remove compat_alloc_user_space()
Subsystems (7)
v4 cover remove compat_alloc_user_space()
From: Arnd Bergmann <[email protected]>
This is the fourth version of my series, now spanning four patches instead of two, with a new approach for handling struct ifreq compatibility after I realized that my earlier approach introduces additional problems.
The idea here is to always push down the compat conversion deeper into the call stack: rather than pretending to be native mode with a modified copy of the original data on the user space stack, have the code that actually works on the data understand the difference between native and compat versions.
I have spent a long time looking at all drivers that implement an ndo_do_ioctl callback to verify that my assumptions are correct. This has led to a series of 29 additional patches that I am not including here but will post separately, fixing a number of bugs in SIOCDEVPRIVATE ioctls, removing dead code, and splitting ndo_do_ioctl into two new ndo callbacks for private and ethernet specific commands.
The patches are identical to v3 that I sent a few weeks ago, except this avoids the build failure when CONFIG_COMPAT is disabled.
Arnd
Changes in v4:
- build fix without CONFIG_INET
- build fix without CONFIG_COMPAT
- style fixes pointed out by hch
Changes in v3:
- complete rewrite of the series
Arnd Bergmann (4):
ethtool: improve compat ioctl handling
net: socket: rework SIOC?IFMAP ioctls
net: socket: simplify dev_ifconf handling
net: socket: rework compat_ifreq_ioctl()
include/linux/compat.h | 82 +++++------ include/linux/ethtool.h | 4 - include/linux/inetdevice.h | 9 ++ include/linux/netdevice.h | 12 +- net/appletalk/ddp.c | 4 +- net/core/dev_ioctl.c | 152 ++++++++++--------- net/ethtool/ioctl.c | 143 ++++++++++++++++-- net/ieee802154/socket.c | 4 +- net/ipv4/af_inet.c | 6 +- net/ipv4/devinet.c | 4 +- net/qrtr/qrtr.c | 4 +- net/socket.c | 289 ++++++++----------------------------- 12 files changed, 338 insertions(+), 375 deletions(-)
Signature · 1 line
2.27.0
PATCH v4 1/4 ethtool: improve compat ioctl handling
From: Arnd Bergmann <[email protected]>
The ethtool compat ioctl handling is hidden away in net/socket.c, which introduces a couple of minor oddities:
- The implementation may end up diverging, as seen in the RXNFC extension in commit 84a1d9c48200 ("net: ethtool: extend RXNFC API to support RSS spreading of filter matches") that does not work in compat mode.
- Most architectures do not need the compat handling at all because u64 and compat_u64 have the same alignment.
- On x86, the conversion is done for both x32 and i386 user space, but it's actually wrong to do it for x32 and cannot work there.
- On 32-bit Arm, it never worked for compat oabi user space, since that needs to do the same conversion but does not.
- It would be nice to get rid of both compat_alloc_user_space() and copy_in_user() throughout the kernel.
None of these actually seems to be a serious problem that real users are likely to encounter, but fixing all of them actually leads to code that is both shorter and more readable.
---
Changes in v2:
- remove extraneous 'inline' keyword (davem)
- split helper functions into smaller units (hch)
- remove arm oabi check with missing dependency (0day bot)
--- include/linux/ethtool.h | 4 -- net/ethtool/ioctl.c | 143 +++++++++++++++++++++++++++++++++++----- net/socket.c | 125 +---------------------------------- 3 files changed, 128 insertions(+), 144 deletions(-)
include/linux/ethtool.h +0 -4
@@ -17,8 +17,6 @@#include <linux/compat.h> #include <uapi/linux/ethtool.h> -#ifdef CONFIG_COMPAT -struct compat_ethtool_rx_flow_spec {u32 flow_type;union ethtool_flow_union h_u;
@@ -38,8 +36,6 @@ struct compat_ethtool_rxnfc {u32 rule_locs[]; }; -#endif /* CONFIG_COMPAT */ -#include <linux/rculist.h> /**
net/ethtool/ioctl.c +127 -16
185 lines not on this page · raw message
net/socket.c +1 -124
143 lines not on this page · raw message
Signature · 1 line
2.27.0
PATCH v4 2/4 net: socket: rework SIOC?IFMAP ioctls
In reply to Arnd BergmannFrom: Arnd Bergmann <[email protected]>
SIOCGIFMAP and SIOCSIFMAP currently require compat_alloc_user_space() and copy_in_user() for compat mode.
