qemu-devel.nongnu.org archive mirror
 help / color / mirror / Atom feed
From: alvise rigo <a.rigo@virtualopensystems.com>
To: "Alex Bennée" <alex.bennee@linaro.org>
Cc: MTTCG Devel <mttcg@listserver.greensocs.com>,
	Claudio Fontana <claudio.fontana@huawei.com>,
	QEMU Developers <qemu-devel@nongnu.org>,
	Paolo Bonzini <pbonzini@redhat.com>,
	Jani Kokkonen <jani.kokkonen@huawei.com>,
	VirtualOpenSystems Technical Team <tech@virtualopensystems.com>,
	Richard Henderson <rth@twiddle.net>
Subject: Re: [Qemu-devel] [RFC v7 13/16] softmmu: Add history of excl accesses
Date: Thu, 18 Feb 2016 15:17:18 +0100	[thread overview]
Message-ID: <CAH47eN2nLdXV4wkB28xCUXkX1jRv1N3OCtS6qdrzyGbB-tzVHA@mail.gmail.com> (raw)
In-Reply-To: <87bn7gy467.fsf@linaro.org>

On Tue, Feb 16, 2016 at 6:07 PM, Alex Bennée <alex.bennee@linaro.org> wrote:
>
> Alvise Rigo <a.rigo@virtualopensystems.com> writes:
>
>> Add a circular buffer to store the hw addresses used in the last
>> EXCLUSIVE_HISTORY_LEN exclusive accesses.
>>
>> When an address is pop'ed from the buffer, its page will be set as not
>> exclusive. In this way, we avoid:
>> - frequent set/unset of a page (causing frequent flushes as well)
>> - the possibility to forget the EXCL bit set.
>
> Why was this a possibility before? Shouldn't that be tackled in the
> patch that introduced it?

Yes and no. The problem happens for instance when a LL will not be
followed by the SC. In this situation, the flag will be set for the
page, but might remain set for the rest of the execution (unless a
complete LL/SC in performed later on in the same guest page).

>
>>
>> Suggested-by: Jani Kokkonen <jani.kokkonen@huawei.com>
>> Suggested-by: Claudio Fontana <claudio.fontana@huawei.com>
>> Signed-off-by: Alvise Rigo <a.rigo@virtualopensystems.com>
>> ---
>>  cputlb.c                | 29 +++++++++++++++++++----------
>>  exec.c                  | 19 +++++++++++++++++++
>>  include/qom/cpu.h       |  8 ++++++++
>>  softmmu_llsc_template.h |  1 +
>>  vl.c                    |  3 +++
>>  5 files changed, 50 insertions(+), 10 deletions(-)
>>
>> diff --git a/cputlb.c b/cputlb.c
>> index 06ce2da..f3c4d97 100644
>> --- a/cputlb.c
>> +++ b/cputlb.c
>> @@ -395,16 +395,6 @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
>>      env->tlb_v_table[mmu_idx][vidx] = *te;
>>      env->iotlb_v[mmu_idx][vidx] = env->iotlb[mmu_idx][index];
>>
>> -    if (unlikely(!(te->addr_write & TLB_MMIO) && (te->addr_write & TLB_EXCL))) {
>> -        /* We are removing an exclusive entry, set the page to dirty. This
>> -         * is not be necessary if the vCPU has performed both SC and LL. */
>> -        hwaddr hw_addr = (env->iotlb[mmu_idx][index].addr & TARGET_PAGE_MASK) +
>> -                                          (te->addr_write & TARGET_PAGE_MASK);
>> -        if (!cpu->ll_sc_context) {
>> -            cpu_physical_memory_unset_excl(hw_addr);
>> -        }
>> -    }
>> -
>
> Errm is this right? I got confused reviewing 8/16 because my final tree
> didn't have this code. I'm not sure the adding of history obviates the
> need to clear the exclusive flag?

We will clear it adding a new item to the history. When an entry is
added, the oldest is removed and cleaned, solving the problem
mentioned above.

