lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:   Fri, 31 May 2019 05:26:30 +0000
From:   Song Liu <songliubraving@...com>
To:     Matt Mullins <mmullins@...com>
CC:     "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "daniel@...earbox.net" <daniel@...earbox.net>,
        "bpf@...r.kernel.org" <bpf@...r.kernel.org>,
        "ast@...nel.org" <ast@...nel.org>,
        "rostedt@...dmis.org" <rostedt@...dmis.org>,
        Andrew Hall <hall@...com>,
        "mingo@...hat.com" <mingo@...hat.com>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        Martin Lau <kafai@...com>, Yonghong Song <yhs@...com>
Subject: Re: [PATCH bpf] bpf: preallocate a perf_sample_data per event fd



> On May 30, 2019, at 5:01 PM, Matt Mullins <mmullins@...com> wrote:
> 
> On Thu, 2019-05-30 at 23:28 +0000, Song Liu wrote:
>>> On May 30, 2019, at 3:55 PM, Matt Mullins <mmullins@...com> wrote:
>>> 
>>> It is possible that a BPF program can be called while another BPF
>>> program is executing bpf_perf_event_output.  This has been observed with
>>> I/O completion occurring as a result of an interrupt:
>>> 
>>> 	bpf_prog_247fd1341cddaea4_trace_req_end+0x8d7/0x1000
>>> 	? trace_call_bpf+0x82/0x100
>>> 	? sch_direct_xmit+0xe2/0x230
>>> 	? blk_mq_end_request+0x1/0x100
>>> 	? blk_mq_end_request+0x5/0x100
>>> 	? kprobe_perf_func+0x19b/0x240
>>> 	? __qdisc_run+0x86/0x520
>>> 	? blk_mq_end_request+0x1/0x100
>>> 	? blk_mq_end_request+0x5/0x100
>>> 	? kprobe_ftrace_handler+0x90/0xf0
>>> 	? ftrace_ops_assist_func+0x6e/0xe0
>>> 	? ip6_input_finish+0xbf/0x460
>>> 	? 0xffffffffa01e80bf
>>> 	? nbd_dbg_flags_show+0xc0/0xc0 [nbd]
>>> 	? blkdev_issue_zeroout+0x200/0x200
>>> 	? blk_mq_end_request+0x1/0x100
>>> 	? blk_mq_end_request+0x5/0x100
>>> 	? flush_smp_call_function_queue+0x6c/0xe0
>>> 	? smp_call_function_single_interrupt+0x32/0xc0
>>> 	? call_function_single_interrupt+0xf/0x20
>>> 	? call_function_single_interrupt+0xa/0x20
>>> 	? swiotlb_map_page+0x140/0x140
>>> 	? refcount_sub_and_test+0x1a/0x50
>>> 	? tcp_wfree+0x20/0xf0
>>> 	? skb_release_head_state+0x62/0xc0
>>> 	? skb_release_all+0xe/0x30
>>> 	? napi_consume_skb+0xb5/0x100
>>> 	? mlx5e_poll_tx_cq+0x1df/0x4e0
>>> 	? mlx5e_poll_tx_cq+0x38c/0x4e0
>>> 	? mlx5e_napi_poll+0x58/0xc30
>>> 	? mlx5e_napi_poll+0x232/0xc30
>>> 	? net_rx_action+0x128/0x340
>>> 	? __do_softirq+0xd4/0x2ad
>>> 	? irq_exit+0xa5/0xb0
>>> 	? do_IRQ+0x7d/0xc0
>>> 	? common_interrupt+0xf/0xf
>>> 	</IRQ>
>>> 	? __rb_free_aux+0xf0/0xf0
>>> 	? perf_output_sample+0x28/0x7b0
>>> 	? perf_prepare_sample+0x54/0x4a0
>>> 	? perf_event_output+0x43/0x60
>>> 	? bpf_perf_event_output_raw_tp+0x15f/0x180
>>> 	? blk_mq_start_request+0x1/0x120
>>> 	? bpf_prog_411a64a706fc6044_should_trace+0xad4/0x1000
>>> 	? bpf_trace_run3+0x2c/0x80
>>> 	? nbd_send_cmd+0x4c2/0x690 [nbd]
>>> 
>>> This also cannot be alleviated by further splitting the per-cpu
>>> perf_sample_data structs (as in commit 283ca526a9bd ("bpf: fix
>>> corruption on concurrent perf_event_output calls")), as a raw_tp could
>>> be attached to the block:block_rq_complete tracepoint and execute during
>>> another raw_tp.  Instead, keep a pre-allocated perf_sample_data
>>> structure per perf_event_array element and fail a bpf_perf_event_output
>>> if that element is concurrently being used.
>>> 
>>> Fixes: 20b9d7ac4852 ("bpf: avoid excessive stack usage for perf_sample_data")
>>> Signed-off-by: Matt Mullins <mmullins@...com>
>>> ---
>>> It felt a bit overkill, but I had to split bpf_event_entry into its own
>>> header file to break an include cycle from perf_event.h -> cgroup.h ->
>>> cgroup-defs.h -> bpf-cgroup.h -> bpf.h -> (potentially) perf_event.h.
>>> 
>>> include/linux/bpf.h       |  7 -------
>>> include/linux/bpf_event.h | 20 ++++++++++++++++++++
>>> kernel/bpf/arraymap.c     |  2 ++
>>> kernel/trace/bpf_trace.c  | 30 +++++++++++++++++-------------
>>> 4 files changed, 39 insertions(+), 20 deletions(-)
>>> create mode 100644 include/linux/bpf_event.h
>>> 
>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>> index 4fb3aa2dc975..13b253a36402 100644
>>> --- a/include/linux/bpf.h
>>> +++ b/include/linux/bpf.h
>>> @@ -467,13 +467,6 @@ static inline bool bpf_map_flags_access_ok(u32 access_flags)
>>> 	       (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG);
>>> }
>>> 
>> 
>> I think we can avoid the include cycle as:
>> 
>> +struct perf_sample_data *sd;
>> struct bpf_event_entry {
>> 	struct perf_event *event;
>> 	struct file *perf_file;
>> 	struct file *map_file;
>> 	struct rcu_head rcu;
>> +	struct perf_sample_data *sd;
>> };
> 
> Yeah, that totally works.  I was mostly doing this so we had only one
> kmalloc allocation, but I'm not too worried about having an extra
> object in kmalloc-64 if it simplifies the code a lot.

