Message ID | 20171110090144.11791-1-ming.lei@redhat.com (mailing list archive) |
---|---|
State | Changes Requested |
Headers | show |
On Fri, 2017-11-10 at 17:01 +0800, Ming Lei wrote: > cmd->cmnd can be allocated/freed dynamically in case of > T10_PI_TYPE2_PROTECTION, > so we should check it in scsi_show_rq() because this request may have > been freed already here, and cmd->cmnd has been set as null. > > We choose to accept read-after-free and dump request data as far as > possible. > > This patch fixs the following kernel crash when dumping request via > block's > debugfs interface: > > [ 252.962045] BUG: unable to handle kernel NULL pointer dereference > at (null) > [ 252.963007] IP: scsi_format_opcode_name+0x1a/0x1c0 > [ 252.963007] PGD 25e75a067 P4D 25e75a067 PUD 25e75b067 PMD 0 > [ 252.963007] Oops: 0000 [#1] PREEMPT SMP > [ 252.963007] Dumping ftrace buffer: > [ 252.963007] (ftrace buffer empty) > [ 252.963007] Modules linked in: scsi_debug ebtable_filter ebtables > ip6table_filter ip6_tables xt_CHECKSUM iptable_mangle ipt_MASQUERADE > nf_nat_masquerade_ipv4 iptable_nat nf_conntrack_ipv4 nf_defrag_ipv4 > nf_nat_ipv4 nf_nat nf_conntrack libcrc32c bridge stp llc > iptable_filter fuse ip_tables sd_mod sg mptsas mptscsih mptbase > crc32c_intel ahci libahci nvme serio_raw scsi_transport_sas libata > lpc_ich nvme_core virtio_scsi binfmt_misc dm_mod iscsi_tcp > libiscsi_tcp libiscsi scsi_transport_iscsi null_blk configs > [ 252.963007] CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0- > rc2.blk_mq_io_hang+ #516 > [ 252.963007] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), > BIOS 1.9.3-1.fc25 04/01/2014 > [ 252.963007] task: ffff88025e6f6000 task.stack: ffffc90001bd0000 > [ 252.963007] RIP: 0010:scsi_format_opcode_name+0x1a/0x1c0 > [ 252.963007] RSP: 0018:ffffc90001bd3c50 EFLAGS: 00010286 > [ 252.963007] RAX: 00000000ffff4843 RBX: 0000000000000050 RCX: > 0000000000000000 > [ 252.963007] RDX: 0000000000000000 RSI: 0000000000000050 RDI: > ffffc90001bd3cd8 > [ 252.963007] RBP: ffffc90001bd3c88 R08: 0000000000001000 R09: > 0000000000000000 > [ 252.963007] R10: ffff880275134000 R11: ffff88027513406c R12: > 0000000000000050 > [ 252.963007] R13: ffffc90001bd3cd8 R14: 0000000000000000 R15: > 0000000000000000 > [ 252.963007] FS: 00007f4d11762700(0000) GS:ffff88027fc40000(0000) > knlGS:0000000000000000 > [ 252.963007] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > [ 252.963007] CR2: 0000000000000000 CR3: 000000025e789003 CR4: > 00000000003606e0 > [ 252.963007] DR0: 0000000000000000 DR1: 0000000000000000 DR2: > 0000000000000000 > [ 252.963007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: > 0000000000000400 > [ 252.963007] Call Trace: > [ 252.963007] __scsi_format_command+0x27/0xc0 > [ 252.963007] scsi_show_rq+0x5c/0xc0 > [ 252.963007] ? seq_printf+0x4e/0x70 > [ 252.963007] ? blk_flags_show+0x5b/0xf0 > [ 252.963007] __blk_mq_debugfs_rq_show+0x116/0x130 > [ 252.963007] blk_mq_debugfs_rq_show+0xe/0x10 > [ 252.963007] seq_read+0xfe/0x3b0 > [ 252.963007] ? __handle_mm_fault+0x631/0x1150 > [ 252.963007] full_proxy_read+0x54/0x90 > [ 252.963007] __vfs_read+0x37/0x160 > [ 252.963007] ? security_file_permission+0x9b/0xc0 > [ 252.963007] vfs_read+0x96/0x130 > [ 252.963007] SyS_read+0x55/0xc0 > [ 252.963007] entry_SYSCALL_64_fastpath+0x1a/0xa5 > [ 252.963007] RIP: 0033:0x7f4d1127e9b0 > [ 252.963007] RSP: 002b:00007ffd27082568 EFLAGS: 00000246 ORIG_RAX: > 0000000000000000 > [ 252.963007] RAX: ffffffffffffffda RBX: 00007f4d1154bb20 RCX: > 00007f4d1127e9b0 > [ 252.963007] RDX: 0000000000020000 RSI: 00007f4d115a7000 RDI: > 0000000000000003 > [ 252.963007] RBP: 0000000000021010 R08: ffffffffffffffff R09: > 0000000000000000 > [ 252.963007] R10: 000000000000037b R11: 0000000000000246 R12: > 0000000000022000 > [ 252.963007] R13: 00007f4d1154bb78 R14: 0000000000001000 R15: > 0000000000020000 > [ 252.963007] Code: c6 e8 1b ca 24 00 eb 8c e8 74 2c ae ff 0f 1f 40 > 00 0f 1f 44 00 00 55 48 89 e5 41 56 41 55 41 54 53 49 89 fd 49 89 f4 > 48 83 ec 18 <44> 0f b6 32 48 c7 45 c8 00 00 00 00 65 48 8b 04 25 28 > 00 00 00 > [ 252.963007] RIP: scsi_format_opcode_name+0x1a/0x1c0 RSP: > ffffc90001bd3c50 > [ 252.963007] CR2: 0000000000000000 > [ 252.963007] ---[ end trace 83c5bddfbaa6573c ]--- > [ 252.963007] Kernel panic - not syncing: Fatal exception > [ 252.963007] Dumping ftrace buffer: > [ 252.963007] (ftrace buffer empty) > [ 252.963007] Kernel Offset: disabled > [ 252.963007] ---[ end Kernel panic - not syncing: Fatal exception > > Fixes: 0eebd005dd07(scsi: Implement blk_mq_ops.show_rq()) > Cc: Bart Van Assche <bart.vanassche@sandisk.com> > Cc: Omar Sandoval <osandov@fb.com> > Cc: Martin K. Petersen <martin.petersen@oracle.com> > Cc: James Bottomley <James.Bottomley@HansenPartnership.com> > Cc: Hannes Reinecke <hare@suse.com> > Signed-off-by: Ming Lei <ming.lei@redhat.com> > --- > V2: > - fix typo > V3: > - prefer to dump data and accept read-after-free > - add some comment > V4: > - read the two variable into local variable first, for avoiding > free between check and passing to __scsi_format_command(). > > > drivers/scsi/scsi_debugfs.c | 14 +++++++++++++- > 1 file changed, 13 insertions(+), 1 deletion(-) > > diff --git a/drivers/scsi/scsi_debugfs.c > b/drivers/scsi/scsi_debugfs.c > index 5e9755008aed..f8437c456458 100644 > --- a/drivers/scsi/scsi_debugfs.c > +++ b/drivers/scsi/scsi_debugfs.c > @@ -8,8 +8,20 @@ void scsi_show_rq(struct seq_file *m, struct request > *rq) > struct scsi_cmnd *cmd = container_of(scsi_req(rq), > typeof(*cmd), req); > int msecs = jiffies_to_msecs(jiffies - cmd- > >jiffies_at_alloc); > char buf[80]; > + unsigned char *cdb = cmd->cmnd; > + unsigned short cdb_len = cmd->cmd_len; > > > - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd- > >cmd_len); > + /* > + * This rq may have been freed, so don't be surprised if > + * read-after-free is reported. I said fix this in sd.c:sd_uninit_command() by setting SCpnt->cmnd to NULL before calling mempool_free(). James