Move the compat handling into the location where the structures are actually used, to avoid using those interfaces and get a clearer implementation.
---
changes in v3:
- complete rewrite
changes in v2:
- fix building with CONFIG_COMPAT disabled (0day bot)
- split up dev_ifmap() into more readable helpers (hch)
- move rcu_read_unlock() for readability (hch)
--- include/linux/compat.h | 18 ++++++------ net/core/dev_ioctl.c | 64 +++++++++++++++++++++++++++++++++--------- net/socket.c | 39 ++----------------------- 3 files changed, 62 insertions(+), 59 deletions(-)
include/linux/compat.h +9 -9
@@ -96,6 +96,15 @@ struct compat_iovec {compat_size_t iov_len; }; +struct compat_ifmap { +compat_ulong_t mem_start; +compat_ulong_t mem_end; +unsigned short base_addr; +unsigned char irq; +unsigned char dma; +unsigned char port; +}; +#ifdef CONFIG_COMPAT #ifndef compat_user_stack_pointer
@@ -314,15 +323,6 @@ typedef struct compat_sigevent { } _sigev_un; } compat_sigevent_t; -struct compat_ifmap { -compat_ulong_t mem_start; -compat_ulong_t mem_end; -unsigned short base_addr; -unsigned char irq; -unsigned char dma; -unsigned char port; -}; -struct compat_if_settings {unsigned int type;/* Type of physical device or protocol */ unsigned int size;/* Size of the data allocated by the caller */
net/core/dev_ioctl.c +51 -13
82 lines not on this page · raw message
net/socket.c +2 -37
@@ -3198,40 +3198,6 @@ static int compat_ifreq_ioctl(struct net *net, struct socket *sock,return err; } -static int compat_sioc_ifmap (struct net *net,unsigned int cmd, -struct compat_ifreq __user *uifr32) -{ -struct ifreq ifr; -struct compat_ifmap __user *uifmap32; -int err; - - uifmap32 = &uifr32->ifr_ifru.ifru_map; - err =copy_from_user (&ifr, uifr32,sizeof (ifr.ifr_name)); - err |=get_user (ifr.ifr_map.mem_start, &uifmap32->mem_start); - err |=get_user (ifr.ifr_map.mem_end, &uifmap32->mem_end); - err |=get_user (ifr.ifr_map.base_addr, &uifmap32->base_addr); - err |=get_user (ifr.ifr_map.irq, &uifmap32->irq); - err |=get_user (ifr.ifr_map.dma, &uifmap32->dma); - err |=get_user (ifr.ifr_map.port, &uifmap32->port); -if (err) -return -EFAULT ; - - err =dev_ioctl (net, cmd, &ifr,NULL ); - -if (cmd ==SIOCGIFMAP && !err) { - err =copy_to_user (uifr32, &ifr,sizeof (ifr.ifr_name)); - err |=put_user (ifr.ifr_map.mem_start, &uifmap32->mem_start); - err |=put_user (ifr.ifr_map.mem_end, &uifmap32->mem_end); - err |=put_user (ifr.ifr_map.base_addr, &uifmap32->base_addr); - err |=put_user (ifr.ifr_map.irq, &uifmap32->irq); - err |=put_user (ifr.ifr_map.dma, &uifmap32->dma); - err |=put_user (ifr.ifr_map.port, &uifmap32->port); -if (err) - err = -EFAULT ; - } -return err; -} -/* Since old style bridge ioctl's endup using SIOCDEVPRIVATE * for some operations; this forces use of the newer bridge-utils that * use compatible ioctls
@@ -3265,9 +3231,6 @@ static int compat_sock_ioctl_trans(struct file *file, struct socket *sock,return compat_dev_ifconf (net, argp);case SIOCWANDEV :return compat_siocwandev (net, argp); -case SIOCGIFMAP : -case SIOCSIFMAP : -return compat_sioc_ifmap (net, cmd, argp);case SIOCGSTAMP_OLD :case SIOCGSTAMPNS_OLD :if (!sock->ops->gettstamp)
@@ -3297,6 +3260,8 @@ static int compat_sock_ioctl_trans(struct file *file, struct socket *sock,case SIOCGIFFLAGS :case SIOCSIFFLAGS : +case SIOCGIFMAP : +case SIOCSIFMAP :case SIOCGIFMETRIC :case SIOCSIFMETRIC :case SIOCGIFMTU :
Signature · 1 line
2.27.0
PATCH v4 3/4 net: socket: simplify dev_ifconf handling
In reply to Arnd BergmannFrom: Arnd Bergmann <[email protected]>
The dev_ifconf() calling conventions make compat handling more complicated than necessary, simplify this by moving the in_compat_syscall() check into the function. The implementation can be simplified further, based on the knowledge that the dynamic registration is only ever used for IPv4.
