Boot mem handling #7039
Boot mem handling #7039
Conversation
jenswi-linaro
force-pushed
the
boot-mem
branch
from
beeb6a6
to
e09cd0a
Compare
There was a problem hiding this comment.
For "mk/clang.mk: -Wno-gnu-alignof-expression":
s/dereferensing/dereferencing/
s/CLANG/Clang/
Reviewed-by: Jerome Forissier <[email protected]>
There was a problem hiding this comment.
- For "core: arm64: increase thread stack size for debug":
Reviewed-by: Jerome Forissier <[email protected]>
There was a problem hiding this comment.
- For "core: mm: add vaddr_to_phys()":
Reviewed-by: Jerome Forissier <[email protected]>
There was a problem hiding this comment.
- For "core: remove CORE_MEM_TA_RAM":
Reviewed-by: Jerome Forissier <[email protected]>
|
OK to squash in the fixes and move up the last commits a bit? |
Fine with me |
jenswi-linaro
force-pushed
the
boot-mem
branch
from
f7c663b
to
893dbc0
Compare
|
Tags applied, squashed in fixes, and moved the last added commits a bit earlier to fix problems before they occur. |
There was a problem hiding this comment.
- For "core: add VCORE_FREE_{PA,SZ,END_PA}":
s/CVORE/VCORE/
s/paractise/practice/
Acked-by: Jerome Forissier <[email protected]>
There was a problem hiding this comment.
- For "core: mm: allow unmapping VCORE_FREE":
Reviewed-by: Jerome Forissier <[email protected]>
core/arch/arm/kernel/kern.ld.S
Outdated
| ASSERT(__vcore_init_ro_start == __vcore_init_ro_end, | ||
| "__vcore_init_ro_start should follow __vcore_init_ro_end") |
There was a problem hiding this comment.
s/__vcore_init_ro_end/__vcore_init_rx_end/? And in the commit description "after __vcore_init_rx_end"?
There was a problem hiding this comment.
You're right, I'll update the commit.
jenswi-linaro
force-pushed
the
boot-mem
branch
from
893dbc0
to
0e3fb1e
Compare
|
Force push to update "core: arm,pager: make __vcore_init_ro_start follow __vcore_init_rx_end" and applied the tags. |
There was a problem hiding this comment.
Very minor review comments.
I was looking at the CFG_BOOT_MEM part and experienced the patches on my Armv7 boards. Commit "core: arm: enable CFG_BOOT_MEM unconditionally" makes my platforms to segfault at primary core init (only when pager is disabled). I dumped the traces below, in case it helps:
I/TC: Early console on UART#4
D/TC:0 add_phys_mem:750 VCORE_UNPG_RX_PA type TEE_RAM_RX 0xde000000 size 0x00076000
D/TC:0 add_phys_mem:750 VCORE_UNPG_RW_PA type TEE_RAM_RW 0xde076000 size 0x0005a000
D/TC:0 add_phys_mem:750 VCORE_UNPG_RW_PA type SEC_RAM_OVERALL 0xde076000 size 0x0005a000
D/TC:0 add_phys_mem:750 VCORE_FREE_PA type TEE_RAM_RW 0xde0d0000 size 0x00130000
D/TC:0 merge_mmaps:703 Merging 0xde076000..0xde0cffff and 0xde0d0000..0xde1fffff
D/TC:0 add_phys_mem:750 VCORE_FREE_PA type SEC_RAM_OVERALL 0xde0d0000 size 0x00130000
D/TC:0 merge_mmaps:703 Merging 0xde076000..0xde0cffff and 0xde0d0000..0xde1fffff
D/TC:0 add_phys_mem:750 next_pa type SEC_RAM_OVERALL 0xde200000 size 0x01e00000
D/TC:0 merge_mmaps:703 Merging 0xde076000..0xde1fffff and 0xde200000..0xdfffffff
D/TC:0 add_phys_mem:750 GIC_BASE type IO_SEC 0xa0000000 size 0x00200000
...