On Fri, Nov 10, 2017 at 08:51:58AM -0800, James Bottomley wrote: > On Fri, 2017-11-10 at 17:01 +0800, Ming Lei wrote: > > cmd->cmnd can be allocated/freed dynamically in case of > > T10_PI_TYPE2_PROTECTION, > > so we should check it in scsi_show_rq() because this request may have > > been freed already here, and cmd->cmnd has been set as null. > > > > We choose to accept read-after-free and dump request data as far as > > possible. > > > > This patch fixs the following kernel crash when dumping request via > > block's > > debugfs interface: > > > > [ 252.962045] BUG: unable to handle kernel NULL pointer dereference > > at (null) > > [ 252.963007] IP: scsi_format_opcode_name+0x1a/0x1c0 > > [ 252.963007] PGD 25e75a067 P4D 25e75a067 PUD 25e75b067 PMD 0 > > [ 252.963007] Oops: 0000 [#1] PREEMPT SMP > > [ 252.963007] Dumping ftrace buffer: > > [ 252.963007] (ftrace buffer empty) > > [ 252.963007] Modules linked in: scsi_debug ebtable_filter ebtables > > ip6table_filter ip6_tables xt_CHECKSUM iptable_mangle ipt_MASQUERADE > > nf_nat_masquerade_ipv4 iptable_nat nf_conntrack_ipv4 nf_defrag_ipv4 > > nf_nat_ipv4 nf_nat nf_conntrack libcrc32c bridge stp llc > > iptable_filter fuse ip_tables sd_mod sg mptsas mptscsih mptbase > > crc32c_intel ahci libahci nvme serio_raw scsi_transport_sas libata > > lpc_ich nvme_core virtio_scsi binfmt_misc dm_mod iscsi_tcp > > libiscsi_tcp libiscsi scsi_transport_iscsi null_blk configs > > [ 252.963007] CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0- > > rc2.blk_mq_io_hang+ #516 > > [ 252.963007] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), > > BIOS 1.9.3-1.fc25 04/01/2014 > > [ 252.963007] task: ffff88025e6f6000 task.stack: ffffc90001bd0000 > > [ 252.963007] RIP: 0010:scsi_format_opcode_name+0x1a/0x1c0 > > [ 252.963007] RSP: 0018:ffffc90001bd3c50 EFLAGS: 00010286 > > [ 252.963007] RAX: 00000000ffff4843 RBX: 0000000000000050 RCX: > > 0000000000000000 > > [ 252.963007] RDX: 0000000000000000 RSI: 0000000000000050 RDI: > > ffffc90001bd3cd8 > > [ 252.963007] RBP: ffffc90001bd3c88 R08: 0000000000001000 R09: > > 0000000000000000 > > [ 252.963007] R10: ffff880275134000 R11: ffff88027513406c R12: > > 0000000000000050 > > [ 252.963007] R13: ffffc90001bd3cd8 R14: 0000000000000000 R15: > > 0000000000000000 > > [ 252.963007] FS: 00007f4d11762700(0000) GS:ffff88027fc40000(0000) > > knlGS:0000000000000000 > > [ 252.963007] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > > [ 252.963007] CR2: 0000000000000000 CR3: 000000025e789003 CR4: > > 00000000003606e0 > > [ 252.963007] DR0: 0000000000000000 DR1: 0000000000000000 DR2: > > 0000000000000000 > > [ 252.963007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: > > 0000000000000400 > > [ 252.963007] Call Trace: > > [ 252.963007] __scsi_format_command+0x27/0xc0 > > [ 252.963007] scsi_show_rq+0x5c/0xc0 > > [ 252.963007] ? seq_printf+0x4e/0x70 > > [ 252.963007] ? blk_flags_show+0x5b/0xf0 > > [ 252.963007] __blk_mq_debugfs_rq_show+0x116/0x130 > > [ 252.963007] blk_mq_debugfs_rq_show+0xe/0x10 > > [ 252.963007] seq_read+0xfe/0x3b0 > > [ 252.963007] ? __handle_mm_fault+0x631/0x1150 > > [ 252.963007] full_proxy_read+0x54/0x90 > > [ 252.963007] __vfs_read+0x37/0x160 > > [ 252.963007] ? security_file_permission+0x9b/0xc0 > > [ 252.963007] vfs_read+0x96/0x130 > > [ 252.963007] SyS_read+0x55/0xc0 > > [ 252.963007] entry_SYSCALL_64_fastpath+0x1a/0xa5 > > [ 252.963007] RIP: 0033:0x7f4d1127e9b0 > > [ 252.963007] RSP: 002b:00007ffd27082568 EFLAGS: 00000246 ORIG_RAX: > > 0000000000000000 > > [ 252.963007] RAX: ffffffffffffffda RBX: 00007f4d1154bb20 RCX: > > 00007f4d1127e9b0 > > [ 252.963007] RDX: 0000000000020000 RSI: 00007f4d115a7000 RDI: > > 0000000000000003 > > [ 252.963007] RBP: 0000000000021010 R08: ffffffffffffffff R09: > > 0000000000000000 > > [ 252.963007] R10: 000000000000037b R11: 0000000000000246 R12: > > 0000000000022000 > > [ 252.963007] R13: 00007f4d1154bb78 R14: 0000000000001000 R15: > > 0000000000020000 > > [ 252.963007] Code: c6 e8 1b ca 24 00 eb 8c e8 74 2c ae ff 0f 1f 40 > > 00 0f 1f 44 00 00 55 48 89 e5 41 56 41 55 41 54 53 49 89 fd 49 89 f4 > > 48 83 ec 18 <44> 0f b6 32 48 c7 45 c8 00 00 00 00 65 48 8b 04 25 28 > > 00 00 00 > > [ 252.963007] RIP: scsi_format_opcode_name+0x1a/0x1c0 RSP: > > ffffc90001bd3c50 > > [ 252.963007] CR2: 0000000000000000 > > [ 252.963007] ---[ end trace 83c5bddfbaa6573c ]--- > > [ 252.963007] Kernel panic - not syncing: Fatal exception > > [ 252.963007] Dumping ftrace buffer: > > [ 252.963007] (ftrace buffer empty) > > [ 252.963007] Kernel Offset: disabled > > [ 252.963007] ---[ end Kernel panic - not syncing: Fatal exception > > > > Fixes: 0eebd005dd07(scsi: Implement blk_mq_ops.show_rq()) > > Cc: Bart Van Assche <bart.vanassche@sandisk.com> > > Cc: Omar Sandoval <osandov@fb.com> > > Cc: Martin K. Petersen <martin.petersen@oracle.com> > > Cc: James Bottomley <James.Bottomley@HansenPartnership.com> > > Cc: Hannes Reinecke <hare@suse.com> > > Signed-off-by: Ming Lei <ming.lei@redhat.com> > > --- > > V2: > > - fix typo > > V3: > > - prefer to dump data and accept read-after-free > > - add some comment > > V4: > > - read the two variable into local variable first, for avoiding > > free between check and passing to __scsi_format_command(). > > > > > > drivers/scsi/scsi_debugfs.c | 14 +++++++++++++- > > 1 file changed, 13 insertions(+), 1 deletion(-) > > > > diff --git a/drivers/scsi/scsi_debugfs.c > > b/drivers/scsi/scsi_debugfs.c > > index 5e9755008aed..f8437c456458 100644 > > --- a/drivers/scsi/scsi_debugfs.c > > +++ b/drivers/scsi/scsi_debugfs.c > > @@ -8,8 +8,20 @@ void scsi_show_rq(struct seq_file *m, struct request > > *rq) > > struct scsi_cmnd *cmd = container_of(scsi_req(rq), > > typeof(*cmd), req); > > int msecs = jiffies_to_msecs(jiffies - cmd- > > >jiffies_at_alloc); > > char buf[80]; > > + unsigned char *cdb = cmd->cmnd; > > + unsigned short cdb_len = cmd->cmd_len; > > > > > > - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd- > > >cmd_len); > > + /* > > + * This rq may have been freed, so don't be surprised