[PATCH v10 0/6] alloc_tag and module codetag section fixes

[PATCH v10 0/6] alloc_tag and module codetag section fixes

[Hao Ge]

With profiling toggled off, the overflow check in
reserve_module_tags() did not run, a module could load with more
tags than the page flags can address, and re-enabling profiling
then silently corrupted /proc/allocinfo. On overflow the fix shuts
profiling down, releases the reservation and returns -EAGAIN, and
the codetag section lands as regular module data in the same load,
so the module loads without profiling.

Review of the earlier series by Sashiko turned up two more problems.

One is a race. layout_sections() and move_module() both asked
codetag_needs_module_section() where a codetag section goes, and
mem_profiling_support can change between the two calls, for instance
when another module load overflows the tag index and shuts profiling
down. move_module() then copied the codetag section to offset 0 of
its regular destination and clobbered the first section placed in
that region.

v7 reworks where codetag sections are allocated, on a prototype by
Petr Pavlu [1]. The allocation runs before layout_sections() and the
placement is decided in one step, so nothing re-asks the question
and the race is gone. The retry is gone too, on -EAGAIN the section
is laid out as regular module data right in the same load.

[1] https://lore.kernel.org/all/499bb60c-c6e3-43a3-bd92-95a0567ece5e@suse.com/

Following review feedback from Petr and Suren the series is now
split into six patches. The two alloc_tag fixes from [2] are folded
in as patches 5 and 6 and replace the versions currently in the mm
tree. Sashiko keeps flagging the percpu counter leak [3], patch 5
fixes it, and now the whole set goes through review again.

[2] https://lore.kernel.org/all/20260817062726.106511-1-hao.ge@linux.dev/
[3] https://lore.kernel.org/all/20260908094736.2B1A61F00A3A@smtp.kernel.org/

Patch 1 moves release_module_tags() above reserve_module_tags(),
since the failure paths now have to call it.

Patch 2 cleans up the populate failure path: the reservation is
released, module_tags.size rolled back and the PTEs a failed
vmap_pages_range() left behind unmapped, so a later populate of the
same range is safe. It carries both Fixes tags so it backports
wherever patch 4 goes, which uses its prev_size.

Patch 3 introduces SH_ENTSIZE_STANDALONE to mark sections with a
separate allocation. The percpu section was previously excluded
from the layout by clearing its SHF_ALLOC, which per the ELF spec
says the section occupies memory during execution, and percpu does,
only outside the regular module layout. The mark lives in sh_entsize
now, find_sec(".data..percpu") gives stable results again and
apply_relocations() goes back to testing only SHF_ALLOC. The
section would show up under /sys/module/*/sections/, but it has one
instance per CPU and no single address, and the entry never
existed, so add_sect_attrs() and add_notes_attrs() skip it.

Patch 4 moves the codetag allocation out of move_module() in front
of layout_sections(), so the placement is decided in one step and
the race is gone. On overflow profiling is shut down, the
reservation released, module_tags.size rolled back and -EAGAIN
returned, the section is laid out as regular module data and the
module loads without profiling instead of failing. Any other error
fails the load. The release and the fallback belong together,
without the release rmmod hits the stale entry and panics.

Patch 5 skips the percpu counter allocation when profiling is off.
After the shutdown modules load their codetag section as regular
data, load_module() still allocated counters for every tag and
release_module_tags() cannot find them on unload, so they leaked
(Suggested by Suren).

Patch 6 defers the /proc/allocinfo removal to a workqueue.
shutdown_mem_profiling() runs under mod_lock, and the synchronous
remove_proc_entry() deadlocks with a reader taking mod_lock for
read in allocinfo_start(). The file is also created at the end of
alloc_tag_init(), a leftover file after a failed init would panic
its readers (Found by Sashiko).

Tested on an x86_64 virtual machine:

Booted without sysctl.vm.mem_profiling=1,compressed:
# cat /proc/allocinfo is fine

Booted with sysctl.vm.mem_profiling=1,compressed:
# cat /proc/allocinfo is fine
# insmod overflow_tag.ko
# dmesg
  With module overflow_tag there are too many tags to fit in 13 page
  flag bits. Memory allocation profiling is disabled!
# rmmod overflow_tag
The module loads without profiling and unloads cleanly.

Also ran continuous LTP stress for a few days, nothing abnormal
so far.

Changes in v10:
- fold in the two alloc_tag fixes from [2] as patches 5 and 6, they
  replace the versions in the mm tree and fix the percpu leak
  Sashiko keeps flagging [3]
- create /proc/allocinfo at the end of alloc_tag_init(), a
  leftover file after a failed init would panic its readers
  (Found by Sashiko)
- count note sections with sect_visible() in add_notes_attrs() too,
  the count has to match the fill loop

Changes in v9:
- do not export .data..percpu under /sys/module/*/sections/ (Petr
  Pavlu).
- move the populate failure cleanup in front of the rework, v8 patch
  4 is patch 2 now. It declares prev_size itself, which the overflow
  path of the rework also uses, so it carries both Fixes tags and the
  two patches backport together

Changes in v8:
- roll back module_tags.size when the reservation is released, so a
  concurrent load which already passed needs_section_mem() does not
  skip populate for the freed gap (Sashiko)
- unmap the PTEs a failed vmap_pages_range() installed, a retry to
  populate the same range would BUG on them
- zero the separately allocated codetag memory for an SHT_NOBITS
  section, the tag area pages are not zeroed on allocation (Sashiko)
- .data..percpu is exported under /sys/module/*/sections/ with the
  boot CPU instance of the per-cpu area, and the interface is
  documented in the ABI docs (Petr Pavlu)
- keep a comment in apply_relocations() on how .data..percpu is
  relocated (Petr Pavlu)
- replace the "Based-on-a-patch-by:" tag with an in-body
  attribution and a numbered Link: (Andrew Morton)

Changes in v7:
- split the rework following review feedback (Petr Pavlu, Suren
  Baghdasaryan)
- new patch 2 marks separately allocated sections with
  SH_ENTSIZE_STANDALONE instead of clearing SHF_ALLOC
  (suggested by Petr Pavlu)
- split the populate failure release into its own patch (suggested
  by Suren Baghdasaryan)

Changes in v6:
- rework on Petr's prototype and allocate codetag sections before
  layout_sections(), the retry and its state resets are gone
- fix the layout_sections()/move_module() race (Found by Sashiko)
- release the reservation on populate failure as well (Found by
  Sashiko)
- only -EAGAIN keeps the fallback, other errors fail the load

Changes in v5:
- add Fixes: and Cc: stable to patch 1/2 as well, since 2/2 does not
  compile without it (Andrew Morton)
- restore frob-adjusted mem[type].size on retry instead of zeroing,
  as s390 and parisc add GOT/PLT space there in
  module_frob_arch_sections() (Reported by Sashiko)
- drop the load_module() mem_profiling_support check; the percpu
  counter leak is pre-existing and orthogonal to this fix

Changes in v4:
- add a new patch (1/2) to move release_module_tags() above
  reserve_module_tags(); the overflow fix is 2/2
- release the reservation on the -EAGAIN path
- return -EAGAIN instead of -ENOMEM so the module can still load
  without profiling (Suren)
- reset sh_addr, mem[type].size and sym/str SHF_ALLOC before retry
- skip percpu counters in load_module() when profiling is off

Changes in v3:
- use pr_warn_once() instead of pr_warn()
- return -ENOMEM instead of -ENOSPC (Suren)
- expand the commit message to describe the /proc/allocinfo impact
  (Andrew)

Changes in v2:
- return an error after shutdown_mem_profiling() to skip
  vm_module_tags_populate()

v1: https://lore.kernel.org/all/20260804064408.105033-1-hao.ge@linux.dev/
v2: https://lore.kernel.org/all/20260804122038.190270-1-hao.ge@linux.dev/
v3: https://lore.kernel.org/all/20260805090633.141001-1-hao.ge@linux.dev/
v4: https://lore.kernel.org/all/20260810093955.153015-1-hao.ge@linux.dev/
v5: https://lore.kernel.org/all/20260812054105.102637-1-hao.ge@linux.dev/
v6: https://lore.kernel.org/all/20260831072104.120197-1-hao.ge@linux.dev/
v7: https://lore.kernel.org/all/20260902081802.146145-1-hao.ge@linux.dev/
v8: https://lore.kernel.org/all/20260907062414.106873-1-hao.ge@linux.dev/
v9: https://lore.kernel.org/all/20260908092412.115953-1-hao.ge@linux.dev/

Hao Ge (6):
  alloc_tag: move release_module_tags() above reserve_module_tags()
  alloc_tag: clean up the populate failure path
  module: introduce SH_ENTSIZE_STANDALONE for separately allocated
    sections
  module: allocate codetag sections before the regular module layout
  alloc_tag: skip percpu counter allocation when profiling is disabled
  alloc_tag: Defer /proc/allocinfo removal to a workqueue

 include/linux/module.h   |   2 +
 kernel/module/internal.h |   8 +++
 kernel/module/kallsyms.c |  13 +---
 kernel/module/main.c     | 133 +++++++++++++++++++-----------------
 kernel/module/sysfs.c    |  17 +++--
 lib/codetag.c            |  10 ++-
 mm/alloc_tag.c           | 141 +++++++++++++++++++++++----------------
 7 files changed, 188 insertions(+), 136 deletions(-)

-- 
2.25.1

[PATCH v10 1/6] alloc_tag: move release_module_tags() above reserve_module_tags()

[Hao Ge]

release_module_tags() is a cleanup helper. reserve_module_tags() can
also fail after storing the reservation in the maple tree, in which
case it should call release_module_tags() to undo it. Move the helper
above reserve_module_tags() so no forward declaration is needed.

No functional change.

Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
Fixes: 0f9b685626da ("alloc_tag: populate memory for module tags as needed")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Cc: stable@vger.kernel.org
Acked-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/alloc_tag.c | 92 +++++++++++++++++++++++++-------------------------
 1 file changed, 46 insertions(+), 46 deletions(-)

diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
index b33410310477..2070e682fe10 100644
--- a/mm/alloc_tag.c
+++ b/mm/alloc_tag.c
@@ -835,6 +835,52 @@ static int vm_module_tags_populate(void)
 	return 0;
 }
 
+static void release_module_tags(struct module *mod, bool used)
+{
+	MA_STATE(mas, &mod_area_mt, module_tags.size, module_tags.size);
+	struct alloc_tag *start_tag;
+	struct alloc_tag *end_tag;
+	struct module *val;
+
+	mas_lock(&mas);
+	mas_for_each_rev(&mas, val, 0)
+		if (val == mod)
+			break;
+
+	if (!val) /* module not found */
+		goto out;
+
+	if (!used)
+		goto release_area;
+
+	start_tag = (struct alloc_tag *)(module_tags.start_addr + mas.index);
+	end_tag = (struct alloc_tag *)(module_tags.start_addr + mas.last);
+	if (!clean_unused_counters(start_tag, end_tag)) {
+		struct alloc_tag *tag;
+
+		for (tag = start_tag; tag <= end_tag; tag++) {
+			struct alloc_tag_counters counter;
+
+			if (!tag->counters)
+				continue;
+
+			counter = alloc_tag_read(tag);
+			pr_info("%s:%u module %s func:%s has %llu allocated at module unload\n",
+				tag->ct.filename, tag->ct.lineno, tag->ct.modname,
+				tag->ct.function, counter.bytes);
+		}
+	} else {
+		used = false;
+	}
+release_area:
+	mas_store(&mas, used ? &unloaded_mod : NULL);
+	val = mas_prev_range(&mas, 0);
+	if (val == &prepend_mod)
+		mas_store(&mas, NULL);
+out:
+	mas_unlock(&mas);
+}
+
 static void *reserve_module_tags(struct module *mod, unsigned long size,
 				 unsigned int prepend, unsigned long align)
 {
@@ -922,52 +968,6 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
 	return (struct alloc_tag *)(module_tags.start_addr + offset);
 }
 
-static void release_module_tags(struct module *mod, bool used)
-{
-	MA_STATE(mas, &mod_area_mt, module_tags.size, module_tags.size);
-	struct alloc_tag *start_tag;
-	struct alloc_tag *end_tag;
-	struct module *val;
-
-	mas_lock(&mas);
-	mas_for_each_rev(&mas, val, 0)
-		if (val == mod)
-			break;
-
-	if (!val) /* module not found */
-		goto out;
-
-	if (!used)
-		goto release_area;
-
-	start_tag = (struct alloc_tag *)(module_tags.start_addr + mas.index);
-	end_tag = (struct alloc_tag *)(module_tags.start_addr + mas.last);
-	if (!clean_unused_counters(start_tag, end_tag)) {
-		struct alloc_tag *tag;
-
-		for (tag = start_tag; tag <= end_tag; tag++) {
-			struct alloc_tag_counters counter;
-
-			if (!tag->counters)
-				continue;
-
-			counter = alloc_tag_read(tag);
-			pr_info("%s:%u module %s func:%s has %llu allocated at module unload\n",
-				tag->ct.filename, tag->ct.lineno, tag->ct.modname,
-				tag->ct.function, counter.bytes);
-		}
-	} else {
-		used = false;
-	}
-release_area:
-	mas_store(&mas, used ? &unloaded_mod : NULL);
-	val = mas_prev_range(&mas, 0);
-	if (val == &prepend_mod)
-		mas_store(&mas, NULL);
-out:
-	mas_unlock(&mas);
-}
-
 static int load_module(struct module *mod, struct codetag *start, struct codetag *stop)
 {
 	/* Allocate module alloc_tag percpu counters */
-- 
2.25.1

[PATCH v10 2/6] alloc_tag: clean up the populate failure path

[Hao Ge]

The reservation is already stored in the maple tree when
vm_module_tags_populate() fails. A failed load never unloads the
module, so nothing releases the entry. Release it and roll
module_tags.size back. Without the rollback a concurrent load that
already passed needs_section_mem() can reuse the freed gap, skip
vm_module_tags_populate() and write to unmapped memory.

vmap_pages_range() may have installed some PTEs before failing. A
retry to populate the same range would BUG on them, so undo them,
but only if vmap actually ran.

Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
Fixes: 0f9b685626da ("alloc_tag: populate memory for module tags as needed")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Cc: stable@vger.kernel.org
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/alloc_tag.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
index 2070e682fe10..95ddf5b743d0 100644
--- a/mm/alloc_tag.c
+++ b/mm/alloc_tag.c
@@ -804,6 +804,13 @@ static int vm_module_tags_populate(void)
 				     next_page, PAGE_SHIFT) < 0) {
 			release_pages_arg arg = { .pages = next_page };
 
+			/*
+			 * vmap_pages_range() only runs once all pages were
+			 * allocated, and it may have installed some mappings
+			 * before failing. Undo them.
+			 */
+			if (nr == more_pages)
+				vunmap_range(phys_end, phys_end + (nr << PAGE_SHIFT));
 			/* Clean up and error out */
 			release_pages(arg, nr);
 			return -ENOMEM;
@@ -947,6 +954,7 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
 		return ret;
 
 	if (module_tags.size < offset + size) {
+		unsigned long prev_size = module_tags.size;
 		int grow_res;
 
 		module_tags.size = offset + size;
@@ -961,6 +969,8 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
 			shutdown_mem_profiling(true);
 			pr_err("Failed to allocate memory for allocation tags in the module %s. Memory allocation profiling is disabled!\n",
 			       mod->name);
+			release_module_tags(mod, false);
+			module_tags.size = prev_size;
 			return ERR_PTR(grow_res);
 		}
 	}
-- 
2.25.1

[PATCH v10 3/6] module: introduce SH_ENTSIZE_STANDALONE for separately allocated sections

[Hao Ge]

SHF_ALLOC means, per the ELF spec, that a section occupies memory
during process execution. Some module sections occupy memory
outside the regular module layout, for example the percpu section
with its per-CPU allocations. The loader currently excludes such
a section from the layout by clearing its SHF_ALLOC, which
overloads the flag with a loader-internal meaning.
apply_relocations() needs a special case for the section, and
find_sec(".data..percpu") returns different results before and
after layout_and_allocate().

Introduce SH_ENTSIZE_STANDALONE to mark sections with a separate
allocation. The percpu section is its first user. layout_sections()
and move_module() skip marked sections, and apply_relocations() goes
back to testing only SHF_ALLOC. Based on a patch by Petr Pavlu [1].

.data..percpu keeps SHF_ALLOC, so it would now show up under
/sys/module/*/sections/. The section has one instance per CPU and no
single address to report, and the entry never existed before, so
skip it in add_sect_attrs(). add_notes_attrs() indexes its attrs[]
array and skips it too. No functional change otherwise.

Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Link: https://lore.kernel.org/all/499bb60c-c6e3-43a3-bd92-95a0567ece5e@suse.com/ [1]
Suggested-by: Petr Pavlu <petr.pavlu@suse.com>
Reviewed-by: Petr Pavlu <petr.pavlu@suse.com>
Cc: stable@vger.kernel.org
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 include/linux/module.h   |  2 ++
 kernel/module/internal.h |  8 ++++++++
 kernel/module/kallsyms.c | 13 +++----------
 kernel/module/main.c     | 32 +++++++++++++++++---------------
 kernel/module/sysfs.c    | 17 +++++++++++++----
 5 files changed, 43 insertions(+), 29 deletions(-)

diff --git a/include/linux/module.h b/include/linux/module.h
index 96cc98568eea..0c6f32ddcbf2 100644
--- a/include/linux/module.h
+++ b/include/linux/module.h
@@ -325,6 +325,8 @@ enum mod_mem_type {
 	MOD_INIT_RODATA,
 
 	MOD_MEM_NUM_TYPES,
+
+	MOD_STANDALONE = -2,
 	MOD_INVALID = -1,
 };
 
diff --git a/kernel/module/internal.h b/kernel/module/internal.h
index 061161cc79d9..4c738074a27b 100644
--- a/kernel/module/internal.h
+++ b/kernel/module/internal.h
@@ -29,6 +29,14 @@
 #define SH_ENTSIZE_TYPE_MASK	((1UL << SH_ENTSIZE_TYPE_BITS) - 1)
 #define SH_ENTSIZE_OFFSET_MASK	((1UL << (BITS_PER_LONG - SH_ENTSIZE_TYPE_BITS)) - 1)
 
+/*
+ * Marker for sections with a separate allocation, which are not placed
+ * into mod->mem[].
+ */
+#define SH_ENTSIZE_STANDALONE					\
+	(((unsigned long)MOD_STANDALONE & SH_ENTSIZE_TYPE_MASK)	\
+	 << SH_ENTSIZE_TYPE_SHIFT)
+
 /* Maximum number of characters written by module_flags() */
 #define MODULE_FLAGS_BUF_SIZE (TAINT_FLAGS_COUNT + 4)
 
diff --git a/kernel/module/kallsyms.c b/kernel/module/kallsyms.c
index f23126d804b2..0dca6d40160e 100644
--- a/kernel/module/kallsyms.c
+++ b/kernel/module/kallsyms.c
@@ -76,7 +76,7 @@ static char elf_type(const Elf_Sym *sym, const struct load_info *info)
 }
 
 static bool is_core_symbol(const Elf_Sym *src, const Elf_Shdr *sechdrs,
-			   unsigned int shnum, unsigned int pcpundx)
+			   unsigned int shnum)
 {
 	const Elf_Shdr *sec;
 	enum mod_mem_type type;
@@ -86,11 +86,6 @@ static bool is_core_symbol(const Elf_Sym *src, const Elf_Shdr *sechdrs,
 	    !src->st_name)
 		return false;
 
-#ifdef CONFIG_KALLSYMS_ALL
-	if (src->st_shndx == pcpundx)
-		return true;
-#endif
-
 	sec = sechdrs + src->st_shndx;
 	type = sec->sh_entsize >> SH_ENTSIZE_TYPE_SHIFT;
 	if (!(sec->sh_flags & SHF_ALLOC)
@@ -131,8 +126,7 @@ void layout_symtab(struct module *mod, struct load_info *info)
 	/* Compute total space required for the core symbols' strtab. */
 	for (ndst = i = 0; i < nsrc; i++) {
 		if (i == 0 || is_livepatch_module(mod) ||
-		    is_core_symbol(src + i, info->sechdrs, info->hdr->e_shnum,
-				   info->index.pcpu)) {
+		    is_core_symbol(src + i, info->sechdrs, info->hdr->e_shnum)) {
 			strtab_size += strlen(&info->strtab[src[i].st_name]) + 1;
 			ndst++;
 		}
@@ -199,8 +193,7 @@ void add_kallsyms(struct module *mod, const struct load_info *info)
 	for (ndst = i = 0; i < kallsyms->num_symtab; i++) {
 		kallsyms->typetab[i] = elf_type(src + i, info);
 		if (i == 0 || is_livepatch_module(mod) ||
-		    is_core_symbol(src + i, info->sechdrs, info->hdr->e_shnum,
-				   info->index.pcpu)) {
+		    is_core_symbol(src + i, info->sechdrs, info->hdr->e_shnum)) {
 			ssize_t ret;
 
 			mod->core_kallsyms.typetab[ndst] =
diff --git a/kernel/module/main.c b/kernel/module/main.c
index c1b34dc1e89a..ae2678ac7840 100644
--- a/kernel/module/main.c
+++ b/kernel/module/main.c
@@ -1619,12 +1619,13 @@ static int apply_relocations(struct module *mod, const struct load_info *info)
 
