crypto: pcrypt - Disallow nesting of the pcrypt wrapper
Subsystems (2)
RFC PATCH 1/1 crypto: pcrypt - Disallow nesting of the pcrypt wrapper
From: Thomas Huth <[email protected]>
When running syzkaller in parallel with "modprobe tcrypt", it's possible to get a call trace with Linux v7.1:
alg: aead: failed to allocate transform for rfc4543(gcm(aes)): -36
alg: self-tests for rfc4543(gcm(aes)) using rfc4543(gcm(aes)) failed (rc=-36)
------------[ cut here ]------------ alg: No test for cbc(cipher_null) (cbc(cipher_null-generic)) alg: self-tests for rfc4543(gcm(aes)) using rfc4543(gcm(aes)) failed (rc=-36) WARNING: crypto/testmgr.c:5798 at alg_test.cold+0x85/0x98, CPU#198: modprobe/65089 Modules linked in: tcrypt(+) adiantum libnh chacha20poly1305 echainiv pcrypt essiv crc32_cryptoapi chacha ip6_vti ip_vti ip_gre ipip sit ip_tunnel geneve macvtap tap ipvlan macvlan 8021q garp mrp hsr xfrm_interface xfrm6_tunnel tunnel4 wireguard libcurve25519 ip6_udp_tunnel udp_tunnel veth vxcan nlmon dummy bonding vcan can_dev bnep bridge stp llc ip6_gre gre hci_vhci cfg80211 ip6_tunnel tunnel6 bluetooth tun crc16 binfmt_misc xcbc cmac blake2b libblake2b rmd160 xxhash_generic ccm camellia_generic camellia_aesni_avx2 camellia_aesni_avx_x86_64 camellia_x86_64 fcrypt pcbc wp512 cast6_avx_x86_64 cast6_generic cast5_avx_x86_64 cast5_generic cast_common serpent_avx2 serpent_avx_x86_64 serpent_sse2_x86_64 serpent_generic lrw twofish_generic twofish_avx_x86_64 twofish_x86_64_3way twofish_x86_64 twofish_common blowfish_generic blowfish_x86_64 blowfish_common md4 des_generic libdes qrtr rfkill vfat amd_atl fat intel_rapl_msr intel_rapl_common amd64_edac edac_mce_amd mlx5_ib kvm_amd ib_uverbs macsec kvm ipmi_si ib_core ipmi_devintf irqbypass ipmi_msghandler wmi_bmof pcspkr rapl acpi_cpufreq mlx5_fwctl i2c_piix4 fwctl ptdma k10temp i2c_smbus joydev i2c_designware_platform i2c_designware_core sg loop xfs nvme_tcp nvme_fabrics nvme_core sd_mod nvme_keyring nvme_auth mlx5_core ahci libahci ast libata ccp mlxfw i2c_algo_bit psample sp5100_tco wmi dm_mirror dm_region_hash dm_log be2iscsi iscsi_boot_sysfs bnx2i cnic uio cxgb4i cxgb4 tls libcxgbi libcxgb iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi fuse i2c_dev dm_multipath dm_mod nfnetlink CPU: 198 UID: 0 PID: 65089 Comm: modprobe Kdump: loaded Not tainted 7.1.0 #34 PREEMPT(full) Hardware name: AMD Corporation DAYTONA_X/DAYTONA_X, BIOS RYM100GA 01/13/2025 RIP: 0010:alg_test.cold+0x8e/0x98 Code: 62 9a fe ff e9 84 16 6a 00 48 c7 c7 20 cf d4 a4 e8 51 9a fe ff e9 17 18 6a 00 48 8d 3d 05 7d 05 02 44 89 e9 48 89 ea 4c 89 e6 <67> 48 0f b9 3a e9 0d 16 6a 00 4d 8b 4d 00 48 8b 54 24 10 41 89 c0 RSP: 0018:ffffcba1b4ad78e8 EFLAGS: 00010297 RAX: 000000000000004d RBX: 0000000000000000 RCX: 00000000ffffffdc RDX: ffffffffc166046f RSI: ffffffffc166046f RDI: ffffffffa567f3a0 RBP: ffffffffc166046f R08: 0000000000000000 R09: 3ffffffffffeffff R10: ffffcba1b4ad7768 R11: 0000000000000003 R12: ffffffffc166046f R13: 