if > > + * read-after-free is reported. > > I said fix this in sd.c:sd_uninit_command() by setting SCpnt->cmnd to > NULL before calling mempool_free(). Hi James, That can't avoid the use-after-free in both reality and theory. When 'cdb = cmd->cmnd' is executed in CPU1, 'cbd' is a valid pointer, just after this instruction is done, this request is completed in CPU2, and mempool_free() is run to complete this buffer. Firstly, both preemption and IRQ can happen inside __scsi_format_command(), when this function is scheduled to rerun, this buffer is freed since mempool_free() has completed already, which is run in irq context. Secondly, the use-after-free can't be avoided too if we disable irq when checking and reading the variable. See following: mempool_free(cmd->cmnd) is often run much quick because we complete request on same CPU with the submission CPU(the CPU for allocation of cmd->cmnd). So from CPU1's review, cmd->cmnd is in a remote NUMA node, __scsi_format_command() is executed much slower than mempool_free(). So when mempool_free() returns, __scsi_format_command() may not fetched the buffer in L1 cache yet, then use-after-free is still triggered. That is why I say this use-after-free is inevitable no matter 'setting SCpnt->cmnd to NULL before calling mempool_free()' or not. Thanks, Ming
On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > On Fri, Nov 10, 2017 at 08:51:58AM -0800, James Bottomley wrote: > > > > On Fri, 2017-11-10 at 17:01 +0800, Ming Lei wrote: > > > > > > cmd->cmnd can be allocated/freed dynamically in case of > > > T10_PI_TYPE2_PROTECTION, so we should check it in scsi_show_rq() > > > because this request may have been freed already here, and cmd- > > > >cmnd has been set as null. > > > > > > We choose to accept read-after-free and dump request data as far > > > as possible. > > > > > > This patch fixs the following kernel crash when dumping request > > > via block's debugfs interface: > > > > > > [ 252.962045] BUG: unable to handle kernel NULL pointer > > > dereference > > > at (null) > > > [ 252.963007] IP: scsi_format_opcode_name+0x1a/0x1c0 > > > [ 252.963007] PGD 25e75a067 P4D 25e75a067 PUD 25e75b067 PMD 0 > > > [ 252.963007] Oops: 0000 [#1] PREEMPT SMP > > > [ 252.963007] Dumping ftrace buffer: > > > [ 252.963007] (ftrace buffer empty) > > > [ 252.963007] Modules linked in: scsi_debug ebtable_filter > > > ebtables > > > ip6table_filter ip6_tables xt_CHECKSUM iptable_mangle > > > ipt_MASQUERADE > > > nf_nat_masquerade_ipv4 iptable_nat nf_conntrack_ipv4 > > > nf_defrag_ipv4 > > > nf_nat_ipv4 nf_nat nf_conntrack libcrc32c bridge stp llc > > > iptable_filter fuse ip_tables sd_mod sg mptsas mptscsih mptbase > > > crc32c_intel ahci libahci nvme serio_raw scsi_transport_sas > > > libata > > > lpc_ich nvme_core virtio_scsi binfmt_misc dm_mod iscsi_tcp > > > libiscsi_tcp libiscsi scsi_transport_iscsi null_blk configs > > > [ 252.963007] CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0- > > > rc2.blk_mq_io_hang+ #516 > > > [ 252.963007] Hardware name: QEMU Standard PC (Q35 + ICH9, > > > 2009), > > > BIOS 1.9.3-1.fc25 04/01/2014 > > > [ 252.963007] task: ffff88025e6f6000 task.stack: > > > ffffc90001bd0000 > > > [ 252.963007] RIP: 0010:scsi_format_opcode_name+0x1a/0x1c0 > > > [ 252.963007] RSP: 0018:ffffc90001bd3c50 EFLAGS: 00010286 > > > [ 252.963007] RAX: 00000000ffff4843 RBX: 0000000000000050 RCX: > > > 0000000000000000 > > > [ 252.963007] RDX: 0000000000000000 RSI: 0000000000000050 RDI: > > > ffffc90001bd3cd8 > > > [ 252.963007] RBP: ffffc90001bd3c88 R08: 0000000000001000 R09: > > > 0000000000000000 > > > [ 252.963007] R10: ffff880275134000 R11: ffff88027513406c R12: > > > 0000000000000050 > > > [ 252.963007] R13: ffffc90001bd3cd8 R14: 0000000000000000 R15: > > > 0000000000000000 > > > [ 252.963007] FS: 00007f4d11762700(0000) > > > GS:ffff88027fc40000(0000) > > > knlGS:0000000000000000 > > > [ 252.963007] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > > > [ 252.963007] CR2: 0000000000000000 CR3: 000000025e789003 CR4: > > > 00000000003606e0 > > > [ 252.963007] DR0: 0000000000000000 DR1: 0000000000000000 DR2: > > > 0000000000000000 > > > [ 252.963007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: > > > 0000000000000400 > > > [ 252.963007] Call Trace: > > > [ 252.963007] __scsi_format_command+0x27/0xc0 > > > [ 252.963007] scsi_show_rq+0x5c/0xc0 > > > [ 252.963007] ? seq_printf+0x4e/0x70 > > > [ 252.963007] ? blk_flags_show+0x5b/0xf0 > > > [ 252.963007] __blk_mq_debugfs_rq_show+0x116/0x130 > > > [ 252.963007] blk_mq_debugfs_rq_show+0xe/0x10 > > > [ 252.963007] seq_read+0xfe/0x3b0 > > > [ 252.963007] ? __handle_mm_fault+0x631/0x1150 > > > [ 252.963007] full_proxy_read+0x54/0x90 > > > [ 252.963007] __vfs_read+0x37/0x160 > > > [ 252.963007] ? security_file_permission+0x9b/0xc0 > > > [ 252.963007] vfs_read+0x96/0x130 > > > [ 252.963007] SyS_read+0x55/0xc0 > > > [ 252.963007] entry_SYSCALL_64_fastpath+0x1a/0xa5 > > > [ 252.963007] RIP: 0033:0x7f4d1127e9b0 > > > [ 252.963007] RSP: 002b:00007ffd27082568 EFLAGS: 00000246 > > > ORIG_RAX: > > > 0000000000000000 > > > [ 252.963007] RAX: ffffffffffffffda RBX: 00007f4d1154bb20 RCX: > > > 00007f4d1127e9b0 > > > [ 252.963007] RDX: 0000000000020000 RSI: 00007f4d115a7000 RDI: > > > 0000000000000003 > > > [ 252.963007] RBP: 0000000000021010 R08: ffffffffffffffff R09: > > > 0000000000000000 > > > [ 252.963007] R10: 000000000000037b R11: 0000000000000246 R12: > > > 0000000000022000 > > > [ 252.963007] R13: 00007f4d1154bb78 R14: 0000000000001000 R15: > > > 0000000000020000 > > > [ 252.963007] Code: c6 e8 1b ca 24 00 eb 8c e8 74 2c ae ff 0f 1f > > > 40 > > > 00 0f 1f 44 00 00 55 48 89 e5 41 56 41 55 41 54 53 49 89 fd 49 89 > > > f4 > > > 48 83 ec 18 <44> 0f b6 32 48 c7 45 c8 00 00 00 00 65 48 8b 04 25 > > > 28 > > > 00 00 00 > > > [ 252.963007] RIP: scsi_format_opcode_name+0x1a/0x1c0 RSP: > > > ffffc90001bd3c50 > > > [ 252.963007] CR2: 0000000000000000 > > > [ 252.963007] ---[ end trace 83c5bddfbaa6573c ]--- > > > [ 252.963007] Kernel