| Message ID | 20200325120345.12946-1-vbabka@suse.cz (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [RFC,v2,1/2] kernel/sysctl: support setting sysctl parameters from kernel command line | expand |
On Wed, Mar 25, 2020 at 01:03:44PM +0100, Vlastimil Babka wrote: > A recently proposed patch to add vm_swappiness command line parameter in > addition to existing sysctl [1] made me wonder why we don't have a general > support for passing sysctl parameters via command line. Googling found only > somebody else wondering the same [2], but I haven't found any prior discussion > with reasons why not to do this. > > Settings the vm_swappiness issue aside (the underlying issue might be solved in > a different way), quick search of kernel-parameters.txt shows there are already > some that exist as both sysctl and kernel parameter - hung_task_panic, > nmi_watchdog, numa_zonelist_order, traceoff_on_warning. A general mechanism > would remove the need to add more of those one-offs and might be handy in > situations where configuration by e.g. /etc/sysctl.d/ is impractical. > Also after 61a47c1ad3a4 ("sysctl: Remove the sysctl system call") the only way > to set sysctl is via procfs, so this would eventually allow small systems to be > built without CONFIG_PROC_SYSCTL and still be able to change sysctl parameters. > > Hence, this patch adds a new parse_args() pass that looks for parameters > prefixed by 'sysctl.' and searches for them in the sysctl ctl_tables. When > found, the respective proc handler is invoked. The search is just a naive > linear one, to avoid using the whole procfs layer. It should be acceptable, > as the cost depends on number of sysctl. parameters passed. > > The main limitation of avoiding the procfs layer is however that sysctls > dynamically registered by register_sysctl_table() or register_sysctl_paths() > cannot currently be set by this method. > > The processing is hooked right before the init process is loaded, as some > handlers might be more complicated than simple setters and might need some > subsystems to be initialized. At the moment the init process can be started and > eventually execute a process writing to /proc/sys/ then it should be also fine > to do that from the kernel. > > [1] https://lore.kernel.org/linux-doc/BL0PR02MB560167492CA4094C91589930E9FC0@BL0PR02MB5601.namprd02.prod.outlook.com/ > [2] https://unix.stackexchange.com/questions/558802/how-to-set-sysctl-using-kernel-command-line-parameter > > Signed-off-by: Vlastimil Babka <vbabka@suse.cz> > --- > v2: - handle any nesting level of parameter name > - add Documentation/admin-guide/kernel-parameters.txt blurb > - alias support for legacy one-off parameters, with first conversion (patch 2) > - still no support for dynamically registed sysctls > > .../admin-guide/kernel-parameters.txt | 9 +++ > include/linux/sysctl.h | 1 + > init/main.c | 21 +++++++ > kernel/sysctl.c | 62 +++++++++++++++++++ > 4 files changed, 93 insertions(+) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index c07815d230bc..5076e288f93f 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -4793,6 +4793,15 @@ > > switches= [HW,M68k] > > + sysctl.*= [KNL] > + Set a sysctl parameter right before loading the init > + process, as if the value was written to the respective > + /proc/sys/... file. Currently a subset of sysctl > + parameters is supported that is not registered > + dynamically. Unrecognized parameters and invalid values > + are reported in the kernel log. > + Example: sysctl.vm.swappiness=40 > + > sysfs.deprecated=0|1 [KNL] > Enable/disable old style sysfs layout for old udev > on older distributions. When this option is enabled > diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h > index 02fa84493f23..62ae963a5c0c 100644 > --- a/include/linux/sysctl.h > +++ b/include/linux/sysctl.h > @@ -206,6 +206,7 @@ struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path, > void unregister_sysctl_table(struct ctl_table_header * table); > > extern int sysctl_init(void); > +int process_sysctl_arg(char *param, char *val, const char *unused, void *arg); > > extern struct ctl_table sysctl_mount_point[]; > > diff --git a/init/main.c b/init/main.c > index ee4947af823f..74a094c6b8b9 100644 > --- a/init/main.c > +++ b/init/main.c > @@ -1345,6 +1345,25 @@ void __weak free_initmem(void) > free_initmem_default(POISON_FREE_INITMEM); > } > > +static void do_sysctl_args(void) > +{ > +#ifdef CONFIG_SYSCTL > + size_t len = strlen(saved_command_line) + 1; > + char *command_line; > + > + command_line = kzalloc(len, GFP_KERNEL); > + if (!command_line) > + panic("%s: Failed to allocate %zu