Large tee ram va size #1669

jenswi-linaro · 2017-07-07T12:16:17Z

Introduces support for tee ram va size larger than the page directory size.

etienne-lms

Thanks for the split in tiny changes.
Reviewed, looks ok. Only minor comments.
Successfully tested on b2260 (dual c-a9) with few configurations spreading vcore range over 2 mmu sections.

etienne-lms · 2017-07-28T10:39:22Z

core/arch/arm/mm/tee_pager.c

+	 * Most addresses are mapped lineary, try that first if possible.
+	 */
+	if (!tee_pager_get_table_info(pa, &ti))
+		return NULL; /* impossible pa */


minor: move this test before the comment which is related to the sequence below.
Also, arg#1 of tee_pager_get_table_info() is a vaddr_t whereas we pass here a paddr_t. This would deserve a cast.

Actually the comment applies to the call of tee_pager_get_table_info() too.
I'll skip the cast as frivolous casting causes more problems than they solve.

etienne-lms · 2017-07-28T10:49:05Z

core/arch/arm/mm/tee_pager.c

+	idx = core_mmu_va2idx(&pt->tbl_info, smem);
+	while (pt <= (pager_tables + ARRAY_SIZE(pager_tables) - 1)) {
+		while (idx < TBL_NUM_ENTRIES) {
+			v = core_mmu_idx2va(&pt->tbl_info, idx);


minor: v is local to this instructions block.

This is a matter of taste. In this case I think the code is easier to read with v declared above together with the other local variables.

etienne-lms · 2017-07-28T10:49:47Z

core/arch/arm/mm/tee_pager.c

+	idx = core_mmu_va2idx(&pager_tables[n].tbl_info, CFG_TEE_RAM_START);
+	while (true) {
+		while (idx < TBL_NUM_ENTRIES) {
+			v = core_mmu_idx2va(&pager_tables[n].tbl_info, idx);


minor: v is local to this instructions block.

This is a matter of taste. In this case I think the code is easier to read with v declared above together with the other local variables.

etienne-lms · 2017-07-28T10:51:29Z

core/arch/arm/mm/tee_pager.c

+	while (true) {
+		while (idx < TBL_NUM_ENTRIES) {
+			v = core_mmu_idx2va(&pager_tables[n].tbl_info, idx);
+			if (v > (CFG_TEE_RAM_START + CFG_TEE_RAM_VA_SIZE))


minor: s/>/>=/

Thanks, I'll fix.

etienne-lms · 2017-07-28T10:56:12Z

core/arch/arm/mm/tee_pager.c

+	struct pager_table *pt = find_pager_table_may_fail(va);
+
+	if (!pt)
+		panic();


minor: i think an assert is enough. Functions that call find_pager_table() will simply page fault when using a null pager table pointer.

etienne-lms · 2017-07-28T11:00:37Z

core/arch/arm/mm/tee_pager.c

-
-	if (smem & SMALL_PAGE_MASK || nbytes & SMALL_PAGE_MASK)
-		panic("invalid area alignment");
-
 	pager_alias_area = mm;


minor: as set_alias_area() shall be called only once, at init, maybe we can restore the test on pager_alias_area == NULL, eventually through an assert().

etienne-lms · 2017-07-28T11:28:26Z

core/arch/arm/mm/tee_pager.c

-			base, size, ti->va_base, tbl_va_size);
-		panic();
+	while (s) {
+		s2 = MIN(CORE_MMU_PGDIR_SIZE - (b & CORE_MMU_PGDIR_MASK), s);


minor: s2 is local to this instruction block.

etienne-lms · 2017-07-28T12:19:05Z

core/arch/arm/mm/core_mmu_lpae.c

+					desc |= base_va - mm->va + mm->pa;
+			}
+
+			if (desc)


minor: if (desc == UNSET_DESC) (here and above) for consistency ?

No, that would be something different as UNSET_DESC is defined to -1

Ok, you are right. Ok for the one above.
For the if (desc) here, I wonder case where desc==UNSET_DESC and attr<0. Minor but still might print a "mapped" trace instead of a "not-mapped" trace. Maybe if (desc != UNSET_DESC && desc).

