with constant key
    and optimize it into constant pointer. Can optimize non-constant
    key into direct pointer arithmetic as well, since pointers and
    value_size are constant for the life of the eBPF program.
    In other words array_map_lookup_elem() may be 'inlined' by verifier/JIT
    while preserving concurrent access to this map from user space

- two main use cases for array type:
  . 'global' eBPF variables: array of 1 element with key=0 and value is a
    collection of 'global' variables which programs can use to keep the state
    between events
  . aggregation of tracing events into fixed set of buckets

- all array elements pre-allocated and zero initialized at init time

- key as an index in array and can only be 4 byte

- map_delete_elem() returns EINVAL, since elements cannot be deleted

- map_update_elem() replaces elements in an non-atomic way
  (for atomic updates hashtable type should be used instead)
Signed-off-by: NAlexei Starovoitov <ast@plumgrid.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>