D/TC:0 add_phys_mem:750 APB1_BASE type IO_NSEC 0x40000000 size 0x00200000
D/TC:0 add_va_space:794 type RES_VASPACE size 0x00a00000
D/TC:0 add_va_space:794 type SHM_VASPACE size 0x02000000
D/TC:0 dump_mmap_table:920 type SEC_RAM_OVERALL va 0xd76ec000..0xd9675fff pa 0xde076000..0xdfffffff size 0x01f8a000 (pgdir)
D/TC:0 dump_mmap_table:920 type IO_SEC va 0xd9800000..0xd99fffff pa 0xa0000000..0xa01fffff size 0x00200000 (pgdir)
...
D/TC:0 dump_mmap_table:920 type IO_NSEC va 0xdb000000..0xdb1fffff pa 0x40000000..0x401fffff size 0x00200000 (pgdir)
D/TC:0 dump_mmap_table:920 type RES_VASPACE va 0xdb200000..0xdbbfffff pa 0x00000000..0x009fffff size 0x00a00000 (pgdir)
D/TC:0 dump_mmap_table:920 type SHM_VASPACE va 0xdbe00000..0xdddfffff pa 0x00000000..0x01ffffff size 0x02000000 (pgdir)
D/TC:0 dump_mmap_table:920 type TEE_RAM_RX va 0xde000000..0xde075fff pa 0xde000000..0xde075fff size 0x00076000 (smallpg)
D/TC:0 dump_mmap_table:920 type TEE_RAM_RW va 0xde076000..0xde1fffff pa 0xde076000..0xde1fffff size 0x0018a000 (smallpg)
D/TC:0 core_mmu_xlat_table_alloc:527 xlat tables used 1 / 4
D/TC:0 core_mmu_xlat_table_alloc:527 xlat tables used 2 / 4
D/TC:0 core_mmu_xlat_table_alloc:527 xlat tables used 3 / 4
D/TC:0 carve_out_core_mem:2654 0xde000000 .. 0xde200000
D/TC:0 boot_mem_release_unused:152 Allocated 0 bytes at va 0xde0d0000 pa 0xde0d0000
D/TC:0 boot_mem_release_unused:156 Tempalloc 4380 bytes at va 0xde1feee4
D/TC:0 boot_mem_release_unused:178 Carving out 0xde000000..0xde0cffff
D/TC:0 boot_mem_release_unused:187 Releasing 1236992 bytes from va 0xde0d0000
I/TC:
I/TC: Embedded DTB found
I/TC: OP-TEE version: 4.3.0-141-g0e3fb1edc (gcc version 11.3.1 20220712 (Arm GNU Toolchain 11.3.Rel1)) #36 Thu Sep 19 07:43:17 UTC 2024 arm
...
I/TC: Primary CPU initializing
...
D/TC:0 0 boot_mem_release_tmp_alloc:227 Releasing 8192 bytes from va 0xde1fe000
...
I/TC: Primary CPU switching to normal world boot
E/TC:0 0
E/TC:0 0 Core data-abort at address 0xde0d0000 (translation fault)
E/TC:0 0 fsr 0x00002a07 ttbr0 0xde0ba000 ttbr1 0x00000000 cidr 0x0
E/TC:0 0 cpu #0 cpsr 0x800001d3
E/TC:0 0 r0 0xde0d0000 r4 0x2fff2d84 r8 0xde078828 r12 0x00000000
E/TC:0 0 r1 0xde1ff018 r5 0x00000000 r9 0xde0cf5c0 sp 0xde0cf5c0
E/TC:0 0 r2 0x00000040 r6 0xc0400000 r10 0xafaeadac lr 0xde00021c
E/TC:0 0 r3 0x0000003f r7 0x00000000 r11 0x00000000 pc 0xde000d24
E/TC:0 0 TEE load address @ 0xde000000
E/TC:0 0 Call stack:
E/TC:0 0 0xde000d24 dcache_cleaninv_range at core/arch/arm/kernel/cache_helpers_a32.S:53
E/TC:0 0 0xde00021c clear_bss at core/arch/arm/kernel/entry_a32.S:588
The data-abort occurs in reset_primary before switch to non-secure world, fluhsing vmem that is not mapped:
core/arch/arm/kernel/entry_a32.S:
/*
* In case we've touched memory that secondary CPUs will use before
* they have turned on their D-cache, clean and invalidate the
* D-cache before exiting to normal world.