 		/*
 		 * Don't bother with non-allocated sections.
-		 * An exception is the percpu section, which has separate allocations
-		 * for individual CPUs. We relocate the percpu section in the initial
-		 * ELF template and subsequently copy it to the per-CPU destinations.
+		 *
+		 * Note that .data..percpu has separate allocations for
+		 * individual CPUs. We relocate the section in the
+		 * initial ELF template and subsequently copy it to the
+		 * per-CPU destinations.
 		 */
-		if (!(info->sechdrs[infosec].sh_flags & SHF_ALLOC) &&
-		    (!infosec || infosec != info->index.pcpu))
+		if (!(info->sechdrs[infosec].sh_flags & SHF_ALLOC))
 			continue;
 
 		if (info->sechdrs[i].sh_flags & SHF_RELA_LIVEPATCH)
@@ -1715,7 +1716,7 @@ static void __layout_sections(struct module *mod, struct load_info *info, bool i
 
 			if ((s->sh_flags & masks[m][0]) != masks[m][0]
 			    || (s->sh_flags & masks[m][1])
-			    || s->sh_entsize != ~0UL
+			    || s->sh_entsize != ~0UL /* offset or standalone */
 			    || is_init != module_init_layout_section(sname))
 				continue;
 
@@ -1745,16 +1746,10 @@ static void __layout_sections(struct module *mod, struct load_info *info, bool i
 /*
  * Lay out the SHF_ALLOC sections in a way not dissimilar to how ld
  * might -- code, read-only data, read-write data, small data.  Tally
- * sizes, and place the offsets into sh_entsize fields: high bit means it
- * belongs in init.
+ * sizes, and place the offsets into sh_entsize fields.
  */
 static void layout_sections(struct module *mod, struct load_info *info)
 {
-	unsigned int i;
-
-	for (i = 0; i < info->hdr->e_shnum; i++)
-		info->sechdrs[i].sh_entsize = ~0UL;
-
 	pr_debug("Core section allocation order for %s:\n", mod->name);
 	__layout_sections(mod, info, false);
 
@@ -2822,7 +2817,8 @@ static int move_module(struct module *mod, struct load_info *info)
 		Elf_Shdr *shdr = &info->sechdrs[i];
 		const char *sname;
 
-		if (!(shdr->sh_flags & SHF_ALLOC))
+		if (!(shdr->sh_flags & SHF_ALLOC)
+		    || shdr->sh_entsize == SH_ENTSIZE_STANDALONE)
 			continue;
 
 		sname = info->secstrings + shdr->sh_name;
@@ -2954,6 +2950,7 @@ core_param(module_blacklist, module_blacklist, charp, 0400);
 static struct module *layout_and_allocate(struct load_info *info, int flags)
 {
 	struct module *mod;
+	unsigned int i;
 	int err;
 
 	/* Allow arches to frob section contents and sizes.  */
@@ -2967,8 +2964,13 @@ static struct module *layout_and_allocate(struct load_info *info, int flags)
 	if (err < 0)
 		return ERR_PTR(err);
 
+	/* Repurpose sh_entsize to track where each section is allocated. */
+	for (i = 0; i < info->hdr->e_shnum; i++)
+		info->sechdrs[i].sh_entsize = ~0UL;
+
 	/* We will do a special allocation for per-cpu sections later. */
-	info->sechdrs[info->index.pcpu].sh_flags &= ~(unsigned long)SHF_ALLOC;
+	if (info->index.pcpu)
+		info->sechdrs[info->index.pcpu].sh_entsize = SH_ENTSIZE_STANDALONE;
 
 	/*
 	 * Mark relevant sections as SHF_RO_AFTER_INIT so layout_sections() can
diff --git a/kernel/module/sysfs.c b/kernel/module/sysfs.c
index 01c65d608873..15c6baeb39c4 100644
--- a/kernel/module/sysfs.c
+++ b/kernel/module/sysfs.c
@@ -62,6 +62,15 @@ static void free_sect_attrs(struct module_sect_attrs *sect_attrs)
 	kfree(sect_attrs);
 }
 
+/*
+ * .data..percpu has a separate allocation per CPU and no single
+ * address to report.
+ */
+static bool sect_visible(const struct load_info *info, unsigned int i)
+{
+	return !sect_empty(&info->sechdrs[i]) && i != info->index.pcpu;
+}
+
 static int add_sect_attrs(struct module *mod, const struct load_info *info)
 {
 	struct module_sect_attrs *sect_attrs;
@@ -72,7 +81,7 @@ static int add_sect_attrs(struct module *mod, const struct load_info *info)
 
 	/* Count loaded sections and allocate structures */
 	for (i = 0; i < info->hdr->e_shnum; i++)
-		if (!sect_empty(&info->sechdrs[i]))
+		if (sect_visible(info, i))
 			nloaded++;
 	sect_attrs = kzalloc_flex(*sect_attrs, attrs, nloaded);
 	if (!sect_attrs)
@@ -92,7 +101,7 @@ static int add_sect_attrs(struct module *mod, const struct load_info *info)
 	for (i = 0; i < info->hdr->e_shnum; i++) {
 		Elf_Shdr *sec = &info->sechdrs[i];
 
-		if (sect_empty(sec))
+		if (!sect_visible(info, i))
 			continue;
 		sysfs_bin_attr_init(sattr);
 		sattr->attr.name =
@@ -159,7 +168,7 @@ static int add_notes_attrs(struct module *mod, const struct load_info *info)
 	/* Count notes sections and allocate structures.  */
 	notes = 0;
 	for (i = 0; i < info->hdr->e_shnum; i++)
-		if (!sect_empty(&info->sechdrs[i]) &&
+		if (sect_visible(info, i) &&
 		    info->sechdrs[i].sh_type == SHT_NOTE)
 			++notes;
 
@@ -181,7 +190,7 @@ static int add_notes_attrs(struct module *mod, const struct load_info *info)
 
 	nattr = &notes_attrs->attrs[0];
 	for (loaded = i = 0; i < info->hdr->e_shnum; ++i) {
-		if (sect_empty(&info->sechdrs[i]))
+		if (!sect_visible(info, i))
 			continue;
 		if (info->sechdrs[i].sh_type == SHT_NOTE) {
 			sysfs_bin_attr_init(nattr);
-- 
2.25.1

[PATCH v10 4/6] module: allocate codetag sections before the regular module layout

[Hao Ge]

Whether a codetag section goes to the codetag region is decided by
layout_sections() and asked again in move_module(). A concurrent
load can shut profiling down in between, and move_module() then
copies the section to offset 0 of its regular destination,
overwriting whatever is there.

Decide and allocate in one pass, before the layout. Allocation
errors fail the load. On a tag area overflow profiling is already
disabled, so -EAGAIN makes the section fall back to regular module
data and the module still loads. The reservation is released and
module_tags.size rolled back, so a concurrent load which already
passed needs_section_mem() does not skip vm_module_tags_populate().

An SHT_NOBITS codetag section is zeroed explicitly, the tag area
pages are not zeroed on allocation.

When profiling was toggled off the overflow check did not run, a
module could load with more tags than the page flags can address,
and re-enabling profiling then silently corrupted /proc/allocinfo.
The check no longer depends on mem_alloc_profiling_enabled().

Based on a patch by Petr Pavlu [1].

Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Link: https://lore.kernel.org/all/499bb60c-c6e3-43a3-bd92-95a0567ece5e@suse.com/ [1]
Reviewed-by: Petr Pavlu <petr.pavlu@suse.com>
Cc: stable@vger.kernel.org
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 kernel/module/main.c | 101 +++++++++++++++++++++++--------------------
 mm/alloc_tag.c       |   9 ++--
 2 files changed, 60 insertions(+), 50 deletions(-)

diff --git a/kernel/module/main.c b/kernel/module/main.c
index ae2678ac7840..b7ebcc40bdda 100644
--- a/kernel/module/main.c
+++ b/kernel/module/main.c
@@ -1723,20 +1723,6 @@ static void __layout_sections(struct module *mod, struct load_info *info, bool i
 			if (WARN_ON_ONCE(type == MOD_INVALID))
 				continue;
 
-			/*
-			 * Do not allocate codetag memory as we load it into
-			 * preallocated contiguous memory.
-			 */
-			if (codetag_needs_module_section(mod, sname, s->sh_size)) {
-				/*
-				 * s->sh_entsize won't be used but populate the
-				 * type field to avoid confusion.
-				 */
-				s->sh_entsize = ((unsigned long)(type) & SH_ENTSIZE_TYPE_MASK)
-						<< SH_ENTSIZE_TYPE_SHIFT;
-				continue;
-			}
-
 			s->sh_entsize = module_get_offset_and_type(mod, type, s, i);
 			pr_debug("\t%s\n", sname);
 		}
@@ -2795,7 +2781,6 @@ static int move_module(struct module *mod, struct load_info *info)
 {
 	int i, ret;
 	enum mod_mem_type t = MOD_MEM_NUM_TYPES;
-	bool codetag_section_found = false;
 
 	for_each_mod_mem_type(type) {
 		if (!mod->mem[type].size) {
@@ -2815,35 +2800,13 @@ static int move_module(struct module *mod, struct load_info *info)
 	for (i = 0; i < info->hdr->e_shnum; i++) {
 		void *dest;
 		Elf_Shdr *shdr = &info->sechdrs[i];
-		const char *sname;
 
 		if (!(shdr->sh_flags & SHF_ALLOC)
 		    || shdr->sh_entsize == SH_ENTSIZE_STANDALONE)
 			continue;
 
-		sname = info->secstrings + shdr->sh_name;
-		/*
-		 * Load codetag sections separately as they might still be used
-		 * after module unload.
-		 */
-		if (codetag_needs_module_section(mod, sname, shdr->sh_size)) {
-			dest = codetag_alloc_module_section(mod, sname, shdr->sh_size,
-					arch_mod_section_prepend(mod, i), shdr->sh_addralign);
-			if (WARN_ON(!dest)) {
-				ret = -EINVAL;
-				goto out_err;
-			}
-			if (IS_ERR(dest)) {
-				ret = PTR_ERR(dest);
-				goto out_err;
-			}
-			codetag_section_found = true;
-		} else {
-			enum mod_mem_type type = shdr->sh_entsize >> SH_ENTSIZE_TYPE_SHIFT;
-			unsigned long offset = shdr->sh_entsize & SH_ENTSIZE_OFFSET_MASK;
-
-			dest = mod->mem[type].base + offset;
-		}
+		dest = mod->mem[shdr->sh_entsize >> SH_ENTSIZE_TYPE_SHIFT].base +
+		       (shdr->sh_entsize & SH_ENTSIZE_OFFSET_MASK);
 
 		if (shdr->sh_type != SHT_NOBITS) {
 			/*
@@ -2875,8 +2838,6 @@ static int move_module(struct module *mod, struct load_info *info)
 	module_memory_restore_rox(mod);
 	while (t--)
 		module_memory_free(mod, t);
-	if (codetag_section_found)
-		codetag_free_module_sections(mod);
 
 	return ret;
 }
@@ -2947,6 +2908,49 @@ static bool blacklisted(const char *module_name)
 }
 core_param(module_blacklist, module_blacklist, charp, 0400);
 