Thank you,
alvise

>
>>      /* refill the tlb */
>>      env->iotlb[mmu_idx][index].addr = iotlb - vaddr;
>>      env->iotlb[mmu_idx][index].attrs = attrs;
>> @@ -517,6 +507,25 @@ static inline bool lookup_and_reset_cpus_ll_addr(hwaddr addr, hwaddr size)
>>      return ret;
>>  }
>>
>> +extern CPUExclusiveHistory excl_history;
>> +static inline void excl_history_put_addr(hwaddr addr)
>> +{
>> +    hwaddr last;
>> +
>> +    /* Calculate the index of the next exclusive address */
>> +    excl_history.last_idx = (excl_history.last_idx + 1) % excl_history.length;
>> +
>> +    last = excl_history.c_array[excl_history.last_idx];
>> +
>> +    /* Unset EXCL bit of the oldest entry */
>> +    if (last != EXCLUSIVE_RESET_ADDR) {
>> +        cpu_physical_memory_unset_excl(last);
>> +    }
>> +
>> +    /* Add a new address, overwriting the oldest one */
>> +    excl_history.c_array[excl_history.last_idx] = addr & TARGET_PAGE_MASK;
>> +}
>> +
>>  #define MMUSUFFIX _mmu
>>
>>  /* Generates LoadLink/StoreConditional helpers in softmmu_template.h */
>> diff --git a/exec.c b/exec.c
>> index 51f366d..2e123f1 100644
>> --- a/exec.c
>> +++ b/exec.c
>> @@ -177,6 +177,25 @@ struct CPUAddressSpace {
>>      MemoryListener tcg_as_listener;
>>  };
>>
>> +/* Exclusive memory support */
>> +CPUExclusiveHistory excl_history;
>> +void cpu_exclusive_history_init(void)
>> +{
>> +    /* Initialize exclusive history for atomic instruction handling. */
>> +    if (tcg_enabled()) {
>> +        g_assert(EXCLUSIVE_HISTORY_CPU_LEN * max_cpus <= UINT16_MAX);
>> +        excl_history.length = EXCLUSIVE_HISTORY_CPU_LEN * max_cpus;
>> +        excl_history.c_array = g_malloc(excl_history.length * sizeof(hwaddr));
>> +        memset(excl_history.c_array, -1, excl_history.length * sizeof(hwaddr));
>> +    }
>> +}
>> +
>> +void cpu_exclusive_history_free(void)
>> +{
>> +    if (tcg_enabled()) {
>> +        g_free(excl_history.c_array);
>> +    }
>> +}
>>  #endif
>>
>>  #if !defined(CONFIG_USER_ONLY)
>> diff --git a/include/qom/cpu.h b/include/qom/cpu.h
>> index 6f6c1c0..0452fd0 100644
>> --- a/include/qom/cpu.h
>> +++ b/include/qom/cpu.h
>> @@ -227,7 +227,15 @@ struct kvm_run;
>>  #define TB_JMP_CACHE_SIZE (1 << TB_JMP_CACHE_BITS)
>>
>>  /* Atomic insn translation TLB support. */
>> +typedef struct CPUExclusiveHistory {
>> +    uint16_t last_idx;           /* index of last insertion */
>> +    uint16_t length;             /* history's length, it depends on smp_cpus */
>> +    hwaddr *c_array;             /* history's circular array */
>> +} CPUExclusiveHistory;
>>  #define EXCLUSIVE_RESET_ADDR ULLONG_MAX
>> +#define EXCLUSIVE_HISTORY_CPU_LEN 256
>> +void cpu_exclusive_history_init(void);
>> +void cpu_exclusive_history_free(void);
>>
>>  /**
>>   * CPUState:
>> diff --git a/softmmu_llsc_template.h b/softmmu_llsc_template.h
>> index b4712ba..b4e7f9d 100644
>> --- a/softmmu_llsc_template.h
>> +++ b/softmmu_llsc_template.h
>> @@ -75,6 +75,7 @@ WORD_TYPE helper_ldlink_name(CPUArchState *env, target_ulong addr,
>>           * to request any flush. */
>>          if (!cpu_physical_memory_is_excl(hw_addr)) {
>>              cpu_physical_memory_set_excl(hw_addr);
>> +            excl_history_put_addr(hw_addr);
>>              CPU_FOREACH(cpu) {
>>                  if (current_cpu != cpu) {
>>                      tlb_flush(cpu, 1);
>> diff --git a/vl.c b/vl.c
>> index f043009..b22d99b 100644
>> --- a/vl.c
>> +++ b/vl.c
>> @@ -547,6 +547,7 @@ static void res_free(void)
>>  {
>>      g_free(boot_splash_filedata);
>>      boot_splash_filedata = NULL;
>> +    cpu_exclusive_history_free();
>>  }
>>
>>  static int default_driver_check(void *opaque, QemuOpts *opts, Error **errp)
>> @@ -4322,6 +4323,8 @@ int main(int argc, char **argv, char **envp)
>>
>>      configure_accelerator(current_machine);
>>
>> +    cpu_exclusive_history_init();
>> +
>>      if (qtest_chrdev) {
>>          qtest_init(qtest_chrdev, qtest_log, &error_fatal);
>>      }
>
>
> --
> Alex Bennée

  reply	other threads:[~2016-02-18 14:17 UTC|newest]