panic - not syncing: Fatal exception > > > [ 252.963007] Dumping ftrace buffer: > > > [ 252.963007] (ftrace buffer empty) > > > [ 252.963007] Kernel Offset: disabled > > > [ 252.963007] ---[ end Kernel panic - not syncing: Fatal > > > exception > > > > > > Fixes: 0eebd005dd07(scsi: Implement blk_mq_ops.show_rq()) > > > Cc: Bart Van Assche <bart.vanassche@sandisk.com> > > > Cc: Omar Sandoval <osandov@fb.com> > > > Cc: Martin K. Petersen <martin.petersen@oracle.com> > > > Cc: James Bottomley <James.Bottomley@HansenPartnership.com> > > > Cc: Hannes Reinecke <hare@suse.com> > > > Signed-off-by: Ming Lei <ming.lei@redhat.com> > > > --- > > > V2: > > > - fix typo > > > V3: > > > - prefer to dump data and accept read-after-free > > > - add some comment > > > V4: > > > - read the two variable into local variable first, for > > > avoiding > > > free between check and passing to __scsi_format_command(). > > > > > > > > > drivers/scsi/scsi_debugfs.c | 14 +++++++++++++- > > > 1 file changed, 13 insertions(+), 1 deletion(-) > > > > > > diff --git a/drivers/scsi/scsi_debugfs.c > > > b/drivers/scsi/scsi_debugfs.c > > > index 5e9755008aed..f8437c456458 100644 > > > --- a/drivers/scsi/scsi_debugfs.c > > > +++ b/drivers/scsi/scsi_debugfs.c > > > @@ -8,8 +8,20 @@ void scsi_show_rq(struct seq_file *m, struct > > > request > > > *rq) > > > struct scsi_cmnd *cmd = container_of(scsi_req(rq), > > > typeof(*cmd), req); > > > int msecs = jiffies_to_msecs(jiffies - cmd- > > > > > > > > jiffies_at_alloc); > > > char buf[80]; > > > + unsigned char *cdb = cmd->cmnd; > > > + unsigned short cdb_len = cmd->cmd_len; > > > > > > > > > - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd- > > > > > > > > cmd_len); > > > + /* > > > + * This rq may have been freed, so don't be surprised if > > > + * read-after-free is reported. > > > > I said fix this in sd.c:sd_uninit_command() by setting SCpnt->cmnd > > to NULL before calling mempool_free(). > > Hi James, > > That can't avoid the use-after-free in both reality and theory. I believe I already gave you the theory of how to avoid it. > When 'cdb = cmd->cmnd' is executed in CPU1, 'cbd' is a valid pointer, > just after this instruction is done, this request is completed in > CPU2, and mempool_free() is run to complete this buffer. > > Firstly, both preemption and IRQ can happen inside > __scsi_format_command(), > when this function is scheduled to rerun, this buffer is freed since > mempool_free() has completed already, which is run in irq context. so disable irq's before taking the cmnd pointer copy ... > Secondly, the use-after-free can't be avoided too if we disable irq > when checking and reading the variable. See following: > > mempool_free(cmd->cmnd) is often run much quick because we complete > request on same CPU with the submission CPU(the CPU for allocation of > cmd->cmnd). only in the add element case, which won't cause a use after free from any tool's point of view. If you want greater assurance, take the pool->lock (with encapsulation so we're not poking around in mempool internals). > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > __scsi_format_command() is executed much slower than mempool_free(). > So when mempool_free() returns, __scsi_format_command() may not > fetched the buffer in L1 cache yet, then use-after-free > is still triggered. > > That is why I say this use-after-free is inevitable no matter > 'setting SCpnt->cmnd to NULL before calling mempool_free()' or not. The bottom line is that there are several creative ways around this but the proposed code is currently broken and simply putting a comment in saying so doesn't make it acceptable. James
Hi James, On Mon, Nov 13, 2017 at 10:55:52AM -0800, James Bottomley wrote: > On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > > On Fri, Nov 10, 2017 at 08:51:58AM -0800, James Bottomley wrote: > > > > > > On Fri, 2017-11-10 at 17:01 +0800, Ming Lei wrote: > > > > > > > > cmd->cmnd can be allocated/freed dynamically in case of > > > > T10_PI_TYPE2_PROTECTION, so we should check it in scsi_show_rq() > > > > because this request may have been freed already here, and cmd- > > > > >cmnd has been set as null. > > > > > > > > We choose to accept read-after-free and dump request data as far > > > > as possible. > > > > > > > > This patch fixs the following kernel crash when dumping request > > > > via block's debugfs interface: > > > > > > > > [ 252.962045] BUG: unable to handle kernel NULL pointer > > > > dereference > > > > at (null) > > > > [ 252.963007] IP: scsi_format_opcode_name+0x1a/0x1c0 > > > > [ 252.963007] PGD 25e75a067 P4D 25e75a067 PUD 25e75b067 PMD 0 > > > > [ 252.963007] Oops: 0000 [#1] PREEMPT SMP > > > > [ 252.963007] Dumping ftrace buffer: > > > > [ 252.963007] (ftrace buffer empty) > > > > [ 252.963007] Modules linked in: scsi_debug ebtable_filter > > > > ebtables > > > > ip6table_filter ip6_tables xt_CHECKSUM iptable_mangle > > > > ipt_MASQUERADE > > > > nf_nat_masquerade_ipv4 iptable_nat nf_conntrack_ipv4 > > > > nf_defrag_ipv4 > > > > nf_nat_ipv4 nf_nat nf_conntrack libcrc32c bridge stp llc > > > > iptable_filter fuse ip_tables sd_mod sg mptsas mptscsih mptbase > > > > crc32c_intel ahci libahci nvme serio_raw scsi_transport_sas > > > > libata > > > > lpc_ich nvme_core virtio_scsi binfmt_misc dm_mod iscsi_tcp > > > > libiscsi_tcp libiscsi scsi_transport_iscsi null_blk configs > > > > [ 252.963007] CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0- > > > > rc2.blk_mq_io_hang+ #516 > > > > [ 252.963007] Hardware name: QEMU Standard PC (Q35 + ICH9, > > > > 2009), > > > > BIOS 1.9.3-1.fc25 04/01/2014 > > > > [ 252.963007] task: ffff88025e6f6000 task.stack: > > > > ffffc90001bd0000 > > > > [ 252.963007] RIP: 0010:scsi_format_opcode_name+0x1a/0x1c0 > > > > [ 252.963007] RSP: 0018:ffffc90001bd3c50 EFLAGS: 00010286 > > > > [ 252.963007] RAX: 00000000ffff4843 RBX: 0000000000000050 RCX: > > > > 0000000000000000 > > > > [ 252.963007] RDX: 0000000000000000 RSI: 0000000000000050 RDI: > > > > ffffc90001bd3cd8 > > > > [ 252.963007] RBP: ffffc90001bd3c88 R08: 0000000000001000 R09: > > > > 0000000000000000 > > > > [ 252.963007] R10: ffff880275134000 R11: ffff88027513406c R12: > > > > 0000000000000050 > > > > [ 252.963007] R13: ffffc90001bd3cd8 R14: 0000000000000000 R15: > > > > 0000000000000000 > > > > [ 252.963007] FS: 00007f4d11762700(0000) > > > > GS:ffff88027fc40000(0000) > > > > knlGS:0000000000000000 > > > > [ 252.963007] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > > > > [ 252.963007] CR2: 0000000000000000 CR3: 000000025e789003 CR4: > > > > 00000000003606e0 > > > > [ 252.963007] DR0: 0000000000000000 DR1: 0000000000000000 DR2: > > > > 0000000000000000 > > > > [ 252.963007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: > > > > 0000000000000400 > > > > [ 252.963007] Call Trace: > > > > [ 252.963007] __scsi_format_command+0x27/0xc0 > > > > [ 252.963007] scsi_show_rq+0x5c/0xc0 > > > > [ 252.963007] ? seq_printf+0x4e/0x70 > > > > [ 252.963007] ? blk_flags_show+0x5b/0xf0 > > > > [ 252.963007] __blk_mq_debugfs_rq_show+0x116/0x130 > > > > [ 252.963007] blk_mq_debugfs_rq_show+0xe/0x10 > > > > [ 252.963007] seq_read+0xfe/0x3b0 > > > > [ 252.963007] ? __handle_mm_fault+0x631/0x1150 > > > > [ 252.963007] full_proxy_read+0x54/0x90 > > > > [ 252.963007] __vfs_read+0x37/0x160 > > > > [ 252.963007] ? security_file_permission+0x9b/0xc0 > > > > [ 252.963007] vfs_read+0x96/0x130 > > > > [ 252.963007] SyS_read+0x55/0xc0 > > > > [ 252.963007] entry_SYSCALL_64_fastpath+0x1a/0xa5 > > > > [ 252.963007] RIP: 0033:0x7f4d1127e9b0 > > > > [ 252.963007] RSP: 002b:00007ffd27082568 EFLAGS: 00000246 > > > > ORIG_RAX: > > > > 0000000000000000 > > > > [ 252.963007] RAX: ffffffffffffffda RBX: 00007f4d1154bb20 RCX: > > > > 00007f4d1127e9b0 > > > > [ 252.963007] RDX: 0000000000020000 RSI: 00007f4d115a7000 RDI: > > > > 0000000000000003 > > > > [ 252.963007] RBP: 0000000000021010 R08: ffffffffffffffff R09: > > > > 0000000000000000 > > > > [ 252.963007] R10: 000000000000037b R11: 0000000000000246 R12: > > > > 0000000000022000 > > > > [ 252.963007] R13: 00007f4d1154bb78 R14: 0000000000001000 R15: > > > > 0000000000020000 > > > > [ 252.963007] Code: c6 e8 1b ca 24 00 eb 8c e8 74 2c ae ff 0f 1f > > > > 40 > > > > 00 0f 1f 44 00 00 55 48 89 e5 41 56 41 55 41 54 53 49 89 fd 49 89 > > > > f4 > > > > 48 83 ec 18 <44> 0f b6 32 48 c7 45 c8 00 00 00 00 65 48 8b 04 25 > > > > 28 > > > > 00 00 00 > > > > [ 252.963007] RIP: scsi_format_opcode_name+0x1a/0x1c0 RSP: > > > > ffffc90001bd3c50 > > > > [ 252.963007] CR2: 0000000000000000 > > > > [ 252.963007] ---[ end trace 83c5bddfbaa6573c ]--- > > > > [ 252.963007] Kernel panic - not syncing: Fatal exception > > > > [ 252.963007] Dumping ftrace buffer: > > > > [ 252.963007] (ftrace buffer empty) > > > > [ 252.963007] Kernel Offset: disabled > > > > [ 252.963007] ---[ end Kernel panic - not syncing: Fatal > > > > exception > > > > > > > > Fixes: 0eebd005dd07(scsi: Implement blk_mq_ops.show_rq()) > > > > Cc: Bart Van Assche <bart.vanassche@sandisk.com> > > > > Cc: Omar Sandoval <osandov@fb.com> > > > > Cc: Martin K. Petersen <martin.petersen@oracle.com> > > > > Cc: James Bottomley <James.Bottomley@HansenPartnership.com> > > > > Cc: Hannes Reinecke <hare@suse.com> > > > > Signed-off-by: Ming Lei <ming.lei@redhat.com> > > > > --- > > > > V2: > > > > - fix typo > > > > V3: > > > > - prefer to dump data and accept read-after-free > > > > - add some comment > > > > V4: > > > > - read the two variable into local variable first, for > > > > avoiding > > > > free between check and passing to __scsi_format_command(). > > > > > > > > > > > > drivers/scsi/scsi_debugfs.c | 14 +++++++++++++- > > > > 1 file changed, 13 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/drivers/scsi/scsi_debugfs.c > > > > b/drivers/scsi/scsi_debugfs.c > > > > index 5e9755008aed..f8437c456458 100644 > > > > --- a/drivers/scsi/scsi_debugfs.c > > > > +++ b/drivers/scsi/scsi_debugfs.c > > > > @@ -8,8 +8,20 @@ void scsi_show_rq(struct seq_file *m, struct > > > > request > > > > *rq) > > > > struct scsi_cmnd *cmd = container_of(scsi_req(rq), > > > > typeof(*cmd), req); > > > > int msecs = jiffies_to_msecs(jiffies - cmd- > > > > > > > > > > jiffies_at_alloc); > > > > char buf[80]; > > > > + unsigned char *cdb = cmd->cmnd; > > > > + unsigned short cdb_len = cmd->cmd_len; > > > > > > > > > > > > - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd- > > > > > > > > > > cmd_len); > > > > + /* > > > > + * This rq may have been freed, so don't be surprised if > > > > + * read-after-free is reported. > > > > > > I said fix this in sd.c:sd_uninit_command() by setting SCpnt->cmnd > > > to NULL before calling mempool_free(). > > > > Hi James, > > > > That can't avoid the use-after-free in both reality and theory. > > I believe I already gave you the theory of how to avoid it. > > > When 'cdb = cmd->cmnd' is executed in CPU1, 'cbd' is a valid pointer, > > just after this instruction is done, this request is completed in > > CPU2, and mempool_free() is run to complete this buffer. > > > > Firstly, both preemption and IRQ can happen inside > > __scsi_format_command(), > > when this function is scheduled to rerun, this buffer is freed since > > mempool_free() has completed already, which is run in irq context. > > so disable irq's before taking the cmnd pointer copy ... It isn't enough. > > > Secondly, the use-after-free can't be avoided too if we disable irq > > when checking and reading the variable. See following: > > > > mempool_free(cmd->cmnd) is often run much quick because we complete > > request on same CPU with the submission CPU(the CPU for allocation of > > cmd->cmnd). > > only in the add element case, which won't cause a use after free from I don't understand why that is only in add-element case, could you explain it a bit? For slub/slab..., I believe there is much optimization in case that the allocation/free is done on one local CPU, which should be quicker than memory read from a remote NUMA node. > any tool's point