| Message ID | 20191205161928.19548.41654.stgit@localhost.localdomain (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm / virtio: Provide support for free page reporting | expand |
On Thu, 2019-12-05 at 08:22 -0800, Alexander Duyck wrote: > This series provides an asynchronous means of reporting free guest pages > to a hypervisor so that the memory associated with those pages can be > dropped and reused by other processes and/or guests on the host. Using > this it is possible to avoid unnecessary I/O to disk and greatly improve > performance in the case of memory overcommit on the host. <snip> > Changes from v14: > https://lore.kernel.org/lkml/20191119214454.24996.66289.stgit@localhost.localdomain/ > Renamed "unused page reporting" to "free page reporting" > Updated code, kconfig, and patch descriptions > Split out patch for __free_isolated_page > Renamed function to __putback_isolated_page > Rewrote core reporting functionality > Added logic to reschedule worker in 2 seconds instead of run to completion > Removed reported_pages statistics > Removed REPORTING_REQUESTED bit used in zone flags > Replaced page_reporting_dev_info refcount with state variable > Removed scatterlist from page_reporting_dev_info > Removed capacity from page reporting device > Added dynamic scatterlist allocation/free at start/end of reporting process > Updated __free_one_page so that reported pages are not always added to tail > Added logic to handle error from report function > Updated virtio-balloon patch that adds support for page reporting > Updated patch description to try and highlight differences in approaches > Updated logic to reflect that we cannot limit the scatterlist from device > Added logic to return error from report function > Moved documentation patch to end of patch set It has been about a week since I posted v15 and haven't heard anything. Consider this a gentle ping. I'm looking for input on patches 3 and 4 in this set as I updated them to address most of the concerns Mel had. Just wondering if the set needs additional work or if we are good with this as a starting point for this feature? Thanks. - Alex
On 05.12.19 17:22, Alexander Duyck wrote: > This series provides an asynchronous means of reporting free guest pages > to a hypervisor so that the memory associated with those pages can be > dropped and reused by other processes and/or guests on the host. Using > this it is possible to avoid unnecessary I/O to disk and greatly improve > performance in the case of memory overcommit on the host. > > When enabled we will be performing a scan of free memory every 2 seconds > while pages of sufficiently high order are being freed. Currently the order > used is pageblock_order so that this feature will not interfere with the > use of Transparent Huge Pages in the case of virtualization. > > Currently this is only in use by virtio-balloon however there is the hope > that at some point in the future other hypervisors might be able to make > use of it. In the virtio-balloon/QEMU implementation the hypervisor is > currently using MADV_DONTNEED to indicate to the host kernel that the page > is currently free. It will be zeroed and faulted back into the guest the > next time the page is accessed. > > To track if a page is reported or not the Uptodate flag was repurposed and > used as a Reported flag for Buddy pages. We walk though the free list > isolating pages and adding them to the scatterlist until we either > encounter the end of the list, processed as many pages as were listed in > nr_free prior to us starting, or have filled the scatterlist with pages to > be reported. If we fill the scatterlist before we reach the end of the > list we rotate the list so that the first unreported page we encounter is > moved to the head of the list as that is where we will resume after we > have freed the reported pages back into the tail of the list. > > Below are the results from various benchmarks. I primarily focused on two > tests. The first is the will-it-scale/page_fault2 test, and the other is > a modified version of will-it-scale/page_fault1 that was enabled to use > THP. I did this as it allows for better visibility into different parts > of the memory subsystem. The guest is running with 32G for RAM on one > node of a E5-2630 v3. The host has had some power saving features disabled > by setting the /dev/cpu_dma_latency value to 10ms. > > Test page_fault1 (THP) page_fault2 > Name tasks Process Iter STDEV Process Iter STDEV > Baseline 1 1208307.25 0.10% 408596.00 0.19% > 16 8865204.75 0.16% 3344169.00 0.60% > > Patches applied 1 1206809.00 0.26% 412558.25 0.32% > 16 8814350.50 0.78% 3420102.00 1.16% > > Patches enabled 1 1201386.25 0.21% 407903.75 0.32% > 16 8880178.00 0.08% 3396700.50 0.54% > > Patches enabled 1 1173529.00 1.04% 409006.50 0.45% > page shuffle 16 8384540.25 