efivars, efi-pstore: Hold off deletion of sysfs entry until the scan is completed
Currently, when mounting pstore file system, a read callback of efi_pstore driver runs mutiple times as below. - In the first read callback, scan efivar_sysfs_list from head and pass a kmsg buffer of a entry to an upper pstore layer. - In the second read callback, rescan efivar_sysfs_list from the entry and pass another kmsg buffer to it. - Repeat the scan and pass until the end of efivar_sysfs_list. In this process, an entry is read across the multiple read function calls. To avoid race between the read and erasion, the whole process above is protected by a spinlock, holding in open() and releasing in close(). At the same time, kmemdup() is called to pass the buffer to pstore filesystem during it. And then, it causes a following lockdep warning. To make the dynamic memory allocation runnable without taking spinlock, holding off a deletion of sysfs entry if it happens while scanning it via efi_pstore, and deleting it after the scan is completed. To implement it, this patch introduces two flags, scanning and deleting, to efivar_entry. On the code basis, it seems that all the scanning and deleting logic is not needed because __efivars->lock are not dropped when reading from the EFI variable store. But, the scanning and deleting logic is still needed because an efi-pstore and a pstore filesystem works as follows. In case an entry(A) is found, the pointer is saved to psi->data. And efi_pstore_read() passes the entry(A) to a pstore filesystem by releasing __efivars->lock. And then, the pstore filesystem calls efi_pstore_read() again and the same entry(A), which is saved to psi->data, is used for resuming to scan a sysfs-list. So, to protect the entry(A), the logic is needed. [ 1.143710] ------------[ cut here ]------------ [ 1.144058] WARNING: CPU: 1 PID: 1 at kernel/lockdep.c:2740 lockdep_trace_alloc+0x104/0x110() [ 1.144058] DEBUG_LOCKS_WARN_ON(irqs_disabled_flags(flags)) [ 1.144058] Modules linked in: [ 1.144058] CPU: 1 PID: 1 Comm: systemd Not tainted 3.11.0-rc5 #2 [ 1.144058] 0000000000000009 ffff8800797e9ae0 ffffffff816614a5 ffff8800797e9b28 [ 1.144058] ffff8800797e9b18 ffffffff8105510d 0000000000000080 0000000000000046 [ 1.144058] 00000000000000d0 00000000000003af ffffffff81ccd0c0 ffff8800797e9b78 [ 1.144058] Call Trace: [ 1.144058] [<ffffffff816614a5>] dump_stack+0x54/0x74 [ 1.144058] [<ffffffff8105510d>] warn_slowpath_common+0x7d/0xa0 [ 1.144058] [<ffffffff8105517c>] warn_slowpath_fmt+0x4c/0x50 [ 1.144058] [<ffffffff8131290f>] ? vsscanf+0x57f/0x7b0 [ 1.144058] [<ffffffff810bbd74>] lockdep_trace_alloc+0x104/0x110 [ 1.144058] [<ffffffff81192da0>] __kmalloc_track_caller+0x50/0x280 [ 1.144058] [<ffffffff815147bb>] ? efi_pstore_read_func.part.1+0x12b/0x170 [ 1.144058] [<ffffffff8115b260>] kmemdup+0x20/0x50 [ 1.144058] [<ffffffff815147bb>] efi_pstore_read_func.part.1+0x12b/0x170 [ 1.144058] [<ffffffff81514800>] ? efi_pstore_read_func.part.1+0x170/0x170 [ 1.144058] [<ffffffff815148b4>] efi_pstore_read_func+0xb4/0xe0 [ 1.144058] [<ffffffff81512b7b>] __efivar_entry_iter+0xfb/0x120 [ 1.144058] [<ffffffff8151428f>] efi_pstore_read+0x3f/0x50 [ 1.144058] [<ffffffff8128d7ba>] pstore_get_records+0x9a/0x150 [ 1.158207] [<ffffffff812af25c>] ? selinux_d_instantiate+0x1c/0x20 [ 1.158207] [<ffffffff8128ce30>] ? parse_options+0x80/0x80 [ 1.158207] [<ffffffff8128ced5>] pstore_fill_super+0xa5/0xc0 [ 1.158207] [<ffffffff811ae7d2>] mount_single+0xa2/0xd0 [ 1.158207] [<ffffffff8128ccf8>] pstore_mount+0x18/0x20 [ 1.158207] [<ffffffff811ae8b9>] mount_fs+0x39/0x1b0 [ 1.158207] [<ffffffff81160550>] ? __alloc_percpu+0x10/0x20 [ 1.158207] [<ffffffff811c9493>] vfs_kern_mount+0x63/0xf0 [ 1.158207] [<ffffffff811cbb0e>] do_mount+0x23e/0xa20 [ 1.158207] [<ffffffff8115b51b>] ? strndup_user+0x4b/0xf0 [ 1.158207] [<ffffffff811cc373>] SyS_mount+0x83/0xc0 [ 1.158207] [<ffffffff81673cc2>] system_call_fastpath+0x16/0x1b [ 1.158207] ---[ end trace 61981bc62de9f6f4 ]--- Signed-off-by: Seiji Aguchi <seiji.aguchi@hds.com> Tested-by: Madper Xie <cxie@redhat.com> Cc: stable@kernel.org Signed-off-by: Matt Fleming <matt.fleming@intel.com>
Showing
Please register or sign in to comment