+/*
+ * Allocate codetag sections separately. They are loaded into preallocated
+ * contiguous memory because they may still be used after the module is
+ * unloaded.
+ *
+ * If the separate allocation overflows, allocate the section normally
+ * so that the module can still be loaded.
+ */
+static int allocate_codetag_sections(struct load_info *info)
+{
+	for (unsigned int i = 1; i < info->hdr->e_shnum; i++) {
+		Elf_Shdr *shdr = &info->sechdrs[i];
+		const char *sname = info->secstrings + shdr->sh_name;
+		void *dest;
+
+		if (!codetag_needs_module_section(info->mod, sname, shdr->sh_size))
+			continue;
+
+		dest = codetag_alloc_module_section(info->mod, sname, shdr->sh_size,
+				arch_mod_section_prepend(info->mod, i), shdr->sh_addralign);
+		if (WARN_ON(!dest)) {
+			codetag_free_module_sections(info->mod);
+			return -EINVAL;
+		}
+		if (dest == ERR_PTR(-EAGAIN))
+			/* Allocate the section as a regular section. */
+			continue;
+		if (IS_ERR(dest)) {
+			codetag_free_module_sections(info->mod);
+			return PTR_ERR(dest);
+		}
+
+		if (shdr->sh_type != SHT_NOBITS)
+			memcpy(dest, (void *)shdr->sh_addr, shdr->sh_size);
+		else
+			memset(dest, 0, shdr->sh_size);
+		shdr->sh_addr = (unsigned long)dest;
+		shdr->sh_entsize = SH_ENTSIZE_STANDALONE;
+	}
+
+	return 0;
+}
+
 static struct module *layout_and_allocate(struct load_info *info, int flags)
 {
 	struct module *mod;
@@ -2979,18 +2983,21 @@ static struct module *layout_and_allocate(struct load_info *info, int flags)
 	 */
 	module_mark_ro_after_init(info->hdr, info->sechdrs, info->secstrings);
 
-	/*
-	 * Determine total sizes, and put offsets in sh_entsize.  For now
-	 * this is done generically; there doesn't appear to be any
-	 * special cases for the architectures.
-	 */
+	/* Allow codetag sections to be allocated separately first. */
+	err = allocate_codetag_sections(info);
+	if (err)
+		return ERR_PTR(err);
+
+	/* Determine total sizes and put offsets in sh_entsize. */
 	layout_sections(info->mod, info);
 	layout_symtab(info->mod, info);
 
 	/* Allocate and move to the final place */
 	err = move_module(info->mod, info);
-	if (err)
+	if (err) {
+		codetag_free_module_sections(info->mod);
 		return ERR_PTR(err);
+	}
 
 	/* Module has been copied to its final place now: return it. */
 	mod = (void *)info->sechdrs[info->index.mod].sh_addr;
diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
index 95ddf5b743d0..5836803898ad 100644
--- a/mm/alloc_tag.c
+++ b/mm/alloc_tag.c
@@ -958,10 +958,13 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
 		int grow_res;
 
 		module_tags.size = offset + size;
-		if (mem_alloc_profiling_enabled() && !tags_addressable()) {
+		if (!tags_addressable()) {
 			shutdown_mem_profiling(true);
-			pr_warn("With module %s there are too many tags to fit in %d page flag bits. Memory allocation profiling is disabled!\n",
-				mod->name, NR_UNUSED_PAGEFLAG_BITS);
+			pr_warn_once("With module %s there are too many tags to fit in %d page flag bits. Memory allocation profiling is disabled!\n",
+				     mod->name, NR_UNUSED_PAGEFLAG_BITS);
+			release_module_tags(mod, false);
+			module_tags.size = prev_size;
+			return ERR_PTR(-EAGAIN);
 		}
 
 		grow_res = vm_module_tags_populate();
-- 
2.25.1

[PATCH v10 5/6] alloc_tag: skip percpu counter allocation when profiling is disabled

[Hao Ge]

After shutdown_mem_profiling() clears mem_profiling_support,
needs_section_mem() returns false, so later modules have their codetag
section placed as regular data and never enter the alloc_tag maple tree.
codetag_load_module() still called load_module(), which allocated a percpu
counter for every tag; release_module_tags() could not find these modules
on unload, so the counters leaked.

Return -EOPNOTSUPP from load_module() when profiling is off:
codetag_module_init() drops the module's cmod, no counters are allocated
and the module loads without its tags. codetag_unload_module() now always
calls free_section_mem(), since a module whose module_load() returned
-EOPNOTSUPP is not in the idr but may still hold a reserved section.

Reported-by: Sashiko <sashiko-bot@kernel.org>
Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
Cc: stable@vger.kernel.org
Suggested-by: Suren Baghdasaryan <surenb@google.com>
Acked-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 lib/codetag.c  | 10 ++++++++--
 mm/alloc_tag.c |  4 ++++
 2 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/lib/codetag.c b/lib/codetag.c
index a9cda4c962a3..a0b600720afc 100644
--- a/lib/codetag.c
+++ b/lib/codetag.c
@@ -240,7 +240,9 @@ static int codetag_module_init(struct codetag_type *cttype, struct module *mod)
 
 	if (err < 0) {
 		kfree(cmod);
-		return err;
+		/* -EOPNOTSUPP means we can load the module without its tag. */
+		if (err != -EOPNOTSUPP)
+			return err;
 	}
 
 	return 0;
@@ -388,7 +390,11 @@ void codetag_unload_module(struct module *mod)
 			++cttype->content_id;
 		}
 		up_write(&cttype->mod_lock);
-		if (found && cttype->desc.free_section_mem)
+		/*
+		 * A module whose module_load() returned -EOPNOTSUPP is not
+		 * in the idr but may still hold reserved section memory.
+		 */
+		if (cttype->desc.free_section_mem)
 			cttype->desc.free_section_mem(mod, true);
 	}
 	mutex_unlock(&codetag_lock);
diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
index 5836803898ad..1ca0409b492b 100644
--- a/mm/alloc_tag.c
+++ b/mm/alloc_tag.c
@@ -988,6 +988,10 @@ static int load_module(struct module *mod, struct codetag *start, struct codetag
 	struct alloc_tag *stop_tag;
 	struct alloc_tag *tag;
 
+	/* Profiling disabled: load the module without its tags. */
+	if (!mem_profiling_support)
+		return -EOPNOTSUPP;
+
 	/* percpu counters for core allocations are already statically allocated */
 	if (!mod)
 		return 0;
-- 
2.25.1

[PATCH v10 6/6] alloc_tag: Defer /proc/allocinfo removal to a workqueue

[Hao Ge]

shutdown_mem_profiling() calls remove_proc_entry() from
reserve_module_tags(), which runs under mod_lock held for write.
remove_proc_entry() waits for readers, and a reader takes mod_lock for
read in allocinfo_start():

  CPU0 (insmod)                      CPU1 (read /proc/allocinfo)
  ----------------                   ----------------------------
  reserve_module_tags()
    down_write(&mod_lock)  [held]
                                     use_pde()            [in_use++]
                                     allocinfo_start()
                                       down_read(&mod_lock)  <- blocks
    shutdown_mem_profiling()
      remove_proc_entry()
        wait for in_use == 0         <- blocks

Move remove_proc_entry() to a workqueue.

The file creation is moved to the end of alloc_tag_init() as well.
If alloc_tag_init() fails with alloc_tag_cttype still NULL or an
error pointer, a concurrent reader of the leftover file would
dereference it in allocinfo_start() and panic.

Reported-by: Sashiko <sashiko-bot@kernel.org>
Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
Cc: stable@vger.kernel.org
Signed-off-by: Hao Ge <hao.ge@linux.dev>
---
 mm/alloc_tag.c | 26 +++++++++++++++++---------
 1 file changed, 17 insertions(+), 9 deletions(-)

diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
index 1ca0409b492b..cfa0fc84b68f 100644
--- a/mm/alloc_tag.c
+++ b/mm/alloc_tag.c
@@ -15,6 +15,7 @@
 #include <linux/seq_file.h>
 #include <linux/string_choices.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <linux/kmemleak.h>
 #include <uapi/linux/alloc_tag.h>
 
@@ -591,6 +592,13 @@ void pgalloc_tag_swap(struct folio *new, struct folio *old)
 	put_page_tag_ref(handle_new);
 }
 
+static void remove_allocinfo_file(struct work_struct *work)
+{
+	remove_proc_entry(ALLOCINFO_FILE_NAME, NULL);
+}
+
+static DECLARE_WORK(remove_allocinfo_work, remove_allocinfo_file);
+
 static void shutdown_mem_profiling(bool remove_file)
 {
 	if (mem_alloc_profiling_enabled())
@@ -600,7 +608,7 @@ static void shutdown_mem_profiling(bool remove_file)
 		return;
 
 	if (remove_file)
-		remove_proc_entry(ALLOCINFO_FILE_NAME, NULL);
+		schedule_work(&remove_allocinfo_work);
 	mem_profiling_support = false;
 }
 
@@ -1358,16 +1366,10 @@ static int __init alloc_tag_init(void)
 		return 0;
 	}
 
-	if (!proc_create(ALLOCINFO_FILE_NAME, 0400, NULL, &allocinfo_proc_ops)) {
-		pr_err("Failed to create %s file\n", ALLOCINFO_FILE_NAME);
-		shutdown_mem_profiling(false);
-		return -ENOMEM;
-	}
-
 	res = alloc_mod_tags_mem();
 	if (res) {
 		pr_err("Failed to reserve address space for module tags, errno = %d\n", res);
-		shutdown_mem_profiling(true);
+		shutdown_mem_profiling(false);
 		return res;
 	}
 
@@ -1375,10 +1377,16 @@ static int __init alloc_tag_init(void)
 	if (IS_ERR(alloc_tag_cttype)) {
 		pr_err("Allocation tags registration failed, errno = %pe\n", alloc_tag_cttype);
 		free_mod_tags_mem();
-		shutdown_mem_profiling(true);
+		shutdown_mem_profiling(false);
 		return PTR_ERR(alloc_tag_cttype);
 	}
 
+	if (!proc_create(ALLOCINFO_FILE_NAME, 0400, NULL, &allocinfo_proc_ops)) {
+		pr_err("Failed to create %s file\n", ALLOCINFO_FILE_NAME);
+		shutdown_mem_profiling(false);
+		return -ENOMEM;
+	}
+
 	return 0;
 }
 module_init(alloc_tag_init);
-- 
2.25.1

Re: [PATCH v10 0/6] alloc_tag and module codetag section fixes

[Suren Baghdasaryan]