0.74% 3288289.25 0.41% > > Patches enabled 1 1193411.00 0.33% 406333.50 0.09% > shuffle w/ RFC 16 8812639.75 0.73% 3321706.25 0.53% > > The results above are for a baseline with a linux-next-20191203 kernel, > that kernel with this patch set applied but page reporting disabled in > virtio-balloon, the patches applied and page reporting fully enabled, the > patches enabled with page shuffling enabled, and the patches applied with > page shuffling enabled and an RFC patch that makes used of MADV_FREE in > QEMU. These results include the deviation seen between the average value > reported here versus the high and/or low value. I observed that during the > test memory usage for the first three tests never dropped whereas with the > patches fully enabled the VM would drop to using only a few GB of the > host's memory when switching from memhog to page fault tests. > > Any of the overhead visible with this patch set enabled seems due to page > faults caused by accessing the reported pages and the host zeroing the page > before giving it back to the guest. This overhead is much more visible when > using THP than with standard 4K pages. In addition page shuffling seemed to > increase the amount of faults generated due to an increase in memory churn. > As seen in the data above, using MADV_FREE in QEMU mostly eliminates this > overhead. > > The overall guest size is kept fairly small to only a few GB while the test > is running. If the host memory were oversubscribed this patch set should > result in a performance improvement as swapping memory in the host can be > avoided. > > A brief history on the background of free page reporting can be found at: > https://lore.kernel.org/lkml/29f43d5796feed0dec8e8bb98b187d9dac03b900.camel@linux.intel.com/ > > Changes from v13: > https://lore.kernel.org/lkml/20191105215940.15144.65968.stgit@localhost.localdomain/ > Rewrote core reporting functionality > Merged patches 3 & 4 > Dropped boundary list and related code > Folded get_reported_page into page_reporting_fill > Folded page_reporting_fill into page_reporting_cycle > Pulled reporting functionality out of free_reported_page > Renamed it to __free_isolated_page > Moved page reporting specific bits to page_reporting_drain > Renamed phdev to prdev since we aren't "hinting" we are "reporting" > Added documentation to describe the usage of unused page reporting > Updated cover page and patch descriptions to avoid mention of boundary > > Changes from v14: > https://lore.kernel.org/lkml/20191119214454.24996.66289.stgit@localhost.localdomain/ > Renamed "unused page reporting" to "free page reporting" > Updated code, kconfig, and patch descriptions > Split out patch for __free_isolated_page > Renamed function to __putback_isolated_page > Rewrote core reporting functionality > Added logic to reschedule worker in 2 seconds instead of run to completion > Removed reported_pages statistics > Removed REPORTING_REQUESTED bit used in zone flags > Replaced page_reporting_dev_info refcount with state variable > Removed scatterlist from page_reporting_dev_info > Removed capacity from page reporting device > Added dynamic scatterlist allocation/free at start/end of reporting process > Updated __free_one_page so that reported pages are not always added to tail > Added logic to handle error from report function > Updated virtio-balloon patch that adds support for page reporting > Updated patch description to try and highlight differences in approaches > Updated logic to reflect that we cannot limit the scatterlist from device Last time Mel said "Ok, I'm ok with how this hooks into the allocator as the overhead is minimal. However, the patch itself still includes a number of optimisations instead of being a bare-boned implementation of the feature with optimisations layered on top." and "Either way, the separate patch could have supporting data on how much it improves the speed of reporting pages so it can be compared to any other optimisation that may be proposed. Supporting data would also help make the case that any complexity introduced by the optimisation is worthwhile." But I was only partially following that discussion. I can see that there is only one additional patch (before the reporting one) compared to the previous series on the MM side. Does that comment no longer apply (I can see e.g., "Removed REPORTING_REQUESTED bit used in zone flags" in the changelog) - IOW, did you drop all the optimizations in question for now? If so, can you share some performance differences with and without the previous optimizations? (just out of personal interest :) ) Christmas is getting closer, and at least in Europe/Germany that usually means that things will slow down ... or however you want to call that. So I wouldn't expect too much review happening before next year (but I might be wrong of course). Cheers!
