Message ID | 20190731153857.4045-1-pasha.tatashin@soleen.com (mailing list archive) |
---|---|
Headers | show |
Series | arm64: MMU enabled kexec relocation | expand |
Hi Pavel, Generally, the cover letter should state up-front what the goal is (or what problem you're trying to solve). It would be really helpful to have that so that we understand what you're trying to achieve, and why. Messing with the MMU is often fraught with danger (and very painful to debug, as you are now aware), and so far we've tried to minimize the number of places where we have to do so. On Wed, Jul 31, 2019 at 11:38:49AM -0400, Pavel Tatashin wrote: > Changelog from previous RFC: > - Added trans_table support for both hibernate and kexec. > - Fixed performance issue, where enabling MMU did not yield the > actual performance improvement. > > Bug: > With the current state, this patch series works on kernels booted with EL1 > mode, but for some reason, when elevated to EL2 mode reboot freezes in > both QEMU and on real hardware. > > The freeze happens in: > > arch/arm64/kernel/relocate_kernel.S > turn_on_mmu() > > Right after sctlr_el2 is written (MMU on EL2 is enabled) > > msr sctlr_el2, \tmp1 > > I've been studying all the relevant control registers for EL2, but do not > see what might be causing this hang: > > MAIR_EL2 is set to be exactly the same as MAIR_EL1 0xbbff440c0400 > > TCR_EL2 0x80843510 > Enabled bits: > PS Physical Address Size. (0b100 44 bits, 16TB.) > SH0 Shareability 11 Inner Shareable > ORGN0 Normal memory, Outer Write-Back Read-Allocate Write-Allocate Cach. > IRGN0 Normal memory, Inner Write-Back Read-Allocate Write-Allocate Cach. > T0SZ 01 0000 > > SCTLR_EL2 0x30e5183f > RES1 : Reserve ones > M : MMU enabled > A : Align check > C : Cacheability control > SA : SP Alignment check enable > IESB : Implicit Error Synchronization event > I : Instruction access Cacheability > > TTBR0_EL2 0x1b3069000 (address of trans_table) > > Any suggestion of what else might be missing that causes this freeze when > MMU is enabled in EL2? > > ===== > Here is the current data from the real hardware: > (because of bug, I forced EL1 mode by setting el2_switch always to zero in > cpu_soft_restart()): > > For this experiment, the size of kernel plus initramfs is 25M. If initramfs > was larger, than the improvements would be even greater, as time spent in > relocation is proportional to the size of relocation. > > Previously: > kernel shutdown 0.022131328s > relocation 0.440510736s > kernel startup 0.294706768s In total this takes ~0.76s... > > Relocation was taking: 58.2% of reboot time > > Now: > kernel shutdown 0.032066576s > relocation 0.022158152s > kernel startup 0.296055880s ... and this takes ~0.35s So do we really need this complexity for a few blinks of an eye? Thanks, Mark.
On Wed, Jul 31, 2019 at 12:33 PM Mark Rutland <mark.rutland@arm.com> wrote: > > Hi Pavel, > > Generally, the cover letter should state up-front what the goal is (or > what problem you're trying to solve). It would be really helpful to have > that so that we understand what you're trying to achieve, and why. > > Messing with the MMU is often fraught with danger (and very painful to > debug, as you are now aware), and so far we've tried to minimize the > number of places where we have to do so. Hi Mark, I understand, this is why I first went another route of solving this problem: pre-reserving contiguous memory, and avoid relocation entirely (the same as what happens during crash reboot). But, that solution was not accepted because it introduces a change to the common code to solve ARM specific problem. So, James Morse, and other suggested that I take a look at the root of the problem, and enable MMU during relocation by doing what is already done during hibernate restore. > > On Wed, Jul 31, 2019 at 11:38:49AM -0400, Pavel Tatashin wrote: > > Changelog from previous RFC: > > - Added trans_table support for both hibernate and kexec. > > - Fixed performance issue, where enabling MMU did not yield the > > actual performance improvement. > > > > Bug: > > With the current state, this patch series works on kernels booted with EL1 > > mode, but for some reason, when elevated to EL2 mode reboot freezes in > > both QEMU and on real hardware. > > > > The freeze happens in: > > > > arch/arm64/kernel/relocate_kernel.S > > turn_on_mmu() > > > > Right after sctlr_el2 is written (MMU on EL2 is enabled) > > > > msr sctlr_el2, \tmp1 > > > > I've been studying all the relevant control registers for EL2, but do not > > see what might be causing this hang: > > > > MAIR_EL2 is set to be exactly the same as MAIR_EL1 0xbbff440c0400 > > > > TCR_EL2 0x80843510 > > Enabled bits: > > PS Physical Address Size. (0b100 44 bits, 16TB.) > > SH0 Shareability 11 Inner Shareable > > ORGN0 Normal memory, Outer Write-Back Read-Allocate Write-Allocate Cach. > > IRGN0 Normal memory, Inner Write-Back Read-Allocate Write-Allocate Cach. > > T0SZ 01 0000 > > > > SCTLR_EL2 0x30e5183f > > RES1 : Reserve ones > > M : MMU enabled > > A : Align check > > C : Cacheability control > > SA : SP Alignment check enable > > IESB : Implicit Error Synchronization event > > I : Instruction access Cacheability > > > > TTBR0_EL2 0x1b3069000 (address of trans_table) > > > > Any suggestion of what else might be missing that causes this freeze when > > MMU is enabled in EL2? > > > > ===== > > > Here is the current data from the real hardware: > > (because of bug, I forced EL1 mode by setting el2_switch always to zero in > > cpu_soft_restart()): > > > > For this experiment, the size of kernel plus initramfs is 25M. If initramfs > > was larger, than the improvements would be even greater, as time spent in > > relocation is proportional to the size of relocation. > > > > Previously: > > kernel shutdown 0.022131328s > > relocation 0.440510736s > > kernel startup 0.294706768s > > In total this takes ~0.76s... > > > > > Relocation was taking: 58.2% of reboot time > > > > Now: > > kernel shutdown 0.032066576s > > relocation 0.022158152s > > kernel startup 0.296055880s > > ... and this takes ~0.35s > > So do we really need this complexity for a few blinks of an eye? Yes, we have an extremely tight reboot budget, 0.35s is not an acceptable waste. > > Thanks, > Mark.