bytes\n", __func__, len); > + > + strcpy(command_line, saved_command_line); No need to open-code this: char *command_line; command_line = kstrdup(saved_command_line, GFP_KERNEL); if (!command_line) panic("%s: Failed to allocate %zu bytes\n", __func__, len); > + > + parse_args("Setting sysctl args", command_line, > + NULL, 0, -1, -1, NULL, process_sysctl_arg); > + > + kfree(command_line); > +#endif > +} > + > static int __ref kernel_init(void *unused) > { > int ret; > @@ -1367,6 +1386,8 @@ static int __ref kernel_init(void *unused) > > rcu_end_inkernel_boot(); > > + do_sysctl_args(); > + > if (ramdisk_execute_command) { > ret = run_init_process(ramdisk_execute_command); > if (!ret) > diff --git a/kernel/sysctl.c b/kernel/sysctl.c > index ad5b88a53c5a..18c7f5606d55 100644 > --- a/kernel/sysctl.c > +++ b/kernel/sysctl.c > @@ -1980,6 +1980,68 @@ int __init sysctl_init(void) > return 0; > } > > +/* Set sysctl value passed on kernel command line. */ > +int process_sysctl_arg(char *param, char *val, > + const char *unused, void *arg) > +{ > + size_t count; > + char *remaining; > + int err; > + loff_t ppos = 0; > + struct ctl_table *ctl, *found = NULL; > + > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) > + return 0; > + > + param += sizeof("sysctl.") - 1; > + > + remaining = param; > + ctl = &sysctl_base_table[0]; > + > + while(ctl->procname != 0) { > + int len = strlen(ctl->procname); > + if (strncmp(remaining, ctl->procname, len)) { > + ctl++; > + continue; > + } I think you need to validate that "len" is within "remaining" here first. > + if (ctl->child) { > + if (remaining[len] == '.') { > + remaining += len + 1; And that "len + 1" is still valid. > + ctl = ctl->child; > + continue; > + } > + } else { > + if (remaining[len] == '\0') { > + found = ctl; > + break; > + } > + } > + ctl++; > + } > + > + if (!found) { > + pr_warn("Unknown sysctl param '%s' on command line", param); > + return 0; > + } > + > + if (!(found->mode & 0200)) { > + pr_warn("Cannot set sysctl '%s=%s' from command line - not writable", > + param, val); > + return 0; > + } Oh yes; good call about this writable mode test. > + > + count = strlen(val); > + err = found->proc_handler(found, 1, val, &count, &ppos); > + > + if (err) > + pr_warn("Error %d setting sysctl '%s=%s' from command line", > + err, param, val); > + > + pr_debug("Set sysctl '%s=%s' from command line", param, val); > + > + return 0; > +} > + > #endif /* CONFIG_SYSCTL */ > > /* > -- > 2.25.1 > Outside of the nits and missing bounds check, I like it! :)
Vlastimil Babka <vbabka@suse.cz> writes: > A recently proposed patch to add vm_swappiness command line parameter in > addition to existing sysctl [1] made me wonder why we don't have a general > support for passing sysctl parameters via command line. Googling found only > somebody else wondering the same [2], but I haven't found any prior discussion > with reasons why not to do this. > > Settings the vm_swappiness issue aside (the underlying issue might be solved in > a different way), quick search of kernel-parameters.txt shows there are already > some that exist as both sysctl and kernel parameter - hung_task_panic, > nmi_watchdog, numa_zonelist_order, traceoff_on_warning. A general mechanism > would remove the need to add more of those one-offs and might be handy in > situations where configuration by e.g. /etc/sysctl.d/ is impractical. > Also after 61a47c1ad3a4 ("sysctl: Remove the sysctl system call") the only way > to set sysctl is via procfs, so this would eventually allow small systems to be > built without CONFIG_PROC_SYSCTL and still be able to change sysctl parameters. > > Hence, this patch adds a new parse_args() pass that looks for parameters > prefixed by 'sysctl.' and searches for them in the sysctl ctl_tables. When > found, the respective proc handler is invoked. The search is just a naive > linear one, to avoid using the whole procfs layer. It should be acceptable, > as the cost depends on number of sysctl. parameters passed. > > The main limitation of avoiding the procfs layer is however that sysctls > dynamically registered by register_sysctl_table() or register_sysctl_paths() > cannot currently be set by this method. > > The processing is hooked right before the init process is loaded, as some > handlers might be more complicated than simple setters and might need some > subsystems to be initialized. At the moment the init process can be started and > eventually execute a process writing to /proc/sys/ then it should be also fine > to do that from the kernel. > > [1] https://lore.kernel.org/linux-doc/BL0PR02MB560167492CA4094C91589930E9FC0@BL0PR02MB5601.namprd02.prod.outlook.com/ > [2] https://unix.stackexchange.com/questions/558802/how-to-set-sysctl-using-kernel-command-line-parameter > > Signed-off-by: Vlastimil Babka <vbabka@suse.cz> > --- > v2: - handle any nesting level of parameter name > - add Documentation/admin-guide/kernel-parameters.txt blurb > - alias support for legacy one-off parameters, with first conversion (patch 2) > - still no support for dynamically registed sysctls > > .../admin-guide/kernel-parameters.txt | 9 +++ > include/linux/sysctl.h | 1 + > init/main.c | 21 +++++++ > kernel/sysctl.c | 62 +++++++++++++++++++ > 4 files changed, 93 insertions(+) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index c07815d230bc..5076e288f93f 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -4793,6 +4793,15 @@ > > switches= [HW,M68k] > > + sysctl.*= [KNL] > + Set a sysctl parameter right before loading the init > + process, as if the value was written to the respective > + /proc/sys/... file. Currently a subset of sysctl > + parameters is supported that is not registered > + dynamically. Unrecognized parameters and invalid values > + are reported in the kernel log. > + Example: sysctl.vm.swappiness=40 > + > sysfs.deprecated=0|1 [KNL] > Enable/disable old style sysfs layout for old udev > on older distributions. When this option is enabled > diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h > index 02fa84493f23..62ae963a5c0c 100644 > --- a/include/linux/sysctl.h > +++ b/include/linux/sysctl.h > @@ -206,6 +206,7 @@ struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path, > void unregister_sysctl_table(struct ctl_table_header * table); > > extern int sysctl_init(void); > +int process_sysctl_arg(char *param, char *val, const char *unused, void *arg); > > extern struct ctl_table sysctl_mount_point[]; > > diff --git a/init/main.c b/init/main.c > index ee4947af823f..74a094c6b8b9 100644 > --- a/init/main.c > +++ b/init/main.c > @@ -1345,6 +1345,25 @@ void __weak free_initmem(void) > free_initmem_default(POISON_FREE_INITMEM); > } > > +static void do_sysctl_args(void) > +{ > +#ifdef CONFIG_SYSCTL > + size_t len = strlen(saved_command_line) + 1; > + char *command_line; > + > + command_line = kzalloc(len, GFP_KERNEL); > + if (!command_line) > + panic("%s: Failed to allocate %zu bytes\n", __func__, len); > + > + strcpy(command_line, saved_command_line); > + > + parse_args("Setting sysctl args", command_line, > + NULL, 0, -1, -1, NULL, process_sysctl_arg); > + > + kfree(command_line); > +#endif > +} > + > static int __ref kernel_init(void *unused) > { > int ret; > @@ -1367,6 +1386,8 @@ static int __ref kernel_init(void *unused) > > rcu_end_inkernel_boot(); > > + do_sysctl_args(); > + > if (ramdisk_execute_command) { > ret = run_init_process(ramdisk_execute_command); > if (!ret) > diff --git a/kernel/sysctl.c b/kernel/sysctl.c > index ad5b88a53c5a..18c7f5606d55 100644 > --- a/kernel/sysctl.c > +++ b/kernel/sysctl.c > @@ -1980,6 +1980,68 @@ int __init sysctl_init(void) > return 0; > } > > +/* Set sysctl value passed on kernel command line. */ > +int process_sysctl_arg(char *param, char *val, > + const char *unused, void *arg) > +{ > + size_t count; > + char *remaining; > + int err; > + loff_t ppos = 0; > + struct ctl_table *ctl, *found = NULL; > + > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) > + return 0; Is there any way we can use a slash separated path. I know in practice there are not any sysctl names that don't have a '.' in them but why should we artifically limit ourselves? I guess as long as we don't mind not being able to set sysctls that have a '.' in them it doesn't matter. > + > + param += sizeof("sysctl.") - 1; > + > + remaining = param; > + ctl = &sysctl_base_table[0]; > + > + while(ctl->procname != 0) { ^^^^^^^^^^^^^^^^^^ Please either test "while(ctl->procname)" or "while(ctl->procname != NULL)" testing against 0 makes it look like procname is an integer. The style in the kernel is to test against NULL, to make it clear when something is a pointer. > + int len = strlen(ctl->procname); You should have done "strchr(remaining)" and figured out if there is another '.' and only compared up to that dot. Probably skipping this entry entirely if the two lengths don't match. > + if (strncmp(remaining, ctl->procname, len)) { > + ctl++; > + continue; > + } > + if (ctl->child) { > + if (remaining[len] == '.') { > + remaining += len + 1; > + ctl = ctl->child; > + continue; > + } > + } else { > + if (remaining[len] == '\0') { > + found = ctl; > + break; > + } > + } > + ctl++; There should be exactly one match for a name a table. If you get here the code should break, not continue on. > + } > + > + if (!found) { > + pr_warn("Unknown sysctl param '%s' on command line", param); > + return 0; > + } > + > + if (!