On Fri, Dec 13, 2019 at 11:00:42AM +0100, David Hildenbrand wrote: > > A brief history on the background of free page reporting can be found at: > > https://lore.kernel.org/lkml/29f43d5796feed0dec8e8bb98b187d9dac03b900.camel@linux.intel.com/ > > > > Changes from v13: > > https://lore.kernel.org/lkml/20191105215940.15144.65968.stgit@localhost.localdomain/ > > Rewrote core reporting functionality > > Merged patches 3 & 4 > > Dropped boundary list and related code > > Folded get_reported_page into page_reporting_fill > > Folded page_reporting_fill into page_reporting_cycle > > Pulled reporting functionality out of free_reported_page > > Renamed it to __free_isolated_page > > Moved page reporting specific bits to page_reporting_drain > > Renamed phdev to prdev since we aren't "hinting" we are "reporting" > > Added documentation to describe the usage of unused page reporting > > Updated cover page and patch descriptions to avoid mention of boundary > > > > Changes from v14: > > https://lore.kernel.org/lkml/20191119214454.24996.66289.stgit@localhost.localdomain/ > > Renamed "unused page reporting" to "free page reporting" > > Updated code, kconfig, and patch descriptions > > Split out patch for __free_isolated_page > > Renamed function to __putback_isolated_page > > Rewrote core reporting functionality > > Added logic to reschedule worker in 2 seconds instead of run to completion > > Removed reported_pages statistics > > Removed REPORTING_REQUESTED bit used in zone flags > > Replaced page_reporting_dev_info refcount with state variable > > Removed scatterlist from page_reporting_dev_info > > Removed capacity from page reporting device > > Added dynamic scatterlist allocation/free at start/end of reporting process > > Updated __free_one_page so that reported pages are not always added to tail > > Added logic to handle error from report function > > Updated virtio-balloon patch that adds support for page reporting > > Updated patch description to try and highlight differences in approaches > > Updated logic to reflect that we cannot limit the scatterlist from device > > Last time Mel said > > "Ok, I'm ok with how this hooks into the allocator as the overhead is > minimal. However, the patch itself still includes a number of > optimisations instead of being a bare-boned implementation of the > feature with optimisations layered on top." > I didn't get the chance to take a close look as I'm trying to clear as much as possible from my table on the run-up to Christmas so I don't come back to a disaster inbox. I also noted that the Acks for earlier patches were not included so I was uncertain if doing a full review would still be a good use of time when time was tight. That said, some optimisations are still included but much reduced. For example, list rotations are still there but it's very straight-forward. The refcount is gone which is good and replaced by a state, which could be be better documented, but is more straight forward and the zone->lock is back protecting the free lists primarily and not zone metadata or prdev metadata (at least not obviously). I didn't put in the time to see if the atomic_set in page_reporting_process() is ok or whether state could be lost but I *think* it's ok because it should be called from just one workqueue request and they shouldn't be stacked. A comment there explaining why atomic_set is definitely correct would be helpful. I'm inclined to decide that yes, this version is potentially ok as a bare minimum but didn't put in the time to be 100% sure.
On Fri, 2019-12-13 at 11:00 +0100, David Hildenbrand wrote: > On 05.12.19 17:22, Alexander Duyck wrote: > > This series provides an asynchronous means of reporting free guest pages > > to a hypervisor so that the memory associated with those pages can be > > dropped and reused by other processes and/or guests on the host. Using > > this it is possible to avoid unnecessary I/O to disk and greatly improve > > performance in the case of memory overcommit on the host. > > > > When enabled we will be performing a scan of free memory every 2 seconds > > while pages of sufficiently high order are being freed. Currently the order > > used is pageblock_order so that this feature will not interfere with the > > use of Transparent Huge Pages in the case of virtualization. > > > > Currently this is only in use by virtio-balloon however there is the hope > > that at some point in the future other hypervisors might be able to make > > use of it. In the virtio-balloon/QEMU implementation the hypervisor is > > currently using MADV_DONTNEED to indicate to the host kernel that the page > > is currently free. It will be zeroed and faulted back into the guest the > > next time the page is accessed. > > > > To track if a page is reported or not the Uptodate flag was repurposed and > > used as a Reported flag for Buddy pages. We walk though the free list > > isolating pages and adding them to the