git: d04c93a2adcc - main - libc: Don't bias DTV entries by TLS_DTV_OFFSET
- Go to: [ bottom of page ] [ top of archives ] [ this month ]
Date: Tue, 06 May 2025 22:15:29 UTC
The branch main has been updated by jrtc27:
URL: https://cgit.FreeBSD.org/src/commit/?id=d04c93a2adccb4c3a17f7391126a9246326e3fea
commit d04c93a2adccb4c3a17f7391126a9246326e3fea
Author: Jessica Clarke <jrtc27@FreeBSD.org>
AuthorDate: 2025-05-06 22:14:50 +0000
Commit: Jessica Clarke <jrtc27@FreeBSD.org>
CommitDate: 2025-05-06 22:14:50 +0000
libc: Don't bias DTV entries by TLS_DTV_OFFSET
PowerPC and RISC-V have a non-zero TLS_DTV_OFFSET. The intent behind
this in the design is that DTV entries are biased by this, as are (in
the other direction) the DTPOFF/DTPREL entries in the GOT. However, this
is pretty pointless in practice, and both FreeBSD and glibc's run-time
linkers don't bother to bias DTV entries, instead just adding the bias
back on at the end in __tls_get_addr. In libc we also have a minimal
implementation of this for statically-linked binaries, which is only in
practice used for code compiled with -fPIC (not -fPIE) that is also
linked without TLS relaxation support. PowerPC supports linker
relaxation for TLS sequences, so this likely never gets hit there, but
RISC-V does not, and so easily does if you compile an executable with
-fPIC. In this implementation we add TLS_DTV_OFFSET both to the DTV
entries in __libc_allocate_tls and to the result of __tls_get_addr,
meaning that any TLS accesses using the General Dynamic model in static
binaries on RISC-V end up off by 0x800.
Historically this also did not matter as __tls_get_addr was a stub that
always returned NULL, so although 6e16d0bc4376 ("Rework alignment
handling in __libc_allocate_tls() for Variant I of TLS layout.") added
this DTV implementation, nothing actually read the entries. However, now
it's a real implementation, and dl_iterate_phdr also now relies on it,
it does matter.
Fix this by not biasing the DTV entries, just like RTLD. We could
instead stop adding TLS_DTV_OFFSET in __tls_get_addr, but being
consistent between libc and RTLD seems better.
(Note this also applies to MIPS on stable/13)
Reviewed by: kib
Fixes: ca46b5698e8a ("libc: implement __tls_get_addr() for static binaries")
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D50181
---
lib/libc/gen/tls.c | 9 ++++-----
1 file changed, 4 insertions(+), 5 deletions(-)
diff --git a/lib/libc/gen/tls.c b/lib/libc/gen/tls.c
index 30b4fc583c98..8d453eccb801 100644
--- a/lib/libc/gen/tls.c
+++ b/lib/libc/gen/tls.c
@@ -151,7 +151,8 @@ libc_free_aligned(void *ptr)
* where TP points (with bias) to TLS and TCB immediately precedes TLS without
* any alignment gap[4]. Only TLS should be aligned. The TCB[0] points to DTV
* vector and DTV values are biased by constant value (TLS_DTV_OFFSET) from
- * real addresses[5].
+ * real addresses. However, like RTLD, we don't actually bias the DTV values,
+ * instead we compensate in __tls_get_addr for ti_offset's bias.
*
* [1] Ulrich Drepper: ELF Handling for Thread-Local Storage
* www.akkadia.org/drepper/tls.pdf
@@ -167,8 +168,6 @@ libc_free_aligned(void *ptr)
* but we must follow this rule due to suboptimal _tcb_set()
* (aka <ARCH>_SET_TP) implementation. This function doesn't expect TP but
* TCB as argument.
- *
- * [5] I'm not able to validate "values are biased" assertions.
*/
/*
@@ -272,7 +271,7 @@ __libc_allocate_tls(void *oldtcb, size_t tcbsize, size_t tcbalign)
/* Adjust the DTV. */
dtv = tcb[0];
- dtv[2] = (Elf_Addr)(tls + TLS_DTV_OFFSET);
+ dtv[2] = (Elf_Addr)tls;
} else {
dtv = __je_bootstrap_malloc(3 * sizeof(Elf_Addr));
if (dtv == NULL) {
@@ -283,7 +282,7 @@ __libc_allocate_tls(void *oldtcb, size_t tcbsize, size_t tcbalign)
tcb[0] = dtv;
dtv[0] = 1; /* Generation. */
dtv[1] = 1; /* Segments count. */
- dtv[2] = (Elf_Addr)(tls + TLS_DTV_OFFSET);
+ dtv[2] = (Elf_Addr)tls;
if (libc_tls_init_size > 0)
memcpy(tls, libc_tls_init, libc_tls_init_size);