Source-Changes-HG archive
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
[src/netbsd-9]: src/sys/kern Pull up following revision(s) (requested by rias...
details: https://anonhg.NetBSD.org/src/rev/d01257b505af
branches: netbsd-9
changeset: 984093:d01257b505af
user: martin <martin%NetBSD.org@localhost>
date: Mon Jun 21 16:14:14 2021 +0000
description:
Pull up following revision(s) (requested by riastradh in ticket #1299):
sys/kern/kern_ksyms.c: revision 1.90
sys/kern/kern_ksyms.c: revision 1.91
sys/kern/kern_ksyms.c: revision 1.92
sys/kern/kern_ksyms.c: revision 1.93
sys/kern/kern_ksyms.c: revision 1.94
sys/kern/kern_ksyms.c: revision 1.95
sys/kern/kern_ksyms.c: revision 1.96
sys/kern/kern_ksyms.c: revision 1.97
ksyms(4): Fix ksymsread synchronization.
Fixes crash on concurrent update and read of /dev/ksyms.
XXX Unclear why we have to skip sd_gone entries here -- it seems like
they should be preserved until ksymsclose.
ksyms(4): Modify ksyms_symtabs only at IPL_HIGH.
This limits the opportunities for ddb to witness an inconsistent
state of the symbol table list.
ksyms(4): Don't skip symbol tables that are soon to be freed.
They will not actually be freed until /dev/ksyms is closed, so
continued access to them remains kosher.
Revert "ksyms(4): Don't skip symbol tables that are soon to be freed."
Apparently the equality kassert this restored doesn't work; to be
analyzed.
Fix regression introduced in rev 1.90
in which the last element of ksyms_symtabs is skipped by mistake.
ksyms(4): Fix race in ksymsread iteration.
TAILQ_NEXT(ksyms_last_snapshot) might change while we are iterating,
but ksyms_last_snapshot itself cannot, so invert the loop structure.
Discussed with rin@.
ksyms(4): Don't skip symbol tables that are soon to be freed, take 2.
They will not actually be freed until /dev/ksyms is closed, so
continued access to them remains kosher.
The previous change was busted because of an off-by-one error in a
previous previous change's iteration over the symtabs; that error has
since been corrected.
ksyms(4): Allow multiple concurrent opens of /dev/ksyms.
First one takes a snapshot; others all agree with the snapshot.
Previously this code path was just broken (could fail horribly if
modules were unloaded after one of the opens is closed), so I just
blocked it off in an earlier commit, but that broke crash(8). So
let's continue allowing multiple opens seeing the same snapshot, but
without the horrible bugs.
diffstat:
sys/kern/kern_ksyms.c | 107 +++++++++++++++++++++++++++++++++++--------------
1 files changed, 75 insertions(+), 32 deletions(-)
diffs (250 lines):
diff -r b0f2dafb5bb2 -r d01257b505af sys/kern/kern_ksyms.c
--- a/sys/kern/kern_ksyms.c Mon Jun 21 16:04:57 2021 +0000
+++ b/sys/kern/kern_ksyms.c Mon Jun 21 16:14:14 2021 +0000
@@ -1,4 +1,4 @@
-/* $NetBSD: kern_ksyms.c,v 1.87.8.1 2020/01/07 11:54:57 martin Exp $ */
+/* $NetBSD: kern_ksyms.c,v 1.87.8.2 2021/06/21 16:14:14 martin Exp $ */
/*-
* Copyright (c) 2008 The NetBSD Foundation, Inc.
