Fix release function in IPMI device model
PATCH 1/1 Fix release function in IPMI device model
Quoted 9 lines · Corey Minyard
Review
Quoted 4 lines
(quoted diff) +4
-0+static void ipmi_bmc_release(struct device *dev) +{ + printk(KERN_DEBUG "ipmi_bmc release\n"); +}eehhhh NO.
Please read the many comments and documentations about why a release function is NOT allowed to be empty. In fact the kernel warned you about that, didn't it?
Arjun, is this ok? I'm not 100% sure I understand all the subtleties of the device model, but I assume all I have to do is keep around the device memory until the release function is called. Also, the device might come from somplace else (if it is on the PCI bus, for instance), so I just free the memory in that case since the memory for the device structure didn't come from this memory.
And "many comments" might be stretching it. I found one comment in the device model documentataion directory, nothing in any of the include files describing the device model. I should have looked more carefully when converting the patch over from Yani, but something non-obvious and import like this deserves a comment in both the documentation and include file by the release function, something like:
/* IMPORTANT: You must define a release function, and the memory for the
device structure *MUST* be kept around until the release function is called. */
Here's the patch...
Arjun van de Ven pointed out that the changeover to the driver model in the IPMI driver had some bugs in it dealing with the release function and cleanup. This patch fixes that and also adds a semaphore around the BMC functions and converts the BMC counter to a kref, which I meant to do earlier, but didn't write down :(.
drivers/char/ipmi/ipmi_msghandler.c +47 -29
@@ -168,7 +168,8 @@ struct bmc_devicestruct ipmi_device_id id;unsigned char guid[16 ];int guid_set; -int interfaces; + +struct kref refcount;/* bmc device attributes */ struct device_attribute device_id_attr;
@@ -352,6 +353,7 @@ static struct device_driver ipmidriver = .name = "ipmi", .bus = &platform_bus_type }; +static DECLARE_MUTEX (ipmidriver_sem);#define MAX_IPMI_INTERFACES 4static ipmi_smi_t ipmi_interfaces[MAX_IPMI_INTERFACES ];
@@ -1800,7 +1802,7 @@ static struct bmc_device *ipmi_find_bmc_static void ipmi_bmc_release (struct device *dev) { -printk (KERN_DEBUG "ipmi_bmc release\n" ); +kfree (to_bmc_device (to_platform_device (dev))); } static ssize_t device_id_show(struct device *dev,
@@ -1913,6 +1915,38 @@ static ssize_t guid_show(struct device * (long long) bmc->guid[8]); } +static void +cleanup_bmc_device (struct kref *ref) +{ +struct bmc_device *bmc; + + bmc = container_of(ref, struct bmc_device, refcount); + +device_remove_file (&bmc->dev.dev, + &bmc->device_id_attr); +device_remove_file (&bmc->dev.dev, + &bmc->provides_dev_sdrs_attr); +device_remove_file (&bmc->dev.dev, + &bmc->revision_attr); +device_remove_file (&bmc->dev.dev, + &bmc->firmware_rev_attr); +device_remove_file (&bmc->dev.dev, + &bmc->version_attr); +device_remove_file (&bmc->dev.dev, + &bmc->add_dev_support_attr); +device_remove_file (&bmc->dev.dev, + &bmc->manufacturer_id_attr); +device_remove_file (&bmc->dev.dev, + &bmc->product_id_attr); +if (bmc->id.aux_firmware_revision_set) +device_remove_file (&bmc->dev.dev, + &bmc->aux_firmware_rev_attr); +if (bmc->guid_set) +device_remove_file (&bmc->dev.dev, + &bmc->guid_attr); +platform_device_unregister (&bmc->dev); +} +static void ipmi_bmc_unregister (ipmi_smi_t intf) {struct bmc_device *bmc = intf->bmc;
@@ -1924,31 +1958,9 @@ static void ipmi_bmc_unregister(ipmi_smi intf->my_dev_name =NULL ; } - bmc->interfaces--; -if (bmc->interfaces ==0 ) { -device_remove_file (&bmc->dev.dev, - &bmc->device_id_attr); -device_remove_file (&bmc->dev.dev, - &bmc->provides_dev_sdrs_attr); -device_remove_file (&bmc->dev.dev, - &bmc->revision_attr); -device_remove_file (&bmc->dev.dev, - &bmc->firmware_rev_attr); -device_remove_file (&bmc->dev.dev, - &bmc->version_attr); -device_remove_file (&bmc->dev.dev, - &bmc->add_dev_support_attr); -device_remove_file (&bmc->dev.dev, - &bmc->manufacturer_id_attr); -device_remove_file (&bmc->dev.dev, - &bmc->product_id_attr); -device_remove_file (&bmc->dev.dev, - &bmc->aux_firmware_rev_attr); -if (bmc->guid_set) -device_remove_file (&bmc->dev.dev, - &bmc->guid_attr); -platform_device_unregister (&bmc->dev); - } +down (&ipmidriver_sem); +kref_put (&bmc->refcount, cleanup_bmc_device); +up (&ipmidriver_sem); }static int ipmi_bmc_register (ipmi_smi_t intf)
@@ -1959,6 +1971,8 @@ static int ipmi_bmc_register(ipmi_smi_tint size;char dummy[1 ]; +down (&ipmidriver_sem); +/* * Try to find if there is an bmc_device struct * representing the interfaced BMC already