*/
588: flush_cache_vrange(cached_mem_start, cached_mem_end)Obviously not reproduced on vexpress-qemu_virt.
| */ | ||
| if (IS_ENABLED(CFG_NS_VIRTUALIZATION)) { | ||
| paddr_t b1 = 0; | ||
| paddr_size_t s1 = 0; | ||
|
|
There was a problem hiding this comment.
Commit "core: merge core_mmu_init_phys_mem() and core_mmu_init_virtualization()":
- Intentionally removed the info level trace
IMSG("Initializing virtualization support");?
(previously printed from core/arch/arm/kernel/boot.c) - inline comment above
... TA are loaded/executedNsec ...is pasted from previous implementation but is is quite buggy. I think it's time to fix it.
There was a problem hiding this comment.
Commit "core: merge core_mmu_init_phys_mem() and core_mmu_init_virtualization()":
- Intentionally removed the info level trace
IMSG("Initializing virtualization support");?
(previously printed from core/arch/arm/kernel/boot.c)
Yes, I saw no point in keeping it.
- inline comment above
... TA are loaded/executedNsec ...is pasted from previous implementation but is is quite buggy. I think it's time to fix it.
Right, I'll remove it.
There was a problem hiding this comment.
Commit "core: merge core_mmu_init_phys_mem() and core_mmu_init_virtualization()":
- Intentionally removed the info level trace
IMSG("Initializing virtualization support");?
(previously printed from core/arch/arm/kernel/boot.c)Yes, I saw no point in keeping it.
Do we have another way of knowing that the OP-TEE binary was compiled with virtualization support? Linux driver traces perhaps?
There was a problem hiding this comment.
I've added "core: arm: add CFG_NS_VIRTUALIZATION boot log" to take care of that.
core/arch/arm/arm.mk
Outdated
| @@ -15,6 +15,10 @@ include mk/cc-option.mk | |||
| # Only applicable when paging is enabled. | |||
| CFG_CORE_TZSRAM_EMUL_SIZE ?= 458752 | |||
|
|
|||
| # Mandatory for arch/arm to avoid ifdefs in the arch specific code | |||
| CFG_BOOT_MEM=y | |||
There was a problem hiding this comment.
This one is needed or we'll get an error:
core/arch/arm/arm.mk:19: *** CFG_BOOT_MEM is set to 'n' (from file) but its value must be 'y'. Stop.
This is due to the default value assigned in mk/config.mk
There was a problem hiding this comment.
I think we should rather have in mk/config.mk something like:
ifeq ($(ARCH),arm)
$(call force,CFG_BOOT_MEM,y)
else
CFG_BOOT_MEM ?= n
endif
core/mm/core_mmu.c
Outdated
| #ifdef TEE_SDP_TEST_MEM_BASE | ||
| if (TEE_SDP_TEST_MEM_SIZE) { | ||
| pa = vaddr_to_phys(TEE_SDP_TEST_MEM_BASE); | ||
| carve_out_core_mem(pa, pa + TEE_SDP_TEST_MEM_BASE); |
There was a problem hiding this comment.