Hmm, it should probably be if (attr >= 0 && desc).
Edit: your test is better, I'll use that instead.

jenswi-linaro · 2017-08-16T08:14:18Z

Update

jenswi-linaro · 2017-08-16T14:44:40Z

Update

etienne-lms

Found 1 remaining case that must be fixed and a potential weakness.

Otherwise, successfully tested on non lpae (using b2260) and lpae (using qemu_armv8) with various secure RAM locations/sizes (with pager enabled of course).

etienne-lms · 2017-08-22T12:43:48Z

core/arch/arm/mm/core_mmu.c

@@ -824,7 +864,8 @@ static void init_mem_map(struct tee_mmap_region *memory_map, size_t num_elems)
 		/* Map non-flat mapped addresses above flat mapped addresses */
 		va = ROUNDUP(va + CFG_TEE_RAM_VA_SIZE, CORE_MMU_PGDIR_SIZE);


It is not very clean how va here is set to the end of the pager va space. Both "pager vaspace" end address and va here are computed using the same sequence, so that's fine now but not clean in the long term.
Maybe add_pager_vaspace() should update end so it can be reused here to define the virtual base address for the non-flat-mapped areas.

jenswi-linaro · 2017-08-22T14:15:25Z

Update, did the add_pager_vaspace() changes slightly different.

etienne-lms · 2017-08-22T14:55:41Z

core/arch/arm/mm/core_mmu.c

@@ -837,12 +843,14 @@ static void init_mem_map(struct tee_mmap_region *memory_map, size_t num_elems)
 	assert(va >= CFG_TEE_RAM_START);
 	assert(end <= CFG_TEE_RAM_START + CFG_TEE_RAM_VA_SIZE);

-	add_pager_vaspace(memory_map, num_elems, va, end, &last);
+	add_pager_vaspace(memory_map, num_elems, va, end, &last,
+			  &next_pgdir_va);



I think here, if pager is enabled, it is mandatory that the pager vmem space is fully reserved to pager. Hence, we should maybe add va = ROUNDDOWN(va, CORE_MMU_PGDIR_SIZE);, the same way we now have the pager vmem end provided through next_pgdir_va.

Maybe more simple to have add_pager_vaspace(memory_map, num_elems, &va, &end, &last); and the function to update va and end content if the virtual memory area reserved to TEE_RAM management must be enlarged due to pager constraint.

No, that's not needed.

why ? What insures the first area mapped right below TEE_RAM (case mapping setup with "tee ram at top") is pgdir aligned ? At this stage, va may not be pgdir aligned.

After few reading and test, I see the issue i'm talking about is not specific to this P-R. So maybe it should be treated outside this P-R.

The reason I don't want to enforce pgdir alignment of va is primary for LPAE. It seems wasteful to forbid usage of otherwise unused entries.

etienne-lms · 2017-08-23T13:57:16Z

core/arch/arm/mm/core_mmu.c

 #ifdef CFG_WITH_PAGER
-	size_t size = CFG_TEE_RAM_VA_SIZE - (end - begin);
+	size_t size = CFG_TEE_RAM_VA_SIZE - (end - begin_pgdir);


This is not ok. size is expected to cover from begin, not begin_pgdir, and next_pgdir_va should be end + size eventually rounded up to the pgdir size. With such changes, we would end to the original code (the one before this [review] patch). I even think the final roundup to pgdir size is useless.

Please propose what you have in mind and we'll take it from there.

Please look at jenswi-linaro#2.

etienne-lms · 2017-08-23T13:58:18Z

core/arch/arm/mm/core_mmu.c


 	if (core_mmu_place_tee_ram_at_top(va)) {
 		/* Map non-flat mapped addresses below flat mapped addresses */
 		for (map = memory_map; !core_mmap_is_end_of_table(map); map++) {
-			if (map_is_flat_mapped(map))
+			if (map_is_flat_mapped(map) ||
+			    map->type == MEM_AREA_PAGER_VASPACE)


Maybe we could replace this with if (map->va) continue; so that only the areas without an assigned virtual address get one assigned.

jenswi-linaro · 2017-08-24T09:20:37Z

core/arch/arm/mm/core_mmu.c

+	mmap[pos].region_size = SMALL_PAGE_SIZE;
+	mmap[pos].attr = core_mmu_type_to_attr(MEM_AREA_PAGER_VASPACE);
+
+	*end += size;


This needs to be updated regardless if pager is enabled or not.