scatterlist until we either > > encounter the end of the list, processed as many pages as were listed in > > nr_free prior to us starting, or have filled the scatterlist with pages to > > be reported. If we fill the scatterlist before we reach the end of the > > list we rotate the list so that the first unreported page we encounter is > > moved to the head of the list as that is where we will resume after we > > have freed the reported pages back into the tail of the list. > > > > Below are the results from various benchmarks. I primarily focused on two > > tests. The first is the will-it-scale/page_fault2 test, and the other is > > a modified version of will-it-scale/page_fault1 that was enabled to use > > THP. I did this as it allows for better visibility into different parts > > of the memory subsystem. The guest is running with 32G for RAM on one > > node of a E5-2630 v3. The host has had some power saving features disabled > > by setting the /dev/cpu_dma_latency value to 10ms. > > > > Test page_fault1 (THP) page_fault2 > > Name tasks Process Iter STDEV Process Iter STDEV > > Baseline 1 1208307.25 0.10% 408596.00 0.19% > > 16 8865204.75 0.16% 3344169.00 0.60% > > > > Patches applied 1 1206809.00 0.26% 412558.25 0.32% > > 16 8814350.50 0.78% 3420102.00 1.16% > > > > Patches enabled 1 1201386.25 0.21% 407903.75 0.32% > > 16 8880178.00 0.08% 3396700.50 0.54% > > > > Patches enabled 1 1173529.00 1.04% 409006.50 0.45% > > page shuffle 16 8384540.25 0.74% 3288289.25 0.41% > > > > Patches enabled 1 1193411.00 0.33% 406333.50 0.09% > > shuffle w/ RFC 16 8812639.75 0.73% 3321706.25 0.53% > > > > The results above are for a baseline with a linux-next-20191203 kernel, > > that kernel with this patch set applied but page reporting disabled in > > virtio-balloon, the patches applied and page reporting fully enabled, the > > patches enabled with page shuffling enabled, and the patches applied with > > page shuffling enabled and an RFC patch that makes used of MADV_FREE in > > QEMU. These results include the deviation seen between the average value > > reported here versus the high and/or low value. I observed that during the > > test memory usage for the first three tests never dropped whereas with the > > patches fully enabled the VM would drop to using only a few GB of the > > host's memory when switching from memhog to page fault tests. > > > > Any of the overhead visible with this patch set enabled seems due to page > > faults caused by accessing the reported pages and the host zeroing the page > > before giving it back to the guest. This overhead is much more visible when > > using THP than with standard 4K pages. In addition page shuffling seemed to > > increase the amount of faults generated due to an increase in memory churn. > > As seen in the data above, using MADV_FREE in QEMU mostly eliminates this > > overhead. > > > > The overall guest size is kept fairly small to only a few GB while the test > > is running. If the host memory were oversubscribed this patch set should > > result in a performance improvement as swapping memory in the host can be > > avoided. > > > > A brief history on the background of free page reporting can be found at: > > https://lore.kernel.org/lkml/29f43d5796feed0dec8e8bb98b187d9dac03b900.camel@linux.intel.com/ > > > > Changes from v13: > > https://lore.kernel.org/lkml/20191105215940.15144.65968.stgit@localhost.localdomain/ > > Rewrote core reporting functionality > > Merged patches 3 & 4 > > Dropped boundary list and related code > > Folded get_reported_page into page_reporting_fill > > Folded page_reporting_fill into page_reporting_cycle > > Pulled reporting functionality out of free_reported_page > > Renamed it to __free_isolated_page > > Moved page reporting specific bits to page_reporting_drain > > Renamed phdev to prdev since we aren't "hinting" we are "reporting" > > Added documentation to describe the usage of unused page reporting > > Updated cover page and patch descriptions to avoid mention of boundary > > > > Changes from v14: > > https://lore.kernel.org/lkml/20191119214454.24996.66289.stgit@localhost.localdomain/ > > Renamed "unused page reporting" to "free page reporting" > > Updated code, kconfig, and patch descriptions > > Split out patch for __free_isolated_page > > Renamed function to __putback_isolated_page > > Rewrote core reporting functionality > > Added logic to reschedule worker in 2 seconds instead of run to completion > > Removed reported_pages statistics > > Removed REPORTING_REQUESTED bit used in zone flags > > Replaced page_reporting_dev_info refcount with state variable > > Removed scatterlist from page_reporting_dev_info > > Removed capacity from page reporting device > > Added dynamic scatterlist allocation/free at start/end of reporting process > > Updated __free_one_page so that reported pages are not always added to tail > > Added logic to handle error from report function > > Updated virtio-balloon