s/TEE_SDP_TEST_MEM_BASE/TEE_SDP_TEST_MEM_SIZE/
core/mm/core_mmu.c
Outdated
| { | ||
| tee_mm_entry_t *mm __maybe_unused = NULL; | ||
|
|
||
| DMSG("%#"PRIxPA" .. %#"PRIxPASZ, pa, end_pa); |
|
To sumup, the cache flush operation at entry_a32.S:588 above flushes a VA range (TEE RAM) that includes VA areas not mapped (released boot-mem areas IIUC), they trigger a data-abort exception in the CPU. Testing a bit, it seems Qemu (at least the tool/config we use) does not trap such accesses as illegals. |
Thanks for the detailed analysis, I hope the update fixes the problem. |
|
Update, comments addressed. |
|
For "core: arm: add CFG_NS_VIRTUALIZATION boot log":
|
core/arch/arm/kernel/boot.c
Outdated
| @@ -934,7 +936,15 @@ static void init_primary(unsigned long pageable_part, unsigned long nsec_entry) | |||
|
|
|||
| core_mmu_save_mem_map(); | |||
| core_mmu_init_phys_mem(); | |||
| boot_mem_release_unused(); | |||
| va = boot_mem_release_unused(); | |||
| if (IS_ENABLED(CFG_ARM32_core) && !IS_ENABLED(CFG_WITH_PAGER)) { | |||
There was a problem hiding this comment.
Sorry, my comment was maybe ambiguous. Don't you expect the same kind of issues on 64 bit platforms?
I've tested on my 64bi platform and indeed the same issue arises: data abort on the data cache flush at core/arch/arm/kernel/entry_a64.S:432 on the VA range unmapped from boot_mem.c.
There was a problem hiding this comment.
You're right, I've pushed an update for that.
core/arch/arm/kernel/boot.c
Outdated
| @@ -540,17 +536,17 @@ static void init_runtime(unsigned long pageable_part) | |||
| asan_memcpy_unchecked(hashes, tmp_hashes, hash_size); | |||
|
|
|||
| /* | |||
| * Need physical memory pool initialized to be able to allocate | |||
| * secure physical memory below. | |||
| * The pager is about the be enabled below, evenual temporary boot | |||
There was a problem hiding this comment.
Thanks, I'll fix.
| __vcore_free_start)) | ||
| #define VCORE_FREE_END_PA ((unsigned long)__vcore_free_end) | ||
| #else | ||
| #define VCORE_FREE_PA PADDR_MAX |
There was a problem hiding this comment.
Add an explicit inline comment like /* Pager consumes all core physical memory: no free area */?
There was a problem hiding this comment.
Sure, I'll add a comment.
|
Update. Let me know if I should squash in the updates. |
|
I'm fine you squash the fixup commits. |
jenswi-linaro
force-pushed
the
boot-mem
branch
from
13036b5
to
45c5964
Compare
|
Squashed in the updates. |
| case MEM_AREA_NEX_RAM_RW: | ||
| case MEM_AREA_TEE_ASAN: | ||
| case MEM_AREA_TA_RAM: | ||
| case MEM_AREA_SEC_RAM_OVERALL: |
There was a problem hiding this comment.
Asan memory is still referenced with MEM_AREA_TEE_ASAN memory area ID when CFG_WITH_PAGER is enabled. Won't this switch case need to be supported?
MEM_AREA_TEE_RAM is still used when CFG_CORE_RWDATA_NOEXEC is disabled. IS that an issue? Maybe it's a bit inconsistent to enable CFG_MEMTAG but not CFG_CORE_RWDATA_NOEXEC.
Is is OK here to no more clear memtags for MEM_AREA_TEE_RAM_RW and MEM_AREA_NEX_RAM_RW areas?
There was a problem hiding this comment.
Asan memory is still referenced with
MEM_AREA_TEE_ASANmemory area ID whenCFG_WITH_PAGERis enabled. Won't this switch case need to be supported?
mk/config.mk:999: *** CFG_CORE_SANITIZE_KADDRESS and CFG_MEMTAG are not compatible. Stop.
I don't think we need to care about MEM_AREA_TEE_ASAN since it's not used with memory tagging enabled.
MEM_AREA_TEE_RAMis still used whenCFG_CORE_RWDATA_NOEXECis disabled. IS that an issue? Maybe it's a bit inconsistent to enableCFG_MEMTAGbut notCFG_CORE_RWDATA_NOEXEC.