@@ -73,7 +73,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: kern_ksyms.c,v 1.87.8.1 2020/01/07 11:54:57 martin Exp $");
+__KERNEL_RCSID(0, "$NetBSD: kern_ksyms.c,v 1.87.8.2 2021/06/21 16:14:14 martin Exp $");
#if defined(_KERNEL) && defined(_KERNEL_OPT)
#include "opt_copy_symtab.h"
@@ -92,6 +92,8 @@
#include <sys/proc.h>
#include <sys/atomic.h>
#include <sys/ksyms.h>
+#include <sys/kernel.h>
+#include <sys/intr.h>
#ifdef DDB
#include <ddb/db_output.h>
@@ -110,7 +112,8 @@
#endif
static int ksyms_maxlen;
-static bool ksyms_isopen;
+static uint64_t ksyms_opencnt;
+static struct ksyms_symtab *ksyms_last_snapshot;
static bool ksyms_initted;
static bool ksyms_loaded;
static kmutex_t ksyms_lock __cacheline_aligned;
@@ -140,7 +143,7 @@
int ksyms_symsz;
int ksyms_strsz;
int ksyms_ctfsz; /* this is not currently used by savecore(8) */
-TAILQ_HEAD(, ksyms_symtab) ksyms_symtabs =
+TAILQ_HEAD(ksyms_symtab_queue, ksyms_symtab) ksyms_symtabs =
TAILQ_HEAD_INITIALIZER(ksyms_symtabs);
static int
@@ -296,6 +299,7 @@
int i, j, n, nglob;
char *str;
int nsyms = symsize / sizeof(Elf_Sym);
+ int s;
/* Sanity check for pre-allocated map table used during startup. */
if ((nmap == ksyms_nmap) && (nsyms >= KSYMS_MAX_ID)) {
@@ -419,7 +423,7 @@
for (new = 0; new < n; new++) {
uint32_t orig = nsym[new].st_size - 1;
uint32_t size = nmap[orig];
-
+
nmap[orig] = new + 1;
/* restore the size */
@@ -428,9 +432,18 @@
}
#endif
- /* ksymsread() is unlocked, so membar. */
- membar_producer();
+ KASSERT(strcmp(name, "netbsd") == 0 || mutex_owned(&ksyms_lock));
+ KASSERT(cold || mutex_owned(&ksyms_lock));
+
+ /*
+ * Ensure ddb never witnesses an inconsistent state of the
+ * queue, unless memory is so corrupt that we crash in
+ * TAILQ_INSERT_TAIL.
+ */
+ s = splhigh();
TAILQ_INSERT_TAIL(&ksyms_symtabs, tab, sd_queue);
+ splx(s);
+
ksyms_sizes_calc();
ksyms_loaded = true;
}
@@ -757,6 +770,8 @@
ksyms_modunload(const char *name)
{
struct ksyms_symtab *st;
+ bool do_free = false;
+ int s;
mutex_enter(&ksyms_lock);
TAILQ_FOREACH(st, &ksyms_symtabs, sd_queue) {
@@ -766,16 +781,26 @@
continue;
st->sd_gone = true;
ksyms_sizes_calc();
- if (!ksyms_isopen) {
+ if (ksyms_opencnt == 0) {
+ /*
+ * Ensure ddb never witnesses an inconsistent
+ * state of the queue, unless memory is so
+ * corrupt that we crash in TAILQ_REMOVE.
+ */
+ s = splhigh();
TAILQ_REMOVE(&ksyms_symtabs, st, sd_queue);
- kmem_free(st->sd_nmap,
- st->sd_nmapsize * sizeof(uint32_t));
- kmem_free(st, sizeof(*st));
+ splx(s);
+ do_free = true;
}
break;
}
mutex_exit(&ksyms_lock);
KASSERT(st != NULL);
+
+ if (do_free) {
+ kmem_free(st->sd_nmap, st->sd_nmapsize * sizeof(uint32_t));
+ kmem_free(st, sizeof(*st));
+ }
}
#ifdef DDB
@@ -854,6 +879,8 @@
struct ksyms_symtab *st;
int i, delta;
+ KASSERT(cold || mutex_owned(&ksyms_lock));
+
ksyms_symsz = ksyms_strsz = 0;
TAILQ_FOREACH(st, &ksyms_symtabs, sd_queue) {
if (__predict_false(st->sd_gone))
@@ -967,10 +994,12 @@
return ENXIO;
/*
- * Create a "snapshot" of the kernel symbol table. Setting
- * ksyms_isopen will prevent symbol tables from being freed.
+ * Create a "snapshot" of the kernel symbol table. Bumping
+ * ksyms_opencnt will prevent symbol tables from being freed.