(found->mode & 0200)) { > + pr_warn("Cannot set sysctl '%s=%s' from command line - not writable", > + param, val); > + return 0; > + } > + > + count = strlen(val); > + err = found->proc_handler(found, 1, val, &count, &ppos); > + > + if (err) > + pr_warn("Error %d setting sysctl '%s=%s' from command line", > + err, param, val); > + > + pr_debug("Set sysctl '%s=%s' from command line", param, val); > + > + return 0; > +} You really should be able to have this code live in fs/proc/proc_sysctl.c and utilize lookup_entry. That should give you the ability to lookup any sysctl. If kernel/sysctl.c is compiled into the kernel proc_sysctl.c is compiled into the kernel. Systems that don't select CONFIG_PROC_SYSCTL won't have any sysctl tables installed at all so they do not make sense to consider or design for. Further it will be faster to lookup the sysctls using the code from proc_sysctl.c as it constructs an rbtree of all of the entries in a directory. The code might as well take advantage of that for large directories. Arguably the main sysctl tables in kernel/sysctl.c should be split up so that things are more localized and there is less global state exported throughout the kernel. I certainly don't want to discourage anyone from doing that just so their sysctl can be used on the command line. Hmm. There is a big gotcha in here and I think it should be mentioned. This code only works because no one has done set_fs(KERNEL_DS). Which means this only works with strings that are kernel addresses essentially by mistake. A big fat comment documenting why it is safe to pass in kernel addresses to a function that takes a "char __user*" pointer would be very good. Eric
On Wed, Mar 25, 2020 at 05:20:40PM -0500, Eric W. Biederman wrote: > Hmm. There is a big gotcha in here and I think it should be mentioned. > This code only works because no one has done set_fs(KERNEL_DS). Which > means this only works with strings that are kernel addresses essentially > by mistake. A big fat comment documenting why it is safe to pass in > kernel addresses to a function that takes a "char __user*" pointer > would be very good. Yeah, I was going to mention this too just now as I went looking through one of the handlers and was reminded that the args are marked __user. :P I suspect we might need to add some __force __user markings or something (with a comment as you say above), to keep sparse from going crazy. :)
On Wed 25-03-20 17:20:40, Eric W. Biederman wrote: > Vlastimil Babka <vbabka@suse.cz> writes: [...] > > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) > > + return 0; > > Is there any way we can use a slash separated path. I know > in practice there are not any sysctl names that don't have > a '.' in them but why should we artifically limit ourselves? Because this is the normal userspace interface? Why should it be any different from calling sysctl? [...] > Further it will be faster to lookup the sysctls using the code from > proc_sysctl.c as it constructs an rbtree of all of the entries in > a directory. The code might as well take advantage of that for large > directories. Sounds like a good fit for a follow up patch to me. Let's make this as simple as possible for the initial version. But up to Vlastimil of course. [...] > Hmm. There is a big gotcha in here and I think it should be mentioned. > This code only works because no one has done set_fs(KERNEL_DS). Which > means this only works with strings that are kernel addresses essentially > by mistake. A big fat comment documenting why it is safe to pass in > kernel addresses to a function that takes a "char __user*" pointer > would be very good. Agreed
On Thu, Mar 26, 2020 at 07:58:29AM +0100, Michal Hocko wrote: > On Wed 25-03-20 17:20:40, Eric W. Biederman wrote: > > Vlastimil Babka <vbabka@suse.cz> writes: > [...] > > > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) > > > + return 0; > > > > Is there any way we can use a slash separated path. I know > > in practice there are not any sysctl names that don't have > > a '.' in them but why should we artifically limit ourselves? > > Because this is the normal userspace interface? Why should it be any > different from calling sysctl? Right. The common method from userspace is dot-separated (which I agree is weird, but it's been like this for ages: see manpages sysctl(8) and sysctl.conf(5) for the details and examples). While "/" is accepted by sysctl, the files shipped in /etc/sysctl.d/ are all using "." separators.