Thread overview: 50+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2016-01-29  9:32 [Qemu-devel] [RFC v7 00/16] Slow-path for atomic instruction translation Alvise Rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 01/16] exec.c: Add new exclusive bitmap to ram_list Alvise Rigo
2016-02-11 13:00   ` Alex Bennée
2016-02-11 13:21     ` alvise rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 02/16] softmmu: Simplify helper_*_st_name, wrap unaligned code Alvise Rigo
2016-02-11 13:07   ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 03/16] softmmu: Simplify helper_*_st_name, wrap MMIO code Alvise Rigo
2016-02-11 13:15   ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 04/16] softmmu: Simplify helper_*_st_name, wrap RAM code Alvise Rigo
2016-02-11 13:18   ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 05/16] softmmu: Add new TLB_EXCL flag Alvise Rigo
2016-02-11 13:18   ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 06/16] qom: cpu: Add CPUClass hooks for exclusive range Alvise Rigo
2016-02-11 13:22   ` Alex Bennée
2016-02-18 13:53     ` alvise rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 07/16] softmmu: Add helpers for a new slowpath Alvise Rigo
2016-02-11 16:33   ` Alex Bennée
2016-02-18 13:58     ` alvise rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 08/16] softmmu: Honor the new exclusive bitmap Alvise Rigo
2016-02-16 17:39   ` Alex Bennée
2016-02-18 14:18     ` alvise rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 09/16] softmmu: Include MMIO/invalid exclusive accesses Alvise Rigo
2016-02-16 17:49   ` Alex Bennée
2016-02-18 14:18     ` alvise rigo
2016-02-18 16:26       ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 10/16] softmmu: Protect MMIO exclusive range Alvise Rigo
2016-02-17 18:55   ` Alex Bennée
2016-02-18 14:15     ` alvise rigo
2016-02-18 16:25       ` Alex Bennée
2016-03-07 18:13         ` alvise rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 11/16] tcg: Create new runtime helpers for excl accesses Alvise Rigo
2016-02-18 16:16   ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 12/16] configure: Use slow-path for atomic only when the softmmu is enabled Alvise Rigo
2016-02-18 16:40   ` Alex Bennée
2016-02-18 16:43     ` Alex Bennée
2016-03-07 17:21     ` alvise rigo
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 13/16] softmmu: Add history of excl accesses Alvise Rigo
2016-02-16 17:07   ` Alex Bennée
2016-02-18 14:17     ` alvise rigo [this message]
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 14/16] target-arm: translate: Use ld/st excl for atomic insns Alvise Rigo
2016-02-18 17:02   ` Alex Bennée
2016-03-07 18:39     ` alvise rigo
2016-03-07 20:06       ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 15/16] target-arm: cpu64: use custom set_excl hook Alvise Rigo
2016-02-18 18:19   ` Alex Bennée
2016-01-29  9:32 ` [Qemu-devel] [RFC v7 16/16] target-arm: aarch64: add atomic instructions Alvise Rigo
2016-02-19 11:34   ` Alex Bennée
2016-02-19 11:44 ` [Qemu-devel] [RFC v7 00/16] Slow-path for atomic instruction translation Alex Bennée
2016-02-19 12:01   ` alvise rigo
2016-02-19 12:19     ` Alex Bennée

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAH47eN2nLdXV4wkB28xCUXkX1jRv1N3OCtS6qdrzyGbB-tzVHA@mail.gmail.com \
    --to=a.rigo@virtualopensystems.com \
    --cc=alex.bennee@linaro.org \
    --cc=claudio.fontana@huawei.com \
    --cc=jani.kokkonen@huawei.com \
    --cc=mttcg@listserver.greensocs.com \
    --cc=pbonzini@redhat.com \
    --cc=qemu-devel@nongnu.org \
    --cc=rth@twiddle.net \
    --cc=tech@virtualopensystems.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).