That shouldn't be a problem, everything should be covered by MEM_AREA_SEC_RAM_OVERALL nonetheless.
Is is OK here to no more clear memtags for
MEM_AREA_TEE_RAM_RWandMEM_AREA_NEX_RAM_RWareas?
The same here, is covered by MEM_AREA_SEC_RAM_OVERALL.
However, while looking at logs etc to double-check I noticed:
D/TC:0 dump_mmap_table:923 type SEC_RAM_OVERALL va 0x12100000..0x122fffff pa 0x0e100000..0x0e2fffff size 0x00200000 (pgdir)
...
D/TC:0 mmap_clear_memtag:392 Clearing tags for VA 0x12000000..0x121fffff
We may have this error in more than one place, see commit f01690c ("core: fix mapping init debug trace") for the dump_mmap_table() print. I'll investigate further.
core/mm/core_mmu.c
Outdated
| @@ -66,7 +66,7 @@ static struct tee_mmap_region static_mmap_regions[CFG_MMAP_REGIONS | |||
| #if defined(CFG_CORE_ASLR) || defined(CFG_CORE_PHYS_RELOCATABLE) | |||
| + 1 | |||
| #endif | |||
| + 2] __nex_bss; | |||
| + 3] __nex_bss; | |||
There was a problem hiding this comment.
(commit "core: mm,pager: map remaining physical memory")
Since this is needed only when pager is enabled, I suggest an explicit case like:
#if defined(CFG_CORE_ASLR) || defined(CFG_CORE_PHYS_RELOCATABLE)
+ 1
#endif
#ifdef CFG_WITH_PAGER
/* Entry to cover the whole physical RAM */
+ 1
#endif
2] __nex_bss;| */ | ||
| ADD_PHYS_MEM(MEM_AREA_TEE_RAM, ram_start, | ||
| VCORE_UNPG_RX_PA - ram_start); | ||
| ADD_PHYS_MEM(MEM_AREA_SEC_RAM_OVERALL, ram_start, |
There was a problem hiding this comment.
This changes to mapping of the physical memory range from TEE_RAM_START to TEE_LOAD_ADDR, from read-only/tagged to read-write/tagged.
I see that some platforms (at least plat-ti with pager enabled) has a 4k page between TEE_RAM_START and TEE_LOAD_ADDR. Most platforms default config also allow to define a CFG_TEE_LOAD_ADDR value different from TEE_RAM_START.
Maybe to discuss with @glneo.
There was a problem hiding this comment.
Would it make more sense only to map the gap between TEE_RAM_START and TEE_LOAD_ADDR in MEM_AREA_SEC_RAM_OVERALL?
There was a problem hiding this comment.
I think it would, unless some platforms load a TEE core image made on core binary image prepend with some data expected to be read-only mapped. Could be strange but current implementation allows that.
Maybe this old scheme could be deprecated. In that case, at least the commit message should explicitly state it IMHO.
core/mm/core_mmu.c
Outdated
| @@ -66,7 +66,7 @@ static struct tee_mmap_region static_mmap_regions[CFG_MMAP_REGIONS | |||
| #if defined(CFG_CORE_ASLR) || defined(CFG_CORE_PHYS_RELOCATABLE) | |||
| + 1 | |||
| #endif | |||
| + 1] __nex_bss; | |||
| + 2] __nex_bss; | |||
There was a problem hiding this comment.
This change adds 1, 3 or 4 cells in static_mmap_regions (depending on CFG_NS_VIRTUALIZATION and CFG_CORE_RWDATA_NOEXEC).
Should this be reflected here? Or maybe add 4 cells, unconditionally to cover all cases? Or use explicit config cases (below, not that very helpful I guess)?
static struct tee_mmap_region static_mmap_regions[CFG_MMAP_REGIONS
#if defined(CFG_CORE_ASLR) || defined(CFG_CORE_PHYS_RELOCATABLE)
+ 1
#endif
+#ifdef CFG_CORE_RWDATA_NOEXEC
+#ifdef CFG_NS_VIRTUALIZATION
+ + 4
+#else
+ + 3
+#endif
+#else
+ + 1
+#endif /*CFG_CORE_RWDATA_NOEXEC*/
+ 1] __nex_bss;There was a problem hiding this comment.