Edit: Maybe not strictly necessary as no code depends on that part of the pgdir being unused.

I agree with the edited comment: when pager is disable, CFG_TEE_RAM_VA_SIZE is meaningless.
Hmm... maybe usefull also for the kasan ?

However, some platform do define CFG_TEE_RAM_VA_SIZE and some piece of code use it. Ok to not condition the sequence here and postponed cleaning of CFG_TEE_RAM_VA_SIZE to later if needed.

jenswi-linaro · 2017-08-29T08:43:28Z

What's next step here, squash and rebase for a final review?

etienne-lms · 2017-08-29T08:58:07Z

I would say remove CFG_WITH_PAGER condition from add_pager_vaspace and sqash rebase.
Reviewed-by: Etienne Carriere <[email protected]> for the current content.
If have successfully tested a rebased version of it on b2260 and qemu_armv8 (with various vcore ranges).

jforissier · 2017-08-29T09:00:09Z

core/arch/arm/plat-hikey/platform_config.h

@@ -120,7 +120,7 @@

 #define CFG_TEE_CORE_NB_CORE	8

-#define CFG_TEE_RAM_VA_SIZE	(1024 * 1024)
+#define CFG_TEE_RAM_VA_SIZE	(2 * 1024 * 1024)


Minor: the memory map above (comment block) needs fixing.

Actually, i don't think this change is required. 1MB is fine for current. Increase to 2MB should not hurt (but the layout description pointed by @jforissier) and allows to test the change.

s/0x3F10_0000/0x3F20_0000/

I think we all agree 2M is a better default value if only for exercising the new code.
Acked-by: Jerome Forissier <[email protected]>

jforissier · 2017-08-29T09:03:01Z

@jenswi-linaro yes I think so, but I'll let @etienne-lms speak up since he has been doing all the review work.

Prior to this patch was tee_mm_vcore initialized to cover the complete page directories covering TEE_RAM. With this patch tee_mm_vcore will only cover TEE_RAM in order to avoid returning unexpected addresses when allocating. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Adds the define CORE_MMU_PGDIR_LEVEL which indicates the level used for page directories. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Hides tee_pager_tbl_info and provides new needed functions. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

All failures in tee_pager_add_core_area() are fatal. Replaces return code with void and panics on errors instead. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Adds MEM_AREA_PAGER_VASPACE which is used to create empty translation tables as needed for the pager. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Moves some internal functions inside the pager code to prepare for a future commit. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Changes tee_pager_release_phys() to handle freeing a range of pages spanning multiple areas. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Deals with CFG_TEE_RAM_VA_SIZE > CORE_MMU_PGDIR_SIZE. This is a special problem as the pages managed by the pager then spans several translation tables. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

Increases CFG_TEE_RAM_VA_SIZE to 2 MiB for the plat-vexpress platforms. Reviewed-by: Etienne Carriere <[email protected]> Tested-by: Jens Wiklander <[email protected]> (QEMU v7) Signed-off-by: Jens Wiklander <[email protected]>

jenswi-linaro · 2017-08-29T10:06:25Z

Rebased and squashed. Tag applied to all commits but the last (plat-hikey) as @etienne-lms had some comment on that.

jenswi-linaro · 2017-08-29T12:05:31Z

@jforissier, did I get that right?

jforissier · 2017-08-29T12:12:07Z

Yes ;)

jenswi-linaro · 2017-08-29T12:39:18Z

Update

jforissier · 2017-08-29T12:40:23Z

For "core: plat-hikey: increase CFG_TEE_RAM_VA_SIZE":
Acked-by: Jerome Forissier <[email protected]>