On Mon, Sep 14, 2026 at 11:59 PM Hao Ge <hao.ge@linux.dev> wrote:
>
> With profiling toggled off, the overflow check in
> reserve_module_tags() did not run, a module could load with more
> tags than the page flags can address, and re-enabling profiling
> then silently corrupted /proc/allocinfo. On overflow the fix shuts
> profiling down, releases the reservation and returns -EAGAIN, and
> the codetag section lands as regular module data in the same load,
> so the module loads without profiling.
>
> Review of the earlier series by Sashiko turned up two more problems.
>
> One is a race. layout_sections() and move_module() both asked
> codetag_needs_module_section() where a codetag section goes, and
> mem_profiling_support can change between the two calls, for instance
> when another module load overflows the tag index and shuts profiling
> down. move_module() then copied the codetag section to offset 0 of
> its regular destination and clobbered the first section placed in
> that region.
>
> v7 reworks where codetag sections are allocated, on a prototype by
> Petr Pavlu [1]. The allocation runs before layout_sections() and the
> placement is decided in one step, so nothing re-asks the question
> and the race is gone. The retry is gone too, on -EAGAIN the section
> is laid out as regular module data right in the same load.
>
> [1] https://lore.kernel.org/all/499bb60c-c6e3-43a3-bd92-95a0567ece5e@suse.com/
>
> Following review feedback from Petr and Suren the series is now
> split into six patches. The two alloc_tag fixes from [2] are folded
> in as patches 5 and 6 and replace the versions currently in the mm
> tree. Sashiko keeps flagging the percpu counter leak [3], patch 5
> fixes it, and now the whole set goes through review again.
>
> [2] https://lore.kernel.org/all/20260817062726.106511-1-hao.ge@linux.dev/
> [3] https://lore.kernel.org/all/20260908094736.2B1A61F00A3A@smtp.kernel.org/
>
> Patch 1 moves release_module_tags() above reserve_module_tags(),
> since the failure paths now have to call it.
>
> Patch 2 cleans up the populate failure path: the reservation is
> released, module_tags.size rolled back and the PTEs a failed
> vmap_pages_range() left behind unmapped, so a later populate of the
> same range is safe. It carries both Fixes tags so it backports
> wherever patch 4 goes, which uses its prev_size.
>
> Patch 3 introduces SH_ENTSIZE_STANDALONE to mark sections with a
> separate allocation. The percpu section was previously excluded
> from the layout by clearing its SHF_ALLOC, which per the ELF spec
> says the section occupies memory during execution, and percpu does,
> only outside the regular module layout. The mark lives in sh_entsize
> now, find_sec(".data..percpu") gives stable results again and
> apply_relocations() goes back to testing only SHF_ALLOC. The
> section would show up under /sys/module/*/sections/, but it has one
> instance per CPU and no single address, and the entry never
> existed, so add_sect_attrs() and add_notes_attrs() skip it.
>
> Patch 4 moves the codetag allocation out of move_module() in front
> of layout_sections(), so the placement is decided in one step and
> the race is gone. On overflow profiling is shut down, the
> reservation released, module_tags.size rolled back and -EAGAIN
> returned, the section is laid out as regular module data and the
> module loads without profiling instead of failing. Any other error
> fails the load. The release and the fallback belong together,
> without the release rmmod hits the stale entry and panics.
>
> Patch 5 skips the percpu counter allocation when profiling is off.
> After the shutdown modules load their codetag section as regular
> data, load_module() still allocated counters for every tag and
> release_module_tags() cannot find them on unload, so they leaked
> (Suggested by Suren).
>
> Patch 6 defers the /proc/allocinfo removal to a workqueue.
> shutdown_mem_profiling() runs under mod_lock, and the synchronous
> remove_proc_entry() deadlocks with a reader taking mod_lock for
> read in allocinfo_start(). The file is also created at the end of
> alloc_tag_init(), a leftover file after a failed init would panic
> its readers (Found by Sashiko).
>
> Tested on an x86_64 virtual machine:
>
> Booted without sysctl.vm.mem_profiling=1,compressed:
> # cat /proc/allocinfo is fine
>
> Booted with sysctl.vm.mem_profiling=1,compressed:
> # cat /proc/allocinfo is fine
> # insmod overflow_tag.ko
> # dmesg
>   With module overflow_tag there are too many tags to fit in 13 page
>   flag bits. Memory allocation profiling is disabled!

Do you have your overflow_tag module posted anywhere in public (github perhaps?)

> # rmmod overflow_tag
> The module loads without profiling and unloads cleanly.
>
> Also ran continuous LTP stress for a few days, nothing abnormal
> so far.
>
> Changes in v10:
> - fold in the two alloc_tag fixes from [2] as patches 5 and 6, they
>   replace the versions in the mm tree and fix the percpu leak
>   Sashiko keeps flagging [3]
> - create /proc/allocinfo at the end of alloc_tag_init(), a
>   leftover file after a failed init would panic its readers
>   (Found by Sashiko)
> - count note sections with sect_visible() in add_notes_attrs() too,
>   the count has to match the fill loop
>
> Changes in v9:
> - do not export .data..percpu under /sys/module/*/sections/ (Petr
>   Pavlu).
> - move the populate failure cleanup in front of the rework, v8 patch
>   4 is patch 2 now. It declares prev_size itself, which the overflow
>   path of the rework also uses, so it carries both Fixes tags and the
>   two patches backport together
>
> Changes in v8:
> - roll back module_tags.size when the reservation is released, so a
>   concurrent load which already passed needs_section_mem() does not
>   skip populate for the freed gap (Sashiko)
> - unmap the PTEs a failed vmap_pages_range() installed, a retry to
>   populate the same range would BUG on them
> - zero the separately allocated codetag memory for an SHT_NOBITS
>   section, the tag area pages are not zeroed on allocation (Sashiko)
> - .data..percpu is exported under /sys/module/*/sections/ with the
>   boot CPU instance of the per-cpu area, and the interface is
>   documented in the ABI docs (Petr Pavlu)
> - keep a comment in apply_relocations() on how .data..percpu is
>   relocated (Petr Pavlu)
> - replace the "Based-on-a-patch-by:" tag with an in-body
>   attribution and a numbered Link: (Andrew Morton)
>
> Changes in v7:
> - split the rework following review feedback (Petr Pavlu, Suren
>   Baghdasaryan)
> - new patch 2 marks separately allocated sections with
>   SH_ENTSIZE_STANDALONE instead of clearing SHF_ALLOC
>   (suggested by Petr Pavlu)
> - split the populate failure release into its own patch (suggested
>   by Suren Baghdasaryan)
>
> Changes in v6:
> - rework on Petr's prototype and allocate codetag sections before
>   layout_sections(), the retry and its state resets are gone
> - fix the layout_sections()/move_module() race (Found by Sashiko)
> - release the reservation on populate failure as well (Found by
>   Sashiko)
> - only -EAGAIN keeps the fallback, other errors fail the load
>
> Changes in v5:
> - add Fixes: and Cc: stable to patch 1/2 as well, since 2/2 does not
>   compile without it (Andrew Morton)
> - restore frob-adjusted mem[type].size on retry instead of zeroing,
>   as s390 and parisc add GOT/PLT space there in
>   module_frob_arch_sections() (Reported by Sashiko)
> - drop the load_module() mem_profiling_support check; the percpu
>   counter leak is pre-existing and orthogonal to this fix
>
> Changes in v4:
> - add a new patch (1/2) to move release_module_tags() above
>   reserve_module_tags(); the overflow fix is 2/2
> - release the reservation on the -EAGAIN path
> - return -EAGAIN instead of -ENOMEM so the module can still load
>   without profiling (Suren)
> - reset sh_addr, mem[type].size and sym/str SHF_ALLOC before retry
> - skip percpu counters in load_module() when profiling is off
>
> Changes in v3:
> - use pr_warn_once() instead of pr_warn()
> - return -ENOMEM instead of -ENOSPC (Suren)
> - expand the commit message to describe the /proc/allocinfo impact
>   (Andrew)
>
> Changes in v2:
> - return an error after shutdown_mem_profiling() to skip
>   vm_module_tags_populate()
>
> v1: https://lore.kernel.org/all/20260804064408.105033-1-hao.ge@linux.dev/
> v2: https://lore.kernel.org/all/20260804122038.190270-1-hao.ge@linux.dev/
> v3: https://lore.kernel.org/all/20260805090633.141001-1-hao.ge@linux.dev/
> v4: https://lore.kernel.org/all/20260810093955.153015-1-hao.ge@linux.dev/
> v5: https://lore.kernel.org/all/20260812054105.102637-1-hao.ge@linux.dev/
> v6: https://lore.kernel.org/all/20260831072104.120197-1-hao.ge@linux.dev/
> v7: https://lore.kernel.org/all/20260902081802.146145-1-hao.ge@linux.dev/
> v8: https://lore.kernel.org/all/20260907062414.106873-1-hao.ge@linux.dev/
> v9: https://lore.kernel.org/all/20260908092412.115953-1-hao.ge@linux.dev/
>
> Hao Ge (6):
>   alloc_tag: move release_module_tags() above reserve_module_tags()
>   alloc_tag: clean up the populate failure path
>   module: introduce SH_ENTSIZE_STANDALONE for separately allocated
>     sections
>   module: allocate codetag sections before the regular module layout
>   alloc_tag: skip percpu counter allocation when profiling is disabled
>   alloc_tag: Defer /proc/allocinfo removal to a workqueue
>
>  include/linux/module.h   |   2 +
>  kernel/module/internal.h |   8 +++
>  kernel/module/kallsyms.c |  13 +---
>  kernel/module/main.c     | 133 +++++++++++++++++++-----------------
>  kernel/module/sysfs.c    |  17 +++--
>  lib/codetag.c            |  10 ++-
>  mm/alloc_tag.c           | 141 +++++++++++++++++++++++----------------
>  7 files changed, 188 insertions(+), 136 deletions(-)
>
> --
> 2.25.1
>

Re: [PATCH v10 2/6] alloc_tag: clean up the populate failure path

[Suren Baghdasaryan]

On Mon, Sep 14, 2026 at 11:59 PM Hao Ge <hao.ge@linux.dev> wrote:
>
> The reservation is already stored in the maple tree when
> vm_module_tags_populate() fails. A failed load never unloads the
> module, so nothing releases the entry. Release it and roll
> module_tags.size back. Without the rollback a concurrent load that
> already passed needs_section_mem() can reuse the freed gap, skip
> vm_module_tags_populate() and write to unmapped memory.
>
> vmap_pages_range() may have installed some PTEs before failing. A
> retry to populate the same range would BUG on them, so undo them,
> but only if vmap actually ran.
>
> Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
> Fixes: 0f9b685626da ("alloc_tag: populate memory for module tags as needed")

If a patch fixes two other patches, it likely needs to be split so
that each part fixes one patch only. But read on please. I'm not sure
if one of the fixes is correct.