--- include/linux/compat.h | 11 +++-- include/linux/inetdevice.h | 9 ++++ include/linux/netdevice.h | 10 +---- net/core/dev_ioctl.c | 92 +++++++++++++++----------------------- net/ipv4/devinet.c | 4 +- net/socket.c | 59 ++++++------------------ 6 files changed, 66 insertions(+), 119 deletions(-)
include/linux/compat.h +5 -6
@@ -105,6 +105,11 @@ struct compat_ifmap {unsigned char port; }; +struct compat_ifconf { +compat_int_t ifc_len;/* size of buffer */ +compat_uptr_t ifcbuf; +}; +#ifdef CONFIG_COMPAT #ifndef compat_user_stack_pointer
@@ -323,12 +328,6 @@ typedef struct compat_sigevent { } _sigev_un; } compat_sigevent_t; -struct compat_if_settings { -unsigned int type;/* Type of physical device or protocol */ -unsigned int size;/* Size of the data allocated by the caller */ -compat_uptr_t ifs_ifsu;/* union of pointers */ -}; -struct compat_ifreq {union {char ifrn_name[IFNAMSIZ ];/* if name, e.g. "en0" */
include/linux/inetdevice.h +9 -0
@@ -178,6 +178,15 @@ static inline struct net_device *ip_dev_find(struct net *net, __be32 addr)int inet_addr_onlink (struct in_device *in_dev,__be32 a,__be32 b);int devinet_ioctl (struct net *net,unsigned int cmd,struct ifreq *); +#ifdef CONFIG_INET +int inet_gifconf (struct net_device *dev,char __user *buf,int len,int size); +#else +static inline int inet_gifconf (struct net_device *dev,char __user *buf, +int len,int size) +{ +return 0 ; +} +#endif void devinet_init (void );struct in_device *inetdev_by_index (struct net *,int );__be32 inet_select_addr (const struct net_device *dev,__be32 dst,int scope);
include/linux/netdevice.h +1 -9
@@ -3163,14 +3163,6 @@ static inline bool dev_validate_header(const struct net_device *dev,return false; } -typedef int gifconf_func_t (struct net_device * dev,char __user * bufptr, -int len,int size); -int register_gifconf (unsigned int family,gifconf_func_t *gifconf); -static inline int unregister_gifconf (unsigned int family) -{ -return register_gifconf (family,NULL ); -} -#ifdef CONFIG_NET_FLOW_LIMIT #define FLOW_LIMIT_HISTORY (1 << 7)/* must be ^2 and !overflow buckets */ struct sd_flow_limit {
@@ -3870,7 +3862,7 @@ void netdev_rx_handler_unregister(struct net_device *dev);bool dev_valid_name (const char *name);int dev_ioctl (struct net *net,unsigned int cmd,struct ifreq *ifr,bool *need_copyout); -int dev_ifconf (struct net *net,struct ifconf *,int ); +int dev_ifconf (struct net *net,struct ifconf __user *ifc);int dev_ethtool (struct net *net,struct ifreq *);unsigned int dev_get_flags (const struct net_device *); int __dev_change_flags(struct net_device *dev, unsigned int flags,
net/core/dev_ioctl.c +36 -56
119 lines not on this page · raw message
net/ipv4/devinet.c +1 -3
@@ -1243,7 +1243,7 @@ int devinet_ioctl(struct net *net, unsigned int cmd, struct ifreq *ifr)return ret; } -static int inet_gifconf (struct net_device *dev,char __user *buf,int len,int size) +int inet_gifconf (struct net_device *dev,char __user *buf,int len,int size) {struct in_device *in_dev =__in_dev_get_rtnl (dev);const struct in_ifaddr *ifa;
@@ -2761,8 +2761,6 @@ void __init devinet_init(void)INIT_HLIST_HEAD (&inet_addr_lst[i]);register_pernet_subsys (&devinet_ops); - -register_gifconf (PF_INET , inet_gifconf);register_netdevice_notifier (&ip_netdev_notifier);queue_delayed_work (system_power_efficient_wq, &check_lifetime_work,0 );
net/socket.c +14 -45
96 lines not on this page · raw message
Signature · 1 line
2.27.0
PATCH v4 4/4 net: socket: rework compat_ifreq_ioctl()
In reply to Arnd BergmannFrom: Arnd Bergmann <[email protected]>
compat_ifreq_ioctl() is one of the last users of copy_in_user() and compat_alloc_user_space(), as it attempts to convert the 'struct ifreq' arguments from 32-bit to 64-bit format as used by dev_ioctl() and a couple of socket family specific interpretations.