On 3/25/20 10:21 PM, Kees Cook wrote: >> --- a/init/main.c >> +++ b/init/main.c >> @@ -1345,6 +1345,25 @@ void __weak free_initmem(void) >> free_initmem_default(POISON_FREE_INITMEM); >> } >> >> +static void do_sysctl_args(void) >> +{ >> +#ifdef CONFIG_SYSCTL >> + size_t len = strlen(saved_command_line) + 1; >> + char *command_line; >> + >> + command_line = kzalloc(len, GFP_KERNEL); >> + if (!command_line) >> + panic("%s: Failed to allocate %zu bytes\n", __func__, len); >> + >> + strcpy(command_line, saved_command_line); > > No need to open-code this: > > char *command_line; > > command_line = kstrdup(saved_command_line, GFP_KERNEL); > if (!command_line) > panic("%s: Failed to allocate %zu bytes\n", __func__, len); > Ah, right. I admit I basically copy_pasted some other parse_args user. >> diff --git a/kernel/sysctl.c b/kernel/sysctl.c >> index ad5b88a53c5a..18c7f5606d55 100644 >> --- a/kernel/sysctl.c >> +++ b/kernel/sysctl.c >> @@ -1980,6 +1980,68 @@ int __init sysctl_init(void) >> return 0; >> } >> >> +/* Set sysctl value passed on kernel command line. */ >> +int process_sysctl_arg(char *param, char *val, >> + const char *unused, void *arg) >> +{ >> + size_t count; >> + char *remaining; >> + int err; >> + loff_t ppos = 0; >> + struct ctl_table *ctl, *found = NULL; >> + >> + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) >> + return 0; >> + >> + param += sizeof("sysctl.") - 1; >> + >> + remaining = param; >> + ctl = &sysctl_base_table[0]; >> + >> + while(ctl->procname != 0) { >> + int len = strlen(ctl->procname); >> + if (strncmp(remaining, ctl->procname, len)) { >> + ctl++; >> + continue; >> + } > > I think you need to validate that "len" is within "remaining" here > first. My reasoning was that if remaining terminates too early, the null byte would be different from non-null byte in ctl->procname and thus strncmp will return it as different? And the reason I used len in strncmp there is only so it doesn't compare the terminating null, because remaning can continue with ".foo" instead. >> + if (ctl->child) { >> + if (remaining[len] == '.') { >> + remaining += len + 1; > > And that "len + 1" is still valid. And since we passed strncmp(..., len), remaining[len] might be null byte, but then we can still compare it with '.'. But C strings are full of landmines.
Michal Hocko <mhocko@kernel.org> writes: > On Wed 25-03-20 17:20:40, Eric W. Biederman wrote: >> Vlastimil Babka <vbabka@suse.cz> writes: > [...] >> > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) >> > + return 0; >> >> Is there any way we can use a slash separated path. I know >> in practice there are not any sysctl names that don't have >> a '.' in them but why should we artifically limit ourselves? > > Because this is the normal userspace interface? Why should it be any > different from calling sysctl? > [...] Why should the kernel command line implement userspace whims? I was thinking something like: "sysctl/kernel/max_lock_depth=2048" doesn't look too bad and it makes things like reusing our kernel internal helpers much easier. Plus it suggest that we could do the same for sysfs files: "sysfs/kernel/fscaps=1" And the code could be same for both cases except for the filesystem prefix. >> Further it will be faster to lookup the sysctls using the code from >> proc_sysctl.c as it constructs an rbtree of all of the entries in >> a directory. The code might as well take advantage of that for large >> directories. > > Sounds like a good fit for a follow up patch to me. Let's make this > as simple as possible for the initial version. But up to Vlastimil of course. I would argue that reusing proc_sysctl.c:lookup_entry() should make the code simpler, and easier to reason about. Especially given the bugs in the first version with a sysctl path. A clean separation between separating the path from into pieces and looking up those pieces should make the code more robust. That plus I want to get very far away from the incorrect idea that you can have sysctls without compiling in proc support. That is not how the code works, that is not how the code is tested. It is also worth pointing out that: proc_mnt = kern_mount(proc_fs_type); for_each_sysctl_cmdline() { ... file = file_open_root(proc_mnt->mnt_root, proc_mnt, sysctl_path, O_WRONLY, 0); kernel_write(file, value, value_len); } kern_umount(proc_mnt); Is not an unreasonable implementation. There are problems with a persistent mount of proc in that it forces userspace not to use any proc mount options. But a temporary mount of proc to deal with command line options is not at all unreasonable. Plus it looks like we can have kern_write do all of the kernel/user buffer silliness. > [...] > >> Hmm. There is a big gotcha in here and I think it should be mentioned. >> This code only works because no one has done set_fs(KERNEL_DS). Which >> means this only works with strings that are kernel addresses essentially >> by mistake. A big fat comment documenting why it is safe to pass in >> kernel addresses to a function that takes a "char __user*" pointer >> would be very good. > > Agreed Eric