of view. If you want greater assurance, take the For example of KASAN[1], each byte of memory is covered by its shadow memory. Once pool->free() in mempool_free(cmd->cmnd) is run before __scsi_format_command() returns, the use-after-free is triggered. And there are several access to this buffer in __scsi_format_command(). KASAN checks if each access is safe before the actual access, and the check is added by GCC, see below words from [1]: Compile-time instrumentation used for checking memory accesses. Compiler inserts function calls (__asan_load*(addr), __asan_store*(addr)) before each memory access of size 1, 2, 4, 8 or 16. These functions check whether memory access is valid or not by checking corresponding shadow memory. [1] Documentation/dev-tools/kasan.rst > pool->lock (with encapsulation so we're not poking around in mempool > internals). The pool->lock isn't needed in fast path of 'pool->free()', and it is required only when 'if (unlikely(pool->curr_nr < pool->min_nr))' is true, so this way doesn't work too. > > > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > > __scsi_format_command() is executed much slower than mempool_free(). > > So when mempool_free() returns, __scsi_format_command() may not > > fetched the buffer in L1 cache yet, then use-after-free > > is still triggered. > > > > That is why I say this use-after-free is inevitable no matter > > 'setting SCpnt->cmnd to NULL before calling mempool_free()' or not. > > The bottom line is that there are several creative ways around this but > the proposed code is currently broken and simply putting a comment in > saying so doesn't make it acceptable. As I explained above, I didn't see one really workable way. Or please correct it if I am wrong. Thanks, Ming
On Tue, 2017-11-14 at 08:55 +0800, Ming Lei wrote: > Hi James, > > On Mon, Nov 13, 2017 at 10:55:52AM -0800, James Bottomley wrote: > > > > On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > > > > > > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > > > __scsi_format_command() is executed much slower than > > > mempool_free(). > > > So when mempool_free() returns, __scsi_format_command() may not > > > fetched the buffer in L1 cache yet, then use-after-free > > > is still triggered. > > > > > > That is why I say this use-after-free is inevitable no matter > > > 'setting SCpnt->cmnd to NULL before calling mempool_free()' or > > > not. > > > > The bottom line is that there are several creative ways around this > > but the proposed code is currently broken and simply putting a > > comment in saying so doesn't make it acceptable. > > As I explained above, I didn't see one really workable way. Or please > correct it if I am wrong. I simply can't believe it's beyond the wit of man to solve a use after free race. About 40% of kernel techniques are devoted to this. All I really care about is not losing the PI information we previously had. I agree with Bart that NULL cmnd is a good indicator, so it seems reasonable to use it. If you have another mechanism, feel free to propose it. James
On Tue, Nov 14, 2017 at 10:14:52AM -0800, James Bottomley wrote: > On Tue, 2017-11-14 at 08:55 +0800, Ming Lei wrote: > > Hi James, > > > > On Mon, Nov 13, 2017 at 10:55:52AM -0800, James Bottomley wrote: > > > > > > On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > > > > > > > > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > > > > __scsi_format_command() is executed much slower than > > > > mempool_free(). > > > > So when mempool_free() returns, __scsi_format_command() may not > > > > fetched the buffer in L1 cache yet, then use-after-free > > > > is still triggered. > > > > > > > > That is why I say this use-after-free is inevitable no matter > > > > 'setting SCpnt->cmnd to NULL before calling mempool_free()' or > > > > not. > > > > > > The bottom line is that there are several creative ways around this > > > but the proposed code is currently broken and simply putting a > > > comment in saying so doesn't make it acceptable. > > > > As I explained above, I didn't see one really workable way. Or please > > correct it if I am wrong. > > I simply can't believe it's beyond the wit of man to solve a use after > free race. About 40% of kernel techniques are devoted to this. All I > really care about is not losing the PI information we previously had. > I agree with Bart that NULL cmnd is a good indicator, so it seems > reasonable to use it. If you have another mechanism, feel free to > propose it. Hi James, This patch is my proposal, no others thought of yet. We can fix the use-after-free easily via lock, rcu and ..., but some cost has to pay. In this case, we can't wait too long in show_rq(), otherwise we may lose important debug info, so I do not have better way. IMO this use-after-free is actually no harm, I don't think we have to fix it, but it should be better to not let utility warn on this case. Thanks, Ming
On Wed, 2017-11-15 at 18:09 +0800, Ming Lei wrote: > On Tue, Nov 14, 2017 at 10:14:52AM -0800, James Bottomley wrote: > > > > On Tue, 2017-11-14 at 08:55 +0800, Ming Lei wrote: > > > > > > Hi James, > > > > > > On Mon, Nov 13, 2017 at 10:55:52AM -0800, James Bottomley wrote: > > > > > > > > > > > > On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > > > > > > > > > > > > > > > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > > > > > __scsi_format_command() is executed much slower than > > > > > mempool_free(). > > > > > So when mempool_free() returns, __scsi_format_command() may > > > > > not fetched the buffer in L1 cache yet, then use-after-free > > > > > is still triggered. > > > > > > > > > > That is why I say this use-after-free is inevitable no matter > > > > > 'setting SCpnt->cmnd to NULL before calling mempool_free()' > > > > > or not. > > > > > > > > The bottom line is that there are several creative ways around > > > > this but the proposed code is currently broken and simply > > > > putting a comment in saying so doesn't make it acceptable. > > > > > > As I