On Wed, Jul 31, 2019 at 12:40:51PM -0400, Pavel Tatashin wrote: > On Wed, Jul 31, 2019 at 12:33 PM Mark Rutland <mark.rutland@arm.com> wrote: > > > > Hi Pavel, > > > > Generally, the cover letter should state up-front what the goal is (or > > what problem you're trying to solve). It would be really helpful to have > > that so that we understand what you're trying to achieve, and why. [...] > > > Here is the current data from the real hardware: > > > (because of bug, I forced EL1 mode by setting el2_switch always to zero in > > > cpu_soft_restart()): > > > > > > For this experiment, the size of kernel plus initramfs is 25M. If initramfs > > > was larger, than the improvements would be even greater, as time spent in > > > relocation is proportional to the size of relocation. > > > > > > Previously: > > > kernel shutdown 0.022131328s > > > relocation 0.440510736s > > > kernel startup 0.294706768s > > > > In total this takes ~0.76s... > > > > > > > > Relocation was taking: 58.2% of reboot time > > > > > > Now: > > > kernel shutdown 0.032066576s > > > relocation 0.022158152s > > > kernel startup 0.296055880s > > > > ... and this takes ~0.35s > > > > So do we really need this complexity for a few blinks of an eye? > > Yes, we have an extremely tight reboot budget, 0.35s is not an acceptable waste. Could you please elaborate on your use-case? Understanfin what you're trying to achieve would help us to understand which solutions make sense. Thanks, Mark.
On Wed, Jul 31, 2019 at 12:50 PM Mark Rutland <mark.rutland@arm.com> wrote: > > On Wed, Jul 31, 2019 at 12:40:51PM -0400, Pavel Tatashin wrote: > > On Wed, Jul 31, 2019 at 12:33 PM Mark Rutland <mark.rutland@arm.com> wrote: > > > > > > Hi Pavel, > > > > > > Generally, the cover letter should state up-front what the goal is (or > > > what problem you're trying to solve). It would be really helpful to have > > > that so that we understand what you're trying to achieve, and why. > > [...] > > > > > Here is the current data from the real hardware: > > > > (because of bug, I forced EL1 mode by setting el2_switch always to zero in > > > > cpu_soft_restart()): > > > > > > > > For this experiment, the size of kernel plus initramfs is 25M. If initramfs > > > > was larger, than the improvements would be even greater, as time spent in > > > > relocation is proportional to the size of relocation. > > > > > > > > Previously: > > > > kernel shutdown 0.022131328s > > > > relocation 0.440510736s > > > > kernel startup 0.294706768s > > > > > > In total this takes ~0.76s... > > > > > > > > > > > Relocation was taking: 58.2% of reboot time > > > > > > > > Now: > > > > kernel shutdown 0.032066576s > > > > relocation 0.022158152s > > > > kernel startup 0.296055880s > > > > > > ... and this takes ~0.35s > > > > > > So do we really need this complexity for a few blinks of an eye? > > > > Yes, we have an extremely tight reboot budget, 0.35s is not an acceptable waste. > > Could you please elaborate on your use-case? > > Understanfin what you're trying to achieve would help us to understand > which solutions make sense. An extremely high availability device with an update story utilizing kexec functionality for a faster kernel update and also for being able to preserve some state in memory without wasting the time of copying it to and from a backing storage. We at Microsoft will be using a fleet of these devices. The total reboot budget is less than half a second, out of which 0.44s is currently spent in kexec relocation. Pasha > > Thanks, > Mark.