We can also do something like

   ee = kzalloc(sizeof(struct bpf_event_entry) + sizeof(struct perf_sample_data));
   ee->sd = (void *)ee + sizeof(struct bpf_event_entry);

Thanks,
Song

> 
>> 
>>> -struct bpf_event_entry {
>>> -	struct perf_event *event;
>>> -	struct file *perf_file;
>>> -	struct file *map_file;
>>> -	struct rcu_head rcu;
>>> -};
>>> -
>>> bool bpf_prog_array_compatible(struct bpf_array *array, const struct bpf_prog *fp);
>>> int bpf_prog_calc_tag(struct bpf_prog *fp);
>>> 
>>> diff --git a/include/linux/bpf_event.h b/include/linux/bpf_event.h
>>> new file mode 100644
>>> index 000000000000..9f415990f921
>>> --- /dev/null
>>> +++ b/include/linux/bpf_event.h
>>> @@ -0,0 +1,20 @@
>>> +/* SPDX-License-Identifier: GPL-2.0 */
>>> +
>>> +#ifndef _LINUX_BPF_EVENT_H
>>> +#define _LINUX_BPF_EVENT_H
>>> +
>>> +#include <linux/perf_event.h>
>>> +#include <linux/types.h>
>>> +
>>> +struct file;
>>> +
>>> +struct bpf_event_entry {
>>> +	struct perf_event *event;
>>> +	struct file *perf_file;
>>> +	struct file *map_file;
>>> +	struct rcu_head rcu;
>>> +	struct perf_sample_data sd;
>>> +	atomic_t in_use;
>>> +};
>>> +
>>> +#endif /* _LINUX_BPF_EVENT_H */
>>> diff --git a/kernel/bpf/arraymap.c b/kernel/bpf/arraymap.c
>>> index 584636c9e2eb..08e5e486d563 100644
>>> --- a/kernel/bpf/arraymap.c
>>> +++ b/kernel/bpf/arraymap.c
>>> @@ -11,6 +11,7 @@
>>> * General Public License for more details.
>>> */
>>> #include <linux/bpf.h>
>>> +#include <linux/bpf_event.h>
>>> #include <linux/btf.h>
>>> #include <linux/err.h>
>>> #include <linux/slab.h>
>>> @@ -659,6 +660,7 @@ static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
>>> 		ee->event = perf_file->private_data;
>>> 		ee->perf_file = perf_file;
>>> 		ee->map_file = map_file;
>> 
>> And do the kzalloc() or some other trick here. 
>> 
>>> +		atomic_set(&ee->in_use, 0);
>>> 	}
>>> 
>>> 	return ee;
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index f92d6ad5e080..a03e29957698 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -6,6 +6,7 @@
>>> #include <linux/types.h>
>>> #include <linux/slab.h>
>>> #include <linux/bpf.h>
>>> +#include <linux/bpf_event.h>
>>> #include <linux/bpf_perf_event.h>
>>> #include <linux/filter.h>
>>> #include <linux/uaccess.h>
>>> @@ -410,17 +411,17 @@ static const struct bpf_func_proto bpf_perf_event_read_value_proto = {
>>> 	.arg4_type	= ARG_CONST_SIZE,
>>> };
>>> 
>>> -static DEFINE_PER_CPU(struct perf_sample_data, bpf_trace_sd);
>>> -
>>> static __always_inline u64
>>> __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
>>> -			u64 flags, struct perf_sample_data *sd)
>>> +			u64 flags, struct perf_raw_record *raw)
>>> {
>>> 	struct bpf_array *array = container_of(map, struct bpf_array, map);
