--- trunk/1.7.x/ccs-patch/README.ccs 2009/12/15 12:13:01 3256 +++ trunk/1.8.x/ccs-patch/README.ccs 2011/03/15 05:23:03 4724 @@ -1723,7 +1723,7 @@ to see process information. The content of /proc/ccs/.execute_handler is identical to /proc/ccs/.process_status . -Version 1.6.5 2008/11/11 Third anniversary release. +Version 1.6.5 2008/11/11 Third anniversary release. Fix 2008/12/01 @@ -2088,7 +2088,7 @@ @ Transit to new domain before do_execve() succeeds. Permission checks for interpreters and environment variables are - done using new domain. In order to be allow ccs-queryd to reach the new + done using new domain. In order to allow ccs-queryd to reach the new domain via global PID, I reverted "Don't transit to new domain until do_execve() succeeds." made on 2008/10/07. @@ -2243,7 +2243,7 @@ /proc/ccs/profile (which results in not protecting the system at all), I added a check for PROFILE_VERSION= . -Version 1.7.1 2009/11/11 Fourth anniversary release. +Version 1.7.1 2009/11/11 Fourth anniversary release. Fix 2009/11/13 @@ -2281,9 +2281,523 @@ Fix 2009/12/15 + @ Allow deleting "quota_exceeded" and "transition_failed" entries. + + To notify users of "this domain has too many entries to hold" and "some + process in this domain was not able to perform domain transition", + "quota_exceeded" and "transition_failed" messages are used respectively. + These messages were not deletable. But it is more convenient for users + to be notified again if such events occurred again after tuning policy. + Thus, I made these messages deletable. + +Fix 2009/12/17 + + @ Don't check read permission in ccs_try_alt_exec(). + + While I was trying to remove ccs_execve_list list for GC optimization + between TOMOYO 1.7.0 and 1.7.1 , I made a mistake which made TOMOYO to + check allow_read permission of the programs specified by execute_handler + and denied_execute_handler keywords. + @ Don't check DAC permission if disabled mode. I was checking DAC permissions regarding directory entry modification operations (e.g. mkdir()) even if mode=disabled . It is a waste of CPU resource to check DAC permissions when MAC permissions are not checked. Thus, I modified to skip DAC permission checks if mode=disabled . + +Fix 2009/12/19 + + @ Fix memory leak in ccs_environ(). + + When I fixed a bug that a permission like + + allow_env PATH if exec.envp["PATH"]="/" + + was not working (2009/11/02), I allocated two buffers but only one buffer + was released. + + This bug will trigger OOM killer if environment variable checking is + enabled. + +Fix 2010/01/17 + + @ Use current domain's name for execute_handler audit log. + + Since 1.6.7 , /proc/ccs/grant_log was by error using next domain's name + when auditing current domain's "execute_handler" line. + +Fix 2010/03/02 + + @ Allow domain transition without execve(). + + To be able to split permissions for Apache's CGI programs which are + executed without execve(), I added special domain transition which is + performed by atomically writing '\0'-terminated binary string to + /proc/ccs/.transition interface. For example, a process which belongs to + " /usr/sbin/httpd" domain will transit to + " /usr/sbin/httpd //app=cgi1\040id=10000" domain by atomically + writing "app=cgi1 id=10000" + '\0' to /proc/ccs/.transition using + Apache's ap_hook_handler() functionality. + + Note that '\0'-terminated binary string is converted to TOMOYO's string + inside kernel and prefix "//" is automatically added to the string so + that domainname does not conflict with domainnames created by execve(). + Without this prefix, if " /usr/sbin/sshd /bin/bash" domain is + allowed to open /proc/ccs/.transition for writing and + " /usr/sbin/sshd /bin/bash /usr/bin/passwd" domain is allowed to + access /etc/shadow , /bin/bash will be able to access /etc/shadow by + atomically writing "/usr/bin/passwd" + '\0' to /proc/ccs/.transition . + Allowing /bin/bash to access /etc/shadow is not what people want. + + Permission for this operation is checked by "allow_transit" keyword. + Unlike "allow_execute" keyword, the string parameter for "allow_transit" + keyword does not refer a real file on filesystem's namespace. Therefore, + you can store any combination of parameters like LDAP's DN entry in the + string parameter for "allow_transit" keyword. + +Fix 2010/03/08 + + @ Allow building as loadable kernel module. + + To be able to minimize filesize increment of vmlinux, I made it + possible to compile