[PATCH] PCI/P2PDMA: Fix a sleeping issue in a RCU read section

[PATCH] PCI/P2PDMA: Fix a sleeping issue in a RCU read section

[Christophe JAILLET]

It is not allowed to sleep within a RCU read section, so use GFP_ATOMIC
instead of GFP_KERNEL here.

Fixes: ae21f835a5bd ("PCI/P2PDMA: Finish RCU conversion of pdev->p2pdma")
Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
---
This patch is speculative.
It is based on a discussion related to another patch. See [1].

It also matches the doc, IIUC. See [2]

[1]: https://lore.kernel.org/all/20240104143925.194295-3-alexis.lothore@bootlin.com/
[2]: https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/Documentation/RCU/whatisRCU.rst#n161
---
 drivers/pci/p2pdma.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
index 0c361561b855..4f47a13cb500 100644
--- a/drivers/pci/p2pdma.c
+++ b/drivers/pci/p2pdma.c
@@ -661,7 +661,7 @@ calc_map_type_and_dist(struct pci_dev *provider, struct pci_dev *client,
 	p2pdma = rcu_dereference(provider->p2pdma);
 	if (p2pdma)
 		xa_store(&p2pdma->map_types, map_types_idx(client),
-			 xa_mk_value(map_type), GFP_KERNEL);
+			 xa_mk_value(map_type), GFP_ATOMIC);
 	rcu_read_unlock();
 	return map_type;
 }
-- 
2.34.1

Re: [PATCH] PCI/P2PDMA: Fix a sleeping issue in a RCU read section

[Logan Gunthorpe]

On 2024-01-04 12:52, Christophe JAILLET wrote:
> It is not allowed to sleep within a RCU read section, so use GFP_ATOMIC
> instead of GFP_KERNEL here.
> 
> Fixes: ae21f835a5bd ("PCI/P2PDMA: Finish RCU conversion of pdev->p2pdma")
> Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>

This makes sense to me. Though, the use of RCU could probably use a
review. Seeing p2pdma is only released through a devm action on the
pdev, I would think it shouldn't be needed if we hold a reference to the
pdev.

Other than that:

Reviewed-by: Logan Gunthorpe <logang@deltatee.com>

Thanks!

Logan

Re: [PATCH] PCI/P2PDMA: Fix a sleeping issue in a RCU read section

[Dan Carpenter]

On Thu, Jan 04, 2024 at 08:52:35PM +0100, Christophe JAILLET wrote:
> It is not allowed to sleep within a RCU read section, so use GFP_ATOMIC
> instead of GFP_KERNEL here.
> 
> Fixes: ae21f835a5bd ("PCI/P2PDMA: Finish RCU conversion of pdev->p2pdma")
> Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
> ---
> This patch is speculative.
> It is based on a discussion related to another patch. See [1].
> 
> It also matches the doc, IIUC. See [2]
> 
> [1]: https://lore.kernel.org/all/20240104143925.194295-3-alexis.lothore@bootlin.com/
> [2]: https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/Documentation/RCU/whatisRCU.rst#n161

Looks good to me.

Smatch is supposed to catch this sort of bug but there are some issues
with xa_store it's normally holding the the xas_lock() but if you pass
a sleeping GFP_ then it drops the lock in __xas_nomem().  Sort of
tricky.  So right now all xa_store() stuff triggers a false positive
but tomorrows version of Smatch will just miss this bug.

regards,
dan carpenter

Re: [PATCH] PCI/P2PDMA: Fix a sleeping issue in a RCU read section

[Bjorn Helgaas]

On Thu, Jan 04, 2024 at 08:52:35PM +0100, Christophe JAILLET wrote:
> It is not allowed to sleep within a RCU read section, so use GFP_ATOMIC
> instead of GFP_KERNEL here.
> 
> Fixes: ae21f835a5bd ("PCI/P2PDMA: Finish RCU conversion of pdev->p2pdma")
> Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>

Applied with Logan's Reviewed-by to pci/p2pdma for v6.9, thanks!

> ---
> This patch is speculative.
> It is based on a discussion related to another patch. See [1].
> 
> It also matches the doc, IIUC. See [2]
> 
> [1]: https://lore.kernel.org/all/20240104143925.194295-3-alexis.lothore@bootlin.com/
> [2]: https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/Documentation/RCU/whatisRCU.rst#n161
> ---
>  drivers/pci/p2pdma.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
> index 0c361561b855..4f47a13cb500 100644
> --- a/drivers/pci/p2pdma.c
> +++ b/drivers/pci/p2pdma.c
> @@ -661,7 +661,7 @@ calc_map_type_and_dist(struct pci_dev *provider, struct pci_dev *client,
>  	p2pdma = rcu_dereference(provider->p2pdma);
>  	if (p2pdma)
>  		xa_store(&p2pdma->map_types, map_types_idx(client),
> -			 xa_mk_value(map_type), GFP_KERNEL);
> +			 xa_mk_value(map_type), GFP_ATOMIC);
>  	rcu_read_unlock();
>  	return map_type;
>  }
> -- 
> 2.34.1
>