Yeah, let's do that to keep things simple.
There was a problem hiding this comment.
Few review tags: Reviewed-by: Etienne Carriere <[email protected]> for commits:
"mk/clang.mk: -Wno-gnu-alignof-expression",
"core: arm64: increase thread stack size for debug",
"core: mm: add vaddr_to_phys()",
"core: remove CORE_MEM_TA_RAM",
"core: add VCORE_FREE_{PA,SZ,END_PA}",
"core: mm: allow unmapping VCORE_FREE",
"core: mm: unify secure core and TA memory",
"core: virt: phys_mem_core_alloc() use both pools",
"core: arm: core_mmu_v7.c: increase MAX_XLAT_TABLES by 2" and
"core: arm,pager: make __vcore_init_ro_start follow __vcore_init_rx_end"
A concern with commit "core: mm: map memory using requested block size".
| panic("Impossible memory alignment"); | ||
|
|
||
| if (map_is_tee_ram(mem_map->map + n)) | ||
| mem_map->map[n].region_size = SMALL_PAGE_SIZE; | ||
| else | ||
| mem_map->map[n].region_size = CORE_MMU_PGDIR_SIZE; |
There was a problem hiding this comment.
Previous implementation allowed to page map cacheable memory outside TEE RAM, (MEM_AREA_RAM_NSEC, MEM_AREA_RAM_SEC, MEM_AREA_ROM_SEC). I fear memory firewall permission violation on speculative accesses for such memories, and likely few other corner issues on some platform specific memories.
There was a problem hiding this comment.
This is the requested granularity. It will not mean mapping more memory than specified. So some ranges will be mapped by of mix of small-page and pgdir sized mappings (level 3 and level 2 with LPAE).
Take QEMU with hafnium as an example:
TEE_RAM_RWX va 0x0e304000..0x0e503fff pa 0x0e304000..0x0e503fff size 0x00200000 (smallpg)
RES_VASPACE va 0x0e600000..0x0effffff pa 0x00000000..0x009fffff size 0x00a00000 (pgdir)
SHM_VASPACE va 0x0f000000..0x10ffffff pa 0x00000000..0x01ffffff size 0x02000000 (pgdir)
IO_SEC va 0x11000000..0x1a080fff pa 0x09000000..0x12080fff size 0x09081000 (pgdir)
SEC_RAM_OVERALL va 0x1a300000..0x1affffff pa 0x0e300000..0x0effffff size 0x00d00000 (pgdir)
IO_SEC is then mapped as:
D/TC:0 [LVL2] VA:0x0011000000 PA:0x0009000000 DEV-RW-XN-S
D/TC:0 [LVL2] VA:0x0011200000 PA:0x0009200000 DEV-RW-XN-S
...
D/TC:0 [LVL2] VA:0x001a000000 TBL:0x000e3fa000
D/TC:0 [LVL3] VA:0x001a000000 PA:0x0012000000 DEV-RW-XN-S
D/TC:0 [LVL3] VA:0x001a001000 PA:0x0012001000 DEV-RW-XN-S
...
D/TC:0 [LVL3] VA:0x001a080000 PA:0x0012080000 DEV-RW-XN-S
D/TC:0 [LVL3] VA:0x001a081000 INVALID
SEC_RAM_OVERALL as:
D/TC:0 [LVL3] VA:0x001a300000 PA:0x000e300000 MEM-RW-XN-S
D/TC:0 [LVL3] VA:0x001a301000 PA:0x000e301000 MEM-RW-XN-S
...
D/TC:0 [LVL3] VA:0x001a3ff000 PA:0x000e3ff000 MEM-RW-XN-S
D/TC:0 [LVL2] VA:0x001a400000 PA:0x000e400000 MEM-RW-XN-S
...