TOMOYO Linux as loadable kernel module. + Although patching the kernel source and recompiling the kernel are + inevitable, this change will make it easier to enable TOMOYO Linux + when there is a filesize limitation on vmlinux (e.g. embedded systems). + +Fix 2010/03/25 + + @ Fix ccs_get_ipv6_address() bug. + + Since 1.7.0 , ccs_get_ipv6_address() was by error returning address of + "struct list_head ccs_address_list" if memory allocation failed. + As a result, ccs_put_ipv6_address() will modify memory near + "struct list_head ccs_address_list" if memory allocation failed. + +Fix 2010/03/26 + + @ Fix ccs_lport_reserved() bug. + + Since 1.7.0 , ccs_lport_reserved() was by error checking wrong port + number. As a result, "deny_autobind" keyword was not working as expected. + +Version 1.7.2 2010/04/01 Feature enhancement release. + +Fix 2010/04/10 + + @ Fix invalid "struct nameidata" to "struct path" conversion macro. + + Regarding kernels 2.6.24 and earlier, I was converting "struct nameidata" + to "struct path" in caller side so that I can unify the callee function's + parameter type. But it turned out that the macro I used did not follow C + standards and did not work with gcc 4.x . As a result, "allow_pivot_root" + keyword was not working as expected. + +Fix 2010/05/05 + + @ Fix incorrect audit on/off control. + + The grant_log= and reject_log= parameters of CONFIG::misc::env were not + used because I forgot to update request type. As a result, those of + CONFIG::file::execute were used for CONFIG::misc::env . + + Those of CONFIG::file::rewrite were not used because I forgot to update + request type. As a result, those of CONFIG::file::truncate were used for + CONFIG::file::rewrite . + +Fix 2010/05/10 + + @ Fix incorrect out of memory warning. + + Out of memory warnings were not printed in some cases by error. + +Fix 2010/05/27 + + @ Add missing rcu_dereference() for ccs_find_execute_handler(). + + Since 1.7.0 , ccs_find_execute_handler() was by error using + list_for_each_entry() rather than list_for_each_entry_rcu(). + This bug affects only Alpha architecture. + +Fix 2010/06/03 + + @ Fix missing sanity check for "file_pattern". + + Since 1.7.0 , ccs_write_pattern_policy() was by error accepting + invalid pathname. + +Fix 2010/06/09 + + @ Add missing ccs_put_name() in ccs_parse_envp(). + + Since 1.7.0 , ccs_parse_envp() was not calling ccs_put_name() if + environment variable's value ('if exec.envp["name"]="value"' condition) + was invalid. + + @ Add missing NULL check in ccs_condition(). + + Since 1.7.0 , if 'if symlink.target=' part was given against non-file + permissions (e.g. allow_env PATH if symlink.target="/"), it triggered + NULL pointer dereference. + +Fix 2010/10/28 + + @ Fix umount() pathname calculation. + + "mount --bind /path/to/file1 /path/to/file2" is legal. + Therefore, "umount /path/to/file2" is also legal. + Do not automatically append trailing '/' if pathname to be unmounted + does not end with '/'. + + @ Add preserve KABI compatibility option. (2.6 kernels only) + + TOMOYO needs "struct ccs_domain_info *" and "u32" for each + "struct task_struct". But embedding these variables into + "struct task_struct" breaks KABI for prebuilt kernel modules (which + means that you will need to rebuild prebuilt kernel modules). + + Since KABI is commonly used (compared to 5 years ago), asking users to + rebuild kernel modules which are not included in kernel package is no + longer preferable. Therefore, I added a new option that keeps + "struct task_struct" unmodified in order to keep KABI. + + Note that you have to use ccs-patch-2.6.\*.diff which patches + kernel/fork.c in order to use this option. Otherwise, TOMOYO will leak + memory whenever "struct task_struct" is released. + + @ Change directives. + + I removed "allow_" prefix from directives. New directives for files are + prefixed with "file ". For example, "allow_read" changed to "file read", + "allow_ioctl" changed to "file ioctl". New directive for "allow_network + TCP" is "network inet stream", "allow_network UDP" is "network inet + dgram", "allow_network RAW" is "network inet raw". New directive for + "allow_env" is "misc env". New directive for "allow_signal" is "ipc + signal". New directive for "allow_capability" is "capability". These new + directives correspond with keywords used by profile's CONFIG lines. + + I removed "deny_rewrite" and "allow_rewrite" directives and introduced + "file