-===================- Kernel Janitor TODO list. $Id: TODO,v 1.52 2006/08/29 04:09:25 dobriyan Exp $ -===================- Before attacking any of these, we suggest sending a few patches in advance to see if you are doing something wrong, or if someone else is already doing same work. Read related threads in the mailing list archive. Where it makes sense, sections are supposed to be ordered by increasing difficulty. Send patches that add/fix items to kernel-janitors@lists.osdl.org. Links should be marked with: D: description/information about the issue E: example patch ----------------------------------------------------------------------------- NOTE: Search & Replace Sometimes all your patch will do is simply changing one function to another. Doing this is especially tempting in "convert to new API" section. History shows this is often wrong: old bugs are not fixed, new bugs are introduced. If all your patch does is search & replace, double check. If in doubt, ask on mailing list. NOTE: scripts/kj.pl was a weekend hack. Don't take it too seriously. - add more checks to kj.pl ----------------------------------------------------------------------------- Audit return codes (and handle failure correctly) for: - request_region() - request_irq() - kmalloc(), vmalloc(), etc - register_netdev() has to be checked as well - misc_register() (yes, it can fail, murphy's law applies here as well) - scsi_register() - create_proc_*() - pci_map_* might return 0 for a valid mapping. Some code tests mapping for a non-zero value, which is incorrect. - __get_free_pages() and __get_free_page() - ioremap() -- Some are using this as a pointer, which is wrong. - put/get_user() - copy_*_user(), unlike put_user() it returns number of bytes failed to copy. - kernel_thread() - IDE drivers, as it can cause a real problem: [D: http://marc.theaimsgroup.com/?l=kernel-janitor-discuss&m=110416923101031] ----------------------------------------------------------------------------- Balancing functions. Make sure calls to certain functions are matched by the relevant function at other end of the function, and also during ALL failure/early return paths. Examples: - pci_alloc_consistent() and pci_free_consistent() - ioremap() must be balanced by an iounmap() (this causes a memory leak). - check for balanced *_lock_* and *_unlock_* along all paths in a function (can cause dead-lock if not). - Make sure no-one is freeing skbs with kfree instead of kfree_skb - check that net_device interrupt functions use dev_kfree_skb_irq and not just dev_kfree_skb ----------------------------------------------------------------------------- Remove unneeded historic code / New API conversions. - Code that depends on LINUX_VERSION_CODE & KERNEL_VERSION < 2.6 can be deleted in most cases. (Mostly/all? done) - checking for NULL on probe routines for net drivers - get rid of init_module / cleanup_module - call SET_MODULE_OWNER or use .owner = THIS_MODULE instead of using MOD_{INC,DEC}_USE_COUNT (only a few remain) - Use pci_set_drvdata() to set dev->driver_data, likewise use pci_get_drvdata() instead of reading directly. - dev->driver_data = pcigame; + pci_set_drvdata(dev, pcigame); - get rid of access_ok with copy_*_user get/put_user, only needed if using __copy_*_user et al - MODULE_PARM must die: [D: http://marc.theaimsgroup.com/?l=linux-kernel&m=109826168201622&w=2] [E: http://www.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.10-rc1/2.6.10-rc1-mm2/broken-out/remove-module_parm-from-allyesconfig-almost.patch] - convert drivers to new PCI API - get rid of check_region, use just request_region checking its return (2.2 request_region returned void) and now the driver init sequence is not to be serialized anymore, so races are possible (look at cardbus/pcihotplug code) [D: http://lists.osdl.org/pipermail/kernel-janitors/2004-January/000346.html] - ALL PCI drivers should call pci_enable_device --before-- reading pdev->irq or pdev->resource[]. irq and resource[] may not have correct values until after PCI hotplug setup occurs at pci_enable_device() time. Many PCI drivers need to be evaluated and checked for this. Many of these