patch that adds support for page reporting > > Updated patch description to try and highlight differences in approaches > > Updated logic to reflect that we cannot limit the scatterlist from device > > Last time Mel said > > "Ok, I'm ok with how this hooks into the allocator as the overhead is > minimal. However, the patch itself still includes a number of > optimisations instead of being a bare-boned implementation of the > feature with optimisations layered on top." > > and > > "Either way, the separate patch could have supporting data on how much > it improves the speed of reporting pages so it can be compared to any > other optimisation that may be proposed. Supporting data would also help > make the case that any complexity introduced by the optimisation is > worthwhile." > > But I was only partially following that discussion. > > I can see that there is only one additional patch (before the reporting > one) compared to the previous series on the MM side. Does that comment > no longer apply (I can see e.g., "Removed REPORTING_REQUESTED bit used > in zone flags" in the changelog) - IOW, did you drop all the > optimizations in question for now? If so, can you share some performance > differences with and without the previous optimizations? (just out of > personal interest :) ) I found alternative approaches and did away with most of them. What I had done is replaced the REPORTING_REQUESTED and reference count logic with the state in order to guarantee that we will make our way through the list and rearm the reporting thread if the work isn't completed. One thing I still think I need to split out based on Mel's comments is the list rotation and probably the new budget value I added. As far as performance this new patch set performs better than the old one did. Most of that is due to the fact that I increased the delay between passes and dropped any optimizations for the shuffling code. > Christmas is getting closer, and at least in Europe/Germany that usually > means that things will slow down ... or however you want to call that. > So I wouldn't expect too much review happening before next year (but I > might be wrong of course). > > Cheers! That is one of the reasons why I wanted to see if there were any comments I could get before the break. It gives me a chance to address them and push one last patch set before I head out on Christmas break myself.
On Fri, 2019-12-13 at 11:08 +0000, Mel Gorman wrote: > On Fri, Dec 13, 2019 at 11:00:42AM +0100, David Hildenbrand wrote: > > > A brief history on the background of free page reporting can be found at: > > > https://lore.kernel.org/lkml/29f43d5796feed0dec8e8bb98b187d9dac03b900.camel@linux.intel.com/ > > > > > > Changes from v13: > > > https://lore.kernel.org/lkml/20191105215940.15144.65968.stgit@localhost.localdomain/ > > > Rewrote core reporting functionality > > > Merged patches 3 & 4 > > > Dropped boundary list and related code > > > Folded get_reported_page into page_reporting_fill > > > Folded page_reporting_fill into page_reporting_cycle > > > Pulled reporting functionality out of free_reported_page > > > Renamed it to __free_isolated_page > > > Moved page reporting specific bits to page_reporting_drain > > > Renamed phdev to prdev since we aren't "hinting" we are "reporting" > > > Added documentation to describe the usage of unused page reporting > > > Updated cover page and patch descriptions to avoid mention of boundary > > > > > > Changes from v14: > > > https://lore.kernel.org/lkml/20191119214454.24996.66289.stgit@localhost.localdomain/ > > > Renamed "unused page reporting" to "free page reporting" > > > Updated code, kconfig, and patch descriptions > > > Split out patch for __free_isolated_page > > > Renamed function to __putback_isolated_page > > > Rewrote core reporting functionality > > > Added logic to reschedule worker in 2 seconds instead of run to completion > > > Removed reported_pages statistics > > > Removed REPORTING_REQUESTED bit used in zone flags > > > Replaced page_reporting_dev_info refcount with state variable > > > Removed scatterlist from page_reporting_dev_info > > > Removed capacity from page reporting device > > > Added dynamic scatterlist allocation/free at start/end of reporting process > > > Updated __free_one_page so that reported pages are not always added to tail > > > Added logic to handle error from report function > > > Updated virtio-balloon patch that adds support for page reporting > > > Updated patch description to try and highlight differences in approaches > > > Updated logic to reflect that we cannot limit the scatterlist from device > > > > Last time Mel said > > > > "Ok, I'm ok with how this hooks into the allocator as the overhead is > > minimal. However, the patch itself still includes a number of > > optimisations instead of being a bare-boned implementation of the > > feature with optimisations layered on top." > > > > I didn't get the