Re: [PATCH] cxl/hdm: fix a warning in devm_remove_action()

[Dan Williams <dan.j.williams@intel.com>]

Sungwoo Kim wrote:
> In the following race scenario, devm_remove_action() can be called
> before devm_add_action(), triggering a warning because there is no
> action to remove.
> 
> To fix this, extend a critical section to embrace both
> __cxl_dpa_reserve() and devm_add_action().
> 
> task 1:
> cxl_dpa_alloc()
>   __cxl_dpa_reserve()    ...(1)
>   devm_add_action()      ...(4)
> 
> task 2:
> cxl_dpa_free()
>   guard()
>   if (!cxled->dpa_res)   ...(2) pass, due to (1)
>     return 0;
>   devm_cxl_dpa_release()
>     devm_remove_action()  ...(3) warning, no action is added yet
> 
> Warning log:
> 
> WARNING: ./include/linux/device/devres.h:160 at devm_remove_action include/linux/device/devres.h:160 [inline], CPU#0: syz.1.6464/25993
> WARNING: ./include/linux/device/devres.h:160 at devm_cxl_dpa_release drivers/cxl/core/hdm.c:290 [inline], CPU#0: syz.1.6464/25993
> WARNING: ./include/linux/device/devres.h:160 at cxl_dpa_free+0x2a4/0x320 drivers/cxl/core/hdm.c:572, CPU#0: syz.1.6464/25993
> Modules linked in:
> CPU: 0 UID: 0 PID: 25993 Comm: syz.1.6464 Not tainted 6.19.0-g15a37b05f387 #10 PREEMPT(lazy)
> Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.17.0-0-gb52ca86e094d-prebuilt.qemu.org 04/01/2014
> RIP: 0010:devm_remove_action include/linux/device/devres.h:160 [inline]
> RIP: 0010:devm_cxl_dpa_release drivers/cxl/core/hdm.c:290 [inline]
> RIP: 0010:cxl_dpa_free+0x2a4/0x320 drivers/cxl/core/hdm.c:572
> Code: 01 31 ff 41 89 c4 89 c6 e8 a9 91 1d fe 45 85 e4 0f 85 55 ff ff ff e8 4b 96 1d fe 90 0f 0b 90 e9 47 ff ff ff e8 3d 96 1d fe 90 <0f> 0b 90 e9 63 ff ff ff bb f0 ff ff ff e9 68 ff ff ff 48 c7 c7 f4
> RSP: 0018:ffff888129f77bb8 EFLAGS: 00010282
> RAX: 00000000000000f3 RBX: ffff888102f64000 RCX: ffffffff837301d3
> RDX: 0000000000080000 RSI: ffffc90001c5a000 RDI: 0000000000000005
> RBP: 00000000fffffffe R08: 0000000000000001 R09: ffff888123f98000
> R10: 00000000fffffffe R11: 0000000000000000 R12: 0000000000000001
> R13: ffff888102f64040 R14: dffffc0000000000 R15: ffff888102f64000
> FS:  00007fae0488d6c0(0000) GS:ffff88827086d000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 0000000000000000 CR3: 000000012f7d6003 CR4: 0000000000770ef0
> PKRU: 80000000
> Call Trace:
>  <TASK>
>  dpa_size_store+0xfb/0x230 drivers/cxl/core/port.c:269
>  dev_attr_store+0x58/0x80 drivers/base/core.c:2437
>  sysfs_kf_write+0xf2/0x150 fs/sysfs/file.c:142
>  kernfs_fop_write_iter+0x3d6/0x5d0 fs/kernfs/file.c:352
>  new_sync_write fs/read_write.c:595 [inline]
>  vfs_write+0x68e/0x1050 fs/read_write.c:688
>  ksys_write+0x12a/0x250 fs/read_write.c:740
>  do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
>  do_syscall_64+0xfc/0x670 arch/x86/entry/syscall_64.c:94
>  entry_SYSCALL_64_after_hwframe+0x77/0x7f
> RIP: 0033:0x7fae0399c669
> Code: ff c3 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 44 00 00 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 c7 c1 e8 ff ff ff f7 d8 64 89 01 48
> RSP: 002b:00007fae0488d028 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
> RAX: ffffffffffffffda RBX: 00007fae03c15fa0 RCX: 00007fae0399c669
> RDX: 000000000000000a RSI: 00002000000001c0 RDI: 0000000000000009
> RBP: 00007fae03a32c71 R08: 0000000000000000 R09: 0000000000000000
> R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
> R13: 00007fae03c16038 R14: 00007fae03c15fa0 R15: 00007ffe53519fb8
>  </TASK>
> irq event stamp: 1021
> hardirqs last  enabled at (1029): [<ffffffffb7d04c4b>] __up_console_sem+0x7b/0x90 kernel/printk/printk.c:347
> hardirqs last disabled at (1038): [<ffffffffb7d04c30>] __up_console_sem+0x60/0x90 kernel/printk/printk.c:345
> softirqs last  enabled at (1016): [<ffffffffb7b73b93>] __do_softirq kernel/softirq.c:656 [inline]
> softirqs last  enabled at (1016): [<ffffffffb7b73b93>] invoke_softirq kernel/softirq.c:496 [inline]
> softirqs last  enabled at (1016): [<ffffffffb7b73b93>] __irq_exit_rcu+0xb3/0xe0 kernel/softirq.c:723
> softirqs last disabled at (1005): [<ffffffffb7b73b93>] __do_softirq kernel/softirq.c:656 [inline]
> softirqs last disabled at (1005): [<ffffffffb7b73b93>] invoke_softirq kernel/softirq.c:496 [inline]
> softirqs last disabled at (1005): [<ffffffffb7b73b93>] __irq_exit_rcu+0xb3/0xe0 kernel/softirq.c:723
> 
> Fixes: cf880423b6a0 ("cxl/hdm: Add support for allocating DPA to an endpoint decoder")
> Fixes: 9c57cde0dcbd ("cxl/hdm: Enumerate allocated DPA")
> Signed-off-by: Sungwoo Kim <iam@sung-woo.kim>

Good find!

However I am not a fan of the fact that this duplicates this subtle
dance in multiple places. It looks like we can just make it by the case
that __cxl_dpa_reserve() establishes the remove action before returning.

> @@ -679,11 +686,28 @@ int cxl_dpa_alloc(struct cxl_endpoint_decoder *cxled, u64 size)
>  	struct cxl_port *port = cxled_to_port(cxled);
>  	int rc;
>  
> +	guard(rwsem_write)(&cxl_rwsem.dpa);
>  	rc = __cxl_dpa_alloc(cxled, size);
>  	if (rc)
>  		return rc;
>  
> -	return devm_add_action_or_reset(&port->dev, cxl_dpa_release, cxled);
> +	/*
> +	 * Add the devres action while still holding cxl_rwsem.dpa to prevent
> +	 * a race with cxl_dpa_free(). Without this, a concurrent cxl_dpa_free()
> +	 * can observe dpa_res set (by __cxl_dpa_reserve()) and attempt
> +	 * devm_remove_action() before devm_add_action() has been called,
> +	 * triggering a WARN_ON in devm_remove_action().
> +	 * Also, devm_add_action_or_reset() cannot be used here because
> +	 * cxl_dpa_release() tries to hold the dpa lock that is already held,
> +	 * causing a self deadlock.
> +	 */

Also, I agree with Alison, this is good detail, but way too much detail
to carry around permanently in the source directly.