fixed already at: http://marc.theaimsgroup.com/?l=linux-kernel&m=98449691919682&w=3 - convert cli/sti/save_flags/save_flags_cli/restore_flags usage to accepted locking primitives; see Documentation/cli-sti-removal.txt and Documentation/spinlocks.txt for details. [E: http://linux.bkbits.net:8080/linux-2.6/cset@419b78fcaLKtZt39OF_dslTbk-b3KQ?nav=index.html|ChangeSet@-8w] [This item is a hard one. Full understanding of what data structures should be protected by introduced locks is required. Testing on real hardware is almost mandatory.] - yield() considered harmful: [D: http://marc.theaimsgroup.com/?t=106554796200005] Don't forget to set_current_state(TASK_{,UN}INTERRUPTIBLE); before schedule*(). Consider using msleep();. - Anything which uses sleep_on*() has a 90% chance of being broken. Fix them all, because we want to remove sleep_on() and friends. [D: http://boudicca.tux.org/hypermail/linux-kernel/2001week05/0305.html] Be sure to read replies too. - Callers of schedule_timeout() who pass in an absolute constant (i.e. with no reference to HZ) may be broken, especially with a variable HZ value. Verify which behavior is intended: if the caller needs the shortest delay possible, then schedule_timeout(1) is correct and a comment can be added; if they actually need a certain amount of time, then use msleep() or schedule_timeout with an appropriate {msecs,usecs,nsecs}_to_jiffies() call. - Replace (un)register_ioctl32_conversion with ioctl_compat: [D: http://lkml.org/lkml/2005/1/5/106] ----------------------------------------------------------------------------- - clean up Documentation/DocBook/ , fix warnings that occur on make *docs. - printk() calls should include appropriate KERN_* constant (of course only at beginning of lines). Also consider using dev_printk() and friends. - pr_debug() from kernel.h could replace a lot of DPRINTK and similar macros. - same for pr_info() - Lots of unnecessary casts in drivers can be removed. Example: - struct netdev_private *np = (struct netdev_private *)dev->priv; + struct netdev_private *np = dev->priv; use netdev_priv() for network devices - min/max macros from kernel.h are safe, a lot of handcrafted MIN/MAX are not. - ARRAY_SIZE macro has duplicates, remove them. - put BIT macro into kernel.h (?) and remove it from drivers - Purely cosmetic, but far nicer to read. - for (list = ymf_devs.next; list != &ymf_devs; list = list->next) { + list_for_each(list, &ymf_devs) { - Convert comments to C99 initializers: - /* 10 */ DECLARE_PIIX_DEV("ICH2"), + [10] = DECLARE_PIIX_DEV("ICH2"), (if order matters) or - 15, /* foo */ + .foo = 15, - make non-global functions static. - convert all explicit lock initializations to spin_lock_init() or rwlock_init(). (Besides consistency this also helps automatic lock validators and debugging code.) [D: http://lwn.net/Articles/109505/] [E: http://linux.bkbits.net:8080/linux-2.6/cset@419a6f292wHnthuDzw7VfgECNLmvLg?nav=index.html|ChangeSet@-8w] - fix compilation warnings/errors - pci_set_dma_mask() and friends should use DMA_{32,64}BIT_MASK instead of 0xffff... This is not 2.4 compatible, so beware of drivers with same code. [D: http://marc.theaimsgroup.com/?t=108001993000001] Don't forget to #include dma-mapping.h - check kmallocs for things like GFP_DMA without a memtype. - check that a freed pointer (*kfree_*) is not used again. - make sure BUG() is used correctly (i.e. if(function()) BUG(); is evil) i.e. even when no-op-ing BUG we still have an if (See also: BUG_ON) - Code in wrong sections (ie. __init code called from __exit): Run `make buildcheck` to find offending code. (read_eeprom in net drivers is a good candidate for __init) [D: http://marc.theaimsgroup.com/?l=kernel-janitor-discuss&m=108640220401558] - Fix gcc 4 warnings. [D: http://marc.theaimsgroup.com/?t=110348353800003] Note: Set CONFIG_DEBUG_INFO=n to make compiles faster and smaller. - Audit all memcpy and memmove calls. If areas overlap memmove should be used, if they don't memcpy. - drivers/char/i8k.c duplicates arch/i386/kernel/dmi_scan.c Better would be to have dmi_scan.c set a flag which the i8k driver checks. - sleeping functions should