00000000ffffffdc R14: 00000000000000bb R15: 0000000000000000 FS: 00007f211da68b80(0000) GS:ffff8aea38d18000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007efc35ca8e28 CR3: 00000001e0fe3005 CR4: 0000000000f70ef0 PKRU: 55555554 Call Trace: <TASK> ? get_page_from_freelist+0x201/0x790 ? srso_alias_return_thunk+0x5/0xfbef5 ? __alloc_frozen_pages_noprof+0x194/0x340 do_test+0xeb4/0x5b80 [tcrypt] do_test+0x5b55/0x5b80 [tcrypt] tcrypt_mod_init+0x79/0xff0 [tcrypt] ? __pfx_tcrypt_mod_init+0x10/0x10 [tcrypt] do_one_initcall+0x5c/0x310 do_init_module+0x60/0x240 init_module_from_file+0xd6/0x130 idempotent_init_module+0x114/0x310 __x64_sys_finit_module+0x71/0xe0 do_syscall_64+0xe3/0x540 ? srso_alias_return_thunk+0x5/0xfbef5 ? ksys_read+0x6b/0xe0 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x11a/0x540 ? __pfx_kfree_link+0x10/0x10 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_sys_openat2+0x9d/0xe0 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x11a/0x540 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x11a/0x540 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_user_addr_fault+0x206/0x680 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x98/0x540 entry_SYSCALL_64_after_hwframe+0x76/0x7e RIP: 0033:0x7f211d32b15d Code: ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 8b ac 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffe9066c218 EFLAGS: 00000246 ORIG_RAX: 0000000000000139 RAX: ffffffffffffffda RBX: 000055adeea20b80 RCX: 00007f211d32b15d RDX: 0000000000000000 RSI: 000055adcb64ae79 RDI: 0000000000000003 RBP: 000055adcb64ae79 R08: 0000000000000070 R09: 000055adeea20b80 R10: 000055adeea218f0 R11: 0000000000000246 R12: 0000000000040000 R13: 000055adeea28690 R14: 0000000000000000 R15: 000055adeea20d20 </TASK> ---[ end trace 0000000000000000 ]--- tcrypt: one or more tests failed!
The problem seems to be that syzkaller instantiated the aes algorithm with a lot of pcrypt(pcrypt(pcrypt(...))) wrappers shortly before tcrypt tries to use it, so adding the rfc4543 prefix in gcm.c can fail with -ENAMETOOLONG when trcypt tries to instantiate it, too.
That means a malicious user space program can block the usage of crypto algorithms for other user space programs by instantiating it with a lot of pcrypt() wrappers, so that the name is close to the length limit.
There does not seem to be any compelling reason for allowing nested pcrypt wrappers, and nesting also lead to problems in the past already, see e.g. commit bbefa1dd6a6d5 ("crypto: pcrypt - Avoid deadlock by using per-instance padata queues"), so let's fix the problem by disallowing nesting here.
---
Note: Marked as "RFC" since I'm not that familiar with the crypto subsystem ... is there a better way to prevent nested pcryptos than searching for the string the the cra_driver_name? "Claude" suggests to set "mask |= CRYPTO_ALG_ASYNC;" in pcrypt_create_aead() instead, similar to what is done in cryptd_type_and_mask(), would that be a better solution?