D/TC:0 [LVL2] VA:0x001ae00000 PA:0x000ee00000 MEM-RW-XN-S
D/TC:0 [LVL2] VA:0x001b000000 INVALID
So pgdir-sized region_size uses as few translation tables as possible, but it will not result in mapping beyond the requested range.
|
Can I squash in the updates? |
Fine with me. |
Add -Wno-gnu-alignof-expression to the warnings flag for Clang in order to avoid warnings like: '_Alignof' applied to an expression is a GNU extension [-Werror,-Wgnu-alignof-expression] when alignof() is applied on an expression like dereferencing a pointer to get the alignment of type. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Jerome Forissier <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
Increase STACK_THREAD_SIZE when CFG_CORE_DEBUG_CHECK_STACKS=y. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Jerome Forissier <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
Add a wrapper function for virt_to_phys() using vaddr_t instead of a void pointer. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Jerome Forissier <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
The buffer attribute CORE_MEM_TA_RAM isn't used to query the status of a buffer anywhere. So remove the attribute to allow future simplifications. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Jerome Forissier <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
Add VCORE_FREE_{PA,SZ,END_PA} defines to identify the unused and free
memory range at the end of TEE_RAM_START..(TEE_RAM_START +
TEE_RAM_VA_SIZE).
VCORE_FREE_SZ is 0 in a pager configuration since all the memory is
used by the pager.
The VCORE_FREE range is excluded from the TEE_RAM_RW area for
CFG_NS_VIRTUALIZATION=y and instead put in a separate NEX_RAM_RW area.
This makes each partition use a bit less memory and leaves the
VCORE_FREE range available for the Nexus.
The VCORE_FREE range is added to the TEE_RAM_RW area for the normal
configuration with CFG_NS_VIRTUALIZATION=n and CFG_WITH_PAGER=n. It's in
practice unchanged behaviour in this configuration.
Signed-off-by: Jens Wiklander <[email protected]>
Acked-by: Jerome Forissier <[email protected]>
Reviewed-by: Etienne Carriere <[email protected]>
Allow unmapping core memory in the VCORE_FREE range when the original boot mapping isn't needed any more. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Jerome Forissier <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
Replace MEM_AREA_TA_RAM with MEM_AREA_SEC_RAM_OVERALL. All read/write secure memory is covered by MEM_AREA_SEC_RAM_OVERALL, sometimes using an aliased map. But secure read-only or execute core memory is not covered as that would defeat the purpose of CFG_CORE_RWDATA_NOEXEC. Since the partition TA memory isn't accessed via MEM_AREA_TA_RAM any longer, don't map it using the partition specific map. This is needed later where unification of OP-TEE core and physical TA memory is possible. Signed-off-by: Jens Wiklander <[email protected]>
In configurations where secure core and TA memory is allocated from the same contiguous physical memory block, carve out the memory needed by OP-TEE core and make the rest available as TA memory. This is needed by later patches where more core memory is allocated as needed from the pool of TA memory. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
With CFG_NS_VIRTUALIZATION=y let phys_mem_core_alloc() allocate from both the core_pool and ta_pool since both pools keep equally secure memory. This is needed in later patches when some translation tables are dynamically allocated from spare physical core memory. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
Increase MAX_XLAT_TABLES by 2 to be able to map all TEE memory with 4k pages. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
TEE memory is always supposed to be mapped with 4k pages for maximum flexibility, but can_map_at_level() doesn't check the requested block size for a region, so fix that. However, assign_mem_granularity() assigns smaller than necessary block sizes on page aligned regions, so fix that by only requesting 4k granularity for TEE memory and PGDIR granularity for the rest. This is needed in later patches where some TEE memory is unmapped. Signed-off-by: Jens Wiklander <[email protected]>