On 3/25/20 11:20 PM, Eric W. Biederman wrote: > Vlastimil Babka <vbabka@suse.cz> writes: >> --- a/kernel/sysctl.c >> +++ b/kernel/sysctl.c >> @@ -1980,6 +1980,68 @@ int __init sysctl_init(void) >> return 0; >> } >> >> +/* Set sysctl value passed on kernel command line. */ >> +int process_sysctl_arg(char *param, char *val, >> + const char *unused, void *arg) >> +{ >> + size_t count; >> + char *remaining; >> + int err; >> + loff_t ppos = 0; >> + struct ctl_table *ctl, *found = NULL; >> + >> + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) >> + return 0; > > Is there any way we can use a slash separated path. I know We could, but as others explained, people and tools are used to the dot separation, so I think the only sensible options are supporting only dot, or both dot and slash. > in practice there are not any sysctl names that don't have > a '.' in them but why should we artifically limit ourselves? Existing tools would probably break (or perhaps sysctl(8) is smarter than I think, dunno). > I guess as long as we don't mind not being able to set sysctls > that have a '.' in them it doesn't matter. Right. >> + >> + param += sizeof("sysctl.") - 1; >> + >> + remaining = param; >> + ctl = &sysctl_base_table[0]; >> + >> + while(ctl->procname != 0) { > ^^^^^^^^^^^^^^^^^^ > > Please either test "while(ctl->procname)" or > "while(ctl->procname != NULL)" testing against 0 makes it look like > procname is an integer. The style in the kernel is to test against > NULL, to make it clear when something is a pointer. OK >> + int len = strlen(ctl->procname); > > You should have done "strchr(remaining)" and figured out if there is > another '.' and only compared up to that dot. Probably skipping this > entry entirely if the two lengths don't match. That's also possible, but AFAICS my code works as intended, as I explained in a reply to Kees, and also below: >> + if (strncmp(remaining, ctl->procname, len)) { >> + ctl++; >> + continue; >> + } >> + if (ctl->child) { >> + if (remaining[len] == '.') { >> + remaining += len + 1; >> + ctl = ctl->child; >> + continue; >> + } >> + } else { >> + if (remaining[len] == '\0') { >> + found = ctl; >> + break; >> + } >> + } >> + ctl++; > > There should be exactly one match for a name a table. > If you get here the code should break, not continue on. If there existed e.g. both "vm.swap" and "vm.swappiness" options and user passed "vm.swappiness=10", but the "swap" ctl entry was encountered first, it will succeed the strncmp(), but then realize "swap" was just a prefix of what user specified (remaining[len] is not '\0') and hence continue serching for other matches. >> + } >> + >> + if (!found) { >> + pr_warn("Unknown sysctl param '%s' on command line", param); >> + return 0; >> + } >> + >> + if (!(found->mode & 0200)) { >> + pr_warn("Cannot set sysctl '%s=%s' from command line - not writable", >> + param, val); >> + return 0; >> + } >> + >> + count = strlen(val); >> + err = found->proc_handler(found, 1, val, &count, &ppos); >> + >> + if (err) >> + pr_warn("Error %d setting sysctl '%s=%s' from command line", >> + err, param, val); >> + >> + pr_debug("Set sysctl '%s=%s' from command line", param, val); >> + >> + return 0; >> +} > > You really should be able to have this code live in > fs/proc/proc_sysctl.c and utilize lookup_entry. > > That should give you the ability to lookup any sysctl. If > kernel/sysctl.c is compiled into the kernel proc_sysctl.c is compiled > into the kernel. Systems that don't select CONFIG_PROC_SYSCTL won't > have any sysctl tables installed at all so they do not make sense to > consider or design for. I see. In fact one reason why I tried to avoid the proc stuff was your commit 61a47c1ad3a4 ("sysctl: Remove the sysctl system call") and this part: > As this removes one of the few uses of the internal kernel mount > of proc I hope this allows for even more simplifications of the > proc filesystem. But if you now suggest using the kernel mount then sure, it I don't object make the code simpler and handle all sysctls. > Further it will be faster to lookup the sysctls using the code from > proc_sysctl.c as it constructs an rbtree of all of the entries in > a directory. The code might as well take advantage of that for large > directories. > > Arguably the main sysctl tables in kernel/sysctl.c should be split up so > that things are more localized and there is less global state exported > throughout the kernel. I certainly don't want to discourage anyone from > doing that just so their sysctl can be used on the command line. Fair point. > Hmm. There is a big gotcha in here and I think it should be mentioned. > This code only works because no one has done set_fs(KERNEL_DS). Which > means this only works with strings that are kernel addresses essentially > by mistake. A big fat comment documenting why it is safe to pass in > kernel addresses to a function that takes a "char __user*" pointer > would be very good. Thanks, didn't realize that. > Eric >