not be called in interrupts or under spinlocks: copy_to/from_user(), put/get_user(), schedule(), kmalloc() except for GPF_ATOMIC, functions with __might_sleep() in body... - go through all the tty/serial drivers and make sure they don't give out excessively useful information to non CAP_SYS_RAWIO users, then loosen permissions. [D: http://lkml.org/lkml/2005/1/17/94] - check that buffers used in copy_to_user() don't leak information. - check for dev_close calls without rntl_lock held (causes assertion failures). - timer_del() vs. timer_del_sync() [D: http://www.uwsg.iu.edu/hypermail/linux/kernel/0005.3/0269.html] A lot of the timer deletion races are hard to fix because of the deadlock problem. - fix watchdog drivers to use link order rather than explicit initialization calls (i810 is particularly broken) - check stack usage (make checkstack) and reduce it in the worst cases. [D: http://bugme.osdl.org/show_bug.cgi?id=386] [D: http://marc.theaimsgroup.com/?l=kernel-janitor-discuss&m=110661227615538] ----------------------------------------------------------------------------- SUSPECTS: copy_to/from_user: - look at drivers/char/n_tty.c Items with longer description in no particular order: ----------------------------------------------------------------------------- Date: Wed, 14 Feb 2001 12:51:53 +0100 From: Manfred Spraul bug in network drivers: * dev->mem_start: NULL means "not command line configuration" 0xffffffff means "default". several drivers only check for NULL, not for 0xffffffff. And then: Date: Wed, 14 Feb 2001 05:54:34 -0600 (CST) From: Jeff Garzik netdev->mem_start is unsigned long... Should the test be for ~0 instead? The value 0xFFFFFFFF seems wrong for 64-bit machines. ----------------------------------------------------------------------------- - drivers that try to find multiple boards, possibly successfully allocating for the first ones, then failing for, lets say, the third board, then it just returns failure for init_module, the module is unloaded but the resources remain allocated... in these cases we need to rollback the allocations, freeing it before returning from init_module or equivalent. - sound/oss/maestro3.c doesn't pci_free_consistent any buffers if one allocation fails, but others succeeded. ----------------------------------------------------------------------------- From: Hans Grobler - audit ioctl functions to make sure there is no way a user can crash the machine or do sensitive stuff. For example, check for the necessary capabilities. Check that no userspace pointers are accessed directly. be careful with this... it takes a while to figure out which ioctls are covered by the parent ioctl function (these can be nested to great depths). ----------------------------------------------------------------------------- Date: Mon, 26 Feb 2001 16:53:50 -0800 (PST) From: "David S. Miller" Andrew Morton writes: > Could this be a driver problem? > netif_rx(skb); ... > new_skb = skb; This is illegal and broken and will never work. Once you give an skb to netif_rx() it is not yours to reference any longer. Donald Becker added: Easier fix: - np->stats.rx_bytes += skb->len; + np->stats.rx_bytes += pkt_len; Grouping the writes to np->stats results in better cache usage. Jeff Garzik added: It makes use of the existing local 'pkt_len', and it checks off another item. ----------------------------------------------------------------------------- From: Jeff Garzik > It is highly recommended to always compile with CONFIG_ISAPNP=y due > to these differences. If you grep around for CONFIG_ISAPNP versus > CONFIG_ISAPNP_MODULE, you'll see that many drivers are woefully > unprepared for isapnp support compiled as a module. Yep.. grep for CONFIG_ISAPNP, look at the code, and evaluate it to make sure that isapnp works for that drivers regardless of whether CONFIG_ISAPNP -or- CONFIG_ISAPNP_MODULE is defined. ----------------------------------------------------------------------------- From Dave Jones - Lots of drivers are doing evil looking things writing to PCI_CACHE_LINE_SIZE This should be done during PCI initialisation. - Many drivers have alignments such