This concerns configurations with CFG_WITH_PAGER=y. Until this patch, even if __vcore_init_ro_size (VCORE_INIT_RO_SZ) is 0 for CFG_CORE_RODATA_NOEXEC=n, __vcore_init_ro_start was using some value smaller than __vcore_init_rx_end. To simplify code trying to find the end of VCORE_INIT_RX and VCORE_INIT_RO parts of the binary, make sure that __vcore_init_ro_start follows right after __vcore_init_rx_end. Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Etienne Carriere <[email protected]>
For CFG_WITH_PAGER=y map the remaining memory following the VCORE_INIT_RO memory to make sure that all physical TEE memory is mapped even if VCORE_INIT_RO doesn't cover it entirely. This will be used in later patches to use the temporarily unused memory while booting. Signed-off-by: Jens Wiklander <[email protected]>
Adds CFG_BOOT_MEM to support stack-like memory allocations during boot before a heap has been configured. Signed-off-by: Jens Wiklander <[email protected]>
Add boot_cached_mem_end in C code, replacing the previous read-only mapped cached_mem_end. This allows updates to boot_cached_mem_end after MMU has been enabled. Signed-off-by: Jens Wiklander <[email protected]>
Enable CFG_BOOT_MEM unconditionally and call the boot_mem_*() functions as needed from entry_*.S and boot.c. The pager will reuse all boot_mem memory internally when configured. The non-pager configuration will unmap the memory and make it available for TAs if needed. __FLATMAP_PAGER_TRAILING_SPACE is removed from the link script, collect_mem_ranges() in core/mm/core_mmu.c maps the memory following VCORE_INIT_RO automatically. Signed-off-by: Jens Wiklander <[email protected]>
With CFG_BOOT_MEM enabled, allocate a temporary memory map array using boot_mem_alloc_tmp() instead of using the global static_mmap_regions[]. core_mmu_save_mem_map() is added and called from boot_init_primary_late() before the temporary memory is reused. Signed-off-by: Jens Wiklander <[email protected]>
Initialize guest physical memory in virt_guest_created() before the first entry into the guest from normal world. This replaces the call to core_mmu_init_phys_mem() in init_tee_runtime(). Remove unused code in core_mmu_init_phys_mem() and the now unused functions core_mmu_get_ta_range() and virt_get_ta_ram(). Signed-off-by: Jens Wiklander <[email protected]>
Moves the implementation of core_mmu_init_virtualization() into core_mmu_init_phys_mem(). This simplifies init_primary() in core/arch/arm/kernel/boot.c. Signed-off-by: Jens Wiklander <[email protected]>
Add a log entry when CFG_NS_VIRTUALIZATION is enabled, for example: D/TC:0 0 boot_init_primary_late:1028 NS-Virtualization enabled, supporting 2 guests Signed-off-by: Jens Wiklander <[email protected]> Reviewed-by: Jerome Forissier <[email protected]>
dump_mmap_table() is updated to use map->va directly as it's always the start of the mapping regardless of map->region_size. Signed-off-by: Jens Wiklander <[email protected]>
Add missing ADD_PHYS_MEM(MEM_AREA_SEC_RAM_OVERALL,...) for CFG_CORE_RWDATA_NOEXEC=n. Signed-off-by: Jens Wiklander <[email protected]>
Addressing a comment. Signed-off-by: Jens Wiklander <[email protected]>
Remove the read-only mapping between TEE_RAM_START and TEE_LOAD_ADDR. Signed-off-by: Jens Wiklander <[email protected]>
jenswi-linaro
force-pushed
the
boot-mem
branch
from
45c5964
to
5cc4816
Compare
|
Tags applied |
|
This pull request has been marked as a stale pull request because it has been open (more than) 30 days with no activity. Remove the stale label or add a comment, otherwise this pull request will automatically be closed in 5 days. Note, that you can always re-open a closed issue at any time. |
|
Should I squash and rebase to continue the review? |
This pull request handles previously unused memory during boot—currently mostly temporary memory allocations, but that will be extended in the next pull requests.
The commit "core: mm: allocate temporary memory map array" demonstrates this by using dynamic allocation instead of hard-coded allocation of
static_mmap_regions[].