chance to take a close look as I'm trying to clear as > much as possible from my table on the run-up to Christmas so I don't come > back to a disaster inbox. I also noted that the Acks for earlier patches > were not included so I was uncertain if doing a full review would still > be a good use of time when time was tight. Sorry about that. I will go back through and make sure to collect the Acks on the earlier patches. I guess I had overlooked them while focusing on rewriting the core functionality. > That said, some optimisations are still included but much reduced. For > example, list rotations are still there but it's very straight-forward. I will go ahead and split the rotations out into a separate patch for v16. I can probably do that and pull the budget bit I had added out and put it together as a "work conserving/limiting" optimization for the patch set. > The refcount is gone which is good and replaced by a state, which could be > be better documented, but is more straight forward and the zone->lock is > back protecting the free lists primarily and not zone metadata or prdev > metadata (at least not obviously). I didn't put in the time to see if > the atomic_set in page_reporting_process() is ok or whether state could > be lost but I *think* it's ok because it should be called from just one > workqueue request and they shouldn't be stacked. A comment there explaining > why atomic_set is definitely correct would be helpful. I will go though and add some more documentation about the state. > I'm inclined to decide that yes, this version is potentially ok as a > bare minimum but didn't put in the time to be 100% sure. Sounds good. I will go through and address the concerns you brought up, and probably post a v16 by the end of next week. Thanks for the feedback. - Alex
> I found alternative approaches and did away with most of them. What I had > done is replaced the REPORTING_REQUESTED and reference count logic with > the state in order to guarantee that we will make our way through the list > and rearm the reporting thread if the work isn't completed. > > One thing I still think I need to split out based on Mel's comments is the > list rotation and probably the new budget value I added. Yeah, both probably make sense. > > As far as performance this new patch set performs better than the old one > did. Most of that is due to the fact that I increased the delay between > passes and dropped any optimizations for the shuffling code. That's interesting - and good :) I probably won't have enough time to look into #4 before Christmas - so I might wait for your resend and review after Christmas.
This series provides an asynchronous means of reporting free guest pages to a hypervisor so that the memory associated with those pages can be dropped and reused by other processes and/or guests on the host. Using this it is possible to avoid unnecessary I/O to disk and greatly improve performance in the case of memory overcommit on the host. When enabled we will be performing a scan of free memory every 2 seconds while pages of sufficiently high order are being freed. Currently the order used is pageblock_order so that this feature will not interfere with the use of Transparent Huge Pages in the case of virtualization. Currently this is only in use by virtio-balloon however there is the hope that at some point in the future other hypervisors might be able to make use of it. In the virtio-balloon/QEMU implementation the hypervisor is currently using MADV_DONTNEED to indicate to the host kernel that the page is currently free. It will be zeroed and faulted back into the guest the next time the page is accessed. To track if a page is reported or not the Uptodate flag was repurposed and used as a Reported flag for Buddy pages. We walk though the free list isolating pages and adding them to the scatterlist until we either encounter the end of the list, processed as many pages as were listed in nr_free prior to us starting, or have filled the scatterlist with pages to be reported. If we fill the scatterlist before we reach the end of the list we rotate the list so that the first unreported page we encounter is moved to the head of the list as that is where we will resume after we have freed the reported pages back into the tail of the list. Below are the results from various benchmarks. I primarily focused on two tests. The first is the will-it-scale/page_fault2 test, and the other is a modified version of will-it-scale/page_fault1 that was enabled to use THP. I did this as it allows for better visibility into different parts of the memory subsystem. The guest is running with 32G for RAM on one node of a E5-2630 v3. The host has had some power saving features disabled by setting the /dev/cpu_dma_latency value to 10ms. Test page_fault1 (THP) page_fault2 Name tasks Process Iter STDEV Process Iter STDEV Baseline 1 1208307.25 0.10% 408596.00 0.19% 16 8865204.75 0.16% 