Also note that the problem does not seem to reproduce (or at least reproduces less likely) in kernel 7.2-rc1 anymore, likely because commit 7524070f26d8 ("crypto: af_alg - Drop support for off-CPU cryptography") dropped the similar algorithms that could be instantiated in parallel? But I think it is likely still a good idea to disallow nesting of the pcrypt wrapper, so I'm sending out this patch here for discussion.
crypto/pcrypt.c | 3 +++ 1 file changed, 3 insertions(+)
crypto/pcrypt.c +3 -0
@@ -228,6 +228,9 @@ static void pcrypt_free(struct aead_instance *inst)static int pcrypt_init_instance (struct crypto_instance *inst,struct crypto_alg *alg) { +if (strstr (alg->cra_driver_name,"pcrypt" )) +return -ELOOP ;/* Nesting of pcrypt is not allowed */ +if (snprintf (inst->alg.cra_driver_name,CRYPTO_MAX_ALG_NAME ,"pcrypt(%s)" , alg->cra_driver_name) >=CRYPTO_MAX_ALG_NAME )return -ENAMETOOLONG ;
Signature · 1 line
2.54.0
Herbert Xu
maintainerQuoted 111 lines · Thomas Huth
From: Thomas Huth <[email protected]>
When running syzkaller in parallel with "modprobe tcrypt", it's possible to get a call trace with Linux v7.1:
alg: aead: failed to allocate transform for rfc4543(gcm(aes)): -36
alg: self-tests for rfc4543(gcm(aes)) using rfc4543(gcm(aes)) failed (rc=-36)------------[ cut here ]------------ alg: No test for cbc(cipher_null) (cbc(cipher_null-generic)) alg: self-tests for rfc4543(gcm(aes)) using rfc4543(gcm(aes)) failed (rc=-36) WARNING: crypto/testmgr.c:5798 at alg_test.cold+0x85/0x98, CPU#198: modprobe/65089 Modules linked in: tcrypt(+) adiantum libnh chacha20poly1305 echainiv pcrypt essiv crc32_cryptoapi chacha ip6_vti ip_vti ip_gre ipip sit ip_tunnel geneve macvtap tap ipvlan macvlan 8021q garp mrp hsr xfrm_interface xfrm6_tunnel tunnel4 wireguard libcurve25519 ip6_udp_tunnel udp_tunnel veth vxcan nlmon dummy bonding vcan can_dev bnep bridge stp llc ip6_gre gre hci_vhci cfg80211 ip6_tunnel tunnel6 bluetooth tun crc16 binfmt_misc xcbc cmac blake2b libblake2b rmd160 xxhash_generic ccm camellia_generic camellia_aesni_avx2 camellia_aesni_avx_x86_64 camellia_x86_64 fcrypt pcbc wp512 cast6_avx_x86_64 cast6_generic cast5_avx_x86_64 cast5_generic cast_common serpent_avx2 serpent_avx_x86_64 serpent_sse2_x86_64 serpent_generic lrw twofish_generic twofish_avx_x86_64 twofish_x86_64_3way twofish_x86_64 twofish_common blowfish_generic blowfish_x86_64 blowfish_common md4 des_generic libdes qrtr rfkill vfat amd_atl fat intel_rapl_msr intel_rapl_common amd64_edac edac_mce_amd mlx5_ib kvm_amd ib_uverbs macsec kvm ipmi_si ib_core ipmi_devintf irqbypass ipmi_msghandler wmi_bmof pcspkr rapl acpi_cpufreq mlx5_fwctl i2c_piix4 fwctl ptdma k10temp i2c_smbus joydev i2c_designware_platform i2c_designware_core sg loop xfs nvme_tcp nvme_fabrics nvme_core sd_mod nvme_keyring nvme_auth mlx5_core ahci libahci ast libata ccp mlxfw i2c_algo_bit psample sp5100_tco wmi dm_mirror dm_region_hash dm_log be2iscsi iscsi_boot_sysfs bnx2i cnic uio cxgb4i cxgb4 tls libcxgbi libcxgb iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi fuse i2c_dev dm_multipath dm_mod nfnetlink CPU: 198 UID: 0 PID: 65089 Comm: modprobe Kdump: loaded Not tainted 7.1.0 #34 PREEMPT(full) Hardware name: AMD Corporation DAYTONA_X/DAYTONA_X, BIOS RYM100GA 01/13/2025 RIP: 0010:alg_test.cold+0x8e/0x98 Code: 62 9a fe ff e9 84 16 6a 00 48 c7 c7 20 cf d4 a4 