The current implementation works correctly when calling dev_ioctl(), inet_ioctl(), ieee802154_sock_ioctl(), atalk_ioctl(), qrtr_ioctl() and packet_ioctl(). The ioctl handlers for x25, netrom, rose and x25 do not interpret the arguments and only block the corresponding commands, so they do not care.
For af_inet6 and af_decnet however, the compat conversion is slightly incorrect, as it will copy more data than the native handler accesses, both of them use a structure that is shorter than ifreq.
Replace the copy_in_user() conversion with a pair of accessor functions to read and write the ifreq data in place with the correct length where needed, while leaving the other ones to copy the (already compatible) structures directly.
--- include/linux/compat.h | 53 ++++++++++---------- include/linux/netdevice.h | 2 + net/appletalk/ddp.c | 4 +- net/ieee802154/socket.c | 4 +- net/ipv4/af_inet.c | 6 +-- net/qrtr/qrtr.c | 4 +- net/socket.c | 100 ++++++++++++++++++++++++-------------- 7 files changed, 101 insertions(+), 72 deletions(-)
include/linux/compat.h +27 -26
@@ -110,6 +110,33 @@ struct compat_ifconf {compat_uptr_t ifcbuf; }; +struct compat_if_settings { +unsigned int type;/* Type of physical device or protocol */ +unsigned int size;/* Size of the data allocated by the caller */ +compat_uptr_t ifs_ifsu;/* union of pointers */ +}; + +struct compat_ifreq { +union { +char ifrn_name[IFNAMSIZ ];/* if name, e.g. "en0" */ + } ifr_ifrn; +union { +struct sockaddr ifru_addr; +struct sockaddr ifru_dstaddr; +struct sockaddr ifru_broadaddr; +struct sockaddr ifru_netmask; +struct sockaddr ifru_hwaddr; +short ifru_flags; +compat_int_t ifru_ivalue; +compat_int_t ifru_mtu; +struct compat_ifmap ifru_map; +char ifru_slave[IFNAMSIZ ];/* Just fits the size */ +char ifru_newname[IFNAMSIZ ]; +compat_uptr_t ifru_data; +struct compat_if_settings ifru_settings; + } ifr_ifru; +}; +#ifdef CONFIG_COMPAT #ifndef compat_user_stack_pointer
@@ -328,32 +355,6 @@ typedef struct compat_sigevent { } _sigev_un; } compat_sigevent_t; -struct compat_ifreq { -union { -char ifrn_name[IFNAMSIZ ];/* if name, e.g. "en0" */ - } ifr_ifrn; -union { -struct sockaddr ifru_addr; -struct sockaddr ifru_dstaddr; -struct sockaddr ifru_broadaddr; -struct sockaddr ifru_netmask; -struct sockaddr ifru_hwaddr; -short ifru_flags; -compat_int_t ifru_ivalue; -compat_int_t ifru_mtu; -struct compat_ifmap ifru_map; -char ifru_slave[IFNAMSIZ ];/* Just fits the size */ -char ifru_newname[IFNAMSIZ ]; -compat_caddr_t ifru_data; -struct compat_if_settings ifru_settings; - } ifr_ifru; -}; - -struct compat_ifconf { -compat_int_t ifc_len;/* size of buffer */ -compat_caddr_t ifcbuf; -}; -struct compat_robust_list {compat_uptr_t next; };
include/linux/netdevice.h +2 -0
@@ -3860,6 +3860,8 @@ int netdev_rx_handler_register(struct net_device *dev,void netdev_rx_handler_unregister (struct net_device *dev);bool dev_valid_name (const char *name); +int get_user_ifreq (struct ifreq *ifr,void __user **ifrdata,void __user *arg); +int put_user_ifreq (struct ifreq *ifr,void __user *arg);int dev_ioctl (struct net *net,unsigned int cmd,struct ifreq *ifr,bool *need_copyout);int dev_ifconf (struct net *net,struct ifconf __user *ifc);