On Thu, Mar 26, 2020 at 07:58:29AM +0100, Michal Hocko wrote: > On Wed 25-03-20 17:20:40, Eric W. Biederman wrote: > > Vlastimil Babka <vbabka@suse.cz> writes: > [...] > > > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) > > > + return 0; > > > > Is there any way we can use a slash separated path. I know > > in practice there are not any sysctl names that don't have > > a '.' in them but why should we artifically limit ourselves? > > Because this is the normal userspace interface? Why should it be any > different from calling sysctl? > [...] Imho, we should use ".". Kernel developers aren't the ones setting these options, admins are and if I think back to the times doing that as a job at uni I'd be very confused if I learned that I get to set sysctl options through the kernel command but need to use yet another format than what I usually do to set those from the shell. Consistency is most of the times to be preferred imho. Also, the kernel docs illustrate that the "." syntax is used for other keys as well (e.g. acpi.<option>) and userspace options passed via the kernel command line have standardized on the "." format as well, e.g. systemd appends in the same format (e.g. systemd.unified_cgroup_hierarchy, systemd.unit what have you). Christian
On Thu 26-03-20 14:30:41, Christian Brauner wrote: > On Thu, Mar 26, 2020 at 07:58:29AM +0100, Michal Hocko wrote: > > On Wed 25-03-20 17:20:40, Eric W. Biederman wrote: > > > Vlastimil Babka <vbabka@suse.cz> writes: > > [...] > > > > + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) > > > > + return 0; > > > > > > Is there any way we can use a slash separated path. I know > > > in practice there are not any sysctl names that don't have > > > a '.' in them but why should we artifically limit ourselves? > > > > Because this is the normal userspace interface? Why should it be any > > different from calling sysctl? > > [...] > > Imho, we should use ".". Kernel developers aren't the ones setting > these options, admins are and if I think back to the times doing that as > a job at uni I'd be very confused if I learned that I get to set sysctl > options through the kernel command but need to use yet another format > than what I usually do to set those from the shell. Consistency is most > of the times to be preferred imho. Absolutely agreed! Even if sysctl can consume / instead of ., which was a news to me btw, the majority of the usage is with `.'
On Thu, Mar 26, 2020 at 07:45:13AM -0500, Eric W. Biederman wrote: > > On Wed 25-03-20 17:20:40, Eric W. Biederman wrote: > plus I want to get very far away from the incorrect idea that you > can have sysctls without compiling in proc support. That is not how > the code works, that is not how the code is tested. Agreed. > It is also worth pointing out that: > > proc_mnt = kern_mount(proc_fs_type); > for_each_sysctl_cmdline() { > ... > file = file_open_root(proc_mnt->mnt_root, proc_mnt, sysctl_path, O_WRONLY, 0); > kernel_write(file, value, value_len); > } > kern_umount(proc_mnt); > > Is not an unreasonable implementation. This: > There are problems with a persistent mount of proc in that it forces > userspace not to use any proc mount options. But a temporary mount of > proc to deal with command line options is not at all unreasonable. > Plus it looks like we can have kern_write do all of the kernel/user > buffer silliness. Is a bit of tribal knowledge worth documenting for the approach taken forward. Vlastimil can you add a little comment mentioning some of this logic? Luis
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt index c07815d230bc..5076e288f93f 100644 --- a/Documentation/admin-guide/kernel-parameters.txt +++ b/Documentation/admin-guide/kernel-parameters.txt @@ -4793,6 +4793,15 @@ switches= [HW,M68k] + sysctl.*= [KNL] + Set a sysctl parameter right before loading the init + process, as if the value was written to the respective + /proc/sys/... file. Currently a subset of sysctl + parameters is supported that is not registered + dynamically. Unrecognized parameters and invalid values + are reported in the kernel log. + Example: sysctl.vm.swappiness=40 + sysfs.deprecated=0|1 [KNL] Enable/disable old style sysfs layout for old udev on older distributions. When this option is enabled diff --git a/include/linux/sysctl.h b/include/linux/sysctl.h index 02fa84493f23..62ae963a5c0c 100644 --- a/include/linux/sysctl.h +++ b/include/linux/sysctl.h @@ -206,6 +206,7 @@ struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path, void unregister_sysctl_table(struct ctl_table_header * table); extern int sysctl_init(void); +int process_sysctl_arg(char *param, char *val, const char *unused, void *arg); extern struct ctl_table sysctl_mount_point[]; diff --git a/init/main.c b/init/main.c