> Reported-by: Sashiko <sashiko-bot@kernel.org>
> Cc: stable@vger.kernel.org
> Signed-off-by: Hao Ge <hao.ge@linux.dev>
> ---
>  mm/alloc_tag.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
>
> diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
> index 2070e682fe10..95ddf5b743d0 100644
> --- a/mm/alloc_tag.c
> +++ b/mm/alloc_tag.c
> @@ -804,6 +804,13 @@ static int vm_module_tags_populate(void)
>                                      next_page, PAGE_SHIFT) < 0) {
>                         release_pages_arg arg = { .pages = next_page };
>
> +                       /*
> +                        * vmap_pages_range() only runs once all pages were
> +                        * allocated, and it may have installed some mappings
> +                        * before failing. Undo them.

If vmap_pages_range() failed to map the range, why should we need to
undo it? If it indeed leaves some partial mapping when failing then I
would argue that vmap_pages_range() should be undoing these partial
mappings itself before returning the error.

> +                        */
> +                       if (nr == more_pages)
> +                               vunmap_range(phys_end, phys_end + (nr << PAGE_SHIFT));
>                         /* Clean up and error out */
>                         release_pages(arg, nr);
>                         return -ENOMEM;
> @@ -947,6 +954,7 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
>                 return ret;
>
>         if (module_tags.size < offset + size) {
> +               unsigned long prev_size = module_tags.size;
>                 int grow_res;
>
>                 module_tags.size = offset + size;
> @@ -961,6 +969,8 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
>                         shutdown_mem_profiling(true);
>                         pr_err("Failed to allocate memory for allocation tags in the module %s. Memory allocation profiling is disabled!\n",
>                                mod->name);
> +                       release_module_tags(mod, false);
> +                       module_tags.size = prev_size;

Yes, this one is a valid fix. It fixes 0f9b685626da commit.

>                         return ERR_PTR(grow_res);
>                 }
>         }
> --
> 2.25.1
>

Re: [PATCH v10 0/6] alloc_tag and module codetag section fixes

[Hao Ge]

Hi Suren

On 2026/9/16 02:23, Suren Baghdasaryan wrote:
> On Mon, Sep 14, 2026 at 11:59 PM Hao Ge <hao.ge@linux.dev> wrote:
>>
>> With profiling toggled off, the overflow check in
>> reserve_module_tags() did not run, a module could load with more
>> tags than the page flags can address, and re-enabling profiling
>> then silently corrupted /proc/allocinfo. On overflow the fix shuts
>> profiling down, releases the reservation and returns -EAGAIN, and
>> the codetag section lands as regular module data in the same load,
>> so the module loads without profiling.
>>
>> Review of the earlier series by Sashiko turned up two more problems.
>>
>> One is a race. layout_sections() and move_module() both asked
>> codetag_needs_module_section() where a codetag section goes, and
>> mem_profiling_support can change between the two calls, for instance
>> when another module load overflows the tag index and shuts profiling
>> down. move_module() then copied the codetag section to offset 0 of
>> its regular destination and clobbered the first section placed in
>> that region.
>>
>> v7 reworks where codetag sections are allocated, on a prototype by
>> Petr Pavlu [1]. The allocation runs before layout_sections() and the
>> placement is decided in one step, so nothing re-asks the question
>> and the race is gone. The retry is gone too, on -EAGAIN the section
>> is laid out as regular module data right in the same load.
>>
>> [1] https://lore.kernel.org/all/499bb60c-c6e3-43a3-bd92-95a0567ece5e@suse.com/
>>
>> Following review feedback from Petr and Suren the series is now
>> split into six patches. The two alloc_tag fixes from [2] are folded
>> in as patches 5 and 6 and replace the versions currently in the mm
>> tree. Sashiko keeps flagging the percpu counter leak [3], patch 5
>> fixes it, and now the whole set goes through review again.
>>
>> [2] https://lore.kernel.org/all/20260817062726.106511-1-hao.ge@linux.dev/
>> [3] https://lore.kernel.org/all/20260908094736.2B1A61F00A3A@smtp.kernel.org/
>>
>> Patch 1 moves release_module_tags() above reserve_module_tags(),
>> since the failure paths now have to call it.
>>
>> Patch 2 cleans up the populate failure path: the reservation is
>> released, module_tags.size rolled back and the PTEs a failed
>> vmap_pages_range() left behind unmapped, so a later populate of the
>> same range is safe. It carries both Fixes tags so it backports
>> wherever patch 4 goes, which uses its prev_size.
>>
>> Patch 3 introduces SH_ENTSIZE_STANDALONE to mark sections with a
>> separate allocation. The percpu section was previously excluded
>> from the layout by clearing its SHF_ALLOC, which per the ELF spec
>> says the section occupies memory during execution, and percpu does,
>> only outside the regular module layout. The mark lives in sh_entsize
>> now, find_sec(".data..percpu") gives stable results again and
>> apply_relocations() goes back to testing only SHF_ALLOC. The
>> section would show up under /sys/module/*/sections/, but it has one
>> instance per CPU and no single address, and the entry never
>> existed, so add_sect_attrs() and add_notes_attrs() skip it.
>>
>> Patch 4 moves the codetag allocation out of move_module() in front
>> of layout_sections(), so the placement is decided in one step and
>> the race is gone. On overflow profiling is shut down, the
>> reservation released, module_tags.size rolled back and -EAGAIN
>> returned, the section is laid out as regular module data and the
>> module loads without profiling instead of failing. Any other error
>> fails the load. The release and the fallback belong together,
>> without the release rmmod hits the stale entry and panics.
>>
>> Patch 5 skips the percpu counter allocation when profiling is off.
>> After the shutdown modules load their codetag section as regular
>> data, load_module() still allocated counters for every tag and
>> release_module_tags() cannot find them on unload, so they leaked
>> (Suggested by Suren).
>>
>> Patch 6 defers the /proc/allocinfo removal to a workqueue.
>> shutdown_mem_profiling() runs under mod_lock, and the synchronous
>> remove_proc_entry() deadlocks with a reader taking mod_lock for
>> read in allocinfo_start(). The file is also created at the end of
>> alloc_tag_init(), a leftover file after a failed init would panic
>> its readers (Found by Sashiko).
>>
>> Tested on an x86_64 virtual machine:
>>
>> Booted without sysctl.vm.mem_profiling=1,compressed:
>> # cat /proc/allocinfo is fine
>>
>> Booted with sysctl.vm.mem_profiling=1,compressed:
>> # cat /proc/allocinfo is fine
>> # insmod overflow_tag.ko
>> # dmesg
>>   With module overflow_tag there are too many tags to fit in 13 page
>>   flag bits. Memory allocation profiling is disabled!
> 
> Do you have your overflow_tag module posted anywhere in public (github perhaps?)
>

This is our simple test program:

---
 overflow_tag/overflow_tag.c | 109 ++++++++++++++++++++++++++++++++++++
 1 file changed, 109 insertions(+)
 create mode 100644 overflow_tag/overflow_tag.c

diff --git a/overflow_tag/overflow_tag.c b/overflow_tag/overflow_tag.c
new file mode 100644
index 000000000000..2745ab766473
--- /dev/null
+++ b/overflow_tag/overflow_tag.c
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * overflow_tag - overflow the compressed allocation tag index
+ */
+
+#include <linux/alloc_tag.h>
+#include <linux/debugfs.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+
+#ifndef CONFIG_MEM_ALLOC_PROFILING
+#error "overflow_tag requires CONFIG_MEM_ALLOC_PROFILING=y"
+#endif
+
+#ifndef NR_OVERFLOW_TAG_BITS
+#define NR_OVERFLOW_TAG_BITS	13
+#endif
+
+#define NR_OVERFLOW_TAGS	(1UL << NR_OVERFLOW_TAG_BITS)
+
+#define MAX_TRIGGER_PAGES	10000
+
+static struct alloc_tag overflow_tags[NR_OVERFLOW_TAGS] __used __aligned(8)
+		__section(ALLOC_TAG_SECTION_NAME) = {
+	[0 ... NR_OVERFLOW_TAGS - 1] = {
+		.ct = {
+			.modname	= KBUILD_MODNAME,
+			.function	= "overflow_tags",
+			.filename	= __FILE__,
+			.lineno		= __LINE__,
+		},
+		.counters = NULL,
+	},
+};
+
+static int overflow_alloc_and_free(unsigned long nr_pages)
+{
+	struct page **pages;
+	unsigned long i;
+
+	pages = kvmalloc_array(nr_pages, sizeof(*pages), GFP_KERNEL);
+	if (!pages)
+		return -ENOMEM;
+
+	for (i = 0; i < nr_pages; i++)
+		pages[i] = alloc_hooks_tag(
+			&overflow_tags[NR_OVERFLOW_TAGS - 1],
+			alloc_pages_noprof(GFP_KERNEL, 0));
+
+	for (i = 0; i < nr_pages; i++)
+		if (pages[i])
+			__free_pages(pages[i], 0);
+
+	kvfree(pages);
+	return 0;
+}
+
+static ssize_t overflow_tag_alloc_write(struct file *file,
+					const char __user *ubuf, size_t count,
+					loff_t *ppos)
+{
+	unsigned long nr_pages;
+	int err;
+
+	err = kstrtoul_from_user(ubuf, count, 0, &nr_pages);
+	if (err)
+		return err;
+
+	if (!nr_pages || nr_pages > MAX_TRIGGER_PAGES)
+		return -EINVAL;
+
+	err = overflow_alloc_and_free(nr_pages);
+	if (err)
+		return err;
+
+	pr_info("allocated and freed %lu pages through the last tag\n",
+		nr_pages);
+
+	return count;
+}
+
+static const struct file_operations overflow_tag_fops = {
+	.owner		= THIS_MODULE,
+	.write		= overflow_tag_alloc_write,
+};
+
+static struct dentry *overflow_tag_dir;
+
+static int __init overflow_tag_init(void)
+{
+	overflow_tag_dir = debugfs_create_dir("overflow_tag", NULL);
+	debugfs_create_file("alloc_and_free", 0200, overflow_tag_dir, NULL,
+			    &overflow_tag_fops);
+
+	pr_info("loaded with %lu allocation tags\n", NR_OVERFLOW_TAGS);
+	return 0;
+}
+
+static void __exit overflow_tag_exit(void)
+{
+	debugfs_remove_recursive(overflow_tag_dir);
+}
+
+module_init(overflow_tag_init);
+module_exit(overflow_tag_exit);
+
+MODULE_AUTHOR("Hao Ge <hao.ge@linux.dev>");
+MODULE_DESCRIPTION("Test module overflowing the compressed allocation tag index");
+MODULE_LICENSE("GPL");
-- 
2.25.1