append" directive. Thus, permission for open(O_WRONLY | O_APPEND) + changed from "allow_write" + "allow_rewrite" to "file append". + + I removed "SYS_MOUNT", "SYS_UMOUNT", "SYS_CHROOT", "SYS_KILL", + "SYS_LINK", "SYS_SYMLINK", "SYS_RENAME", "SYS_UNLINK", "SYS_CHMOD", + "SYS_CHOWN", "SYS_IOCTL", "SYS_PIVOT_ROOT" keywords from capabilities + because these permissions can be checked by other directives (e.g. + "file mount", "ipc signal"). + + I also removed "conceal_mount" keyword from capabilities because this + check requires hooks in filesystem part while almost all hooks for + filesystem part have moved to LSM by Linux 2.6.34. + + New directive for "execute_handler" is "task auto_execute_handler", + "denied_execute_handler" is "task denied_execute_handler". + + @ Distinguish send() and recv() operations. + + Until now, it was impossible for UDP and IP sockets to allow either + only sending or only receiving because permissions were aggregated with + "connect" keyword. I broke "connect" keyword into "send" and "recv" + keywords so that you can keep access control for send() operation enabled + when you have to disable access control for recv() operation due to + application breakage by discarding incoming datagram. + + @ Add Unix domain socket restriction support. + + Until now, it was possible to restrict only inet domain sockets (i.e. + TCP/UDP/RAW). I added restriction for Unix domain sockets (i.e. stream/ + dgram/seqpacket). New directive "network unix" is added as well as + "network inet" directive. + + @ Allow specifying multiple permissions in a line. + + Until now, only "allow_read/write" can be specified for combination of + "allow_read" + "allow_write". Now, you can combine other permissions as + long as type of parameters for these permissions is same. For example, + "file read/write/append/execute/unlink/truncate /tmp/file" is correct + but "file read/write/create /tmp/file" is wrong because "file create" + requires create mode whereas "file read" and "file write" do not. + + @ Allow wildcard for execute permission and domainname. + + Until now, to execute programs with temporary names, "aggregator" is + needed. To simplify code, I modified to accept wildcards for execute + permission and domainname. Now, you can directly specify + "file execute /tmp/logrotate.\?\?\?\?\?\?" and use + "/tmp/logrotate.\?\?\?\?\?\?" within domainnames. + + @ Change pathname for non-rename()able filesystems. + + LSM version of TOMOYO wants to use /proc/self/ rather than /proc/$PID/ if + $PID matches current thread's process ID in order to prevent current + thread from accessing other process's information unless needed. + But since procfs can be mounted on various locations (e.g. /proc/ /proc2/ + /p/ /tmp/foo/100/p/ ), LSM version of TOMOYO cannot tell that whether the + numeric part in the string returned by __d_path() represents process ID + or not. + + Therefore, to be able to convert from $PID to self no matter where procfs + is mounted, I changed pathname representations for filesystems which do + not support rename() operation (e.g. proc, sysfs, securityfs). + + Now, "/proc/self/mounts" changed to "proc:/self/mounts" and + "/sys/kernel/security/" changed to "sys:/kernel/security/" and + "/dev/pts/0" changed to "devpts:/0". + + @ Add a new keyword "any" for domain transition control. + + To be able to make it easier to apply auto_execute_handler on each + domain, I added "any" keyword to domain transition control keywords. Now, + "initialize_domain /usr/sbin/sshd" changed to + "initialize_domain /usr/sbin/sshd from any" and + "keep_domain /usr/sbin/sshd /bin/bash" changed to + "keep_domain any from /usr/sbin/sshd /bin/bash". + + "keep_domain /path/to/auto_execute_handler from any" will allow you to + apply auto_execute_handler for any domains without creating domains for + auto_execute_handler. + + @ Change buffering mode for reading policy. + + To be able to read() very very long lines correctly, I changed the way + TOMOYO buffers policy for reading. + + @ Introduce "acl_group" keyword. + + Until now, it was possible to specify only "allow_read" and "allow_env" + keywords in the exception policy. + + Since some operations like "file read/write/append /dev/null" and + "network UDP send/recv @DNS_SERVER 53" are very common and should be + permitted to all domains, I introduced "acl_group" keyword for giving + such permissions. + + For example, specify "acl_group 0 file