net/appletalk/ddp.c +2 -2
@@ -666,7 +666,7 @@ static int atif_ioctl(int cmd, void __user *arg)struct rtentry rtdef;int add_route; -if (copy_from_user (&atreq, arg,sizeof (atreq))) +if (get_user_ifreq (&atreq,NULL , arg))return -EFAULT ; dev =__dev_get_by_name (&init_net, atreq.ifr_name);
@@ -865,7 +865,7 @@ static int atif_ioctl(int cmd, void __user *arg)return 0 ; } -return copy_to_user (arg, &atreq,sizeof (atreq)) ? -EFAULT :0 ; +return put_user_ifreq (&atreq, arg); }static int atrtr_ioctl_addrt (struct rtentry *rt)
net/ieee802154/socket.c +2 -2
@@ -129,7 +129,7 @@ static int ieee802154_dev_ioctl(struct sock *sk, struct ifreq __user *arg,int ret = -ENOIOCTLCMD ;struct net_device *dev; -if (copy_from_user (&ifr, arg,sizeof (struct ifreq ))) +if (get_user_ifreq (&ifr,NULL , arg))return -EFAULT ; ifr.ifr_name[IFNAMSIZ -1 ] =0 ;
@@ -143,7 +143,7 @@ static int ieee802154_dev_ioctl(struct sock *sk, struct ifreq __user *arg,if (dev->type ==ARPHRD_IEEE802154 && dev->netdev_ops->ndo_do_ioctl) ret = dev->netdev_ops->ndo_do_ioctl (dev, &ifr, cmd); -if (!ret &©_to_user (arg, &ifr,sizeof (struct ifreq ))) +if (!ret &&put_user_ifreq (&ifr, arg)) ret = -EFAULT ;dev_put (dev);
net/ipv4/af_inet.c +3 -3
@@ -949,10 +949,10 @@ int inet_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)case SIOCGIFNETMASK :case SIOCGIFDSTADDR :case SIOCGIFPFLAGS : -if (copy_from_user (&ifr, p,sizeof (struct ifreq ))) +if (get_user_ifreq (&ifr,NULL , p))return -EFAULT ; err =devinet_ioctl (net, cmd, &ifr); -if (!err &©_to_user (p, &ifr,sizeof (struct ifreq ))) +if (!err &&put_user_ifreq (&ifr, p)) err = -EFAULT ;break ;
@@ -962,7 +962,7 @@ int inet_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)case SIOCSIFDSTADDR :case SIOCSIFPFLAGS :case SIOCSIFFLAGS : -if (copy_from_user (&ifr, p,sizeof (struct ifreq ))) +if (get_user_ifreq (&ifr,NULL , p))return -EFAULT ; err =devinet_ioctl (net, cmd, &ifr);break ;
net/qrtr/qrtr.c +2 -2
@@ -1157,14 +1157,14 @@ static int qrtr_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg) rc =put_user (len, (int __user *)argp);break ;case SIOCGIFADDR : -if (copy_from_user (&ifr, argp,sizeof (ifr))) { +if (get_user_ifreq (&ifr,NULL , argp)) { rc = -EFAULT ;break ; } sq = (struct sockaddr_qrtr *)&ifr.ifr_addr; *sq = ipc->us; -if (copy_to_user (argp, &ifr,sizeof (ifr))) { +if (put_user_ifreq (&ifr, argp)) { rc = -EFAULT ;break ; }
net/socket.c +63 -37
130 lines not on this page · raw message
Signature · 1 line
2.27.0
RE: [PATCH v4 2/4] net: socket: rework SIOC?IFMAP ioctls
In reply to Arnd BergmannFrom: Arnd Bergmann
Quoted 13 lines
Sent: 24 November 2020 15:18
SIOCGIFMAP and SIOCSIFMAP currently require
compat_alloc_user_space()andcopy_in_user()for compat mode.Move the compat handling into the location where the structures are actually used, to avoid using those interfaces and get a clearer implementation.