index ee4947af823f..74a094c6b8b9 100644 --- a/init/main.c +++ b/init/main.c @@ -1345,6 +1345,25 @@ void __weak free_initmem(void) free_initmem_default(POISON_FREE_INITMEM); } +static void do_sysctl_args(void) +{ +#ifdef CONFIG_SYSCTL + size_t len = strlen(saved_command_line) + 1; + char *command_line; + + command_line = kzalloc(len, GFP_KERNEL); + if (!command_line) + panic("%s: Failed to allocate %zu bytes\n", __func__, len); + + strcpy(command_line, saved_command_line); + + parse_args("Setting sysctl args", command_line, + NULL, 0, -1, -1, NULL, process_sysctl_arg); + + kfree(command_line); +#endif +} + static int __ref kernel_init(void *unused) { int ret; @@ -1367,6 +1386,8 @@ static int __ref kernel_init(void *unused) rcu_end_inkernel_boot(); + do_sysctl_args(); + if (ramdisk_execute_command) { ret = run_init_process(ramdisk_execute_command); if (!ret) diff --git a/kernel/sysctl.c b/kernel/sysctl.c index ad5b88a53c5a..18c7f5606d55 100644 --- a/kernel/sysctl.c +++ b/kernel/sysctl.c @@ -1980,6 +1980,68 @@ int __init sysctl_init(void) return 0; } +/* Set sysctl value passed on kernel command line. */ +int process_sysctl_arg(char *param, char *val, + const char *unused, void *arg) +{ + size_t count; + char *remaining; + int err; + loff_t ppos = 0; + struct ctl_table *ctl, *found = NULL; + + if (strncmp(param, "sysctl.", sizeof("sysctl.") - 1)) + return 0; + + param += sizeof("sysctl.") - 1; + + remaining = param; + ctl = &sysctl_base_table[0]; + + while(ctl->procname != 0) { + int len = strlen(ctl->procname); + if (strncmp(remaining, ctl->procname, len)) { + ctl++; + continue; + } + if (ctl->child) { + if (remaining[len] == '.') { + remaining += len + 1; + ctl = ctl->child; + continue; + } + } else { + if (remaining[len] == '\0') { + found = ctl; + break; + } + } + ctl++; + } + + if (!found) { + pr_warn("Unknown sysctl param '%s' on command line", param); + return 0; + } + + if (!(found->mode & 0200)) { + pr_warn("Cannot set sysctl '%s=%s' from command line - not writable", + param, val); + return 0; + } + + count = strlen(val); + err = found->proc_handler(found, 1, val, &count, &ppos); + + if (err) + pr_warn("Error %d setting sysctl '%s=%s' from command line", + err, param, val); + + pr_debug("Set sysctl '%s=%s' from command line", param, val); + + return 0; +} + #endif /* CONFIG_SYSCTL */ /*
A recently proposed patch to add vm_swappiness command line parameter in addition to existing sysctl [1] made me wonder why we don't have a general support for passing sysctl parameters via command line. Googling found only somebody else wondering the same [2], but I haven't found any prior discussion with reasons why not to do this. Settings the vm_swappiness issue aside (the underlying issue might be solved in a different way), quick search of kernel-parameters.txt shows there are already some that exist as both sysctl and kernel parameter - hung_task_panic, nmi_watchdog, numa_zonelist_order, traceoff_on_warning. A general mechanism would remove the need to add more of those one-offs and might be handy in situations where configuration by e.g. /etc/sysctl.d/ is impractical. Also after 61a47c1ad3a4 ("sysctl: Remove the sysctl system call") the only way to set sysctl is via procfs, so this would eventually allow small systems to be built without CONFIG_PROC_SYSCTL and still be able to change sysctl parameters. Hence, this patch adds a new parse_args() pass that looks for parameters prefixed by 'sysctl.' and searches for them in the sysctl ctl_tables. When found, the respective proc handler is invoked. The search is just a naive linear one, to avoid using the whole procfs layer. It should be acceptable, as the cost depends on number of sysctl. parameters passed. The main limitation of avoiding the procfs layer is however that sysctls dynamically registered by register_sysctl_table() or register_sysctl_paths() cannot currently be set by this method. The processing is hooked right before the init process is loaded, as some handlers might be more complicated than simple setters and might need some subsystems to be initialized. At the moment the init process can be started and eventually execute a process writing to /proc/sys/ then it should be also fine to do that from the kernel. [1] https://lore.kernel.org/linux-doc/BL0PR02MB560167492CA4094C91589930E9FC0@BL0PR02MB5601.namprd02.prod.outlook.com/ [2] https://unix.stackexchange.com/questions/558802/how-to-set-sysctl-using-kernel-command-line-parameter Signed-off-by: Vlastimil Babka <vbabka@suse.cz> --- v2: - handle any nesting level of parameter name - add Documentation/admin-guide/kernel-parameters.txt blurb - alias support for legacy one-off parameters, with first conversion (patch 2) - still no support for dynamically registed sysctls .../admin-guide/kernel-parameters.txt | 9 +++ include/linux/sysctl.h | 1 + init/main.c | 21 +++++++ kernel/sysctl.c | 62 +++++++++++++++++++ 4 files changed, 93 insertions(+)