explained above, I didn't see one really workable way. Or > > > please correct it if I am wrong. > > > > I simply can't believe it's beyond the wit of man to solve a use > > after free race. About 40% of kernel techniques are devoted to > > this. All I really care about is not losing the PI information we > > previously had. I agree with Bart that NULL cmnd is a good > > indicator, so it seems reasonable to use it. If you have another > > mechanism, feel free to propose it. > > Hi James, > > This patch is my proposal, no others thought of yet. > > We can fix the use-after-free easily via lock, rcu and ..., but some > cost has to pay. In this case, we can't wait too long in show_rq(), > otherwise we may lose important debug info, so I do not have better > way. > > IMO this use-after-free is actually no harm, I don't think we have to > fix it, but it should be better to not let utility warn on this case. Fine, so lose the snide comment and set the cmnd to NULL *before* calling free so we narrow the race window. James
On Wed, Nov 15, 2017 at 07:28:00PM +0900, James Bottomley wrote: > On Wed, 2017-11-15 at 18:09 +0800, Ming Lei wrote: > > On Tue, Nov 14, 2017 at 10:14:52AM -0800, James Bottomley wrote: > > > > > > On Tue, 2017-11-14 at 08:55 +0800, Ming Lei wrote: > > > > > > > > Hi James, > > > > > > > > On Mon, Nov 13, 2017 at 10:55:52AM -0800, James Bottomley wrote: > > > > > > > > > > > > > > > On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > > > > > > > > > > > > > > > > > > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > > > > > > __scsi_format_command() is executed much slower than > > > > > > mempool_free(). > > > > > > So when mempool_free() returns, __scsi_format_command() may > > > > > > not fetched the buffer in L1 cache yet, then use-after-free > > > > > > is still triggered. > > > > > > > > > > > > That is why I say this use-after-free is inevitable no matter > > > > > > 'setting SCpnt->cmnd to NULL before calling mempool_free()' > > > > > > or not. > > > > > > > > > > The bottom line is that there are several creative ways around > > > > > this but the proposed code is currently broken and simply > > > > > putting a comment in saying so doesn't make it acceptable. > > > > > > > > As I explained above, I didn't see one really workable way. Or > > > > please correct it if I am wrong. > > > > > > I simply can't believe it's beyond the wit of man to solve a use > > > after free race. About 40% of kernel techniques are devoted to > > > this. All I really care about is not losing the PI information we > > > previously had. I agree with Bart that NULL cmnd is a good > > > indicator, so it seems reasonable to use it. If you have another > > > mechanism, feel free to propose it. > > > > Hi James, > > > > This patch is my proposal, no others thought of yet. > > > > We can fix the use-after-free easily via lock, rcu and ..., but some > > cost has to pay. In this case, we can't wait too long in show_rq(), > > otherwise we may lose important debug info, so I do not have better > > way. > > > > IMO this use-after-free is actually no harm, I don't think we have to > > fix it, but it should be better to not let utility warn on this case. > > Fine, so lose the snide comment and set the cmnd to NULL *before* > calling free so we narrow the race window. Hi James, Given we can't avoid the use-after-free, how about not do that way so we won't lose the precious debug info too early? Thanks, Ming
On Wed, Nov 15, 2017 at 08:04:49PM +0800, Ming Lei wrote: > On Wed, Nov 15, 2017 at 07:28:00PM +0900, James Bottomley wrote: > > On Wed, 2017-11-15 at 18:09 +0800, Ming Lei wrote: > > > On Tue, Nov 14, 2017 at 10:14:52AM -0800, James Bottomley wrote: > > > > > > > > On Tue, 2017-11-14 at 08:55 +0800, Ming Lei wrote: > > > > > > > > > > Hi James, > > > > > > > > > > On Mon, Nov 13, 2017 at 10:55:52AM -0800, James Bottomley wrote: > > > > > > > > > > > > > > > > > > On Sat, 2017-11-11 at 10:43 +0800, Ming Lei wrote: > > > > > > > > > > > > > > > > > > > > > So from CPU1's review, cmd->cmnd is in a remote NUMA node, > > > > > > > __scsi_format_command() is executed much slower than > > > > > > > mempool_free(). > > > > > > > So when mempool_free() returns, __scsi_format_command() may > > > > > > > not fetched the buffer in L1 cache yet, then use-after-free > > > > > > > is still triggered. > > > > > > > > > > > > > > That is why I say this use-after-free is inevitable no matter > > > > > > > 'setting SCpnt->cmnd to NULL before calling mempool_free()' > > > > > > > or not. > > > > > > > > > > > > The bottom line is that there are several creative ways around > > > > > > this but the proposed code is currently broken and simply > > > > > > putting a comment in saying so doesn't make it acceptable. > > > > > > > > > > As I explained above, I didn't see one really workable way. Or > > > > > please correct it if I am wrong. > > > > > > > > I simply can't believe it's beyond the wit of man to solve a use > > > > after free race. About 40% of kernel techniques are devoted to > > > > this. All I really care about is not losing the PI information we > > > > previously had. I agree with Bart that NULL cmnd is a good > > > > indicator, so it seems reasonable to use it. If you have another > > > > mechanism, feel free to propose it. > > > > > > Hi James, > > > > > > This patch is my proposal, no others thought of yet. > > > > > > We can fix the use-after-free easily via lock, rcu and ..., but some > > > cost has to pay. In this case, we can't wait too long in show_rq(), > > > otherwise we may lose important debug info, so I do not have better > > > way. > > > > > > IMO this use-after-free is actually no harm, I don't think we have to > > > fix it, but it should be better to not let utility warn on this case. > > > > Fine, so lose the snide comment and set the cmnd to NULL *before* > > calling free so we narrow the race window. > > Hi James, > > Given we can't avoid the use-after-free, how about not do that way so > we won't lose the precious debug info too early? Hi James, Are you fine with V4? As I explained, the use-after-free can't be avoided, we have to make scsi_show_rq() to survive that, so we don't need to touch code in free path. Also we won't lose debug info too early in this way, not like 'set the cmnd to NULL *before* calling free'. Thanks, Ming