Signed-off-by Arnd Bergmann <[email protected]>---
changes in v3:
- complete rewrite
...
Quoted 11 lines
include/linux/compat.h | 18 ++++++------ net/core/dev_ioctl.c | 64 +++++++++++++++++++++++++++++++++--------- net/socket.c | 39 ++----------------------- 3 files changed, 62 insertions(+), 59 deletions(-)include/linux/compat.h +0
-0@@ -96,6 +96,15 @@ struct compat_iovec {compat_size_t iov_len;
include/linux/compat.h struct compat_iovec { · line 106
Quoted 10 lines
include/linux/compat.h +8
-0}; +struct compat_ifmap { +compat_ulong_t mem_start; +compat_ulong_t mem_end; +unsigned short base_addr; +unsigned char irq; +unsigned char dma; +unsigned char port; +};
Isn't the only difference the number of pad bytes at the end? If you don't copy these (in or out) then the compat version isn't special at all.
Not copying the pad in or out would ensure you don't leak kernel stack to userspace.
OTOH you may want to write the padding zero.
So a CT_ASSERT(offsetof (struct ifmap, port) == offsetof (struct compat_ifmap, port)) would suffice.
Maybe a CT_ASSERT_EQ_OFFSET(struct ifmap, struct compat_ifmap, port); Would make the code easier to read.
Although you might want the version that adds an offset
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
RE: [PATCH v4 1/4] ethtool: improve compat ioctl handling
In reply to Arnd BergmannFrom: Arnd Bergmann
Quoted 5 lines
Sent: 24 November 2020 15:18
The ethtool compat ioctl handling is hidden away in
net/socket.c, which introduces a couple of minor oddities:
...
net/ethtool/ioctl.c static noinline_for_stack int ethtool_get_sset_info(struct net_device *dev, · line 826
Quoted 5 lines
net/ethtool/ioctl.c +5
-0+ +static int ethtool_rxnfc_copy_from_compat (struct ethtool_rxnfc *rxnfc, +const struct compat_ethtool_rxnfc __user *useraddr, +size_t size) +{
I think this (and possibly others) want a 'noinline_for_stack'. So that both the normal and compat structures aren't both on the stack when the real code is called.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
Re: [PATCH v4 2/4] net: socket: rework SIOC?IFMAP ioctls
In reply to David LaightQuoted 29 lines · David Laight
From: Arnd Bergmann
Quoted 13 lines
Sent: 24 November 2020 15:18
SIOCGIFMAP and SIOCSIFMAP currently require
compat_alloc_user_space()andcopy_in_user()for compat mode.Move the compat handling into the location where the structures are actually used, to avoid using those interfaces and get a clearer implementation.
Signed-off-by Arnd Bergmann <[email protected]>---
changes in v3:
- complete rewrite...
Quoted 13 lines
include/linux/compat.h | 18 ++++++------ net/core/dev_ioctl.c | 64 +++++++++++++++++++++++++++++++++--------- net/socket.c | 39 ++----------------------- 3 files changed, 62 insertions(+), 59 deletions(-)include/linux/compat.h +0
-0@@ -96,6 +96,15 @@ struct compat_iovec {compat_size_t iov_len; };
Quoted 10 lines
include/linux/compat.h struct compat_iovec { · line 106
Quoted 8 lines
include/linux/compat.h +8
-0+struct compat_ifmap { +compat_ulong_t mem_start; +compat_ulong_t mem_end; +unsigned short base_addr; +unsigned char irq; +unsigned char dma; +unsigned char port; +};Isn't the only difference the number of pad bytes at the end?
No, the main difference is in the first two fields, which are 'unsigned long' and therefore different. The three-byte padding is in fact the same on all architectures (including x86) that have a compat mode, though it might be different on m68k and arm-oabi, which have slightly special struct alignment rules.
It could be done with two assignments and a memcpy, but I like the individual assignments better here.
Arnd
Re: [PATCH v4 1/4] ethtool: improve compat ioctl handling
In reply to David LaightQuoted 8 lines · David Laight
From: Arnd Bergmann
Quoted 5 lines
Sent: 24 November 2020 15:18
The ethtool compat ioctl handling is hidden away in
net/socket.c, which introduces a couple of minor oddities:...