read/write/append /dev/null" in + the exception policy and specify "use_group 0" from the domains in the + domain policy. + + "ignore_global_allow_read" and "ignore_global_allow_env" directives were + removed from domain policy and "use_group" keyword was added. + + @ Remove "if" and "; set" keyword. + + I removed need for specifying these keyword. + You can simply specify like below. + + file read /etc/shadow task.uid=0 + + @ Remove "file_pattern" keyword. + + I removed "file_pattern" keyword because it is impossible to predefine + all possible pathname patterns. Also, learning pathnames using incomplete + patterns makes it difficult to later replace using "path_group" keyword. + + @ Replace verbose= parameter with statistic interface. + + Since it is noisy if a lot of policy violation messages are printed, + I removed printk(). To be able to check whether policy violation occurred + or not, I introduced /proc/ccs/stat interface which counts number of + policy violations occurred. You can firstly check /proc/ccs/stat and then + check /proc/ccs/reject_log . + + @ Remove global preference. + + I removed global preference in order to make code simpler. + + @ Allow controlling generation of access granted logs for per an entry + basis. + + I added per-entry flag which controls generation of grant logs because + Xen and KVM issues ioctl requests so frequently. For example, + + file ioctl /dev/null 0x5401 grant_log=no + + will suppress /proc/ccs/grant_log even if preference says grant_log=yes . + + file ioctl /dev/null 0x5401 grant_log=yes + + will generate /proc/ccs/grant_log even if preference says grant_log=no . + + file ioctl /dev/null 0x5401 + + will generate /proc/ccs/grant_log only if preference says grant_log=yes . + + This flag is intended for frequently accessed resources like + + file read /var/www/html/\{\*\}/\*.html grant_log=no + + . + + @ Automatically create domain by execve() even if enforcing mode. + + Until now, new domains are not created if the domain was not defined and + current domain is enforcing mode ("CONFIG::file::execute=enforcing"). + + To be able to restrict shell session without using "keep_domain", + I changed to create new domains automatically even if current domain is + enforcing mode. + + @ Replace "task.state" with "auto_domain_transition". + + task.state is difficult to use. Thus, I replaced task.state with + auto_domain_transition which performs domain transition instead of + changing current process's state variables. + + If domain transition failed, current process will be killed by SIGKILL + signal. This should not happen in normal circumstances, for you know the + domain to transit to and thereby you will define the domain beforehand + when you use "auto_domain_transition" keyword. + + @ Replace "allow_transit" with "task manual_domain_transition". + + I changed this directive to specify absolute domainname (e.g. + " /usr/sbin/httpd //app=cgi1\040id=10000") rather than virtual + pathname (e.g. "//app=cgi1\040id=10000") because you know the domain to + transit to and thereby you will define the domain beforehand when you use + "task manual_domain_transition" directive. + + This change allows you to jump to arbitrary domain. + + Note that this change also reverts "Change /proc/ccs/info/self_domain ." + made on 2006/10/24. Now, 'cat < /proc/ccs/info/self_domain' will act like + 'cat /proc/ccs/info/self_domain'. Programs depending on old assumption + need to be updated. + + @ Add "task auto_domain_transition". + + This is similar to "task manual_domain_transition", but is automatically + applied whenever conditions are met. For example, + + task auto_domain_transition //./non-root task.uid!=0 + + will automatically jump to " //./non-root" domain if current + process's UID is not 0 whereas + + task manual_domain_transition //./non-root task.uid!