The test machine uses 13 bits for tag index. Relevant kernel config is as follows:
CONFIG_NUMA_BALANCING=y
CONFIG_NR_CPUS=32
CONFIG_NODES_SHIFT=10
CONFIG_SPARSEMEM=y
CONFIG_SPARSEMEM_VMEMMAP=y
CONFIG_ZONE_DEVICE=y
# CONFIG_LRU_GEN is not set
CONFIG_MEM_ALLOC_PROFILING=y
# CONFIG_MEM_ALLOC_PROFILING_DEBUG is not set

Steps to reproduce the corruption on the unfixed kernel:
# boot with sysctl.vm.mem_profiling=0,compressed
# insmod overflow_tag.ko
# echo 1 > /proc/sys/vm/mem_profiling
# echo 10 > /sys/kernel/debug/overflow_tag/alloc_and_free
# cat /proc/allocinfo

Observed corruption after a 10-page run:
       40960       10 overflow_tag.c:30 [overflow_tag] func:overflow_tags
       -40960 18446744073709551606 overflow_tag.c:41 [overflow_tag] func:overflow_alloc_and_free

Thanks
Best Regards
Hao

>> # rmmod overflow_tag
>> The module loads without profiling and unloads cleanly.
>>
>> Also ran continuous LTP stress for a few days, nothing abnormal
>> so far.
>>
>> Changes in v10:
>> - fold in the two alloc_tag fixes from [2] as patches 5 and 6, they
>>   replace the versions in the mm tree and fix the percpu leak
>>   Sashiko keeps flagging [3]
>> - create /proc/allocinfo at the end of alloc_tag_init(), a
>>   leftover file after a failed init would panic its readers
>>   (Found by Sashiko)
>> - count note sections with sect_visible() in add_notes_attrs() too,
>>   the count has to match the fill loop
>>
>> Changes in v9:
>> - do not export .data..percpu under /sys/module/*/sections/ (Petr
>>   Pavlu).
>> - move the populate failure cleanup in front of the rework, v8 patch
>>   4 is patch 2 now. It declares prev_size itself, which the overflow
>>   path of the rework also uses, so it carries both Fixes tags and the
>>   two patches backport together
>>
>> Changes in v8:
>> - roll back module_tags.size when the reservation is released, so a
>>   concurrent load which already passed needs_section_mem() does not
>>   skip populate for the freed gap (Sashiko)
>> - unmap the PTEs a failed vmap_pages_range() installed, a retry to
>>   populate the same range would BUG on them
>> - zero the separately allocated codetag memory for an SHT_NOBITS
>>   section, the tag area pages are not zeroed on allocation (Sashiko)
>> - .data..percpu is exported under /sys/module/*/sections/ with the
>>   boot CPU instance of the per-cpu area, and the interface is
>>   documented in the ABI docs (Petr Pavlu)
>> - keep a comment in apply_relocations() on how .data..percpu is
>>   relocated (Petr Pavlu)
>> - replace the "Based-on-a-patch-by:" tag with an in-body
>>   attribution and a numbered Link: (Andrew Morton)
>>
>> Changes in v7:
>> - split the rework following review feedback (Petr Pavlu, Suren
>>   Baghdasaryan)
>> - new patch 2 marks separately allocated sections with
>>   SH_ENTSIZE_STANDALONE instead of clearing SHF_ALLOC
>>   (suggested by Petr Pavlu)
>> - split the populate failure release into its own patch (suggested
>>   by Suren Baghdasaryan)
>>
>> Changes in v6:
>> - rework on Petr's prototype and allocate codetag sections before
>>   layout_sections(), the retry and its state resets are gone
>> - fix the layout_sections()/move_module() race (Found by Sashiko)
>> - release the reservation on populate failure as well (Found by
>>   Sashiko)
>> - only -EAGAIN keeps the fallback, other errors fail the load
>>
>> Changes in v5:
>> - add Fixes: and Cc: stable to patch 1/2 as well, since 2/2 does not
>>   compile without it (Andrew Morton)
>> - restore frob-adjusted mem[type].size on retry instead of zeroing,
>>   as s390 and parisc add GOT/PLT space there in
>>   module_frob_arch_sections() (Reported by Sashiko)
>> - drop the load_module() mem_profiling_support check; the percpu
>>   counter leak is pre-existing and orthogonal to this fix
>>
>> Changes in v4:
>> - add a new patch (1/2) to move release_module_tags() above
>>   reserve_module_tags(); the overflow fix is 2/2
>> - release the reservation on the -EAGAIN path
>> - return -EAGAIN instead of -ENOMEM so the module can still load
>>   without profiling (Suren)
>> - reset sh_addr, mem[type].size and sym/str SHF_ALLOC before retry
>> - skip percpu counters in load_module() when profiling is off
>>
>> Changes in v3:
>> - use pr_warn_once() instead of pr_warn()
>> - return -ENOMEM instead of -ENOSPC (Suren)
>> - expand the commit message to describe the /proc/allocinfo impact
>>   (Andrew)
>>
>> Changes in v2:
>> - return an error after shutdown_mem_profiling() to skip
>>   vm_module_tags_populate()
>>
>> v1: https://lore.kernel.org/all/20260804064408.105033-1-hao.ge@linux.dev/
>> v2: https://lore.kernel.org/all/20260804122038.190270-1-hao.ge@linux.dev/
>> v3: https://lore.kernel.org/all/20260805090633.141001-1-hao.ge@linux.dev/
>> v4: https://lore.kernel.org/all/20260810093955.153015-1-hao.ge@linux.dev/
>> v5: https://lore.kernel.org/all/20260812054105.102637-1-hao.ge@linux.dev/
>> v6: https://lore.kernel.org/all/20260831072104.120197-1-hao.ge@linux.dev/
>> v7: https://lore.kernel.org/all/20260902081802.146145-1-hao.ge@linux.dev/
>> v8: https://lore.kernel.org/all/20260907062414.106873-1-hao.ge@linux.dev/
>> v9: https://lore.kernel.org/all/20260908092412.115953-1-hao.ge@linux.dev/
>>
>> Hao Ge (6):
>>   alloc_tag: move release_module_tags() above reserve_module_tags()
>>   alloc_tag: clean up the populate failure path
>>   module: introduce SH_ENTSIZE_STANDALONE for separately allocated
>>     sections
>>   module: allocate codetag sections before the regular module layout
>>   alloc_tag: skip percpu counter allocation when profiling is disabled
>>   alloc_tag: Defer /proc/allocinfo removal to a workqueue
>>
>>  include/linux/module.h   |   2 +
>>  kernel/module/internal.h |   8 +++
>>  kernel/module/kallsyms.c |  13 +---
>>  kernel/module/main.c     | 133 +++++++++++++++++++-----------------
>>  kernel/module/sysfs.c    |  17 +++--
>>  lib/codetag.c            |  10 ++-
>>  mm/alloc_tag.c           | 141 +++++++++++++++++++++++----------------
>>  7 files changed, 188 insertions(+), 136 deletions(-)
>>
>> --
>> 2.25.1
>>

Re: [PATCH v10 2/6] alloc_tag: clean up the populate failure path

[Hao Ge]

Hi Suren

Thanks for you review.

On 2026/9/16 05:09, Suren Baghdasaryan wrote:
> On Mon, Sep 14, 2026 at 11:59 PM Hao Ge <hao.ge@linux.dev> wrote:
>>
>> The reservation is already stored in the maple tree when
>> vm_module_tags_populate() fails. A failed load never unloads the
>> module, so nothing releases the entry. Release it and roll
>> module_tags.size back. Without the rollback a concurrent load that
>> already passed needs_section_mem() can reuse the freed gap, skip
>> vm_module_tags_populate() and write to unmapped memory.
>>
>> vmap_pages_range() may have installed some PTEs before failing. A
>> retry to populate the same range would BUG on them, so undo them,
>> but only if vmap actually ran.
>>
>> Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
>> Fixes: 0f9b685626da ("alloc_tag: populate memory for module tags as needed")
> 
> If a patch fixes two other patches, it likely needs to be split so
> that each part fixes one patch only. But read on please. I'm not sure
> if one of the fixes is correct.
>

Kept them in one patch because they only make sense together.

After a failed populate the reservation is released and the
size rolled back, so the next load takes the same gap and calls
vm_module_tags_populate() again, and that retry would trip over
the PTEs the failed vmap left behind.

Without the rollback a load that already passed needs_section_mem()
just reuses the gap, skips populate and writes to unmapped memory.
So splitting them leaves each half broken.                                   

>> Reported-by: Sashiko <sashiko-bot@kernel.org>
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Hao Ge <hao.ge@linux.dev>
>> ---
>>  mm/alloc_tag.c | 10 ++++++++++
>>  1 file changed, 10 insertions(+)
>>
>> diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
>> index 2070e682fe10..95ddf5b743d0 100644
>> --- a/mm/alloc_tag.c
>> +++ b/mm/alloc_tag.c
>> @@ -804,6 +804,13 @@ static int vm_module_tags_populate(void)
>>                                      next_page, PAGE_SHIFT) < 0) {
>>                         release_pages_arg arg = { .pages = next_page };
>>
>> +                       /*
>> +                        * vmap_pages_range() only runs once all pages were
>> +                        * allocated, and it may have installed some mappings
>> +                        * before failing. Undo them.
> 
> If vmap_pages_range() failed to map the range, why should we need to
> undo it? If it indeed leaves some partial mapping when failing then I
> would argue that vmap_pages_range() should be undoing these partial
> mappings itself before returning the error.
>
Yeah it does - none of the paths under vmap_pages_range() undo themselves on error.
pcpu_map_pages does perform cleanup on its own as well. There may be other similar sites.
https://elixir.bootlin.com/linux/v7.3-rc3/source/mm/percpu-vm.c#L255
so I did the same here.

If we really want vmap_pages_range() to handle this cleanup internally, I believe that would
be a relatively large change, and we would need to consider many more cases. So I think we keep
the current behaviour for now.