e8 51 9a fe ff e9 17 18 6a 00 48 8d 3d 05 7d 05 02 44 89 e9 48 89 ea 4c 89 e6 <67> 48 0f b9 3a e9 0d 16 6a 00 4d 8b 4d 00 48 8b 54 24 10 41 89 c0 RSP: 0018:ffffcba1b4ad78e8 EFLAGS: 00010297 RAX: 000000000000004d RBX: 0000000000000000 RCX: 00000000ffffffdc RDX: ffffffffc166046f RSI: ffffffffc166046f RDI: ffffffffa567f3a0 RBP: ffffffffc166046f R08: 0000000000000000 R09: 3ffffffffffeffff R10: ffffcba1b4ad7768 R11: 0000000000000003 R12: ffffffffc166046f R13: 00000000ffffffdc R14: 00000000000000bb R15: 0000000000000000 FS: 00007f211da68b80(0000) GS:ffff8aea38d18000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007efc35ca8e28 CR3: 00000001e0fe3005 CR4: 0000000000f70ef0 PKRU: 55555554 Call Trace: <TASK> ? get_page_from_freelist+0x201/0x790 ? srso_alias_return_thunk+0x5/0xfbef5 ? __alloc_frozen_pages_noprof+0x194/0x340 do_test+0xeb4/0x5b80 [tcrypt] do_test+0x5b55/0x5b80 [tcrypt] tcrypt_mod_init+0x79/0xff0 [tcrypt] ? __pfx_tcrypt_mod_init+0x10/0x10 [tcrypt] do_one_initcall+0x5c/0x310 do_init_module+0x60/0x240 init_module_from_file+0xd6/0x130 idempotent_init_module+0x114/0x310 __x64_sys_finit_module+0x71/0xe0 do_syscall_64+0xe3/0x540 ? srso_alias_return_thunk+0x5/0xfbef5 ? ksys_read+0x6b/0xe0 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x11a/0x540 ? __pfx_kfree_link+0x10/0x10 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_sys_openat2+0x9d/0xe0 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x11a/0x540 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x11a/0x540 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_user_addr_fault+0x206/0x680 ? srso_alias_return_thunk+0x5/0xfbef5 ? do_syscall_64+0x98/0x540 entry_SYSCALL_64_after_hwframe+0x76/0x7e RIP: 0033:0x7f211d32b15d Code: ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 8b ac 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffe9066c218 EFLAGS: 00000246 ORIG_RAX: 0000000000000139 RAX: ffffffffffffffda RBX: 000055adeea20b80 RCX: 00007f211d32b15d RDX: 0000000000000000 RSI: 000055adcb64ae79 RDI: 0000000000000003 RBP: 000055adcb64ae79 R08: 0000000000000070 R09: 000055adeea20b80 R10: 000055adeea218f0 R11: 0000000000000246 R12: 0000000000040000 R13: 000055adeea28690 R14: 0000000000000000 R15: 000055adeea20d20 </TASK> ---[ end trace 0000000000000000 ]--- tcrypt: one or more tests failed!The problem seems to be that syzkaller instantiated the aes algorithm with a lot of pcrypt(pcrypt(pcrypt(...))) wrappers shortly before tcrypt tries to use it, so adding the rfc4543 prefix in
gcm.ccan fail with-ENAMETOOLONGwhen trcypt tries to instantiate it, too.That means a malicious user space program can block the usage of crypto algorithms for other user space programs by instantiating it with a lot of
pcrypt()wrappers, so that the name is close to the length limit.There does not seem to be any compelling reason for allowing nested pcrypt wrappers, and nesting also lead to problems in the past already, see e.g. commit
bbefa1dd6a6d5("crypto: pcrypt - Avoid deadlock by using per-instance padata queues"), so let's fix the problem by disallowing nesting here.Signed-off-by Thomas Huth <[email protected]>---