I will send a new version soon, so please do not spend time reviewing this work. In the new version I will fix MMU at EL2 issue by doing what we are doing in hibernation: reduce to EL1 to do the copying, and escalate back to to EL2 to branch to new kernel. Also, this will simplify copying function by actually doing the linear copy as ttbr1 and ttbr0 are always available this way. Thank you, Pasha On Wed, Jul 31, 2019 at 11:38 AM Pavel Tatashin <pasha.tatashin@soleen.com> wrote: > > Changelog from previous RFC: > - Added trans_table support for both hibernate and kexec. > - Fixed performance issue, where enabling MMU did not yield the > actual performance improvement. > > Bug: > With the current state, this patch series works on kernels booted with EL1 > mode, but for some reason, when elevated to EL2 mode reboot freezes in > both QEMU and on real hardware. > > The freeze happens in: > > arch/arm64/kernel/relocate_kernel.S > turn_on_mmu() > > Right after sctlr_el2 is written (MMU on EL2 is enabled) > > msr sctlr_el2, \tmp1 > > I've been studying all the relevant control registers for EL2, but do not > see what might be causing this hang: > > MAIR_EL2 is set to be exactly the same as MAIR_EL1 0xbbff440c0400 > > TCR_EL2 0x80843510 > Enabled bits: > PS Physical Address Size. (0b100 44 bits, 16TB.) > SH0 Shareability 11 Inner Shareable > ORGN0 Normal memory, Outer Write-Back Read-Allocate Write-Allocate Cach. > IRGN0 Normal memory, Inner Write-Back Read-Allocate Write-Allocate Cach. > T0SZ 01 0000 > > SCTLR_EL2 0x30e5183f > RES1 : Reserve ones > M : MMU enabled > A : Align check > C : Cacheability control > SA : SP Alignment check enable > IESB : Implicit Error Synchronization event > I : Instruction access Cacheability > > TTBR0_EL2 0x1b3069000 (address of trans_table) > > Any suggestion of what else might be missing that causes this freeze when > MMU is enabled in EL2? > > ===== > Here is the current data from the real hardware: > (because of bug, I forced EL1 mode by setting el2_switch always to zero in > cpu_soft_restart()): > > For this experiment, the size of kernel plus initramfs is 25M. If initramfs > was larger, than the improvements would be even greater, as time spent in > relocation is proportional to the size of relocation. > > Previously: > kernel shutdown 0.022131328s > relocation 0.440510736s > kernel startup 0.294706768s > > Relocation was taking: 58.2% of reboot time > > Now: > kernel shutdown 0.032066576s > relocation 0.022158152s > kernel startup 0.296055880s > > Now: Relocation takes 6.3% of reboot time > > Total reboot is x2.16 times faster. > > Previous approaches and discussions > ----------------------------------- > https://lore.kernel.org/lkml/20190709182014.16052-1-pasha.tatashin@soleen.com > reserve space for kexec to avoid relocation, involves changes to generic code > to optimize a problem that exists on arm64 only: > > https://lore.kernel.org/lkml/20190716165641.6990-1-pasha.tatashin@soleen.com > The first attempt to enable MMU, some bugs that prevented performance > improvement. The page tables unnecessary configured idmap for the whole > physical space. > > Pavel Tatashin (8): > kexec: quiet down kexec reboot > arm64, mm: transitional tables > arm64: hibernate: switch to transtional page tables. > kexec: add machine_kexec_post_load() > arm64, kexec: move relocation function setup and clean up > arm64, kexec: add expandable argument to relocation function > arm64, kexec: configure transitional page table for kexec > arm64, kexec: enable MMU during kexec relocation > > arch/arm64/Kconfig | 4 + > arch/arm64/include/asm/kexec.h | 24 ++- > arch/arm64/include/asm/pgtable-hwdef.h | 1 + > arch/arm64/include/asm/trans_table.h | 66 ++++++ > arch/arm64/kernel/asm-offsets.c | 10 + > arch/arm64/kernel/cpu-reset.S | 4 +- > arch/arm64/kernel/cpu-reset.h | 8 +- > arch/arm64/kernel/hibernate.c | 261 ++++++------------------ > arch/arm64/kernel/machine_kexec.c | 168 ++++++++++++--- > arch/arm64/kernel/relocate_kernel.S | 238 +++++++++++++++------- > arch/arm64/mm/Makefile | 1 + > arch/arm64/mm/trans_table.c | 272 +++++++++++++++++++++++++ > kernel/kexec.c | 4 + > kernel/kexec_core.c | 8 +- > kernel/kexec_file.c | 4 + > kernel/kexec_internal.h | 2 + > 16 files changed, 756 insertions(+), 319 deletions(-) > create mode 100644 arch/arm64/include/asm/trans_table.h > create mode 100644 arch/arm64/mm/trans_table.c > > -- > 2.22.0 >