diff --git a/drivers/scsi/scsi_debugfs.c b/drivers/scsi/scsi_debugfs.c index 5e9755008aed..f8437c456458 100644 --- a/drivers/scsi/scsi_debugfs.c +++ b/drivers/scsi/scsi_debugfs.c @@ -8,8 +8,20 @@ void scsi_show_rq(struct seq_file *m, struct request *rq) struct scsi_cmnd *cmd = container_of(scsi_req(rq), typeof(*cmd), req); int msecs = jiffies_to_msecs(jiffies - cmd->jiffies_at_alloc); char buf[80]; + unsigned char *cdb = cmd->cmnd; + unsigned short cdb_len = cmd->cmd_len; - __scsi_format_command(buf, sizeof(buf), cmd->cmnd, cmd->cmd_len); + /* + * This rq may have been freed, so don't be surprised if + * read-after-free is reported. + * + * __scsi_format_command() should be written as not being broken + * with gargabe 'cdb' input. + */ + if (cdb && cdb_len) + __scsi_format_command(buf, sizeof(buf), cdb, cdb_len); + else + strcpy(buf, ""); seq_printf(m, ", .cmd=%s, .retries=%d, allocated %d.%03d s ago", buf, cmd->retries, msecs / 1000, msecs % 1000); }
cmd->cmnd can be allocated/freed dynamically in case of T10_PI_TYPE2_PROTECTION, so we should check it in scsi_show_rq() because this request may have been freed already here, and cmd->cmnd has been set as null. We choose to accept read-after-free and dump request data as far as possible. This patch fixs the following kernel crash when dumping request via block's debugfs interface: [ 252.962045] BUG: unable to handle kernel NULL pointer dereference at (null) [ 252.963007] IP: scsi_format_opcode_name+0x1a/0x1c0 [ 252.963007] PGD 25e75a067 P4D 25e75a067 PUD 25e75b067 PMD 0 [ 252.963007] Oops: 0000 [#1] PREEMPT SMP [ 252.963007] Dumping ftrace buffer: [ 252.963007] (ftrace buffer empty) [ 252.963007] Modules linked in: scsi_debug ebtable_filter ebtables ip6table_filter ip6_tables xt_CHECKSUM iptable_mangle ipt_MASQUERADE nf_nat_masquerade_ipv4 iptable_nat nf_conntrack_ipv4 nf_defrag_ipv4 nf_nat_ipv4 nf_nat nf_conntrack libcrc32c bridge stp llc iptable_filter fuse ip_tables sd_mod sg mptsas mptscsih mptbase crc32c_intel ahci libahci nvme serio_raw scsi_transport_sas libata lpc_ich nvme_core virtio_scsi binfmt_misc dm_mod iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi null_blk configs [ 252.963007] CPU: 1 PID: 1881 Comm: cat Not tainted 4.14.0-rc2.blk_mq_io_hang+ #516 [ 252.963007] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.9.3-1.fc25 04/01/2014 [ 252.963007] task: ffff88025e6f6000 task.stack: ffffc90001bd0000 [ 252.963007] RIP: 0010:scsi_format_opcode_name+0x1a/0x1c0 [ 252.963007] RSP: 0018:ffffc90001bd3c50 EFLAGS: 00010286 [ 252.963007] RAX: 00000000ffff4843 RBX: 0000000000000050 RCX: 0000000000000000 [ 252.963007] RDX: 0000000000000000 RSI: 0000000000000050 RDI: ffffc90001bd3cd8 [ 252.963007] RBP: ffffc90001bd3c88 R08: 0000000000001000 R09: 0000000000000000 [ 252.963007] R10: ffff880275134000 R11: ffff88027513406c R12: 0000000000000050 [ 252.963007] R13: ffffc90001bd3cd8 R14: 0000000000000000 R15: 0000000000000000 [ 252.963007] FS: 00007f4d11762700(0000) GS:ffff88027fc40000(0000) knlGS:0000000000000000 [ 252.963007] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 252.963007] CR2: 0000000000000000 CR3: 000000025e789003 CR4: 00000000003606e0 [ 252.963007] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 252.963007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 252.963007] Call Trace: [ 252.963007] __scsi_format_command+0x27/0xc0 [ 252.963007] scsi_show_rq+0x5c/0xc0 [ 252.963007] ? seq_printf+0x4e/0x70 [ 252.963007] ? blk_flags_show+0x5b/0xf0 [ 252.963007] __blk_mq_debugfs_rq_show+0x116/0x130 [ 252.963007] blk_mq_debugfs_rq_show+0xe/0x10 [ 252.963007] seq_read+0xfe/0x3b0 [ 252.963007] ? __handle_mm_fault+0x631/0x1150 [ 252.963007] full_proxy_read+0x54/0x90 [ 252.963007] __vfs_read+0x37/0x160 [ 252.963007] ? security_file_permission+0x9b/0xc0 [ 252.963007] vfs_read+0x96/0x130 [ 252.963007] SyS_read+0x55/0xc0 [ 252.963007] entry_SYSCALL_64_fastpath+0x1a/0xa5 [ 252.963007] RIP: 0033:0x7f4d1127e9b0 [ 252.963007] RSP: 002b:00007ffd27082568 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 [ 252.963007] RAX: ffffffffffffffda RBX: 00007f4d1154bb20 RCX: 00007f4d1127e9b0 [ 252.963007] RDX: 0000000000020000 RSI: 00007f4d115a7000 RDI: 0000000000000003 [ 252.963007] RBP: 0000000000021010 R08: ffffffffffffffff R09: 0000000000000000 [ 252.963007] R10: 000000000000037b R11: 0000000000000246 R12: 0000000000022000 [ 252.963007] R13: 00007f4d1154bb78 R14: 0000000000001000 R15: 0000000000020000 [ 252.963007] Code: c6 e8 1b ca 24 00 eb 8c e8 74 2c ae ff 0f 1f 40 00 0f 1f 44 00 00 55 48 89 e5 41 56 41 55 41 54 53 49 89 fd 49 89 f4 48 83 ec 18 <44> 0f b6 32 48 c7 45 c8 00 00 00 00 65 48 8b 04 25 28 00 00 00 [ 252.963007] RIP: scsi_format_opcode_name+0x1a/0x1c0 RSP: ffffc90001bd3c50 [ 252.963007] CR2: 0000000000000000 [ 252.963007] ---[ end trace 83c5bddfbaa6573c ]--- [ 252.963007] Kernel panic - not syncing: Fatal exception [ 252.963007] Dumping ftrace buffer: [ 252.963007] (ftrace buffer empty) [ 252.963007] Kernel Offset: disabled [ 252.963007] ---[ end Kernel panic - not syncing: Fatal exception Fixes: 0eebd005dd07(scsi: Implement blk_mq_ops.show_rq()) Cc: Bart Van Assche <bart.vanassche@sandisk.com> Cc: Omar Sandoval <osandov@fb.com> Cc: Martin K. Petersen <martin.petersen@oracle.com> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Cc: Hannes Reinecke <hare@suse.com> Signed-off-by: Ming Lei <ming.lei@redhat.com> --- V2: - fix typo V3: - prefer to dump data and accept read-after-free - add some comment V4: - read the two variable into local variable first, for avoiding free between check and passing to __scsi_format_command(). drivers/scsi/scsi_debugfs.c | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-)