Note: Marked as "RFC" since I'm not that familiar with the crypto subsystem ... is there a better way to prevent nested pcryptos than searching for the string the thecra_driver_name? "Claude" suggests to set "mask |=CRYPTO_ALG_ASYNC;" inpcrypt_create_aead()instead, similar to what is done incryptd_type_and_mask(), would that be a better solution?Also note that the problem does not seem to reproduce (or at least reproduces less likely) in kernel 7.2-rc1 anymore, likely because commit
7524070f26d8("crypto:af_alg- Drop support for off-CPU cryptography") dropped the similar algorithms that could be instantiated in parallel? But I think it is likely still a good idea to disallow nesting of the pcrypt wrapper, so I'm sending out this patch here for discussion.crypto/pcrypt.c | 3 +++ 1 file changed, 3 insertions(+)crypto/pcrypt.c +0
-0
crypto/pcrypt.c static void pcrypt_free(struct aead_instance *inst) · line 232
Quoted 5 lines
crypto/pcrypt.c +2
-0@@ -228,6 +228,9 @@ static void pcrypt_free(struct aead_instance *inst)static int pcrypt_init_instance (struct crypto_instance *inst,struct crypto_alg *alg) { +if (strstr (alg->cra_driver_name,"pcrypt" )) +return -ELOOP ;/* Nesting of pcrypt is not allowed */
This doesn't fix the problem completely since you can nest in other ways, e.g., pcrypt(cryptd(pcrypt(...))). How about handling the name- too-long error more gracefully?
Cheers,
Signature · 3 lines
Email: Herbert Xu <[email protected]> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
Quoted 3 lines · Herbert Xu
This doesn't fix the problem completely since you can nest in other ways, e.g., pcrypt(cryptd(pcrypt(...))). How about handling the name- too-long error more gracefully?
Could we just delete pcrypt instead of continuing to try to fix all the weird problems it has? A web search for pcrypt just finds CVEs and advice not to use it, e.g.
https://github.com/libreswan/libreswan/wiki/Internals:-Cryptographic-Acceleration#obsoleted-ipsec-accelerations
What is the rationale for keeping it? Who cares about it, other than people looking for vulnerabilities?
- Eric
Herbert Xu
maintainerQuoted 8 lines · Eric Biggers
Could we just delete pcrypt instead of continuing to try to fix all the weird problems it has? A web search for pcrypt just finds CVEs and advice not to use it, e.g.
https://github.com/libreswan/libreswan/wiki/Internals:-Cryptographic-Acceleration#obsoleted-ipsec-accelerationsWhat is the rationale for keeping it? Who cares about it, other than people looking for vulnerabilities?
I wouldn't object. Last time I tried it the start-up time was about 20,000 cycles, which is just pointless these days.
Please send a patch.
Thanks,
Signature · 3 lines
Email: Herbert Xu <[email protected]> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
Steffen Klassert
maintainerQuoted 9 lines · Eric Biggers
Quoted 3 lines · Herbert Xu
This doesn't fix the problem completely since you can nest in other ways, e.g., pcrypt(cryptd(pcrypt(...))). How about handling the name- too-long error more gracefully?
Could we just delete pcrypt instead of continuing to try to fix all the weird problems it has? A web search for pcrypt just finds CVEs and advice not to use it, e.g.
https://github.com/libreswan/libreswan/wiki/Internals:-Cryptographic-Acceleration#obsoleted-ipsec-accelerations
The comments on pctypt there are not quite right, but it is obsolte, that's correct. It was usefull back in the days when the overhead of (slow) software crypto was high compared to an IPI. I guess there are not many users left, so I'm OK with removing it.