Quoted 9 lines
net/ethtool/ioctl.c static noinline_for_stack int ethtool_get_sset_info(struct net_device *dev, · line 826
Quoted 5 lines
net/ethtool/ioctl.c +5
-0+ +static int ethtool_rxnfc_copy_from_compat (struct ethtool_rxnfc *rxnfc, +const struct compat_ethtool_rxnfc __user *useraddr, +size_t size) +{I think this (and possibly others) want a '
noinline_for_stack'. So that both the normal and compat structures aren't both on the stack when the real code is called.
Yes, makes sense. I checked that the difference is small unless CONFIG_KASAN_STACK is enabled, and that gcc is smart enough not to inline these by default, but adding noinline_for_stack is both consistent with the rest of the file and the safe bet here.
Arnd
Jakub Kicinski
maintainerRe: [PATCH v4 3/4] net: socket: simplify dev_ifconf handling
In reply to Arnd BergmannQuoted 8 lines · On Tue, 24 Nov 2020 16:18:27 +0100 Arnd Bergmann wrote:
From: Arnd Bergmann <[email protected]>
The
dev_ifconf()calling conventions make compat handling more complicated than necessary, simplify this by moving thein_compat_syscall()check into the function. The implementation can be simplified further, based on the knowledge that the dynamic registration is only ever used for IPv4.
Looks like this one breaks bisection (/breaks build which patch 4 then fixes):
In file included from ../arch/x86/include/asm/sigframe.h:8,
from ../arch/x86/kernel/asm-offsets.c:17:
../include/linux/compat.h:348:29: error: field ‘ifru_settings’ has incomplete type
348 | struct compat_if_settings ifru_settings;
| ^~~~~~~~~~~~~
../include/linux/compat.h:352:8: error: redefinition of ‘struct compat_ifconf’
352 | struct compat_ifconf {
| ^~~~~~~~~~~~~
../include/linux/compat.h:108:8: note: originally defined here
108 | struct compat_ifconf {
| ^~~~~~~~~~~~~make[2]: *** [arch/x86/kernel/asm-offsets.s] Error 1 make[1]: *** [prepare0] Error 2
make: *** [__sub-make] Error 2
RE: [PATCH v4 2/4] net: socket: rework SIOC?IFMAP ioctls
In reply to Arnd BergmannFrom: Arnd Bergmann
Quoted 50 lines
Sent: 24 November 2020 19:06
Quoted 29 lines · David Laight
From: Arnd Bergmann
Quoted 13 lines
Sent: 24 November 2020 15:18
SIOCGIFMAP and SIOCSIFMAP currently require
compat_alloc_user_space()andcopy_in_user()for compat mode.Move the compat handling into the location where the structures are actually used, to avoid using those interfaces and get a clearer implementation.
Signed-off-by Arnd Bergmann <[email protected]>---
changes in v3:
- complete rewrite...
Quoted 13 lines
include/linux/compat.h | 18 ++++++------ net/core/dev_ioctl.c | 64 +++++++++++++++++++++++++++++++++--------- net/socket.c | 39 ++----------------------- 3 files changed, 62 insertions(+), 59 deletions(-)include/linux/compat.h +0
-0@@ -96,6 +96,15 @@ struct compat_iovec {compat_size_t iov_len; };Quoted 10 lines
include/linux/compat.h struct compat_iovec { · line 106
Quoted 8 lines
include/linux/compat.h +8
-0+struct compat_ifmap { +compat_ulong_t mem_start; +compat_ulong_t mem_end; +unsigned short base_addr; +unsigned char irq; +unsigned char dma; +unsigned char port; +};Isn't the only difference the number of pad bytes at the end?
No, the main difference is in the first two fields, which are 'unsigned long' and therefore different. The three-byte padding is in fact the same on all architectures (including x86) that have a compat mode, though it might be different on m68k and arm-oabi, which have slightly special struct alignment rules.
Quoted 2 lines
It could be done with two assignments and a memcpy, but I like the individual assignments better here.
Gah my brain hurts today.
I was just thinking of the alignment and padding, not the sizes.
You could read the compat structure to 'mem_end' and then move the first two fields forward.
But, I guess, this structure doesn't have many fields.
What you really need for these copies is COBOL's 'move corresponding'. (Which wasn't implemented by the only COBOL compiler I've used.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
Just curious: what is the status of the various compat_alloc_user_space removal patches? It would be great to kill the function off for good.