as.. dev->priv = (void *)(((unsigned long)dev->priv + 7) & ~7); jgarzik: What they are doing here is DMA'ing into their private board (dev->priv), which is wrong -- they should be using PCI DMA instead. They can't use init_etherdev because it doesn't pass GFP_DMA to kmalloc for dev->priv, and since they are DMA'ing into dev->priv, they have to do things manually. - Signedness issues. We have _lots_ of these. Adding a -W to the kernel build line shows them. We have lots of code paths that are not executed because comparisons <0 on an unsigned int always evaluates to true. We also need to check for things like... char foo[4]={1,2,3,133}; /* 133 in signed char!*/ ----------------------------------------------------------------------------- From: Jeff Garzik 1) The string form [const] char *foo = "blah"; creates two variables in the final assembly output, a static string, and a char pointer to the static string. The alternate string form [const] char foo[] = "blah"; is better because it declares a single variable. For variables marked __initdata, the "*foo" form causes only the pointer, not the string itself, to be dropped from the kernel image, which is a bug. Using the "foo[]" form with regular 'ole local variables also makes the assembly shorter. 2) "unsigned int" is preferred to "int", it generates better asm code on all platforms except sh5. This replacement needs to be done manually, because often 'int' is required due to negative values -Exxx commonly passed as error values. ----------------------------------------------------------------------------- From: Greg KH - delete all pci_find_* functions from the kernel tree. Instead of pci_find_device use pci_get_device().. > Looking at pci.txt, it appears that if I use pci_get_device(), I will > also need to use pci_dev_put(). also check using the proper shutdown callback. - convert from pci_module_init to pci_register_driver - don't do it for drivers that are 2.4 compatible ----------------------------------------------------------------------------- 1. Fix sparse warnings. Read section "Where to get sparse" in Documentation/sparse.txt. $ make allyesconfig or whatever config you like. Set CONFIG_DEBUG_INFO=n to make compiles faster and smaller. $ make -k C=1 2>&1 | tee ../W_sparse or $ make -k C=1 CF=-D__CHECK_ENDIAN__ 2>&1 | tee ../W_sparse _Plenty_ of patches here (well, not all of them fix sparse warnings ;-) ): [E: http://linux.bkbits.net:8080/linux-2.6/search/?expr=sparse&search=ChangeSet+comments] 2. For portability reasons pointers should be printed as ("%p", ptr), not ("%08x", (int) ptr) or similar. 3. $ make randconfig $ make -k 2>&1 | tee ../W_randconfig Find out why it doesn't build. Fix warnings that aren't seen with usual "allyesconfig", "allmodconfig", ... P. S.: Run it several times, think a little and take bets on whether it'll build or not on the next party among local linuxoids. :-) 4. Some drivers do: if ((jiffies - data->last_updated > HZ * 2) || (jiffies < data->last_updated)) Should be: #include if (time_after(jiffies, data->last_updated + HZ * 2)) [E: http://linux.bkbits.net:8080/linux-2.6/gnupatch@42261f25iCdhlwTMywIQxvBDrIHj1A] 5. Some function prototypes (in both .h and .c files) specify attributes like __init and __exit in the prototype. gcc (at least at 3.3.3) uses the last such attribute that is actually specified, without issuing a warning. So we can have: * Prototype declarations that use one attribute and a function body that uses another attribute. * Functions that from the .c code appear to be normal text but the .h file is silently setting a special attribute. Both are potential sources of programmer confusion or bugs. Identify: * all places where attributes in prototype and function definition don't match (choose the correct one, move it to declaration), * all places where prototype contains attribute but definition doesn't (move it to declaration). The same task should be done for extern data declarations. Once that is done, remove #include from all .h files. Only .[cS] files should specify which section the data and text are stored in, .h files should only define the C language information.