@@ -1979,7 +1993,8 @@ static int ipmi_bmc_register(ipmi_smi_t intf->bmc = old_bmc; bmc = old_bmc; - bmc->interfaces++; +kref_get (&bmc->refcount); +up (&ipmidriver_sem);printk (KERN_INFO "ipmi: interfacing existing BMC (man_id: 0x%6.6x,"
@@ -1992,14 +2007,17 @@ static int ipmi_bmc_register(ipmi_smi_t bmc->dev.id = bmc->id.device_id; bmc->dev.dev.driver = &ipmidriver; bmc->dev.dev.release = ipmi_bmc_release; - bmc->interfaces =1 ; +kref_init (&bmc->refcount); rv =platform_device_register (&bmc->dev); +up (&ipmidriver_sem);if (rv) {printk (KERN_ERR "ipmi_msghandler:" " Unable to register bmc device: %d\n" , rv); +/* Don't go to out_err, you can only do that if +the device is registered already. */ return rv; }
drivers/char/ipmi/ipmi_si_intf.c +17 -13
@@ -231,6 +231,7 @@ struct smi_infostruct list_head link; }; +#define to_smi_device (device) container_of(device, struct smi_info, idev)static int try_smi_init (struct smi_info *smi);
@@ -1161,7 +1162,6 @@ static void port_cleanup(struct smi_inforelease_region (addr, mapsize); } -kfree (info); }static int port_setup (struct smi_info *info)
@@ -1270,7 +1270,6 @@ static void mem_cleanup(struct smi_inforelease_mem_region (addr, mapsize); } -kfree (info); }static int mem_setup (struct smi_info *info)
@@ -2235,7 +2234,7 @@ static int is_new_interface(struct smi_istatic void si_release (struct device *dev) { -printk (KERN_DEBUG "ipmi_si release\n" ); +kfree (to_smi_device (to_platform_device (dev))); }static int try_smi_init (struct smi_info *new_smi)
@@ -2395,7 +2394,7 @@ static int try_smi_init(struct smi_infoprintk (KERN_ERR "ipmi_si: Unable to register device: error %d\n" , rv); - goto out_err_unreg_dev; + goto out_err_stop_timer; } rv = ipmi_smi_add_proc_entry(new_smi->intf, "type",
@@ -2405,7 +2404,7 @@ static int try_smi_init(struct smi_infoprintk (KERN_ERR "ipmi_si: Unable to create proc entry: %d\n" , rv); - goto out_err_unreg_dev; + goto out_err_stop_timer; } rv = ipmi_smi_add_proc_entry(new_smi->intf, "si_stats",
@@ -2415,7 +2414,7 @@ static int try_smi_init(struct smi_infoprintk (KERN_ERR "ipmi_si: Unable to create proc entry: %d\n" , rv); - goto out_err_unreg_dev; + goto out_err_stop_timer; }list_add_tail (&new_smi->link, &smi_infos);
@@ -2426,10 +2425,6 @@ static int try_smi_init(struct smi_inforeturn 0 ; - out_err_unreg_dev: -if (new_smi->dev_registered) -platform_device_unregister (&new_smi->idev); - out_err_stop_timer:atomic_inc (&new_smi->stop_operation);wait_for_timer_and_thread (new_smi);
@@ -2456,6 +2451,11 @@ static int try_smi_init(struct smi_infoif (new_smi->io_cleanup) new_smi->io_cleanup (new_smi); +if (new_smi->dev_registered) +platform_device_unregister (&new_smi->idev); +else +kfree (new_smi); +up (&smi_infos_lock);return rv;
@@ -2583,9 +2583,6 @@ static void __devexit cleanup_one_si(str rv); } -if (to_clean->dev_registered) -platform_device_unregister (&to_clean->idev); - to_clean->handlers->cleanup (to_clean->si_sm);kfree (to_clean->si_sm);
@@ -2594,6 +2591,13 @@ static void __devexit cleanup_one_si(str to_clean->addr_source_cleanup (to_clean);if (to_clean->io_cleanup) to_clean->io_cleanup (to_clean); + +if (to_clean->dev_registered) +/* We registered the device, let the release free it. */ +platform_device_unregister (&to_clean->idev); +else +/* The device came from elsewhere, we free now. */ +kfree (to_clean); } static __exit void cleanup_ipmi_si(void)
drivers/char/ipmi/ipmi_msghandler.c static struct bmc_device *ipmi_find_bmc_ · line 1808
Quoted 9 lines · Corey Minyard
drivers/char/ipmi/ipmi_msghandler.c +1
-1@@ -1800,7 +1802,7 @@ static struct bmc_device *ipmi_find_bmc_static void ipmi_bmc_release (struct device *dev) { -printk (KERN_DEBUG "ipmi_bmc release\n" ); +kfree (to_bmc_device (to_platform_device (dev))); } static ssize_t device_id_show(struct device *dev,
Okay, so the release function is in the same module as the code which is unregistering it. No, wrong.
What this means is that if you rmmod this module, but userspace has some file associated with the platform device still open, and then it drops that refcount, your release function will be called.
But wait, the module code has been unmapped. Instant oops.
Use the additions to the platform device API to sanely handle platform devices, and don't try to wrap a platform device up into some other data structure.
There's comments on lkml about this additional platform device API (search for platform_device_alloc) or in the git commit comments. I don't have any references to either to hand, sorry.
Signature · 3 lines
Russell King Linux kernel 2.6 ARM Linux - http://www.arm.linux.org.uk/ maintainer of: 2.6 Serial core