3344169.00 0.60% Patches applied 1 1206809.00 0.26% 412558.25 0.32% 16 8814350.50 0.78% 3420102.00 1.16% Patches enabled 1 1201386.25 0.21% 407903.75 0.32% 16 8880178.00 0.08% 3396700.50 0.54% Patches enabled 1 1173529.00 1.04% 409006.50 0.45% page shuffle 16 8384540.25 0.74% 3288289.25 0.41% Patches enabled 1 1193411.00 0.33% 406333.50 0.09% shuffle w/ RFC 16 8812639.75 0.73% 3321706.25 0.53% The results above are for a baseline with a linux-next-20191203 kernel, that kernel with this patch set applied but page reporting disabled in virtio-balloon, the patches applied and page reporting fully enabled, the patches enabled with page shuffling enabled, and the patches applied with page shuffling enabled and an RFC patch that makes used of MADV_FREE in QEMU. These results include the deviation seen between the average value reported here versus the high and/or low value. I observed that during the test memory usage for the first three tests never dropped whereas with the patches fully enabled the VM would drop to using only a few GB of the host's memory when switching from memhog to page fault tests. Any of the overhead visible with this patch set enabled seems due to page faults caused by accessing the reported pages and the host zeroing the page before giving it back to the guest. This overhead is much more visible when using THP than with standard 4K pages. In addition page shuffling seemed to increase the amount of faults generated due to an increase in memory churn. As seen in the data above, using MADV_FREE in QEMU mostly eliminates this overhead. The overall guest size is kept fairly small to only a few GB while the test is running. If the host memory were oversubscribed this patch set should result in a performance improvement as swapping memory in the host can be avoided. A brief history on the background of free page reporting can be found at: https://lore.kernel.org/lkml/29f43d5796feed0dec8e8bb98b187d9dac03b900.camel@linux.intel.com/ Changes from v13: https://lore.kernel.org/lkml/20191105215940.15144.65968.stgit@localhost.localdomain/ Rewrote core reporting functionality Merged patches 3 & 4 Dropped boundary list and related code Folded get_reported_page into page_reporting_fill Folded page_reporting_fill into page_reporting_cycle Pulled reporting functionality out of free_reported_page Renamed it to __free_isolated_page Moved page reporting specific bits to page_reporting_drain Renamed phdev to prdev since we aren't "hinting" we are "reporting" Added documentation to describe the usage of unused page reporting Updated cover page and patch descriptions to avoid mention of boundary Changes from v14: https://lore.kernel.org/lkml/20191119214454.24996.66289.stgit@localhost.localdomain/ Renamed "unused page reporting" to "free page reporting" Updated code, kconfig, and patch descriptions Split out patch for __free_isolated_page Renamed function to __putback_isolated_page Rewrote core reporting functionality Added logic to reschedule worker in 2 seconds instead of run to completion Removed reported_pages statistics Removed REPORTING_REQUESTED bit used in zone flags Replaced page_reporting_dev_info refcount with state variable Removed scatterlist from page_reporting_dev_info Removed capacity from page reporting device Added dynamic scatterlist allocation/free at start/end of reporting process Updated __free_one_page so that reported pages are not always added to tail Added logic to handle error from report function Updated virtio-balloon patch that adds support for page reporting Updated patch description to try and highlight differences in approaches Updated logic to reflect that we cannot limit the scatterlist from device Added logic to return error from report function Moved documentation patch to end of patch set --- Alexander Duyck (7): mm: Adjust shuffle code to allow for future coalescing mm: Use zone and order instead of free area in free_list manipulators mm: Add function __putback_isolated_page mm: Introduce Reported pages virtio-balloon: Pull page poisoning config out of free page hinting virtio-balloon: Add support for providing free page reports to host mm: Add free page reporting documentation Documentation/vm/free_page_reporting.rst | 41 ++++ drivers/virtio/Kconfig | 1 drivers/virtio/virtio_balloon.c | 87 +++++++- include/linux/mmzone.h | 44 ---- include/linux/page-flags.h | 11 + include/linux/page_reporting.h | 25 ++ include/uapi/linux/virtio_balloon.h | 1 mm/Kconfig | 11 + mm/Makefile | 1 mm/internal.h | 1 mm/page_alloc.c | 169 +++++++++++---- mm/page_isolation.c | 6 - mm/page_reporting.c | 336 ++++++++++++++++++++++++++++++ mm/page_reporting.h | 54 +++++ mm/shuffle.c | 12 + mm/shuffle.h | 6 + 16 files changed, 700 insertions(+), 106 deletions(-) create mode 100644 Documentation/vm/free_page_reporting.rst create mode 100644 include/linux/page_reporting.h create mode 100644 mm/page_reporting.c create mode 100644 mm/page_reporting.h --