Thanks
Best Regards
Hao

>> +                        */
>> +                       if (nr == more_pages)
>> +                               vunmap_range(phys_end, phys_end + (nr << PAGE_SHIFT));
>>                         /* Clean up and error out */
>>                         release_pages(arg, nr);
>>                         return -ENOMEM;
>> @@ -947,6 +954,7 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
>>                 return ret;
>>
>>         if (module_tags.size < offset + size) {
>> +               unsigned long prev_size = module_tags.size;
>>                 int grow_res;
>>
>>                 module_tags.size = offset + size;
>> @@ -961,6 +969,8 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
>>                         shutdown_mem_profiling(true);
>>                         pr_err("Failed to allocate memory for allocation tags in the module %s. Memory allocation profiling is disabled!\n",
>>                                mod->name);
>> +                       release_module_tags(mod, false);
>> +                       module_tags.size = prev_size;
> 
> Yes, this one is a valid fix. It fixes 0f9b685626da commit.
> 
>>                         return ERR_PTR(grow_res);
>>                 }
>>         }
>> --
>> 2.25.1
>>

Re: [PATCH v10 2/6] alloc_tag: clean up the populate failure path

[Suren Baghdasaryan]

On Tue, Sep 15, 2026 at 11:00 PM Hao Ge <hao.ge@linux.dev> wrote:
>
> Hi Suren
>
> Thanks for you review.
>
> On 2026/9/16 05:09, Suren Baghdasaryan wrote:
> > On Mon, Sep 14, 2026 at 11:59 PM Hao Ge <hao.ge@linux.dev> wrote:
> >>
> >> The reservation is already stored in the maple tree when
> >> vm_module_tags_populate() fails. A failed load never unloads the
> >> module, so nothing releases the entry. Release it and roll
> >> module_tags.size back. Without the rollback a concurrent load that
> >> already passed needs_section_mem() can reuse the freed gap, skip
> >> vm_module_tags_populate() and write to unmapped memory.
> >>
> >> vmap_pages_range() may have installed some PTEs before failing. A
> >> retry to populate the same range would BUG on them, so undo them,
> >> but only if vmap actually ran.
> >>
> >> Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
> >> Fixes: 0f9b685626da ("alloc_tag: populate memory for module tags as needed")
> >
> > If a patch fixes two other patches, it likely needs to be split so
> > that each part fixes one patch only. But read on please. I'm not sure
> > if one of the fixes is correct.
> >
>
> Kept them in one patch because they only make sense together.
>
> After a failed populate the reservation is released and the
> size rolled back, so the next load takes the same gap and calls
> vm_module_tags_populate() again, and that retry would trip over
> the PTEs the failed vmap left behind.
>
> Without the rollback a load that already passed needs_section_mem()
> just reuses the gap, skips populate and writes to unmapped memory.
> So splitting them leaves each half broken.

Yeah, that's why I think vmap_pages_range() cleanup should be a
completely separate patch with vmap_pages_range() doing cleanup
itself.

>
> >> Reported-by: Sashiko <sashiko-bot@kernel.org>
> >> Cc: stable@vger.kernel.org
> >> Signed-off-by: Hao Ge <hao.ge@linux.dev>
> >> ---
> >>  mm/alloc_tag.c | 10 ++++++++++
> >>  1 file changed, 10 insertions(+)
> >>
> >> diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
> >> index 2070e682fe10..95ddf5b743d0 100644
> >> --- a/mm/alloc_tag.c
> >> +++ b/mm/alloc_tag.c
> >> @@ -804,6 +804,13 @@ static int vm_module_tags_populate(void)
> >>                                      next_page, PAGE_SHIFT) < 0) {
> >>                         release_pages_arg arg = { .pages = next_page };
> >>
> >> +                       /*
> >> +                        * vmap_pages_range() only runs once all pages were
> >> +                        * allocated, and it may have installed some mappings
> >> +                        * before failing. Undo them.
> >
> > If vmap_pages_range() failed to map the range, why should we need to
> > undo it? If it indeed leaves some partial mapping when failing then I
> > would argue that vmap_pages_range() should be undoing these partial
> > mappings itself before returning the error.
> >
> Yeah it does - none of the paths under vmap_pages_range() undo themselves on error.

Hmm. I wonder if this behavior is intentional or requires a fix. The
only possible reason I can think of is performance but I can't imagine
a failure to map is a performance-critical case.

+David Hildenbrand (Red Hat), +Lorenzo Stoakes (Oracle) what do you
guys think? Shouldn't vmap_pages_range() undo its possible partial
mapping when it fails?


> pcpu_map_pages does perform cleanup on its own as well. There may be other similar sites.
> https://elixir.bootlin.com/linux/v7.3-rc3/source/mm/percpu-vm.c#L255
> so I did the same here.
>
> If we really want vmap_pages_range() to handle this cleanup internally, I believe that would
> be a relatively large change, and we would need to consider many more cases. So I think we keep
> the current behaviour for now.

Yeah, I understand it would require a deeper cut but if that's the
right way to handle this we might as well do that. Let's see what MM
folks think.

>
> Thanks
> Best Regards
> Hao
>
> >> +                        */
> >> +                       if (nr == more_pages)
> >> +                               vunmap_range(phys_end, phys_end + (nr << PAGE_SHIFT));
> >>                         /* Clean up and error out */
> >>                         release_pages(arg, nr);
> >>                         return -ENOMEM;
> >> @@ -947,6 +954,7 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
> >>                 return ret;
> >>
> >>         if (module_tags.size < offset + size) {
> >> +               unsigned long prev_size = module_tags.size;
> >>                 int grow_res;
> >>
> >>                 module_tags.size = offset + size;
> >> @@ -961,6 +969,8 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
> >>                         shutdown_mem_profiling(true);
> >>                         pr_err("Failed to allocate memory for allocation tags in the module %s. Memory allocation profiling is disabled!\n",
> >>                                mod->name);
> >> +                       release_module_tags(mod, false);
> >> +                       module_tags.size = prev_size;
> >
> > Yes, this one is a valid fix. It fixes 0f9b685626da commit.
> >
> >>                         return ERR_PTR(grow_res);
> >>                 }
> >>         }
> >> --
> >> 2.25.1
> >>

Re: [PATCH v10 2/6] alloc_tag: clean up the populate failure path

[Suren Baghdasaryan]

On Wed, Sep 16, 2026 at 9:31 AM Suren Baghdasaryan <surenb@google.com> wrote:
>
> On Tue, Sep 15, 2026 at 11:00 PM Hao Ge <hao.ge@linux.dev> wrote:
> >
> > Hi Suren
> >
> > Thanks for you review.
> >
> > On 2026/9/16 05:09, Suren Baghdasaryan wrote:
> > > On Mon, Sep 14, 2026 at 11:59 PM Hao Ge <hao.ge@linux.dev> wrote:
> > >>
> > >> The reservation is already stored in the maple tree when
> > >> vm_module_tags_populate() fails. A failed load never unloads the
> > >> module, so nothing releases the entry. Release it and roll
> > >> module_tags.size back. Without the rollback a concurrent load that
> > >> already passed needs_section_mem() can reuse the freed gap, skip
> > >> vm_module_tags_populate() and write to unmapped memory.
> > >>
> > >> vmap_pages_range() may have installed some PTEs before failing. A
> > >> retry to populate the same range would BUG on them, so undo them,
> > >> but only if vmap actually ran.
> > >>
> > >> Fixes: 4835f747d3ed ("alloc_tag: support for page allocation tag compression")
> > >> Fixes: 0f9b685626da ("alloc_tag: populate memory for module tags as needed")
> > >
> > > If a patch fixes two other patches, it likely needs to be split so
> > > that each part fixes one patch only. But read on please. I'm not sure
> > > if one of the fixes is correct.
> > >
> >
> > Kept them in one patch because they only make sense together.
> >
> > After a failed populate the reservation is released and the
> > size rolled back, so the next load takes the same gap and calls
> > vm_module_tags_populate() again, and that retry would trip over
> > the PTEs the failed vmap left behind.
> >
> > Without the rollback a load that already passed needs_section_mem()
> > just reuses the gap, skips populate and writes to unmapped memory.
> > So splitting them leaves each half broken.
>
> Yeah, that's why I think vmap_pages_range() cleanup should be a
> completely separate patch with vmap_pages_range() doing cleanup
> itself.
>
> >
> > >> Reported-by: Sashiko <sashiko-bot@kernel.org>
> > >> Cc: stable@vger.kernel.org
> > >> Signed-off-by: Hao Ge <hao.ge@linux.dev>
> > >> ---
> > >>  mm/alloc_tag.c | 10 ++++++++++
> > >>  1 file changed, 10 insertions(+)
> > >>
> > >> diff --git a/mm/alloc_tag.c b/mm/alloc_tag.c
> > >> index 2070e682fe10..95ddf5b743d0 100644
> > >> --- a/mm/alloc_tag.c
> > >> +++ b/mm/alloc_tag.c
> > >> @@ -804,6 +804,13 @@ static int vm_module_tags_populate(void)
> > >>                                      next_page, PAGE_SHIFT) < 0) {
> > >>                         release_pages_arg arg = { .pages = next_page };
> > >>
> > >> +                       /*
> > >> +                        * vmap_pages_range() only runs once all pages were
> > >> +                        * allocated, and it may have installed some mappings
> > >> +                        * before failing. Undo them.
> > >
> > > If vmap_pages_range() failed to map the range, why should we need to
> > > undo it? If it indeed leaves some partial mapping when failing then I
> > > would argue that vmap_pages_range() should be undoing these partial
> > > mappings itself before returning the error.
> > >
> > Yeah it does - none of the paths under vmap_pages_range() undo themselves on error.
>
> Hmm. I wonder if this behavior is intentional or requires a fix. The
> only possible reason I can think of is performance but I can't imagine
> a failure to map is a performance-critical case.
>
> +David Hildenbrand (Red Hat), +Lorenzo Stoakes (Oracle) what do you

Huh, my gmail thinks you are still at your previous companies :)

> guys think? Shouldn't vmap_pages_range() undo its possible partial
> mapping when it fails?
>
>
> > pcpu_map_pages does perform cleanup on its own as well. There may be other similar sites.
> > https://elixir.bootlin.com/linux/v7.3-rc3/source/mm/percpu-vm.c#L255
> > so I did the same here.
> >
> > If we really want vmap_pages_range() to handle this cleanup internally, I believe that would
> > be a relatively large change, and we would need to consider many more cases. So I think we keep
> > the current behaviour for now.
>
> Yeah, I understand it would require a deeper cut but if that's the
> right way to handle this we might as well do that. Let's see what MM
> folks think.
>
> >
> > Thanks
> > Best Regards
> > Hao
> >
> > >> +                        */
> > >> +                       if (nr == more_pages)
> > >> +                               vunmap_range(phys_end, phys_end + (nr << PAGE_SHIFT));
> > >>                         /* Clean up and error out */
> > >>                         release_pages(arg, nr);
> > >>                         return -ENOMEM;
> > >> @@ -947,6 +954,7 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
> > >>                 return ret;
> > >>
> > >>         if (module_tags.size < offset + size) {
> > >> +               unsigned long prev_size = module_tags.size;
> > >>                 int grow_res;
> > >>
> > >>                 module_tags.size = offset + size;
> > >> @@ -961,6 +969,8 @@ static void *reserve_module_tags(struct module *mod, unsigned long size,
> > >>                         shutdown_mem_profiling(true);
> > >>                         pr_err("Failed to allocate memory for allocation tags in the module %s. Memory allocation profiling is disabled!\n",
> > >>                                mod->name);
> > >> +                       release_module_tags(mod, false);
> > >> +                       module_tags.size = prev_size;
> > >
> > > Yes, this one is a valid fix. It fixes 0f9b685626da commit.
> > >
> > >>                         return ERR_PTR(grow_res);
> > >>                 }
> > >>         }
> > >> --
> > >> 2.25.1
> > >>