=0 + + will jump to " //./non-root" domain if current process's UID is + not 0 and current process wrote " //./non-root" to + /proc/ccs/self_domain interface. + + If domain transition failed, current process will be killed by SIGKILL + signal. + + @ Optimize for object's size. + + I merged similar code in order to reduce object's filesize. + +Version 1.8.0 2010/11/11 Fifth anniversary release. + +Fix 2010/12/01 + + @ Use same interface for audit logs. + + To be able to perform fine grained filtering by /usr/sbin/ccs-auditd , + I merged /proc/ccs/grant_log and /proc/ccs/reject_log as + /proc/ccs/audit and added granted=yes or granted=no to audit logs. + +Fix 2010/12/17 + + @ Split ccs_null_security into ccs_default_security and ccs_oom_security. + + ccs_null_security is used by preserve KABI compatibility option and is + used for providing default values against threads which have not yet + allocated memory for their security contexts. + + If current thread failed to allocate memory for current thread's security + context, current thread uses ccs_null_security. Since current thread is + allowed to modify current thread's security context, current thread might + modify ccs_null_security which should not be modified for any reason. + + Therefore, I split ccs_null_security into ccs_default_security and + ccs_oom_security and use ccs_oom_security when current thread failed to + allocate memory for current thread's security context. + + Threads which do not share ccs_oom_security are not affected by threads + which share ccs_oom_security. Threads which share ccs_oom_security will + experience temporary inconsistency, but such threads are about to be + killed by SIGKILL signal. + +Fix 2011/01/11 + + @ Use filesystem name for unnamed devices when vfsmount is missing. + + "Change pathname for non-rename()able filesystems." changed to use + "$fsname:" if the filesystem does not support rename() operation and + "dev($major,$minor):" otherwise when vfsmount is missing. But it turned + out that it is useless to use "dev($major,$minor):" for unnamed devices + (filesystems with $major == 0). Thus, I changed to use "$fsname:" rather + than "dev($major,$minor):" for filesystems with $major == 0 when vfsmount + is missing. + +Fix 2011/02/07 + + @ Fix infinite loop bug when reading /proc/ccs/audit or /proc/ccs/query . + + In ccs_flush(), head->r.w[0] holds pointer to string data to be printed. + But head->r.w[0] was updated only when the string data was partially + printed (because head->r.w[0] will be updated by head->r.w[1] later if + completely printed). However, regarding /proc/ccs/audit and + /proc/ccs/query , an additional '\0' is printed after the string data was + completely printed. But if free space for read buffer became 0 before + printing the additional '\0', ccs_flush() was returning without updating + head->r.w[0]. As a result, ccs_flush() forever reprints already printed + string data. + +Fix 2011/03/01 + + @ Run garbage collector without waiting for /proc/ccs/ users. + + Currently TOMOYO holds SRCU lock upon open() and releases it upon close() + because list elements stored in the "struct ccs_io_buffer" instances are + accessed until close() is called. However, such SRCU usage causes lockdep + to complain about leaving the kernel with SRCU lock held. Therefore, + I changed to hold/release SRCU upon each read()/write() by selectively + deferring kfree() by keeping track of the "struct ccs_io_buffer" + instances. + +Fix 2011/03/05 + + @ Support built-in policy configuration. + + To be able to start using enforcing mode from the early stage of boot + sequence, I added support for built-in policy configuration and + activating access control without calling external policy loader program. + + This will be useful for systems where operations which can lead to the + hijacking of the boot sequence are needed before loading the policy. + For example, you can activate immediately after loading the fixed part of + policy which will allow only operations needed for mounting a partition + which contains the variant part of policy and verifying (e.g. running GPG + check) and loading the variant part of policy. Since you can start using + enforcing mode from the beginning, you can reduce the possibility of + hijacking the boot sequence. + +Fix 2011/03/10 + + @ Remove /proc/ccs/meminfo interface. + + Please use /proc/ccs/stat interface instead. + +Fix 2011/03/15 + + @ Pack policy when printing via /proc/ccs/ interface. + + The kernel side is ready for accepting packed input like + + file read/write/execute /path/to/file + + but was using unpacked output like + + file read /path/to/file + file write /path/to/file + file execute /path/to/file + + because most of userland tools were not ready for accepting packed input. + + The advantages of using packed policy are that it makes policy files + smaller and it speeds up loading/saving policy files. + + Since most of userland tools are ready for accepting packed input by now, + I changed to use packed policy for both input and output.