Increases CFG_TEE_RAM_VA_SIZE to 2 MiB for the plat-hikey platform. Acked-by: Jerome Forissier <[email protected]> Tested-by: Jens Wiklander <[email protected]> (Hikey) Signed-off-by: Jens Wiklander <[email protected]>

jenswi-linaro · 2017-08-29T12:48:59Z

Tag applied

jenswi-linaro mentioned this pull request Jul 10, 2017

Confusion about set_alias_area() during pager initialization #1676

Closed

etienne-lms reviewed Jul 28, 2017

View reviewed changes

etienne-lms reviewed Aug 22, 2017

View reviewed changes

etienne-lms reviewed Aug 23, 2017

View reviewed changes

etienne-lms mentioned this pull request Aug 24, 2017

proposal for the large_tee_ram_va_size jenswi-linaro/optee_os#2

Merged

jenswi-linaro commented Aug 24, 2017

View reviewed changes

jforissier reviewed Aug 29, 2017

View reviewed changes

jenswi-linaro added 9 commits August 29, 2017 11:56

core: add CORE_MMU_PGDIR_LEVEL

c007125

Adds the define CORE_MMU_PGDIR_LEVEL which indicates the level used for page directories. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

core: pager: hide tee_pager_tbl_info

f09ea43

Hides tee_pager_tbl_info and provides new needed functions. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

core: pager: update tee_pager_add_core_area()

76be0a1

All failures in tee_pager_add_core_area() are fatal. Replaces return code with void and panics on errors instead. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

core: add MEM_AREA_PAGER_VASPACE

e4e77e2

Adds MEM_AREA_PAGER_VASPACE which is used to create empty translation tables as needed for the pager. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

core: pager: move some internal functions

76a75c2

Moves some internal functions inside the pager code to prepare for a future commit. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

core: pager: free pages across multiple areas

eaf1eaf

Changes tee_pager_release_phys() to handle freeing a range of pages spanning multiple areas. Reviewed-by: Etienne Carriere <[email protected]> Signed-off-by: Jens Wiklander <[email protected]>

core: plat-vexpress: increase CFG_TEE_RAM_VA_SIZE

11c0b0d

Increases CFG_TEE_RAM_VA_SIZE to 2 MiB for the plat-vexpress platforms. Reviewed-by: Etienne Carriere <[email protected]> Tested-by: Jens Wiklander <[email protected]> (QEMU v7) Signed-off-by: Jens Wiklander <[email protected]>

jenswi-linaro force-pushed the large_tee_ram_va_size branch from 6a84f54 to 4292368 Compare August 29, 2017 10:05

core: plat-hikey: increase CFG_TEE_RAM_VA_SIZE

6516fef

Increases CFG_TEE_RAM_VA_SIZE to 2 MiB for the plat-hikey platform. Acked-by: Jerome Forissier <[email protected]> Tested-by: Jens Wiklander <[email protected]> (Hikey) Signed-off-by: Jens Wiklander <[email protected]>

jenswi-linaro force-pushed the large_tee_ram_va_size branch from 8001472 to 6516fef Compare August 29, 2017 12:48

jforissier merged commit 8446c47 into OP-TEE:master Aug 29, 2017

jenswi-linaro deleted the large_tee_ram_va_size branch August 29, 2017 13:17

		@@ -824,7 +864,8 @@ static void init_mem_map(struct tee_mmap_region *memory_map, size_t num_elems)
		/* Map non-flat mapped addresses above flat mapped addresses */
		va = ROUNDUP(va + CFG_TEE_RAM_VA_SIZE, CORE_MMU_PGDIR_SIZE);

Large tee ram va size #1669

Large tee ram va size #1669

Conversation

jenswi-linaro commented Jul 7, 2017

etienne-lms left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jenswi-linaro Aug 16, 2017 • edited Loading

Choose a reason for hiding this comment

jenswi-linaro commented Aug 16, 2017

jenswi-linaro commented Aug 16, 2017

etienne-lms left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jenswi-linaro commented Aug 22, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jenswi-linaro Aug 24, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jenswi-linaro commented Aug 29, 2017

etienne-lms commented Aug 29, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jforissier commented Aug 29, 2017

jenswi-linaro commented Aug 29, 2017

jenswi-linaro commented Aug 29, 2017

jforissier commented Aug 29, 2017

jenswi-linaro commented Aug 29, 2017

jforissier commented Aug 29, 2017

jenswi-linaro commented Aug 29, 2017

jenswi-linaro Aug 16, 2017 •

edited

Loading

jenswi-linaro Aug 24, 2017 •

edited

Loading