Re: [GIT PULL] NTB bug fixes for v5.11

From: Dan Carpenter
Date: Mon Jan 04 2021 - 03:31:46 EST


On Sun, Dec 27, 2020 at 09:38:23AM -0800, Linus Torvalds wrote:
> On Sun, Dec 27, 2020 at 6:16 AM Jon Mason <jdmason@xxxxxxxx> wrote:
> >
> > Wang Qing (1):
> > ntb: idt: fix error check in ntb_hw_idt.c
>
> So this patch seems to be at least partially triggered by a smatch
> warning that is a bit questionable.
>
> This part:
>
> if (IS_ERR_OR_NULL(dbgfs_topdir)) {
> dev_info(&ndev->ntb.pdev->dev, "Top DebugFS directory absent");
> - return PTR_ERR(dbgfs_topdir);
> + return PTR_ERR_OR_ZERO(dbgfs_topdir);
> }
>
> works, but is very non-optimal and unnecessary.
>
> The thing is, "PTR_ERR()" works just fine on a IS_ERR_OR_NULL pointer.
> It doesn't work on a _regular_ non-NULL and non-ERR pointer, and will
> return random garbage for those. But if you've tested for
> IS_ERR_OR_NULL(), then a regular PTR_ERR() is already fine.
>
> And PTR_ERR_OR_ZERO() potentially generates an extraneous pointless
> tests against zero (to check for the ERR case).
>
> A compiler may be able to notice that the PTR_ERR_OR_ZERO() is
> unnecessary and remove it (because of the IS_ERR_OR_NULL() checks),
> but in general we should assume compilers are "not stupid" rather than
> "really smart".
>
> So while this patch isn't _wrong_, and I've already pulled it, the
> fact that apparently some smatch test triggers these pointless and
> potentially expensive patches is not a good idea.
>
> I'm not sure what the smatch tests should be (NULL turns to 0, which
> may be confusing), but I'm cc'ing Dan in case he has ideas.
>

The most common bug that this check finds is the other part of that same
commit 91b8246de859 ("ntb: idt: fix error check in ntb_hw_idt.c"):

/* Allocate the memory for IDT NTB device data */
ndev = idt_create_dev(pdev, id);
- if (IS_ERR_OR_NULL(ndev))
+ if (IS_ERR(ndev))
return PTR_ERR(ndev);

idt_create_dev() never returns NULL, but if it did then we don't want
to return success.

For the debugfs stuff, the caller doesn't check the return value anyway.
Just make it a void function. A lot of this debugfs code could be
simplified. It's not a bug to pass an error pointer or a NULL dbgfs_topdir
pointer to debugfs_create_file(). There isn't any benefit in checking
debugfs_initialized().

diff --git a/drivers/ntb/hw/idt/ntb_hw_idt.c b/drivers/ntb/hw/idt/ntb_hw_idt.c
index e7a4c2aa8baa..710c17b2a923 100644
--- a/drivers/ntb/hw/idt/ntb_hw_idt.c
+++ b/drivers/ntb/hw/idt/ntb_hw_idt.c
@@ -2504,28 +2504,14 @@ static ssize_t idt_dbgfs_info_read(struct file *filp, char __user *ubuf,
*
* Return: zero on success, otherwise a negative error number.
*/
-static int idt_init_dbgfs(struct idt_ntb_dev *ndev)
+static void idt_init_dbgfs(struct idt_ntb_dev *ndev)
{
char devname[64];

- /* If the top directory is not created then do nothing */
- if (IS_ERR_OR_NULL(dbgfs_topdir)) {
- dev_info(&ndev->ntb.pdev->dev, "Top DebugFS directory absent");
- return PTR_ERR_OR_ZERO(dbgfs_topdir);
- }
-
/* Create the info file node */
snprintf(devname, 64, "info:%s", pci_name(ndev->ntb.pdev));
ndev->dbgfs_info = debugfs_create_file(devname, 0400, dbgfs_topdir,
- ndev, &idt_dbgfs_info_ops);
- if (IS_ERR(ndev->dbgfs_info)) {
- dev_dbg(&ndev->ntb.pdev->dev, "Failed to create DebugFS node");
- return PTR_ERR(ndev->dbgfs_info);
- }
-
- dev_dbg(&ndev->ntb.pdev->dev, "NTB device DebugFS node created");
-
- return 0;
+ ndev, &idt_dbgfs_info_ops);
}

/*
@@ -2792,7 +2778,7 @@ static int idt_pci_probe(struct pci_dev *pdev,
goto err_deinit_isr;

/* Initialize DebugFS info node */
- (void)idt_init_dbgfs(ndev);
+ idt_init_dbgfs(ndev);

/* IDT PCIe-switch NTB driver is finally initialized */
dev_info(&pdev->dev, "IDT NTB device is ready");
@@ -2904,9 +2890,7 @@ static int __init idt_pci_driver_init(void)
{
pr_info("%s %s\n", NTB_DESC, NTB_VER);

- /* Create the top DebugFS directory if the FS is initialized */
- if (debugfs_initialized())
- dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);
+ dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);

/* Register the NTB hardware driver to handle the PCI device */
return pci_register_driver(&idt_pci_driver);
--
2.29.2