*/
mutex_enter(&ksyms_lock);
+ if (ksyms_opencnt++)
+ goto out;
ksyms_hdr.kh_shdr[SYMTAB].sh_size = ksyms_symsz;
ksyms_hdr.kh_shdr[SYMTAB].sh_info = ksyms_symsz / sizeof(Elf_Sym);
ksyms_hdr.kh_shdr[STRTAB].sh_offset = ksyms_symsz +
@@ -979,8 +1008,8 @@
ksyms_hdr.kh_shdr[SHCTF].sh_offset = ksyms_strsz +
ksyms_hdr.kh_shdr[STRTAB].sh_offset;
ksyms_hdr.kh_shdr[SHCTF].sh_size = ksyms_ctfsz;
- ksyms_isopen = true;
- mutex_exit(&ksyms_lock);
+ ksyms_last_snapshot = TAILQ_LAST(&ksyms_symtabs, ksyms_symtab_queue);
+out: mutex_exit(&ksyms_lock);
return 0;
}
@@ -989,25 +1018,35 @@
ksymsclose(dev_t dev, int oflags, int devtype, struct lwp *l)
{
struct ksyms_symtab *st, *next;
- bool resize;
+ TAILQ_HEAD(, ksyms_symtab) to_free = TAILQ_HEAD_INITIALIZER(to_free);
+ int s;
/* Discard references to symbol tables. */
mutex_enter(&ksyms_lock);
- ksyms_isopen = false;
- resize = false;
- for (st = TAILQ_FIRST(&ksyms_symtabs); st != NULL; st = next) {
- next = TAILQ_NEXT(st, sd_queue);
+ if (--ksyms_opencnt)
+ goto out;
+ ksyms_last_snapshot = NULL;
+ TAILQ_FOREACH_SAFE(st, &ksyms_symtabs, sd_queue, next) {
if (st->sd_gone) {
+ /*
+ * Ensure ddb never witnesses an inconsistent
+ * state of the queue, unless memory is so
+ * corrupt that we crash in TAILQ_REMOVE.
+ */
+ s = splhigh();
TAILQ_REMOVE(&ksyms_symtabs, st, sd_queue);
- kmem_free(st->sd_nmap,
- st->sd_nmapsize * sizeof(uint32_t));
- kmem_free(st, sizeof(*st));
- resize = true;
+ splx(s);
+ TAILQ_INSERT_TAIL(&to_free, st, sd_queue);
}
}
- if (resize)
+ if (!TAILQ_EMPTY(&to_free))
ksyms_sizes_calc();
- mutex_exit(&ksyms_lock);
+out: mutex_exit(&ksyms_lock);
+
+ TAILQ_FOREACH_SAFE(st, &to_free, sd_queue, next) {
+ kmem_free(st->sd_nmap, st->sd_nmapsize * sizeof(uint32_t));
+ kmem_free(st, sizeof(*st));
+ }
return 0;
}
@@ -1035,9 +1074,9 @@
* Copy out the symbol table.
*/
filepos = sizeof(struct ksyms_hdr);
- TAILQ_FOREACH(st, &ksyms_symtabs, sd_queue) {
- if (__predict_false(st->sd_gone))
- continue;
+ for (st = TAILQ_FIRST(&ksyms_symtabs);
+ ;
+ st = TAILQ_NEXT(st, sd_queue)) {
if (uio->uio_resid == 0)
return 0;
if (uio->uio_offset <= st->sd_symsize + filepos) {
@@ -1048,6 +1087,8 @@
return error;
}
filepos += st->sd_symsize;
+ if (st == ksyms_last_snapshot)
+ break;
}
/*
@@ -1055,9 +1096,9 @@
*/
KASSERT(filepos == sizeof(struct ksyms_hdr) +
ksyms_hdr.kh_shdr[SYMTAB].sh_size);
- TAILQ_FOREACH(st, &ksyms_symtabs, sd_queue) {
- if (__predict_false(st->sd_gone))
- continue;
+ for (st = TAILQ_FIRST(&ksyms_symtabs);
+ ;
+ st = TAILQ_NEXT(st, sd_queue)) {
if (uio->uio_resid == 0)
return 0;
if (uio->uio_offset <= st->sd_strsize + filepos) {
@@ -1068,6 +1109,8 @@
return error;
}
filepos += st->sd_strsize;
+ if (st == ksyms_last_snapshot)
+ break;
}
/*
Home |
Main Index |
Thread Index |
Old Index