>>> 	unsigned int cpu = smp_processor_id();
>>> 	u64 index = flags & BPF_F_INDEX_MASK;
>>> 	struct bpf_event_entry *ee;
>>> 	struct perf_event *event;
>>> +	struct perf_sample_data *sd;
>>> +	u64 ret;
>>> 
>>> 	if (index == BPF_F_CURRENT_CPU)
>>> 		index = cpu;
>>> @@ -439,13 +440,22 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
>>> 	if (unlikely(event->oncpu != cpu))
>>> 		return -EOPNOTSUPP;
>>> 
>>> -	return perf_event_output(event, sd, regs);
>>> +	if (atomic_cmpxchg(&ee->in_use, 0, 1) != 0)
>>> +		return -EBUSY;
>> 
>> And we only need xchg() here, so we can eliminate in_use. 
>> 
>> Does this make sense?
> 
> You mean xchg a null-pointer or something in there while it's in-use,
> then xchg the slab back?  Makes sense to me.  I'll try that and see
> where it gets me.
> 
>> 
>> Thanks,
>> Song
>> 
>>> +
>>> +	sd = &ee->sd;
>>> +	perf_sample_data_init(sd, 0, 0);
>>> +	sd->raw = raw;
>>> +
>>> +	ret = perf_event_output(event, sd, regs);
>>> +
>>> +	atomic_set(&ee->in_use, 0);
>>> +	return ret;
>>> }
>>> 
>>> BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
>>> 	   u64, flags, void *, data, u64, size)
>>> {
>>> -	struct perf_sample_data *sd = this_cpu_ptr(&bpf_trace_sd);
>>> 	struct perf_raw_record raw = {
>>> 		.frag = {
>>> 			.size = size,
>>> @@ -456,10 +466,8 @@ BPF_CALL_5(bpf_perf_event_output, struct pt_regs *, regs, struct bpf_map *, map,
>>> 	if (unlikely(flags & ~(BPF_F_INDEX_MASK)))
>>> 		return -EINVAL;
>>> 
>>> -	perf_sample_data_init(sd, 0, 0);
>>> -	sd->raw = &raw;
>>> 
>>> -	return __bpf_perf_event_output(regs, map, flags, sd);
>>> +	return __bpf_perf_event_output(regs, map, flags, &raw);
>>> }
>>> 
>>> static const struct bpf_func_proto bpf_perf_event_output_proto = {
>>> @@ -474,12 +482,10 @@ static const struct bpf_func_proto bpf_perf_event_output_proto = {
>>> };
>>> 
>>> static DEFINE_PER_CPU(struct pt_regs, bpf_pt_regs);
>>> -static DEFINE_PER_CPU(struct perf_sample_data, bpf_misc_sd);
>>> 
>>> u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
>>> 		     void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
>>> {
>>> -	struct perf_sample_data *sd = this_cpu_ptr(&bpf_misc_sd);
>>> 	struct pt_regs *regs = this_cpu_ptr(&bpf_pt_regs);
>>> 	struct perf_raw_frag frag = {
>>> 		.copy		= ctx_copy,
>>> @@ -497,10 +503,8 @@ u64 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
>>> 	};
>>> 
>>> 	perf_fetch_caller_regs(regs);
>>> -	perf_sample_data_init(sd, 0, 0);
>>> -	sd->raw = &raw;
>>> 
>>> -	return __bpf_perf_event_output(regs, map, flags, sd);
>>> +	return __bpf_perf_event_output(regs, map, flags, &raw);
>>> }
>>> 
>>> BPF_CALL_0(bpf_get_current_task)
>>> -- 
>>> 2.17.1
>>> 
>> 
>> 

Powered by blists - more mailing lists